Does the trading of CDS increase firm’s bankruptcy risk through the empty creditor problem?

(1)

Does the trading of CDS increase firm’s bankruptcy risk through

the empty creditor problem?

Master thesis

By Elenka Stefanova

MSc Business Economics: Finance track Amsterdam Business School

University of Amsterdam

June 2015

(2)

Statement of Originality

This document is written by Elenka Stefanova who declares to take full responsibility for the contents of this document.

I declare that the text and the work presented in this document is original and that no sources other than those mentioned in the text and its references have been used in creating it.

The Faculty of Economics and Business is responsible solely for the supervision of completion of the work, not for the contents.

(3)

Abstract

The separation of control from cash flow rights caused by the presence of the Credit Default Swap (CDS) contracts is referred in the academic literature as the “empty creditor problem”. The current research verifies the existence of the problem by examining the changes in firm’s bankruptcy risk when unfavorable events, such as the recent financial crisis and a downgrading event, occur. In particular, the methodology is based on a difference-in-difference estimation, where the dependent variable is the Merton’s distance-to-default. The regression results from the economic downturn setting confirm the hypothesis that the bankruptcy risk of CDS entities increased more steeply than that of non-CDS companies during the crisis. Although the decline in the default measure caused by a negative rating change does not differ between both groups, the probability of CDS company to be downgraded is statistically higher than that of non-CDS one during the financial crisis period. Following the findings of Campello & Matta (2013), it could be argued that the inferences about the existence of the empty creditor problem are valid also during an economic expansion. Finally, the regression estimates do not reveal any difference in the increase of bankruptcy risk between CDS and non-CDS entities when the rating classification is taken into account.

Keywords:

Credit Default Swaps, empty creditor problem, bankruptcy risk, distance-to-default, financial crisis, downgrading

(4)

The Credit Default Swap (CDS) is considered to be one of the major financial innovations in the recent decades. Its market reaches exceptionally high volume of trading in the pre-crisis period, namely $58 trillion in notional amount.1 Therefore, understanding the nature and specific characteristics of the security is crucial for future investment strategies. One of the recently discussed features of the derivative contract is that it can potentially trigger empty creditor problem. In essence, theory predicts that creditors insured with CDS contract would resist any out-of-court renegotiations, and thus, drive the debtor into bankruptcy in order to subtract higher surplus from liquidation. Therefore, the separation of control from cash flow rights caused by the CDS trading alters the debtor-creditor relationship and results in socially inefficient outcome, i.e. firm’s default.

Having said that, the current research aims to confirm empirically the hypothesis for the existence of empty creditors investigating entities’ bankruptcy risk when unfavorable events, such as the recent financial crisis or downgrading, occur. In particular, the study hypothesizes that the problem of insured bondholders arises only when the debtor faces financial difficulties and hence, we should observe steeper increase in bankruptcy risk of reference entities during these events. In order to verify the hypothesis, the thesis applies the distance-to-default based on Merton’s model, which is a continuous measure of credit quality and enables the computation of bankruptcy risk changes. Furthermore, the methodology is grounded on a difference-in-difference estimation, which analyzes the different decline in the default measure from pre to post event period between the treatment (CDS) and control (non-CDS) groups. In order to compare the entities, one-to-one propensity score matching based on size, leverage and industry is applied one year prior to the sample time period. In essence, the empty creditor problem is confirmed if the distance-to-default of CDS firm during the recent financial crisis or during a downgrading event decreases more steeply than that of non-CDS one.

The contribution of the current research is in the proposed empirical procedure, because it takes into account the findings of the related literature and overcomes the problems faced in other studies. In particular, the distance-to-default measure enables an investigation of the changes in bankruptcy risk ex-ante, whereas the logistic models applied for instance in the researches by Subrahmanyam et al. (2012) and Peristiani and Savino (2011) could be conducted only ex post, when the outcome, i.e. firm bankruptcy, is already observed. In this sense, the current

(6)

2 methodology allows the inclusion of large dataset, and the derived conclusions are not grounded on a limited number of bankruptcy filings. In addition, the proportional hazard model implemented in the latter two articles assumes that the empty creditor problem is persistent in time, though, the theory predicts that lenders rely on the CDS insurance only when the firm faces financial difficulties. Since the thesis examines the recent financial crisis and the downgrading event as a trigger of the problem this feature is taken into account. Moreover, following the findings of Narayanan and Uzmanoglu (2015), we should not expect higher incidence of bankruptcy, but rather only increased default risk, because the debtor anticipates potential empty creditors. In this sense, the current research produces more accurate results for the existence of the problem compared to the related literature, which generally studies default or debt restructuring frequencies. Finally but not least, since the difference-in-difference estimation investigates the changes in the dependent variable, the methodology overcomes the problem of non-random assignment of treatment group, i.e. of CDS trading. In the setting of the current research, where the characteristics of reference and non-reference firms differ substantially, this feature limits the effect of potential endogeneity problems.

Taking into account the considerations mentioned above, the thesis analyzes the empty creditor problem applying a data sample from 2004 until 2010. In summary, the regression results of both settings, namely the financial crisis and downgrading events, justify the hypothesis that the bankruptcy risk of CDS entities increased more steeply during the recent financial crisis than that of non-CDS firms. In particular, the distance-to-default measure of reference firm declines on average by around 8% more than that of non-reference one during the economic downturn. Moreover, the probability of downgrade in the treatment group is 16% higher than that in the control group during the same period. If we assume validity of the inferences derived by Campello and Matta (2013) that over-insurance is pro-cyclical and by Bolton and Oehmke (2011) that the empty creditor problem is triggered by over-insurance, we can conclude that the estimates in the current research verify the existence of the problem not only during an economic downturn, but also during an expansion.

The rest of the thesis is structured as follows. The theoretical background of the CDS contract and the empirical findings of the related articles are discussed in Section 2. Section 3 develops the hypotheses and the methodologies applied to test them. The data and the descriptive statistics are analyzed in Section 4. The empirical results from the regression analyses are introduced in Section 5. Further, two robustness checks are conducted in Section 6 and finally, the thesis discusses potential endogeneity problems (Section 7) and concludes in Section 8.

(7)

3 2. Literature review and background

By definition the CDS is a financial swap agreement, which serves as an insurance contract against the default of a debtor. In this sense, a firm creditor buys the contract from a protection seller for a pre-specified premium, also called credit spread. In return, the insurer commits to pay the difference between the par value and the recovery value of the bond if credit event occurs. The definition of the latter is viewed more broadly, since it ranges from entity’s bankruptcy, failure to pay to debt restructuring. Accordingly, several types of CDS contracts existed, whose enforcement was triggered by these different events. The most common form of insurance were against the restructuring as a type of credit event, called Modified Restructuring (MR), and against only the default of the debtor, called No Restructuring (NR) contracts. However, due to the Big Bang protocol in April 2009 restructuring events are no longer considered as credit events on the CDS market.

The main purpose of the CDS contract is to secure better credit risk allocation, since the risk is transferred from the firm’s bondholder to the CDS protection seller. As a consequence, entity’s creditor still holds voting rights in case of credit event, but now the cash flows stem from other source. In the academic literature this problem is referred to as “empty creditor problem” and is firstly introduced by Hu and Black (2008). The academics state that the separation of control from cash flow rights could create significant scope for strategic behavior. The conclusion is based on the premise that because of the insurance contract the debt holder has no incentive to monitor the entity, moreover, no incentive to roll the debt, grant new financing, or agree to voluntary debt restructuring in case of financial difficulties. This is due to the fact that the surplus from liquidation provided by the CDS seller might exceed the value of the restructured debt. As a consequence, the firm defaults, although the optimal social outcome would be to continue its operations. On the opposite side, if there aren’t any CDSs traded on a firm, its bondholders would benefit of debt restructuring and the borrower would potentially avoid bankruptcy. Overall, the empty creditor problem suggests that the trading of CDS can alter the debtor-creditor relationship leading to inefficient outcome.

According to the derivative pricing theory, the value of the CDS can be replicated by a portfolio including a long position in the underlying bond and a short one in the Treasury bond. Therefore, in a complete market the security does not improve the risk allocation, it only decreases the transaction costs investors face. Thus, the derivative is in theory characterized as a redundant security, i.e. it doesn’t have any impact on the price or performance of the

(8)

4 underlying firm. However, the empty creditor problem implies that the presence of the CDS trading changes the economic outcome. In particular, it suggests that the probability of firm’s bankruptcy increases because of the tough bondholders, who are protected with CDS contracts. In this sense, Bolton and Oehmke (2011) provide anecdotal evidences from the recent years, where reference entities eventually filed for Chapter 7 bankruptcy after failed renegotiations with “empty creditors”.

The researchers also establish a formal model, which theoretically explains the problematic behind the amended debtor-creditor relationship through the introduction of CDS. In particular, the model assumes the presence of limited commitment problem, which means that the borrower could refuse to meet the contractual payments because of strategic behavior. According to the researchers, the introduction of CDS will insure the fulfillment of obligations and decrease the number of strategic defaults, i.e. serve as a commitment device. Consequently, because of the enhanced bargaining power of creditors in renegotiations the efficiency of the firm’s investments increases, as well as its borrowing capacity. Opposed to the potential ex ante benefits, the ex post analysis suggests that the presence of CDS triggers the empty creditor problem, and therefore, a large number of inefficient bankruptcies compared to the social optimum. According to the analysis, there are two driving factors behind the empty creditor issue. First, the separation of lender’s control rights from cash flow rights creates an incentive to reject any debt renegotiation due to higher surplus from liquidation. Second, the lender’s optimal level of insurance exceeds the social one, which leads to inefficient equilibria with empty creditors and consequently, firm bankruptcy. In addition, Bolton and Oehmke (2011) argue that the effect of over-insurance is more severe, when the debt is held by multiple creditors. Overall, the academics conclude that the presence of CDS contract could theoretically increase the efficiency and the amount of financing a firm receives, but could also raise the incidence of bankruptcies.

While the research by Bolton and Oehmke (2011) concentrates on a theoretical cost-benefit analysis of the CDS trading, other academics investigate the negative implications of the derivative. The study by Subrahmanyam et al. (2012) aims to provide empirical insights into the changes following the inception of the CDS contracts. According to their results, the probability of default, the leverage ratio and the number of creditors of the reference entity increase after the introduction of the insurance contract. In addition, Subrahmanyam et al. (2012) provide evidence for the existence of the empty creditor problem using proportional hazard model. According to their estimates, the higher the amount of CDS contracts

(9)

5 outstanding relative to the total debt, the higher the probability of bankruptcy. Peristiani and Savino (2011) apply the same methodology to test the empty creditor issue with the difference being the explanatory variable, which is a dummy equal to one if the particular firm has listed CDSs. However, the estimated results suggest that there is no statistically significant relationship between the probability of default and the empty creditor problem with the exception of year 2008.

Although the regression analysis in the study by Subrahmanyam et al. (2012) is based on an extensive time frame, the number of bankruptcies in the sample constitutes only 0.14% of the total observations. Similarly, the research by Peristiani and Savino (2011) is based on a sample of only 43 bankruptcy filings of CDS companies. In this sense, the proportional hazard model, which is applied in both studies, might not be the appropriate empirical setting to test the hypothesis. In contrast to these two researches, the current thesis applies a linear model instead of proportional one. In order to do so, the distance-to-default (DD) is introduced as dependent variable, which is designed as a continuous measure of credit quality and enables an investigation of the changes in the firm’s bankruptcy risk following a particular event. Peristiani and Savino (2011) attempt to confirm the existence of empty creditors applying similar methodology; however, the level of the default measure does not differ for CDS and non-CDS entities according to their estimates.

While the two latter papers try to analyze the empty creditor problem from the perspective of probability of default, other studies concentrate on the successful debt restructurings. This approach assumes that the creditors of a firm do not participate in out-of-court renegotiations, if they are insured by CDS contracts. However, Mengle (2009) and Bedendo, Cathcart, and El-Jahel (2011) do not report any significant results for the former hypothesis when applying time-series or cross-sectional data. On the opposite, Danis (2013) provides evidence for the resistance of insured creditors applying the participation rate in the offer as a dependent variable. However, the inference is based on a sample of bonds with diverse seniority and the fact that CDS contracts are usually traded only on senior unsecured debt is not taken into account. In this sense, the study by Narayanan and Uzmanoglu (2015) is supplementary to the previous one, since it takes the debtor’s side into consideration. The academics demonstrate that the entities anticipate the potential empty creditors and therefore, restructure their debt in a proper manner, i.e. disproportionally targeting senior and junior bondholders. Having said that, trying to identify the empty creditor issue by examining out-of-court renegotiations might lead to false conclusions or to insignificant results as seen in previous studies. Thus, the

(10)

6 methodology suggested in the current research, which concentrates on firm’s default probability, might be more appropriate to verify the existence of empty creditors.

In addition, Narayanan and Uzmanoglu (2015) argue that the empty creditor holdout increases firm’s default probability, but because of the debtor’s response it does not increase the number of bankruptcies. This finding is crucial for evaluating the methodology applied in the current paper. In particular, it does not concentrate only on firms, which are in financial distress or eventually filed for Chapter 7 or 11 bankruptcy, but rather on events, which overall increased entity’s bankruptcy risk, but not necessarily lead to default. Hence, it takes into account that firms anticipate creditors’ behavior.

Particularly important is the research by Campello and Matta (2013), which is the basis for the empirical settings applied in the current paper. The academics argue that the over-insurance of creditors is more pronounced in economic boom as it is in recessions, i.e. lower effect of empty creditors during the recent financial crisis should be expected. Consequently, if a significant relationship between the presence of CDS contracts and firm’s bankruptcy risk is found when examining the downturn of 2007-2009, then it could be concluded that the empty creditor problem exists also during expansions.

There is a large number of papers, which study different implications of the CDS contracts and their impact on firm’s characteristics. Saretto and Tookes (2013) examine the effect of CDS inception on firm’s capital structure and infer that reference firms are characterized with increased leverage ratio, as well as extended bond maturities. Danis and Gamba (2014) derive the same conclusion by using a single-period debt renegotiation model. On the opposite, Ashcraft and Santos (2009) find no evidences for decreased debt financing cost after the inception of CDS contracts.

Stulz (2010) analyses the potential impact of the derivative on the recent financial crisis and states that the insurance contracts cannot be blamed for the severe downturn in the financial sector as other academics suggest. At the same conclusion arrives Jarrow (2010), who indicates that the derivate was actually beneficial for investors, since it facilitates better risk allocation. In this sense, one should expect that the magnitude with which the financial crisis affected reference and non-reference companies is similar.

(11)

7 3. Methodology

In order to confirm the existence of empty creditors, one should firstly analyze when the problem arises. The debt holders insure themselves against unfavorable events and they receive the par value of the bond only if these events occur. Hence, insured creditors could become empty only when the firm is in or close to financial distress. In essence, it should not be expected that the empty creditor problem is persistent in time, but rather that it appears in “bad times”, i.e. when the overall economic conditions are unfavorable or when the firm faces difficulties on individual level. One of the drawbacks of the proportional hazard model applied in previous studies is that it assumes that the problem is evident during the entire period of the sample. (Peristani and Savino (2011)). In order to overcome this issue, the current research examines two particularly bad events: recent financial crisis and credit rating downgrading. The first affected negatively the entire economy, the second one is an indicator that the entity is performing poorly. In both cases, the creditors would be exposed to potential losses and instead of bearing them, they would chose to become empty creditors and receive the surplus from liquidation, i.e. let the firm defaults. In contrast, bond holders of non-CDS companies would welcome the opportunity to restore at least some of the debt value. All of the above leads to the conclusion that if the empty creditor problem exists, in these two bad events the bankruptcy risk of CDS entities increases more steeply than the one of non-CDS. This does not imply that reference firms have higher default probability overall, but rather that its change is larger.

To analyze the change in bankruptcy risk, the current research applies the distance-to-default (DD). First, since the variable is continuous and not discrete measure of credit quality, one could compute its changes. Second, it could be calculated in every time period even if the firm still operates. In contrast, the proportional hazard model used in previous studies could be conducted only ex post, when the outcome is already determined, i.e. the firm has already gone into bankruptcy. This leads to the third advantage of applying the distance-to-default, namely the research does not concentrate only on cases, where firms eventually filed for bankruptcy. As mentioned earlier, analysis based on default frequencies are executed on a limited number of observations. In contrast, the current methodology enables the inclusion of a large data sample. Finally, as discussed above, Narayanan and Uzmanoglu (2015) provide evidence that the debtor is aware of potential empty creditors and act accordingly in debt restructuring. As a result, the firm succeeds in avoiding bankruptcy. This leads to two conclusions: first, examining ex post default rates and second, testing the hypothesis analyzing debt

(12)

8 renegotiations might not prove the existence of empty creditors. Hence, the current methodology aims to confirm the empty creditor problem by investigating bankruptcy risk when the firm still operates. In this sense, the debtor’s actions are also taken into consideration. All in all, the current research intend to validate the following hypothesis:

In case of bad event the distance-to-default of a firm decreases more steeply if it has listed CDSs due to the empty creditor problem.

A specific characteristic of the derivative, which contradicts to the insurance theory, is that the protection buyer does not need to have a direct exposure to the underlying bond, i.e. does not need to have an “insurable interest”. Therefore, in order to verify the hypothesis, the assumption is made that at least some of the CDS holders are also bondholders of the reference entity.

3.1 Financial crisis

The negative impact of the recent financial crisis is present in the entire economy, not only in the financial sector, and thus, it could be instantly considered as “bad” event. In order to confirm the hypothesis of empty creditors, larger negative drop in the distance-to-default of CDS compared to non-CDS entities is expected during the crisis period.

However, before discussing the methodology, it should be analyzed whether the crisis affected reference and non-reference entities in the same way. Since the crisis originated from the financial industry, the protection sellers were subject to high losses and financial difficulties, which in some cases lead to bankruptcies. In this sense, bond holders insured with CDS contracts had to bear the cost of counterparty risk and presumably not be able to use their insurance, i.e. debt holders might not have had the opportunity to become empty creditors. As a consequence, one should expect that if the problem exists, then its magnitude during the recent financial crisis is rather small. Furthermore, Campello and Matta (2013) develop a model, which tries to predict the optimal insurance level of creditors taking the economic conditions into account. The academics conclude that the over-insurance ispro-cyclical, i.e. debt holder tends to buy more protection during expansions. As Bolton and Oehmke (2011) argue, one of the driving factors of the empty creditor problem is exactly the over-insurance. Altogether, if the level of protection during recessions is relatively small, then we should not expect any evidences for empty creditors during the crisis. Consequently, if the proposed

(13)

9 methodology confirms the steeper increase in the bankruptcy risk of CDS companies even in a downturn, then the empty creditor problem should also exist in good economic times.

Overall, the research hypothesizes that:

There is a larger decrease in the distance-to-default of CDS companies during the recent financial crisis and it is due to the empty creditor problem.

In order to compute the distance-to-default, the thesis applies Merton’s (1974) model. In particular, a system of the following two equations in equity E and stock volatility σe is

simultaneously solved, so that the approximate firm value V and volatility σv could be

estimated.

𝐸 = 𝑉𝑁(𝑑₁) − 𝐷е−𝑟𝑇_𝑁(𝑑

2) (1)

𝜎_𝐸 =𝑉

𝐸𝑁(𝑑1)𝜎𝑉 (2)

Afterwards, the distance-to-default is calculated using the following formula:

𝐷𝐷 =

log(𝑉 𝐷⁄ )+(𝑟𝑓−𝜎𝑣

2_{⁄ )𝑇}₂

𝜎𝑣 √𝑇 (3)

Where the debt value D is computed as the sum of current liabilities and one half of the long-term debt. The equity value of the entity E is the product of the closing price and the shares outstanding as of the end of the year. The volatility σe is calculated based on daily returns in

the corresponding year and then annualized. The risk-free rate rf is assumed to be the Treasury

three-month constant maturity rate as provided by the Federal Reserve Bank. The time horizon of the default measure is one year.

Furthermore, propensity score matching is applied before the regression analysis so that to each CDS firm one non-CDS company is assigned. The selection criteria are the industry the entities are operating in, the natural logarithm of total assets and leverage ratio. It is controlled for industry, because it might be correlated with the presence of CDS contracts. In addition, the academic literature has proven that the latter two differ significantly in reference and non-reference entities. In particular, CDS companies are characterized as large and mature firms, which implies that there are substantial differences in the size of the two groups. Finally, several studies document increased leverage ratio of CDS firms due to the ease of rising debt (Saretto and Tookes (2013) and Danis and Gamba (2014)). The values for the propensity score are taken one year before the sample period.

(14)

10 Finally, the thesis applies the following difference-in-difference estimation based on panel data with distance-to-default as dependent variable to test the existence of empty creditors:

𝐷𝐷_𝑖.𝑡 = 𝛼_𝑖,𝑡 + 𝐶𝐷𝑆_𝑖 ∗ 𝐹𝑖𝑛𝑎𝑛𝑐𝑖𝑎𝑙 𝑐𝑟𝑖𝑠𝑖𝑠_𝑡+ 𝐶𝐷𝑆_𝑖+ 𝐹𝑖𝑛𝑎𝑐𝑖𝑎𝑙 𝑐𝑟𝑖𝑠𝑖𝑠_𝑡+ 𝑉𝑜𝑙𝑎𝑡𝑖𝑙𝑖𝑡𝑦_𝑖,𝑡+ 𝐸𝑥𝑐𝑒𝑠𝑠 𝑟𝑒𝑡𝑢𝑟𝑛_𝑖,𝑡 + 𝐿𝑒𝑣𝑒𝑟𝑎𝑔𝑒 _𝑖,𝑡+ ln (𝐴𝑠𝑠𝑒𝑡𝑠)_𝑖.𝑡+ 𝑢_𝑖,𝑡 (4) Where the dummy CDS indicates whether the firm has listed insurance contracts. The Financial crisis variable is equal to one if the corresponding year is in the crisis period, i.e. 2007, 2008 or 2009, and zero otherwise. The reasoning for defining these years as a crisis period is provided in the next section and is based on the data applied in the current research. In this case, the intercept represents the level of DD of a non-CDS company in the before crisis period. The variable of interest is the interaction term, which measures the difference in the slope of distance-to-default of CDS compared to non-CDS companies due to the bad event. If the regression estimate is negative and significant, then the drop in bankruptcy risk of CDS firm is larger compared to non-reference entity, which confirms the hypothesis of empty creditors. Following Das, Hanouna and Sarin (2009), who conclude that market-based variables are better predictors of bankruptcy risk, a number of variables is included to control for: return volatility, calculated based on daily observations and then annualized; profitability, measured by the annual excess return over the CRSP value-weighted index; capital structure, represented by leverage ratio; and size, denoted by the natural logarithm of total assets.

3.2 Credit rating downgrading

As already discussed, in order to test the existence of empty creditors, one should analyze situations, where the firm undergoes particularly bad events. Although the credit downgrade is not defined as a credit event in the sense of the CDS contract and therefore it is not a direct trigger of the empty creditor problem, it has an indirect impact on the subsequent firm’s performance and it is also an indicator of problems before the rating change. The credit downgrade demonstrates that the company underperforms and most importantly, that its probability of default has increased. As a consequence, the creditors of an entity are less likely to receive the interest payments and the principal of the bond back. On the CDS market this also means that the spread paid to the protection seller raises. Moreover, it is rarely the case that an entity files for bankruptcy without any preliminary signals, one of which could be the negative change in credit quality. In the sense of the empty creditor problem, we should expect that insured creditors will anticipate the potential negative impact of the rating change and therefore, will aim to use their protection, i.e. will become empty creditors.

(15)

11 In the related literature there is a number of studies, which investigate the relationship between a credit downgrade and the market reaction to its announcement. The results from previous articles, which concentrate on the hypothesis that these events are anticipated and therefore, there is no market reaction afterwards, are mixed. According to Katz (1974), the prices of the underlying are affected already before the announcement of the downgrade, and hence, no additional response is observed. Wansley, Glascock and Clauretie (1992) and Dynkin,Hyman and Konstantinovsky (2002) argue that bond prices are influenced negatively before as well as after the announcement. Steiner and Hanke (2001) though confirm that negative bond returns are realized after the credit change. Micu, Remolona and Wooldridge (2004) on the other hand analyze the effect of credit downgrade on the CDS spread and conclude that there is still some information added by the announcement. Finally and most importantly, Aggarwal, Singh and Thomas (2012) provide evidence that the change in distance-to-default in the period before the downgrade is informative about the following rating changes. Having said that, in order to test the empty creditor hypothesis, the thesis assumes that first, market participants are aware or at least foresee the subsequent downgrading and second, the poor firm performance is reflected in the DD before the event.

All taken together, the current setting aims to confirm the following hypothesis:

The decrease in distance-to-default during a credit downgrade of a CDS firm is steeper than that of a non-CDS one due to the empty creditor problem.

In order to verify the hypothesis, the thesis first identifies all of the credit rating downgrades and upgrades. The latter are applied to control for different effects between positive and negative changes in credit quality. According to Aggarwal, Singh and Thomas (2012), the change in distance-to-default in the preceding three quarters can explain the subsequent credit rating change. Hence, the default measure is computed on a quarterly basis in the three quarters before and after the event of announcement. In addition, it is calculated in the same manner as in the previous section (3.1. Financial crisis) with the difference being that quarterly information is applied. In particular, the debt is the sum of current liabilities plus one half of the long-term debt as of the end of the corresponding quarter. The market value is computed as shares outstanding times the closing price in the end of every third month before and after the month of the rating change. The return volatility equals the annualized standard deviation of daily returns in these three months. The risk-free rate is assumed to be the Treasury three-month constant maturity rate as provided by the Federal Reserve Bank. As a result, for each

(16)

12 observation of credit downgrade, respectively upgrade, seven annualized values of the distance-to-default are computed.

Similarly to the financial crisis specification, propensity score is applied one year before the sample period. The CDS and non-CDS firms are matched according to the industry they are operating in, the logarithm of total assets and the leverage ratio. Finally, a triple difference-in-difference estimation based on panel data is used to test the hypothesis of the empty creditor problem:

𝐷𝐷_𝑖.𝑡 = 𝛼 + 𝐶𝐷𝑆_𝑖 ∗ 𝑃𝑜𝑠𝑡_𝑖,𝑡∗ 𝐷𝑜𝑤𝑛𝑔𝑟𝑎𝑑𝑖𝑛𝑔_𝑖.𝑡+ 𝐶𝐷𝑆_𝑖 ∗ 𝑃𝑜𝑠𝑡_𝑖,𝑡+ 𝐶𝐷𝑆_𝑖 ∗ 𝐷𝑜𝑤𝑛𝑔𝑟𝑎𝑑𝑖𝑛𝑔_𝑖.𝑡 + 𝑃𝑜𝑠𝑡_𝑖,𝑡∗ 𝐷𝑜𝑤𝑛𝑔𝑟𝑎𝑑𝑖𝑛𝑔_𝑖.𝑡 + 𝐶𝐷𝑆_𝑖 + 𝑃𝑜𝑠𝑡_𝑖,𝑡+ 𝐷𝑜𝑤𝑛𝑔𝑟𝑎𝑑𝑖𝑛𝑔_𝑖.𝑡+ 𝑉𝑜𝑙𝑎𝑡𝑖𝑙𝑖𝑡𝑦_𝑖,𝑡+ 𝐸𝑥𝑐𝑒𝑠𝑠 𝑟𝑒𝑡𝑢𝑟𝑛_𝑖,𝑡 + 𝐿𝑒𝑣𝑒𝑟𝑎𝑔𝑒 _𝑖,𝑡+

ln (𝐴𝑠𝑠𝑒𝑡𝑠)𝑖.𝑡+ 𝑢𝑖,𝑡 (5)

Where the dummy CDS indicates whether the firm has listed insurance contracts. The Post variable is equal to one if the observation covers the quarter of announcement or one of the following three quarters. The dummy Downgrading takes the value of one if there is a negative change in the credit rating and zero if there is a positive one. Thus, the intercept depicts the level of distance-to-default of a non-CDS firm before it gets a higher rating. Furthermore, the interaction term of the latter two (𝑃𝑜𝑠𝑡𝑖,𝑡∗ 𝐷𝑜𝑤𝑛𝑔𝑟𝑎𝑑𝑖𝑛𝑔𝑖.𝑡) is the additional change in DD

after the announcement of a downgrade of a non-CDS firm relative to an upgrade. The product of the CDS and Downgrading dummies is interpreted as the additional effect on bankruptcy risk of a CDS company when compared with a non-CDS one due to a negative rating change in the before period. Whereas the interaction 𝐶𝐷𝑆_𝑖 ∗ 𝑃𝑜𝑠𝑡_𝑖,𝑡 represents the added impact on DD of an upgrade of a CDS firm compared to a non-CDS after the announcement. Finally, the triple interaction term (𝐶𝐷𝑆_𝑖∗ 𝑃𝑜𝑠𝑡_𝑖,𝑡∗ 𝐷𝑜𝑤𝑛𝑔𝑟𝑎𝑑𝑖𝑛𝑔_𝑖.𝑡) is the variable of interest, because it represents the difference in the slope of the distance-to-default curve between the two groups when a downgrade takes place. In other words, if the analysis confirms that the coefficient is negative and significant, then the bankruptcy risk of a CDS entity increases more steeply in case of bad event due to the empty creditors. Lastly, the same controls are included in the regression as in the case of the financial crisis, namely return volatility, excess return, leverage ratio and logarithm of total assets. As already discussed, the analysis is based on quarterly data, hence, the variables are computed accordingly.

(17)

13 4. Data and descriptive statistics

As explained earlier, there were different types of CDS contracts, though only one of them is implemented in the current research. The logic suggests that the empty creditor problem arises only by CDS contracts, which recognize the firm’s bankruptcy alone as credit event (“No Restructuring” clause). If creditors hold insurance also against possible debt restructuring (“Modified restructuring”), they will receive the difference between the par value and the renegotiated one in any case. Therefore, we should not expect them to act as empty creditors. Having said that, the data in the current research concentrates only on CDS contracts, which are subject to the “No Restructuring” clause. In addition, the analysis is based on CDSs issued on senior debt with maturity of five years. The information about the type of contracts and the firms, on which they are traded, is gathered from Datastream. All of the financial and balance sheet data is subtracted from CRSP/Compustat merged database. The information about the risk-free rate is taken from the Federal Reserve Bank. The records for credit rating up- and downgrades are downloaded from Compustat.

4.1 Financial crisis

In order to test the hypothesis in the setting of the recent financial crisis, the research is based on a time window from 2004 until 2010. The initial sample collection includes data for all U.S. based entities excluding financials. Since the computation of distance-to-default requires a number of input variables (stock price, shares outstanding, current liabilities, long-term debt) all observations with missing values for these variables are excluded from the sample. As a result, the data contains observations for 3600 firms in total, 450 of which have CDS contracts traded on their debt. As explained above, the control group (non-CDS firms) is assigned to the treatment group (CDS firms) applying a propensity score matching at the end of 2003, i.e. one year before the sample period. The selection criteria are the industry the entities are operating in, the natural logarithm of total assets and leverage ratio. The industry is represented by the first two digits of the SIC code. The leverage is calculated as liabilities divided by total assets. In addition, the latter two are winsorized to the 1% level on the right side. The pseudo R-squared from the matching amounts to 53.34%. However, there are substantial differences between the CDS and non-CDS companies. Table 1 displays the mean values and the standard deviations of the assets, liabilities, leverage ratio and earnings per share of both groups in the end of 2003. Although their capital structure and profitability are similar, the length of their balance sheets differs significantly. As expected, CDS companies are much larger entities,

(18)

14 probably because insurer will be willing to sell protection only to the debt of these firms that have proven to be mature and stable, and whose probability of default converges to zero.

The same conclusion can be drawn by observing firms main characteristics throughout the sample window. Figure 1 plots the annual average values of balance sheet and market-based variables for both CDS and non-CDS firms from 2004 until 2010. The values are winsorized to the 1% level one or both sided depending on their distribution. The plotted data confirms that CDS companies are larger compared to non-CDS ones based on their total assets, liabilities and market value. The differences are significant to the 1% level when tested with two-sample t-test in every year of the estimation period. Although the leverage ratio of both groups in 2004 seems to be quite similar, its path changes drastically in the sample period. The liabilities-to-assets ratio of reference entities increases, whereas it decreases on average in the non-reference group. This observation is in alignment with the hypothesis that the presence of the derivative increases entity’s borrowing capacity (Bolton and Oehmke (2011)). In terms of stock performance however, both groups do not differ significantly. The annualized return and standard deviation based on daily observations suggest that on average the market perceives both groups similarly. In addition, it should be pointed out that in most of the variables there is an obvious effect of the financial crisis and as expected, it is in the same direction for both groups.

Applying Merton’s (1974) model, the distance-to-default is calculated on annual basis as of the end of each calendar year. The path of its mean values throughout the sample period is plotted in Figure 2for both groups. The first insight of the graph is that the distance-to-default of CDS companies is higher on average than that of non-CDS ones in every year of the window. In general, this observation is consistent with the fact that CDS companies are larger and more stable than non-CDS ones as seen in Figure 1, which means that the derivative is traded only on firms, which are less likely to file for bankruptcy.

The second intuition of the plot is the magnitude of the decrease in the default measure during the financial crisis. The average values in 2007 and 2008 suggest that the decline of the DD in the reference group is steeper than that in non-reference one. The annual changes in bankruptcy risk are plotted in Figure 3. As discussed earlier, the existence of empty creditors would be confirmed if the regression analysis verifies that the difference in the decline of distance-to-default between both groups is significant. Finally, both graphs (Figure 2 and Figure 3) suggest that the effect of the crisis is first evident in 2007, the lowest point is reached in 2008, whereas

(19)

15 the recovery phase starts in 2009 and in 2010 the distance-to-default is almost at its initial values. This observation will be later taken into account when defining the financial crisis period.

Although initially the data contains 450 CDS and 450 non-CDS entities, some of them drop from the sample. The total number of companies, which went inactive each year during the time period, and the reasons for that as provided by Compustat are presented in Table 2. According to the table, bankruptcy or liquidation is rather infrequent reason for firms to be delisted. This observation is in line with the argument that models, which test the hypothesis of empty creditors based on firms’ default frequencies, will produce insignificant or inaccurate results. However, the larger number of M&A activities could not only be attributed to strategic behavior, but rather one might suggest that at least some of these entities were problematic. In this sense, the current methodology takes this presumption into account. Furthermore, Table 2 visualizes that the probability of firm to become inactive decreases when CDS contracts are traded on its debt. This argument holds particularly in the first three years of the sample period, though, this effect diminishes during the crisis. This finding is an indirect evidence for the existence of empty creditors, i.e. there is steeper increase in the probability of being delisted for CDS firms during the downturn.

4.2 Credit downgrading

The analysis of the credit downgrading as a trigger of the empty creditor problem starts with identifying the firms, which are up- or downgraded in the period from 2004 until 2010. The information for the changes in issuer rating is taken from Compustat. As in the financial crisis case, all of the observations with missing values for current liabilities, long-term debt, shares outstanding and stock price, as well as the financials, are excluded from the data. As a result, the sample includes 753 entities in total, 321 of which have listed CDS contracts on their debt. These are then matched with the non-CDS firms based on the first two digits of the SIC code, the logarithm of total assets and the leverage ratio in 2003, i.e. one year before the sample period. The pseudo R-squared of the matching process amounts to only 25.8%, but this could be easily explained with the low number of companies in total. The average values of assets, liabilities, leverage ratio and EPS for both groups in 2003 are displayed in Table 3. Overall, the information confirms the inference derived so far, namely CDS companies are significantly larger than non-CDS ones. Moreover, in the credit downgrading specification the differences

(20)

16 in the total assets and liabilities between both groups are even larger than these in the financial crisis case.

Due to the one-to-one matching the sample includes equal number of treatment and control entities, namely 321 CDS and 321 non-CDS firms. However, Table 4 demonstrates that there are substantial differences in the number of up- and downgrades per group in the same time window. In particular, the CDS entities are less likely to receive a higher rating and more likely to receive a lower rating when compared with the non-CDS companies during the recent financial crisis. Though, this observation does not hold in the pre-crisis period. In addition, Figure 4 illustrates the number of companies per rating category, which are subsequently downgraded. According to the graph, the distribution of the CDS firms is skewed to the left, i.e. to the investment grade area. On the opposite, most of the non-CDS companies are characterized with a “junk” rating. In general, this finding infers the conclusion that both groups are substantially different. Furthermore, it confirms the insight that CDS sellers provide protection only to mature and stable entities, which are also marked with higher credit quality. All in all, the difference in the rating distribution will be taken into account when testing the model, because it will be applied to two subsamples of companies with investment and “junk” grade.

As already explained, the distance-to-default is calculated in the quarter of announcement of the rating change, as well as in the three quarters before and after it. The average values of the default measure in case of up- and downgrading for both CDS and non-CDS companies are depicted in Figure 5. According to the graph, the CDS entities are marked with lower bankruptcy risk on average regardless whether there is a positive or negative change in credit quality. Nonetheless, it should be pointed out that in case of an upgrade the distance-to-default does not follow the expected pattern. In particular, no distinct increase of the variable in the chosen window is observed, which might suggest that there is a different time reaction in the default measure against up- and downgrades, or just that there is a potential delay in the upgrades provided by the rating agencies. On the opposite, the path of the distance-to-default in case of a downgrade by both groups gives the impression that the assumption made in Section 3.2 holds, namely market participants anticipate the rating change. As a result, the DD decreases before the event and flattens afterwards. However, if we observe the average values of the default measure in the subsamples before and during the financial crisis, which are displayed in Figure 6 and Figure 7, we can infer that this is not the case. According to the graph based on the expansion period, the distance-to-default continues to decline even after the

(21)

17 announcement of the downgrade. In contrast, the bankruptcy risk decreases on average after the “bad” event, if it is during the crisis, which can be attributed to the recovery period as already seen in the financial crisis case. All taken together, there are substantial differences in the path of the default measure between both time windows, hence, the regression analysis will be conducted on the entire sample as well as on both periods separately. Finally, but most importantly, the three figures do not provide any evidence for the steeper increase in the bankruptcy risk of CDS-entities when compared with non-CDS ones. In contrast, the paths of the average default measure of both groups seem to be parallel, which is an indication that the hypothesis of the empty creditor problem might not be confirmed.

5. Regression results

In the following two subsections the methodologies discussed so far are applied to the data in the financial crisis, respectively in the downgrading, specification in order to confirm the existence of empty creditors.

5.1 Financial Crisis

As Figure 2illustrates the distance-to-default of CDS companies is higher on average in every year of the sample period. In order to confirm this observation, a regression analysis is conducted, where the dependent variables is the DD and the explanatory one a CDS dummy equal to one if the underlying firm has the derivative traded on its debt, and zero otherwise. Additionally, a set of controls is included. The estimates based on the entire sample are displayed in Table 5, where the last three columns include time, entity and/or industry fixed effects. The coefficient of the variable of interest, i.e. the CDS dummy, is positive and highly significant in all of the specifications. Moreover, the R-squared suggest that most of the variation in the dependent variable is explained by the models. All in all, the reference firms demonstrate significantly lower bankruptcy risk, which, as already discussed, confirms the logic that a CDS seller would provide insurance only to the debt of safe entities. The results are also in agreement with the conclusions derived in the study by Subrahmanyam et al. (2012). According to the academics, the inception of the derivative contract generally increases the default probability of an entity, though, the bankruptcy risk of CDS firms is overall lower. The estimates of the controls are in the same direction as expected and are highly significant as well. This observation is not surprising, since the variables are direct or indirect inputs for the calculation of the distance-to-default and their effect on firm’s default risk is anticipated. In

(22)

18 particular, the volatility of the stock returns and the leverage ratio have a negative relationship with the dependent variable, on the opposite the excess return over the CRSP value-weighted index and the size of the entity affect the default measure positively.

As discussed in Section 3.1, the current research hypotheses that the empty creditor problem arises only when unfavorable circumstances occur. The recent financial crisis can be considered as such an event, because it affected the entire economy negatively. Furthermore, there is no obvious link between the crisis and the reference entities, thus, we can assume that it generally had the same impact on CDS and non-CDS firms. Moreover, we should expect that if the empty creditor problem exists, its effect during an economic downturn is relatively low. All of this taken together leads to the conclusion that if any difference in the decline of the distance-to-default due to the crisis is found, it can be attributed to the presence of the derivative contract, and therefore, to the empty creditor problem. The hypothesis is tested applying regression equation (2) and the results are presented in Table 6, where the second column includes industry and the third one entity fixed effects.

As already mentioned, Figure 2 visualizes that the bankruptcy risk of the companies increases already in 2007 due to the crisis and its effect diminishes completely in 2010. Hence, the hypothesis for the existence of empty creditors is tested on data period from 2004 until 2009, i.e. 2010 is excluded from the estimation. Furthermore, the dummy variable Financial crisis is equal to one, if the corresponding year is 2007, 2008 or 2009, and zero if the observation is in the first three years of the sample. As a consequence, the interaction term in equation (2) (𝐶𝐷𝑆𝑖 ∗ 𝐹𝑖𝑛𝑎𝑛𝑐𝑖𝑎𝑙 𝑐𝑟𝑖𝑠𝑖𝑠𝑡) represents the difference in the slope of the dependent variable

between the reference and non-reference groups from the first to the second period. According to the results, the estimated coefficient is negative and significant in all of the regression specifications, which means that the bankruptcy risk of CDS entities increased more steeply during the crisis compared to that of non-CDS ones, i.e. the hypothesis is confirmed. In the setting of the current research, this implies that the presence of the empty creditors decreased the distance-to-default of CDS firms with around 0.6 points on average. Holding constant the other covariates, the presence of the contract causes an additional decline in the bankruptcy risk measure of 8%. In comparison, the effect of the financial crisis itself amounts to 11%, i.e. the trading of the derivative almost doubles the magnitude of crisis. From an economic perspective, the higher decline of the DD suggests that although firms with CDS contracts are safer in general, they become more fragile when facing financial difficulties. All in all, the performance of these companies seems to be pro-cyclical. In addition, the coefficient of the

(23)

19 CDS dummy is positive and significant and with almost the same magnitude as the one on the interaction term with the exception of the third specification, where it is controlled for entity fixed effects. This implies that these two effects, the positive and the negative one, cancel each other on average, i.e. CDS and non-CDS firms possess the same bankruptcy riskin “bad” times. This observation could be also seen in Figure 2.

Furthermore, the coefficient of the Financial crisis dummy and the estimates of the control variables are as anticipated with the exception of the natural logarithm of total assets. The coefficient is negative and significant when controlled for entity fixed effects, which suggests that the larger the company, the higher its probability of default. According to Gilson, John, and Lang (1990), the negative term can be explained by the fact that the relationship between firm’s bankruptcy risk and its size is concave. This is due to the fact that the debt structure gets more complicated with the length of the balance sheet. Finally, the explanatory powers of the models indicate that the larger part of the variation in DD is explained. In conclusion, the theory of empty creditors is confirmed when investigating the changes in bankruptcy risk due to the financial crisis.

5.2 Credit downgrading

In order to test the hypothesis of the empty creditor problem in the downgrading setting, the thesis applies regression equation (3). The results based on the entire sample period, namely from 2004 until 2010, are provided in the first column of Table 7. As already discussed, the distance-to-default has a distinct pattern before and during the crisis (without year 2010), hence the same analysis is conducted separately in these two periods, the estimates of which are reported in the second and third columns of the table. In addition, Figure 4 displays that the rating distribution of CDS and non-CDS firms differs, thus, the model is tested in subsamples of companies with investment grade and with “junk” grade, i.e. higher, respectively lower, than a BB+ rating. The regression results of the latter two are provided in the last columns of Table 7. Entity and time fixed effects are included in all of the specifications.

With regard to the proposed methodology, the hypothesis of the empty creditor problem would be confirmed, if the triple interaction term, i.e. 𝐶𝐷𝑆_𝑖 ∗ 𝑃𝑜𝑠𝑡_𝑖,𝑡 ∗ 𝐷𝑜𝑤𝑛𝑔𝑟𝑎𝑑𝑖𝑛𝑔_𝑖.𝑡, is negative and significant. According to the results however, the coefficient is not statistically different from zero in any of the specifications, which means that the distance-to-default of reference and non-reference entities declines with the same magnitude when the issuer is downgraded.

(24)

20 In this sense, it appears that the creditors of a CDS firm behave in the manner as these of a non-CDS one, and consequently they do not turn into empty creditors.

Even though the hypothesis is not confirmed when applying the distance-to-default as a bankruptcy risk measure, the coefficients of the other variables provide some important insights. First, the CDS entities are characterized with lower default probability when the estimation period is based on the recent financial crisis and when the sample includes only companies with investment grade. The first finding is in alignment with the results based on the crisis as a trigger of the empty creditor problem. The second, however, suggests that even if both groups are classified by the rating agencies to have similar credit quality, the CDS firms still appear to be safer to invest in. In contrast, when the model is tested applying the entire sample, the bankruptcy risk of CDS entities seems to be higher. Overall, it could be inferred that there are inherent differences between the treatment and the control groups, which are supposedly amplified by the recent crisis.

The second intuition of the results is that there is indeed an increase in the default measure due to an upgrade, which is measured by the coefficient of the Post dummy. The estimate is positive and significant in the specifications based on the entire and on the pre-crisis periods, which was not particularly visible in Figure 5 and Figure 6. Third, according to the coefficient of the variable Downgrading in the subsample based on the crisis, the level of DD is higher when a downgrade follows, and not when an upgrade. This observation might be explained by the fact that positive rating changes during this time window occurred after the entities have been downgraded first. In addition, the magnitude and the significance of the control variables are as anticipated. Lastly, in all of the specifications more than 80% of the variation in the default measure is explained by the regression.

As noticed from Table 4, the CDS firms have been subject to more downgrading and less upgrading events when compared with the same number of non-CDS companies. Having said that, the thesis applies a different methodology to test the hypothesis of the empty creditor problem, namely it uses the following logistic model:

Pr(𝑌_𝑖,𝑡 = 1|𝑋_{𝑖,𝑡−1}) = 1

1+exp (−𝛼−𝛽′_𝑋

𝑖,𝑡−1) (6)

Where the dependent variable is a dummy equal to one if the firm is downgraded, and zero if upgraded. The matrix X includes the explanatory variables CDS, the distance-to-default and the same set of controls (return volatility, excess return, leverage ratio and logarithm of total

(25)

21 assets). The lagged values of the independent variables are taken in the quarter prior to the event. In essence, the empty creditor problem would be confirmed if the coefficient of the CDS dummy is positive and significant.

Following the considerations explained above, the model is estimated based on the entire sample, on the pre and post financial crisis periods, and on the investment and “junk” grade subsamples. The results from the corresponding specifications are reported in Table 8. According to the estimates, the reference group is indeed more likely to be downgraded during the financial crisis. In particular, the marginal effect of the CDS contract during this period is 15.8%, i.e. the probability of a CDS entity to be downgraded is almost 16% higher than that of a non-CDS firm. In addition, this effect is also visible when the analysis is performed on the entire sample, though its magnitude decreases to almost 10% due to the inclusion of the pre-crisis period. The same conclusion is derived by Subrahmanyam et al. (2012) using the same methodology. According to the academics however, the marginal effect of the derivative on the probability of downgrades amounts to only 0.8%. The imparity can be explained with the different data periodicity and time periods applied in the analysis. Furthermore, Peristiani and Savino (2011) find out that the bankruptcy risk of CDS entities is higher than that of non-CDS ones only in year 2008, which is also in alignment with the results provided in the current research. All in all, even if the methodology based on distance-to-default does not confirm the empty creditor problem, an investigation of the downgrading rates manifests that the CDS entities are indeed subject to the “treat” insured creditors. Finally, the estimates of the CDS dummy in the last two specifications, i.e. based only on investment or on “junk” graded entities, are insignificant. Possible explanation is the different distribution of downgrading events in the treatment and control groups, which is also displayed in Figure 4. According to the graph, the reference entities are usually characterized with investment, whereas the non-reference ones with “junk” grade. Consequently, the comparison of CDS and non-CDS companies within one of the two credit quality categories might be subject to omitted variable problem and therefore, produces insignificant results.

Since the number of downgrades of reference firms significantly exceeds this of non-reference ones, we should expect that this is also reflected in the distance-to-default. In this sense, the default measure of CDS firm should decrease indeed more steeply than that of non-CDS entity, and the hypothesis would be confirmed also by applying the first methodology. Nonetheless, this is not the case and there is a potential explanation, namely delay in the rating change provided by the rating agencies. As academics and practitioners in the field argue, the credit

(26)

22 quality of issuers is not frequently revised and hence, up- and downgrades are held over. Anecdotal evidences from the recent financial crisis, where triple A rated bonds are downgraded to a single B rating in a time window of several months, imply that these considerations are indeed crucial for the analysis. Actually these conjectures are confirmed by the results conducted by the logistic model (Table 8). According to the coefficient of the Distance-to-default in the financial crisis specification, the values of the default measure cannot predict subsequent rating changes. Since the models used by rating agencies to assign companies to particular credit quality categories are all grounded on Merton’s model, which is also applied in the current research, it is implausible to assume that the DD in the previous period cannot explain following up- or downgrades. All taken together, it appears that the announcement of the rating change does not coincide with the actual change in firm’s bankruptcy risk. Consequently, methodology which concentrates on the time subjective rating change as an event might not produce accurate results as seen in the current research.

The estimates in Table 8 though imply that the driving factor of a subsequent change in credit quality during the crisis is the return volatility. In contrast, there is no significant relationship between the standard deviation of the returns and the probability of downgrades during the expansion or when the model is tested on different credit qualities. It should be also mentioned that increased leverage ratio is the main trigger for a subsequent negative rating change in the subsample of investment graded entities, tough, the coefficient is insignificant in the other specifications. The estimates of the size measured by the log of total assets are insignificant as well. Finally, it should be pointed out that the explanatory power of the model is relatively low in all of the specifications, namely 20% on average. Nonetheless, the results provided in the study by Subrahmanyam et al. (2012) report an R-squared of only around 10%. In this sense, the higher marginal effect of the CDS contract measured in the current research might be more plausible.

All taken together, the results infer that the CDS entities are characterized with higher increase in bankruptcy risk during the recent financial crisis, when applying the crisis, as well as the downgrading as a “bad” event. Additionally, the regression results do not suggest any difference between entities with investment or “junk” grade. Consequently, it can be concluded that the existence of the empty creditor problem is only confirmed, when the economy is in a downturn. Since the estimates in the setting of a rating change do not provide any evidences for the “tough” creditors during the expansion, no general inferences can be derived about the existence of the problem. On the opposite, if we follow the arguments discussed by Campello

(27)

23 and Matta (2013) that the over-insurance is pro-cyclical and by Bolton and Oehmke (2011) that over-insurance is one of the driving factors of empty creditors, we can conclude that the problem is more severe during economic boom. Having said that, from the current results we can argue that if the hypothesis is confirmed during the crisis, then it should be also valid during the expansion. As already discussed, the insignificant results can be explained by the delay in rating changes provided by the rating agencies.

6. Robustness checks

In order to test the validity of the results, the thesis applies two robustness checks to the financial crisis specification. In the first one the model is tested on a sample of middle-sized entities, in the second, the change in distance-to-default is applied as a dependent variable.

6.1 Robustness check on the middle 50% of the sample

One of the main concerns when analyzing the results from the financial crisis is that the CDS entities are significantly larger than non-CDS. The differences in firms’ main characteristics are evident in the propensity matching process as well as in the summary statistics during the sample period. Therefore, one should cautiously draw conclusions when there is an inherent distinction between control and treatment group. In order to overcome this problem, the thesis applies a robustness check to the financial crisis case based only on the middle 50% of the entities as measured by their size. In particular, CDS firms whose total assets lie above or below the 25% threshold of the distribution are excluded from the sample. Afterwards, propensity score matching is conducted based again on the industry, natural logarithm of total assets and leverage ratio at the end of 2003. The pseudo R-squared of the regression does not differ substantially from the one in the main analysis (53%), namely it amounts to 57%. Table 9 reports the mean values and standard deviation of total assets and liabilities, leverage ratio and earnings per share for the 225 CDS and 225 non-CDS firms. According to the table, both groups indicate similar characteristics in the new setting, nonetheless, the differences are still significant to the 1% level when tested with two-sample t-test.

Following the same methodology as in Section 3.1 (Financial crisis), Table 10 presents the regression results based on equation (2) and on the middle half of the sample. The coefficient of the interaction term is negative and significant, which implies that the empty creditor problem is confirmed here as well. Its magnitude is slightly lower in the three specifications compared to the results in Table 6. However, it should be pointed out that the estimate of the

(28)

24 CDS dummy is negative and significant when entity fixed effects are included. This finding implies that on individual level the distance-to-default of middle-sized CDS companies is actually lower than that of non-CDS ones. Furthermore, the bankruptcy risk and the size of an entity have been proven to have concave relationship (Gilson, John, and Lang (1990)), which means that exactly the middle-sized firms are supposed to have a zero coefficient in the regression analysis. This argument is confirmed by the estimates of the natural logarithm of total assets, which are insignificant in all three regression specifications. Finally, when comparing the results based on the entire sample and on the middle half, we see that there is only a slight difference in their explanatory powers. Apparently, the difference-in-difference estimation, which takes inherent differences in treatment and control groups into consideration, produces accurate results even when we do not control for entities size. All in all, the robustness check verifies that the conclusions derived from the entire sample are not biased due to differences in firms’ characteristics.

6.2 Annual change in distance-to-default

As already discussed, one of the main advantages of using the distance-to-default as a bankruptcy risk measure is that we can calculate its changes in time. Therefore, the current robustness check aims to verify the empty creditor problem by investigating the annual change in the default measure, which is plotted in Figure 3. The main insight of the graph is that the bankruptcy risk doesn’t alter in the first three years of the sample period (i.e. in 2004, 2005 and 2006), then it increases significantly in the next two, and improves again in 2009 and 2010. Having said that, the financial crisis in the robustness analysis is defined as year 2007 and 2008, whereas the last two years of the sample are marked as a recovery period. In contrast, the analysis presented in Section 3.1 (Financial Crisis) does not exclude year 2010 in the regression, and the crisis is assumed to cover 2009 as well. By defining three types of periods the hypothesis of the financial crisis as a trigger of the empty creditor problem can be tested against two other phases: before and post crisis.

In order to examine whether the change in distance-to-default due to the crisis is significantly different in both groups the following regression equation is applied:

∆𝐷𝐷_𝑖.𝑡 = 𝛼_𝑖,𝑡+ 𝐶𝐷𝑆_𝑖∗ 𝐹𝑖𝑛𝑎𝑛𝑐𝑖𝑎𝑙 𝑐𝑟𝑖𝑠𝑖𝑠_𝑡+ 𝐶𝐷𝑆_𝑖 + 𝐹𝑖𝑛𝑎𝑐𝑖𝑎𝑙 𝑐𝑟𝑖𝑠𝑖𝑠_𝑡+ ∆𝑉𝑜𝑙𝑎𝑡𝑖𝑙𝑖𝑡𝑦_𝑖,𝑡 + ∆𝐸𝑥𝑐𝑒𝑠𝑠 𝑟𝑒𝑡𝑢𝑟𝑛_𝑖,𝑡+ ∆𝐿𝑒𝑣𝑒𝑟𝑎𝑔𝑒 _𝑖,𝑡 + ∆ln (𝐴𝑠𝑠𝑒𝑡𝑠)_𝑖.𝑡+ 𝑢_𝑖,𝑡 (7)

Does the trading of CDS increase firm’s bankruptcy risk through the empty creditor problem?