Systemic Risk and Credit Risk in Bank Loan Portfolios

Systemic Risk and Credit Risk in Bank Loan Portfolios Yu Shan 1 Department of Economics and Finance, Zicklin School of Business, Baruch College, New York, NY 10010, USA Aug 27, 2017 Abstract I investigate the causal relationship between systemic risk (using two complementary measures of systemic risk, ΔCoVaR and CATFIN) and credit risk in the bank s loan portfolio (the borrower s distance-to-default). Systemically risky banks originate new loans with greater default risk exposure even after controlling for self-selection of lenders by borrowers. Lead-lag analysis and dynamic panel GMM estimation address causality concerns. Using within-loan regressions to remove the impact of demand side factors, I find a positive (negative) relationship between bank systemic risk and borrower default risk for risky (safe) borrowers, such that systemically risky banks increase their loan participation in loans with higher credit risk. 1 Tel.: (01) 213-618-6006; E-mail: yu.shan@baruch.cuny.edu. Address: Department of Economics and Finance, Zicklin School of Business, Baruch College, New York, NY 10010.

1. Introduction An important lesson that policymakers and academicians have taken away from the global financial crisis of 2007-2009 is the importance of distinguishing between individual financial institutions risks in isolation and their systemic risk exposure. Thus, bank regulations that focus exclusively on the individual institutions fail to recognize systemic interrelationships that should form an important component of macro-prudential regulation. Basel III capital requirements incorporate this recognition through the imposition of cyclical capital buffers and TLAC (total loss absorbing capacity) requirements related to aggregate market conditions. However, neither academic research nor regulatory policy papers have addressed the direct connection between individual bank risk taking and systemic risk. In this paper, I consider the impact of systemic risk exposures on the credit risk of the bank s loan portfolio. Credit risk is a major source of risk in bank portfolios. Thus, I examine whether systemically risky banks increase the credit risk in their loan portfolios as either the individual micro-level or the aggregate macro-level measure of systemic risk increases. Banks control the credit risk in their loan portfolios by adjusting their credit and underwriting standards. However, borrowers may also respond to bank risk taking. For example, riskier borrowers may be more dependent on stable refinancing sources especially during economic downturns, and therefore they may tend to request loans from systemically important banks, which may be protected by public guarantees and therefore more likely to survive during periods of economic turmoil. Therefore, I empirically control for both the supply and demand effects of lending at systemically important banks. In this paper, I test the three alternative responses to systemic risk in the credit risk of the loan portfolio. That is, there may be no connection between credit risk and systemic risk, or alternatively banks may either reduce or increase their credit risk in their loan portfolios as their systemic risk increases. Adrian and Brunnermeier (2016) suggest that there is only a very loose cross-sectional link between institution s risk/tail risk and its systemic risk (Adrian and Brunnermeier, 2016). Although not testing for credit risk explicitly, this could be considered evidence that there is no connection between individual bank risk taking and systemic risk. On the other hand, banks may reduce their own risk exposures when a systemic crisis looms in order to pull back from the brink of insolvency. This automatic stabilizer could reduce systemic risk, thereby mitigating some of the negative impact. This may be induced by managerial incentives (Schwarcz (2017) and clawbacks (Allen and Li (2011)). Charter values may also induce banks to control the credit risk in their loan portfolios when systemic crises are more likely to occur. Either implicit or explicit government bail-out guarantees may result in higher charter values for protected 1

banks due to lower refinancing costs (Keeley, 1990; Gropp, Hakenes, and Schnabel, 2011). Systemically risk banks may reduce the credit risk in their loan portfolios in order to protect valuable bank charter values. Alternatively, however, public guarantees and expectations of public guarantees can increase the protected banks risk taking by reducing market discipline (Flannery, 1998; Sironi, 2003; Gropp, Vesala, and Vulps, 2006; Acharya, Anginer, and Warburton, 2016), thereby suggesting the third alternative hypothesis in this paper. That is, banks may exacerbate their credit risk exposure as the likelihood of systemic bailouts is increased in order to reap the benefits of government bailouts, particularly for too-big-to-fail (TBTF) institutions. Banks loan underwriting response to systemic risk is an empirical question. Indeed, focusing on bank charter values (rather than credit risk in the loan portfolio), Cordella and Yeyati (2003) and Hakenes and Schnabel (2010) argue that the net effect of systemic risk and public guarantees is ambiguous. However, if banks regularly mitigate systemic risk by reducing own bank risk exposure, then systemic risk regulations can be less drastic due to this automatic stabilizer. Alternatively, if banks regularly exacerbate their own bank risk exposure by exploiting potential moral hazard advantages, systemic risk regulation will be insufficient to address system-wide crises. Therefore, investigating the endogenous control of systemically important banks over their risk positions has important policy implications. This is particularly relevant in light of recent regulatory interventions during the 2007-2009 crisis. The explicit and implicit cost of funds from government support may actually be quite high. First, the injected preferred equity has priority over common equity, which may amplify the losses of common equity holders in the event of failure, resulting in a reduction in the value of common equity, and increasing the difficulty of raising equity (Berge, Roman and Sedunov 2016). Berger and Roman (2015) suggest that the TARP funds may be relatively expensive. If banks shareholders are concerned about the negative implications of bailouts, they may induce the bank to adjust the credit risk of their loans in order to pull back from the brink of delinquency and bailout. Second, the supported banks may have to undertake certain social welfare responsibilities through lending expansion, which may not be an equilibrium choice during economic downturns. Recent experience with government intervention in the financial system may induce banks to reduce their credit risk exposure in order to mitigate the bank s own risk of insolvency, thereby reducing the likelihood that 2

bailouts will be needed. 2 This paper empirically examines the relationship between systemic risk and credit risk exposure in bank loan portfolios to address these important public policy issues. Using data on syndicated bank loans, this paper investigates whether financial institutions adjust the credit risk in their loan portfolios risks in response to levels of systemic risks. I use two complementary measures of systemic risk in this paper, the macro-level systemic risk and the microlevel systemic risk exposure of each individual bank. To measure the macro-level aggregate systemic risk, I utilize CATFIN (Allen, Bali and Tang 2012). CATFIN is a cross-sectional measure that identifies the overall level of systemic risk in the financial system at each point in time. This allows me to examine whether high levels of aggregate systemic risk that are likely to lead to government bailouts or other interventions induce banks to adjust the credit risk in their loan portfolios upward or downward. To measure the micro-level systemic risk, I utilize ΔCoVaR (Adrian and Brunnermeier 2016) to determine the impact of an individual bank s insolvency on the overall financial system. The greater the contribution of an individual bank s insolvency to market-wide systemic risk, the greater the individual bank s imposition of systemic risk onto the macroeconomy. I utilize this measure to determine whether the bank s systemic risk exposure impacts its ability to lend to borrowers of differing risk levels. There are several findings in this paper. I find that, in syndicated loan markets, systemically risky banks are matched with risky borrowers, implying banks with higher levels of systemic risk choose to lend to borrowers with higher levels of default risk. The result is robust to a number of different specifications, which control for loan, borrower, and bank characteristics, as well as a series of fixed effects and endogeneity controls. However, the relationship between bank loan portfolios credit risk and systemic risk may be induced by either supply or demand effects. Thus, I test my findings of a direct relationship between systemic risk and credit risk using a within-loan regression analysis that controls for borrower self-selection lenders (see Chu, Zhang and Zhao (2017)). Taking advantage of the unique feature that a syndicated loan often has multiple lenders, I examine how the systemic risks of banks that fund the same loan affects their contributions to the loan, i.e., within-loan estimation, which eliminates any fluctuation on the demand side. My major result persists after using within-loan 2 In the wake of the 2007-2009 financial crisis, greater regulatory costs, stricter security, and higher disciplinary pressure has been imposed on systemically important banks. For example, G-SIBs are subject to higher capital buffer requirements, Total Loss-Absorbing Capacity (TLAC) standards, resolvability requirements, and higher supervisory expectations. With all these higher regulatory requirements, systemically important banks may be at a competitive disadvantage relative to other banks, leading them to reduce their lending to high risk borrowers that may encourage regulatory scrutiny. For example, regulatory warnings regarding leveraged loans in 2014 induced banks to cut back on lending to high risk borrowers. Alternatively, however, systemically risky banks may choose to lend to riskier borrowers in order to recoup their higher costs of bank capital. 3

estimation to control for borrower self-selection of lenders. I find that for risky (safe) borrowers, the systemic risk of banks participating in the same syndicated loan are positively (negatively) associated with the bank s contribution to the loan, thereby implying that systemically risky banks as loan suppliers increase their credit risk-taking in their loan portfolios. To alleviate the reverse causality concern, I follow Wintoki, Linck, and Netter (2012) by utilizing a dynamic panel GMM estimation, which confirms my previous findings. The rest of the paper is organized as follows. Section 2 describes the data and sample construction. Section 3 presents the empirical results, including a two-stage least squares analysis controlling for borrower self-selection, a lead-lag analysis, and the within-loan estimation. Section 4 provides robustness checks, including a different version of within-loan regression and a dynamic panel GMM analysis. Section 5 concludes. 2. Data and Variable Constructions 2.1. Borrowing Firm Distance-to-default In this paper, I use the Merton distance-to-default as a measure of borrower default risk. I follow Bharath and Shumway (2004), Crosbie and Bohn (2003), and Drucker and Puri (2009) in calculating Merton s distance-to-default. The market equity value of a company is modeled as a call option on the company s assets: V E = V A e dt N(d 1 ) Xe rt N(d 2 ) + (1 e dt )V A (1) d 1 = log (V A X ) + (r + s 2 A 2 )T ; d s 2 = d 1 s A T (2) A where V E is the market value of a firm s equity, which is calculated from the CRSP database as the product of share price at the end of the quarter and the number of shares outstanding. X is the face value of debt maturing at time T, which is calculated as debt in current liabilities (COMPUSTAT data item 45) plus one half of long term debt (COMPUSTAT data item 51). V A is the value of the firm s assets. r is the risk-free rate, which is defined as the 1-year Treasury Constant Maturity Rate obtained from the Board of Governors of the Federal Reserve system. s A is the volatility of the value of assets. I simultaneously solve the above two equations to find the values of V A and s A. Quarterly Merton s distance-to-default is finally computed as: 4

Distance to Default = log (V A X ) + (m s 2 A 2 )T s A T (3) The default probability is the normal transform of the quarterly distance-to-default measure, defined as PD = F( Distance to Default). As a robustness check, I also calculate quarterly distance-to-default by first calculating the monthly distance-to-default and then taking quarterly average. Both methods generate similar results in my following analysis. 2.2. Systemic Risks To measure the micro-level systemic risk, I follow the methodology used in Adrian and Brunnermeier (2016) to generate time-varying ΔCoVaR. First, I run the following quantile regressions in the weekly data (where j is a financial institution): X j t = α j q + γ j q M t 1 + ε j q,t, (4) X system j t = α system j q M t 1 + β system j q M t 1 + β system j q X j t + ε system j q,t, (5) where X j t is the weekly return of institution j in week t, X system j t is the financial sector return in week t, and M t 1 is a vector of seven systematic state variables in week t, including three-month yield change, term spread change, TED spread, credit spread change, market return, real estate excess return, equity volatility. Then I generate the predicted values from these regressions to obtain j VaR q,t = α q j + γ q j M t 1, (6) j CoVaR q,t = α q system j + r q system j M t 1 + β q system j VaR j q,t, (7) j Finally, compute the ΔCoVaR q,t for each institution: j ΔCoVaR q,t j = CoVaR q,t j CoVaR 50,t = β system j j (VaR q,t j VaR 50,t ) (8) From these regressions, I get a panel of weekly ΔCoVaR j q,t. Then obtain a quarterly time series of j ΔCoVaR q,t by averaging the weekly risk measures within each quarter. Throughout the paper, I use q equals 99%, but I ve test all the analysis using 95%, and get similar results. 5

To construct the macro-level systemic risk, CATFIN 3, I follow Allen, Bali, and Tang (2012) and first estimate VaR at the 99% confidence level using three different methodologies - the generalized Pareto distribution (GPD), the skewed generalized error distribution (SGED) and the non-parametric estimation method based on the left tail of the actual empirical distribution without any assumptions about the underlying return distribution. CATFIN is defined as the arithmetic average of the GPD, SGED and non-parametric VaR measures. Allen, Bali, and Tang (2012) suggest that the risk of macroeconomic downturns increases when CATFIN is above some early warning level, where early warning level is determined by using Chicago Fed National Activity Index (CFNAI) as a benchmark. CFNAI is an index of eighty-five existing monthly economic indicators. The Federal Reserve Bank of Chicago denotes the three-month moving average of CFNAI (CFNAI-MA3) value of -0.7 as a turning point indicating economic contraction, and Allen, Bali, and Tang (2012) show that when CATFIN is above some early warning level, it can significantly predict lower economic activity (CFNAI-MA3) one month to twelve months in advance of the downturn. Therefore, CATFIN offers an early warning to alert regulators to the risk of economic recessions. In this paper, I also test whether the banks loan risk-taking is affected by whether CATFIN breaches the early warning level. Following Allen, Bali, and Tang (2012), I construct an early warning dummy, which is equal to 1 if CATFIN is above the early warning level, and 0 otherwise. For each quarter t, the early warning level is calculated as the median CATFIN using all observations up to quarter t in which CFNAI-MA3 falls below -0.7. To address possible reverse causality, I utilize both contemporaneous and lagged measures of systemic risk to examine the following quarter s credit risk in the bank s loan portfolio. A historical monthly CATFIN, CFNAI-MA3, Early Warning Level, and Warning dummy are presented in Table 12. 2.3. Control Variables I use a set of control variables to control for loan characteristics, borrower characteristics and lender characteristics. To ensure that statistical results are not heavily influenced by outliers, I set all observations higher than the 99th percentile of each variable to that value, and all values lower than the 1 st percentile of each variable are winsorized in the same manner. The definitions for all variables are defined in the appendix. 3 I thank Linda Allen and Yi Tang for providing the data on their CATFIN systemic risk measure and Tobias Adrian and Markus Brunnermeier for making their measure of systemic risk ( CoVaR) available. 6

The first set of control variables include bank characteristics, such as Bank Total Assets, Bank Capital Ratio, Return on Equity, Liquidity, Loan Charge-offs, Loan Loss Allowance, and Risk-Weighted Assets. Since ΔCoVaR is constructed using market data and most public banks are bank holding companies, I measure bank financial information at the bank holding company level. ln(bank Total Assets) is defined as the natural logarithm of bank total assets (in billions); Bank Capital Ratio is defined as the bank s total capital over bank s total assets; Bank ROE is defined as bank net income over book equity; Bank Liquidity is defined as the sum of cash and available-for-sale securities divided by bank total assets; Loan Charge-offs is defined as the total charge-offs on loans and leases divided by bank total assets; Loan Loss Allowance is defined as the total allowance for loan and lease losses divided by bank total assets; Risk-Weighted Assets is defined as total risk weighted assets divided by bank total assets. Although I include both lead lenders and participants in all my regressions, I add a lead bank dummy in the within-loan regressions to account for possible unobservable differences between lead and non-lead banks. I define a bank as a lead lender if its lend arranger credit variable is Yes in Dealscan. The second set of control variables is borrower characteristics. ln(borrower Total Assets) is defined as the natural logarithm of total assets (in billions); Tangibility is defined as total property, plant, and equipment divided by total assets; Leverage is defined as the total debt divided by total assets. I also include a lending relationship measure, which is borrowed from Bharath et al. (2007). The reason to include a lending relationship measure is that the intensity of the lending relationship can have great impact on the borrowerlending matching. The lending relationship between borrower i and bank j is defined as the dollar amount of loans to borrower i by bank j in last 5 years over the total dollar amount of loans by borrower i in last 5 years. Finally, I control for an array of loan characteristics. ln(package Amount) is defined as the natural log of the package amount, where package amount is measured in millions. ln(package Maturity) is defined as the natural log of the maturity of the deal in months. Package maturity is calculated as the value-weighted average of the facility maturities. ln(no. of Lead Banks) is the natural log of the number of lead lenders in the deal syndicate. 2.4. Sample Construction The sample period of my study spans Q1 1995 to Q4 2013. Banks financial data are collected from the Consolidated Financial Statements for Holding Companies ( FR Y-9C ) available on the Federal Reserve Bank of Chicago website, and all market data are obtained from CRSP. Borrowing firms financial data are collected from Compustat. Syndicated loan data is obtained from Loan Pricing Corporation s (LPC) 7

Dealscan loan database. The Dealscan database contains historical information on the terms and conditions of deals in the global commercial loan market. Borrower financial data are linked to Dealscan using the Dealscan-Compustat linking data provided by Roberts and Sufi (2009). All observations are quarterly. To construct the sample, I first start with a sample of 219,023 deal packages newly originated between Jan. 1995 and Dec. 2013 from the LPC Dealscan database. Since in this paper I conduct my analysis at the bank holding company level, I need to first identify the lenders in my sample and their ultimate parent companies. I utilize the information provided by the Federal Reserve System via its National Information Center (NIC) database to identify financial institutions acting as lenders in my sample. I do not use the identity variables for lender and ultimate owner in Dealscan because Dealscan overwrites the ultimate owner of the lenders after mergers and acquisitions, i.e., the ultimate owner in Dealscan is the ultimate owner at the end of the merger chain. Whereas, for this paper, I need to identify the ultimate owners at the time of the loan issuance for analysis at the holding company level. The NIC database provides detailed information about financial institutions, including types of institutions, establishment time, ownership information, address changes, name changes, merger and acquisition history. NIC also provides each financial institution s RSSD ID, a unique identifier assigned to each financial institution by the Federal Reserve System. Based on the lender information provided by Dealscan, including name, location and lending history, I manually find each lender s RSSD ID. Using their RSSD ID, which is item RSSD9001 in Call Report and Y-9C, the lender s ultimate owner at the time of loan origination is determined by cross-checking the information contained in Call Report items RSSD ID of Regulatory High Holder 1 (RSSD9348), Financial Higher Holder ID (RSSD9364), and Financial High Holder Percent of Equity (RSSD9365). The three items provide the RSSD ID (RSSD9001) of a lender s ultimate bank holding company during the time when a lead arranger has a RSSD ID. For a bank that was acquired by another bank and lost its RSSD ID but kept its lending activity, the acquirer s RSSD ID was applied to the acquired bank as its new RSSD ID, and the new ultimate owner can be found from the Call Report and Y-9C items RSSD9348, RSSD9364, and RSSD9365. The full bank and bank holding companies merger and acquisition history is obtained from Federal Bank of Chicago. Using the RSSD ID, I link Dealscan to the Y-9C to obtain bank financial data. I also link Dealscan to CRSP to collect market data through the PERMCO-RSSD link table provided by the Federal Reserve Bank of New York. I collect bank characteristics data at the bank holding company level using Y-9C reports. Using the PERMCO-RSSD link, I merge bank holding companies with their systemic risk measure, ΔCoVaR. This process reduces my sample to 80,193 packages (140,609 package-lender pairs). Next, I merge borrower characteristics from Compustat with the information on corporate loans in Dealscan using the linking table provided by Roberts and Sufi (2009). This table matches loan facilities from Dealscan with the borrower s GVKEY identifier in Compustat. Due to differences in 8

capital structures and financing strategies between financial and non-financial firms, I exclude loans to financial companies (SIC between 6000 and 6999) from the sample. I also exclude the utility firms (SIC code falls between 4900 and 4999) because they may have different operating and reporting environments. This leads to a sample of 10,915 packages (22,595 package-lender pairs), which include 3,603 unique borrowers, and 214 unique banks, which are owned by 66 unique public traded bank holding companies with ΔCoVaR data. Since systemic risk ΔCoVaR is mostly measured at the bank holding company level, in my subsequent analysis, I run all regressions at the loan-bank holding company level. 2.5. Summary Statistics for Baseline Regressions I present the summary statistics in Table 1. There are in total 22,595 lender-package observations in the baseline regressions. I only use package level data in my baseline regression. The key dependent variable is borrower s Distance to Default, which has a mean of 6.772, a median of 6.002, and a standard deviation of 4.903. The key independent variables are ΔCoVaR, which has a mean of 5.216, a median of 4.351, and a standard deviation of 2.550, and CATFIN, which has a mean of 2.393, a median of 2.286, and a standard deviation of 0.926. The average deal amount is 878 million, with average maturity of 49 months. On average, there are 4.141 lead lenders in each package. The Lending Relationship between borrower i and lender j, which is defined as the dollar amount of loans to borrower i by bank j in last 5 years over the total dollar amount of loans by borrower i in last 5 years, has a mean of 0.47, indicating that on average each bank engaged in 47% of the total amount a typical firm borrowed in the 5 years preceding the loan origination. Borrowers have an average size of 6.752 billion, with a mean tangibility of 0.310 and leverage of 0.322. Banks have a mean size of 779 billion, with mean capital ratio of 8.5% and return on equity of 8.3%. Table 2 presents the spearman correlation matrix for the variable included in the baseline regressions. As shown in Table 2, borrower distance-to-default is negatively correlated with both ΔCoVaR (the bank-specific measure of systemic risk) and CATFIN (the aggregate measure of systemic risk), with correlation coefficients of -0.220 and -0.289, respectively. This provide a preliminary evidence that banks with higher levels of systemic risk, or banks that are lending during periods of high aggregate systemic risk, lend to borrowers with higher credit risk. 2.6. Sample Construction for Within-loan Regressions To investigate whether the relationship between systemic risks and borrower default risk is driven by the bank side, in Section 3 and Section 4, I apply the within-loan estimations methodology from Chu, Zhang and Zhao (2017). By adding package or facility fixed effects, I can remove the impact of the demand side factors from the supply side factors. 9

The sample of within-loan regressions is constructed slightly differently with the one used in the baseline regressions. First, since all borrower, loan, and macroeconomic characteristics drop out in a package or facility fixed effects regression, only lender characteristics are relevant to the within-loan regressions. Therefore, I don t require a loan to have borrower characteristics to be included in the sample. Second, to control for an array of lender characteristics, I add more lender variables. Besides the bank size, capital ratio, return on equity, I also include liquidity, loan charge-offs, loan loss allowance, and risk-weighted assets into the within-loan regressions. These two differences to some extent change the observations of my sample. Although I include both lead lenders and participants in all my regressions, I add a lead bank dummy in the within-loan regressions to account for possible unobservable differences between lead and non-lead banks. I define a bank as a lead lender if its lend arranger credit variable is Yes in Dealscan. To construct the sample for within-loan estimation, I start with a sample of 315,963 loan facilities between Jan. 1995 and Dec. 2013 from the Loan Pricing Corporation Dealscan database. These 315,963 facilities belong to 219,023 deal packages. Then I manually find the RSSD ID of the bank lenders through NIC database, based on lender names and the dates of earliest and latest lending behavior, and drop all observations that doesn t have an identifiable lender RSSD ID. Next, I identify the ultimate holding companies of those banks from Y9-C, following the same methodologies introduced in Section 2.4. This reduce the sample size to 147,224 facilities (101,409 packages). Then I restrict my sample to loans from banks with non-missing market equity data, systemic risk data and Y-9C data in quarter t and t-1, where t denotes the quarter of loan origination. This leads to a sample of 93,055 facilities (63,070 packages), which was originated by 173 distinct banks. These 173 banks belong to 75 distinct bank holding companies. I then further exclude loans to non-u.s. firms, utilities and financial sectors (two-digit SIC code equal to 49 or between 60 and 69), and restrict sample to firms that have non-missing distance-to-default in quarter t and t-1, which reduce the sample to 18,507 facilities (12,899 packages). The key variable of interest is bank s allocation share in a loan. For a portion of loans, Dealscan reports lending bank s allocation share, which captures the bank s contribution toward funding the loan. Since I use a bank s allocation share to measure bank risk-taking at the package or facility level, I exclude loans whose bank allocation share information is missing. Finally, I made two adjustments to my sample. First, I exclude facilities/packages that have only one participant from the sample, because their allocation share is always 100 percent. Second, since I include lender fixed effects, to make sure there is enough variation for each lender, I only keep lenders who at least have five loan 10

observations in my sample. This leads to a final sample of 3,307 packages (11,886 package-bank holding company pairs) and 4,182 facilities (15,167 facility-bank holding company pairs). To conduct within-loan regressions, I sort borrowing firms into two groups based on whether their distanceto-default is higher or low than the median level in each year, and run within-loan regressions for the two groups separately. If systemically risky banks increase their risk taking by increasing their lending to riskier borrowers, then I should observe a significant positive between systemic risk and allocation share for loans to low distance-to-default (high credit risk) borrowers, and a significant negative relationship for high distance-to-default default borrowers. Panel A (Panel B) in Table 3 present the summary statistics for the two groups at the package (facility) level. Panel A1 (A2) reports the summary statistics for packages borrowed by the low (high) distance-to-default firms. Panel B1 (B2) reports the summary statistics for facilities borrowed by the low (high) distance-to-default firms. The mean and median of key variables, e.g. ΔCoVaR and other bank characteristics are close to those in the baseline regressions. 3. Empirical Results 3.1. Baseline Results I first present the baseline results. I estimate the following model: Distance to Default i,j,k,t = α 0 + α 1 ΔCoVaR j,t + α 2 CATFIN t + α 3 ΔCoVaR j,t CATFIN t + α 4 ΔCoVaR j,t 1 + α 5 Distance to Default i,t 1 + α 6 Loan Controls k,t + Bank Controls j,t + α 8 Borrower Controls i,t + α 9 Lending Relationship i,j,t + α 10 Macro Control t + Year FE + Industry FE + Bank FE + ε i,j,k,t (9) where subscript i, j, k and t indicate the firm, the bank, the package, and the time (quarter), respectively. The regressions are run at the package level. The dependent variable, Distance to Default, is borrower s distance-to-default at the quarter of loan origination. The explanatory variables of interest are ΔCoVaR and CATFIN, as well as their interaction. The vectors of variables Loan Controls, Bank Controls, and Borrower Controls contain loan, bank, and firm-specific control variables from the quarter of loan origination. In all regressions, I include calendar year fixed effects to remove time trends, and industry and bank fixed effects to remove time-invariant factors that drive matching between borrowers and lenders. I control for bank size to distinguish the effects of size and systemic risk on borrower-lending choice. 11

I also control for bank capital ratio and return on equity since undercapitalized or less profitable banks may refrain from lending to risky borrowers. I include controls for borrower characteristics (asset tangibility, size, and leverage) to mitigate the impact of demand-side factors that are correlated with both the firms distance to default and firm s choice of bank. Loan Controls include the natural logarithm of deal amount, maturity, and number of lead lenders. To distinguish secured loans from unsecured loans, I also add a dummy variable, which is equal to 1 if the loan is secured, and 0 otherwise. Since relationship lending affects lender-borrower matching, I include the lending relationship measure from Bharath et al. (2007). The lending relationship between borrower i and bank j is defined as the dollar amount of loans to borrower i by bank j in last 5 years over the total dollar amount of loans by borrower i in last 5 years. Finally, I add quarterly GDP growth rate to control for macroeconomic conditions. Table 4 reports the regression results. In Column I and II, I only include the variables of interest and fixed effects in the regressions. In Column III and IV, I add all other control variables. The results in Column I to IV show a significant inverse relation between borrower distance-to-default and the bank s microlevel measure of systemic risk. The coefficients on ΔCoVaR and lagged ΔCoVaR in all four columns are negative and significant at 1% level. The results show that, higher bank systemic importance is associated with lower borrower distance-to-default. That is, systemically important banks tend to lend to borrowers with high default risk. The coefficients on both lagged and contemporaneous CATFIN are also negative and significant at the 1% level, indicating that credit risk in bank loan portfolios is high (i.e., borrowers distance to default is low) during periods of high aggregate systemic risk. Economically, taking the coefficients of -0.078 and -0.452 in Column IV, starting from mean value, one standard deviation increase (2.550) in ΔCoVaR is associated with a decrease in distance-to-default by 0.199, which a 0.041 standard deviation (4.903) decrease in borrower s distance-to-default and a 2.937% decrease from the mean value (6.772) of distance-to-default; One standard deviation increase (0.926) in CATFIN is associated with a decrease in distance-to-default by 0.419, which is a 0.085 standard deviation (4.903) decrease in borrower s distance-to-default and 6.190% decrease from the mean value (6.772) of distance-to-default. To perform a quasi-diff-in-diff analysis, I interact ΔCoVaR and CATFIN in Column V to investigate how aggregate systemic risk affects the relation between ΔCoVaR and borrower distance to default. The signs, magnitude, and significance of the coefficients on ΔCoVaR and CATFIN remain. The coefficient on the interaction term is positive and significant at 1% level, although its magnitude is much smaller than the coefficients on ΔCoVaR and CATFIN. Actually, even though the sign of the interaction term is positive, unless CATFIN is extremely high, the net effect of changes in ΔCoVaR on borrower distance-to-default is still significantly negative with large economic magnitude. For 12

example, when CATFIN is at its median level (2.286), one standard deviation increase in ΔCoVaR is associated with 0.073 standard deviation decrease in borrower s distance-to-default, which is an even larger effect than the ΔCoVaR in Column IV. That means only when CATFIN is much higher can it reduce the negative relationship between borrower distance-to-default and ΔCoVaR, which indicates that banks with high individual levels of systemic risk reduce the credit risk in their loan portfolios during periods of very high levels of aggregate systemic risk. The aggregate effect of an increase in both ΔCoVaR and CATFIN is also strong. For example, a simultaneous one standard deviation increase in both ΔCoVaR and CATFIN is associated with a 0.135 standard deviation decrease in borrower s distance-to-default. These results suggest that banks reduce their individual credit risk exposure when aggregate systemic risk is high, thereby lending to safer borrowers when the risk of recession increases. That is, the negative relation between ΔCoVaR and borrower distance to default is weakened during periods of high CATFIN. As robustness checks, I perform another two tests. First, in Column VI, I replace CATFIN by a Recession dummy, which is equal to 1 if the quarter of loan origination is an economic recession quarter, and 0 otherwise. The result is similar as in Column VI. The relationship between ΔCoVaR and borrower distance to default is weakened but still remains. In Column VII, I replace CATFIN by an Early Warn dummy, which equal to 1 when CATFIN exceeds its early warning level, and 0 otherwise. As shown in Column VII, the result remains similar as Column VII. Economically, taking the coefficient of -0.195 on ΔCoVaR in Column VII, one standard deviation increase in ΔCoVaR is associated with 0.101 standard deviation decrease in borrower s distance-to-default. If the loan is originated during periods when the early warning level is breached, then one standard deviation increase in ΔCoVaR is only associated with 0.047 standard deviation decrease in borrower s distance-to-default. The results in Column V, VI, and VII indicate that, during periods of very high levels of aggregate systemic risk, recession or when the early warning has been triggered, banks with high individual levels of systemic risk reduce the credit risk in their loan portfolios. This result is consistent with some mitigation of moral hazard by systemically important banks during crisis (or imminent crisis) periods, as these banks attempt to pull back from the brink by reducing the credit risk in their loan portfolios. Note that the coefficients on bank total assets remain insignificant in all columns of Table 4. Large banks do not necessarily take on higher credit risk in their loan portfolios. This result, together with the coefficient on ΔCoVaR, indicates that moral hazard incentives impact all systemically important banks rather than only big banks. 3.2. 2SLS: Self-selection and Lender-borrower Matching 13

In the baseline regressions, the borrower self-selection and lender-borrower matching is endogenous. In other words, the probability of initiating a new loan between firm i and bank j is endogenous. It could be that some unobservable borrower characteristics induce borrowing firms to choose to apply for loans from certain banks, thereby introducing selection bias into the OLS analysis. Therefore, I must control for the probability of originating a loan between firm i and bank j that is independent of the channel of systemically risky banks selecting borrower with certain default risks. That is, the lending relationship measure used as a control variable in my baseline regressions is itself endogenous. In order to reduce the endogeneity, I employ the geographic distance and number of banks in the state of the borrower as instruments for observed lending relationships. Geographic distance measured as the distance in thousands of kilometers between the location of the firm and the location of the parent company of the lending bank in the quarter of loan origination. Geographic distance proxies for information asymmetries and transportation costs (Degryse and Ongena, 2005) which impede the bank s ability to monitor a firm s financial condition. This may reduce the likelihood of a loan origination. The number of banks in the borrowing firm s state of incorporation is measured as the number of financial institutions that filed Call Report during the quarter of loan origination. Both instruments should affect the probability of originating loans between a certain pair of borrower and lender, but is unlikely to affect how systemically risky banks are matched with borrowers with differing default probabilities. In the first stage, I regress the lending relationship variable on geographic distance and number of banks, and other independent variables. Table 5 reports the regression diagnostics for all specifications in Table 6, and Table 6 reports the results for the first stage regression. In Table 5, I first report the Anderson LM test statistic for tests of identification. The null hypothesis tested is that the instruments and endogenous variable are not correlated and, in addition, that the overidentifying restrictions are valid. The p-values is close to 0, which strongly rejects the null hypothesis. Then, I report the Sargan's chi-square statistic, which tests the joint null hypothesis that the excluded instruments are valid instruments (i.e., uncorrelated with the error term) and correctly excluded from the estimated equation. The Sargan s chi-square test statistics are insignificant in all specifications. This implies that the excluded instruments are valid instruments and correctly excluded from the estimated equation. Table 6 shows that the distance coefficient is significantly negative in the lending relationship first stage regression. The further the borrower is from the lender, the less likely that they will have a lending relationship. The coefficient on No. of Banks is negative but insignificant. The negative sign 14

is very intuitive, the more banks operating in the state of the borrower, the less likely that the borrower and lender will have a lending relationship. Table 7 reports the results for the second stage regression. I use the fitted value of lending relationship in all specifications. Generally, the results in all columns still supports my previous finding. Although the lagged systemic risk, CoVaR j,t 1, is less significant, the contemporaneous systemic risks, CoVaR j,t and CATFIN t, remains significant with similar economic magnitude as found in the baseline regressions. Taking the coefficients of -0.107 and -0.540 in Column IV as an example, starting from mean value, one standard deviation increase (2.550) in ΔCoVaR is associated with a decrease in distance-to-default by 0.273, which a 0.055 standard deviation (4.903) decrease in borrower s distance-todefault and a 4.030% decrease from the mean value (6.772) of distance-to-default; one standard deviation increase (0.926) in CATFIN is associated with a decrease in distance-to-default by 0.500, which is a 0.102 standard deviation (4.903) decrease in borrower s distance-to-default and 7.380% decrease from the mean value (6.772) of distance-to-default; One standard deviation increase (0.926) in CATFIN is associated with a decrease in distance-to-default by 0.500, which is a 0.102 standard deviation (4.903) decrease in borrower s distance-to-default and 7.380% decrease from the mean value (6.772) of distance-to-default. 3.3. Lead-lag Analysis Previous results indicate that systemically risky banks are matched with risky borrowers in the syndicated bank loan market. However, it is unclear whether the causality extends from systemic risk to credit risk or vice versa. That is, borrower s may self-select on the basis of the bank s systemic risk, particularly if government bailouts may protect bank customers from the impact of bank insolvency. To address this, I use a set of lead-lag regressions, using contemporaneous borrower s distance-to-default or lender s ΔCoVaR as the dependent variables, and test whether they are associated with lagged lender s ΔCoVaR or lagged borrower s distance-to-default. The intuition is that if higher lagged systemic risk leads the higher borrower credit risk measured by distance-todefault, it suggests that banks are very likely making endogenous decision on their loan portfolio credit risk-taking conditioning on their systemic risk. In contrast, if borrower s higher credit risk leads the bank s systemic risk exposure, it suggests that borrowers are endogenously choosing banks. Table 8 presents the results of the lead-lag analysis. The dependent variable in Column I, II, III, and IV is Distance to Default i,t, the borrower s distance-to-default at the quarter of loan origination, and the dependent variable in Column V, VI, VII, and VII is CoVaR j,t, bank j s systemic risk during the quarter of loan origination. The key variable of interest is CoVaR j,t 1 and 15

Distance to Default i,t 1. In Column I, II, V, and VI, I only include variables of interest as independent variables, while in Column III, IV, VI, and VII, I include all characteristics of loans, banks, and borrowers as control variables. From Column I to IV, the coefficients on CoVaR j,t 1 are all significant at 1% level, indicating that lagged bank systemic risk significantly is related to contemporaneous borrower distance-to-default. Economically, taking the coefficient on CoVaR j,t 1 in Column III, one standard deviation increase in CoVaR j,t 1 is associated with 0.063 standard deviation decrease in borrower s distance-to-default at loan origination. However, in Column V to VIII, the coefficients on Distance to Default i,t 1 are significant in three out of four specifications. This suggest that while higher lender s systemic risk in quarter t-1 is associated with higher borrower s credit risk in quarter t, higher borrower credit risk in quarter t-1 is not necessarily associated with higher lender s systemic risk in quarter t. In other words, the borrowers that are riskier in the quarter preceding loan origination do not necessarily borrow from systemically risky banks at loan origination, and systemically risky banks in the quarter preceding loan origination lend to risky borrowers at loan origination. Even though this test cannot separate the supply-side effect from the demand-side, it still provides evidence indicating that systemically risky banks may endogenously select risky borrowers, whereas risky borrowers may not intentionally choose banks with greater exposure to systemic risk. In the next section, I employ within-loan estimations to remove the impact of the demand-side factors from the supply-side in order to address these causal effects. 3.4. Within-loan Regressions Even though previous results show a negative relationship between bank systemic risk and borrower default risk, I still cannot identify the causal effect. Both supply side and demand side factors can lead to my results. First, systemically risky banks may increase their risk taking by proactively choosing risky borrowers. Second, due to borrowers self-selection, risky borrowers may tend to apply for loans specifically from systemically risky banks that have a higher likelihood of receiving government support to survive through crisis period. It is difficult to separate the effect of systemic risk on borrower default risk from demand side factors. To solve that problem, I use the within-loan estimations methodology from Chu, Zhang and Zhao (2017) to remove the impact of the demand side factors, so that I can only focus on the supply side. Taking advantage of the unique feature that a syndicated loan often has multiple lenders, I examine how the systemic risks of different banks that fund the same loan impact their loan share percentages. Because all lenders lend to the same borrower, this removes demand side factors from the analysis. 16

The within-loan estimation methodology takes advantage of the underwriting process of a syndicated loan. In a syndicated loan, the lead lender (arranger) originates the loan, negotiates the spreads and terms with the borrower, and attracts other banks to participate. Conditional on the loan demand, a participant in the syndicated loan can therefore determine its own contribution to funding the total loan amount. I regress the natural logarithm of bank share in a package/facility on the banks systemic risk, control variables for bank characteristics, and package/facility fixed effects: Log(Bank Share) i,j,k,t = α k + α 1 ΔCoVaR j,t 1 + α 2 Bank Controls j,t 1 + ε i,j,k,t (10) where subscript i, j, k, and t indicate the firm, the bank, the packages/facilities, and the time (quarter), respectively. α k denotes the package/facility fixed effects. Bank Controls is a vector of control variables for bank characteristics. Since all borrower, loan, and time characteristics drop out when package/facility fixed effects are included, the within-loan regression removes any confounding effects due to unobservable borrower characteristics or endogenously determined loan characteristics. Therefore, any remaining differences in the relative loan shares between different lenders within a package/facility are likely to be driven by lender-side factors. To account for possible unobservable differences between lead and non-lead banks, I add a lead bank dummy in the regressions, which is equal to 1 if a bank is the lead bank in the package/facility, and 0 otherwise. I define a bank as a lead lender if its lender credit variable is Yes in Dealscan. I also include the lending relationship measure borrowed from Bharath et al. (2007) to account for the effect of relationship lending on banks share. Each year, I sort borrowing firms into two groups based on whether their distance-to-default is higher or low than the median level in that year, and run equation (10) for each of the groups. If systemically risky banks increase their risk taking by increasing their lending to riskier borrowers, then we should observe a significant positive α 1 for loans to low distance-to-default (high credit risk) firms, and a significant negative α 1 for loans to high distance-to-default default firms. Table 9 shows the results of the estimation. As hypothesized, the coefficient estimates for α 1 are positive in all low distance-to-default regressions and negative in all high distance-to-default regressions. In Column I to IV of Table 9, I estimate the within-loan model at the facility level. In Column I and II, I include the control variables for bank characteristics used in Equation (9), and in Column III and IV, I add more control variables to mitigate the concern that bank systemic risk may be correlated with other bank characteristics. All variables are defined in the appendix. In Column I, I run equation (10) for loans borrowed by firms with distance to default lower than the median, 17