Covenant Violations, Loan Contracting, and Default Risk of Bank Borrowers

Transcription

1 Covenant Violations, Loan Contracting, and Default Risk of Bank Borrowers Felix Freudenberg Björn Imbierowicz Anthony Saunders* Sascha Steffen March 2012 Abstract This paper investigates the consequences of covenant violations for subsequent loans to the same borrower using a hand-collected sample of loans extended to U.S. firms during the 1996 to 2010 period. We find that borrowers who violate covenants have a default probability of 30% on the day following the violation. We further show that loan spreads increase by 18 basis points in the next loan. We also find that the new contract includes more and tighter financial covenants and is overall more restrictive. These results are consistent with a higher degree of active monitoring following covenant violations. Lenders increase the number of contingencies and situations in which they gain control rights over the firm as a response to the higher default risk of the borrower. We also find that borrowers who have violated financial covenants in the previous contract are significantly more likely to violate covenants again in the next loan. Moreover, we find that borrowers who have violated covenants in the previous contract are more likely to switch to new lenders. Keywords: Covenant Violation, Financial Default, Syndicated Loans. Goethe University Frankfurt. felix.freudenberg@hof.uni-frankfurt.de. Tel: Goethe University Frankfurt. imbierowicz@finance.uni-frankfurt.de. Tel: * Stern School of Business, New York University. asaunder@stern.nyu.edu. Tel: European School of Management and Technology (ESMT). steffen@esmt.org. Tel:

2 Covenants are restrictions in credit agreements that dictate, to varying degrees, how borrowers can operate, financially and otherwise. LSTA Handbook of Loan Syndication and Trading 1. Introduction The previous literature on the use of debt covenants has focused on giving shareholders incentives to monitor the actions of the management (Jensen and Meckling (1976) and Smith and Warner (1979)) rather than creditors which is consistent with much of the corporate governance literature that classifies creditors as passive investors (Townsend (1979), Gale and Hellwig (1985) and Hart and Moore (1998)). More recent research, however, recognizes the role of creditors in monitoring the management by renegotiating loan contracts before firms are in default (Shleifer and Vishni (1997)). These papers analyze the implications of covenant violations on corporate financial policy (Roberts and Sufi (2009a)) and investment (Chava and Roberts (2008)), the role of covenants and other contingencies to accelerate renegotiations (Roberts and Sufi (2009b)), to limit risk shifting (Nini et al. (2009)) and as governance mechanism (Nini et al. (2011)). This paper adds to this literature examining the impact of covenant violations on a borrower`s probability of default, covenant packages and contract design of subsequent loans to the same firm. Do covenant violations predict default rates, thus providing a signal of deteriorating credit quality of borrowers to lenders? How do lenders react? Do they become more active monitors? Do they structure the contracts such that control rights are transferred to lenders in more states of the world? For example, do they make loan contracts stricter? Also, do covenant violations increase the cost of debt for borrowers? We analyze these questions for a sample of U.S. firms who are recurring borrowers in loan markets and focus on the contract design of newly issued loans. Covenants are part of virtually all private credit agreements (Roberts and Sufi (2009a)). At the core of our analysis is a novel and hand collected dataset of loans and covenants which we construct from original loan contracts from the borrowers SEC filings. Our sample comprises 3,813 loans over the 1996 to 2010 period after applying a large number of filters and after matching these contracts to LPC Dealscan and borrowers to the merged CRSP/Compustat database. We collect more than 80 2

3 different covenant types and definitions from these contracts and classify them into 17 groups of financial covenants. We also know step-up and step-down provisions for each covenant. We use all information available to us from these contracts to calculate covenant violations on a quarterly basis. In a first set of tests we ask whether covenant violations increase a borrower`s likelihood to default. We augment our dataset with Chapter 11 filings from the UCLA-LoPucki bankruptcy database to examine this question using three different approaches. First, we investigate the impact of a covenant violation in the previous loan contract on the probability to default on the subsequent loan. Second, we construct an indicator variable equal to 1 if the firm violates a covenant between 1,080 to 180 days before the default date. The contracts with covenant violations from 180 days prior to default until the default date are excluded to be conservative in our analysis. The reason is that covenants are violated almost inevitably the closer the default and may not serve as an early warning signal any more but pick up only a secondary effect of the approaching default. Third, instead of focusing on the last contract violation before default, we use the entire universe of loan contracts of our sample borrowers. This allows us to assess whether violations in the past still have some predictive power. Overall, we find that covenant violations significantly increase the likelihood of future default. However, the likelihood of default decreases substantially the more time has passed after the last covenant violation. The conditional probability of default (PD) is 30% on the first day after the covenant violation and it continues to be substantially higher compared to the unconditional PD for the first 100 days after covenant violation. Taken together, our results suggest that there is an important role for covenants in monitoring borrowers and that covenant violations provide an early warning signal for a severe deterioration of borrower credit quality. We then ask, how do lenders respond to this signal? Do they design subsequent loan contracts that reflect the higher default risk of these borrowers? If they extend further loans, do they demand compensation for this risk increasing the cost of debt for these firms? Do they design the covenants such that they become more active monitors of the firm? We find that borrowers who violated covenants pay, on average, 18 bps higher loan spreads compared to borrowers that have not violated covenants in their previous contract. Lenders also increase the number of covenants in subsequent loan contracts. In addition, they set average covenant 3

4 thresholds stricter. Interestingly, we find that negative effects attenuate if borrowers repeatedly borrow. Using Bradley and Roberts (2004) covenant index, we also find that contracts become stricter after violations. These results are consistent with a higher degree of active monitoring following covenant violations. Lenders increase the number of contingencies and situations in which they gain control rights over the firm reacting to the higher default risk of the borrower. We follow Christensen and Nikolaev (2011) and classify covenants into capital-based and profitability-based covenants and assess their changes as a function of previous covenant violations. Capital-based covenants are used to align incentives of shareholders and debtholders such that shareholders have sufficient incentives to monitor. Profitability-based covenants, however, are used to ex-post allocate control rights over the firm. We find that lenders increase the number of profitability-based covenants relative to capital-based covenants. How is more active monitoring reflected in subsequent loan-lending relationships? Are more renegotiations more likely? That is, are borrowers more likely to violate covenants again after having violated financial covenants in the previous loan contract? We find that borrowers who violated covenants in the previous contract are about 30% more likely to violate again in the current loan, and they also violate sooner than other borrowers. Larger firms are less likely to violate while highly leveraged firms are more likely to violate covenants. How do borrowers react following covenant violations? If they anticipate more actively involved debt holders and higher cost of debt they might be willing to switch lenders if they expect better lending terms. Starting in 1987, we have a complete history of borrower and lender pairs in this market, and we use this information in the last section of the paper to classify borrowers into whether or not they switch lenders after violations. Roberts and Sufi (2009a) find that borrowers hardly switch lenders after covenant violations. We study the effect of switching via matched pairs of equal borrowers only differing whether they have violated a covenant in the past loan contract. Our results show that borrowers switch to new lenders more often after a covenant violation. Overall, we find that lenders substantially increase the number and intensity of covenants after violations which is consistent with the interpretation that covenant violations are used as an early warning signal for deteriorating borrower creditworthiness and that lenders increase monitoring efforts making loan contracts stricter. 4

5 Our paper adds to the literature studying the effects of covenants and covenant violations. Nini et al. (2009) find a decline in acquisitions and capital expenditures after a violation. Furthermore, borrowers decrease their leverage and CEO turnover increases. Chava and Roberts (2008) discover a sharp decline in investment spending, which is particularly pronounced if information problems are more severe. Roberts and Sufi (2009a) and Demiroglu and James (2010) show that tighter covenants decrease investment spending and net debt issuances. The importance of covenant violations is further emphasized by Dyreng (2009). He highlights that borrowers engage in earnings management in order to prevent covenant violations. None of these papers however, studies the role of covenant violations as early warning indicator of default and monitoring instrument. Moreover, we study violations not only in the cross-section of borrowers but also in the time series using a history of hand-collected loans for borrowers who recurrently borrow in the loan market. The paper proceeds as follows. The next section describes how we construct the dataset and provides some descriptive statistics. Section 3 shows the results relating contract violations to default rates. Section 4 analyzes active monitoring by debtholders following covenant violations. Section 5 concludes. 2. Data and Descriptive Statistics 2.1 Data To investigate the effect of covenant violations on borrower default probability and subsequent loans, we construct a new data set collecting original loan contracts directly from the Security and Exchange Commission (SEC) filings of public firms using EDGAR (Electronic Data-Gathering, Analysis and Retrieval). Material loan contracts have to be reported as required by the SEC and can be found as an exhibit to a 10-K, 10-Q or 8-K report. We start with the set of private credit agreements provided by Greg Nini, David Smith and Amir Sufi who collected these contracts for the 1996 to 2005 period and extend this set of contracts for 5 more years until the 5

6 end of Their sample includes 3,720 contracts amended in our data set with 1,276 additional loans. We start with the universe of 4,996 private credit agreements for the 1996 to 2010 period available from Dealscan. We exclude all observations where we cannot identify a contract in EDGAR as well as loans specified as amendments in Dealscan or in the loan contract following Roberts (2010). In other words, all contracts are new loans. 96% of these loans can be matched to the contracts from EDGAR. Although Dealscan already provides some information on negative and financial covenants which is also more complete particularly starting in 2000, we still find that several covenants are missing from our contracts. Furthermore, the definition of seemingly similar covenants differs substantially between contracts and is aggregated in Dealscan without further information. Additionally, only one threshold for financial covenants is recorded in the database, but thresholds frequently change. 2 We therefore manually build a novel set of covenants in private credit agreements collecting all covenants from 3,813 contracts. We do not use any text-search program to avoid possible misspecification of the algorithm. Private loan agreements typically include both negative and financial covenants. 3 Negative covenants prevent the borrower from certain actions such as excessive investments, distribution of too high dividends, sale of assets, changes in company control, entering sale-andlease-back transactions, or a change in business activities. Financial covenants are often termed performance hurdles or trip wires (Dichev and Skinner (2002)) due to their ability to shift control rights. Examples of financial covenants are accounting-based amounts and ratios which can be found in the reporting data of the company (e.g., Taylor and Sansone (2007), Nini et al. (2009)). There also exist maintenance and incurrence covenants. The former imply that the borrower has to meet certain criteria on a regular basis where the latter refer to a predetermined event, such as the issuance of new debt or the acquisition of another company. We record the covenant threshold for each quarter from origination to final maturity of each loan because it changes several times in many loans. These step-down or step-up provisions cannot be found in Dealscan. Furthermore, we find about 80 different definitions of covenants 1 Nini, Smith and Sufi (2009) and Roberts and Sufi (2009). 2 Appendix I shows an example of a financial covenant section in a loan contract. 3 We do not include affirmative covenants such as punctual payment of interest and principal, delivery of financial statements, property and equipment maintenance, compliance to accounting standards, or paying insurance and taxes, as these are often not observable following, for example, Bradley and Roberts (2004) and Demiroglu and James (2010). 6

7 and classify them into 17 main covenant types. 4 They are described in Table I. However, we use the definitions of all 80 covenants to identify covenant violations using the corresponding information from the company s financial statements. To construct our data set, we merge the contracts from EDGAR with several other data sources. We obtain loan contract information from Dealscan including loan spread (AISD), maturity, loan amounts and lender identity. To identify repeated borrowing from the same lender as well as switching between lenders, we construct the merger history for each lender in Dealscan using information obtained from the FDIC and the National Information Center (NIC). Using Robert s Dealscan-Compustat Linking Database (Chava and Roberts (2008)), we collect quarterly financial statement information from Compustat and merge it to each loan contract. We exclude all loans for which this information is not available. Finally, we obtain borrower default information via the Chapter 11 filings in the UCLA-LoPucki bankruptcy research database. All variables are described in Table I. The final dataset includes 3,813 loans with 5,411 loan facilities. [Table I] 2.2 Covenants and Covenant Violations Number of Covenants Using the covenants collected from the SEC filings, we construct several proxies as to the strictness of contracts and covenant violations. We define the Number of Financial Covenants simply as the number of financial covenants in each loan contract. These are all part of the 17 main covenant types shown in Table I. A contract with more covenants is more restrictive compared to a contract with fewer covenants Covenant Looseness We propose a new measure for the strictness of covenants which, because of the way it has been derived, we call Looseness of covenants. Most prior studies only concentrate on 4 The substantially larger number of covenants can be explained by the variety of definitions of the respective variables. Consider for example a debt to capitalization covenant. Debt can be senior, long term or defined as the total value. Capitalization can refer to net worth plus equity or tangible net worth plus equity. 7

8 specific covenant types when they talk about covenant strictness (e.g., Dichev and Skinner (2002), Chava and Roberts (2008), Drucker and Puri (2009), Gow (2009), Zhang (2011)). Our measure is closest to that proposed in Murfin (2011) but we focus only on the average covenant strictness, irrespective of the number of covenants. We first calculate the standard deviation for each of the accounting variables that are part of our 80 different covenants using the previous 12 quarters prior to the loan origination date. We then derive the slack for each covenant which is the (absolute) difference between the observed accounting value/ratio and the covenant threshold that is specified in the loan contract. Each slack is normalized by its respective standard deviation. The value thus reflects the number of standard deviations an accounting value/ratio may deteriorate before the covenant threshold is violated. To derive the average looseness of all covenants in a loan contract we use two aggregation levels. First, calculate the mean ratio of all ratios within each of our 17 main covenant types. We then average again across all covenant types to calculate our measure of looseness. A contract with looser covenants is less restrictive compared to a contract with tighter covenants. As an example for our covenant looseness variable consider the covenants in Gray Communications Systems loan contract on July 31 st, 1998 shown in Appendix II. It contains 6 covenants in 5 main covenant types: i. an adjusted debt service coverage ratio of 1.1, ii. a senior debt to adjusted EBITDA ratio of 4.25, iii. an adjusted fixed charge coverage ratio of 1, iv. an adjusted interest coverage ratio of 1.5, v. a debt to adjusted EBITDA ratio of 6.9, and vi. an adjusted debt to adjusted EBITDA ratio of 6.75 where the two latter both belong to the Debt to EBITDA covenant main type. 5 We then derive the slack and divide it by the variable s standard deviation. The accounting value for the adjusted debt service coverage ratio which is the cash flow to interest and principal payment on July 31st, 1998 is Subtracting the covenant threshold of 1.1 and dividing the result by a standard deviation of gives a value of It reflects the fact that the Cash Flow to Interest and Principal Payment ratio may decrease times its standard 5 Adjusted refers to any definition different from the variable on a stand-alone basis. 8

9 deviation of , i.e. 0.52, before the covenant is violated. The calculation for the other covenants follows the same notion. These values are then aggregated to their main financial covenant type implying that only the values for debt to adjusted EBITDA of and adjusted debt to adjusted EBITDA ratio of are averaged to where the remaining covenants all belong to different main types. Finally, the total covenant looseness of is derived as the mean of all main financial covenant types. The reason that we use two aggregation levels is that covenants within the same main covenant type otherwise could receive too much weight. Using only one aggregation level that is weighting all individual covenants equally the covenant looseness for Gray Communications Systems would change to although debt to adjusted EBITDA and adjusted debt to adjusted EBITDA differ only marginally Contract Intensity Index Finally, our Contract Intensity Index reflects the overall restrictiveness of the loan on the actions of the borrower s management following Bradley and Roberts (2004). It includes not only financial but also negative covenants. The index ranges from zero to six with high values indicating intense contracts. It is constructed summing the indicator variables for dividend restriction, equity sweep, asset sweep, debt sweep, securitization, and a binary variable that is one if the contract includes two or more financial covenants Covenant Violation and Days to Covenant Violation The borrower has to comply with most financial covenants on a quarterly basis (Roberts and Sufi (2009a)). A covenant is violated if the corresponding accounting value is above (a max. threshold type in Appendix II) or below (a min. threshold type in Appendix II) its respective threshold. Days to Covenant Violation is measured as the difference between inception of the contract until the end of the quarter during which a financial covenant is violated for the first time. 6 Note that as a robustness check, we also used in all analyses only the least loose covenant in each loan contract as a measure for covenant looseness, i.e for Gray Communications Systems. The results are similar with a statistically stronger outcome in some cases. We do not report them for reasons of brevity. 9

10 2.3 Descriptive Statistics The final data set consists of 5,411 facilities (3,813 loans) over the 1996 to 2010 period. Table II provides detailed summary statistics on loan and borrower characteristics. All data are measured in real terms with 2000 as the base year. [Table II] The average borrower default rate is 2.5%, the average loan facility is $298 million with an All-In-Spread-Drawn (AISD) of 183 basis points (bps) and 2.55 financial covenants. The covenant looseness is 3.95, that is, covenants can change, on average, by 3.95 standard deviations before a contract violation occurs. The contract intensity index is derived following Bradley and Roberts (2004) and loans contain, on average, 4.55 out of 6 possible restrictions. Note that we need to rely on data reported by Dealscan when collecting the restrictions. As we require all 6 indicator variables to be observable but information about asset sweep, equity sweep or secured status is missing in many cases it reduces the number of observations by more than 70%. 55% of the loans are violated and these violations occur on average 14 months (427 days) after the loan origination date. Borrowers switch banks in 35.1% of all cases and violate a financial covenant in more than half (57.2%) of all of their previous loan contracts. The average borrower size is $3,291 million with a profitability of 17%, a current ratio of 1.84, a leverage ratio of 0.33, an interest coverage ratio of 15.44, and a market-to-book ratio of More than one half of the loans are rated and 24.1% are classified as investment and 34.5% as non-investment grade. Panel C of Table II shows the distribution of covenants across all risk classes. More than 60% of all loan contracts contain Debt/EBITDA covenants followed by Interest Coverage (44%) and Fixed Charge Coverage (42%). 3. Covenant Violations and the Likelihood to Default To examine the implications of covenant violations we first ask whether and how covenant violations predict default rates. Moreover, how do default rates evolve over time conditional on borrowers having (or not having) violated covenants? 10

11 We augment our loan data with Chapter 11 filings obtained from the UCLA-LoPucki bankruptcy research database which ultimately collects its information from court files or SEC filings. Most importantly, we derive the exact default date of each borrower from LoPucki. To assess the incident of covenant violations as a signal of deteriorating borrower quality for lenders, we need a starting point from which we estimate a borrower`s probability of default (PD). If a covenant is violated we take the violation date. In case no covenant violation occurs we choose the contract end date as the signal for repaying the loan without violating a covenant. We first explore this graphically. Figure 1 plots the default probability for borrowers after no covenant violation (Panel A) and after a covenant was violated (Panel B). [Figure 1] Panel A in Figure 1 plots the development of a borrower`s PD over time after no covenant has been violated. Directly after the full repayment of the loan borrowers have a PD of almost zero. In line with Flannery (1994), the PD increases (for levered firms) the longer the time period. Panel B plots the development of a borrower`s PD over time after a covenant violation occurred. Borrowers have, on average, a PD of 30.26% at the first day after a covenant violation (not shown for scaling purposes). Figure 1 shows that borrower`s PD is substantially higher in the period directly following the violation and decreases as a convex function over time. Borrowers exhibit a substantially higher likelihood to default especially in the first 100 days after a violation. Figure 1 implies that it takes more than two years (882 days) until the PD decays to a level comparable with the PD of borrowers with no violations. This is consistent with covenant violations being a signal of elevated borrower risk. To investigate the impact of covenant violation on borrower PD in a regression model we use three different approaches. First, we investigate the impact of a covenant violation in the previous loan contract on the probability to default on the subsequent loan. Second, we construct an indicator variable equal to 1 if the firm violates a covenant between 1,080 to 180 days before the default date. The contracts with covenant violations from 180 days prior to default until the default date are excluded to be conservative in our analysis. The reason is that covenants are violated almost inevitably the closer the default and may not serve as an early warning signal any 11

12 more but pick up only a secondary effect of the approaching default. Third, instead of focusing on the last contract violation before default, we use the entire universe of loan contracts of our sample borrowers. For each borrower, all prior loan contracts are matched to the most recent contract. We construct an indicator variable Violation in Past Contract which is 1 if there is a violation in any of the prior contracts. It is reasonable to assume that the quality of violations as early warning signal should be highest the closer the violation occurs to the default date, however, violations in the past might have some predictive power. We specifically control for the time aspect by interacting the variable Violation in Past Contract with a new variable Ln(Days since Past Contract (Violation)). It equals the natural logarithm of the days between the past covenant violation and the current loan if there is a violation, and the days between the end of the past contract and the current loan if there is no violation in the past loan. The results are reported in Table III. In all tests, we control for loan characteristics including maturity (months), a dummy variable when the loan is secured, and the natural logarithm of the facility size. We also include borrower characteristics such as profitability, current ratio, leverage, coverage, market to book ratio, total assets, and borrower credit rating. Furthermore, we account for time (calendar year), industry (one-digit SIC industry classification), loan type, and loan purpose fixed effects as well as clustering of error terms at the firm level and heteroscedasticity. [Table III] Column (1) of Table III reports the first model. The coefficient of Previous Covenant Violation is positive and significant at the 5% level, that is, covenant violations in the previous loan contract increase the likelihood to default. The second model is reported in Column (2) of Table III. Covenant violations 1,080 to 180 days before the default date also substantially increase the probability of default. As mentioned above we exclude the last half year prior to default to remain conservative. 7 Columns (3) and (4) report the results from the third model. The results in Column (3) suggest that also prior contracts contain information about a borrower`s default probability. The most recent performance and, also, covenant violation should be the best signal for the lender, that is, we should observe in the data that the predictive power of covenant violations for default rates decays over time. To examine this, we interact Ln(Days since Past Contract (Violation)) with the covenant violation variable. We find a negative coefficient of the 7 Note that the coefficient increases to 3.319, significant at the 1% level, if this time period is included. 12

13 interaction term which is significant at the 1% level. In other words, having violated financial covenants in the past increases a borrower s default probability substantially, however, the default likelihood is significantly lower the more time has passed since the covenant violation date. All regression specifications indicate that covenant violations serve as an early warning signal for an increased borrower default probability. To further examine whether covenant violations indicate higher borrower risk we conduct two additional analyses. First, we determine in each year the loan contracts with and without a covenant violation. That is we construct annual cohorts of loan contracts which are (in each cohort) split by covenant violation. We then compare for each cohort the default probability between borrowers who violated a covenant and for those who did not over the following years. Note that this facilitates comparing default rates in different economic conditions. And second, we match borrowers who violate a financial covenant to non-violating borrowers with the same rating class to control for differences in credit quality when the information about a covenant violation becomes public. That is, we use the rating of violators at the time of covenant violation and the rating of non-violators at the time when the contract is repaid without having violated a covenant over the maturity and investigate each group s default probability in the following years. The results are shown in Table IV. [Table IV] Panel A in Table IV reports the results by loan cohort year. As an example consider the year We have 1,765 ongoing loans in our data set of which 1,520 are violated and 245 do not show a covenant violation. Comparing these two groups shows that in the following year, the year 2001, the difference in default probability is -1.6%. This implies that borrowers who violated a financial covenant in 2000 have a higher (historical) default probability of 1.6% one year later compared with borrowers who did not violate a covenant in 2000, significant at the 5% level. For the year 2002, we observe that the difference is still 1%, significant at the 1% level, although firms which defaulted in 2001 already dropped out of the sample. For the following years we do not observe a statistically significant difference any more. This is in line with general expectations as during 2003 to 2005 borrower default rates have been low due to positive 13

14 economic conditions. Accordingly, Panel A in Table IV shows that independent of the time period when the covenant was violated, borrower default rates in the following years are in general higher for violating firms compared to firms without a covenant violation in the same year. Panel B in Table IV compares default rates of borrowers with the same credit rating at the time they violated (no) financial covenants. It is consistent with our finding of covenant violations indicating higher borrower risk. Independent of the credit quality of borrowers at the time when the covenant violation becomes public information, borrower default probability is on average higher when a covenant is violated. That is borrowers with a rating of BBB have a higher (historical) default probability in the following year if they have violated a financial covenant compared with BBB-rated borrowers which repaid their loan in adherence to contract terms. This pattern is reflected for all rating classes. Accordingly, covenant violations indicate a higher default probability in the subsequent periods. Overall, our results are consistent with the interpretation that covenant violations provide an early warning signal for elevated borrower risk. How do lenders respond to this signal? Do they design subsequent loan contracts that reflect the higher default risk of these borrowers? If they extend further loans, do they demand compensation for this risk increasing the cost of debt for these firms? Do they design the covenants such that they become more active monitors of the firm? We analyze these questions in the next section. 4. Covenant Violations and Monitoring 4.1 Univariate Results To get a better understanding of how spreads and covenant restrictions evolve conditional on covenant violation, we graphically explore average loan spreads, number of covenants in loan contracts, covenant looseness, and contract strictness as a function of repeated borrowing in the loan market. Figure 2 shows how the four measures develop for a borrower who obtains loans for the first until the fourth time. The number of observations is provided in parentheses. [Figure 2] 14

15 A first time borrower pays on average 178 bps above LIBOR which decreases to 108 bps in the fourth contract conditional on never having violated a covenant. Borrowers also become less opaque as they frequently borrow in the loan market. At the same time, always violating a covenant in the first three loans increases loan spreads to 238 bps in the fourth loan. A violation in the first loan increases loan spreads to 228 bps, not violating a covenant in the second loan reduces spreads to 148 bps in the third loan. Interestingly, a borrower who violates a covenant in the second but not the first loan pays, on average, 174 bps if she borrows for the third time. In other words, the effect of violations on spreads dissipates over time which is consistent with the decreasing effect on default risk in Table III. Overall, Figure 2 shows that covenant violations result in higher spreads for borrowers in subsequent loans. A similar pattern is observable with respect to the number of financial covenants in new loan contracts. Panel B shows how the average number of the financial covenants develops over time as borrowers frequently return to the loan market. Not violating covenants reduces the number of covenants in subsequent loans, whereas the number of covenants increases following a covenant violation in the previous contract. Again, we observe that the effect of covenant violations is more pronounced in the contract following the violation. Panel C and D show the results for covenant looseness and contract intensity. The figures depict comparable patterns again. Covenant violations result in stricter covenant thresholds and contracts. Not violating a financial covenant causes thresholds and contracts to be less strict. Note that decreasing borrower opaqueness due to repeated borrowing seems to cause covenant thresholds to adjust to comparable levels in the fourth loan obliterating to some extent the effect of violating a covenant. In Table V, we segregate the entire sample based on whether or not the borrower violated a covenant in the previous contract. [Table V] Columns (A) and (B) of Table V show mean and median characteristics for borrowers who did not violate a covenant in the previous contract and for those who did. The last column reports the parametric t-statistic (nonparametric z-statistic) of the difference in means (medians) 15

16 test. Table V shows that the differences between both groups are substantial. In line with the results in the previous section, borrowers violating a covenant in the previous loan contract have a much higher default probability. On average, they have to pay a higher spread of 98 bps, accept 0.6 more financial covenants which are in addition significantly more restrictive. Loan contracts also become more intense after a covenant violation. On the other hand, borrowers who violate a covenant in the previous loan contract again violate contract terms in 70% of all loans where the violation occurs within a shorter time period from the contract start date. Table V also shows that borrowers who violated covenants are more likely to switch lenders in the next loan. We furthermore find that loan amounts decrease after a covenant violation. Moreover, the percentage of secured loans almost doubles. Borrowers who violate covenants are also smaller, higher leveraged with lower interest coverage and market-to-book ratio and they are also lower rated. Overall, the univariate results are consistent with the interpretation that covenant violations provide an early warning signal for deteriorating borrower performance and that covenants are used by banks to monitor borrowers. 4.2 Repeated Borrowing Table VI tabulates various contract characteristics (spread, covenant measures, violations and defaults) grouped by the number of loans of the same borrower. [Table VI] Panel A of Table VI focuses on loans of borrowers from the same lender. That is, we observe how contract characteristics develop over a bank-borrower relationship. Loan spreads are decreasing which is consistent with the literature (e.g., Bharath et al. (2011)). We also observe a lower number of covenants, looser covenants and lower default rates as the relationship evolves consistent with the monitoring role of relationship lenders. Interestingly, the contract intensity is increasing. One possible interpretation is that an increase in contract intensity associated with covenant violations (note that both violations and non-violations are included) is dominating possible positive relationship effects. 16

17 Panel B of Table IV shows the results also grouped by the number of loans from the same borrower but not conditional on borrowing from the same lender. In other words, we observe both relationship borrowers as well as firms that switch lenders. Interestingly, we do not find lower loan spreads for borrowers who borrow repeatedly in the loan market. We observe fewer and looser covenants and a lower likelihood to violate covenants. However, contract intensity is increasing as well as the likelihood to default. Overall, we find evidence consistent with relationship benefits with regard to lower loan spreads. Repeated borrowing alone does not reduce loan spreads. The result of fewer and looser covenants over successive contracts does not depend on the relationship. 4.3 Covenant Violations and the Design of Subsequent Loan Contracts To analyze the effect of covenant violations on the design of subsequent loan contracts, we use a regression framework of the following form: LCT = a + b * Previous Covenant Violation + c * Loan Characteristics + d * Borrower Characteristics + e * Other Controls + ε LCT ( Loan Contract Terms ) refers to two different groups of dependent variables. The first group contains a proxy to measure the cost of debt (AISD), the second group describes the new covenant package as measured by i) the Number of Covenants, ii) Covenant Looseness, and iii) Contract Intensity. The results are reported in Table VII. [Table VII] Table VII reports four models with different dependent variables. It also shows the regression methodology used in each regression. In addition to the reported variables, all regressions further include year and industry fixed effects, indicator variables for the different rating classes of borrowers considering also unrated firms, loan type and loan purpose fixed effects. Column 1 of Table VII reports the results of an OLS regression relating AISD to previous covenant violations and our other control variables. We find that previous covenant violations 17

18 increase loan spreads in subsequent loans, on average, by 18 bps which is significant at the 1% level and economically meaningful. It translates into $0.54 million higher annual loan costs for a borrower who violated covenants in the previous contract. Most of the other control variables are also highly significant and carry the expected signs. For example, larger loans and loans containing a performance pricing grid have lower spreads, secured loans and highly leveraged loans carry larger spreads. All variables are defined in Table I. Standard errors are heteroscedasticity robust, clustered at the borrower level. We then explore the implications of covenant violations on the covenant package of new loans. If lenders become more active monitors, we expect to find that the covenant package overall becomes more restrictive. Columns 2 to 4 of Table VII report the results. First, we relate the number of financial covenants to previous covenant violations and our control variables using ordered logit regressions. 8 We find that previous violations increase the number of financial covenants used by lenders in the subsequent loan contract. The coefficient is significant at the 1% level. The OLS regression results for covenant looseness are shown in Column 3. A covenant violation in the previous contract leads to stricter covenants with thresholds set 1.5 standard deviations closer to the actual accounting value in subsequent contracts. Column 4 in Table V reports the impact of covenant violations in the previous loan on contract intensity using an ordered logit regression. It shows that contract intensity is increasing, the coefficient, however, is only weakly significant. Note that the number of observations also drops to 848 in Column 4 because we need to rely on data recorded in Dealscan in order to calculate the index in a similar fashion as in Bradley and Roberts (2004), Demiroglu and James (2010) and Bharath et al. (2011). Information about asset sweep, equity sweep or the secured status of the loan is missing in many cases. Overall, the results are consistent with a higher degree of active monitoring following covenant violations. Lenders increase the number of contingencies and situations in which they gain control rights over the firm in reaction to the higher default risk of the borrowers both by increasing the number of financial covenants and by setting the thresholds tighter to the actual accounting values. They further demand a higher compensation for the increase in default risk. 8 We use ordered logit regressions because the number of covenants is an ordinal measure in our context. In robustness tests, we also use OLS models as well as Poisson models and get similar results. We do not report these tests for brevity. 18

19 4.4 Profitability-Based versus Capital-Based Covenants The previous section shows that, on average, the number of financial covenants is larger if borrowers have violated covenants in their previous loan contract. Christensen and Nikolaev (2011) argue that covenants can be broadly classified into two groups, profitability-based covenants and capital-based covenants. Profitability-based covenants include: (1) Cash interest coverage ratio; (2) Debt service coverage ratio; (3) Level of EBITDA; (4) Fixed charge coverage ratio; (5) Interest coverage ratio; (6) Debt to EBITDA; and (7) Senior debt to EBITDA. Capital-based covenants include: (1) Quick ratio; (2) Current ratio; (3) Debt-toequity ratio; (4) Loan-to-value ratio; (5) Debt-to-tangible net worth ratio; (6) Leverage ratio; (7) Senior leverage ratio; and (8) Net Worth requirement. Lenders use capital-based covenants essentially to restrict leverage in order to ensure that firms maintain a specific level of equity capital within the firm. Following the arguments in Smith and Warner (1979), tying up sufficient shareholder wealth inside the firm aligns incentives of shareholders and debtholders and reduces the need for debtholders to actively monitor the management. Profitability-based covenants, on the other hand, are used as tripwires to ex-post allocate control rights to the lender. Since they are based on the current performance of the borrower (that is why they are also called performance covenants), lenders can immediately respond to a deteriorating performance of the firm. We thus hypothesize that lenders will increase the number of profitability-based covenants following covenant violations to increase the states of the world in which they gain control rights over the firm. In other words, we expect to find a positive correlation between previous covenant violations and the number of profitability based covenants if lenders want to be more actively involved in the monitoring of the firm. We test this hypothesis formally and reports the results in Table VIII. [Table VIII] 19

20 Column (1) [Column (2)] of Table VIII reports the results using the Number of Profitability Covenants [Number of Capital Covenants] as dependent variable. We use ordered logit models for estimation purposes and include the same control variables we also use in Table VII. Consistent with our hypothesis we find that the coefficient of the previous violation indicator is positive and significant at the 1% level but does only explain the elevated use of profitability based covenants in loan contracts of borrowers who have violated covenants relative to those who have not. There is no difference with respect to capital-based covenants which is consistent with more monitoring from debtholders. The switch in signs between the coefficients from Column (1) to Column (2) is important and is suggestive of the differential use of both types of covenants and consistent with the correlations reported in Christensen and Nikolaev (2011). For example, longer maturity loans have more profitability-based and fewer capital-based covenants relative to short-term loans which is plausible since lenders cannot get access to their investment unless borrowers violate 1 or more of the covenants. Monitoring is thus more important if loans have longer maturities. This argument extends to more levered firms as well as firms with higher market to book ratios. Loan contracts that include performance pricing provisions, on the other hand, include a larger number of capital-based covenants and fewer profitability-based covenants. As the consequences of contingencies specified in the performance pricing provision are contracted upon ex-ante, lenders have already relinquished control in these states to the borrower. In other words, lenders try to align their incentives with those of shareholders by including capital-based covenants, giving the latter higher incentives to monitor because of their higher exposure to the firm. 4.5 Covenant Violations, Loan Contract Terms, and Information Asymmetry If lenders increase their monitoring activity after firms have violated covenants included in their loan contracts, this effect might even be amplified if information asymmetries between lenders and borrowers are more pronounced. We test this hypothesis using different measures of information asymmetry. There are: (i) Small which is a dummy variable equal to 1 if a borrower s asset size falls into the 25 th percentile of the distribution of our sample firms, and (ii) Young which is a dummy variable equal to 1 if borrowers are less than 3 years listed on a stock exchange. The regression model for this analysis is of the following form. 20

21 LCT = a + b * Previous Covenant Violation + c * Borrower Opacity + d * Previous Covenant Violation * Borrower Opacity + e * Loan Characteristics + f * Borrower Characteristics + g * OtherControls + ε LCT ( Loan Contract Terms ) again refers to different dependent variables which are: AISD, Number of Covenants, Covenant Looseness, and Contract Intensity. The results are reported in Table IX. [Table IX] Column (2) shows that opaque (that is young) firms pay larger spreads consistent with prior literature (e.g., Bharath et al. (2011), Saunders and Steffen (2011)). However, Column (1) reports that opaque firms have higher borrowing costs after a covenant violation due to their opacity. Overall, our results for the AISD suggest larger spreads for opaque firms. Our opacity proxies do not show a differential effect for opaque relative to transparent firms with regard to covenant number and contract intensity. Column (6) indicates looser covenants for opaque (that is young) borrowers. This is almost compensated if a covenant was violated in the previous loan. However, the results with regard to covenant looseness are not confirmed for both opacity variables. Note that Table IX supports and even amplifies our previous results from Table VII. The coefficient for our variable indicating a covenant violation in the previous loan contract increases in level and significance for contract intensity while it shows the same economic and statistical significances as in Table VII for the remaining loan contract term variables. Taken together, our tests show that covenant violations in the previous loan contract have substantial effects on the design of subsequent loan contracts. Violations increase borrowing costs, the number of financial covenants, the strictness of covenants, and also covenant intensity. In line with the literature, borrowing costs are higher for opaque borrowers. Overall we observe that covenants are used by banks to monitor borrowers irrespective of the level of information asymmetry. 21

22 4.6 The Propensity to Violate Covenants In the previous section we find that lenders design covenant packages of subsequent loans to become more actively involved monitors when covenants have been violated. How is this reflected in subsequent loan-lending relationships? Are more renegotiations more likely? That is, are borrowers more likely to violate covenants again after having violated financial covenants in the previous loan contract? To answer these questions we examine the likelihood of covenant violations in the loan contract following a violation using the following logit regression model. Pr(VIOL) = a + b * PreviousCovenantViolation + c * LoanCharacteristics + d * BorrowerCharacteristics + e * OtherControls + ε Pr(VIOL) is an indicator that is 1 if the borrower violates at least 1 covenant in the new loan. The results are reported in Table X. [Table X] Column 1 in Table X shows a coefficient for previous covenant violation of significant at the 1% level. This implies that borrowers who have violated covenants in the previous contract are 30% more likely to violate covenants in the subsequent loan. The control variables are as expected, larger firms with a higher market to book ratio are less likely to violate, whereas firms with a higher leverage have a higher probability of covenant violation. Conditional on violating covenants in the 2 nd, 3 rd, x th loan, when are covenants violated? More precisely, we focus on the time period between contract initiation date and covenant violation (DAYS) in an OLS model with the following specification. DAYS = a + b * Previous Covenant Violation + e * LoanCharacteristics + f * BorrowerCharacteristics + g * OtherControls + ε 22