An empirical analysis of the effects of trading CDS on the bond market

(1)

1

An Empirical Analysis of the Effects of Trading CDS

on the Bond Market *

Janine Liu

July 7, 2016

University of Amsterdam

Supervisor: Tomislav Ladika

Abstract

Using DTCC CDS weekly position data, this research examines the empirical effects of trading Credit Default Swaps (CDS) on the underlying bond pricing and trading volume, testing the liquidity-based model derived by Oehmke and Zawadowski (2015). Utilizing the IMM dates as an identification strategy, we find evidence that trading CDS is associated with a positive effect on daily bond returns, both statistically and economically significant. When running the same pooled regressions at weekly interval, the effect remains positive however statistical power weakens. Trading CDS is also found to be more beneficial to the prices of relatively more illiquid bonds than liquid ones in the same firm. On bond trading volume regressions, overall statistical power is weak, suggesting trading CDS does not have a strong effect on bond trading volume on weekly level.

*Author Janine Liu, with supervision from Tomislav Ladika. This research is a master thesis as part of the graduation program - MSc Business Economics, Finance track, in the academic year 2015 - 2016, at the University of Amsterdam, Amsterdam Business School.

(2)

2

Statement of Originality

This document is written by Janine Liu 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)

3

I. Introduction

The U.S. corporate bond market represents an important source of capital for firms, however it is not very well understood when compared to other financial markets, such as the equity market. First, the corporate bond market is an over-the-counter market with relatively low transparency. Second, many corporate bonds were bought shortly after issuance and held to maturity, leaving only a small portion of bonds available for trade (Alexander, Edwards, and Ferri (2000)). Therefore secondary bond market is often rather illiquid. Trading credit risks in the bond market hence is hindered by low transparency, low liquidity and high transaction costs. The introduction of Credit Default Swaps (CDS) provides an alternative opportunity for investors to trade credit risk. Compared to the bond market, the contractual nature of the CDS market makes it relatively easier to trade (Augustin et al (2014)). Investors intending to hedge or speculate credit risks can enter long or short a credit default swap contract relatively flexibly, so long as they find a counterparty who is willing to be on the other side of the contract. Therefore, there is no supply constraint in the CDS market compared to the bond market. In light of the close relation between these two markets, subsequent questions arise such as how are the CDS market and the cash bond market related? Does trading credit risk using CDS affect cash bond market in any way? Such as pricing, borrowing cost, liquidity, and market efficiency among others? Does trading CDS benefit or harm the referencing bond market on those parameters? These become interesting questions to answer considering the size and importance of the CDS market.

The objective of this research is to directly assess if trading CDS have any effects on bond pricing and trading volume, at the individual bond level. It is interesting to study the effects on pricing and trading volume, because bond pricing directly reflects the borrowing costs and trading volume is an important indicator for bond liquidity. We test the theories of the liquidity-based model

derived by Oehmke and Zawadowski (2015)1. The key of their model is based on the empirical

observation that CDS market features lower transaction costs than the bond market. Investors based on different liquidity needs, sort themselves into the bond or the CDS market. Investors that are with frequent liquidity shocks choose to trade in the CDS market because of the lower transaction costs, while buy-and-hold investors choose to remain in the bond market. This suggests that when CDS is traded, bond market might be affected in a few ways. First, investors previously holding a long position in the bond can switch to taking a short position in the CDS. This leads to less demand for

1_{Oehmke and Zawadowski (2015) Synthetic or Real? The equilibrium effects of credit default swaps on bond markets,}

(5)

2

bond and should have a downward pressure on bond prices. Second, investors previously holding a short position in the bond can take a long position in CDS instead. This leads to higher demand for bond and should have a positive effect on bond prices. These are the crowding out effects and together bond prices are expected to decline. Third, arbitragers can participate in the negative basis trades which involve holding both a long position in the underlying bond and the CDS. This is referred to as the basis-trade effect. If arbitragers do not have any funding constraints, the third effects is strong and overall trading CDS is expected to push up bond prices. However, if market funding is constrained, the third effect becomes weak and the overall effect of trading CDS on bond prices is expected to be negative. Trading CDS might also affect bonds of different liquidity differently. Relatively more liquid bonds are held more by investors with frequent liquidity needs who trade more often and prefer to trade in a venue with lower transaction costs. Hence, when CDS market becomes available, the net crowding out effect is more prominent for the liquid bonds and subsequently bond demand and prices fall. On the other hand, more illiquid bonds are in general held more by investors who invest with longer horizons, such as pension funds and insurance companies with social responsibilities. Trading CDS increases the protection for these investors and allows them to holder longer and even larger amount of bond credit risks. Therefore bond prices are more likely to rise. Trading CDS is also expected to have a direct effect on bond trading volume, since liquidity trades represent an important portion of the overall trading volume in the cash bond market, trading CDS is expected to have a negative effect on bond trading volume.

The weekly CDS data underlying this study is the novel net notional CDS position data from the Depository Trust & Clearing Corporation (DTCC). We focus on the single-name net notional CDS position data provided at the individual reference entity level. DTCC reports both weekly net notional and weekly gross notional CDS position data. Nevertheless as already argued by Oehmke and Zawadowski (2015), net notional CDS position compared to the gross notional CDS position, is a more accurate measure of the actual stock of credit risk transferred in the CDS market. We merge the DTCC data with bond transaction data from TRACE Enhanced and North America Compustat Quarterly data. Our final dataset includes 2,091 unique bond issues from 411 U.S. public companies, and ranges between December 2008 and September 2014, hence the sample size is large and covers recent developments in both the CDS and bond markets.

To carry out this study, it is important to recognize an identification strategy that enables changes in the CDS outstanding amounts independent from changes in firm financials or changes in macroeconomic conditions. We are able to use the IMM date as the identification strategy. Since the standardization of the CDS market in 2002, all CDS contracts mature or terminate on one of the four IMM dates every year, which are March 20, June 20, September 20 and December 20. We observe

(6)

3

patterns that the net notional CDS position always drops sharply in the near weeks before the IMM dates and goes up again after. Since the schedule of the IMM dates is fixed, firm financial changes or macroeconomic changes are unlikely to be the reason that causes the changes of net CDS position around every IMM date. Although U.S. public firms release their quarterly earnings around similar time, they do not follow strictly a fixed schedule as the IMM dates do. We argue that the pattern of the changes in net CDS should be forced by the IMM dates and further provide supporting evidence. Utilizing this feature of the IMM dates, we find a statistically significant increase in the average net notional CDS position in the first following week after IMM dates. On average, net notional CDS position increases by around 8.61% in week 1 after the IMM date than week 0 – the week that

includes the IMM date2. This provides us an ideal window to carry out empirical analysis. Our final

sample includes data around 24 IMM dates.

We perform pooled regressions at weekly interval between week 0 and week 1 around each IMM date in the sample period, and a few findings emerge. First, using daily bond returns as the dependent variable, we find that an increase in the weekly net notional CDS position is associated with a both statistically and economically significant effect on bond returns. On average, a 1 percentage point increase in the weekly net CDS position is associated with a 0.35 basis point (bp) increase in bond daily returns. Average bond daily returns are small around 1 basis point only, this increase raises the average returns by about 35%. Since daily returns can be noisy, alternatively when using weekly bond returns as the dependent variable and including fewer observations than before (15,626 observations, roughly a reduction of 75% compared to the daily level tests), the coefficient remains positive however becomes significant only at the 10 percent statistical level. This suggests, considering the large number of observations included in the empirical analyses, we ought to interpret with care the statistical significance of the coefficients. Results are not highly robust compared to the number of observations included. Second, we find evidence that trading more CDS can be beneficial for bond prices when the overall capital market is well functioning and there are no funding constraints for investors to forgo arbitrage opportunities. Trading CDS during times with tight funding conditions such as the financial crisis, can even dampen bond prices. We interact a dummy variable crisis that denotes year 2008 and 2009 with the variable for weekly change in net CDS position as a proportion. The coefficient on the interaction term remains negative in alternate specifications however is not statistically significant overall. This is perhaps due to the relatively small number of observations in 2008 and 2009 included in the sample.

(7)

4

Third, we find evidence that trading CDS has differential effects between relatively more liquid and illiquid bonds in the same firm, using different liquidity proxies such as bond issuance size, maturity, and issuance yield spread in the same credit rating category. Results for issuance size and issuance yield spread are not highly robust, while results show that trading CDS seems to benefit more the prices of bonds with longer maturity in the same firm. This suggests that with CDS traded, firms may switch to issue proportionately more long-maturity bonds and more tailor-made bonds to cater to client interests as also argued by Oehmke and Zawadowski (2015), or in order to maintain a good relationship with lenders for availability of funding (Petersen and Rajan (1994)). Fourth, the regressions of changes in net CDS position on bond trading volume in general show weak statistical power. Following standard bond trading volume literature, we use the weekly number of trades and weekly average trade size as dependent variables. Results are overall not statistically robust. We use weekly volume data over daily data because bonds in general are illiquid, and a considerable number of bonds only trade once a day. It is therefore problematic to calculate the intra-day changes in bond trading volume. Weekly interval data, even though still imperfect, can improve to some extent the situation and better pronounce the effects of weekly changes in net CDS on bond trading volume. The sign on the coefficient of the weekly net CDS position is negative when regressed against the weekly number of trades, suggesting investors trade bonds less frequently when the alternative CDS market is available. The sign on the coefficient of the weekly net CDS position is positive when regressed against the weekly average trade size, suggesting that investors trade larger amount of bonds now that they can hedge their positions with CDS contracts.

Our findings are in general consistent with the theories of the liquidity-based model derived by Oehmke and Zawadowski (2015). Their model predicts that trading CDS is expected to push up bond prices when there are no funding constraints and when the trading costs in the CDS market is significantly lower than in the cash bond market. In addition, because liquid and illiquid bonds are held by different type of investors, trading CDS is likely to affect the prices of these bonds differently. Specifically, their model shows that trading CDS benefits the prices of illiquid bonds generally more than those of liquid ones. Both of these hypotheses are consistent with the findings presented in this paper. On bond trading volume, their model predicts that trading CDS should lead to less trading volume in the bond market because liquidity trades are drawn to the CDS market due to lower transaction cost. Results in this research however do not find that there is a strong effect of trading more CDS on weekly bond trading volume. Our findings add to the empirical literature on the dynamic effects of the CDS market on the bond market, and shed additional light on how firms may structure their financing portfolios when credit default swaps are traded. This paper is organized as follows. Section 2 reviews existing literature and presents empirical hypotheses. Section 3

(8)

5

discusses the dataset and details the construction of the sample. Section 4 examines in detail the identification strategy. Section 5 presents the empirical findings, and Section 6 concludes.

II. Literature review and hypotheses

This paper is directly related to the existing literature on the effects of trading CDS on the bond markets. Despite the size of the CDS market and its importance, studying CDS however was previously hindered by the limited data availability. Intuitionally, being able to trade CDS should allow investors to hedge credit risks and reduce exposure to default risks. In terms of cost of debt, the greater protection achieved through CDS for lenders therefore should affect bond market in a similar way as restrictive covenants and we should expect a reduction in firms’ cost of debt. A few early studies that focused on the direct impact of trading CDS on bond market however find opposing results. Ashcraft and Santos (2009) using a sample of 111 firms find that the onset of CDS trading does not lower the cost of capital for the average firm and does lead to a small reduction of bond and

loan spreads for the safer and more informationally transparent firms3. Das et al. (2014) later using

data between 2002 to 2008, find that trading CDS has a detrimental effect on bond market efficiency and liquidity, and does not help reducing bond market pricing errors.

On the other hand, there is a strand of research that find trading CDS is beneficial to the cash bond market in terms of efficiency, liquidity and borrowing terms. Blanco et al. (2005) investigate the effect of trading CDS on the cash bond market, from the perspective of market efficiency. Using a small sample of 33 U.S. and European investment-grade firms between January 2001 to June 2002, they find that CDS prices lead credit spreads in the price discovery process. The introduction of the CDS market improves price discovery, and hence the efficiency of the bond market. Nashikkar et al. (2011) are able to exploit a larger dataset and extend the sample period from July 2002 to June 2006 covering 1,167 firms. They find that there is a positive liquidity spillover effect from the CDS market to the corporate bond market. The liquidity of the CDS contract influences both the liquidity of the bond and the bond price. Saretto and Tookes (2013) studying nonfinancial firms between 2002 and 2010, find that firms with traded CDS contracts on their debt are able to maintain higher leverage ratios and issue longer debt maturities.

To improve the transparency of the CDS market, the Depository Trust & Clearing Corporation (DTCC) since 2008 start to report reference entity level CDS data. Such detailed level data greatly helps to facilitate a better understanding of the CDS market. Recent empirical researches

3

(9)

6

examining the relation between the CDS and the bond market and the Bond-CDS basis are able to exploit this database. Using DTCC data, Oehmke and Zawadowski (2015) are probably the first to provide a comprehensive overview on the fundamental relation between the CDS market and the underlying bond market on individual reference entity level. Their dataset is large and recent. It covers a period from October 2008 to December 2012 and includes 496 firms. They provide evidence that both CDS and bond market are used for hedging purposes, but when it comes to speculating trades, CDS market is the main trading venue. They show the market segmentation is due to that CDS market is superior to the bond market in liquidity and hence attracts more liquidity investors to trade in the CDS market. Nevertheless, their paper is largely descriptive and does not identify any causal effects. Shachar (2011) and Siriwardane (2015) also using DTCC data, examine how dealer inventory and seller capital fluctuations affect credit spreads.

This research differs itself from previous research on a few aspects. First, it exploits the DTCC Single Name CDS data at the individual reference entity level and constructs a large sample

from December 2008 to September 20144. Both bond market and CDS market have evolved

significantly in the last 10 years, therefore our dataset includes more recent data and captures the current dynamics between these two markets. Second, it utilizes the feature of IMM dates around which CDS contracts mature, causing a change in the weekly net notional CDS position. Relying on this identification strategy, this research is able to draw on some causal effects of trading CDS on the underlying cash bond. Third, it examines the direct effects of trading CDS on bond market in terms of both bond price and trading volume. Fourth, it finds evidence that trading CDS affects differently a firm’s bonds that are of differential liquidity level. Through this link, this paper is also related to another area of existing literature on bond illiquidity, such as Collin-Dufresne et al. (2001), Long at al.(2007) and Bao et al. (2011) among others. Bond illiquidity has long been an interest for researchers however are probably among the least well-understood topics. The empirical finding that trading CDS may affect bonds with different liquidity level differently can shed light on corporate financing portfolio build-ups.

This paper does not derive its own theory but uses the liquidity-based theoretical model of Oehmke and Zawadowski (2015) based on the empirical observation that transaction costs are lower in the CDS market than the bond market. With regards to the relation between trading CDS and bond price, they argue that there are three forces together determine the effect of trading CDS on bond price and the overall effect is ambiguous. First, investors previously taking a long position in bond credit risk can switch to directly selling the CDS. This reduces the demand for bond and bond price

4 TRACE Enhanced reports bond transaction data with a lag of 18 months. Therefore we could trace data back to September 2014 the latest.

(10)

7

is expected to decline. Second, investors who previously hold a short position in the bond can purchase CDS instead, which is a much easier trade to enter. This increases the demand for bond and bond price is expected to increase. Together they are the crowding out effects. Since the first effect is larger than the second as shorting bonds are difficult, bond prices are expected to decline overall from the crowing out effect. Third, arbitragers can take part in negative basis trades which involve taking a long position in both the bond and the CDS. Depending on whether the basis traders can take sufficient leverage, their trading activities can have a strong positive effect on the bond price if they can take levered positions or have a weak positive effect when there are capital constraints. This third force is referred to as the CDS-Bond basis-trade effect. The third effect is stronger and the overall effect is positive when basis traders can take large levered positions and when there are limited capital constraints. Therefore we expect a positive relation between a larger net CDS position and bond prices overall during times when funding environment is lenient. And expect a negative relation when there are funding constraints – such as the period during the recent financial crisis, between 2007-2009.

Furthermore, their model shows that through different clientele effects, trading CDS should affect differently liquid and illiquid bonds of the same firm. Liquid bonds are held relatively more by investors with frequent liquidity shocks and the crowding out effect is thus more prominent. Therefore the demand for liquid bonds should disproportionately decrease on trading the firm’s CDS and this is likely to lead to a decrease in bond prices. Illiquid bonds on the other hand are held relatively more by buy-and-hold investors, and the basis-trade effect is more dominant. Therefore, illiquid bonds should benefit disproportionately more from trading the firm’s CDS and bond price is more likely to increase. To examine this relation, we use a dummy variable equal to 1 when the bond is relatively more liquid, and 0 when the bond is relatively more illiquid, and interact it with the net CDS position variable. The interaction term therefore measures the different effects of trading CDS on bond prices between a firm’s more liquid and more illiquid bonds. We use 3 alternative dummy variables to identify liquid and illiquid bonds of a firm, bond issuance size, maturity, and issuance yield spread of bonds in the same rating category. Issuance size and maturity are among the common proxies for bond liquidity. It is well documented that bonds with larger issuance amounts are more

liquid than smaller issues, and long-maturity bonds are less liquid compared to short-maturity bonds5.

For the firm’s bonds in the same rating category, difference in bond yield spread should be mainly driven by the illiquidity component (see Bao et al. (2011) and Chen, Lesmond, and Wei (2007)).

5 Various studies as early as Fisher (1959), Sarig and Warga (1989) and Amihud and Mendelson (1991) already document that bonds with larger issuance amounts are more liquid. Mahanti et al.(2008) using their own developed liquidity measure – the latent liquidity, find evidence that bonds with lower original maturity (5, or 7 years) are more liquid than bonds with original maturity of 10 years or 30 years.

(11)

8

This provides a way to identify liquid and illiquid bonds over and above credit risks. We expect a

negative sign on the interaction term of the liquid bond dummy and net CDS position variable when

regressing against bond returns.

Finally, based on the frequency of liquidity needs, investors choose to trade in the CDS or the bond markets. Intuitionally, compared to buy-and-hold investors, liquidity traders generate a significant portion of the bond market trading volume, and thus when they switch to the CDS market, trading volume in the bond market should unambiguously decline. Hence, we expect trading CDS to have a negative effect on the bond market trading volume.

III. Data

The primary source for bond trading data is Trade Reporting and Compliance Engine (TRACE) Enhanced. First, we eliminate cancelled trades and small trades that are less than $100,000 following Bessembinder at al. (2008) because these small trades tend to be noninstitutional

trades6. Most of the bonds do not trade every day. The 25th percentile bonds trade 92 days per year,

while the 75th percentile bonds trade 227 days per year in TRACE Enhanced during 2008-2014. To

account for this illiquidity, we require bonds to be traded at least 3 days in each week. If there are only 1 day or 2 days trade records for a specific bond in the whole week, the transaction records for that week is eliminated. Next, trading data is matched with Mergent Fixed Income Securities Database (FISD) via the 9-digit Committee on Uniform Security Identification Procedures (CUSIP) identifier. FISD contains bond characteristic details such as coupon, issue size, issue date, maturity, and bond credit rating.

Daily bond price is calculated by weighting each transaction by trade size. From daily price, daily bond returns is calculated as:

𝐷𝑅 = 𝑃𝑡− 𝑃𝑡−1 𝑃𝑡−1

(1) If bond price is missing in the previous trading day, we allow transactions no longer than 3 days to be carried over in order to calculate bond returns. Transaction price reported in TRACE Enhanced is the clean price measure. Hence, the daily bond dirty price is calculated by adding to the clean price accrued interest, which is calculated as:

𝐴𝐼 = 𝑐𝑜𝑢𝑝𝑜𝑛 ×𝑑𝑎𝑦𝑠 𝑠𝑖𝑛𝑐𝑒 𝑙𝑎𝑠𝑡 𝑐𝑜𝑢𝑝𝑜𝑛 𝑑𝑎𝑡𝑒₃₆₅

𝑖𝑛𝑡𝑒𝑟𝑒𝑠𝑡 𝑓𝑟𝑒𝑞𝑢𝑒𝑛𝑐𝑦 (2)

6_{Bessembinder at al. (2008) find that using daily bond return data and eliminate small noninstitutional trades increases}

(12)

9

The last coupon date can be inferred from the first interest date reported in FISD. Daily bond returns based on dirty price is calculated then similarly to the clean price measure. Besides daily bond returns, we also compound the daily returns to a weekly returns measure as an alternative dependent variable.

Bonds without S&P credit rating are excluded as it is not possible to control for credit rating and credit risk. Bonds that are effectively defaulted are excluded. Bonds that have zero outstanding amounts are excluded. Short-term bonds with maturity less than a year are excluded since they are “on-the-run” and their prices might behave differently than the “off-the-run” bonds. The trades of bonds that are in the same month of a bond issuance and a bond maturity are also excluded. Bonds with less than $1000 face value are excluded. Retail notes and foreign government and agency bonds are also dropped from the sample. We also exclude bonds that are denominated in foreign currency or backed by assets. Finally, bonds issued by firms that have already filed for bankruptcy petition are also dropped.

The primary source for CDS data is the Depository Trust & Clearing Corporation (DTCC). DTCC reports weekly CDS open position data. CDS weekly position data (Section I, Table 6), available since October 31, 2008 include both gross notional amounts and net notional amounts outstanding per week. Oehmke and Zawadowski (2015) explain the difference between these two measures. Net notional amounts adjust the gross notional amounts for offsetting positions. For example, if investor A purchases $10 million worth of CDS contracts from a CDS seller C and sells $10 million worth of CDS contracts to buyer B. The total net notional position is $0 in the market, but total gross notional position is equal to $20 million. Therefore, net CDS notional amounts is a more accurate measure of the actual amount of credit risk transferred in the CDS market. Around each IMM date, we collect 11 weeks data –the IMM week and +/- 5 weeks. DTCC and North America Compustat Quarterly data are hand matched by reference entity names and double checked using Wikipedia and Bloomberg. Only exact matches are included, meaning that only parent companies that are available in Compustat and are traded CDS reference entities are kept in the sample. Subsidiaries that are traded CDS reference entities but are not present in Compustat are excluded.

To also control for firm-level financial variables that we might expect to have a significant impact on bond return and trading volume, we further match TRACE Enhanced and Compustat data. First, we utilize the identifier file from S&P Capital IQ on Wharton Research Data Services (WRDS) which contains the past issuer CUSIP of the same firm. Issuer CUSIP changes when there is a corporation action taking place and sometimes it changes even when the company name changes. Therefore we obtain all issuer CUSIP of the same company in order to reduce mismatch due to issuer

(13)

10

CUSIP changes when merging different databases. In this way, we try to maximize sample observations. Next, we use the Center for Research in Security Prices (CRSP) database to obtain parent firm information, because some bond issuers change parent firms during the sample period due to mergers and acquisitions. We manually check the parent firm information where possible and cross check online again using Wikipedia and Bloomberg. Firms that are subsidiaries of the same parent firm can be identified via the parent ID provided by FISD. We then use the 6-digit issuer

CUSIP of the corresponding parent firm at the time to match with Compustat7. In cases where data is

missing, we use data available as of the previous quarter. In order to make sure that all market participants have the firm information available when trading, we lag all Compustat variables by one quarter when including them in regressions.

Data on interest rates, such as the 10-year treasury yield, 10-year minus 2- year treasury spread, and the TED spread are obtained from the Federal Reserve. These data are matched with the bond and CDS data by week. The final sample period is from December 2008 to September 2014. In total, data around 24 IMM dates are collected. The final sample includes 2,091 unique bond issues out of 411 firms.

IV. Identification strategy

A. Net notional CDS position around the IMM dates

The validity of this research relies on the identification strategy – the IMM dates. Before moving on to the empirical analysis, in this section we include a few analysis to evaluate and present support for the identification strategy. Since the standardization of the CDS market in 2002, all CDS contracts mature on one of the four IMM dates, which are March 20, June 20, September 20 and December 20. Figure 1 illustrates the average weekly net notional CDS amounts in +/-5 weeks around the IMM dates in the sample period. Net CDS notional position sees a sharp drop until reaches the lowest point in the IMM week (week 0), and increases again afterwards when investors presumably renew the contracts. Since IMM dates follow a fixed schedule, firm financial changes or macroeconomic changes are unlikely to be the reason that causes the change of CDS net position around every IMM date. Such a natural change in the CDS outstanding amounts – first decrease and then increase right around the IMM dates provides a good opportunity to study the empirical effects of trading CDS on the underlying bond markets.

7

Here I utilize the one-to-one match file between issuer CUSIP and GVKEY created by Tomislav Ladika. I am thankful to him for letting me use the file to improve the merging result.

(14)

11

To examine the magnitude of the week-to-week changes of net CDS position around the IMM dates, we run a series of regressions with results summarized in Table 1. We choose to examine only a few weeks (+/-2 weeks) closest to the IMM dates because changes in these weeks are more likely to be forced by the IMM dates than also due to changes of firm financials or macroeconomic conditions. Despite the fact that in Figure 1 net CDS position on average increases significantly in the third week after IMM dates, that is also close to the time when public companies typically start to release quarterly earnings (early to mid January, April, July, October), which can

Figure.1 Weekly Net CDS Notional Amounts around IMM dates. This figure shows the weekly average CDS Net notional amounts ($ bln) in +/- 5 weeks window around the IMM dates in the sample period (December 2008 to September 2014). Week 0, or IMM week is the week that includes the IMM date. CDS amounts drop before week 0 and goes up afterwards.

generate significant amount of volatilities and can affect trading in both the CDS and the bond market at the same time. Hence it is a more cautious choice to stick closely around the IMM dates. For each week, we use a week indicator that equals 1 in the mentioned week and 0 in the previous week. This allows us to directly compare the weekly changes in net CDS position. For example, a

week 0 indicator equals 1 when it is week 0 (the IMM week), and 0 when it is the previous week, week -1. We also include relevant bond, firm and macroeconomic control variables suggested by

finance theory. Bond variables include S&P credit rating and bond amount outstanding for each unique bond issue. Investors should hold proportionately more CDS for bonds with lower credit rating and larger outstanding amounts. Macroeconomic variables include TED spread and the 10-year treasury yield. TED spread indicates perceived credit risks and long-term interest rate influences

.9 2 .9 4 .9 6 .9 8 1 C D S n e t n o ti o n a l -5 0 5

Weeks around IMM dates

(15)

12

bond default risks which in turn, should have an impact on the net CDS position. We also include firm control variables from NA Compustat Quarterly. All Compustat variables are lagged one quarter when included in the regressions. Firm control variables include the natural logarithm of market value, market leverage ratio and proportion of tangible assets. There should be proportionately larger net CDS outstanding amounts for larger firms and highly leveraged firms. Since leverage ratio for financial firms are much higher than that for non-financial firms, it is not meaningful to control leverage ratio for all universe of firms. Financial firms are therefore excluded in all regressions that include firm level financial ratios.

Results in Table 1 show that all weekly changes in the net CDS position are statistically significant at the 1% level, with the exception of the change between week 1 and week 2. Column (1) – column (4) control for bond and macroeconomic measures, in addition to time and industry fixed effects. The coefficient on both week indicators week -1 (column(1)) and week 0 (column(2)) are negative, suggesting that net CDS decreases before week 0. The coefficient 0.0861 on indicator week

1 (column (3)) is positive and highly statistically significant, meaning that compared to week 0, net

CDS amounts increase by about 8.61% in week 1. The coefficient on indicator week 2 in column (4) is not different from 0 and statistically insignificant. To see if the results vary when also control for firm financials, we run regressions in column (5) to (8) repeating the previous tests however without financial firms. Surprisingly, the coefficient on indicator week 0 in column(6) turns positive and is significant at 1% level, suggesting that for non-financial firms, net CDS amounts may start to go up in week 0 already. Results are similar in column (5), (7) and (8) compared to before. Especially, the coefficient 0.0861 in column (7) is the same compared to column (3) and is significant at 1% level. Result of this analysis is important because first, it confirms that the changes of net CDS around the IMM dates are valid, statistically speaking. Second, it helps with the selection of the regression timing for the empirical analysis later. Specifically, net CDS position has a highly robust increase in the week following the IMM dates, which provides a good window to carry out empirical regressions later. Despite net CDS position is lower overall in week 1 compared to the previous weeks, it should not cause any major issue because we set to examine the incremental effects on bond returns from the incremental increases in net CDS.

B. Is IMM date the forcing variable?

Another question should still be answered is that, are the changes of net CDS outstanding amounts around the IMM dates indeed due to the feature of IMM dates and not due to changes of firm financials or changes of macroeconomic characteristics that result in changes of the default risk

(16)

13 T ab le 1: Chan ge s of Ne t C DS aroun d IM M d at es T h is tab le ex a m in es th e w e ek ly c h an g e s o f lo g (Ne t CDS) b etw ee n we ek T -1 a n d we ek T a ro u n d I M M d ates . Eac h w ee k d u m m y v ari a b le eq u als 1 i n t h at w ee k a n d 0 in th e w ee k b ef o re . F o r ex am p le, d u m m y W e e k 0 e q u als 1 i n we ek 0 , an d e q u als 0 i n we ek -1 . C o lu m n (1 ) – (4 ) co n tro ls b o n d a n d m ac ro ec o n o m ic m ea su re s. Co lu m n (5 ) – (8 ) ad d it io n all y co n tro l fo r fir m c h ara cteristics . F in an cial firms a re re m o v ed in C o lu m n (5 ) – (8 ). S tan d ar d e rro rs are c lu ste re d a t firm le v el. Z -sta ts are in p are n th ese s. * * * sig n if ic an t at th e 1 p erc en t lev el. * * sig n if ic an t at th e 5 p erc en t lev el. * si g n if ica n t at th e 1 0 p erc en t lev el (1 ) (2 ) (3 ) (4 ) (5 ) (6 ) (7 ) (8 ) L o g (Ne t CD S) L o g (Ne t CDS) L o g (Ne t CDS) L o g (Ne t CDS) L o g (Ne t CDS) L o g (Ne t CDS) L o g (Ne t CDS) L o g (Ne t CDS) W ee k -1 -1 .0 8 4 1 * * * -1 .0 2 5 2 * * * (-1 0 .9 9 ) (-9 .9 5 ) W ee k 0 -0 .0 4 1 5 * * * 0 .0 4 4 1 * * * (-3 .5 7 ) (4 .5 1 ) W ee k 1 0 .0 8 6 1 * * * 0 .0 8 6 1 * * * (5 .5 0 ) (4 .3 9 ) W ee k 2 -0 .0 0 3 8 -0 .0 0 3 6 (-0 .6 0 ) (-0 .5 5 ) Bo n d o u tstan d in g 0 .0 1 2 2 0 .0 1 5 0 .0 2 0 0 * * .0 1 8 7 0 .0 0 7 6 0 .0 1 0 4 0 .0 1 6 2 0 .0 1 6 6 (1 .3 9 ) (1 .6 9 ) (2 .2 2 ) (2 .0 4 ) (0 .4 9 ) (0 .6 7 ) (1 .0 7 ) (1 .0 6 ) S & P Cre d it Ra ti n g 0 .0 1 6 7 .0 1 7 1 0 .0 1 6 6 .0 1 6 5 0 .0 1 4 9 0 .0 1 6 9 0 .0 1 8 2 0 .0 2 0 1 (1 .1 7 ) (1 .1 8 ) (1 .1 3 ) (1 .1 1 ) (0 .7 0 ) (0 .7 8 ) (0 .8 2 ) (0 .9 0 ) lo g (M ark et v alu e) 0 .0 8 9 9 0 .0 7 6 6 0 .0 7 4 8 0 .0 7 2 0 (1 .2 2 ) (1 .0 7 ) (1 .0 7 ) (1 .0 3 ) M ark et lev er ag e 0 .9 4 8 1 0 .8 9 3 6 0 .8 5 9 5 0 .7 8 7 5 (1 .3 6 ) (1 .3 2 ) (1 .2 6 ) (1 .1 3 ) T an g ib le as se ts -0 .2 8 0 3 -0 .3 2 7 3 -0 .3 6 2 1 -0 .3 3 4 2 (-0 .8 8 ) (-1 .0 5 ) (-1 .1 5 ) (-1 .0 0 ) T ED sp re ad -0 .7 2 2 8 * * * 0 .3 1 7 8 * * * 0 .3 1 5 0 * * * 0 .6 4 8 7 -0 .6 6 1 2 * * * 0 .7 1 1 2 * * * 0 .2 3 0 9 * * 0 .6 7 9 6 * * * (-7 .8 0 ) (8 .2 1 ) (6 .7 7 ) (1 0 .9 8 ) (-6 .7 5 ) (9 .1 4 ) (2 .5 6 ) (1 1 .8 7 ) 10 -y ea r trea su ry y ie ld 1 4 .1 4 5 2 * * * -0 .6 5 9 1 * * * -0 .6 9 3 4 * * * 0 .3 9 6 8 1 3 .4 6 0 8 * * * 0 .9 3 9 4 * * * -0 .9 7 9 6 * * * 0 .4 2 0 4 * * * (1 1 .0 2 ) (-1 2 .4 4 ) (-1 2 .6 1 ) (5 .4 3 ) (9 .8 7 ) (4 .1 1 ) (-3 .2 9 ) (5 .0 0 ) T im e fix ed e ff e cts Ye s Ye s Ye s Ye s Ye s Ye s Ye s Ye s In d u stry f ix ed e ff ec ts Ye s Ye s Ye s Ye s Ye s Ye s Ye s Ye s o b s. 3 6 ,0 7 9 3 6 ,2 7 7 3 3 ,6 1 6 3 1 ,3 3 7 2 9 ,3 7 1 2 9 ,6 0 6 2 7 ,3 6 7 2 5 ,4 8 6 R -sq u are d 0 .3 9 4 3 0 .4 1 4 9 0 .4 0 7 7 0 .3 8 2 2 0 .3 9 0 4 0 .4 1 3 5 0 .4 0 6 2 0 .3 7 8 6

(17)

14

of the reference entities? First, CDS is an over-the-counter product with contractual nature, meaning that investors intending to hedge or speculate credit risks can enter long or short a credit default swap contract relatively flexibly – if not at any time they want. To cancel the existing contract, they can simply enter an opposite direction trade. Hence, investors are not bound to trade around an IMM date only. Second, it is also unlikely that firm credit risk or macroeconomic conditions change drastically and systematically around each IMM period – which are merely a few weeks. What could be problematic as discussed earlier is that quarterly earnings release of public firms might be around the same time, hence speculative and informed trades may contribute to the changes in net CDS position, and simultaneously these information might also affect bond pricing and trading volume. The first following week after the IMM dates are usually the last week in March, June, September and

December, between 27th and 3rd of the next month. Hence it is not unreasonable to argue that

investors’ choices of whether to renew the CDS contracts could be partly contingent on firm specific information. Nevertheless earnings release do not strictly follow a fixed schedule as the IMM dates for CDS do, and it is not unusual to see companies release quarterly earnings in completely different windows. IMM date is still the compelling force that could explain the large changes of net CDS around these fixed dates.

To provide additional support, we complement the analysis by also examining variables that may cause such changes of net CDS around the IMM dates, suggested by existing literature. Both Oehmke and Zawadowski (2015), and Da Silva et al. (2015) argue that hedging credit risk and speculating on changes of firm financials are motives for investors to take positions in the CDS market. Da Silvia et al. (2015) find evidence that credit risk is a determinant of CDS open interest amounts. Collin-Dufresne et al. (2001) and Blanco et al. (2005) argue that the possibility of default is one reason that explains the credit spreads in corporate bond and CDS prices from the contingent-claims approach. It is reasonable to think that the potential default possibility is also a key factor that drives changes in the CDS open interest amounts. Based on these literature, we choose to examine the following variables around the IMM dates:

1. TED spread. TED spread is an indicator of perceived credit risk in the general economy. A rise in the TED spread indicates an increase in credit risks. Therefore, if there is a consistent pattern of a decrease of the TED spread followed by an increase around the IMM dates, it may dampen the validity of the identification strategy.

2. 10-year treasury yield. A higher spot interest rate increases the risk-neutral discount factor of the firm’s valuation process, which in turn reduces the risk-neutral (but not necessarily the actual) possibility of default. (Blanco et al. (2005)) Therefore, if there is a common trend of

(18)

15

Panel A – TED SPREAD Panel B – DAILY CHANGE (TED)

Panel C – 10-YEAR TREASURY Panel D – DAILY CHANGE (10Y)

Panel E – 10- MINUS 2-YEAR TREASURY Panel D – DAILY CHANGE (10-2Y)

Figure 2. All graphs show daily data in +/- 25 business days around IMM dates, between 2008 and 2014. IMM weeks are denoted by days between 0 and 4. Panel A shows the daily TED spread, Panel B shows the daily change in Ted spread. Panel C shows the daily 10-year treasury yield. Panel D shows its daily change. Panel E shows the 10- minus 2-year treasury yield, and Panel F shows its daily change.

.3 5 .4 .4 5 .5 .5 5 .6 T ED sp re a d -30 -20 -10 0 10 20 30

Days around IMM date

-. 0 4 -. 0 2 0 .0 2 .0 4 T ED sp re a d d a il y ch a n g e -20 -10 0 10 20 30

Days around IMM dates

2 .7 2 .7 5 2 .8 2 .8 5 1 0 -y e a r tre a su ry yie ld -30 -20 -10 0 10 20 30

-. 0 4 -. 0 2 0 .0 2 .0 4 1 0 -y e a r tre a su ry yie ld d a ily ch a n g e -20 -10 0 10 20 30

2 2 .0 5 2 .1 1 0 - min u s 2 -y e a r tre a su ry yie ld -30 -20 -10 0 10 20 30

-. 0 2 -. 0 1 0 .0 1 .0 2 1 0 - min u s 2 -y e a r tre a su ry yie ld d a ily ch a n g e -20 -10 0 10 20 30

(19)

16

first an increase of the 10-year treasury yield followed by a decrease, it may post validity issue for the identification strategy.

3. Spread on 10- and 2- year Treasury yield. The 10Y-2Y yield spread gives indication on the expected real economic growth and inflation. A flattened yield curve indicates a slowing down of the future growth. Therefore we want to observe if the 10Y-2Y yield spread first increases and then decreases consistently around the IMM dates.

Panel A in figure 2 shows the average daily TED spread in +/- 5 weeks or +/-25 business days around the IMM dates, between 2008 and 2014. The average TED spread shows a slight upward trend before the IMM dates. Panel B shows the average daily changes of Ted spread in the same period. Daily changes in TED spread generally fluctuate around zero however seems to be slightly more volatile in the weeks after the IMM dates compared to before, this could be due to the impact of firms quarterly earnings release. Panel C plots the average daily 10-year treasury yield around the IMM dates. And Panel D plots its daily change. Both graphs do not seem to suggest that there is a large and sudden change in the long-term interest rate that can affect bond default risks. Panel E further plots the average daily 10- and 2-year treasury spread, and panel F plots its average daily changes. Examining the graphs of these macroeconomic variables around the IMM dates, the overall impression is that there seems not to be a significant pattern that could cause drastic changes in credit risks or default risks around every IMM date. The graphs illustrating the average daily changes in the macroeconomic variables all fluctuate around mean zero. Figure 3 in Appendix A includes disaggregated plots of the same variables around separate IMM dates and there does not appear to be any persistent patterns.

To be able to estimate the magnitude of these changes, Table 2 further summarizes the mean and standard deviation of these variables in the same time period, +/- 5 weeks or +/-25 business days around the IMM dates, between 2008 and 2014. Specifically, the average daily changes in TED spread near the IMM dates is around 0.0005. When using the one-sample t-test to check if the average change is statistically significant from zero, we do not find evidence to reject that average daily change in TED spread is different from zero with the p-value reported to be 0.7725. The average daily changes in the longterm interest rate and the slope of the yield curve are around -0.0017 and -0.0001. Conducting the same t-test for both the 10-year treasury yield and the 10- minus 2-year treasury spread, we also do not find statistical evidence that the average daily changes in these variables around the IMM dates are different from zero. The evidence from this exercise provides support that the movements of net CDS around the IMM dates are more likely to be driven by the

(20)

17

feature of IMM dates than by changes in credit or default risks. It provides comfort to the main assumption that change of CDS net notional amounts around the IMM dates are exogenous.

TABLE 2: TED SPREAD, 10-YEAR TREASURY YIELD AND 10- MINUS 2-YEAR TREASURY SPREAD

This table reports the mean and standard deviation of TED spread, 10-year treasury yield, 10- minus 2-year treasury spread, and the daily changes of each variable. Time period is the same as in Figure 2: +/- 5 weeks around the IMM dates between 2008 and 2014. P-value is reported from the one-sample t-tests if the daily changes in each of these variables are statistically different from 0.

Mean s.d. One-sample t

test p-value

TED spread 0.4702 0.5739

Daily change 0.0005 0.0665 0.7725

10-year treasury yield 2.7722 0.7144

Daily change -0.0017 0.0702 0.3674

Slope of yield curve 2.0618 0.4649

Daily change -0.0001 0.0519 0.9262

N 1,362

V. Empirical Analysis

A. Descriptive statistics

Having examined the identification strategy, we move on next to the empirical analysis. Table 3 summarizes the descriptive statistics of the sample in the baseline week, week 0. Reporting statistics in the baseline week is useful, because it helps to measure the magnitude of the effects on bond market from the incremental increases in net CDS position between week 0 to week 1 (the regression window identified earlier in Section IV). The mean daily bond returns calculated from the clean price is small, around 1 basis point in the IMM week. For the median firm, the daily bond returns are around 0.2 basis point. Since mean returns more accurately reflect the aggregate returns of investors (Bessembinder et al (2008), p.4222), we will mainly use mean returns as a reference when measuring economic magnitude. The mean daily bond returns based on dirty price is around 5 basis points. The median bond has 7 transactions per day. Nevertheless the standard deviation for daily no. of trades is large around 23.8, suggesting that the number of transactions vary a lot per bond. The mean of daily average trade size is 587,000 units, and the median of daily average trade size is 108,700 units. This further shows that daily trading volume varies significantly from liquid to

(21)

18

TABLE 3: DESCRIPTIVE STATISTICS

Table 2 reports descriptive statistics of the main variables in the sample in week 0, or the IMM week (that includes the IMM date) between December 2008 and September 2014 . All-bond level data. #Daily bond returns (clean price) and (dirty price) are daily bond returns calculated from bond clean price and dirty price respectively. Weekly bond returns are compounded from daily bond returns. Bond price (weekly) and Bond yield (weekly) are the volume weighted average weekly trading price and

yield from TRACE Enhanced. Bond outstanding is total outstanding amounts per parent firm in the IMM week (include both

parent and subsidiary issues). Bond trading volume (by parent firms) is trading volume consolidated by parent firms. Bond turnover is Bond trading volume divided by bond outstanding amounts. Net and Gross CDS are Net and Gross CDS notional

amounts in the IMM week, reported by DTCC. CDS trading volume is CDS weekly activity reported by DTCC Section IVa.

CDS turnover is CDS trading volume divided by net CDS. S&P Credit Rating is rating reported by S&P on each individual bond. AAA+ is given a numerical value of 1, AAA is 2 and so on. D is given a numerical value of 22. Years till maturity is the years left till the bond is mature. Total assets is total book assets atq in Compustat. Total debt is dlc + dltt in Compustat. Market value is calculated as atq – book equity(seqq + txditcq – pstkq) + market equity(cshoq*prccq). Market leverage is the ratios of debt to market value of assets. Market/Book is defined as market value of assets normalized by their book value. Tangible assets ratio is defined as the ratio of property, plant, and equipment (PPE) ppentq divided by book assets in Compustat. All ratios are winsorized at 1% level.

Mean Median sd p25 p75 N

Daily level bond data

Daily bond returns (clean price) 0.0001 0.0000 0.0105 -0.0038 0.0039 70640 Daily bond returns (dirty price) 0.0005 0.0004 0.0102 -0.0034 0.0042 70349

Daily No. of trades 13.8133 7.0000 23.7734 3.0000 15.0000 70640

Daily average trade size (mln units) 0.5873 0.1087 3.3644 0.0287 0.5000 70640

Weekly level bond data

Weekly bond returns (clean price) 0.0003 0.0004 0.0161 -0.0063 0.0075 70640 Weekly bond returns (dirty price) 0.0020 0.0019 0.0157 -0.0045 0.0090 70640

Weekly No. of trades 66.4394 35.0000 109.1986 18.0000 71.0000 70640

Weekly average trade size (mln units) 0.5897 0.2468 1.9983 0.0831 0.6792 70640

Bond price (weekly) 107.7025 106.7184 12.5740 100.2975 114.6925 70640

Bond yield (weekly) 4.1634 4.1639 1.9095 2.7795 5.3180 66303

Bond outstanding amounts (by parent firm) 30.3273 17.9817 37.2689 7.4500 37.7312 70640 Bond trading volume (mln units, by parent firm) 1.1932 0.4141 2.7460 0.1206 1.1003 70640

Bond turnover 0.0351 0.0220 0.0533 0.0123 0.0410 70640

S&P Credit Rating 7.7883 7.0000 2.5265 6.0000 9.0000 70640

Coupon 5.4121 5.7500 1.7990 4.2500 6.5000 70508

Maturity (yrs) 16.9328 10.0521 11.7294 10.0164 30.0247 70140

Years to Maturity (yrs) 13.4001 8.7205 10.8402 6.1534 21.9945 70140

Issuance size (bln$) 0.9487 0.7500 0.7394 0.5000 1.2000 70640

Bond amount outstanding (bln$) 0.8696 0.6525 0.7049 0.4360 1.0000 70640

Weekly level CDS data

Net CDS (bln$) 1.2480 0.9258 1.0371 0.5890 1.5459 70640

Gross CDS (bln$) 16.1075 11.7434 14.7619 6.3772 20.0599 70640

Net CDS/Total debt 0.1009 0.0580 0.1337 0.0260 0.1219 67351

CDS trading volume (by parent firm) 0.1718 0.0890 0.2236 0.0322 0.2212 59871

CDS turnover 0.1491 0.1013 0.1568 0.0447 0.2012 59871

Firm-level data (lagged a quarter)

Total assets (bln$) 195.9519 48.7394 400.6471 23.8770 122.6210 70636

Total debt (bln$) 62.4105 13.7463 145.3189 6.9400 32.8480 70110

Market Value (bln$) 219.1621 70.7450 389.3780 34.7827 201.9698 69982

Market leverage 0.2387 0.2173 0.1344 0.1343 0.3161 69456

Market-to-Book 1.4622 1.2740 0.5786 1.0072 1.7327 69982

(22)

19

illiquid bonds. Compounded from daily returns, the mean weekly bond returns based on clean price in the IMM week is around 3 basis points. The mean weekly bond returns based on dirty price is around 20 basis points. It is interesting to see that the average number of trades in the IMM week is around 66 times, and it is 71 times for the 75th percentile bond in the sample, while for the median bond, the weekly number of trades is around 35 only. This suggests that a large bulk of bond trading volume concentrate on the most liquid bonds. We should note that the bonds in our sample are in general safer and more liquid than the average bond included in TRACE enhanced. 90% of the bonds in the sample are investment grade bonds. And since we only include bonds that trade at least 3 times a week, this makes our sample more liquid than general. Although less than ideal, this should not cause any significant bias in the results and may even help to strengthen part of the analysis that related to the differential effects of trading CDS on liquid and illiquid bonds.

The mean CDS net notional amounts in the IMM week is around $1.25 billion and the median is around $0.93 billion. The mean Gross notional amounts in the IMM week is around $16.11 billion and its median is around $11.77 billion. Offsetting bilateral positions reduces CDS notional positions by about a factor of 13. For the median firm in the sample, there are $5.8 mln net CDS notional amounts on every $100 mln outstanding debt. Because CDS data is only reported on a weekly basis, to directly compare bond turnover and CDS turnover, we compute bond turnover as the total traded volume in a week of all bonds issued (incl. issues from subsidiaries) by a firm divided by the total bond outstanding amounts. For the median firm in the sample, the weekly bond turnover in the IMM week is around 2.2%, while in comparison, the weekly CDS turnover in the IMM week is around 10.1%. For the median firm, net CDS notional amounts are around $0.93 bln and bond outstanding amounts are around $18 bln. CDS market is therefore much smaller in size however accommodates higher liquidity, suggesting that investors trading in the bond market are likely to demand a higher liquidity premium than when trading in the CDS market. Since TRACE Enhanced does not report bond transaction costs such as the bid-ask spread, the turnover measure in a way confirms that CDS market is more liquid and should be relatively cheaper to trade than the underlying bond market. It is also consistent with the prediction of the model from Oehmke and Zawadowski (2015).

B. Net notional CDS position and Bond returns

Next, we move on to investigate the dynamic effects of a larger net CDS notional position on the underlying bond markets. Specifically, we investigate if the sudden increases in CDS net notional

(23)

20

amounts from week 0 (the IMM week) to week +1 around the IMM dates have any effects on bond prices. To tease out the main hypotheses, we estimate the variants of the following pooled regression:

𝑅_𝑗,1= 𝛽₁∆𝐶𝐷𝑆𝑟,1

𝐶𝐷𝑆_𝑟,0 + 𝛽2𝑋 + 𝜀 (3)

Where the dependent variable is the daily returns of individual bond 𝑗, in the week following the IMM dates, the main independent variable is the change of net CDS by reference entity between week 0 and week 1 as a proportion of net CDS position in week 0. 𝑋 includes a set of controls, including bond specific characteristics, quarterly firm financial ratios and weekly change in macroeconomic variables. Bond characteristics include coupon rate, maturity, years to maturity and log(issuance size) that could affect bond prices. Quarterly firm level financial ratios include market leverage and the proportion of tangible assets to total assets that affect bond default risks. Finally, macroeconomic variables include the weekly change in 10-year treasury yield and the 10-year minus 2-year treasury spread that could also influence bond prices.

We perform a series of pooled regressions and Table 4 summarizes the results showing the relationship between changes in net notional CDS amounts and daily bond returns for a range of alternate specifications. All regressions are on individual bond level. Column (1) –(4) use the proportional change of net CDS in week 1 as the independent variable. Because bond returns are very small in magnitude, we keep the independent variable as in proportional change rather than percentage change. Across all 4 specifications, the coefficient on the CDS variable are positive and statistically significant, indicating that trading CDS does have a positive effect on bond prices. Column (1) controls for bond credit rating, bond outstanding amounts, time fixed effects and

industry fixed effects. To put into perspective quantitatively, the coefficient 0.00358 in column (1)

can be interpreted as: holding everything else the same, a 1 percentage point increase in the weekly net notional CDS amounts is associated with a 0.35 basis point increase in bond daily returns (clean measure). To gauge the economic magnitude, we compare it to the mean daily bond returns (clean measure) in the sample, which is around 1 basis point (bp) during the IMM week. This means, a 0.35 basis point increase even though small in absolute terms, raises the average daily bond return by about 35%. The median bond in the sample has a daily returns of about 0.2 bp, therefore for the median bond, a 0.35 bp increase in bond return translates into an even larger effect. Nevertheless this

8

A one unit increase in weekly net notional CDS amounts is in fact an 100 percentage points increase. To give a meaningful interpretation, hence, a 1 percentage point increase in the weekly net CDS, is associated with a 0.0035*0.01 point, or 0.35 basis point increase in bond daily returns. We choose to use the independent variable in proportion, rather than in percentage for the convenience of reporting since bond daily returns are very small in magnitude (slightly higher than 1 basis point). In fact, when using the weekly change in net CDS as a percentage measure, the coefficient is reported as 0.000035 which is troublesome to report.

(24)

21

Table 4. Pooled regressions of daily bond returns on net notional CDS position

This table reports the weekly regression results (between the IMM week, or week 0 and week 1) of pooled regressions under six different regression specifications. The dependent variable is daily bond returns in week 1 based on clean price measure. In Column(1) – (4), the independent CDS variable is weekly change in net CDS position as a proportion, defined as the difference between net CDS in week 1 and week 0, divided by net CDS in week 0. Other explanatory variables include individual bond characteristics: coupon, maturity, years to maturity, log(issuance size), S&P credit rating, and bond amount outstanding; and macroeconomic variables: weekly change in 10-year treasury constant maturity rate, and weekly change in the slope of yield curve – defined as the 10-year minus the 2-year treasury spread. Explanatory variables also include firm level ratios such as market leverage and proportion of tangible assets, both lagged one quarter. When including firm level ratios, financial firms are removed from the sample and this reduces the sample size by around 20%. In Column(5) – (6), the independent CDS variable is the weekly change in the ratio of net CDS normalized by total debt, calculated as net CDS/total debt in week 1 minus net CDS/total debt in week 0. The numbers in the parentheses are z-statistics. *** significant at the 1

percent level. ** significant at the 5 percent level.* significant at the 10 percent level.

Daily bond returns (clean price)

(1) (2) (3) (4) (5) (6)

Net CDS proportional change 0.0035** 0.0030* 0.0033** 0.0033* (2.10) (1.75) (1.98) (1.80)

∆ Net CDS/Total debt 0.0293* 0.0226

(1.73) (1.31)

Coupon 0.0000 -0.0000 -0.0000

(0.67) (-0.08) (-0.13)

Maturity (yrs) 0.0000 0.0000 0.0000

(0.36) (0.22) (0.37)

Years to maturity (yrs) 0.0000** 0.0000** 0.0000**

(2.13) (2.45) (2.25)

log (issue size) -0.0001 -0.0001 -0.0001

(-0.95) (-0.78) (-0.63)

Change in 10-year treasury yield -0.0267*** -0.0245*** -0.0231***

(-5.50) (-4.70) (-4.37)

Change in slope of yield curve 0.0068 0.0079 0.0070

(1.17) (1.21) (1.06)

Market leverage 0.0005 0.0004

(0.72) (0.59)

Tangible assets ratio -0.0009* -0.0008

(-1.74) (-1.32)

S&P Credit Rating -0.0001*** -0.0002*** -0.0001*** -0.0001*** -0.0001*** -0.0001*** (-4.71) (-3.12) (-5.14) (-2.98) (-3.73) (-2.74) Bond amount outstanding -0.0002** -0.0002** -0.0000 -0.0001 -0.0002** -0.0001

(-2.39) (-2.26) (-0.39) (-0.43) (-2.58) (-0.69)

Time fixed effects Yes Yes Yes Yes Yes Yes

Industry fixed effects Yes Yes Yes Yes Yes

Firm fixed effects Yes

SE clustered at firm level Yes Yes Yes Yes Yes Yes

Observations 59,173 59,173 58,713 48,084 55,161 45,864

(25)

22

is a less meaningful measurement. In Section 4, net CDS position is identified to increase by around 8.61 percentage points from week 0 to week 1. Translating this magnitude into bond means that daily bond return on average increases by around 2.9 bp in week 1 compared to week 0 around the IMM dates. For indication purpose, compound this daily increase in bond returns into an annual measure suggests an increase of 6.7% approximately, nevertheless it is a less realistic measure since net CDS position will not sustain a weekly increase of 8.61 percentage points. When we alternatively include firm fixed effects in column (2), the coefficient however becomes smaller in magnitude and is only significant at the 10% level. The coefficient for bond outstanding amounts are negative and statistically significant in both column (1) and column (2) as expected. This shows that investors indeed require a liquidity premium to compensate for holding less liquid bonds.

Specification (3) controls for bond characteristics, as well as macroeconomic variables that might affect bond returns such as change in the level of long-term interest rate and change in the slope of the yield curve. The coefficient is similar in magnitude to column (1), and remains both statistically and economically significant. Specification (4) in addition to (3), also controls for firm characteristics. Since firms with higher market leverage and less tangible assets are riskier, investors trading bonds from these risky firms are likely to demand a higher return in order to compensate for the risk. And since financial firms in general have a much higher leverage ratio, we remove financial firms from the sample according to the SIC classification when including firm specific variables. This reduces the sample size by roughly about 20%. The coefficient 0.0033 in specification (4) remains statistically significant and similar in magnitude to previous specifications, however statistical power weakens. This suggests that trading CDS have a positive effect on bond prices in non-financial firms and the result is not solely driven by financial firms in the sample. The controls for bond characteristics are insignificant statistically, and are due to perhaps the rather short time span in daily bond returns. The controls for firm leverage are also not statistically robust however the signs still enter the regression as expected. Bonds from firms that have a lower proportion of tangible assets to total assets have higher bond returns.

Besides the independent variable of weekly proportional change in net CDS position amounts, we include an alternative independent variable measured by the weekly change of net CDS net notional amounts normalized by total debt of the reference entity. It is calculated as the ratio of net CDS/total debt in week 1 minus that in week 0. Specification (5) is comparable to (1). The coefficient 0.0293 in specification (5) remains statistically significant however is now significant only at the 10 percent level. The 8.61 percentage points increase in net CDS increases the net CDS/debt ratio by roughly around 1 percentage point as outstanding debt is much larger than net CDS. Hence compared to previous specifications the economic magnitude remains consistent, a 1

(26)

23

percentage point increase in weekly net CDS normalized by debt is associated with a 2.9 bp increase on average in daily bond returns. When controlling for bond, firm and macroeconomic characteristics in specification (6), the coefficient becomes smaller to 0.0226 and statistically insignificant. This is perhaps because the effect from trading CDS on the bond market is related directly to the absolute change in the CDS market itself and hence how active the CDS market is, regardless of how large the change is relative to the counterparty’s credit risk exposure.

When we run the same pooled regressions but use instead daily bond returns based on bond dirty price with accrued interest added as dependent variable, the results remain positive nevertheless become smaller in magnitude and in general statistically insignificant. Bond prices in the U.S. are conventionally quoted in clean prices, that is without accrued interest added. Investors therefore tend to think of quotes directly in the clean measure. Probably the stronger statistical relation when using bond returns based on clean price as the dependent variable is because it reflects more directly the price changes for investors. For brevity, we report the details of bond returns based on dirty price as the dependent variable in Table A in the appendix.

One advantage of using daily bond returns is that it increases the number of observations in the regressions. When identifying abnormal bond performance for event studies, Bessembinder at al. (2008) find using daily bond return data can significantly increase the power of tests compared to using monthly data. The disadvantage is that, daily returns incorporate more trading “noise” and approximates less well to a normal distribution. Therefore we compound the daily returns to weekly bond returns and see if the same results still hold when alternatively run the regressions on weekly interval. The number of observations included reduces to approximately 25% of the sample observations before. Table 5 shows that the relation between the weekly change in net CDS position as proportion and weekly bond returns remains positive however across specifications, statistical power weakens. The coefficient 0.0121 in column (1) means that a 1 percentage point increase in net notional CDS amounts is associated with a 0.0121 percentage point increase in weekly bond returns, or 1.2 bp. The mean weekly bond return in the sample during the IMM week is around 3 basis point, therefore the increase in bond returns is roughly equal to 40%, which is comparable to previous

results. The other specifications repeat Table 4before, magnitudes of the coefficients remain similar

however overall with weak statistical significance. This suggests that, despite the statistically significant results shown in both Table 4 and Table 5, we ought to interpret with care the effects because overall the statistical power is not strong in comparison to the number of observations included in the regressions.

An empirical analysis of the effects of trading CDS on the bond market