Loan supply, credit markets and the euro area financial crisis

Journal Pre-proof

Loan supply, credit markets and the euro area financial crisis Carlo Altavilla , Matthieu Darracq Paries , Giulio Nicoletti PII: DOI: Reference:

S0378-4266(19)30233-X https://doi.org/10.1016/j.jbankfin.2019.105658 JBF 105658

To appear in:

Journal of Banking and Finance

Received date: Accepted date:

4 April 2017 23 September 2019

Please cite this article as: Carlo Altavilla , Matthieu Darracq Paries , Giulio Nicoletti , Loan supply, credit markets and the euro area financial crisis, Journal of Banking and Finance (2019), doi: https://doi.org/10.1016/j.jbankfin.2019.105658

This is a PDF file of an article that has undergone enhancements after acceptance, such as the addition of a cover page and metadata, and formatting for readability, but it is not yet the definitive version of record. This version will undergo additional copyediting, typesetting and review before it is published in its final form, but we are providing this version to give early visibility of the article. Please note that, during the production process, errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain. © 2019 Published by Elsevier B.V.

Loan supply, credit markets and the euro area financial crisis *

Carlo Altavilla European Central Bank

Matthieu Darracq Paries European Central Bank

Giulio Nicoletti European Central Bank

Abstract We derive a measure of loan supply shocks from proprietary bank-level information on credit standards from the euro area Bank Lending Survey (BLS) controlling for both macroeconomic and bank-specific demand factors. Using this indicator as an external instrument in a Bayesian vector autoregressive (BVAR) model, we find that a tightening of credit standards – i.e. banks‟ internal guidelines or loan approval criteria – leads to a protracted contraction in credit volumes intermediated by banks and higher lending margins. This fosters firms‟ incentives to substitute bank loans with market finance, ultimately producing a significant increase in debt securities issuance and higher corporate bond spreads. We also show that widely-used measures of financial uncertainty do not influence or drive our results.

JEL classification: E51, E44, C32 Keywords: Credit supply, non-bank financing, lending standards, bank lending survey, external instruments

We would like to thank Martin Bodenstein, Karel Mertens and Haroon Mumtaz, as well as some participants in the 30th Annual Congress of the European Economic Association, the International Conference on Computational and Financial Econometrics, and the 2nd Annual Conference of the International Association for Applied Econometrics for their comments and suggestions. Anca-Mihaela Paraschiv provided excellent research assistance. All errors and omissions are our own responsibility. The opinions in this paper are those of the authors and do not necessarily reflect the views of the European Central Bank and the Eurosystem. Please address any comments to Carlo Altavilla, [email protected]; Matthieu Darracq Paries, [email protected]; or Giulio Nicoletti, [email protected]. *

1 Introduction This paper analyses how changes in bank loan supply influence firm decisions and real economic outcomes. In principle, firms might react to a contraction in bank lending by either reducing their activity and/or seeking alternative sources of funding through corporate bond markets. Given the prominent role played by credit institutions in financing euro area firms, the change in firms‟ behaviour following bank supply shocks has important implications for both monetary policy and financial stability. Identifying loan supply shocks is complicated by the presence of many confounding factors that can lead to changes in credit volumes not necessarily driven by changes in loan supply. For instance, the significant contraction in credit and the concomitant surge in firms‟ borrowing costs observed over the recent recessionary episodes might be consistent with a restriction in the supply of bank loans.1 However, many different factors mostly associated with general economic conditions can also be at play. In a context of falling aggregate demand and poor macroeconomic performance banks might be induced to protect their balance sheet by tightening their loan approval criteria. In these circumstances, the contraction in credit would be therefore driven by demand (rather than supply) conditions and banks could be seen as responding to adverse economic conditions rather than generating them. These considerations suggest that in order to assess the macroeconomic impact of credit supply shocks, it is crucial to correctly identify changes in credit origination due to bank lending decisions not related to other confounding factors. In this paper, we use unique bank-level information on credit standards from the ECB‟s bank lending survey (BLS) to construct an indicator of loan supply shocks for the euro area which we call loan supply indicator (LSI). Specifically, the LSI is derived from changes in bank credit standards which are orthogonal to bank-specific demand factors and already prevailing macro-financial conditions. Importantly, the LSI accounts for both easing and tightening of loan supply. This indicator is then used as an external instrument in a Bayesian VAR model to identify the impact of loan supply shocks on real economic activity. Our results show that an adverse loan supply shock leads to a protracted contraction in lending volumes and a marked increase in bank lending margins. A historical decomposition indicates that this shock is able to explain a significant share of observed movements in prices, quantities and economic activity over the two latest euro area recessions.

The first recession we refer to started in 2008Q1 and took place in the context of the post-Lehman global financial crisis until 2009Q2. The second recession started in 2011Q3 and ended in 2013Q1. See Appendix 1 for some stylised facts on the euro area economy. 1

2

In addition, the paper empirically investigates the use of alternative sources of firms‟ external financing (i.e. bonds versus loans) in periods of financial distress. We find that a reduction in loan supply fosters firms‟ incentives to substitute bank loans with bonds thereby producing a significant increase in debt securities issuance and higher bond spreads (as in Adrian et al. 2012). Historical decomposition analysis shows that, especially during the Great Recession, loan supply tightening explains both the increases in credit spreads and the stronger bond issuance by firms. The substitution between loans and bonds is found to be only partial (as in Crouzet, 2018; and Becker and Ivashina, 2014). Results are robust to considering alternative shocks such as the variance risk premium (Bekaert and Hoerova, 2014), the VIX, the news shock (Barsky and Sims 2011), an uncertainty indicator based on gold prices (Piffer and Podstawski 2018), and a measure of excess bond premium (EBP) for the euro area that we construct following the seminal contributions of Gilchrist et al. (2009) and Gilchrist and Zakrajšek (2012a, 2012b, 2013). The paper relates and contributes to two main strands of literature. The first one focuses on the identification of the macroeconomic effects of credit supply shocks identified by bank lending standards as reported in surveys. Bassett et al. (2014) use information from the Senior Loan Officer Opinion Survey collected by the Federal Reserve to construct a credit supply indicator. Lown and Morgan (2006) study the macroeconomic impact of a shock to lending standards using information contained in the same survey. Ciccarelli et al. (2013, 2015) estimate a panel VAR including variables from the BLS aggregated at country level and found that, especially in financially distressed countries, the credit channel (both the bank lending and the non-financial borrower balance sheet channel) might act as an amplification mechanism for the impact of monetary policy on the real economy. Compared with this strand of literature, we are the first to rely on individual bank data therefore adopting a more granular approach, in the context of the euro area. The second strand of literature focuses on the alternative financing sources of non-financial firms and their choice between bank and non-bank financing in periods of financial distress. Early studies on the endogenous choice between banks and market finance (Holmstrom and Tirole, 1997; Repullo and Suarez, 2000) argue that a contraction in firms‟ net worth, as the one observed during the crisis, should lead to a shift from bond to more bank finance. Becker and Ivashina (2014) highlight instead that banks‟ credit tightening can induce firms to substitute bank credit with additional issuance of corporate bonds. A theoretical rationale for this evidence is proposed by Adrian et al. (2012). In their model, banks follow a “Value-at-Risk” approach. When the default risk of NFCs increases, the bank‟s optimal choice is to delever sharply and thus reduce lending. Given that the demand for credit from NFCs has limited elasticity, risk-averse bond investors need to be

3

encouraged to increase their credit supply. This also requires a widening of the spreads on corporate bonds. More recently, Crouzet (2018) also provides an analysis where firms hit by loan supply restrictions substitute bank finance with bond finance, but not completely. The substitution between bonds and loans is also described in De Fiore and Uhlig (2011) where the shift from bank loans to bonds can be the result of NFCs‟ optimal financing choices in the face of a negative bank supply shock. We lend further empirical support to these studies by quantifying how loan supply disruptions influence bond issuance in the euro area. The rest of the paper is organised as follows. In Section 2 we exploit bank-level information in the BLS to construct an indicator of loan supply restrictions independent from prevailing economic conditions. Section 3 provides evidence on the macroeconomic effects of loan supply shocks using the new loan supply indicator as an external instrument in a BVAR model. Sensitivity analysis to the instrument choice is also discussed. Section 4 extend the framework to the largest euro area countries and derive macroeconomic implications. Section 5 provides concluding remarks.

2 A proxy for loan supply shocks We use individual bank-level information available from the bank lending survey (BLS) maintained at the European Central Bank: most central banks, such as the Federal Reserve and the Bank of England, conduct similar surveys to extract soft information on banking sector behaviour.. In the BLS, a representative sample of euro area banks2 is asked, among other things, about the credit standards (CS) they apply for the provision of loans to non-financial firms: we consider CS as source of information to construct our indicator on loan supply conditions (see Basset et al., 2014; and Lown and Morgan, 2006). Answers to the CS question are typically displayed in the form of a net diffusion index of credit standards. However, relying on the simple net diffusion index of credit standards to gauge credit supply factors is problematic since such an index captures many factors affecting credit besides pure credit supply effects. A visual comparison between the CS net diffusion index and the corresponding index on credit demand factors 3 shows that the two indices are highly (negatively) correlated (Figure 1, left panel). This feature is common also to the surveys conducted in the UK and US (Figure 1, middle and right panels): the decrease (increase) in the net demand for loans to enterprises is associated with the net tightening (easing) of credit standards. This evidence suggests that the macroeconomic effect of credit supply shocks might be difficult to identify as many factors simultaneously influence both the demand for and the supply of credit. See Appendix 2 for a more detailed description of the bank lending survey. Changes in the demand for loans or credit lines to enterprises as expressed by the question: “Over the past three months, how has the demand for loans or credit lines to enterprises changed at your bank, apart from normal seasonal fluctuations?” 2 3

4

Figure 1: Net tightening of credit standards and net demand for loans to enterprises according to the surveys in the euro area, UK and US (percentages of banks, inverted scale for demand)

Note: The figure reports the results on credit supply and demand contained in the bank lending survey for the euro area, the Credit Conditions Survey in the UK, and the Senior Loan Officer Opinion Survey on Bank Lending Practices for the US. The blue solid line depicts the change in overall bank lending standards (net tightening) as proxy for credit supply; the red dashed line depicts the diffusion index for overall loan demand. A positive value of the diffusion index indicates a net tightening/increase in credit standards/demand, while a negative value indicates a net easing/decrease in credit standards/demand. The sample period extends from Q1 2003 to Q4 2016.

To deal with this endogeneity issue, we exploit bank-level information in the BLS to extract our innovations of pure credit supply factors, i.e. shocks that affect credit supply but are independent from already prevailing economic conditions. Soft information from surveys has been used in previous studies (see Ciccarelli et al., 2013 and 2015, for the euro area, and Bassett et al., 2014, for the US) to disentangle credit supply from demand. Unlike in Ciccarelli et al. (2013), this paper uses granular information (i.e. bank-level responses) contained in the BLS to identify the effects of credit supply disturbances on the euro area economy. Similar to our study, Basset et al. (2014) use individual data from the Senior Loan Officer Opinion Survey. However, as the BLS is fully anonymised at bank level, we cannot match banks‟ balance sheet items to our individual BLS responses as proposed in Basset et al. (2014). To disentangle the possible causal effect of a tightening in credit standards, we borrow from the literature on treatment effects (Imbens, 2004) and use an inverse propensity score method. We formulate a propensity score model explaining how likely it is for each bank to tighten credit based on prevailing economic conditions; the model includes bank-level information from the BLS as well as country-level and euro area-wide economic conditions. This information is then used to reweight each bank‟s response so as to mimic the conditions of a randomised experiment. The measure obtained by reweighting the individual credit standard responses is labelled the loan supply indicator measure for each individual bank “i”, i.e.

5

. Finally, the individual

are aggregated

across banks and across countries to obtain a measure of changes in lending standards for the euro area. The aggregation method employed is the same as the one used to obtain the CS net diffusion index for the BLS. We assume a dummy variable D, which corresponds to the different changes in credit conditions: D can take values

according to whether conditions ease, there is no change, or

credit conditions tighten, respectively. The number of categorical answers in the BLS is reduced to three cases, down from the original five. In fact, although the BLS‟s answers range from “strong easing” to “strong tightening”, most of the individual responses are concentrated in practice in three bins (see Figures A.2 and A.3 in Appendix 2). We then group the BLS‟s answers according to three main categories. The first one, called “easing”, considers “easing” and “strong easing”; the second one, “tightening”, jointly considers the answers of “tightening” and “strong tightening”; and the third – “neutral” – considers the answer “no change”. We can then consider a true “tightening” or “easing” of bank credit supply as a treatment (with respect to the no-change case) that influences what banks declare. More specifically, the true “tightening” of credit standards is assimilated to the treatment a patient receives and whose causal effect we aim at estimating when observing the outcome of declared “tightening” from the survey. The indicator variable

(

) denotes the setting of the supply-side

treatment, while the credit standard declared by banks is denoted CS. The three “potential outcomes” are

(

), with the average treatment effect (ATE) of either ease (j = -1) or

tighten (j=1) being the expected difference between the two counterfactual paths of treatment versus no treatment: (

)

(

)

(1)

The lack of treatment randomisation leads to selection bias. Indeed, if the selection of treated and non-treated units were made as in a fully randomised experiment, no endogeneity issue would arise and the net diffusion index of credit standards would resemble a measure of the ATE. However, as banks are likely to tighten or ease credit in a way which is driven by determinants that also influence the potential outcomes, the assumption valid for a randomised experiment cannot hold. To overcome this problem, as usually done in this literature, the key assumption we make is that the experiment is random conditional on a set of observable factors

. This is the “selection on

observables” condition (see Imbens 2004): ( )

(2)

6

The treatment (or no treatment) is orthogonal to the potential outcomes, once we control for factors

. When equation (2) holds, one can use the inverse propensity score estimator to recover

the causal effect of treatment on observed variables.4 The new credit standard indicator (or „loan supply indicator‟) of bank i for event j, where j = -1, 1 (easing or tightening), is denoted as

and it is computed as the inverse propensity weighted

estimator of each bank response (see Imbens, 2004). This corresponds to: ( ( (

where (

))

( ( (

)

)) )

-

) is the indicator function of a (single) bank declaring tightening and (

(3) ) is

the probability (or propensity score) that the bank declares tightening given a set of conditioning variables Z. To estimate the propensity score of the different cases, we use an ordered probit as a model for the probability distribution of tightening versus easing. Angrist et al. (2011) and Angrist et al. (2013) provide a similar application of ordered probit but to the case where monetary policy rates are raised or lowered. The main determinants Z of individual credit standards considered in the estimation are the bank-level demand factor from the BLS, GDP growth and unemployment of the country of the reporting bank, expectations about economic activity as reported by Consensus Economics, VSTOXX (i.e. the Dow Jones EURO STOXX 50 volatility measured for the euro area), a proxy for the monetary policy stance (as consistent with the risk-taking channel; see Maddaloni and Peydrò, 2013) both in terms of the current stance (proxied by the EONIA spot rate) and in terms of the expected future stance (summarised by the three-month overnight index swap forward rate in one year). Gertler and Karadi (2015) use a similar forward rate for the US to capture both standard and non-standard monetary policy. Finally, we incorporate an EBP for euro area NFCs as developed for the United States by Gilchrist et al. (2012a).5 In robustness checks we also include the first lag of the BLS credit standards. Table 1 provides the estimates of the marginal effects of our determinants with standard errors clustered at bank level. In the simplest model, individual bank demand factors are a significant determinant of credit standards. In particular, individual demand conditions contribute to increasing the probability of having a tightening by some 8 percentage points (p.p.). The specification in the column denoted by “No monetary policy” shows that both actual and expected economic conditions also significantly contribute to explaining tightening of lending standards, as well as financial factors. When a country experiences adverse economic conditions, unemployment is high (or GDP consensus forecasts are low) and banks also tend to reduce their credit supply. If The inverse propensity score estimator is a more general case of the linear framework used by Bassett et al. (2014) for the US and was used in the context of monetary policy shocks by Angrist and Kuersteiner (2011). 5 We describe its construction in section 4. 4

7

such macroeconomic factors are omitted, the effects of credit supply tightening can be polluted by confounding factors. Financial risk factors also influence credit standards: as suggested by Gilchrist et al. (2012b) financial market spreads – proxied by the EBP for non-financial firms – significantly explain credit tightening: an increase of the EBP of about 100 basis points makes tightening more likely by almost 10 percentage points. Compared with the EBP, the volatility of euro area stock markets (VSTOXX) is also significant in our estimates, but once the EBP is included, its impact becomes more muted.6 The richer baseline specification (denoted “Baseline” in the table) includes also monetary policy indicators. The EONIA and the OIS forward rate are highly significant (as in Maddaloni and Peydrò, 2013) suggesting that monetary policy is an important determinant of tightening as expressed in our preferred specification of the ordered probit. This evidence corroborates the existence of a risk-taking channel of monetary policy for the euro area. In a nutshell, when interest rates are low, banks are induced to take on more risk, thereby relaxing credit supply standards. A tightening of credit conditions is much easier to explain in an environment of tight monetary policy, than when policy rates are extremely low. The order of magnitude of the effects of the EONIA is similar to that of the EBP after taking into account the different scale of the variables.7 We undertake additional statistical analyses (reported in appendix 3 to ease readability) on the ordered probit based on the specification test proposed by Angrist et al. (2013): this suggests that the baseline ordered probit is correctly specified. In particular, the test shows that the shock implicitly defined by the difference between actual outcomes („easing‟, „neutral‟, „tightening‟) and their propensity scores from the ordered probit is close to being randomly assigned. The most important variables to obtain the correct specification are bank level demand, the VSTOXX and the Excess Bond Premium. Results of the test suggest two points. First, banking groups face specific demand schedules which are not fully captured by country-level or euro area level information on activity. Second, and not surprisingly, existing financial market uncertainty and the EBP play a relevant role in shaping credit supply. Finally, we have also tested for the inclusion of additional variables in the baseline specification. More specifically, including house prices and a proxy for US monetary policy (see Rey, 2016 and Miranda-Agrippino and Rey, 2015) does not change the results.8

We do not enter here into the discussion about whether uncertainty or the excess bond premium is causal for economic conditions (see for example Caldara et al. (2016) and Gilchrist, Sim and Zakrajšek (2014), where they show that financial frictions are able to explain uncertainty), but we simply add both measures to our probit regressions. However, using only the excess bond premium does not change the results substantially. 7 Appendix 3 presents a statistical analysis based on Angrist et. al (2013), testing whether the estimated order probit is correctly specified. Results suggest that individual bank level demand, EBP, and VSTOXX are crucial to capture the dynamic behaviour of credit standards. 8 Results are not reported for saving space. They are however available upon request. 6

8

Table 1: Factors affecting changes in banks‟ credit standards Marginal effects

Variables

Base Coefficients

No monetary policy

Baseline

Baseline Dynamic

Sample until 2011Q3

Country dummies

-0.0803***

-0.0569***

-0.0542***

-0.0324***

-0.0605***

-0.0560***

-0.265***

(0.0128)

(0.0121)

(0.0117)

(0.00866)

(0.0165)

(0.0109)

(0.0581)

(0.0364)

Country level lagged GDP Growth

-0.0159***

-0.0205***

-0.0111***

-0.0285***

-0.0203***

-0.100***

-0.106***

(0.00293)

(0.00337)

(0.00273)

(0.00445)

(0.00350)

(0.0167)

(0.0172)

Country-level GDP Forecast (Consensus)

-0.0154**

-0.0139**

-0.00618

-0.0187*

-0.0107*

-0.0681**

-0.0721***

(0.00699)

(0.00673)

(0.00384)

(0.0103)

(0.00597)

(0.0323)

(0.0255)

0.0260

0.0399**

0.015

0.0485*

0.0125

0.195**

0.103

(0.0183)

(0.0186)

(0.0118)

(0.0250)

(0.0148)

(0.0897)

(0.0645)

0.0910***

0.0795***

0.134***

0.0829***

0.445***

0.449***

(0.0216)

(0.0176)

(0.0280)

(0.0216)

(0.102)

(0.102)

0.0715***

0.0526***

0.0804***

0.0675***

0.350***

0.359***

Bank level Demand factor

Country level Changes in Unemployment Euro area Lag EONIA Euro area Lagged change in 3m-in-1y OIS Forward

-0.306***

(0.0139)

(0.0141)

(0.0175)

(0.0134)

(0.0657)

(0.0889)

0.00737***

0.00370***

0.00882***

0.00758***

0.0361***

0.0387***

Euro area VSTOXX

0.00275*** (0.000682)

(0.00102)

(0.000711)

(0.00137)

(0.000997)

(0.00434)

(0.00428)

Euro area Change in EBP

0.000923***

0.000957***

0.000644***

0.00123***

0.000884***

0.00468***

0.00484***

(0.000168)

(0.000837)

(0.000890)

YES

NO

(0.000155)

(0.000176)

(0.000147)

(0.000214)

Random Effects

NO

NO

NO

NO

NO

Bank level Lag Credit Standard

NO

NO

NO

0.244***

NO

YES

NO

NO

NO

NO

4,483

4,483

(0.0159) Country dummy

NO

NO

NO

NO

NO

YES

Pseudo-R2

0.02

0.07

0.08

0.24

0.09

0.11

# Observation

4,483

4,483

4,483

4,483

2,664

4,483

Note: Ordered probit specification for the expected change of credit standards. This table reports selected marginal effects on the probability of a credit tightening. Sample period: Q1 2003-Q1 2016Q1; number of banks: 137. The dependent variable in the ordered probit regression is a discrete variable ΔCSit = (tight, neutral, ease), representing the change in credit standards reported by bank i in quarter t in the Bank Lending Survey. Lag credit standard is the lagged dependent variable (i.e. ΔCSit-1); the bank-level demand factor is a discrete variable (tight, neutral, ease) taking values (1,0,-1). Tight reports a reduction of demand; we invert the sign of the coefficient to show it as in Basset et al. (2015). VSTOXX is the Dow Jones EURO STOXX 50 Volatility measured for the euro area; EBP is the excess bond premium; 3m-in-1y OIS forward is the forward rate retrieved from the threemonth OIS rate one year ahead. GDP forecast is taken from Consensus Economics forecasts. Robust asymptotic standard errors are clustered at the bank level and are reported in parentheses. *** Statistical significance at 1%, ** statistical significance at 5%, * statistical significance at 10%.

The fourth column (“Baseline dynamic”) includes also past (bank-specific) tightening conditions, which provide a strong contribution (about 25 p.p.) to the probability of having a further tightening.9 Table 2: Marginal effects in ordered probit regression with sub-questions of the Credit Standards related to supply factors: Capital Positions, Market Funding, and Liquidity Position as dependent variables Marginal Effects Variables Bank Level Demand Factor Country Level Lagged GDP Growth Country Level Changes in Unemployment Country Level Lagged GDP Forecast (Consensus) Euro area Lagged Eonia Euro area Lagged Changes in 3,-in-1y OIS Forward Euro area Changes in VSTOXX Euro area Change in Excess Bond Premium Lagged Capital Position

Capital Position Market Funding Liquidity Position 0.0314*** (0.00944) -0.00462** (0.00214) 0.0164* (0.00979) -0.00568 (0.00381) 0.0133*** (0.00296) -0.00109 (0.00959) 0.000427 (0.00100) 0.000257 (0.000202)

0.0172** (0.00835) -0.00280 (0.00235) 0.0218*** (0.00762) -0.00366 (0.00343) 0.0144*** (0.00267) 0.00725 (0.00848) 0.000275 (0.000957) 0.000393* (0.000207)

0.192*** (0.0147)

Lagged Market Funding

0.169*** (0.0108)

Lagged Liquidity Position Pseudo-R2 Observations

0.0164** (0.00794) -0.00109 (0.00229) 0.0155*** (0.00596) -0.00180 (0.00370) 0.00785*** (0.00244) 0.00594 (0.00870) -0.000392 (0.000878) 0.000470** (0.000190)

0.160*** (0.0125) 0.30 4,792

0.38 4,792

0.34 4,792

Note: Sample period: Q1 2003-Q1 2016; number of banks: 137. Coefficients refer to the average marginal effects on the dependent variable. The dependent variable in the probit regressions is the discrete variable ΔCSit = (tight, neutral), referring respectively to the capital position, market funding and liquidity position of the question “factors affecting credit standards” and it represents the change in credit standards reported by bank i in quarter t. Lag capital position, lag market funding and lag liquidity position are the lagged dependent variables (i.e. ΔCSit-1) in each respective regression; the bank-level demand factor is a discrete variable (tight, neutral, ease) taking values (1,0,-1). Tight reports a reduction of demand; we invert the sign of the coefficient, as in Basset et al. (2015). VSTOXX is the Dow Jones EURO STOXX 50 Volatility measured for the euro area; EBP is the excess bond premium; 3m-in-1y OIS forward is the forward rate retrieved from the three-month OIS rate one year ahead. GDP forecast is taken from Consensus Economics forecasts. Robust asymptotic standard errors are clustered at the bank level and are reported in parentheses. *** Statistical significance at 1%, ** statistical significance at 5%, * statistical significance at 10%.

In terms of results, impulse response functions obtained using the “Baseline” or the “Baseline dynamic” are not significantly different. To ease presentation, we report the simpler case. 9

The remaining specifications provide some robustness checks of our ordered probit. When we cut the sample to the pre-Q3 2011 period, thus excluding the sovereign debt crisis period, we find that the results are broadly unchanged and, if anything, the role of individual demand factors as well as macroeconomic factors are stronger. Additional country dummies do not change the results significantly, as seen with random effects for single banks. 10 Our main results are also robust to using alternative measures of credit standards (i.e alternative questions of the BLS) are used. In particular, results are confirmed when credit standards are measured by the BLS sub-questions on capital, liquidity position, and market funding used by Ciccarelli et al. (2013).11 This reinforces the argument that credit standards suffer from endogeneity problems, to be corrected as we do in this paper. As we show in Table 2, individual components of the overall index of credit standards also respond to past bank-level demand with a marginal effect that is only slightly smaller than for the baseline credit standards. Economic determinants are also significant, albeit less so than in our baseline case12. Finally, we aggregate individual LSI into a euro area measure of loan supply using weights that reflect each country‟s share of total euro area loans. This is the same aggregation method as the one used to compute the BLS net diffusion index. Figure 2 summarises the time evolution of the two BLS-based supply indicators. The top panel shows the LSI, whereas the bottom panel shows BLS credit standards for NFCs. In general, the two indicators exhibit similar cyclical patterns. Nevertheless, there are a number of important differences. First, there is a relatively large easing of LSI around 2005 compared with the BLS measure: this is consistent with the view that credit was relatively slack over the period, with credit booming and the ECB raising interest rates. Then, the subprime crisis in early 2007 came as a surprise, in the context of favourable (current and expected) macroeconomic conditions: both indicators signal a tightening in credit conditions, although the LSI has a more marked increase. The LSI also indicates that the tightening of credit conditions at the end of 2008 was important although it was partially an endogenous response to the already manifesting deterioration in macroeconomic and financial market conditions, as indicated by the significant increase in the stock As marginal effects are not computed for such a case, model (8) reports estimates of the full specification without random effects; models (7) and (8) display very similar estimates, irrespective of whether bank-level random effects are included. 11 More specifically, the sub-questions are the following: “Over the past three months, how have the following factors affected your bank‟s credit standards as applied to the approval of loans or credit lines to enterprises (as described in question 1 in the column headed “Overall”)”? i) Costs related to your bank‟s capital position; ii) Your bank‟s ability to access market financing (e.g. money or bond market financing); iii) Your bank‟s liquidity position. 12 Even the LSI derived from the capital position sub-question (which is the one that seems to be the least endogenous) is quite similar to our baseline LSI (results available upon request). 10

11

market volatility over that period. Finally, the sovereign crisis was associated with a significant and persistent effect on credit supply in LSI as compared with the BLS indicator: the LSI remains high almost until 2014. Following the most recent policy measures by the ECB, both indicators of credit standards provide a similar indication of easing.

Figure 2: Estimated loan supply indicator (blue dashed line) vs. BLS net index of tightening of credit standards on loans to enterprises (solid blue line), net percentage changes.

Note: Sample period: 2003Q1-2016Q1, The figure depicts the headline BLS credit standards indicator (bottom panel) compared to our aggregated LSI (top panel). The econometric model used obtain the LSI is the baseline ordered probit model estimated on pooled individual BLS responses from the main Credit Standards. Individual BLS responses are corrected by taking into account bank-specific loan demand (BLS), macroeconomic conditions (actual and expected) at country level, the riskiness conditions of non-financial corporations in the euro area and monetary policy conditions (EONIA and forward rates).

3. The macroeconomic impact of loan supply shocks This section evaluates the macroeconomic impact of credit supply shocks as measured by the LSI in a time-series modelling framework. More precisely, we use a BVAR model to analyse the dynamics of a set of macroeconomic variables, which links real economic activity with credit and financial markets. In particular, the model includes measures of firms‟ costs of financing derived from both bonds and bank loans (BBB/AA spreads and lending margins respectively), volumes of

12

bank lending to new business (adjusted for sales and securitisation) and corporate securities (as notional stocks), as well as the policy interest rate, GDP and the GDP deflator 13. The model is estimated in log-levels for real GDP, the GDP deflator, loans to NFCs, and debt issuance by NFCs. The short-term interest rate, the composite lending rate to NFCs and the spread between the NFCs‟ BBB bonds and the AAA government bond are instead taken in levels. The autoregressive order used in the estimation is five. For the estimation of the VAR, we address the curse of dimensionality problem by using Bayesian shrinkage, as suggested in De Mol, Giannone and Reichlin (2008). The methodology follows closely Giannone, Lenza and Primiceri (2015).14 In more detail, priors are standard Normal-Inverse Wishart and impose the so-called Minnesota prior, according to which each variable follows a random walk process, possibly with drift (Litterman, 1979). We impose two sets of prior distributions on the sum of the coefficients of the VAR model: the “sum-of-coefficients” prior, originally proposed by Doan, Litterman and Sims (1984), and an additional prior that was introduced by Sims (1993), known as the “dummy-initial-observation” prior. The hyper-parameters controlling for the informativeness of the prior distributions are treated, as suggested in Giannone, Lenza and Primiceri (2015), as random variables and are drawn from their posterior distribution, so that we also account for the uncertainty surrounding the prior set-up in our evaluation. The sample period goes from the first quarter of 1999 to first quarter of 2016, with the beginning of the ECB‟s corporate bond purchase programme, which significantly affected the yields and the issuance of corporate bonds. Results are similar when enlarging the sample, i.e. starting in 1996, or restricting it to the period of the availability of the BLS (starting in the first quarter of 2003).

3.1 The identification strategy Identification of the credit supply shock is achieved by using the LSI as an external instrument in the VAR. Simple identification schemes based, for example, on Cholesky decomposition have been widely criticised in the literature on financial frictions and on monetary policy for being typically mis-specified with respect to generic data-generating processes.15 The external instrument methodology that we use follows recent contributions by Stock and Watson (2012), and Mertens and Ravn (2013). Credit supply developments are identified using the Results are robust to considering the HICP instead of the GDP deflator. The EONIA is taken as the policy interest rate; results are similar using the three-month EURIBOR measure. 14 The same type of BVAR model, including the choice of the priors, was used in Miranda-Agrippino and Rey (2015), Altavilla, Giannone, Lenza (2016). 15 A recent discussion on the US case regarding how identification with external instruments can deliver more reliable results compared with other identification schemes can be found in Mumtaz et al. (2018). 13

13

additional information from the instrument without imposing predefined sign restrictions on the rotation matrix of the variance-covariance matrix. The method allows the sample over which the LSI is calculated (i.e. the sample over which the BLS is active) to be much shorter than the sample used for the estimation of the BVAR16 and –differently from including the instrument directly into the VAR-- it is robust to measurement error which is a frequent issue when using information from surveys. Identification through external instrument follows the approach in Stock and Watson (2012). The BVAR model produces reduced-form residuals shocks

which are a linear combination of structural

: (4)

where H is the

identification matrix. Without loss of generality, we consider credit supply as

the first shock,

. Consider now for simplicity having a single instrument Z for the credit supply

shock. In order to be a valid instrument, Z should satisfy three conditions: ,

(5)

,

(6) (7)

[

]

Condition (5) is an exogeneity assumption: instrument Z should be orthogonal with respect to the structural shocks different from the one we want to identify. Condition (6) specifies that Z is valid as it is correlated to the shock we want to identify, with covariance

. Finally, condition (7)

states that structural shocks are not cross-correlated. Under assumptions (5-7) the structural shock can be estimated by the linear projection of the reduced-form innovations

on the instrument Z. More in detail, the linear projection is given by: ( ( (

)

)

) (

(8) )

is the inverse of the variance-covariance matrix of reduced-form residuals. In the first step of (8), relation (4) is used to factorise

into the product of the identification matrix H and the

diagonal matrix V. In the second step, equation (6) replaces the expected product between

and

We are aware of the criticisms made by Ramey (2016) about having different samples for the VAR and instruments. In our case, however, the results are not significantly affected by shortening the sample of our VAR. 16

14

the proxy Z. Finally, the last result follows from the orthogonality condition in equation (5): this step makes it clear that the identification of

is obtained up to a normalisation constant of

Once the structural shock

has been identified from equation (8),

identified structural shock

on the reduced-form innovations

. 17

is derived by regressing the

, i.e. by equation (4) In this way,

we compute the impulse responses and historical decomposition for the identified structural shock. The case of multiple instruments used to identify a single shock can also be handled by substituting the OLS regression in equation (8) with a reduced-rank regression, as we discuss in Section 4, where we empirically evaluate whether the LSI and the EBP are instruments for the same underlying shocks, by testing for overidentifying restrictions as in Stock and Watson (2012). Finally, in order to test for quality of the LSI instrument, we also use the “reliability statistic” of Mertens and Ravn (2013): this provides a metric for evaluating how closely the instruments are related to the true shocks, providing a sense of the quality of identification. When there is only one instrument, as in our case, this is equivalent to computing the fraction of the variance of the LSI that residuals are able to explain (R2) and the procedure is similar to using the first-stage regression to evaluate the strength of the instruments in the context of weak-instruments. In the case of our LSI the test statistic equals to 23%, which is above the relevant threshold used to evaluate strength of instruments of 15% (Stock and Yogo, 2005).

3.2 Measuring the effects of loan supply shocks The impulse response functions (IRFs) of a loan supply shock corresponding to a tighter LSI are presented in Figure 3, volume variables are reported in growth rates to facilitate the reading. Real GDP growth declines for about one year with a peak loss of 0.4 in year-on-year terms. The associated loss for the GDP level is about 0.6 percentage points at the trough. The impact on the inflation rate as measured by the GDP deflator is also negative, but comparably small. Regarding loan volumes and prices, the loan supply shock translates into a gradual increase of bank lending margins (here the difference between lending rates and the EONIA) and into lower lending volumes. Year-on-year loan growth reduces by about double the decrease in economic activity and reaches a trough 3-4 quarters after GDP; the overall effect of a loan supply shock on lending is more gradual and persistent than on output. In terms of relative magnitudes, the peak effect on loans reaches 0.8 percentage points, significantly larger than the one on GDP. The lagged impact of credit standards on bank loans compared to GDP confirms previous findings from euro area data

17

Whenever the identification of a shock is compared under two different instruments, we normalised the impact of the identified shock on the credit spreads.

15

(Ciccarelli et al., 2015). Similar evidence is available for the US in Lown and Morgan (2006) using a recursive identification scheme and in Bassett et al. (2014)

Figure 3: Response to an adverse loan supply shock

Note: The figure reports the impact of a tightening of the loan supply indicator on selected macroeconomic variables. The blue solid line represents the posterior median. The red-dashed area represent the 95% intervals based on 2,000 MCMC draws from the posterior distribution. Y-o-y results derived from original levels.

Concerning firms‟ debt composition, the negative loan supply shock translates into stronger recourse to bond financing as the contraction in loan volumes and the higher bank lending spreads increase firms‟ incentives to substitute bank loans with bonds. This is consistent with the observed substitution between loans and bonds, particularly significant during the two euro area crises (see Charts C and D figure A1 in Appendix 1). The increase in debt security issuance however only partially makes up for the reduction in loans. This is in line with the model of Crouzet (2018),

16

where the bond-loans substitution induced by bank loan supply is partial and not enough to avoid a decline in real economic activity. Finally, the BBB-AA corporate bond spread also increases: access to corporate debt markets provides a financial buffer to firms against the reduction in loan supply, but investors will require some risk compensation to hold the additionally issued bonds. The surge of bond spreads is however relatively short-lived when compared to the quite persistent increase of lending rate margins (i.e. the difference between lending rates to NFCs and the 3 M Euribor).

3.3 Sensitivity to alternative shocks This subsection discusses the relation between loan supply shocks identified by our LSI and structural shocks identified in the literature. For robustness, we consider both the correlation across different proxies (as in Piffer and Podstawski, 2018) and the correlation between the identified shocks as in Stock and Watson (2012). The two criteria give a comprehensive picture on the instrument exogeneity condition (5) in the data. The measures considered are the following: the total factor productivity (TFP); fiscal policy measures (the tax revenues and public expenditure developed in Blanchard and Perotti, 2002); the uncertainty indicator based on the price of gold (Piffer and Podstawski, 2018); the economic policy uncertainty index (Baker et al. 2016); the excess bond premium (EBP) that we construct for the euro area using the methodology developed by Gilchrist and Zakrajšek (2012); the stock market volatility index, VIX; the variance risk premium (Bekaert and Hoerova, 2014); and the news shock (Barsky and Sims, 2011). Table 3 reports results for the indicators mentioned above. Table 3: Exogeneity of the proxy

Own computation

Correlation of LSI with other instuments 0.27***

Correlation of loan supply shocks with other shocks 0.58***

Policy Uncertainty

Bloom et al. (2016)

0.07

0.49*

VSTOXX

Bloomberg

0.41***

0.71***

Uncertainty (gold)

Piffer and Podstawski (2018)

-0.07**

0.70**

News Shocks

Barski and Sims (2011)

-0.24*

-0.23

TFP

Eurostat data, production function

-0.12

0.19

Public expenditure

Blanchard and Perotti (2002)

-0.04

0.02

Tax Revenues

Blanchard and Perotti (2002)

0.00

-0.13

Instruments/Shocks

Reference

Excess Bond Premium

Note: The table reports correlations between different instruments/shocks and the loan supply indicator. The methodology used to compute the excess bond premium for the euro area is reported in Appendix 5. Standard errors in parentheses: * p<0.1, ** p<0.05, *** p<0.01.

17

The results shown in the table indicate that the LSI is not correlated with policy uncertainty, TFP and fiscal policy measures. The news shock proxy has only a muted correlation with the LSI shock. The uncertainty measure based on gold prices (Piffer and Podstawski, 2018) shows a significant but negative correlation with the LSI measure. This would indicate that a reduction in global uncertainty would decrease when loan supply becomes tighter. This result might potentially be due to the fact that gold prices were decreasing in the aftermath of the great recession, while the euro area was entering into sovereign debt crisis, with adverse consequences on banks‟ supply of loans. Finally, the LSI does correlate positively with measures of financial uncertainty and risk premia (the VIX and the EBP) which could thereby be a potential confounding factor for our results. Concerning the VIX, the observed correlations between the loan supply indicator and stock market volatility can stem from two potential sources: either financial uncertainty is a confounding factor for credit supply, or both financial uncertainty measures and credit supply are influenced by a deeper common shock. The former explanation would see uncertainty as a confounding factor for the LSI. However increases in uncertainty alone do not seem to be able to explain firms substituting bonds to loans in their source of finance, lacking a restriction on the supply of loans by banks.18 Economic theory would then rather corroborate the second explanation: the model by He and Krishnamurty (2013), for example, explains how a surge in financial spreads and in the VIX coincide with a reduced risk-taking capacity by financial intermediaries; Adrian et al (2012) also provide a model that explains the correlation between changes in credit supply and increase in financial spreads. To understand the positive correlation of the LSI with the VIX, we augment the VAR with the two components of the VIX (see Bekaert and Hoerova, 2014, 2018): the variance risk premium (VRP) and conditional stock market variance (CV). 19 Results reported in Figure 4, indicate that the LSI shock loads differently onto the two components and affect the CV more substantially: the shock measured by the LSI is closer to a volatility shock rather than one affecting risk premia. This is consistent with the finding in Bekaert and Hoerova (2014) that CV rather than VRP play a larger role in predicting the real economy.

Figure A.1D suggests that firms substitute bonds to loans more strongly over those periods when banks reduce loan supply such as during the global financial crisis and the sovereign crisis. 19 We thank Marie Hoerova for providing us with the series for the euro area. 18

18

Figure 4: Variance risk premium and Conditional Volatility responses to loan supply shock (LSI)

Note: The figure reports the impact of a tightening of the loan supply indicator variance risk premium (VRP) and stock market conditional volatility (CV) as in Bekaert and Hoerova (2018). The blue solid line represents the posterior median. The red-dashed area represent the 95% intervals based on 2,000 MCMC draws from the posterior distribution.

Finally, concerning the positive correlation with the EBP, we include the two proxies (EBP and LSI) as instruments in the proxy-SVAR and use narrative sign restrictions (Antolin-Diaz and Rubio-Ramirez, 2018) to disentangle the impact of the underlying shocks. We impose sign restrictions over a specific historical episode where the two proxies showed a clearly different pattern. Between mid-2015 until the end of the sample, financial spreads (and EBP) were increasing amid bad news from the US, but banks‟ conditions proxied by the LSI remained accommodative. We can then use this period to identify two shocks. First, a positive loan supply shock – proxied by the LSI – reduced lending margins as well as financial spreads. Second, a financial market shock had instead a positive impact on financial spreads. The financial market shock is assumed to have an effect on the financial spread that is comparatively larger than the effect of the loan supply shock. Results are reported in Figure 5. The figure shows that the estimated impacts under this identification scheme largely coincide with the ones obtained when using the LSI (black solid line in the charts).

19

Figure 5: Responses to an adverse loan supply shock, robustness to identifying a financial market shock

Note: The figure reports the impact of a rise in the EBP of euro area NFCs on selected macroeconomic variables. The blue solid line represents the posterior median. The red-dashed area represents the 95% intervals based on 2,000 MCMC draws from the posterior distribution. The black line reports the median impulse to the same shock, disentangled from financial uncertainty.

3.4 Historical decomposition of loan supply shock Having identified and described the propagation mechanism of shocks to loan supply, we now turn to the historical decomposition of macroeconomic and financial variables over the two recessionary episodes in order to assess the role of loan supply structural shocks in the euro area macroeconomic developments during the financial crises. First we concentrate on how much loan supply has contributed to negative developments in economic activity over the two recessionary episodes. Second, we focus on how much interaction between banks and market financing was at play due to restrictions in banks‟ lending conditions. Overall, credit supply restrictions capture the dynamics of real GDP and loans relatively well in the Great Recession. Figure 6 shows that over that period the counterfactual dynamics of variables (dashed line) remains relatively close to the actual developments (solid line).

20

Figure 6: Historical decomposition of selected macroeconomic and financial variables

Note: Starting date of the decomposition is 2008Q1. The panels of the figure depict the contribution of the credit supply shock (black dashed line) compared to the historical developments of selected macroeconomic variables (red solid bar). Real GDP, the GDP deflator, Loans to NFCs and Debt securities issuance by NFCs are expressed as y-o-y growth rates in deviation from the baseline of their unconditional forecast. Lending margins and corporate bond spreads are expressed in level deviation from unconditional forecasts.

For GDP we explain more than one third of the fall of y-o-y GDP growth and about half of the dynamics of the loans. For the GDP deflator the performance is instead less remarkable. For what concerns lending margins, the tightening of lending standards is able to explain most of the increase in the price of loans, suggesting that we are well capturing a good share of the effect of a loan supply restriction. Finally, the reduction in loan supply explains a large part of the increase in bond issuance around 2010 and more than half of the increase in spreads observed over the Great Recession. The counterfactual exercise captures less of the dynamics of variables over the sovereign crisis, suggesting that also other shocks (e.g. fiscal adjustments for economic activity) played a role. Over the sovereign crisis period the counterfactual exercise captures only a limited part of relation between loans and bonds as it fails to capture a large part of the surge of bond issuance in 2013.

21

This result is confirmed also when looking at country-specific developments, in particular for the four largest euro area countries (Germany, France, Italy and Spain). Country-level analysis, reported in Appendix 3, shows however that at least for the case of France, loan supply shocks are able to explain at least part of the increase in bond issuance slightly better than for the euro area case. The counterfactual dynamics of both loans and GDP are also rather muted. Still the counterfactual exercise is able to capture most of the dynamics of the lending margins, suggesting that the identification of credit supply restrictions using the LSI is appropriate. Finally, all counterfactual variables, except prices, capture the overall trends after 2014-2015: an improvement of lending conditions with associated improved dynamics of loans; a decrease in lending margins and bond spreads; a deceleration of bond issuance. Overall, the ability of loan supply shocks to explain the BBB-AA spreads appears to be remarkable over the whole sample. This suggests that the rise in credit spreads for euro area nonfinancial corporations might have been driven by the same sources as tightening of loan supply.

4. Cross-country heterogeneity This section analyses possible cross-country heterogeneity in the transmission of loan supply shocks.20 Country specific analysis allows us to control for potentially heterogeneous transmission mechanisms of loan supply shocks in different countries, which could have been a relevant feature especially over the sovereign crisis period, characterised by fragmentation, i.e. a very different behaviour of Northern versus Southern countries. We concentrate on the effects of loan supply tightening on key macroeconomic variables for the four largest euro area countries: Germany, France, Italy and Spain. Country-specific models have the exact same endogenous variables used for the euro area specification. We exclude lending rates and short term policy rate as monetary policy in the euro area is not supposed to react to countryspecific developments. Including the monetary policy rate with country-specific information would provide a distorted estimate of the systematic behaviour of monetary policy. Country-specific LSIs are used as instruments and are reported in the appendix (Figure A4). A possible caveat that should be taken into consideration is that Bank Lending Survey answers measure conditions at banking groups whose perimeter can go much beyond national borders. In this respect, responses of large banks might be considered as more representative of their overall perimeter of operations (Euro area), than the country where their mother company belongs to. Similar considerations hold for bond issuance of large non-financial corporations, which typically issue new bonds in countries where conditions are most favourable through financial subsidiaries. In this sense both bond Appendix 4 reports credit standards indicators for the largest four euro area economies as obtained with the LSI (dashed blue line) and BLS (solid blue line). 20

22

market issuance and loan supply measures can be more representative of the euro area than of conditions prevailing in individual countries. Figure 7 reports for each country the estimated reaction of real GDP volumes (Panel A) and loans to non-financial corporations (Panel B). All variables are expressed in year-on-year growth rate. A negative loan supply shock produces adverse effects on both GDP and loans in all countries. The size of the impact is however heterogeneous across jurisdictions. More precisely, a supply shock that induces an increase of 20bps in bond spreads leads to a reduction in the growth rate of real GDP of about 0.5% in Italy, Spain and Germany. The reaction seems to be much more muted in France, where the decline in GDP is almost halved (i.e. 0.2%). The weaker responsiveness of real activity in France is in line with the higher possibility to substitute among different sources of firms‟ financing, given the more developed market for private bonds of France.21 The effects of credit supply factors in the global recession are gauged to be sizeable in Germany, Italy and Spain, where they can explain overall about half of the fall in GDP (Figure A5 in appendix); the French economy seems to be less affected by loan supply developments. In terms of the dynamic reaction of loans to a negative loan supply shock, the negative effects reach a peak after 8 quarters in all countries. The size of the peak effect is however quite heterogeneous across countries, ranging from less than -0.5% in Germany to about -1.5% in Spain. The lower sensitivity of loans to the credit supply shock in Germany reflects the historically more muted dynamics of lending compared to other European countries.

Notice that French corporate debt securities account for about 50% of the entire amount of bond outstanding in the euro area (see Giuzio and Nicoletti, 2018). 21

23

Figure 7: Impulse response for selected euro area countries Panel (A): Real GDP

Panel (B): Loans to NFC

Note: the figure reposts the response of real GDP and lending growth to a negative credit supply shock that reduces bond spreads by 20 basis points.

24

5 Conclusions This paper analysed the macroeconomic impact of loan supply shocks in the euro area. The relevance of this issue relates to the prominent role played by banks in financing euro area firms. Loan supply shocks are identified using a new indicator of credit supply conditions derived from bank-specific confidential information on credit standards in the ECB‟s bank lending survey. The analysis disentangled credit supply developments from credit demand and other concurrent macroeconomic and financial factors. The results indicated that credit supply developments are substantially influenced by bank-specific demand, financial market uncertainty, and corporate bond spreads. The loan supply indicator (LSI) was then used as external instrument in a Bayesian VAR model to identify the macroeconomic consequence of a contraction in bank credit supply. We found that adverse shocks to bank loans lead to a contraction in real activity and credit volumes as well as to a widening in bank lending spreads. The results also highlighted that following a credit tightening, firms tend to substitute bank loans with debt securities issuance. More specifically, tighter supply of bank loans to non-financial corporations has been found to explain a large part of the substitution between loans and bonds (i.e. surge in corporate bond issuance and increase in bond spreads) during the financial crisis. Firms substitute loans with bonds at times of banks‟ distress. This substitution is however incomplete. Focusing on cross-country differences, the analysis showed that countries with more developed corporate bond markets have been more resilient to negative loan supply shocks.

25

References Adrian, T., Colla, P. and Shin, H. S. (2012), “Which Financial Frictions? Parsing the Evidence from the Financial Crisis of 2007 to 2009”, NBER Chapters in: NBER Macroeconomics Annual 2012, Vol. 27, National Bureau of Economic Research, pp. 159-214. Altavilla, C., Giannone D. and Lenza, M. (2016), “The Financial and Macroeconomic Effects of the OMT Announcements”, International Journal of Central Banking, vol. 12(3), pages 29-57, September. Angrist, J. D. and Kuersteiner, G. D. (2011), “Causal Effects of Monetary Shocks: Semiparametric Conditional Independence Tests with a Multinomial Propensity Score”, Review of Economics and Statistics, Vol. 93(3), pp. 725-747. Angrist, J. D., Jordà, O. and Kuersteiner, G. D. (2013), “Semiparametric Estimates of Monetary Policy Effects: String Theory Revisited”, NBER Working Paper No 19355. Antolín-Díaz, J and Rubio-Ramírez, Juan F. (2018), “Narrative Sign Restrictions for SVARs“ American Economic Review, vol. 108(10), pages 2802-2829, October. Baker Scott R., Nicholas Bloom, Steven J. Davis (2016), Measuring Economic Policy Uncertainty, The Quarterly Journal of Economics, vol. 131(4), pages 1593-1636. Barsky, Robert B. & Sims, Eric R., 2011. “News shocks and business cycles”, Journal of Monetary Economics, vol. 58(3), pages 273-289. Bassett, W. F., Chosak, M. B., Driscoll, J. C. and Zakrajšek, E. (2014), “Changes in bank lending standards and the macroeconomy”, Journal of Monetary Economics, Vol. 62(C), pp. 23-40. Bekaert, Geert and Hoerova, M (2014), “The VIX, the Variance Premium and Stock Market Volatility”, Journal of Econometrics, Vol. 183, No. 2, pp. 181-192. Bekaert, G., M. Hoerova and N. Xu (2018), “Risk, Uncertainty and Monetary Policy in a Global World,” mimeo. Becker, B. and Ivashina, V. (2014), “Cyclicality of credit supply: Firm level evidence”, Journal of Monetary Economics, Vol. 62, pp. 76-93. Blanchard O. and Perotti, R. (2002), “An Empirical Characterization of the Dynamic Effects of Changes in Government Spending and Taxes on Output”, The Quarterly Journal of Economics, vol. 117(4), pages 1329-1368. Bleaney, M., Mizen, P. and Veleanu, V. (2016), “Bond Spreads and Economic Activity in Eight European Economies”, Economic Journal (doi:10.1111/ecoj.12288). Caldara, D., Fuentes-Albero, C., Gilchrist, S. and Zakrajšek, E. (2016), “The Macroeconomic Impact of Financial and Uncertainty Shocks”, European Economic Review. Volume 88, pp.185-207.

26

Christiano, L. J., Motto, R. and Rostagno, M. (2014), “Risk Shocks”, American Economic Review, Vol. 104(1), pp. 27-65. Christiano, L. J., Eichenbaum, M. S. and Trabandt, M. (2014), “Understanding the Great Recession”, NBER Working Paper No 20040. Ciccarelli, M., Maddaloni, A. and Peydró, J.-L. (2013), “Heterogeneous Transmission Mechanism: Monetary Policy and Financial Fragility in the Euro Area”, Economic Policy, Vol. 28(75), pp. 459-512. Ciccarelli, M., Maddaloni, A. and Peydró, J.-L. (2015), “Trusting the Bankers: A New Look at the Credit Channel of Monetary Policy”, Review of Economic Dynamics, Volume 18, Issue 4, pp.979-1002. Crouzet N. (2018) Aggregate Implications of Corporate Debt Choices, Review of Economic Studies (Forthcoming) De Fiore, F. and Uhlig, H. (2011), “Bank Finance versus Bond Finance”, Journal of Money, Credit and Banking, Vol. 43(7), pp. 1399-1421. De Mol, C., Giannone, D. and Reichlin, L. (2008), “Forecasting using a large number of predictors: Is Bayesian shrinkage a valid alternative to principal components?”, Journal of Econometrics, Vol. 146(2), pp. 318-328. Doan, T., Litterman, R. and Sims, C. A. (1984), “Forecasting and Conditional Projection Using Realistic Prior Distributions”, Econometric Reviews, Vol. 3, pp. 1-100. Giannone, D., Lenza, M. and Primiceri, G. (2015), “Priors for Vector Autoregressions”, Review of Economics and Statistics, Vol. 2(97), pp. 436-451. Gertler, M. and Karadi, P. (2015), “Monetary Policy Surprises, Credit Costs and Economic Activity”, American Economic Journal: Macroeconomics, Vol. 7(1), pp. 44-76. Gilchrist, S. and Mojon, B. (2018), “Credit Risk in the Euro Area”, The Economic Journal, Vol. 128, Issue608, pp. 118-158. Gilchrist, S., Sim, J. W. and Zakrajšek, E. (2014), “Uncertainty, Financial Frictions, and Investment Dynamics”, NBER Working Paper No 20038. Gilchrist, S., Yankov, V. and Zakrajšek, E. (2009), “Credit market shocks and economic fluctuations: Evidence from corporate bond and stock markets”, Journal of Monetary Economics, Vol. 56(4), pp. 471-493. Gilchrist, S. and Zakrajšek, E. (2012a), “Credit Spreads and Business Cycle Fluctuations”, American Economic Review, Vol. 102(4), pp. 1692-1720. Gilchrist, S. and Zakrajšek, E. (2012b), “Bank Lending and Credit Supply Shocks”, proceedings of the 16th World Congress of the International Economic Association, Vol. III: The Global Macro Economy and Finance, edited by Allen, F., Aoki, M., Kyotaki, N., Gordon, R. and Stiglitz, J.

27

Gilchrist, S. and Zakrajšek, E. (2013), “The Impact of the Federal Reserve‟s Large‐Scale Asset Purchase Programs on Corporate Credit Risk”, Journal of Money, Credit and Banking, Vol. 45(s2), pp. 29-57. Giuzio M. and Nicoletti G. (2018) Integrating euro area corporate bond markets: benefits and potential financial stability challenges, Special feature B, ECB Financial integration in Europe, May 2018 He, Z., Krishnamurthy, A., (2013), Intermediary Asset Pricing, American Economic Review, 103(2): 732–770. Holmstrom, B. and Tirole, J. (1997), “Financial Intermediation, Loanable Funds, and the Real Sector”, Quarterly Journal of Economics, Vol. 112(3), pp. 663-691. Imbens, G. (2004), “Non parametric Estimation of Average Treatment Effects under Exogeneity: A Review”, Review of Economics and Statistics, Vol. 86(1), pp. 4-29. Litterman, R. (1979), “Techniques of forecasting using vector autoregressions”, Federal Reserve of Minneapolis Working Paper No 115. Lown, C. and Morgan, D. P. (2006), “The Credit Cycle and the Business Cycle: New Findings Using the Loan Officer Opinion Survey”, Journal of Money, Credit and Banking, Vol. 38(6), pp. 15751597. Maddaloni, A. and Peydro, J.-L. (2013), “Monetary Policy, Macroprudential Policy, and Banking Stability: Evidence from the Euro Area”, International Journal of Central Banking, Vol. 9(1), pp. 121169. Mertens, K. and Ravn, M. O. (2013), “The Dynamic Effects of Personal and Corporate Income Tax Changes in the United States”, American Economic Review, Vol. 103(4), pp. 1212-1247. Miranda-Agrippino, Silvia and H Rey (2015) “World Asset Markets and the Global Financial Cycle”, NBER Working Papers 21722. Mumtaz, H., Pinter, G. and Theodoridis, K. (2018), “What do VARs Tell Us about the Impact of a Credit Supply Shock? An Empirical Analysis”, International Economic Review, Volume59, Issue2, Pages 625-646. Piffer, Michele. and Podstawski, M. 2018, “Identifying Uncertainty Shocks Using the Price of Gold”, Economic Journal, vol. 128(616), pages 3266-3284, December. Ramey, V (2016) “Macroeconomic Shocks and their Propagation” in Handbook of Macroeconomics, Edited by John B. Taylor, Harald Uhlig, Volume 2, Pages 1-2693 Repullo, R. and Suarez, J. (2000), “Entrepreneurial moral hazard and bank monitoring: A model of the credit channel”, European Economic Review, Vol. 44(10), pp. 1931-1950. Rey, H. (2016), International Channels of Transmission of Monetary Policy and the Mundellian Trilemma, IMF Economic Review 64(1), 6-35.

28

Sims, C. A. (1993), “A Nine-Variable Probabilistic Macroeconomic Forecasting Model”, Business Cycles, Indicators and Forecasting, NBER Chapters, National Bureau of Economic Research, pp. 179212. Stock J. H. and Yogo M. (2005), “Testing for Weak Instruments in Linear IV Regression”, in: Andrews DWK Identification and Inference for Econometric Models. New York: Cambridge University Press, pp. 80-108. Stock, J. H. and Watson, M. W. (2012), “Disentangling the Channels of the 2007-2009 Recession”, NBER Working Paper No 18094, National Bureau of Economic Research.

29

Appendix 1 Stylised facts The euro area has recently experienced two severe recessions. The first recession, usually referred to as the “Great Recession”, started in 2008Q1 and took place in the context of the post-Lehman global financial crisis until 2009Q2. The second recession started in 2011Q3, ended in 2013Q1, and was instead mainly determined by euro area specific events following the European sovereign debt crisis22. Real GDP losses associated with the recent recessionary episodes appear to be substantial.23 Panel A of Figure A.1 suggests that in the absence of the two recessionary periods, real GDP would have been about 12% higher by the beginning of 2016. The reduction in economic activity was however smaller than the associated fall in credit to nonfinancial corporations (NFCs), suggesting that a strong deleveraging process occurred. Panel B in Figure A.1 shows such a deleveraging process in terms of the nominal loan-to-GDP ratio, which has fallen considerably since the start of the financial crisis. At the same time, bond issuance of non-financial corporations surged as firms tried to make up for lost funding opportunities from banks (Panel C). Panel D puts the relationship between loans and bonds into a historical context. In a first period of the euro area‟s existence, 1999-2004, both loans and bond issuance grew at a relatively stable pace, while periods observed with more buoyant loan supply (2005-2008) are associated with a reduced pace of growth of bond issuance and an acceleration of loan growth. The two severe contractions (2009-2010 and 2012-2014) are instead characterised by a reduction of loans and an expansion of bond issuance by NFCs. Overall, the strong negative correlation between loans and bonds was a specific feature of the two recent recessions. However, some milder form of substitution between the two sources of financing might have been at play also in normal times depending also on credit supply dynamics.

See CEPR for an official dating of the euro area business cycle, http://cepr.org/content/euro-areabusiness-cycle-dating-committee 23 In order to compare the observed historical outcomes with plausible estimates of trend growth that would have occurred in the absence of the crises, we follow Christiano, Eichenbaum and Trabandt (2014). The uncertainty surrounding the trend output estimates is accounted for by computing several (log-)linear trends of variables at different points in time between Q1 1990 and Q2 2008. We then extrapolate each of these trends until Q1 2016. Gaps are then constructed as the percentage difference between the projected trend value of the variable and its observed realisation. 22

30

Figure A.1: Deviations from (log)-linear trend of macro variables since 2008 and joint dynamics of loans and bonds since the beginning of the euro (A) Real GDP– detrended (gap from trend in percent)

(B) Loans to NFCs relative to GDP– detrended (gap from trend in percent)

Note: The figure reports the median (solid red line) and the 95th-5th percentile (light grey) and 84th-16th percentile (dark grey) range of the detrended data.

Note: The figure reports the distribution of the detrended loans to NFCs-to-nominal GDP ratio. It shows the median (solid red line) and the 84th-16th percentile (dark grey) and 95th-5th percentile (light grey) range of the detrended data.

(C) Debt securities issuance – detrended (gap in percent)

(D) Substitution between loans and bonds in the euro area 1200

1200

2014Q2- 2016Q1 1000

1000

NFCs Securities (euro bn)

2012 - 2014 800

800

2009 - 2010

600

600

2005 - 2008

400

400

1999 - 2004

200

0 1000

200

1500

2000

2500

3000

3500

4000

4500

0 5000

MFI Loans to NFCs (euro bn)

Note: The figure reports the distribution of the detrended debt securities issuance. It shows the median (solid red line) and the 84th-16th percentile (dark grey) and 95th-5th percentile (light grey) range of the detrended data.

Note: The figure reports the evolution of the notional stock of bonds issued by non-financial firms (y-axis) against the notional stock of their loans (x-axis) in the euro area. Both quantities are reported in billions of euro.

31

Appendix 2 Bank lending survey The BLS is addressed to senior loan officers of a sample of euro area banks and is conducted four times a year. The survey is representative of the banking sector in the euro area as it collects individual answers from 137 banks and takes into account the characteristics of national banking structures. The survey contains 17 specific questions on past and expected credit market developments. The former covers developments over the past three months, while the latter focus on the next three months. Questions are classified according to the two borrower sectors that are the central focus of the survey, i.e. enterprises and households. The definitions and classifications used in the survey are consistent with other ECB statistics. For both enterprises (i.e. non-financial corporations) and households, the questionnaire covers both loan demand and loan supply factors. Among the supply factors, attention is given to credit standards and credit terms and conditions, as well as to the various factors that may be responsible for their changes. Credit standards are the internal guidelines or criteria that guide a bank‟s loan policy. The terms and conditions of a loan refer to the specific obligations agreed upon by the lender and the borrower, such as the interest rate, collateral required and maturity. All in all, ten questions centre on supply factors, of which seven look at credit standards and three at terms and conditions. Of the questions on credit standards, three refer to banks‟ liquidity position, access to market funding and capital position, which were the most acute sources of bank vulnerabilities during the financial crisis affecting credit supply. Seven questions focus explicitly on loan demand, of which three look at demand from enterprises and four at demand from households.

32

Table A.1: Representativeness of the BLS sample Country

Loans to euro area resident nonfinancial enterprises

Loans to euro area resident households for house purchase

Loans to euro area resident households for consumer credit and other lending

Total main assets

Number of banks in the BLS sample*

2014 2013 2012 2011 2010 2014 2013 2012 2011 2010 2014 2013 2012 2011 2010 2014 2013 2012 2011 2010

euro area BE DE EE IE GR ES FR IT CY LV LU LU (domestic) MT NL AT PT SI SK FI

56.2 79.3 37.3 84.4 72.9 86.8 63.0 53.0 60.3 81.7 72.5 42.8 66.5 69.4 69.7 43.4 69.3 47.6 68.2 73.7

56.2 78.1 37.1 85.2 69.0 80.6 61.8 54.2 59.7 79.4

54.7 77.5 36.6 86.1 67.2 64.3 54.6 55.6 60.7 78.3

54.0 78.6 34.8 88.5 67.8 61.5 50.3 56.1 61.8 84.0

51.7 78.3 34.6

44.6 67.8 68.5 69.8 43.8 68.7 55.0 68.6 73.9

45.5 69.0 92.5 71.8 32.3 68.1 56.4 46.0 74.9

41.8 55.6 74.4 71.8 32.2 69.7 56.2 46.5 76.5

42.2 60.8 73.6 60.5 34.2 68.7 56.8 46.1 77.1

68.7 62.4 44.1 55.8 61.5 81.8

64.3 54.5 33.2 97.4 84.6 90.7 62.7 85.6 61.8 89.0 84.7 88.6 90.2 94.7 93.8 29.0 75.4 61.8 76.0 90.5

64.9 52.8 33.5 97.7 82.4 80.3 62.4 86.5 62.4 88.8

63.8 54.5 33.8 97.5 79.4 66.8 57.6 86.3 63.3 87.7

63.8 51.0 34.4 97.6 83.7 66.8 54.0 86.4 64.8 88.0

61.0 45.9 34.2

88.4 90.1 94.7 96.1 29.5 75.4 62.5 76.1 90.8

88.9 89.7 95.2 97.0 27.7 75.4 62.8 61.7 91.1

75.2 75.3 96.3 97.1 27.3 78.8 64.4 61.0 90.8

75.0 75.2 97.9 93.1 27.9 80.0 64.7 61.5 90.6

74.3 67.4 48.5 86.0 66.5 87.9

47.4 55.7 31.4 85.0 74.0 91.5 53.9 58.3 44.0 93.6 63.5 32.2 82.8 91.8 81.6 22.1 47.9 55.4 78.9 86.8

47.4 57.3 31.3 85.8 59.9 85.7 54.9 57.6 44.6 93.2

46.1 57.5 31.1 75.2 58.5 67.2 49.3 57.4 46.4 91.0

45.5 56.3 30.4 90.4 67.7 66.3 45.1 57.5 47.1 91.5

40.0 56.9 30.6

30.5 85.0 93.3 81.0 22.7 48.8 55.4 80.9 87.9

30.8 86.0 94.6 80.7 20.5 49.3 56.5 70.6 87.9

26.8 73.7 95.7 79.3 17.9 52.0 57.8 72.2 88.0

32.2 73.0 97.0 70.0 19.0 53.2 60.0 72.0 88.1

76.5 65.3 40.6 50.0 47.1 91.8

57.0 60.0 38.2 86.1 51.4 90.8 59.8 67.5 54.4 72.6 49.5 18.9 18.9 32.1 92.4 41.2 69.7 53.5 70.4 83.4

55.7 60.4 38.2 87.8 49.9 84.4 57.6 67.5 55.3 70.2

59.6 63.5 37.3 87.3 49.6 73.0 52.9 65.4 57.1 69.6

53.8 68.2 43.8 89.4 51.4 70.6 54.1 65.3 58.9 71.2

17.5 17.5 26.1 80.8 41.3 69.5 54.2 70.3 80.7

18.1 18.1 26.4 81.1 32.7 69.5 56.9 57.0 79.8

18.8 18.8 53.6 83.7 30.5 73.4 56.7 54.5 86.2

73.4 39.1 59.1 71.0 50.2 68.1 62.4 59.9 19.4 19.4 60.0 73.6 32.5 76.5 58.4 55.9 81.6

137 4 34 4 5 4 10 15 8 4 4 7 7 4 8 7 5 5 5 4

Source: Eurosystem. *Number of banks in the October 2014 survey round.

The cross-sectional analysis focuses on the individual responses of banks to questions related to changes in credit standards and loan demand. Given BLS data availability, with the first wave of the survey taking place in Q1 2003, we only consider the 11 countries that had joined the euro area as of Q1 2003 with the exclusion of Luxembourg: i.e. Austria, Belgium, Finland, France, Germany, Greece, Ireland, Italy, the Netherlands, Portugal and Spain. The analysis focuses on the following two questions. The first one relates to changes in credit standards applied to the approval of loans or credit lines to enterprises: “Over the past three months, how have your bank‟s credit standards as applied to the approval of loans or credit lines to enterprises changed?” The second one relates to changes in the demand for loans or credit lines to enterprises: “Over the past three months, how has the demand for loans or credit lines to enterprises changed at your bank, apart from normal seasonal fluctuations?” Figures A.2 and A.3 report the distributions of (i) changes in credit standards for loans to euro area NFCs, and (ii) changes in net demand for loans to euro area NFCs.

33

Figure A.2: Distribution of changes in credit standards for loans to euro area NFCs (percentages of banks)

Source: ECB (BLS).

Figure A.3: Distribution of changes in net demand for loans to euro area NFCs (percentages of banks)

Source: ECB (BLS).

34

Appendix 3 Ordered Probit specification test This appendix presents a statistical analysis and tests whether the order probit estimated in the main text is correctly specified. The test is based on the orthogonality between explanatory variables and difference between events and propensity scores. In particular, it is used the nonparametric test of Angrist et al (2013) on whether, after controlling for observable factors, credit tightening outcomes can be considered as „randomly assigned‟. The null hypothesis of the test is given by the condition that the shock defined as difference between outcomes and their propensity score is orthogonal to the information set z used in the ordered probit: [ (

( ))]

(A.1)

Where, following the notation we have used in the main part of the paper, D are the outcomes („ease‟, „neutral‟, „tightening‟), z the explanatory covariates and p(z) the propensity score computed with the ordered probit. When the null hypothesis in equation (A.1) cannot be rejected, there is no information in variables

that can be used to explain the „shocks‟ (D-p(z)), i.e. the probit model is

well-specified. Following Angrist et al (2013), Table A.2 shows p-values for tests that changes of credit standards are independent of the covariates listed, once conditioned upon the propensity score. Test results shown focus on tightening credit standards, i.e. D=tight. Concerning the easing, tests were not fully conclusive, also owing to the limited observations in the sample. A model which includes only macroeconomic activity variables (reported in column 1) falls short of being correctly specified as its „shocks‟ can still be predicted by using bank level demand factors 24 (p-value close to 0) as well as the VSTOXX and the EBP. Adding VSTOXX and EBP as controls improves the specification, although bank level demand factor still rejects the null hypothesis. Finally, monetary policy indicators (EONIA and OIS Forward) are orthogonal to the credit tightening and they tend to reject the test only when testing the easing part, suggesting a potentially asymmetric risk-taking channel of monetary policy.

Such result is robust to considering a „binned‟ version of macroeconomic variables. In our „binned‟ version a dummy {1,0,-1} is attributed to the variables according to whether they are below (above) their 25 (75) percentile. 24

35

Table A.2 (1)

(2)

Country level Macro variables only

Adding VSTOXX and Change in EBP

0

0

Euro area Lag EONIA

0.254

0.38

Euro area Lagged change in 3m-in-1y OIS Forward

0.198

0.226

Euro area VSTOXX

0.006

-

0

-

Bank level Demand factor

Euro area Change in EBP

NOTE: Table reports P-values results for variables that appear in propensity score models other than the one being tested. Model (1) is an ordered probit predicting changes in credit standards using only country level GDP, expected GDP and unemployment. Model (2) also includes VSTOXX and Change in EBP. The reported P-values have been computed by using a Wild Bootstrap as in Angrist et al (2013).

Finally, the external instrument procedure already controls for possible correlation between the instrument and the lagged variables in the VAR, i.e. if the instrument were perfectly explained by lagged variables in the VAR, VAR residuals would tend to have no or little predicting power on the instrument. However, we also use the test of Angrist et al (2013) to check whether changes of credit standards are also independent of the (lagged) variables which are included in the VAR but were not included in the ordered probit. This is a further confirmation that the bank lending instrument we use cannot easily be predicted using the variables contained in the VAR, irrespective of whether such variables appear in the specification of the ordered probit. In particular, table A.3 reports p-values of testing equation (A.1) when the term z is the lagged endogenous variables of the BVAR: Table A.3 Lagged Outcome Variables

Baseline ordered probit

Bank Credit Growth

0.08

Lending Margin

0.07

Bonds growth

0.12

NOTE: Table reports P-values results for variables that are not contained in any propensity score model but are in the BVAR. Variables are stationarised by taking first log-difference when needed. Model used is the Baseline ordered probit model.

The ordered probit seems also to be well-specified, i.e. condition (A.1) is fulfilled, although with slightly less confidence, also when the first lag of variables in the VAR are included.

36

Appendix 4 Country-specific LSI Figure A.4: Credit standards for the largest four euro area economies as obtained with the LSI (dashed blue line) and BLS (solid blue line) Germany

France

Italy

Spain

Note: Sample period: 2003Q1-2016Q1, The figure depicts the standardised LSI (dashed blue line) together with the BLS change of credit standards (solid blue line). The econometric model used for the LSI is the baseline ordered probit model estimated on pooled BLS data. Individual BLS responses are corrected by taking into account bank-specific loan demand (BLS), macroeconomic conditions (actual and expected) at country level, the riskiness conditions of non-financial corporations in the euro area and monetary policy conditions (EONIA and forward rates).

37

Figure A5: Historical decomposition of selected macro-variables in the 4 largest EA countries: Effects of loan supply shock Germany France

Italy

Spain

Notes: Starting date of the decomposition is 2008Q1. The panels of the figure depict the contribution of the credit supply shock (black dashed line) compared to the historical developments of selected macroeconomic variables (red solid bar). Real GDP, the GDP deflator, Loans to NFCs and Debt securities issuance by NFCs are expressed as y-o-y growth rates in deviation from the baseline of their unconditional forecast.

Appendix 5 The excess bond premium in the euro area The EBP measures the premium requested by investors additional to what available information on bonds and their issuers would be able to explain in normal times (see Gilchrist and Zakrajšek, (2012a), and more recently for Europe, Bleaney et al. (2016) and Gilchrist and Mojon (2018)). We construct our EBP following the bottom-up approach of Gilchrist et al. (2009). In terms of the dataset, we select individual bonds which constitute the Markit iBOXX indices. Our selection criteria exclude bonds with embedded options and bonds that have a residual maturity of less than one year or that are illiquid for other reasons. We focus on euro-denominated investment-grade bonds with a minimum issuance size of 250 million euro. The iBOXX dataset starts in 2004 at a daily frequency and it has been then complemented with information from Moody‟s to date it back to 1999. In terms of bond characteristics, Markit collects for each ISIN the country of origin of each company‟s head office, the sector of the issuer (i.e. whether financial or

38

non-financial), the bond yield, the rating, the residual maturity, the coupon and information on outstanding amounts. We include both financial and non-financial firms in our database and we control for the country of the issuer. Similar to Bleaney et al. (2016), the euro area is considered as an aggregation of seven countries: Austria, Belgium, Germany, Spain, France, Italy and the Netherlands. Additionally, we use the Moody‟s KMV database to gauge one-year-ahead probabilities of default for firms. Individual corporate bond spreads are computed as the difference between the effective yield on the security and the overnight index swap (OIS) rate with the same maturity as the residual maturity of the respective bond. Gilchrist and Mojon (2018) instead use the German Bund zero coupon interest rates. The euro area sovereign debt crisis led to a strong fragmentation of the sovereign markets across the region, with significant “safe-haven” flows especially towards German sovereign bonds. In this respect, the OIS curve may provide better benchmark risk-free interest rates for the euro area and avoid contaminating the corporate bond spreads with specific factors related to liquidity premia in the German sovereign market. Turning to the econometric specification, we employ log-linear regressions explaining logs of credit spreads 1+

by individual bond characteristics

(coupon, residual maturity and logistic

transformation of the rating scale); one-year-ahead probabilities of default (EDF1) and a set of dummies for sectors and countries: (

)

(

)

∑

∑

(9)

where i refers to the individual bond and t to the time period. The regression is performed for the whole sample of bonds, including both financial and non-financial sectors using a sector dummy. Standard errors have been computed by clustering errors by time and country. The sample period runs from January 1999 to March 2016. Table A.4 shows the results. Regarding the bond characteristics

, residual maturity has a positive and statistically significant

effect at 1%. The level of a coupon paid by a bond also significantly affects corporate bond spreads. Ratings also have a sizeable and significant effect. Finally, probabilities of default significantly explain the spreads. To construct the EBP, we aggregate the residuals using as weights the value of the outstanding amount of bond i over total bonds outstanding at time t. Figure A.6 plots the EBP for the nonfinancial sector. The EBP tracks the main episodes of financial stress for non-financial corporations in the euro area starting from 1999. The environment of very low risk perception between 2004 and 2007 is also captured by the EBP indicator. Finally, the EBP peaks the month after the Lehman event and during the sovereign debt crisis in the euro area, in the summer of 2011 and then again in 2012.

39

Table A.4: Corporate bond OLS regression VARIABLES

log(1+S)

Ratings

0.00317*** (0.000203) 0.000855*** (0.000104) 0.000346*** (1.77e-05) 0.00268*** (0.000775)

Coupon Residual maturity Sector dummy

log(1+EDF1)

0.0132*** (0.00178) 0.0156 (0.0118) YES

Constant Country dummies

Observations 45,733 R-squared 0.271 Robust standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1 Note: The dependent variable in each column is log(1+bond spread). The residuals of this regression make up the excess bond premium. Standard errors (in parentheses) are clustered along the country and time dimensions. Country dummies are included, but not reported.

Figure A6: Excess bond premium for NFCs in the euro area

Sources: Markit iBOXX, Moody‟s KMV, Bank of America Merrill Lynch, Bloomberg and authors‟ calculations. Last observation: 2016Q1.

40