MSc Business Economics, Finance track
The Empty Creditor Problem: Does Credit Default trading influence the use of
bondholder protective covenants in debt contracts?
ABSTRACT
I use covenant data on bond issues between 2004 to March 2015 to examine the effect of credit default swap trading on the number of bondholder protective covenants used in debt contracts. Uninsured creditors might demand more protection through the use of covenants, to compensate for the increased credit risk arising from the Empty Creditor problem first described in 2008 by Hu & Black. Using debt contract data on US firms, I find that the average number of bondholder protective covenants is 20% higher when credit default swaps are traded on a firms debt at the moment of issuance. Furthermore, my findings are consistent with the theoretical predictions on credit default swaps made by Bolton & Oehmke (2011) and built further upon the empirical results of Subrahmanyam, Tang, and Wang (2014).
Master Thesis by Hans van der Woude
1
Statement of originality
This thesis is the work of Hans van der Woude. Please do not copy without permission. This is to certify that to the best of my knowledge, the content of this thesis is my own work. I certify that the intellectual content of this thesis is the product of my own work and that all the assistance received in preparing this thesis and sources have been acknowledged.
Acknowledgements
For providing me with access to essential data, providing me quick and clear feedback, I would like to thank my supervisor Dr Tomislav Ladika. Furthermore I would like to thank Milian Bachem, Davy Westhof and Emma Meessen for their useful and critical feedback.
2
Table of Content
Introduction ... 3
Theoretical Background ... 5
A. Credit Default Swaps ... 5
B. Covenants ... 6
Current literature ... 7
A. Previous studies on credit default swaps ... 7
B. Previous studies on covenant design ... 9
C. Channel mechanism ... 11
Data & Methodology ... 13
A. Collecting and preparing data ... 13
B. Summary of dataset ... 14
E. Methodology ... 17
F. Regression by Industry ... 19
Empirical Evidence ... 20
A. Baseline regression ... 20
B. Baseline regression on matched sample ... 23
C. Fixed effects regression ... 26
D. Robustness check ... 29
E. Empty Creditors and debt contract design ... 31
F. Discussion of the results ... 33
Conclusion ... 36
References ... 38
3
Introduction
The relatively new market of Credit Default Swap (CDS) trading and the role the CDS market played in the financial crisis of 2007-2009 have many aspects which are not well understood. Empirical research has so far mainly focused on the default risk (Subrahmanyam, Tang, and Wang (2014), Peristiani and Savino (2011), Bedendo, Cathcart and El-Jahel (2011), Danis (2012)), while the theoretical model of Bolton & Oehmke (2011) shows important ex ante benefits for both debtors and creditors. With a notional amount outstanding of $32 trillion in 2011 according to BIS statistics, a misunderstanding of the consequences could have severe implications on the stability of the financial markets.
While the main purpose of Credit Default Swaps is to shift default risk from one party to another, CDS contracts have also been used as a speculation device, causing a change in the relationship between certain types of creditors. On the one hand there are insured creditors, while on the other hand there are uninsured creditors. The trading of CDS may affect the incentives of the uninsured creditors up to a point where they favour bankruptcy. This phenomenon was first described by Hu and Black (2008) as the ‘empty creditor
problem’. The recent financial crisis was characterized by an unusual growth in CDS trading, which could indicate an increase in the existence of Empty Creditors. In response to this financial crisis and the exponential growth of notional amount of CDS outstanding, $800 billion in 2000 versus $32 trillion in 2009 (Hull, 2006), researchers started focussing on the relative default frequency of firms with CDS traded on their debt. The main hypothesis was that due to empty creditors, the credit risk significantly increases when CDS trading
commences on a firms debt. This hypothesis and the existence of empty creditors is mostly confirmed by empirical research. Subrahmanyam, Tang, and Wang (2014), Peristiani and Savino (2011) and Danis (2012) all find evidence for the existence of empty creditors.
Where previous studies focused on the conflict and the changing relationship between debtors and creditors, little to no research has been done in to the possible conflict arising from CDS trading between different creditors. On the one hand there are insured creditors who can become empty creditors while on the other hand there are uninsured creditors who are exposed to increased default risk. One possible way through which uninsured creditors could protect themselves against the increased default risk would be trough demanding more bondholder protective covenants. This paper will empirically test whether the empty creditor problem, as defined by Hu and Black (2008), gives rise to a conflict between insured and
4
uninsured creditors. As Berlin and Mester (1992) show, the strictness of covenants depends on the default risk and the default risk increases when CDS are traded on a firms debt as shown by Subrahmanyam, Tang, and Wang (2014). This leads to the following hypothesis:
Hypothesis #1: Uninsured creditors demand more bondholder protective covenants when
CDS is traded on a firms debt
I assume there is a certain fraction of the participating agents in the bond market without access to credit default insurance which is large enough to influence the debt contract design. In order to compensate for the increased credit risk, this group of uninsured creditors demands more bondholder protective covenants. To relate this to the Empty Creditor problem however, we need further proof. One way of testing the relation between covenant use and the Empty Creditor problem is relating the use of covenants to the general state of the economy. During stressed economic periods we would expect more corporate distress and naturally more empty creditors using the corporate distress to force bankruptcy. This results in the second
hypothesis:
Hypothesis #2: The number of bondholder protective covenants for CDS firms increases with
the level of Empty Creditors
As a proxy for the level of Empty Creditors I use GDP growth in the year prior to the debt issue. Corporate distress is related to the general state of the economy. During corporate distress we would expect relatively more debt renegotiations and thus more activity from Empty Creditors to force bankruptcy. If uninsured creditors are aware of the existence of Empty Creditors, the uninsured creditors should show a significant reaction to a decline in economic activity by demanding more protection. The interaction between GDP growth and CDS trading should therefore be able to support the second hypothesis and relate the
increased use of covenants to the existence of Empty Creditors.
When using propensity score matching, I find that on average the amount of
bondholder protective covenants in a contract significantly increases with almost 20% when CDS is traded on a firms debt on an average of 4.548 bondholder protective covenants. I conclude that uninsured creditors are at least partially compensated for the increased credit risk through more bondholder protective covenants. I am among the first to empirically test whether there is a conflict between creditors arising from CDS trading. This paper is built on
5
the theoretical predictions of Bolton and Oehmke (2011) and further supports the empirical
results of Subrahmanyam, Tang, and Wang (2014).From here on the thesis is structured as
follows. In section 2 will give a theoretical background about credit default swaps, credit default swap trading and covenants. I outline the current literature in a literature review in section 3. I then describe my data and methodology in section 4. In section 5 I will outline the results and I will provide a discussion of the results. Finally section 6 will conclude the paper.
Theoretical Background
A. Credit Default Swaps
A credit default swap is essentially a contract between a protection seller and a protection buyer. The credit default swap allows the protection buyer to insure against the default risk of a firm. If the firm is not able to repay the outstanding obligation, the firm is in default. When a firm defaults on its debt it is described as a credit event. In the case of a credit event the protection seller is obliged to pay the face value of the debt to the protection buyer. In return the protection buyer hands over the ownership of the obligation. In practice, most of the time there is a simple exchange of money which is the difference between the face value and the recovery value of the obligation (Hull, 2006).
J.P. Morgan is said to be the inventor of the credit default swaps used today. In 1994 J.P. Morgan sold their credit risk on a loan to Exxon on order to hedge their risk and lower the amount of regulatory capital needed for this loan (Shan, Tang, & Yan, 2014). Since then the growth of credit default swaps has been exponential. According to BIS statistics the notional amount of CDS contracts outstanding has grown from $180 billion in 1997 to over $32 trillion in 2011. Unlike ordinary insurance contracts, a buyer of credit default protection is not required to have an economic interest in the insured object. In other words, the buyer of credit default protection can insure against default risk, while not actually being exposed to this default risk (Bolton and Oehmke, 2013). In addition, unlike the general stock market, credit default swaps are not traded through a centralized stock exchange, but rather through several fractured over the counter (OTC) markets. OTC markets are characterized by very little transparency and regulation. These two particular features of credit default swaps came under heavy criticism after the financial crisis of 2007-2009, because it allowed a build-up of systematically financial risk (Bolton and Oehmke, 2013).
6
Recent examples by which CDS became known to the general public is the failure of AIG during the credit crisis and the big trading losses J.P. Morgan had to take in 2012. In both instances banks were documented to use CDS trading mainly for non-hedging purposes (Minton, Stulz & Williamson, 2009). Following these significant events regulators became more concerned about the risk taking behaviour of banks. This resulted in the Dodd-Frank act which came in existence in 2010. Part of the Dodd-Frank act prohibit propriety trading by banks and requires more capital for derivative trading activities. Subsequently derivative activities became more capital intensive, which led some major Wall Street banks to cut back on proprietary trading. In response to significant systematic credit events during the existence of Credit Default Swaps, several models have been developed and empirically tested which will be discussed in the next section.
B. Covenants
Over the past decades, the volume of credit issuances has fluctuated heavily from year to year. In sync with the fluctuations of credit issuance the use of covenants in debt contracts has also fluctuated heavily. The use of covenants seems to be influenced by the market perception of risk. During boom periods of 2002 to 2006 the use of covenants receded significantly, while after 2008 the use of covenants became a popular instrument in debt contract design (Murfin, 2012).
Covenants are viewed as a mean of allocating control rights between the management of a firm and the creditors of the firm (Aghion & Bolton, 1988). The goal of these covenants is to mitigate agency problems. In particular these covenants can be used to stop firms from under investment in projects that increase firm value in difficult market conditions. During difficult market conditions firms are also more likely to experience default, which can induce a firms management to take on high risk projects. Since this phenomenon is known to
creditors, debtors either pay a higher interest rate or agree to covenants (Berlin & Mester, 1992). The use of covenants is generally a trade-off. During prosperous economic times the covenants designed for difficult economic conditions can constrain management in
undertaking positive net present value projects.
Covenants can be subdivided in three categories. The first category is ‘bondholder protective’ that include covenants such as ‘negative pledge’, ‘cross default’ and ‘cross acceleration’ covenants. The second category is ‘issuer restrictive’ that includes covenants such as ‘indebtedness’, ‘sale of assets’ and ‘restricted payments’. The third category is
7
‘subsidiary restrictive’ that includes covenants like ‘dividend related payments’,
‘indebtedness’ and ‘fixed charge coverage’1
. It is proven by Berlin & Mester (1992) that the strictness of the covenants depends on the credit risk of the borrowing firm. It is generally believed that private placements and loans typically have more and stricter covenants then publicly traded bonds. Smith and Warner (1979) argued that this arises from the fact that private placements and loans are more easily renegotiated due to the limited and concentrated group of creditors. The research of this paper will particularly focus on the bondholder
protective covenants, since this type of covenant should be providing uninsured creditors with additional protection against default.
Current literature
A. Previous studies on credit default swaps
The empty creditor problem was first described by Hu and Black in 2008 as a creditor which has hedged its economic risk, but retains its rights as a credit holder and thereby causes a conflict of interest between insured and uninsured creditors. Bolton and Oehmke (2011)
then used this concept in their theoretical model.The main conclusions of Bolton and
Oehmke (2011) state that CDS increases the bargaining power of the creditor and this has both a positive and a negative effect. On the positive side the bargaining power of the creditor serves as a commitment device for the debtor to repay. On the negative side CDS trading may lead to over-insurance. A debtor firm might opt for a strategic default if the expected future cash flows are low enough. This limited commitment of repaying causes two inefficiencies. First, the limited commitment reduces the firm’s ability to borrow even for positive present value projects and second, there exists a level of initial investment where financing is possible only with strategic default. The increased bargaining power of creditors has two effects. The first is that protected creditors are tougher in renegotiation and therefore able to extract a higher sum during renegotiations. The second is that the debtors incentive to strategically default is reduced, because the debtor anticipates on the increased bargaining power. This in turn increases the debtors ability to borrow for positive present value projects which were previously characterized as strategic default projects. On the contrary to these positive effects, Bolton and Oehmke (2011) also show that under certain circumstances creditors will buy
1
8
credit insurance up to a point at which it will prevent renegotiation all together. At that point creditors would prefer a credit event instead of a successful renegotiation. This over insurance is classified as the Empty Creditor problem.
Following the predictions made by the theoretical model of Bolton and Oehmke (2011), several researchers empirically tested the existence of Empty Creditors.
Subrahmanyam, Tang, and Wang (2014) examine the effect of CDS trading on the credit risk of reference firms using data ranging from 1997 until 2009. After controlling for the
endogeneity of CDS trading and other factors affecting credit risk they find supportive evidence for the existence of Empty Creditors. Once a firm is in distress, they find that the likelihood of a firm filing for bankruptcy increases when CDS are traded on the firms debt relative to non-CDS traded firms. In addition, Peristiani and Savino (2011) found that from 2004 to 2008, firms with CDS trading have a higher expected default frequency than comparable firms without CDS traded. Finally, supportive for the existence of Empty Creditors, Danis (2012) finds that CDS traded firms have more difficulty to restructure their debt out of court.
There would be no reason for discussion on this topic however, if all researchers found supportive evidence for the existence of empty creditors. Bedendo, Cathcart and El-Jahel (2011) who focused their research on non-financial US firms over a period ranging from 2007 until 2011, find no evidence of the existence of empty creditors. Similar to the research conducted by Subrahmanyam, Tang, and Wang (2014), they focused on restructuring outcome conditional on financial distress. Contrary to the Empty Creditor hypothesis, no evidence is found for a significant difference in restructuring outcome. Moreover they find that the restructuring choice is driven by the same factors for reference firms and non-reference firms. Although these results indicate there is no significant existence of Empty Creditors, it does not contradict the theoretical model of Bolton and Oehmke (2011). It simply implies that the assumptions made by Bolton and Oehmke (2011) under which creditors over-insure do not hold in the real world.
To further test the effects of CDS trading predicted by the Bolton and Oehmke (2011) model, research has also focused on the increased borrowing ability of reference firms. Supportive for the increased lending ability of reference firms ,Subrahmanyam, Tang, and Wang (2014) find that the leverage of reference firms increases significantly after the
inception of CDS trading. In addition Saretto and Tookes (2013) consider the impact of CDS trading on the capital structure of reference firms. Examining data on capital structures of S&P500 firms from 2002 until 2010 they find supportive evidence for the prediction that CDS
9
trading increases the reference firms ability to borrow. Their results indicate that a reference firms’ leverage increases by between 6% and 22% and their debt maturity increases by between 8% and 21%. It is worth noting however that these effects are mainly non-price effects, Ashcraft and Santos (2009) do not find significant evidence that CDS trading lowers the cost of capital for reference firms. Saretto and Tookes (2013) suggest this could be due to the relatively flat demand curve for credit, which makes an outward shift of the supply curve have a greater impact on quantities than on prices. This has an important implication for non-price debt contract design, which I elaborate in the next section. Further evidence for the increased access to financing comes from Shan, Tang and Winton (2014). In their paper they examine the impact of CDS on private loan covenant design. They hypothesise that banks which use CDS as a method of risk hedging should become less risky and in turn increases the banks’ perceived risk capacity. This would increase the borrowing capacity of reference firms. In fact they find that on average the loan size increases by 15-19%.
Given the empirical results of increased default risk, tougher debt renegotiations and increased leverage, uninsured creditors are exposed to increased risk for which they are not compensated through interest rates. One possible way unprotected creditors could protect themselves against increased default risk would be trough demanding more covenants. To test for the effect of CDS on covenant strictness Shan Tang and Winton (2014) examine the covenants strictness of reference firms. Their results indicate that when CDS is traded on a firms debt, debt covenants become significantly less strict. This could indicate that CDS provide creditors with a way of default protection that replaces the need for active monitoring.
B. Previous studies on covenant design
Several studies have focused on the drivers of debt contract design and the strictness of covenants in specific. Covenants provide creditors with a contingent way of control. The effects of these covenants vary from limited access to other sources of financing, to an increased creditor influence on the financial direction of a firm (Murfin, 2012). Berlin & Mester (1992) examined why some firms place their debt in the public market and others choose to place their debt privately. They propose a model which states that the ease of renegotiation is a major factor in deciding to issue public or private debt. As mentioned earlier, private loans have more and stricter covenants, but can be renegotiated more easily
10
due to the limited number and concentrated creditors. Zinbarg (1975) even reported that private loans in the portfolio of one specific company in his study are renegotiated once a year on average, while also mentioning that this is not only due to financial distress. In addition, Lummer & McConnell (1989) report that 50% of the debt agreements published in the Wall Street Journal are revisions to prior debt agreements.
As has been shown by Fama and Miller (1972), it is easy to come up with a business plan that increases the expected payoff for the stockholders, but decreases the expected payoff for bondholders, known as ‘asset substitution’. These agency problems can be controlled by including covenants in the debt contract. Aghion and Bolton (1988) view the central trade-off in debt contract design as the trade-off between managing agency problems versus the ability of the firm’s management to make the most efficient decisions. Berlin and Mester (1992) propose a similar trade-off in their model and additionally include the option of renegotiation in the contract design. The conclusion derived from their model is that covenants become less strict when the creditworthiness of a firm increases. An interesting addition is that when creditworthiness is considerably low, an increase in creditworthiness will increase the value of the option to renegotiate the debt contract only up to a minimum level. After this level of creditworthiness has been reached the value of debt renegotiation declines. It is this dependency on creditworthiness and the relationship between CDS trading and credit risk which makes it interesting to study the relationship between CDS trading and covenant design. In similar vein to Berlin and Mester (1992), Smith and Warner (1979) examine how contracts are designed to control the agency problems. They mention two hypotheses, the first one being the ‘Irrelevance Hypothesis’ which follows Modigliani and Miller (1958) in stating that the total value of the firm will not change by including covenants as long as the
investment policy remains fixed. Even when the investment policy is not fixed, Smith and Warner (1979) state that the choice of financial contracts does not influence the firm value. The second hypothesis is the ‘Costly Contracting Hypothesis’. This hypothesis states that the inclusion of covenants can increase the value of the firm by reducing the agency costs. Even though including covenants in a contract can be costly, the Costly Contracting Hypothesis states that these covenants can be used to mitigate the agency costs to such an extent that including these covenants increases the value of the firm. After empirical research, Smith and Warner (1979) found that debt covenants reduce the cost associated with agency problems, indicating support for the Costly Contracting Hypothesis and rejecting the Irrelevance Hypothesis. Consequently they state that there is a unique level of contract design which maximizes firm value. Given the empirical evidence that CDS reference firms experience an
11
increase in leverage which could cause agency problems in itself, including covenants could be a good lever to reduce these agency costs arising from CDS trading.
In contrast to Berlin & Mester (1992) and Smith & Warner (1979), Murfin (2012) investigated the effect of defaults in the loan portfolio of a credit supplier on the contract design in subsequent credit issuances of non-defaulted firms. Murfin (2012) based the motivation for his study on findings by Chava and Purnanandam (2011) that creditors
significantly reduced lending after experiencing a major default in their portfolio. In addition Murfin was motivated by the study of Gopalan, Nanda and Yerramilli (2011) who show that single corporate defaults have an effect on the lead arranger activity in the syndicated loan market. Murfin (2012) used data on syndicated and bilateral loans issued from 1984 to 2008. He proposed his own measure of covenant strictness as being the probability of covenant violation. Using this measure he plotted the covenant strictness from 1996 to 2008 and shows that the strictness recedes in prosperous economic times and significantly increases in difficult economic times. When examining the recent default experience of creditors, Murfin (2012) finds that creditors tighten their covenant strictness in new loans in response to defaults in their loan portfolios. These results hold even when the defaults occur in different industries and different geographical regions. After robustness checks Murfin (2012) finds that next to recent default experience, creditor capital is also an important driver for covenant strictness. Murfin (2012) reasons the tightening of covenants in subsequent debt issues is due to the revision of creditors screening ability. Experiencing a default indicates the screening ability of the creditor failed at the moment the debt contract was designed. The results of Murfin
indicate that a default of one debtor can influence the contract design of another debtor. With the increased use of CDS and their influence on credit risk, the difference in covenant design between CDS firms and non-CDS firms could be small, given the suggestions that default effects spill over to non-default debtors.
C. Channel mechanism
Given the results by previous empirical research which show reference firms have an increased credit risk, uninsured creditors face a risk for which they do not get compensated through higher interest rates. Saretto and Tookes (2013) and Ashcraft and Santos (2009) find CDS trading mainly affects non-price factors. I therefore hypothesise that including more bondholder protective covenants in the bond issuance of reference firms could compensate the
12
uninsured creditor. This would lead to reference firms facing significantly more bondholder protective covenants in their bond issues relative to non-reference firms. Publicly issued bonds are issued through investment banks which can place them at several different types of investors. These bonds can be traded relatively easy and can be held by investors which have limited or no access to credit default insurance. I assume there is a certain fraction of the participating agents in the bond market without access to credit default insurance which is large enough to influence the debt contract design. To the best of my knowledge I am the first to empirically test whether the empty creditor problem results in a conflict between insured and uninsured creditors in which uninsured creditors demand more bondholder protective covenants due to the increased credit risk of reference firms.
13
Data & Methodology
A. Collecting and preparing data
I employ three main datasets for this paper. Bond issuance and covenant data is obtained from CapitalIQ, firm specific information is obtained from the Compustat database and CDS spread data is obtained from Thomson Reuters. In addition, GDP data is retrieved the Bureau of Economic Analysis.
The dataset used to obtain corporate bond issuance data including the specified covenants is the CapitalIQ database. All corporate bond issuances in North-America from January 1997 to March 2015 are downloaded. The focus of this paper is on firms for which a gvkey is available in order to match this dataset to the Compustat dataset. The gvkey is not provided by CapitalIQ and therefore manually inserted by matching on issuer and ultimate parent name. This results in a dataset of 8,174 observations. I then delete 325 observations due to a missing gvkey. In the resulting database of 7,849 observations, 5,188 observations had credit default swaps traded at some point during 1997 to 2015.
The dataset used to obtain company specific data is Compustat. The company specific variables which are downloaded from January 1997 until March 2015 are; gvkey, Fiscal year end, Total Assets, Total long term debt, EBITDA, SIC industry code, yearend market value, working capital and Sales. The CapitalIQ and Compustat databases are then combined using the gvkey and fiscal year end variables to match on. This results in a dataset with 7,528 observations. Every observation for which one or more company specific variable is missing is then deleted (683). Next all observations before 2004 are deleted (671), leaving 6,174 observations. In this dataset a dummy variable CDS is created to indicate whether a CDS contract is being traded at the moment of issuance. This dataset is then matched by gvkey and prior fiscal year end to a dataset containing CDS spreads ranging from 2004 until 2015. If a CDS spread is found, the CDS dummy is changed to 1 and 0 otherwise. The resulting dataset consists of 6,174 observations in which 3,535 observations had a CDS contract traded at the time of the bond issuance. In this dataset I create four covenant data variables; BPT, IRT, SRT and CovT which are the number of bondholder protective covenants, issuer restrictive
14
B. Summary of dataset
Panel A presents summary statistics of the complete sample for each year in the sample. Column 2 from panel A shows that the number of bond issuances rises from 102 issues in 2004 to 1,081 in 2014, an almost tenfold rise in 10 years. Column 3 presents the number of bond issues by year for firms which had credit default swap trading at the moment of issuance. The number of CDS issues rises together with the total number of issues. The fourth column presents the average number of bondholder protective covenants included in the debt contracts. The average number in 2004 was 2.49, which rose to 3.78 in 2015. A similar trend is visible for CDS and non-CDS firms separately. On average, non-CDS firms had 0.1 more bondholder protective covenants included in their debt contracts, although this is not a weighted average.
Panel A. Summary of bond issuance by year
(1) (2) (3) (4) (5) (6) (7) (8) (9) Year Issues CDS-firms Average BPT total Average BPT CDS Average BPT non-CDS Average duration Average issue size ($ million) GDP growth 2004 102 44 2.49 2.52 2.47 25.37 289.48 6.639 2005 210 96 2.99 3.03 2.95 21.88 344.80 6.6705 2006 271 161 3 2.93 3.12 22.45 508.93 5.8211 2007 413 281 3.66 3.69 3.58 20.17 484.94 4.4869 2008 304 217 3.72 3.82 3.48 16.82 582.62 1.6646 2009 383 286 3.78 3.82 3.67 14.13 565.85 -2.0376 2010 521 327 4.12 3.84 4.58 13.08 521.02 3.7847 2011 666 382 4.69 4.35 5.14 12.58 579.97 3.6988 2012 916 530 4.69 4.6 4.82 12.25 589.36 4.1584 2013 1,081 564 4.65 4.47 4.85 11.74 592.87 3.7425 2014 1,081 524 4.61 4.66 4.57 11.14 606.35 3.8812 2015 226 123 3.78 3.91 3.63 10.46 689.94 3.1000* Average 515 295 3.85 3.8 3.9 16.01 529.68 3.4462
Panel A presents summary results of the complete dataset by year of issuance. (1) is the year of issuance. (2) is the number of bond issuances in the given year. (3) is the number of bond issuances of firms which had credit default swaps traded at their debt at the moment of issuance. (4) is the average number of bondholder protective covenants for the complete sample. (5) is the average number of bondholder protective covenants for CDS firms. (6) is the average number of bondholder protective covenants for non-CDS firms. (7) is the average duration of the bond issuances by issuance year. (8) is the average bond issue size in millions of US Dollars. (9) is the percentage US nominal GDP growth. The appendix contains a graphical representation of this table. *represents the US nominal GDP growth forecast from OECD.
Column 7 presents the average duration of the debt issuance. Although the number of issues increased over the years, the average duration decreased by more than 50% from 2004 to 2015. Column 8 shows the average size of the bond issues in millions of US Dollars. The average increased from $289.48 million in 2004 to $689.94 million in 2015. The last column,
15
column 9, presents the nominal GDP growth of the United States. The column clearly shows the effect of the financial crisis during 2008 and 2009.
Panel B presents the same summary statistics as panel A, but now by industry.
Panel B. Summary of bond issuance by industry
(1) (2) (3) (4) (5) (6) (7) (8) (9) Industry Issues CDS-firms Average BPT total Average BPT CDS Average BPT non-CDS Average duration Average issue size ($ million) Average Z-score Agriculture 24 20 5.92 6.1 5 16.79 396.98 12.5095 Mining, Oil&Gas 548 234 4.95 4.26 5.46 13.96 594.68 3.9556 Utilities 907 557 3 2.85 3.25 18.37 366.87 1.4501 Construction 98 74 4.86 4.77 5.13 9.33 273.39 0.2951 Manufacturing 1,904 1,122 4.58 4.54 4.64 11.99 590.71 5.0433 Wholesale Trade 124 59 5.49 5.8 5.22 11.69 445.82 5.7079 Retail Trade 280 209 4.91 4.86 5.06 12.84 688.58 6.7058 Transportation 405 284 4.16 4.04 4.45 17.13 442.55 2.1496 Information 621 366 4.27 3.96 4.71 13.1 854.41 2.9396 Finance&Insurance 696 302 3.31 3.29 3.33 15.64 567.8 0.41601 Real Estate 146 65 4.88 5.31 4.53 10.2 417.72 0.9809 Professional&Technical Services 89 40 3.97 4.05 3.9 8.27 534.14 4.874 Administrative&Support&Waste Management 75 30 4.96 5.57 4.56 11.52 473.91 4.5311 Educational Services 2 - 6 - 6 7.5 207.18 4.9333 Health Care 92 71 5.29 5.15 5.76 12.76 670.12 2.9887 Entertainment 20 7 5.85 6.57 5.46 8.6 434.52 1.4313 Accomodation&Food Services 78 41 5.38 5.44 5.32 10.4 575.36 3.5527 Other 12 8 4.67 4.5 5 10.5 256.33 5.572 Public Administration 53 46 4.23 4.15 4.71 14.7 910.84 - Average - - 4.77 4.73 4.81 12.38 510.63 3.891
Panel B presents summary results for the entire dataset by issuing industry from 2004 to March 2015. A complete overview of the different industries is given in the appendix. (1) indicates the industry. (2) is the number of bond issuances in the given industry. (3) is the number of bond issuances of firms which had credit default swaps traded at their debt at the moment of issuance. (4) is the average number of bondholder protective covenants for the complete sample. (5) is the average number of bondholder protective covenants for CDS firms. (6) is the average number of bondholder protective covenants for non-CDS firms. (7) is the average duration of the bond issuances by issuance industry. (8) is the average bond issue size in millions of US Dollars. (9) is the average Altman Z-score per industry. The appendix contains a graphical representation of this table.
Column 1 indicates the industry of issuance. The second column shows that the Utilities industry and Manufacturing industry have issued the most bonds during the sample. Likewise, the most CDS-firms are in these industries. However, relatively speaking Agriculture and Public Administration have the most CDS firms. The third column presents the average number of bondholder protective covenants by industry. When excluding Wholesale Trade due to the low number of observations, debt contracts from the Agriculture industry contain the most covenants on average. When only looking at the average number of covenants for
16
CDS firms Arts & Entertainment has the most covenants on average. For non-CDS firms the Healthcare industry has the most covenants on average. Column 7 presents the average
duration for each industry. Agriculture has the longest duration on average, while Professional & Technical Services has the shortest duration on average. The last column contains the average issue size in million US Dollars. The Information industry the Public Administration industry have the largest bond issues on average. Column 9 presents the average Altman Z-score by industry. The lowest Z-score comes by far from the Financial Industry. The reason is that working capital is nonexistent for financial firms and thus missing in the Altman Z-score.
Panel C presents a before-after comparison between CDS firms and non-CDS firms. The results on duration and issue size confirm the predictions made by the theoretical model of Bolton and Oehmke (2011), which predicts that credit default swaps increase the access to finance. The average duration increases from 12.67 to 14.77, an increase of 17%, which lies in between the 8% to 21% increase found by Saretto and Tookes (2013).
Panel C. Before and after comparison
Variable Non-CDS CDS Difference
Total covenants 9.572 8.656 0.916***
Bondholder protective covenants 4.391 4.140 0.252***
Duration 12.67 14.77 -2.1***
Issue size ($ million) 468.26 635.95 -167.69***
Panel C presents the before after comparison for CDS versus non-CDS firms. A CDS firm is defined as a firm which had a credit default swap traded at the moment of debt issuance. The non-CDS sample contains 2,639 observations and the CDS sample contains 3,535 observations. The difference column presents the absolute differences. The apteryxes indicate the significance of the results, where *** = p<0.01, ** = p<0.05, * = p<0.1.
The average issue size increases from an average of $468.26 million to an average of $635.95 million. Although this does not directly imply an increase in leverage found by Saretto and Tookes (2013), it does comply with the increased access to finance prediction made by the Bolton and Oehmke model (2011). The total number of covenants and the number of bondholder protective covenants for CDS firms is significantly lower than non-CDS firms. When only looking at the averages these results contradict the first hypothesis in this paper. To draw conclusions however, would need addressing of the selection into CDS trading and other explanatory variables.
17
E. Methodology
This section will describe how the hypotheses will be empirically tested and how I control for possible endogeneity problems. The sample consists of a number of firms by year, but is not necessarily followed by data on the same firm the year after. Therefore the baseline regression model will be an OLS regression on cross-sectional data. To follow Ashcraft and Santos (2009) two CDS variables are constructed. The first CDS variable is a dummy variable CDS which indicates whether a CDS contract is traded on a firms debt at the moment of the debt contract design. The second CDS variable is CDSTI, which is the time invariant dummy to indicate if a CDS contract has been traded on the firms debt during the complete sample period. The inclusion of two CDS variables helps to detect the time invariant factors affecting the debt contract design of CDS firms. Furthermore, the fact that the number of bondholder protective covenants varies from year to year warrants the use of the separate CDS variables to control for the trends in covenants use. In essence, the CDSTI variable will indicate time invariant differences in the number of bondholder protective covenants (BPT) between CDS firms and non-CDS firms. The baseline specification for the regression is constructed as:
(1)
Where CDS is the dummy variable indicating if a CDS contract is traded on a firms debt at the moment the debt contract is designed. To distinguish between effect of CDS and CDSTI on BPT I perform the regression in (1) twice, substituting CDS for CDSTI. This allows me to demonstrate the differences between the effect CDS trading has on the number of bondholder protective covenants and the effect of having CDS firm characteristics (CDSTI). In addition, when using the CDS variable as an explanatory variable, I consider two cases. The first is where I include all firms where CDSTI is either 1 or 0 and for second I exclude firms where CDSTI is equal to 0. is a vector of control variables. These control variables are selected following previous research on covenants strictness such as that of Shan, Tang and Winton (2014). The loan-characteristic variables are; DURATION, which is the duration of the debt issue and IDURATION where the latter is the former interacted with the CDS variable. The borrower-characteristic variables are; DEBTTOCAPITAL which is the total level of long term debt divided by the total assets, ROA which is the earnings before interest, tax, depreciation and amortization divided by the total assets. SALES which is the total amount of revenues generated by each firm during the fiscal year and TOTASSETS which represents the
18
total assets of a firm minus the total cash. Finally, as a measure of firm distress, I include the Altman Z-score for each firm, ZSCORE. As an additional control for year fixed effects and
industry fixed effects I include dummy variables, where is the dummy variable for the
industry and is the dummy variable for the year of issuance.
One concern in this specification and many other empirical studies is the possibility of endogenous effects. Endogeneity could result from two factors, the first factor is reversed causality. A high number of bondholder protective covenants could for instance induce credit default swap trading, or the other way around, a low number of covenants could induce credit default swap trading (Parlour & Winton, 2013). Similarly Subrahmanyam, Tang, and Wang (2014) point out, when studying the effect of CDS trading on credit risk, a high level of credit risk induces CDS trading and CDS trading increases the credit risk. The second endogeneity problem arises from omitted variable bias. CDS trading is not commenced on a firms debt on a random basis, but rather caused by certain factors such as credit risk, industry, size and other factors. Moreover these factors tend to change over time. Although the industry and other firm characteristic variables are observable, some factors such as credit risk are less easily observed and some other variables are not observable at all. Observable determinants are for instance the current debt ratio and the amount of physical assets. Unobservable determinants could again be credit risk. To account for the endogeneity problems I employ two variations on the baseline regression. Since CDS trading selection is non-random, I cannot measure the number of bondholder protective covenants of the treatment group were they not to receive treatment. To control for observable differences between CDS firms and non-CDS firms I employ a propensity score matching model. Propensity score matching became a widely used technique due to its simplicity. Propensity score matching estimates the chance of receiving treatment, conditional on a certain set of covariates. The covariates I use to match the sample are: DEBTTOCAPITAL, ROA, SALES, INDSTR, YEAR, DURATION and TOTASSETS, where YEAR indicates the year of issuance. After matching each CDS firm with a non-CDS firm using nearest neighbour matching I attempt to filter out the selection problem. Summary statistics on the matched sample can be found in the appendix in figure 1. On the resulting matched sample I perform the baseline OLS regression from specification (1). To control for unobservable omitted variables I employ a fixed effects regression model on panel data. To control for firm specific characteristics which differ between firms but are constant over time I include firm fixed effects. To further control for effects which are the same for each firm, but change over time I include time fixed effects.
19
In order to test for the second hypothesis I construct two additional variables; GDP and GDPI. The variable GDP is the nominal US GDP growth in each year. The GDPI variable is the GDP variable interacted with the CDS variable. These two variables are included in the baseline regression to arrive at the following model:
(2)
As explained in the introduction, GDP growth acts as a proxy for the level of Empty
Creditors. According to the second hypothesis the GDPI ratio should be negative, indicating more covenants are used for CDS firms when the level of Empty Creditors increases.
F. Regression by Industry
Given the contradicting results of Subrahmanyam, Tang, and Wang (2014) who do not distinguish between industry versus those of Bedendo, Cathcart and El-Jahel (2011) who only consider non-financial US companies I consider the financial industry and the non-financial industry separately. An additional reason why it is interesting to consider the Financial Industry separately is that Empty Creditors need a particular level of influence to distort the debt renegotiation and force bankruptcy. When there are relatively few bonds outstanding and the issues are relatively small, Empty Creditors can easily gain a significant influence.
Financial firms, like banks, typically have a relative high number of bonds outstanding and a relative high leverage ratio. This could indicate a dispersed base of creditors, reducing the relative influence of Empty Creditors during debt renegotiation. A reduced influence of Empty Creditors should in turn mitigate the demands by uninsured creditors.
According to the hypothesis of this paper, if the Financial Industry does not suffer from the existence of Empty Creditors I should not find a significant positive treatment coefficient, i.e. uninsured creditors do not face an increased credit risk from CDS trading and don’t have to be compensated. I first run the baseline regression on the dataset excluding the Financial Industry. After that I run the same regression for the Financial Industry separately.
20
Empirical Evidence
A. Baseline regression
The main specification in this analysis is the baseline regression on cross-sectional data. To distinguish between the effects of CDS trading and CDS traded the baseline model (1) is performed using CDS and CDSTI as explanatory variables separately. The dependent variable is the number of bondholder protective covenants written in the debt contract at the moment of initiation. The variable of interest in the main specification is the CDSTI/CDS variable. For each specification multiple regressions have been performed to include control variables and interacted control variables. Table 1A shows the baseline OLS regression results on the cross-sectional data. The dependent variable for all three models is the number of bondholder protective covenants (BPT). When excluding control variables in model 1A the effect of CDS traded on BPT is 0.459 less BPT covenant on an average of 4.687, representing a 10% decrease. This coefficient increases to 0.0890 when control variables are added in model 2A. Interacting these control variables with CDS in model 3A increases this coefficient to a significant 0.319 or a 7% increase in the number of bondholder protective covenants. This effect is both statistically and economically significant. These results indicate that firms with CDS traded characteristics tend to have more bondholder protective covenants to start with. The interesting observation is the fact that interacting the controls has a significant effect on the coefficient of interest, indicating that covenants react differently for CDS firms versus non-CDS firms. As expected, firms with high debt to capital ratio’s face more covenants, given the positive coefficient on DEBTTOCAPITAL. The coefficient on SALES is negative, meaning firms with higher sales have less BPT covenants. An interesting observation is the positive sign on the ROA coefficient, indicating more BPT covenants with a high return on assets. This could be due to the fact that high ROA assets tend to be more risky relative to low ROA assets. The coefficient on ZSCORE changes from negative to positive, while we would expect a negative coefficient in all cases, although the negative coefficient is insignificant and possibly related to the ROA variable. Table 1B shows the same baseline OLS regression as in table 1A, but now performed only on the Financial Industry sample. The first model (1B) shows the baseline specification without additional control variables. The resulting coefficient indicates 0.107 more BPT covenants for CDS traded financial firms. Adding control variables to the specification in model 2B increases the coefficient to a significant 0.420. Including the interacted control variables in model 3B results in a significantly positive coefficient of 0.521
21
or a 16% increase in the number of bondholder protective covenants. This result is both statistically and economically significant.
Table 1A Table 1B
Baseline regression excluding Financial Industry Baseline regression on Financial Industry sample
VARIABLES Model 1A Model 2A Model 3A Model 1B Model 2B Model 3B
Constant 4.687*** 3.220*** 3.005*** 3.218*** 2.205*** 2.235*** (0.0521) (0.590) (0.611) (0.0783) (0.236) (0.246) CDSTI -0.459*** 0.0890 0.319*** 0.107 0.420*** 0.521*** (0.0608) (0.0636) (0.0726) (0.104) (0.123) (0.168) DEBTTOCAPITAL 0.100*** 0.0934*** -0.0155 -0.000169 (0.0111) (0.0143) (0.0250) (0.0282) SALES -0.118*** -0.0447** 0.00557 0.0157 (0.0122) (0.0180) (0.0188) (0.0193) ROA 0.119*** 0.138*** 0.115*** 0.0793** (0.0121) (0.0164) (0.0289) (0.0396) ZSCORE 0.0136** -0.00857 0.0715** 0.101*** (0.00659) (0.00925) (0.0309) (0.0333) DURATION -0.0167*** -0.0165*** -0.0171*** -0.0173*** (0.00220) (0.00218) (0.00306) (0.00304) Interacted Controls X X
Industry Fixed Effects X
Year Fixed Effects X X X X
Control variables X X X X
Observations 5,254 5,252 5,252 923 923 923
R-squared 0.012 0.241 0.254 0.001 0.212 0.220
Table 1A presents the regression results of the unmatched sample excluding the Financial Industry. The first model contains only the CDSTI dummy variable without additional control variables. The second model includes the control variables. Model 3 includes the variables DEBTTOCAPITAL, ROA, ZSCORE and SALES interacted with CDS. Table 1B presents the regression results of the unmatched Financial Industry sample. The first model contains only the CDST dummy variable without additional control variables. The second model includes the control variables. Model 3 includes the variables DEBTTOCAPITAL, ROA, ZSCORE and SALES interacted with CDS. Robust standard errors in parentheses where *** = p<0.01, ** = p<0.05, * = p<0.1.
As in the regression in table 1A, the interacting controls has a large impact on the coefficient, again suggesting debt contract design reacts differently on borrower characteristics for CDS firms versus non-CDS firms. The interesting observation is the positive sign in front of the coefficient for both Financial and non-Financial firms, suggesting in general more covenants are used in debt contracts of firms with CDS characteristics. Table 1C shows the baseline regression results of the model using CDS as the explanatory variable. In all six models the dependent variable is the number of bondholder protective covenants. Excluding control variables in model 1 results in a negative coefficient of 0.397 on an average of 4.616, or equal to a decrease of 9%. When including control variables in model 2 the coefficient increases to -0.135. Using interacted controls results in a highly significant coefficient of 2.120, indicating an increase of 46% in the number of bondholder protective covenants. When dropping the non-CDS firms for which CDSTI equals 0, the coefficient becomes 0.694 and is significant at
22
the 5% level. This represents an increase of 15%. An interesting observation comes from model 5 and 6, which present the results of the baseline regression using CDS on the Financial Industry sample. In model 5 the coefficient on CDS becomes insignificant and when dropping the non-CDS firms in model 6 this coefficient becomes significantly negative at the 1% level, indicating 2.515 less covenants are used for Financial firms when CDS is traded at the
moment the debt contract is designed. In contrast to the model using CDSTI, the observations for the Financial Industry using CDS differ from those of the non-Financial firms.
Table 1C
Baseline regression using CDS as explanatory variable
VARIABLES Model 1 Model 2 Model 3 Model 4 Model 5 Model 6
Constant 4.616*** 3.659*** 2.394*** 3.440*** 2.447*** 5.025*** (0.0464) (0.587) (0.653) (0.745) (0.271) (0.563) CDS -0.397*** -0.135** 2.120*** 0.694** -0.00593 -2.515*** (0.0569) (0.0606) (0.216) (0.310) (0.316) (0.655) DEBTTOCAPITAL 0.104*** 0.139*** -0.00489 -0.00223 0.0474 (0.0112) (0.0154) (0.0299) (0.0290) (0.0524) SALES -0.0955*** 0.00747 -0.139*** 0.0269 -0.252*** (0.0125) (0.0181) (0.0291) (0.0202) (0.0485) ROA 0.121*** 0.166*** 0.177*** 0.0669 -0.0919 (0.0121) (0.0165) (0.0347) (0.0411) (0.133) ZSCORE 0.0150** -0.000325 -0.0272 0.0984*** -0.0249 (0.00663) (0.00928) (0.0184) (0.0330) (0.0719) DURATION -0.0165*** -0.0167*** -0.0134*** -0.0176*** -0.0145*** (0.00220) (0.00217) (0.00246) (0.00305) (0.00389) Excluding non-CDS X X
Interacted control variables X X X X
Industry Fixed Effects X X X - -
Year Fixed Effects X X X X X
Control variables X X X X X
Observations 5,256 5,254 5,254 3,493 923 459
R-squared 0.010 0.241 0.264 0.313 0.211 0.341
Table 1C presents the regression results of the unmatched sample. The first model contains only the CDS dummy variable without additional control variables. The second model includes the control variables. Model 3 includes the variables DEBTTOCAPITAL, ROA, ZSCORE and SALES interacted with CDS. In model 4 the observations for which no CDS contract has been traded during the entire sample period have been dropped, resulting in 3,493 observations. Model 4 presents the results for the Financial Industry sample including CDS firms. Model 6 presents the results for the Financial Industry sample excluding non-CDS firms. Robust standard errors in parentheses where *** = p<0.01, ** = p<0.05, * = p<0.1.
Combining the results of table 1A, B and C provides evidence for the main hypothesis in this paper, though further econometric specifications are needed to rule out endogeneity problems.
23
B. Baseline regression on matched sample
To address the possible endogeneity problems with CDS trading the baseline
regressions are also performed on a matched sample. The sample is matched using propensity score matching on the following covariates; DEBTTOCAPITAL, ROA, SALES, INDSTR, YEAR, DURATION and TOTASSETS. By using propensity score matching I assume all factors determining debt contract design can be observed using the covariates above. The baseline OLS regression from (1) has been run on the matched sample, leading to the results in table 2A, B and C. Again the dependent variable in the regressions is the number of
bondholder protective covenants. The variable of interest is the CDS/CDSTI variable. As with the unmatched sample, three regressions have been performed to show the effect of including control variables and interacted control variables. Without any control variables model 1 of table 2A shows a decrease in the number of covenants of 0.440 on an average of 4.674, equivalent to a 9% decrease. Including the control variables in model 2 increases this coefficient to 0.176. Including both controls and interacted controls in model 3 results in a positive coefficient of 0.349 which is significant at the 1% level and represents an increase in the number of BPT of 7%. After controlling for observable selection differences, these results are consistent with the results from table 1A. The coefficient signs on the control variables are consistent with the unmatched results. Table 2B shows the results from the baseline
regression exclusively on the Financial Industry sample. Model 1B presents the results without additional control variables, resulting in a positive coefficient 0.261. Including the control variables in model 2B increases the coefficient to 0.483. Including the interacted control variables in model 3B decreases the coefficient to 0.434 which is insignificant, but still positive, representing an increase of 14% in the average number of Bondholder Protective Covenants for Financial firms with time invariant CDS characteristics.
24
Table 2A Table 2B
Baseline regression excluding Financial Industry Baseline regression on Financial Industry sample
VARIABLES Model 1A Model 2A Model 3A Model 1B Model 2B Model 3B
Constant 4.674*** 3.382*** 3.139*** 3.019*** 2.296*** 2.766*** (0.0733) (0.566) (0.573) (0.170) (0.358) (0.480) CDSTI -0.440*** 0.176** 0.349*** 0.261 0.483*** 0.434 (0.0802) (0.0748) (0.0876) (0.184) (0.176) (0.273) DEBTTOCAPITAL 0.0655*** 0.0486** -0.0766** -0.0787 (0.0135) (0.0207) (0.0339) (0.0491) SALES -0.181*** -0.103*** 0.00206 0.0125 (0.0139) (0.0229) (0.0258) (0.0286) ROA 0.0733*** 0.0912*** 0.176*** 0.138*** (0.0142) (0.0245) (0.0318) (0.0523) ZSCORE 0.0191** -0.00664 0.0506 0.105** (0.00783) (0.0157) (0.0415) (0.0437) DURATION -0.0166*** -0.0165*** -0.0147*** -0.0150*** (0.00240) (0.00237) (0.00388) (0.00392) Interacted Controls X X
Industry Fixed Effects X
Year Fixed Effects X X X X
Control variables X X X X
Observations 4,136 4,136 4,136 506 506 506
R-squared 0.009 0.274 0.282 0.005 0.291 0.297
Table 2A presents the results of the regressions on the matched sample excluding the Financial Industry. The sample has been matched on the variables; DEBTTOCAPITAL, ROA, SALES, INDSTR, YEAR, DURATION and TOTASSETS using propensity score matching. The resulting sample contains 4,136 matched observations. Model 1 is the basic model without additional control variables. Model 2 is the basic model including the control variables. Model 3 is the model with control variables and interacted control variables; DEBTTOCAPITAL, ROA, ZSCORE and SALES interacted with CDS. Table 2B presents the results of the regressions on the matched sample when the INDSTR = 10 (Financial Industry). The sample has been matched on the variables;
DEBTTOCAPITAL, ROA, SALES, INDSTR, YEAR, DURATION and TOTASSETS using propensity score matching. The resulting sample contains 506 matched observations. Model 1 is the basic model without additional control variables. Model 2 is the basic model including the control variables. Model 3 is the model with control variables and interacted control variables; DEBTTOCAPITAL, ROA, ZSCORE and SALES interacted with CDS. The robust standard errors are in parentheses where *** = p<0.01, ** = p<0.05, * = p<0.1.
As with table 2A, matching the sample produces similar results as the unmatched sample. In general the results from table 2A and B confirm that both Financial firms and non-Financial firms with CDS traded characteristics tend to have more bondholder protective covenants as a basis. An interesting observation is the negative sign for DEBTTOCAPITAL for the Financial Industry. This indicates less covenants are used when leverage increases. An explanation for this observation could be the fact that banks naturally have high leverage ratios and this ratio might not reflect credit risk the same way as for non-Financial firms.
25
Table 2C
Baseline regression using CDS as explanatory variable on matched sample
VARIABLES Model 1 Model 2 Model 3 Model 4 Model 5 Model 6
Constant 4.653*** 3.998*** 2.594*** 3.364*** 3.183*** 6.746*** (0.0678) (0.566) (0.642) (0.689) (0.538) (0.986) CDS -0.434*** -0.0395 1.733*** 0.930** -0.329 -3.691*** (0.0753) (0.0731) (0.314) (0.449) (0.485) (0.941) DEBTTOCAPITAL 0.0673*** 0.106*** 0.00183 -0.102** -0.0820 (0.0136) (0.0253) (0.0397) (0.0497) (0.111) SALES -0.164*** -0.0423* -0.122*** -0.00549 -0.306*** (0.0145) (0.0255) (0.0406) (0.0405) (0.0658) ROA 0.0731*** 0.135*** 0.193*** 0.0965 -0.194 (0.0142) (0.0255) (0.0468) (0.0589) (0.194) ZSCORE 0.0200** 0.00318 -0.00709 0.106*** -0.0147 (0.00793) (0.0159) (0.0308) (0.0403) (0.0716) DURATION -0.0165*** -0.0165*** -0.0139*** -0.0150*** -0.0147*** (0.00241) (0.00237) (0.00256) (0.00388) (0.00400) Excluding non-CDS X X
Interacted control variables X X X X
Industry Fixed Effects X X X - -
Year Fixed Effects X X X X X
Control variables X X X X X
Observations 4,137 4,137 4,137 3,250 506 403
R-squared 0.009 0.273 0.286 0.313 0.293 0.326
The table above presents the regression results of the matched sample using CDS as an explanatory variable. The first model contains only the CDS dummy variable without additional control variables. The second model includes the control variables. Model 3 includes the variables DEBTTOCAPITAL, ROA, ZSCORE and SALES interacted with CDS. In model 4 the observations for which no CDS contract has been traded during the entire sample period have been dropped, resulting in 3,250 observations. Model 5 presents the results for the Financial Industry sample including non-CDS firms. Model 6 presents the results for the Financial Industry sample excluding non-CDS firms. Robust standard errors in parentheses where *** = p<0.01, ** = p<0.05, * = p<0.1.
Table 2C presents the results of the baseline regression using CDS as the explanatory variables. For all six models the dependent variable is the number of bondholder protective covenants. The first four models are performed on the complete sample excluding the Financial Industry and the last two models are performed exclusively on the Financial Industry. Excluding control variables in model 1 results in a negative coefficient of 0.434. Including the control variables in model 2 increases the coefficient to -0.0395. As in the results of the previous tables, including interacted controls in model 3 changes the coefficient significantly to 1.733, or an increase of 37% on the average of 4.653. When dropping the non-CDS traded firms in model 4 the coefficient reduces to 0.930 which is significant at the 5% level and represents an increase of 20% in the number of bondholder protective covenants used. As with the unmatched regression, the coefficient for the Financial Industry is significantly negative at the 1% level, indicating less covenants are used for Financial Industry debt when CDS is traded at the moment of debt issuance. Overall the results of the
26
matched sample comply with the results of the unmatched sample. These results provide enough statistical evidence to conclude more bondholder protective covenants are used in the debt contracts of non-financial CDS firms. For Financial Industry firms the results suggest the opposite effect, which raises an interesting question of why these firms might be less
susceptible to Empty Creditors. A discussion will be provided in a separate section following after robustness checks.
C. Fixed effects regression
Even though the regressions above include control variables and are matched using propensity score matching, propensity score models only control for observable factors. The regressions could still suffer from omitted unobservable variable bias. To check for the possibility of unobservable omitted variables I run a fixed effects regression on panel data. The sample I collected in this study contains certain firms on which no CDS contracts were traded at the beginning of the sample, but had CDS contracts traded on their debt in a later stage of the sample. This allows me to directly compare these firms including time fixed effects and firm fixed effects. The model to estimate the effects of CDS trading on covenants is specified as:
(3)
Where represents the firm fixed effects, represents the time fixed effects and
represents the vector of control variables; DEBTTOCAPITAL, ROA, ZSCORE,
TOTALASSETS, SALES and DURATION. I run three regressions, one on the complete sample, one on the sample excluding the Financial Industry and one exclusively on the Financial Industry. The results of these regressions are displayed below. In all models the dependent variable is the number of bondholder protective covenants. The variable of interest is CDS. For each table four regression specifications have been run to consider the effects of including fixed effects and control variables. Table 3A shows the results of the fixed effects regression on the complete sample. Model 1 is the basic regression without fixed effects. The coefficient on CDS is 0.660 which is significant at the 1% level and represents an increase of 19% on an average of 3.534. Including firm fixed effects in model 2 reduces this coefficient to 0.402. Including interacted controls in model 3 increases the coefficient to 0.659 which is not significant, but positive and represents an increase of 19%.
27
Table 3A
Fixed effects regression model complete sample
VARIABLES Model 1 Model 2 Model 3 Model 4
Constant 3.534*** 2.228*** 1.972** 3.148*** (0.408) (0.619) (0.797) (0.850) CDS 0.660*** 0.402*** 0.659 0.231 (0.132) (0.144) (0.567) (0.596) SALES -0.0462 0.172* -0.0149 0.0170 (0.0561) (0.0905) (0.101) (0.100) DEBTTOCAPITAL -0.0233 0.0790 0.149** 0.155*** (0.0347) (0.0512) (0.0583) (0.0582) ROA 0.173*** 0.0985* 0.0310 0.000904 (0.0419) (0.0556) (0.0747) (0.0747) ASSETS -0.0407 -0.0212 0.193 0.0124 (0.0598) (0.101) (0.119) (0.125) ZSCORE 0.0205 -0.0157 -0.00149 -0.00130 (0.0250) (0.0289) (0.0284) (0.0281) DURATION -0.0245*** -0.0206*** -0.0180*** -0.0172*** (0.00429) (0.00437) (0.00429) (0.00452)
Time fixed effects X
Interacted control variables X X
Firm fixed effects X X X
Control variables X X X X
Observations 533 533 533 533
R-squared 0.0796 0.116 0.173 0.222
Number of firms 92 92 92 92
This table presents the results of the fixed effects model on the complete sample. The panel data set contains 533 observations with 92 different firms. Model 1 shows the difference in difference regression results without fixed effects, but including control variables which are divided in 10 quintiles. Model 2 shows the difference in difference results including firm fixed effects and control variables. Model 3 shows the results of the same estimation as model 2, including the interacted control variables. The interacted control variables are: DEBTTOCAPITAL, ROA, SALES, ZSCORE and ASSETS. Model 4 shows the results of the regression including control variables, interacted control variables, firm fixed and time fixed effects. The standard errors are in parentheses where *** = p<0.01, ** = p<0.05, * = p<0.1.
Model 4 includes time fixed effects and results in an insignificant coefficient of 0.231. Although the first two models present results which were expected, including interacted controls causes the coefficient to become insignificant and including time fixed effects causes the coefficient to decline to 0.231. The time fixed effects control for unobservable variables that differ over time, but influence each firm equally. One such unobservable variable could be investor sentiment or appetite for risk. As shown by previous research the use of covenants varies over time and is correlated with the state of the economy. Table 3B presents the results of the panel data regression on the sample excluding the financial industry. The results of the first three columns are in line with table 3A. The interesting observation here is that when including time fixed effects, the coefficient becomes not only insignificant, but also negative.
28
Although no statistical conclusions can be drawn, it is clear that time fixed effects affect the results. I will discuss these results further in section F.
Table 3B
Fixed effects regression model excluding Financial Industry
VARIABLES Model 1 Model 2 Model 3 Model 4
Constant 2.806*** 1.660** 1.921* 3.459*** (0.522) (0.696) (0.980) (1.054) CDS 0.813*** 0.463*** 0.392 -0.0431 (0.164) (0.176) (0.893) (0.904) SALES 0.0531 0.247** 0.0982 0.0844 (0.0683) (0.107) (0.131) (0.129) DEBTTOCAPITAL 0.0147 0.0896* 0.130* 0.132* (0.0431) (0.0534) (0.0686) (0.0685) ROA 0.216*** 0.111* 0.0307 0.0138 (0.0478) (0.0603) (0.0875) (0.0874) ASSETS -0.138* -0.0291 0.147 0.0357 (0.0740) (0.122) (0.160) (0.161) ZSCORE 0.0202 -0.0117 -0.00249 0.00426 (0.0265) (0.0295) (0.0296) (0.0295) DURATION -0.0253*** -0.0191*** -0.0191*** -0.0217*** (0.00679) (0.00693) (0.00692) (0.00705)
Time fixed effects X
Interacted control variables X X
Firm fixed effects X X X
Control variables X X X X
Observations 381 381 381 381
R-squared 0.0905 0.146 0.173 0.232
Number of firms 76 76 76 76
This table presents the results of the fixed effects model on the sample excluding the Financial Industry. The panel data set contains 381 observations with 76 different firms. Model 1 shows the difference in difference regression results without fixed effects, but including control variables. Model 2 shows the difference in difference results including firm fixed effects and control variables. Model 3 shows the results of the same estimation as model 2, including the interacted control variables. The interacted control variables are: DEBTTOCAPITAL, ROA, SALES, ZSCORE and ASSETS. Model 4 shows the results of the regression including control variables, interacted control variables, firm fixed and time fixed effects. The standard errors are in parentheses where *** = p<0.01, ** = p<0.05, * = p<0.1.
Table 3C presents the results of the panel data regression exclusively on the Financial Industry sample. Model 1 shows the results without fixed effects, but including control variables. The coefficient is 0.390. When including firm fixed effects this coefficient increases to 0.786 which is significant at the 1% level. Including the interacted control variables in model 3 decreases the coefficient to an insignificant 0.145. When further including time fixed effects the coefficient increases to 1.721, which is insignificant at the 10% level. Where the results of the baseline regression differed significantly for Financial Industry firms and non-Financial Industry firms, the results of the fixed effects regression mainly show consistent results. Just as for the complete sample and the sample excluding the
