The effect of credit rating downgrades on CDS trading : an event study analysis of the current CDS market

The effect of credit rating downgrades on CDS trading: An

event study analysis of the current CDS market

Name: Frans IJkelenstam

Student Number: 10650067

Program: Master Finance

Track: Quantitative Finance

Supervisor: dr. T. Ladika

Date: 01-07-2018

Abstract

This paper analyzes the effect of credit rating downgrades on credit default swap spreads and volatility, estimated by a GARCH model. If banks use CDS in order to hedge credit risk, prevent unfavorable regulation (regulatory arbitrage), or to speculate on default, trading activity in CDS should increase following a credit rating downgrade. The effect should be larger for speculative graded firms. In line with previous research, evidence was found for the existence of cumulative abnormal spreads before the date of the downgrade. The effect is larger for speculative firms. Only limited evidence was found for abnormal spread volatility. This means that no conclusions can be drawn regarding the effect of rating downgrades on

Statement of Originality This document is written by Frans IJkelenstam, who declares to take full responsibility for the contents of this document. I declare that the text and the work presented in this document are 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.

Table of Contents

1. Introduction ... 3 2.0 Literature review ... 8 2.1 The CDS market ... 8 2.2 Empirical research on CDS pricing ... 9 2.3 CDS trading motives ... 11 2.4 Credit ratings and CDS ... 12 2.5 The volatility of CDS spreads and trading ... 15 2.6 Objective in comparison to existing papers ... 15 3.0 Methodology ... 16 3.1 Hypothesis testing ... 16 3.2 The Event Study Methodology ... 18 3.2.1 Setting up the model and the event window ... 19 3.2.2 Creating a benchmark for normal behavior ... 20 3.2.3 Calculating abnormal spreads and (average) cumulative abnormal spreads ... 21 3.2.4 Hypothesis Testing ... 21 3.3 Stochastic CDS Volatility ... 22 4.0 Data ... 23 4.1 Credit default swap spreads ... 23 4.2 Credit ratings ... 24 4.3. Stochastic CDS volatility ... 25 5.0. Results ... 28 5.1 Hypothesis testing ... 28 5.1.1 Hypothesis 1: the effect of a downgrade on CDS spreads ... 29 5.1.2 Hypothesis 2: the effect of a downgrade on CDS volatility ... 31 5.1.3 Hypothesis 3: the distinction between speculative and investment grade firms .. 34 5.1.4 Hypothesis 4: the effect of an upgrade on CDS spreads ... 35 5.2 Multivariate Results ... 37 6.0 Conclusion ... 40 6.1 Summary of findings ... 40 6.2 Limitations and improvements ... 41

1. Introduction

The credit default swap (CDS) and the market for these credit derivatives has only recently been introduced. Since its introduction in the early 90’s by the investment bank J,P. Morgan, CDSs have become increasingly important in the world of finance, with the CDS market experiencing exponential growth. As of June 30, 2008, the CDS market amounted to a total of $57 trillion in notionals outstanding (Arora, Gandhi & Longstaff, 2012). In a CDS contract, the buyer insures himself against credit default of the reference company. The seller of the contract takes over the credit risk by ensuring to pay the protection buyer the face value of the underlying bond in case of default. Effectively, a CDS contract thus serves as insurance against default of the reference company. Credit risk can thus be hedged using a CDS contract. The price of the contract therefore increases when credit becomes more risky. Research on the pricing determinants of CDS contracts is abundant. However, despite the large growth and size of the CDS market, research on what motivates CDS trading is limited. One of the reasons for this is that, although efforts have been made to standardize CDS contracts, CDSs are normally traded over the counter. This makes it harder for researchers to analyze the market structure. Where most of the existing literature focuses on the hedging motive of CDS trading (see e.g. Duffee & Zhou, 2001), some more recent articles point to other underlying trading reasons. Yorulmazer (2013) notes that 95% of current trading is executed by large institutional investors. These institutions are normally very well hedged already, indicating that there may be other reasons behind CDS trading. One of these reasons may be to reduce the amount of regulatory capital the institution needs to hold, so it can make additional loans. By hedging the credit risk of a risky loan with a CDS contract, the loan is treated as risk-free under the Basel accords. However, in case of a positive joint default probability of the reference firm and the risk seller, the combination of a bond and CDS may not be entirely risk free (Arora et al., 2012). Making use of such flaws is formally known as regulatory arbitrage and Yorulmazer (2013) argues that this is one of the foremost reasons of current CDS trading by institutional investors, and that it may have systemic risk consequences It remains open for debate what causes CDS trading. Peltonen, Scheicher and Vuillemey (2014) perform research on the structure of the CDS market. One of their findings suggests that CDS volatility is a much more important predictor of the size and

trading level of the CDS market than absolute levels of spreads. A larger number of traders and higher activity is documented for reference entities whose potential fluctuations in creditworthiness and systematic risk levels are larger. During periods of higher trading activity, CDS spreads become more volatile. However, it is still unclear what causes CDS spreads to become more volatile. The authors’ arguments include hedging, regulatory arbitrage, and speculation on default. Furthermore, the trading activity should in theory increase with deteriorating credit quality. If the credit quality of a reference company, roughly approximated by its credit rating, increases, then its credit becomes more risky. Traders holding its credit then have more reasons to trade its CDS for various reasons. First, hedging becomes more relevant when credit risk is more volatile. Second, when considering regulatory arbitrage, banks are more likely to successfully reduce their value-at-risk, and thus their capital charge, when the more volatile credit exposures are mitigated using CDSs. Third, speculating on default becomes more convenient when credit becomes more volatile, because more volatile credit usually results in a larger probability of default. Above reasons could lead to higher CDS trading activity following a rating downgrade. However, in the current-day CDS market, this may not be obvious. Typical market making by banks is performed by buying CDSs and selling them to an end user. If the market maker cannot find such an end user, it may choose not to trade the CDS in the first place. The inability to offset the CDS exposure can increase the amount of regulatory capital the bank has to hold. Choosing not to make the market could lead to a reduction in the overall size and trading activity of the market. A large drawback of the article of Peltonen et al., (2014) and other articles related to CDS trading activity is that CDS volatility is measured unconditionally. In other words, the value of their volatility variable is not changing over time. Castellano & D’Eclassia (2013) argue that the assumption of constant CDS volatility over time is invalid, because volatility clustering of CDS spreads is existent. They show that there are certain periods of relatively high and low CDS volatility. This means that the accuracy of the CDS volatility estimates that are oftentimes used in research articles could possibly be improved by using a different volatility measure. One that takes into account the existence of volatility clustering. Doing so would provide us with a more profound understanding of the determinants of CDS trading activity.

Castellano & D’Ecclasia (2013) tackle the problem of CDS spread volatility clustering by fitting a GARCH(1,1) model, established by Bollerslev (1986) on each time series of CDS data in their sample. The model provides a way of using shocks in CDS spreads, i.e. large day-to-day differences, to update volatility daily. Doing so yields a time series for volatility that can account for volatility clustering. With that benefit, this volatility measure can be a more accurate measure. Its flexibility with respect to static measures which usually only adjust each month, quarter or year is a benefit because the model can interpret signals from the market and exploit new information incorporated in CDS quotes immediately. The aim of this thesis is to investigate whether deteriorating credit quality indeed causes CDSs to become more actively traded. Doing so can increase our understanding of what causes current day CDS market activity. The research is conducted by means of an event study. First, I test whether CDS spreads actually increase after a credit rating downgrade, so that it can be verified whether credit ratings are a good proxy for the creditworthiness of a credit company. Next, I use the same analysis to test whether CDS volatility increases following a credit rating downgrade. CDS volatility here is measured per reference firm using a GARCH(1,1) model to account for stochastic volatility. This should increase the accuracy of the results. The following research question is centralized: “Does CDS trading, approximated by CDS volatility following a GARCH(1,1) process, increase following a credit rating downgrade?” To answer this question, data from reference firm’s CDS spreads will be collected, along with ratings from the S&P 500. By the means of an event study, abnormal spread and abnormal spread volatility around credit rating changes will be estimated. Research on credit ratings and the stock market is abundant. Also, some research has been conducted to estimate the effect of credit ratings on CDS spreads (Hull et al., 2004 ; Norden & Weber, 2004, to cite a couple of relevant papers). However, to my knowledge, no research has been done on the effect on the volatility of CDS spreads. By examining this relation, our knowledge of current-day CDS trading activity can be extended. Thus, the main goal of this thesis lies in estimating the effect of a credit rating downgrade on CDS volatility. It is expected that a rating downgrade causes spread

volatility to go up, which is interpreted as an increase in trading activity. The increase in trading activity can happen through various channels, including hedging, regulatory arbitrage and speculation. This thesis does not aim to differentiate between these reasons. That is out of the scope of the research. Separate tests will be performed for investment graded firms (S&P rating BBB- or higher) and speculative firms (S&P rating BB+ or lower). It will be interesting to see if the effect of a downgrade on CDS volatility is the same for these two types of firms. If CDS trading is indeed conducted to hedge credit risk, to increase risk via regulatory arbitrage or to speculate on a potential default, the expectation is that the effect of a downgrade on CDS volatility is larger for speculative firms. Four hypotheses have been formulated that are correlated with the central research question. The main hypothesis, which concerns the effect of a rating downgrade on CDS volatility, has not been confirmed. Only limited evidence was found for an increase in CDS volatility around the rate of the downgrade. The effect amounts to 4.68 percentage points yearly volatility in the window 60 days before the downgrade. This effect is not found to be statistically significantly. Therefore, no real conclusions can be drawn as to what increases CDS trading activity. The other three hypotheses have been confirmed. First, consistent with previous research (e.g. Hull et al., 2004; Norden and Weber, 2004), evidence for an increase in spread levels prior to the rating downgrade was found. On average, spreads increase with 58 bps in the 60 days prior to a rating downgrade. Second, this effect is substantially larger for speculative graded firms, who show an increase of 121 bps in the same window, compared to investment grade firms, whose increase amounts to 13 bps. Lastly, the analysis was repeated for rating upgrades, showing evidence for a decrease in spread levels after a rating upgrade. The documented effect was smaller in absolute sense in all test windows when compared to the effect of a rating downgrade. The rest of the thesis is organized as follows. Chapter 2 contains a comprehensive overview of literature that is related to this research, and it covers the objective of this thesis compared to existing literature. In chapter 3, the methodology section, four different hypotheses are formulated that are correlated with the research question as stated above. The necessary methods to test these hypotheses will be derived thoroughly. Chapter 4 contains a description of the data. Chapter 5 shows the results of

the tests of the hypotheses and in chapter 6 conclusions from these results are summarized along with the limitations of this thesis.

2.0 Literature review

This part of the paper covers the related literature and aims at providing a comprehensive overview of empirical papers covering credit default swaps. First, existing literature on the pricing of CDSs will be discussed. Secondly, the current functioning of the CDS market will be discussed, shedding light on a possible shift in the motivation of CDS trading. Thereafter, literature of credit ratings and CDSs will be discussed. Then, current day CDS trading volatility will be discussed, and the motivation for focusing on CDS volatility will be explained. Lastly, the objective of the thesis in comparison to existing literature will be discussed. 2.1 The CDS market A credit default swap is in a way an insurance contract. One company, the protection buyer, pays a risk premium, the CDS spread, to shift the credit risk of the reference entity to the protection seller. The notional principal is the amount that is insured. If the reference company experiences a credit event, the protection seller has to take over the defaulted loan, whereas the protection buyer receives the notional principal. Examples of such credit events are default, debt restructuring or non-paying. CDSs are thus a tool to shift credit risk. However, one does not necessarily need to own the underlying debt to enter a CDS contract. Therefore, CDSs can also be used to speculate on the default of the reference entity. In exchange for credit default protection, the providers of such protection require a recurring payment, also known as the CDS-spread. The relative magnitude of the CDS-spread naturally depends on the creditworthiness of the reference company - an increased probability of default results in a higher required premium by the protection provider. CDS spreads are therefore often used in literature to measure credit risk. CDS contracts are traded over the counter (OTC). This means that the contracts are privately negotiated between the two parties involved. Most contracts are thus not sold on an exchange. The reasoning for this is that each CDS contract has its unique features, which are tailor made to the protection buyer’s wishes. Because the contracts are sold over the counter, there is a chance that when a credit event happens, the

protection seller is not able to repay the notional principal. This form of risk is called counterparty risk. After the default of the banks Bear Sterns and The Lehman Brothers in 2008, the market started caring more about counterparty risk. The thought of ‘too big to fail ‘ banks was no longer valid, and a defaulting counterparty suddenly became a real possibility (Arora et al., 2012). Several authors, among who Arora et al. (2012) found significant evidence for the presence of counterparty credit risk in the price of a CDS contract. To mitigate the effect of counterparty risk, regulators have opted to reduce risk through central clearing. This means that the contract is split up into two contracts with a mediating party, also called a central clearing counterparty (CCP). By doing so, counterparty risk can be reduced by netting trades, posting collateral and ensuring payment. According to the Bank of International Settlements, the use of CCPs has increased over the years (Woolridge, 2017). The CDS market has experienced explosive growth since its introduction by J.P. Morgan in the late 1990’s. As mentioned before, its peak was in 2008 with $57 trillion in notionals outstanding. Since then the market has shrunk to $11.5 trillion outstanding in 2017 (Bank of International Settlements, 2018). These numbers seem extremely large, but they are somewhat misleading. These amounts refer to the total notionals from all contracts outstanding, without netting. When looking at the net notionals outstanding, which is the maximum amount that is due after a credit event, the amount outstanding equals $550 million (Bank of International Settlements, 2018). 2.2 Empirical research on CDS pricing Collin-Dufresne et al. (2001) examine how credit spread changes are influenced by several proxies for changes in default probability. Intuitively, any variable that increases the probability of default should have a positive relation with credit spreads. Models based on this insight are known as ‘structural models’ of default that use this intuition, build on the theory of Merton (1974). The key assumption in structural models is that a firm enters default when firm value falls below a certain threshold. This threshold is not static, but a function of the firm’s outstanding debt. Using a structural model approach, Collin-Dufresne et al. (2001) find that variables that both in theory and in proposed literature should determine credit spread changes, including changes in the spot rate,

variation. Furthermore, they also find that treasury yield changes can explain a substantial fraction of the variation in bond yield changes. The same does not hold when running the same analysis using credit spread changes. When analyzing the residuals, they conclude that a significant part of them is contained by a systemic factor, which is not included in the by theory proposed explanatory variables. The authors conclude that our knowledge of yield changes is larger than our understanding of credit spread changes. The results of Collin-Dufresne et al. (2001) pose challenges to modeling credit spread changes using structural models. Avramov, Jostova and Philipov (2007) argue that we do not necessarily need to discard the use of structural models in the empirical analysis of credit risk changes because of these results. They use a contingent-claims framework that assumes both equity and debt to be an option on a company’s value. This indicates that variables affecting company value should have an effect on credit spreads through default probabilities and default recovery rates. Using this approach, they were able to explain 67% of the CDS-spread variation of lower level bonds. However, their results were not able to explain a similar amount of variation in higher level bonds. This finding is in contrast with the findings of Galil et al. (2014), whose model better explains the credit spreads of investment-grade firms than those of speculative firms. The use of different sample periods may account for their different findings. Ericsson, Jacobs and Oviedo (2009) also investigate the relationship between CDS-spreads and theoretical determinants of credit risk. In contrast with Collin-Dufresne et al. (2001), they do find that volatility, leverage, and the risk-free rate explain a substantial and statistically significant part of the variation in CDS-spreads. Similarly, Avramov et al. (2007) also reject the hypothesis proposed by Dufresne et al. (2001) that the residuals have a common factor. Galil et al. (2014) find, in contrast to earlier studies, that both firm-specific and market variables account for a significant part in the variation of credit spreads. They use stock return and changes in stock return volatility, which are both firm-specific factors, and the change in median CDS spread in the rating class, which is a market factor. They find that these factors can explain changes in credit risk using their dataset of 718 US firms in the period 2002-2013. Zhang, Zhou and Zhu (2009) use a different approach than the often used

structural model approach. They estimate volatility and jump risk measures using high-frequency data on equity returns. On its own, these two measures account for respectively 48% and 19% of the credit risk variation. After adding firm-specific and market variables to the equation, the results remain in tact and the explained variation increases to 73%. 2.3 CDS trading motives One of the main purposes of entering a CDS contract is risk hedging. When a certain entity is long on a bond, it faces the risk that the counterparty will not be able to pay back the face value of that bond. When the same party enters a long CDS contract with the reference company in the contract being the same entity as the bond seller, the default risk shifts from the protection buyer, who is long the CDS contact, to the protection seller, which is short the contract. This feature of CDS contracts provides entities with a way to hedge default risk, and theoretically it is the primary role of entering a CDS contract. Existing literature focuses on the hedging role of CDS contracts. However, 95% of all trading activity in CDSs is executed between financial institutions, who may use CDSs for other purposes, such as regulatory arbitrage, instead (Yorulmazer, 2013). Regulatory arbitrage is a practice in which firms use a set of tactics to make use of flaws in regulatory systems in order to prevent facing unfavorable regulation. An example of how firms may make use of such tactics, is by trying to lower their capital requirements, so that they can increase investment. The Basel capital regulations state that when a bank or other financial institution uses a CDS to cover the credit risk of a risky investment, it may hold less capital as a safety net against that same investment. CDS contracts could be used by financial institutions to lower capital to dangerously low levels with consequences for systemic risk (Yorumazer, 2013). Especially in distress periods, banks’ cost of capital may rise to levels at which they are forced to pass on good investment projects due to capital requirements. When this happens, financial institutions may enter in CDS contracts to lower capital requirements which allows them to enter good investment projects that would have been disregarded otherwise. In the presence of counterparty default risks, if this default risk is correlated with the probability of failure of the project that is hedged using the CDS from that same counterparty, the CDS does not provide full insurance. This means that the CDS contract

only partially hedges an investment project. Therefore, regulatory arbitrage may have adverse systemic risk consequences. It is surprising that although 95% of trading activity is executed between financial institutions, most literature still focuses on the hedging motive of CDS contracts. Financial institutions usually are well diversified already, indicating that this large portion of trading activity might be for regulatory arbitrage purposes, or other reasons besides hedging. Oehmke & Zawadowski (2017) argue that although the CDS market has grown from a niche product to a well developed venue for credit risk transfer, relatively little is known about what motivates CDS trading. They use disaggregated CDS data on net notional amounts and trading volume of individual entities. Using these data, they find that firms with more outstanding bonds also trade CDSs more frequently, and have larger net CDS positions. This proves that firms may use CDSs to hedge their credit risk arising from bond positions. Next, they find that companies who have a larger earnings forecast uncertainty, as measured by larger analyst forecast dispersion, also possess larger CDS notionals. This indicates that companies may use CDSs for speculation purposes. However, these results do not directly explain why investors trade in CDSs instead of in the underlying bond itself. The main conclusion is that the CDS market serves a standardization role. CDS markets are more attractive as opposed to bond markets when the underlying bonds have heterogeneous contractual terms. 2.4 Credit ratings and CDS This subsection contains content of some relevant literature regarding the impact of credit ratings on CDS spreads. Table one contains a description of the different classes of credit ratings that are provided by the S&P 500. The first column contains the grade. Apart from the “AAA” and the “CC” grades, each grade class can be modified into 3 grade classes by adding a plus of a minus sign to the rating. These signs are added to divide the classes further, in order to show the relative differences within each class. In total, this results in 20 different classes. The second column displays the rating’s corresponding category. The distinction between investment graded and speculative graded firms is made to form a rough cutoff proxy for firms that fall in a speculative or less speculative category with respect to credit quality. Note that banks can choose their own measure of risk and are not required to follow these categories in their risk measurement. In this paper, following Hull et al. (2004) and Norden and Weber (2004), it is hypothesized that

Table 1: S&P 500 ratings and their descriptions

This table shows the different rating classes of the S&P 500, and their corresponding category and description.

Source: “https://www.spratings.com/en_US/understanding-ratings”

Grade Grade Category Description

AAA Investment Grade "Extremely strong capacity to meet financial commitments" AA Investment Grade "Very strong capacity to meet financial commitments" A Investment Grade "Strong capacity to meet financial commitments, but somewhat susceptible to adverse economic conditions and changes in circumstances" BBB Investment Grade "Adequate capacity to meet financial commitments, but more subject to adverse economic conditions" BB Speculative Grade "Less vulnerable in the near-term but faces major ongoing uncertainties to adverse business, financial and economic conditions" B Speculative Grade "More vulnerable to adverse business, financial and economic conditions but currently has the capacity to meet financial commitments" CCC Speculative Grade "Currently vulnerable and dependent n favorable business, financial and economic conditions to meet financial commitments" CC Speculative Grade "Highly vulnerable; default has not occurred, but is expected to be a virtual certainty" firms that belong to the lower, i.e. speculative rating class, will experience a larger effect of a credit rating downgrade, both in terms of abnormal spread changes, and abnormal spread volatility changes. The reason for this is twofold. First, firms with a lower credit rating is that on average difference in the implied probability of default increases as the rating deteriorates (Norden & Weber, 2004). Second, as a firm approaches default, buying a CDS may become more beneficial. Parties seeking to hedge their risk get an increased incentive to do so. Institutions seeking to lower their capital requirements

through regulatory arbitrage get an increased incentive to hedge the underlying bond after a downgrade. Also, speculation may increase as a firm approaches default. Credit ratings provided by rating agencies play an important role in capital markets. They can mitigate the existing problem of asymmetric information between different participants in the capital market. Investors and analysts can take credit ratings as a valuable sign of creditworthiness. Hull et al.(2004) argue that this is holds for the creditworthiness of the reference company, not only of the underlying bond. The reason for this is that it is very uncommon that two different bonds that are issued by the same company have different credit ratings. When rating agencies announce a rating change they refer to the rating of the company, and not of an individual bond, most of the time (Hull et al., 2004). Therefore, in this paper, the focus is on the ratings of a company, not its individually issued bonds. There has been some literature on the effect of credit ratings on CDS spreads. Hull et al. (2004) and Norden and Weber (2004) conduct research on the effect of credit rating downgrades that are provided by three different organizations, namely Moody’s, Fitch and S&P 500, on CDS spreads. By the means of an event study, both articles find that the CDS market can anticipate credit events quicker that the stock market can. More specific: credit rating downgrades are all anticipated most of the time before the event itself actually happens. They find no evidence for a delayed effect by the means of post-announcement abnormal spreads. This evidence supports the theory that CDS markets are efficient in the sense that they already incorporate the latest information regarding credit risk in the prices for shifting credit risk. Some authors also investigate the effect of credit rating upgrades. Norden & Weber (2004) hypothesize an asymmetric market reaction to credit rating upgrades and downgrades. They thus expect that, in contrast to downgrades, upgrades cause no significant abnormal returns around the event date. Two possible explanations for this are biases in the processing of information (Dichev & Piotrowski, 2001), or a disciplinary effect on firm management (Vassalou & Xing, 2003). The dataset in Norden & Weber, however, is very limited with only 12 upgrades. Therefore, they cannot conclude anything with certainty. The effect is substantially smaller than for downgrades.

2.5 The volatility of CDS spreads and trading The introduction of credit derivatives has provided useful insights regarding a company’s creditworthiness. Obviously, the compensation required by a protection seller will increase when the creditworthiness of the reference company lowers. Typically, in CDS data, we see that not only the credit spreads, but also trading in credit derivatives increases when bad news regarding creditworthiness of a single entity or the financial system in general enters the market. This causes credit spreads to become more volatile in certain periods compared to others. Peltonen et al. (2014) find that CDS volatility is a much better predictor of the size and activity of a CDS market than the absolute level of spreads. The trading activity on CDS contracts is larger when the perceived creditworthiness is more volatile. Therefore, it makes sense to not focus research solely on what events determine credit spreads, but also on what events determine credit spread volatility. Castellano & D’Ecclesia (2013) do just that. They argue that the clustering of volatility in credit spreads around certain periods (e.g. crises) makes the assumption of constant CDS volatility, which is made in almost all other research papers performing event studies on credit spreads, invalid. In their paper, they use event study methodology combined with stochastic credit spread volatility provided by an exponential Generalized Autoregressive Conditional Heteroskedasticity (GARCH) model, established by Bollerslev (1986). They find that credit rating events have a significant effect on conditional credit spread volatility in their sample of 60 international firms during 2004-2009. However, both the small sample size and the limited sample period, which is centered around the financial crisis, cast some doubt on their results. Also, the authors only use the stochastic volatility to normalize each observation, whereas in this research, the volatility variable will be used as the independent variable, a proxy for trading activity. 2.6 Objective in comparison to existing papers As mentioned before, the pricing of CDS contracts has been studied extensively, but the activity of CDS trading has not. Peltonen et al. (2014) find that CDS volatility is a key indicator for the degree of activity of the CDS market. Volatility may increase with deteriorating credit quality and the authors name three possible reasons for this, namely hedging, regulatory arbitrage and speculation. However, in their article no actual

Also, they use an unconditional volatility measure that does not change over time. Castellano & D’Ecclasia (2013) argue that using an unconditional measure of volatility is unfavorable because of the clustering of volatility around certain periods. There has also been some research on the effect of deteriorating credit quality, indicated by a credit rating downgrade, on CDS spreads. Most articles, among which Hull et al. (2004) and Norden & Weber (2004) find that CDS spreads significantly increase prior to a rating downgrade. This is a pretty well established fact, indicating that rating changes are anticipated by the market by incorporating the most recent information into the CDS price. In this thesis, the event study methodology used in Hull et al., (2004) and Norden & Weber (2004) will be combined with the time-varying CDS volatility measure as in Castellano & D’Ecclasia (2013), based on the GARCH model of Bollerslev (1986). Combining these methods, the effect of deteriorating credit quality on CDS volatility will be estimated. By doing so, this thesis aims to fill part of the gap of information regarding what factors increase trading activity in CDS. Should CDS volatility increase significantly following a rating downgrade, this will be considered as evidence for an increase in trading activity in CDSs.

3.0 Methodology

This section describes the methodology used in this research to investigate whether credit rating downgrades cause CDS spreads and CDS volatility to increase. This will be done in the form of an event study. Therefore, event study methodology with respect to credit default swaps will be introduced first. The second part concerns the matter of CDS volatility. Most papers treat CDS volatility to be constant over time. However, the post-crisis CDS market tends to fluctuate in trading activity, making this assumption questionable. In this section, it will be clarified how the GARCH(1,1) model of Bollerslev (1986) can be used as a way to derive stochastic volatility. 3.1 Hypothesis testing Using these methods, 4 hypotheses will be tested. These hypotheses are based on the research question depicted in the introduction. The first hypothesis concerns the effect of deteriorating credit quality, indicated by a rating downgrade, on the spread level

around the date of the downgrade. This relation has been measured before by, amongst others, Hull et al. (2004) and Norden & Weber (2004). However, these studies date back to the pre-crisis period, and therefore it will be helpful to test them again in a post-crisis setting, since the CDS market has changed much in terms of size and activity. By doing so, it can also be confirmed whether or not credit ratings are still a good proxy for creditworthiness. Before the crisis this was more clear, but now this may need renewed verification. The expected effect, which is measured in abnormal spreads, is expected to occur before the date of the downgrade, because the market should already incorporate most information associated with the downgrade in the pricing of the CDS. The second hypothesis will be tested thereafter. It covers the effect of a rating downgrade on spread volatility, which has not yet been tested in empirical literature. Volatility is measured using a GARCH model, which accounts for volatility clustering of CDS spreads. Because in typical CDS data, volatility clustering is present (see Castellano & D’Ecclasia, 2013), this volatility measure is preferred over a static measure. It is expected that a rating downgrade causes CDS trading activity to increase through increased hedging, regulatory arbitrage and speculation. This is the main contribution of the thesis. The thesis does not aim to differentiate between the different channels through which trading activity increases. The third hypothesis covers the difference between the abnormal spreads in each test period between speculative graded firms and investment grade firms. It is expected that the effect is larger for speculative graded firms with respect to investment graded firms, for two reasons. First, a firm that has a lower initial credit rating relative to another firm may experience a larger increase in the probability of default after a rating downgrade, because the distance to default becomes smaller. Second, a lower credit rating increases the motives to buy the CDS for hedging, regulatory arbitrage, or speculation issues, increasing the demand for the CDS. The increase in demand may be reflected by a higher price, and thus a higher spread. The fourth and last hypothesis concerns the effect of a rating upgrade on absolute spread levels. The opposite effect of hypothesis 1 is expected. However, the expected absolute effect of a rating upgrade is expect to be smaller compared to a rating downgrade. This can be explained by informational processing biases and disciplinary effects on firm management. A negative effect on spreads is still expected because a rating upgrade reduces the demand to hold the spread, because it can reduce the need to

trade the CDS. Hedging, regulatory arbitrage and speculation are all reasons for a decreasing demand in the CDS in case of a rating upgrade. Norden & Weber (2004) investigate the relation between rating upgrades and spread levels. However, their sample contains only 12 upgrade, which is very little. They therefore cannot conclude anything. The four hypotheses that have just been described are depicted below: Hypothesis 1: Credit rating downgrades cause significant cumulative abnormal spreads changes before the date of the downgrade Hypothesis 2: Credit rating downgrades cause significant abnormal spread volatility before the date of the downgrade Hypothesis 3: If cumulative abnormal spread changes are present, they are larger for speculative graded firms compared to investment graded firms. Hypothesis 4: Credit rating upgrades cause significant cumulative abnormal spreads around the date of the upgrade. The effect is smaller compared to downgrades. 3.2 The Event Study Methodology Standard event study methodology tests are tests of market efficiency before, during and after a certain event. Should the event of interest convey new information to market participants, then prices should react after the event date. In the opposite case, the event of interest conveys no new information at all. In this case, market participants already fully anticipated the event prior to the event date. Concerning credit default swaps, this would suggest that in a fully efficient market, if ratings contain no new information, a credit rating downgrade of a firm is already fully anticipated by the market, so that no significant abnormal spread changes are realized before the event date itself. In event study methods and literature, the term abnormal means compared with a certain standard, or period of normal behavior. In this section, the methods of calculating normal and abnormal spreads will be described in detail. The event study methodology

used in this paper are oftentimes used in existing literature, and it is explained in detail in the syllabus of De Goeij & De Jong (2011). 3.2.1 Setting up the model and the event window Starting from raw daily CDS quotes, the first step is to simply calculate daily changes in CDS quotes: 𝐶𝐷𝑆𝑐ℎ𝑎𝑛𝑔𝑒_!,! = 𝐶𝐷𝑆_!,!− 𝐶𝐷𝑆_!,!!! (1) The CDS-change variable calculated in equation (1) will be used for calculating both the benchmark and the abnormal behavior of CDS quotes per company. However, prior to those calculations it is necessary to set up the event window. Consider a time interval [n1, n2] lasting from n1 business days after the event to n2 business days after the event.

Note that both n1 and n2 can be positive and negative. Negative values therefore indicate

that the interval happens before the downgrade. For example, the interval [-60, -31] takes place from 60 days before the event to 31 day before the event. The interval [2, 10] happens between 2 and 10 days after the event. The event window is chosen to be 110 days. The test period start 60 trading days before the event. This number chosen for two reasons. First, it is not too large to cause a large amount of events to overlap. Second, in Hull et al. (2004) and Norden and Weber (2004), the abnormal returns are largest up to 60 trading days prior to the event itself. Figure 1 provides a graphical illustration of the event windows for every separate event in the sample. Figure 1: Event Window This figure shows the event window. The event window can be divided further into the estimation period and the test period. Moreover, the test period can again be further divided into various subperiods.

t=-100

Estimation Period

Event Window

Test Period

t=-60

t=-30

t=-1

t=0

t=1

t=10

As can be seen from figure 1, we can separate the whole event window in the estimation period and the test periods, which can further be divided into 7 subperiods Ij, j = {1, … 7}. Note again that the event itself takes place at time t=0. If a credit rating downgrade is already incorporated in the price of the CDS prior to the downgrade itself, it is expected that the abnormal spread changes are largest in the earlier periods. If this not the case, i.e. the downgrade is not anticipated by the market, the reaction should be largest around the event itself, e.g. in the interval [-1, 1]. Alternatively, if there is a delayed response, there should be an effect present in the period [2, 10]. As can be seen from the figure, the total event window lasts 110 trading days. This corresponds to around 5½ months. Some credit rating downgrades for the same reference entity may follow each other quickly when a firm experiences a period of credit risk increase. This may cause different event windows to overlap. Therefore, following Norden & Weber (2004), Hull et al. (2004) and Castellano & D’Ecclasia (2013), all events that overlap each other are eliminated from the sample once the abnormal returns are calculated. 3.2.2 Creating a benchmark for normal behavior To calculate abnormal spread changes, we must first set a benchmark. Any deviations from this benchmark will be considered as ‘abnormal’. Following Castellano & D’Ecclesia (2013), in this paper, the benchmark for normal behavior is chosen to be the mean relative CDS change during the estimation period: 𝐸[𝑆_!,!− 𝑆_!,!!!] = 𝐶𝐷𝑆𝑐ℎ𝑎𝑛𝑔𝑒!,! !!,!" !!,!" 𝑛_!,!"− 𝑛_!,!" t = {n1,EP, n2,EP} (2)

Where 𝐸[𝑆_!,!] represents the normal spread change of firm i at period t, and [n1,EP, n2,EP]

represents the time interval of the estimation period prior to the event window. Intuitively, one would expect the normal spread change to be zero on average if the probability of an up- or downward movement in CDS spreads were to be equal. However, spreads tend to move in the same direction across firms within an economic cycle. Therefore, a mean change is more appropriate as benchmark than assuming a zero change on average.

3.2.3 Calculating abnormal spreads and (average) cumulative abnormal spreads In the end, we are interested whether the rating downgrade causes spreads to change with respect to the benchmark we have just set. Again, the effect could be present prior to, around, or after the credit event happens. In order to investigate this effect we calculate the abnormal spread. The abnormal spread is calculated by subtracting the mean changes in the estimation period from the relative daily CDS changes during the test period. There is no need to adjust for any index or market adjustments, because CDS spread changes are assumed to be 0 on average over the long run. Instead, as described in the previous section, the spreads are adjusted for the mean change over the estimation period to account for the current normal behavior. 𝐴𝑆_!,! = (𝐶𝐷𝑆𝑐ℎ𝑎𝑛𝑔𝑒_!,!− 𝐸[𝑆_!,! − 𝑆_!,!!!]) (3) Now that the abnormal spreads are defined, we sum the abnormal spreads over the different subperiods in the test period to obtain the cumulative abnormal spreads. This variable measures the total effect of the credit event during the different test periods Ij. It is of particular interest because it helps in answering the question whether the market cares about credit ratings: 𝐶𝐴𝑆_! 𝐼_! = 𝐴𝑆_!,! !∈!! (4) Finally, the average cumulative abnormal spread can be derived by summing all the different cumulative abnormal spreads during each subperiod Ij. This cross-sectional average can be seen as the average effect of a credit rating downgrade on the relative spread change during the subperiod: 𝐴𝐶𝐴𝑆 𝐼_! = 𝐶𝐴𝑆_! 𝐼_! !∈!! (5) 3.2.4 Hypothesis Testing After constructing the cumulative abnormal returns per event, we are interested if these cumulative abnormal returns are significantly different from 0 on average. This will be done by the means of a simple t-test:

𝐻_!: 𝛽_!"# = 0 𝐻_!: 𝛽_!"# ≠ 0 𝑡 = 𝛽!"# 𝑆𝐸!"# (6) Since event study methodology is applied in this thesis, there will be no regression analysis. This means that none of the classical ordinary least squares assumptions matter (De Goeij & De Jong, 2011). Therefore, no tests for autocorrelation nor heteroskedasticity will be provided. 3.3 Stochastic CDS Volatility The paper of Castelano and D’Ecclesia (2013) shows some evidence that credit rating announcements may not be reflected by changes in CDS spreads. This is especially true during crisis periods. The fact that the CDS market does not react to a credit rating change may imply numerous things. The first possible explanation would be that markets do not care about or trust the ratings provided by rating agencies. Indeed, the recent past has shown us that ratings provided by the big agencies are often inaccurate or even biased (Poon, 2003). If investors collectively believe that the ratings are of little value, then markets do not react to ratings changes because they convey little or no information. Another possible explanation could be that, since most trading activity is done between financial institutions, these institutions are so well hedged that rebalancing after a credit event is somewhat unnecessary. The assumption of constant CDS volatility may be invalid due to abovementioned reasons. Volatility may differ across different time periods or per company. Therefore, stochastic volatility of CDS quoted will be estimated using a GARCH(1,1) procedure. The family of Generalized Autoregressive Conditional Heteroskedasticity (GARCH) models was founded by Bollerslev (1986). GARCH models have been found to provide adequate estimates of stock volatility over time. The main feature of this class of models is that it can account for volatility clustering in asset returns. A large return shock will cause the estimated volatility for the following day to increase, after which it reverts gradually back to its unconditional value. The model will be estimated for each entity separately using statistical programming software and looks as follows:

𝜎_!,!! _{= 𝜔 + 𝛼 ∗ %𝐶𝐷𝑆 𝑐ℎ𝑎𝑛𝑔𝑒} !,! + 𝛽 ∗ 𝜎!,!!!! 𝜔, 𝛼, 𝛽 > 0 (7) As can be seen from the model, volatility today depends on the volatility of yesterday plus a correction which depends on the return, or in this case CDS quote change, today. The coefficients in equation (6) will be estimated using programming software and used to create a time series for volatility. When taking expectations left and right, one can calculate the estimated unconditional variance: 𝜎! = 𝜔 1 − 𝛼 − 𝛽 (8) Equation (7) shows how the coefficients can be used in order to estimate the unconditional, long run variance of CDS quotes. Notice that the time subscript drops out of the equation since the unconditional variance is constant per firm over time. It is expected that a credit rating downgrade causes credit volatility to increase, leading to a deviation from its unconditional value, after which it gradually converges back to that value.

4.0 Data

This section describes the data that is used for this research. The various sources from which the data have been retrieved will be described. Also, a brief description of how these different sources have been linked together will be provided. Next, the use and reasoning of the different variables that have been constructed by the use of these data will be described. 4.1 Credit default swap spreads One of the main variables of this research are CDS spreads. The spreads that are used for this research are 5-year daily CDS spreads for 558 reference firms. Like in almost any other research paper, the 5-year spreads are chosen because these spreads are by far the most liquid and frequently traded contracts. The spreads are retrieved from Datastream and cover the period from November 2010 until the present. The initial sample consists of 631,590 spreads for North-American reference firms. A liquidity problem arises when using daily CDS-spread data, which is especially

a day-by-day basis. However, taking monthly data is not an option when conducting an event study. Therefore, when conducting the event study on CDS volatility, the most illiquid CDS spreads are eliminated from the sample. This is done by deleting any reference firm that sees its daily CDS-spread unchanged with respect to the day before over 50% of the time in the sample. This results in a loss of 50 observations. Furthermore, I decided not to eliminate firms from the sample based on the level of the CDS spreads or other variables being seemingly very small or very large. This is consistent with Norden & Weber (2004) and Hull et al. (2004). Since the sample only contains large and established firms that are listed on important and well-known indices, it should not affect the dataset nor the results in a significant manner. 4.2 Credit ratings Another important variable of this research are credit ratings. These ratings were hand-collected from the website of the S&P 500. The ratings of the S&P 500 range from “AAA” to “CC”, amounting to a total of 20 different ratings before a company has a default grading: “D” or “SD”. Companies with a rating of “NR” are non-rated and are deleted from the sample during trading days where they are non rated. For example: if company A has a credit rating of “BBB” during November 2010 until January 2013, after which it is non-rated, only the observations up until January 2013 are used in the sample. The data on these downgrades range from November 2010 until the present. This is convenient since this thesis tries to focus on the post-crisis role of the CDS market. The sample exists of a total of 631,590 daily spreads and corresponding credit ratings. Table 2 provides some descriptive statistics regarding these credit ratings. As expected, CDS spreads generally are inversely related to credit ratings. As we move down the table, creditworthiness as estimated by the S&P rating declines. We see that, apart from some notable differences, the mean of the CDS spread increases when the ratings become lower. These notable differences are AA+ and CCC+. For these ratings classes, the inverse relation is violated. One possible explanation could be that these firms only contain 2850 and 466 ratings observations respectively. Since most firms in the dataset have data for 7 years of trading days, these numbers are quite small. Therefore, we can consider this randomness and continue the analysis, since it should not influence the outcomes much.

Table 2: Spread statistics per rating class

This table shows summary statistics for CDS spreads over the sample period 2011-2018. Spreads are divided into rating classes. The ratings and the rating classes are obtained from the S&P 500 website. They range from AAA to CC, amounting to 20 classes total.

S&P Rating N Mean Min Max Median

AAA 8475 29,11259 11,14 55,31 29,59999 AA+ 2850 67,6309 25,90599 337,6 53,37 AA 12713 35,33007 9,92 108,03 33,34 AA- 16480 41,58907 14,22 200,5 35,06999 A+ 22630 43,96883 11,59 126,84 41,46 A 64918 52,54034 10,19 395,96 50,13349 A- 53183 60,76918 14,48 424,1819 53,15799 BBB+ 89255 89,15575 12,17 870,8799 78,31 BBB 118606 115,8088 14,92 916,7898 100,455 BBB- 80652 148,9447 25,32999 1036,78 127,531 BB+ 33829 206,4586 20,702 1181,88 181,76 BB 30402 219,0756 14,6 1160,29 186,45 BB- 28930 472,8836 26,25 6222,941 310,2798 B+ 22766 433,3435 26,25 6222,941 341,3474 B 22787 564,823 28,20999 6222,941 453,8599 B- 14437 1729,525 36,29999 13366,96 690,3499 CCC+ 5965 1139,386 216,69 11058,34 939,0098 CCC 1638 2103,678 618,24 11334,36 1687,74 CCC- 466 2227,839 694,7898 13366,96 2057,685 CC 608 1560,227 241,51 3312,94 1355,175 Total 631590 206,1101 9,92 13366,96 95,81 4.3. Stochastic CDS volatility Most authors do not account for differences in CDS volatility over time. Castellano & D’Eclassia (2013) argue that the assumption of constant CDS volatility over time is invalid, and therefore use a GARCH(1,1) model to estimate, for each reference entity, time-varying CDS volatility. In this article, this procedure is also done for a larger sample

and longer time period. To show that CDS volatility indeed fluctuates over time, and in a different way around event dates per firm, two graphical examples are included. Consider figure 2 and figure 3. The blue line in both graphs shows the estimates of the dynamic CDS volatility during the period of November 2010 until January 2018 for two different reference companies in the sample. What stands out in both graphs, is that the CDS volatility is high around the event date with respect to other periods. The fact that spreads are more volatile around credit rating downgrades may indicate that they are also traded more often around these periods.

Figure 2: the Stochastic CDS volatility for AK Steel Corp. This graph shows the credit default swap spread volatility for AK Steel Corp for the period of nov 2017 – jan 2018, estimated by a GARCH(1,1) model, over time. On the day of 12/20/2015, AK Steel Corp experienced a credit rating downgrade, represented by the red bar in the graph. Figure 3: the Stochastic CDS volatility for Enbridge Energy Partners. This graph shows the credit default swap spread volatility for Enbridge Energy Partners corp for the period of nov 2017 – jan 2018, estimated by a GARCH(1,1) model, over time. On the day of 6/19/2015, Enbridge Energy Partners experienced a credit rating downgrade, represented by the red bar in the graph.

5.0. Results

In this section, the necessary tests that are described extensively in the methodology section will be performed on the panel data to test the hypotheses that are also formulated in the methodology section. On the basis of these results, the hypotheses will either be confirmed or rejected. The obtained results will be discussed in terms of statistical significance and economical importance. Furthermore, the results will be compared, if possible, with the results of academic literature that is described in the methodology section. The tables that contain the results of the event study are set up in the same manner. Each table contains 3 panels. Panel A contains all companies in the sample. In panel B and panel C, the sample is divided into two subsamples according to the credit rating of the company. Panel B contains the event study results for the subsample of companies that have an “investment grade” credit rating. This means that the company receives a S&P 500 grade in the category “BBB-“ or higher. Panel C contains the results for the subsample of companies that belong to the “speculative grade” rating, which indicates that the company receives a S&P 500 grade in the category BB+ or lower. This distinction is made to test the third hypothesis. Furthermore, the cumulative abnormal spreads for all three different (sub)samples are tested according the method described in the methodology section. Five different test periods [n1, n2] are defined, namely [-60, -2], [-60, -31], [-30, -2]. [-1, 1] and [2, 10]. For every period, the coefficient, standard error and corresponding t-statistic and p-value, along with the amount of observations, are given. The results section is organized as follow: first, the four different hypotheses derived in the methodology section will be tested and discussed in section 5.1. Next, in section 5.2, I will try to go one step further. CDS spreads and volatility will be regressed on variables typically used structural models in order to see if the main results can be put in context. 5.1 Hypothesis testing In the methodology, four hypotheses have been formulated which will be tested in terms of statistical significance in this section. The main goal of the thesis is to estimate the post-crisis effect of a credit rating downgrade on the CDS spreads and CDS

spread volatility, as estimated by the GARCH equation. Doing this helps us to better understand the current CDS market. 5.1.1 Hypothesis 1: the effect of a downgrade on CDS spreads Hypothesis 1 concerns the effect of a credit rating downgrade on CDS spreads. It is hypothesized that the credit rating event is already (partially) incorporated into the price of the CDS before the event actually occurs. Table 3 contains regression results for the event study, estimating mean cumulative abnormal spread changes (CASCs) for all different test periods. Panel A contains the results for the whole sample. The first coefficient for the mean CASCs equals approximately 58.49. This coefficient indicates that over the whole sample of firms that experienced a credit rating change, CDS spreads changed on average with 58.49 basis points per year of the notional outstanding amount during the time window [-60, -2]. This increase amounts to 28.49% of the mean spread in the sample. Regarding panel A, it can be seen that for the time period [-60, -2], the CACSs amount to 58.49 basis points on average. Its coefficient is statistically significant at the 1% level. This can be interpreted as strong evidence that markets anticipated the credit rating change by incorporating (part of) the information associated with the credit rating change into the price of the CDS up to 60 days prior to the event date. When the full pre-event test window is divided into two sub windows, the evidence suggests that the market reaction is strongest in the window [-30, -2], where the mean CASC equals 84.52 basis points, compared to the window [-60, -31], where the mean CASC equals 30.91 basis points. This is also reflected by the t-statistic and corresponding p-value, which is significant at the 1% level in the window [-30, -2], but insignificant in the window [-60, -31]. Furthermore, no significant effect is found at the later test periods [-1, 1] and [2, 10]. The results found in panel A are in line with earlier literature. Hull et al. (2004), Norden and Weber (2004) and Castellano & D’Ecclassia (2013) all find significant (cumulative) abnormal spread changes in test windows prior to the actual rating change, but little or no similar evidence for post-event windows. Therefore, it can be concluded that the rating change itself is of little informational value to the market.

Table 3: Results of the event study for downgrades on spread changes This table shows the results of the conducted event study for all time periods. The results are divided into three panels: (A) All firms, (B) Investment grade firms and (C) Speculative firms. The cumulative abnormal spread change (CASC) is the sum of all abnormal spreads during the periods [n1, n2] measured in basis points. The t-statistic and corresponding p-value is given for the test whether these CAR’s are significantly different from 0. The significance levels are given by * and ** indicating statistical significance at the 5% and 1% level, respectively. Time periods [n1, n2] [-60, -2] [-60, -31] [-30, -2] [-1, 1] [2, 10] Panel A: All Firms Mean CASC 58,48836 30,91143 84,5178 30,47971 17,26198 Standard Error 17,90555 20,1759 29,06415 19,59696 17,10822 T-Value 3,27** 1,53 2,91** 1,56 1,01 p-value 0,001 0,127 0,004 0,122 0,314 N 187 187 187 187 187 Panel B: Investment Grade Firms Mean CASC 12,96715 14,74674 11,23522 4,470128 4,53478 Standard Error 3,802675 5,069476 5,671205 2,543477 2,03722 T-Value 3,41** 2,91** 1,98* 1,76 2,23* p-value 0,001 0,004 0,049 0,082 0,029 N 105 105 105 105 105 Panel C: Speculative Firms Mean CASC 121,3648 54,09499 182,2279 65,0494 30,13718 Standard Error 41,88192 48,64303 66,12165 45,39011 34,39911 T-Value 2,90** 1,11 2,76** 1,43 0,08 p-value 0,004 0,27 0,007 0,156 0,383 N 82 82 82 82 82

Panel B and C contain the results for the subsamples of investment grade firms and speculative firms respectively. These results are provided to distinguish between these types of firms, since rating downgrades between these different classes can have different impact in terms of spread changes and volatility. As can be seen, for the investment grade firms, the coefficient for the time windows [-60, -2] and [-60, -31] are significant at the 1% level, whereas the coefficients for [-60, -2] and [2, 10] are significant at the 5% level. For the speculative firms, the coefficients for [-60, -2] and [-30, -2] are significant on the 1% level. These results reconfirm that overall, the events are anticipated by the market prior to the event itself happening. This means that the first hypothesis is confirmed. 5.1.2 Hypothesis 2: the effect of a downgrade on CDS volatility The second hypothesis concerns daily credit default swap volatility, as estimated by the GARCH model. It is hypothesized here that a credit rating downgrade increases the conditional volatility. This is expected because a rating downgrade causes higher trading activity, which causes higher trading volatility. For example, hedging becomes more relevant when the credit becomes more risky. Furthermore, a financial institution may trade CDS more in the case of a rating downgrade because it lowers the amount of regulatory capital it has to hold. When it does not buy a CDS, the credit is not hedged in terms of credit risk and therefore the position becomes more risky. Its capital charges, which are based on the credit ratings, then rise. The effect becomes stronger as credit ratings become lower and in effect credit risk becomes more prominent. Lastly, also speculation on default becomes more relevant when the reference firm approaches default. Table 4 contains the results for this event study, calculating cumulative abnormal spread volatility changes (CASVCs). Note that the amount of observations is lower, because the most illiquid spreads have been dropped from the sample. This is done because the GARCH model does not perform well when no changes in spreads occur. Therefore, spreads with over fifty percent daily spread changes within the event window are dropped. This results in a total loss of 50 observations. Regarding panel A, it can be seen that the sample as a whole do not show consistent CASVCs over the test period. The coefficient of CASVC of the full pre-event window [-60, -2] is 0.295, which indicates that during this window, daily volatility

points. However, the coefficient is insignificant. Only one coefficient is significant at the 5% level, for the window [-60, -31], which equals 0.561. In yearly terms, this amounts to roughly 8.91 percentage points. The coefficient for the window [-1, -1] is not significant on the 5% level, but it is close. Especially with the relatively small sample size, this can also be regarded as limited evidence for an increase in volatility in the 3 days surrounding the downgrade. Next, the sample is again split up into investment grade and speculative firms. For investment grade firms, only the window [-1, 1] is significant at the 5% level. It can be seen that the speculative firms show some limited evidence for pre-event volatility changes, with the coefficients for the test windows [-60, -2] and [-60, -31] being significant at the 5% level. The estimated effect on volatility is fairly limited. There are several possible explanations for this. Firstly, the expected effect of an increase in trading activity through increased hedging, regulatory arbitrage and speculation might just not be present. Secondly, the reduction in the sample size due to liquidity problems eliminates a substantial fraction of the statistical power of the model. Thirdly, because of a lack of liquidity in the CDS data, the GARCH model may not always perform adequately. Lastly, typical market making is performed by financial institutions who buy a CDS and then sell it to an entity who seeks to hedge some form of credit risk. However, if the market maker cannot find a buyer, it becomes harder for the institution to offset their CDS exposure. In this case, following a credit rating downgrade, the institution may choose to not trade the CDS in the first place, because it would increase the amount of regulatory capital the institution has to hold, which is unfavorable. This is turn could lead to a reduction in the size and liquidity of the market. This is a trend which may have contributed to CDS dealers leaving the market and the before mentioned shrinkage of the CDS market in recent years. To conclude, some evidence in support of hypothesis 2 has been found. Especially the data for speculative firms seem to provide some evidence in favor of the hypothesis. However, the evidence is not strong enough to confirm the hypothesis with certainty. Therefore, no definite conclusion can be drawn from these results. This in turn means that on the basis of these results, it can not be said that CDS trading increases after a rating downgrade through increased hedging, regulatory arbitrage or speculation.

Table 4: Results of the event study for downgrades on spread volatility changes This table shows the results of the conducted event study for all time periods. The results are divided into three panels: (A) All firms, (B) Investment grade firms and (C) Speculative firms. The cumulative abnormal spread volatility change (CASVC) is the sum of all abnormal spread volatility changes during the periods [n1, n2] measured in daily percentage points. The t-statistic and corresponding p-value is given for the test whether these CAR’s are significantly different from 0. The significance levels are given by * and ** indicating statistical significance at the 5% and 1% level, respectively. Time periods [n1, n2] [-60, -2] [-60, -31] [-30, -2] [-1, 1] [2, 10] Panel A: All Firms Mean CASVC 0.2948425 0.5607739 0.0346096 0.276588 -0.1128256 Standard Error 0.1808568 0.2344595 0.2736537 0.1462786 0.1071595 T-Value 1.63 2.39* 0.13 1.90 -1.05 p-value 0.104 0.018 0.900 0.061 -.294 N 137 137 137 137 137 Panel B: Investment Grade Firms Mean CASVC 0.10068 0.3407438 -0.3390826 0.2583302 -0.11895 Standard Error 0.1588935 0.2213113 0.2236561 0.1238101 0.1217889 T-Value -0.63 1.54 -1.52 2.09* -0.98 p-value 0.529 0.127 0.133 0.040 0.332 N 93 93 93 93 93 Panel C: Speculative Firms Mean CASVC 0.8897568 1.010613 0.7740432 0.4192671 -0.1052637 Standard Error 0.4361413 0.502792 0.676900 0.350212 0.1880315 T-Value 2.04* 2.01* 1.14 1.20 -0.56 p-value 0.044 0.046 0.259 0.237 0.578 N 44 44 44 44 44

Table 5: Mean difference in CASC between investment grade and speculative firms This table shows the mean difference in cumulative abnormal spread change (ΔCASC) between investment grade and speculative during the periods [n1, n2], which is measured in basis points. The t-statistic and corresponding p-value is given for the test whether these CAR’s are significantly different from 0. The significance levels are given by * and ** indicating statistical significance at the 5% and 1% level, respectively. Time periods [n1, n2] [-60, -2] [-60, -31] [-30, -2] [-1, 1] [2, 10] Mean ΔCASC 108.3976 39.34825 170.9927 60.57927 25.6024 Standard Error 35.89769 41.02187 57.58784 39.44501 34.26101 T-Value 3.0196** 0.9592 2.9693** 1.5358 0.7473 p-value 0.0027 0.3387 0.0034 0.1263 0.4559 N 187 187 187 187 187 5.1.3 Hypothesis 3: the distinction between speculative and investment grade firms The third hypothesis concerns the difference between the cumulative abnormal spread changes of speculative and investment graded firms. As mentioned before, table 3 contains the results for the event study, estimating CASCs. It is hypothesized that following a credit rating downgrade, spreads for firms that are speculative graded react stronger than spreads for investment graded firms. As mentioned before, there are two reasons for this. The first reason is that the firms with a lower initial credit rating are expected to experience a larger difference in implied default probability after a rating downgrade compared to firms that are initially relatively higher rated. The second reason is that as a reference firm approaches default, its credit becomes riskier and thus there are more motives to purchase the CDS for hedging, regulatory arbitrage, or to speculate on default. Hedging increases because of the increased benefits of hedging. Regulatory arbitrage increases because when the credit of a more risky firm is hedged, this will more likely reduce the value-at-risk, and thus the capital charge a financial institution faces, compared to a when the credit of a less risky firm is hedged. This demand effect can generate a larger increase in spread levels for speculative firms, compared to investment grade firms. Table 5 contains a mean-comparison t-test for CASCs between investment grade and speculative firms, which are provided in table 3. As expected, all the coefficients are