The Anatomy of the CDS Market

March 1, 2012
Abstract
Using novel position data for single-name credit default swaps (CDS), this paper investigates the
determinants of the amount of credit protection bought (or equivalently sold) in the CDS market. Our
results support the view of CDS markets as alternative trading venues that are used by investors for
both hedging and speculation. CDS markets are more likely to emerge and more heavily used when
the bonds of the underlying rm are fragmented and hard to trade. CDS positions are increasing in
insurable interest (a proxy for hedging needs) and disagreement (a proxy for speculation). These eects
are stronger when the underlying bond is hard to trade. We also nd that rms with a more negative
CDS-bond basis (i.e., the bond is undervalued relative to the CDS) have more CDS outstanding,
suggestive of arbitrage activity.
1 Introduction
The market for credit default swaps (CDS) has grown tremendously in recent years. According to the
Bank for International Settlements (BIS), the CDS market has grown from a total notional amount
of $6 trillion in 2004 to a peak of $57 trillion by mid 2008an almost seven-fold increase in size.
While it has since then decreased in size to around $30 trillion in mid 2010, these numbers show that
the CDS market has developed from an exotic niche market to an enormous and very active market
for credit risk transfer. At the same time, CDS markets are currently at the center of a number of
policy debates, including their role during the recent nancial crisis (for a summary of the issues, see
Stulz, 2010) and, more broadly, their impact on the debtor-creditor relationship through the potential
emergence of empty creditors (see, e.g., Hu and Black, 2007; Bolton and Oehmke, 2011).
Despite this, very little is known about the positions taken in the CDS market. This paper aims
to ll this void. Using newly available, disaggregated data on individual single-name CDS positions,
our paper makes three main contributions. First, we establish a number of basic stylized facts about
the positions taken in the single-name CDS market. For example, we show that net notional CDS
positions as a fraction of debt are generally smaller than is often claimednot many rms have net
notional CDS amounts outstanding that exceed their debt. Second, our analysis sheds light on the
determinants of CDS market existence. In other words, what characteristics determine whether a
rm becomes a traded entity in the CDS market? Third, we investigate the determinants of the
size of CDS positions. In other words, given that a rm is a traded reference entity in the CDS
market, what determines the amount of credit risk that is transferred in those markets? The data
underlying our analysis of CDS positions are the newly available weekly market statistics that are
released by the Depository Trust & Clearing Corporation (DTCC), which is the most comprehensive
and disaggregated CDS position data available.
In a CDS, a protection seller agrees to make a payment to the protection buyer in the case of a
credit event on a prespecied reference entity. In exchange for this promised payment, the protection
seller receives a periodic premium payment (and potentially an upfront payment) from the protection
buyer. What exactly determines a credit event is dened in the initial CDS contract. Generally
speaking, credit events include a bankruptcy ling of the reference entity, non-payment of debt, and
2
in some CDS contracts, debt restructuring or a credit-rating downgrade. When a credit event occurs,
the protection seller pays the protection buyer the dierence between the face value of a pre-specied
reference bond and the recovery value of that bond, which is typically determined in a CDS settlement
auction.
Our empirical analysis of CDS positions focuses on the net notional amounts of CDS outstanding
on individual reference entities. We focus on the net notional amount, because it provides a much more
accurate picture of the amount of credit risk transferred in the CDS market than the gross notional
amount, which is the main alternative measure of CDS market size. The net notional amount is
calculated as the sum of net protection bought by counterparties that are net buyers of protection
for a particular reference entity (or equivalently, the sum of net protection sold by all counterparties
that are net sellers of protection for a particular reference entity). In other words, when calculating
the amount of CDS outstanding, the net notional amount corrects for osetting positions within
counterparties. One way to interpret the net notional outstanding is to think of it as the maximum
amount of payments that need to be made between counterparties in the case of a credit event on
a particular reference entity.
1
The gross notional amount, on the other hand, simply sums up all
outstanding contracts: it is the sum of all protection bought, or equivalently all protection sold. The
gross notional amount may increase even as counterparties in the CDS market reduce their exposure,
for example by taking on osetting positions, and is thus economically less meaningful.
In order to investigate the determinants of both existence and positioning in the CDS market, we
combine DTCC data on net notional amounts outstanding in the CDS market with a number of other
data sources (e.g., Compustat, Mergent FISD, Bloomberg, Capital IQ). We then run probit regressions
to estimate the determinants of CDS market existence, and censored regressions to investigate the
determinants of CDS position sizes. We use a censored regression approach because the DTCC reports
notional CDS amounts outstanding only for the 1,000 largest traded reference entities in terms of gross
notional. This means that some traded reference entities with small amounts of CDS outstanding do
not make it into the DTCC data.
Our results broadly suggest that CDS markets function as alternative trading venues for both
1
It is the maximum amount of payments, because actual payments will usually be less than the par value of the CDS,
reecting non-zero recovery rates on the defaulted bonds as well as previous marking-to-market by counterparties.
3
hedging and speculation on the underlying bond. They are alternative trading venues in the sense
that investor could often make the same economic trade directly in the underlying bond, but choose to
use the CDS market. This interpretation of CDS markets as alternative trading venues is supported
by the nding that CDS markets more likely to emerge when the underlying bonds of the rm are
illiquid and thus hard or expensive to trade. Similarly, conditional on a CDS market existing for a
reference entity, the positions taken in the CDS market are larger when the underlying bond is illiquid.
More specically, we rst document that rms with more assets are more likely to be traded refer-
ence entities in the CDS market. However, conditional on being a traded reference entity in the CDS
market, more assets are not associated with larger CDS positions. Similarly, rms with more debt
outstanding are more likely to be traded reference entities. In addition, if they are traded reference
entities, rms with more debt tend to have more CDS outstanding. When looking at disaggregated
balance sheet data, we nd that the eect of debt on CDS positions is mainly driven by bonds out-
standing. We also nd that accounts payable are associated with more CDS. The positive coecients
on debt and accounts payable suggests that insurable interest (i.e., debt outstanding, payments to
be received from trading partners) is a determinant of CDS positions. Similarly, rms that provide
credit guarantees (e.g., monoline insurers) tend to have more CDS outstanding. Taken together, these
ndings are consistent with the view that at least some debtholders use CDS markets to hedge their
debt, bond, or counterparty exposure.
Second, we document eects of credit quality on CDS positions. In particular, investment grade
rms have more CDS outstanding than rms that are not investment grade. This is particularly
true for rms that are investment grade, but close to the investment grade cuto. In addition, fallen
angels (i.e., rms that have lost investment grade status) have more CDS outstanding, suggesting that
investors exposed to these rms use the CDS market to hedge their exposure.
Third, for traded reference entities disagreement, measured as earnings forecast dispersion, is
associated with larger net notional amounts of CDS outstanding. This suggests that investors use the
CDS markets to speculate by taking views on the default probabilities of traded reference entities.
Similarly, equity turnover is also associated with larger positions in the CDS market.
While these results suggest that both hedging and speculation are determinants of CDS positions,
they do not explain why investors prefer to hedge or speculate in the CDS market, as opposed to
4
directly trading the underlying bond. Here our analysis suggests that CDS markets emerge as alter-
native trading venues for credit risk, in particular if the underlying bonds are illiquid or hard to trade.
Specically, controlling for the amount of bonds outstanding, we show that if a rms bond issues are
fragmented, as proxied by the Herndahl index, CDS markets are more likely to emerge. In addition,
for rms that are traded reference entities in the CDS market, a more fragmented bond market is
associated with larger positions in the CDS market. Both of these results suggests that CDS contracts
create a unied liquid market for multiple individual bond issues.
The interpretation of the CDS market as an alternative trading venue receives further support when
we interact proxies for hedging or speculation with measures of trading frictions in the market for the
underlying bond. In particular, we nd that the eect of both disagreement, a proxy for speculation,
and lost investment grade status, a proxy for hedging demand, on the net notional amounts of CDS
outstanding are stronger for whose bonds are illiquid. This gives further support for the view that
investors use CDS markets to take views or hedge exposures when trading in the underlying bond is
dicult or expensive.
Finally, we document that net notional CDS positions are increasing in the CDS-bond basis, which
is the dierence between the CDS spread and the spread over the risk-free rate on the reference bond.
When the CDS-bond basis is negative (as it has been form many reference entities since the nancial
crisis), the bond is cheap relative to a synthetic bond formed out of a CDS and the risk-free rate
(a situation that gives rise to the so-called negative basis trade). Our analysis shows that rms
which have a more negative CDS-bond basis have more CDS outstanding. This result suggests that
arbitrageurs lean against the negative CDS-bond basis. Moreover, if via the negative basis trade the
presence of CDS can help to (partially) eliminate mispricings in the bond market and potentially
compress spreads for bond issuers, this may improve rms access to nancing. This interpretation
echoes the arguments in Saretto and Tookes (2010), who maintain that the presence of CDS allows
rms to borrow more and at longer maturities.
The remainder of the paper is structured as follows. In Section 2 we discuss the related literature
and draw a number of empirical hypotheses from it. In Section 3 we describe our data sources. Section
4 discusses our empirical strategy and presents the main empirical ndings. Section 5 contains a
number of robustness checks. Section 6 concludes.
5
2 Related Literature and Hypothesis Development
Despite the recent growth in the literature on CDS markets, relatively little is known about positioning
and the determinants of position sizes in these markets. To our knowledge, this is the rst paper that
systematically investigates positioning in the CDS market using position data at the reference entity
level. In his survey of CDS markets and their role in the recent crisis, Stulz (2010) provides a number
of summary statistics based on aggregate position data from the DTCC and survey data from the
Bank for International Settlements (BIS), which was the main source of position information before
the DTCC data became available. Stulz (2010)also provides a good introduction to the basics of CDS
markets and the current policy debates about CDS.
A couple of other recent empirical studies that look at CDS positions or transaction volume.
Shachar (2011) uses detailed transaction level data to investigate price eects of traded volume, order
imbalances and dealer inventories in the CDS market. Lee (2011) uses data from the DTCC to
document a predictive eect of net notional amounts outstanding in the CDS market as a fraction of
debt on a reference on stock prices and CDS spreads of that reference entity. Neither of these papers
investigate the determinants of positions taken in the CDS market. In addition, a number of recent
papers have investigated the so-called CDS-bond basis, which is the dierence between the CDS spread
and the spread over the risk-free rate of the underlying bond (for details on the CDS-bond basis and
conditions under which it should be zero by no arbitrage, see Due, 1999). Recent empirical papers
on the CDS-bond basis include Blanco et al. (2005), Nashikkar et al. (2010), Bai and Collin-Dufresne
(2010), and Fontana (2011). Our paper contributes to this literature by linking the CDS bond basis
to quantities in the CDS market.
More broadly, there is a growing empirical literature that investigates the eects of CDS markets
on information transmission, risk transfer, and credit market outcomes. For example, Acharya and
Johnson (2007) document informed trading in CDS markets that leads equity markets in response to
negative credit news. Qiu and Yu (2012) examine liquidity provision in CDS markets by investigating
the number of distinct dealers that provide quotes for certain reference entities. They link their
liquidity proxy to obligor size, credit rating and to information ow from the CDS market to the stock
market. Minton et al. (2009) document the use of CDSs as risk transfer instruments by banks.
6
Another strand of recent empirical literature has investigated the eect of CDS markets on credit
supply. Ashcraft and Santos (2009) nd that the introduction of CDSs has lead to an improvement
in borrowing terms for safe and transparent rms, where banks monitoring incentives are not likely
to play a major role. Hirtle (2009) shows that greater use of CDSs leads to an increase in bank credit
supply and an improvement in credit terms, such as maturity and required spreads, for large loans
that are likely to be issued by companies that are named credits in the CDS market. Saretto and
Tookes (2010) document that rms with traded CDS contracts can sustain higher leverage and borrow
at longer debt maturities. While it is not their main focus, Saretto and Tookes (2010) also investigate
the emergence of CDS market.
Finally, our paper relates to a growing theory literature on the use of CDSs. Much of that literature
has focused on why investors may have an incentive to trade in the CDS markets. One strand of
literature links trading in CDSs to hedging demands by banks or investors. For example, in Duee
and Zhou (2001), CDS contracts allow for the decomposition of credit risk into components that are
more or less information sensitive, thus potentially helping banks overcome a lemons problem when
hedging credit risk. Thompson (2007) and Parlour and Winton (2008) analyze how banks lay o
credit risk when they have a choice between loan sales and purchasing CDS protection. Parlour and
Plantin (2008) analyze under which conditions liquid markets for credit risk transfer can emerge when
there is asymmetric information about credit quality. Their model implies that markets for credit
risk transfer are more likely to be active for high-quality issuers. In Bolton and Oehmke (2011) CDS
contracts strengthen the ex-post bargaining position of creditors and thus allow the rm to raise more
nancing ex-ante. At the same time, CDSs may also lead to inecient liquidation of the rm at an
interim date. In Zawadowski (2011) investors can use CDS to insure counterparty risk.
A second strand of literature uses heterogeneous beliefs models to analyze trading in CDSs. For
example, Che and Sethi (2011) use a model with heterogeneous beliefs to analyze the impact of
credit derivatives on rms cost of capital. In their model, the presence of the CDS market can lead
some investors to take CDS positions rather than buying the bond, which can raise a rms cost of
capital. Geanakoplos and Fostel (2011) argue, within a heterogeneous beliefs model, that tranching
and CDS may have contributed to the recent boom bust cycle. One general implication of models
with heterogeneous beliefs is that an increase in disagreement among investors usually leads to more
7
CDS outstanding.
Hypotheses. We now draw on the related literature to develop a number of hypotheses. Taken
together, the theory literature thus suggests two main motives for trading in CDS: hedging and
speculation. This leads to the rst two hypotheses. First, to the extent that CDS are used for hedging,
insurable interest should play a role in determining the net notional amounts of CDS outstanding.
Consider, for example, a setting in which investors insure a constant fraction of their bonds. When
more bonds are outstanding (i.e., insurable interest rises) we should expect to see a larger net notional
amount of CDS outstanding. A similar eect should be present when with other forms of exposures.
For example, if a rm owes suppliers payments (accounts payable or trade credit), the suppliers may
choose to lay o some of that risk by purchasing credit protection. Speculation, on the other hand,
should not be directly related to insurable interest, since it is a pure bet on future changes in credit
quality and thus does not directly depend, for example, on the amounts of bonds outstanding or the
size of the reference entity.
Second, to the extent that CDSs are used as speculative instruments (i.e., investors may use CDS
contracts as speculative vehicles in order to express views about a reference entitys default prospects,
even if they does not own the bond or have any other exposure to the reference entity), reference
entities on which investors beliefs dier more should have larger CDS positions outstanding than
reference entities with less disagreement, as in the models of Che and Sethi (2011) and Geanakoplos
and Fostel (2011).
Finally, because investors can choose between trading in the CDS market or directly in the un-
derlying bond, we expect the eects of speculation or hedging motives on the amount of net notional
CDS protection outstanding to be present in particular for rms with less liquid bonds. The rationale
is that, while in principle investors can place bets on default risk both in the bond and in the CDS
market, they may have a preference for using the CDS market when the underlying bond is illiquid
and thus expensive to trade. In a similar manner, investors that want to hedge themselves against
the deteriorating credit risk have a choice of selling the bond or buying a CDS that hedges their
exposure. Also in this case, hedging through CDS as opposed to selling the underlying bond may be
particularly attractive when the underlying bond is illiquid and thus hard to sell. By this logic, CDS
markets should be more likely to emerge and more heavily used for rms whose bonds are hard to
8
trade, for example because they are illiquid or fragmented into many separate bond issues. In this
case, a CDS market may oer investors a more unied alternative trading venue that is more liquid
than the underlying bonds.
3 Data
3.1 DTCC data
Our data on CDS positions comes from the Depository Trust & Clearing Corporation (DTCC). The
DTCC provides clearing, settlement and trade conrmation in a number of markets, such as equities,
corporate and municipial bonds, and over-the-counter derivatives. In the CDS market, the DTCC
provides trade processing and trade registration services. All major dealers register their standard
CDS trades with the DTCC. The DTCC then enters these trades into a Trade Information Warehouse
(TIW). Since October 31, 2008, the DTCC has published weekly reports on CDS positions based on
the trades recorded in the TIW. The positions recorded in the TIW capture almost the entire market
for standard single-name CDS.
2
According to the DTCC (2009), the TIW captures around 95% of
globally traded CDS. The data is thus the most accurate and comprehensive dataset that is available
on CDS positions. In this study we use 27 months of DTCC data, from October 2008 to December
2010.
Prior to the release of position data by the DTCC, the survey data from the BIS (which is available
from 2004 onward and can be found at http://www.bis.org/statistics/derstats.htm) was the main
source of information about position sizes in the CDS market. However, relative to the DTCC data,
the BIS data has a number of disadvantages. First, the BIS data only provides aggregate market
statistics, while the DTCC data provides positioning at the individual reference entity level. Second,
the BIS data is based on surveys as opposed to actual registered positions in the market. Third,
because of its survey-based nature, the BIS data is prone to double counting. This issue arises
because as part of the survey, the same CDS transaction may be reported both by the buyer and the
2
The DTCC also registers CDS positions on sovereigns, indices, and structured nancial instruments. In this study we
focus on single-name CDS on companies. According to Stulz (2010), in June 2008 single-name CDS made up about 58% of
the overall CDS market.
9
seller to the transaction, resulting in a double count.
In its weekly reports, the DTCC discloses both the aggregate gross notional as well as the aggregate
net notional amounts outstanding on a particular reference entity, where notional refers to the par
amount of credit protection that is bought or sold. The gross notional amount outstanding is the sum
of all notional CDS contracts on a given reference entity. The gross notional amount thus reects the
total par amount of credit protection bought (or equivalently sold). It is dened as either the sum
of all long or, equivalently, the sum of all short CDS contracts outstanding. With the exception of
occasional compression trades, in which osetting CDS positions are eliminated, the gross notional
amount outstanding increases with every trade. In particular, the gross notional position will increase
even if a trade osets an existing trade and thus reduces the overall amount of credit risk transfer
in the CDS market. This makes the gross notional amount outstanding an imprecise proxy for the
amount of credit risk that is transferred in the CDS market.
The net notional amount outstanding adjusts the gross notional amount for osetting positions in
order to better reect the actual economic protection purchased. It is calculated as the sum of net
protection bought by counterparties that are net buyers of protection for a particular reference entity
(or equivalently, as the sum of net protection sold by all counterparties that are net sellers of protection
for a particular reference entity). One intuitive way to interpret the net notional outstanding is to
think of it as the maximum amount of payments that need to be made between counterparties in the
case of a credit event on a particular reference entity. It is the maximum amount of payments, because
actual payments will usually be less than the par value of the CDS, reecting non-zero recovery rates
on the defaulted bonds as well as previous marking-to-market by counterparties.
The dierence between gross notional amounts and net notional amounts is best illustrated via a
simple example. For a graphical illustration of this example, see Figure 1. Consider three banks, A, B
and C, that are trading CDS on a given reference entity. Assume that initially B has purchased $10m
in credit protection from A. The gross notional amount outstanding, calculated as either the sum of
all gross protection bought or sold, is given by $10m: Bank A has sold a gross amount of $10m to
bank B, and bank B has bought a gross amount of $10m from bank A. By a similar reasoning, the net
notional amount, calculated as the sum of all net protection bought or sold, is also given by $10m.
This is illustrated in Panel (a) of Figure 1. Now assume that B osets the initial trade by selling
10
$10m in protection to C. This raises the gross notional amount to $20m: Bank B has bought a gross
amount of protection from bank A, and bank C has bought a gross amount of protection from bank
B. The net notional amount, however, remains at $10m. The reason for this is that B is now fully
hedged, such that the only net payment to be made in the case of default is a payment of $10m from
A to C. This is illustrated in Panel (b) of Figure 1. Finally, in Panel (c), B sells $10m in protection to
A, such that all three parties have a net zero position. This means that the net notional outstanding,
the sum of all long or short net positions, is also $0. The gross notional, on the other hand, which is
given by the sum of all gross long or short positions, is now given by $30m.
3.2 Other data sources
We combine the DTCC position data with a number of other data sources. We take balance sheet data,
credit ratings, and industry codes from Compustat. For more detailed capital structure information,
we hand-collect information from Capital IQ. We use CRSP for equity market data, we gather data
on outstanding bonds from Mergent FISD, and obtain bond trading data from TRACE.
In terms of dataset construction, we start with the universe of US and international rms contained
in Compustat. We use consolidated balance sheet data and drop all subsidiaries. Debt is dened as the
sum of current and long-term debt. We then merge this data to all single-name reference entities in the
DTCC data (using rm name as an identier). We drop companies that are not uniquely merged (i.e.,
cases where merging by name gives us two Compustat companies for one reference entity from DTCC).
We hand check these merges to make sure that the reference names in DTCC are correctly assigned
to companies in Compustat. If the DTCC reference entity is not in Compustat but the parent is, we
match to the parent. Later we perform robustness checks excluding such indirect matches. We drop
DTCC reference entities that we could not nd in either the domestic or the international Compustat
dataset. We also drop all companies for which both parent and subsidiary appear in the DTCC data,
since we cannot perfectly net CDS outstanding in these cases: CDS written on subsidiaries might
be indirectly written on the parent. We drop companies that have assets less than 100 million (the
smallest company that has CDS in the DTCC data has assets of about 250 million). Finally, we check
whether a certain reference entity is part of a major CDS index (CDX.NA.IG, CDX.NA.HY, Itraxx)
based on the Markit manual. We also hand-match our data to balance sheet data from Capital IQ.
11
We then add data on the bonds outstanding of US companies using Mergent FISD. We exclude
all short-term bonds; only bonds with at least 366 days of original maturity are considered. We also
exclude bonds in the month of their issuance and the month of their redemption. Pass-through notes
are also dropped. In Mergent FISD we drop bonds that have been eectively recalled or decrease
the amount outstanding by the recall amount. We also drop bonds with zero or unrecorded oering
amount. We calculate bond trading for the bonds in Mergent FISD using Trace and match the two
using the CUSIP of the bond issues. In Trace 1MM+ is replaced by 1 million and 5MM+ by 5
million.
Given the complicated legal structure of companies, we construct two dierent measures of bonds
outstanding. The rst measure, bonds outstanding includes all the bonds of a given limited liability
entity including all bonds issued by companies that have been acquired and fully dissolved. This
is important because in case of mergers and acquisitions, the new parent inherits the bonds of the
old company. The second measure, consolidated bonds outstanding includes all bonds issued by all
companies with the same ultimate parent. This distinction is important but imperfect. For example,
ExxonMobil in its 2010 annual report (which is consolidated), reports bonds issued by the acquired
companies Mobil, SeaRiver Maritime Financial Holdings, and XTO Energy as liabilities. However
while ExxonMobil is legally liable for Mobil bonds since it is a not a separate legal entity, this is
not necessarily the case for the bonds of XTO Energy since it is a limited liability subsidiary. On
the other hand, while SeaRiver Maritime Financial Holdings is a limited liability entity and as such
ExxonMobil is not necessarily liable in case of a default on these bonds. However, in its 2010 annual
report ExxonMobil explicitly states that it guarantees bonds issued by SeaRiver Maritime Financial
Holdings, while no explicit guarantees are stated in case of XTO Energy (which of course does not
rule out implicit guarantees). Note that in the DTCC database we do nd CDS outstanding on
both ExxonMobil and XTO Energy but not on the other subsidiaries of ExxonMobil. Thus parent
companies may or may not be liable for the bonds of their subsidiaries leading to potential joint
defaults. While it is extremely cumbersome to verify all such guarantees, in our regression analysis we
analyze whether they matter. Note that Compustat and Capital IQ both look at consolidated balance
sheets thus they treat the bond issuances of all subsidiaries as that of the parent or acquirer.
The matching between bond issues and Compustat companies is done along two dimensions. First,
12
since most of the companies issuing debt also have traded equity, we use the CRSP les to match
old cusips to new cusips. In case of a merger or acquisition, we use the same le to nd the new
parent company. We then hand-check all the matches and verify whether the acquired companies (or
subsidiaries) are limited liability entities or not, i.e. whether the parent is liable for the obligations.
Second, we use the Mergent FISD parent identier to consolidate companies with the same parent. To
exclude potential erroneous matches between Compustat and FISD Mergent we exclude companies
that have more than twice as many bonds than debt.
3
We then use the rst six digits of CUSIP
(which identies the issuer) to match our data to bond data from Mergent FISD. In a second round
of matching all unmatched issues in Mergent FISD are, if possible, hand-matched to Compustat.
We compute three dierent measures of bond liquidity. First, for the bond Herndahl measure
sum the squared shares that each bond issue contributes to the overall amount of bonds a limited
liability entity (issuer) has issued. We use the Herndahl index to measure the fragmentation of a
rms total outstanding bonds into separate bond issues. In a similar spirit, Choi et al. (2012) use
the Herndahl index to measure how spread out the maturity dates of a rms bonds are. Second,
we compute bond turnover as the monthly trading volume from Trace for all bonds outstanding over
all bonds of a given reference entity. Our third liquidity measure, number of trades is the number of
all the trades in all bonds of a given issuer. When using the latter two measures as a liquidity proxy,
we use the rolling average of the preceding 12 months in order to minimize the confounding eect of
current trading demand.
For all companies with at least one bond issue identied in the FISD Mergent database we hand-
search Bloomberg for 5 year CDS spreads. We use 5 year CDS spreads because they are the most liquid.
The months CDS spread is the last CDS quote recorded in that given month. For all companies with
a CDS spread we then search for xed-coupon bonds without any embedded options
4
and a remaining
maturity of 1 to 10 years. For every issuer we rank these bonds by dollar volume of trade. This
enables us to concentrate on the most liquid bond issuances of a given company. We then hand-search
Bloomberg for the CDS-bond basis of these bonds, and take the bond with the highest dollar-volume
with a reported CDS-bond basis. Bloomberg calculates this as the dierence of the CDS spread
3
Note that a company might have somewhat more bonds because Compustat and Mergent FISD data are not perfectly
synchronized in time.
4
Thus we exclude all bonds with oating coupon, all bonds that are callable, putable, redeemable or convertible.
13
(interpolated to the exact remaining time to maturity) and the z-spread (the parallel shift in the
treasury yield curve that gives that matches the price of the bond).
We use IBES earnings analyst forecasts to calculate measures of disagreement. We take monthly
data on the two-year earnings per share forecast since it has the most forecasts. For the analyst
disagreement: std./mean measure we calculate standard deviation above mean estimate of 2 year
earnings if the mean estimate is above 5 cents per share otherwise use missing value and there are
at least 3 forecasts; analyst disagreement: std./price is calculated by dividing standard deviation of
forecasts by CRSP stock price if the stock price is above one dollar.
We drop companies with SIC industry code 9995 (non-operating establishments) and companies
with no assets. We also drop companies for which we have no SIC codes. To avoid possibly erro-
neous matches with Capital IQ and Mergent FISD that result in outliers, we lter our matches. We
exclude Capital IQ observations for which the the total amount of borrowing measured by Capital IQ
exceeds the total amount of debt measured by Compustat by more than 50% of assets (measured by
Compustat). Similarly we exclude all Mergent FISD observations for which the total amount bonds
outstanding measured by Mergent FISD exceeds the total amount of debt measured by Compustat by
more than 50% of assets. The results are not sensitive to the exact specication of such data ltering.
Companies with SIC code 9997 are hand-assigned to industries.We winsorize all variables computed as
ratios at the 99% level. We winsorize the CDS-bond basis at the 5% and 95% level because of outliers
but the winsorization does not eect the results. Finally, we drop (quasi) state-owned companies (Fan-
nie Mae, Federal Home Loan Mortgage Corporation, United States Postal Service). These companies
have large asset bases but no CDS, and thus behave very dierently from the regular sample.
4 Empirical Analysis and Results
4.1 Summary Statistics
Table 1 provides summary statistics for our data. The table is split in three parts. First we provide
summary statistics for our Compustat sample. We then present summary statistics for the subsample
of U.S. rms that have at least one bond outstanding in Mergent FISD. We restrict our attention to
14
U.S. rms in Mergent FISD because non-US rms are likely to issue a substantial fraction of their
bonds in other jurisdictions and would thus are not be captured in the dataset. Finally, we provide
summary statistics for the subset of rms for which we have detailed (annual) balance sheet data from
Capital IQ.
Overall, we have data on net and gross CDS positions for 14,784 rm-month observations. The
mean gross notional amount of CDS outstanding on a reference entity in our sample is $14.63bn.
The mean net notional amount outstanding is given by $1.19bn. Hence, on average netting within
counterparties reduces the amount of CDS outstanding by a factor of more than ten. The average
number of CDS contracts outstanding on rms in our sample is 2,185.
Another way of looking at the amount of credit protection outstanding is to normalize the CDS
protection bought or sold by either total assets or total debt of the reference entity. The net notional
amount of CDS outstanding for the average rm in our DTCC sample is equal to 7.99% of total
assets. The 90th percentile of net CDS as a fraction of assets is 19.9%, the 10th percentile 0.7%.
When looking at net notional as a fraction of the reference entitys outstanding debt, we nd that for
the average rm in our sample the amount of net notional outstanding is CDS is equal to 28.6% of
the companys debt. Even at the 90th percentile, net notional CDS outstanding are equal to only
64.6% of debt. For companies where we have access FISD data on bonds outstanding, the mean ratio
of CDS to bonds is given by 66.2% when only looking at bonds issued directly by that rm, and 42.7%
when we consolidate bonds to also include bonds issued by subsidiaries. The 90th percentiles given by
135.0% and 105.3%, respectively. While these are signicant amounts, the data does not conrm the
conventional wisdom that the amounts outstanding in CDS markets usually vastly exceed insurable
interest (at least not when looking at the economically more meaningful quantity of net protection
bought or sold).
5
Table 1 also provides summary statistics for total assets and debt data (from Compustat), ratings
data (from Compustat Ratings Express), bond turnover (from TRACE), two disagreement measures
5
Nonetheless, there are a few companies for which the amount of net notional of CDS outstanding exceeds debt or the
amount of bonds outstanding, sometimes by signicant amounts. However, these companies are exceptions. Some of the
companies that have high CDS as a fraction of their debt outstanding are potential buyout targets (with low current debt,
but potentially large future debt if a leveraged buyout is to take place). Examples from this group are the clothing retailer
Gap, or the electronics distributer Arrow Electronics. Other types of companies with high CDS as a fraction of debt are
homebuilders, mortgage insurers and suppliers for the automobile industry.
15
(calculated from IBES), and the disaggregated balance sheet data obtained from Capital IQ.
Figure 2 documents the evolution of total net CDS positions outstanding over the course of our
sample. The top solid line is the total amount of net CDS outstanding on all single-name reference
entities, as reported by the DTCC. It thus captures the net notional outstanding in the entire single-
name CDS market. The dashed line below the solid line is the total net notional in CDS protection
written on the top 1,000 single name entities. Comparing this line to the total single-name CDS market
demonstrates that the top 1,000 reference entities make up a large fraction of the overall single-name
CDS market, at least when measured in net notional CDS outstanding. The dotted line plots the
total net notional amounts of outstanding CDSs for the reference entities that we managed to match
with Compustat. While we lose some reference entities (e.g. sovereigns) in the matching process, our
matched reference entities still constitute roughly 50% of the total single-name CDS market.
Figure 3 plots the ratio of gross to net amounts in CDS outstanding over time. While overall
positions in CDS markets decline over our sample, the ratio of gross to net increases. This is true for
the median ratio of gross to net, but also for the 10th and 90th percentile. Hence, net positions have
been decreasing faster than gross positions.
4.2 Censoring in the DTCC Data
For single-name CDS, the DTCC provides weekly position data (gross and net notional) for the top
1,000 traded reference entities in terms of aggregate gross notional amounts outstanding. This implies
that there is a censoring issue in the data: we do not observe CDS positions for rms that have gross
notional amounts outstanding that are too small to make it into the top 1,000 reference entities.
The censoring issue is illustrated in Figure 4. The gure plots the logarithm of net notional
amounts in CDS outstanding as a function of log assets. The gure plots the reference entities for
which we have CDS position data from the DTCC and censored observations, for which we do not
have CDS positions. In the gure we set censored observations to the log of the minimum contract
size of $10m. The censored observations either do not make it into the top 1,000 reference entities,
or are not traded entities in the CDS market. For some of these censored reference entities we know
that a CDS market exists because CDS quotes are available in the same month on Bloomberg. Hence,
we know that these rms have gross notional amounts outstanding that lie below this cuto. In our
16
empirical analysis, we thus use a censored regression approach, since otherwise our empirical results
would be biased (see, e.g., Wooldridge, 2010).
In adjusting for the censoring bias, an additional complication arises. Our analysis focuses not on
gross notional outstanding, but the net notional outstanding, because, as explained above, the net
notional outstanding is an economically more meaningful quantity. The DTCC, however, determines
the cuto as to which reference entity makes the top 1,000 list in terms of gross notional. Of course, the
resulting censoring problem carries over to net notional values: Reference entities that have low gross
notional amounts of CDSs outstanding, are also likely to have low net notional amounts outstanding.
Hence, because of the cuto in terms of gross notional outstanding, our data is also likely to leave out
reference entities with small amounts of net notional CDS outstanding. However, because the DTCC
cuto is in gross notional, in adjusting for this bias we have to make an assumption on the relation
between gross notional and net notional amounts of CDS outstanding.
We make this adjustment by exploiting the empirical relationship between gross notional and net
notional amounts. This relationship is illustrated in Figure 5. The Figure plots a sample histogram
of the net notional divided by the gross notional outstanding in a given observation month. From
Figure 5, we see that on average net notional amounts outstanding that are roughly 10% of the gross
notional amount outstanding. However, the gure also shows that there is signicant variation in this
relationship across reference entities. To make the adjustment from gross notional to net notional we
assume that for companies that are left out of our data because their gross notional amount outstanding
are too small, the relationship between gross notional and net notional amounts has the same mean
and standard deviation as the empirical relationship in the same month, which is illustrated in Figure
5.
4.3 Regression Specication
In our regression specication we focus on the log of the net notional amount of CDS outstanding,
y
i
= log (net cds
i
) as our main left-hand side variable. Our regression specication is thus given by
y
i,t
= log (Net CDS
i,t
) = X
i,t
+
i,t
, (1)
17
where the vector X contains our explanatory variables and a constant, and is an error term.
Because of the censoring in the DTCC data, we cannot estimate Equation 1 via OLS, since this
would bias our results. We thus run a maximum likelihood estimation that corrects for the cuto in
the reference entities that we observe in the data. The likelihood function is constructed as follows.
We observe y
i
= log (Net CDS
i,t
) for all rms for which y
i,t
exceeds the threshold y
i,t
, where
y
i,t
= log (NetCuto
t
) = log(GrossCuto
t
) +
n2g,t
, (2)
where
n2g
is the average observed log(net/gross) in a given month. Given this observation specic
censoring cuto, we can write the likelihood function as:
L
t
=
n

i=1
_
1

_
y
i,t
X
i,t

__
d
i,t

_
_

_
_
y
i,t
X
i,t
_

2
+
2
n2g,t
_
_
_
_
1d
i,t
, (3)
where d
i,t
is an indicator for observing net notional CDS outstanding,
d
i,t
=
_

_
1 if y
i,t
y
i,t
0 if y
i,t
< y
i,t
.
(4)
X is a vector that contains our explanatory variables and a constant. is the pdf of the standard
normal distribution, and the cdf of the standard normal distribution. This specication assumes
that NetCuto is given by the average ratio of net to gross for a given date t plus an error term with
the standard deviation equal to the standard deviation of net to gross ratios on that date. In the
specication, this additional error term is captured by
n2g
, the standard deviation of the observed
log(net/gross) ratio in a given month.
A reference entity may be left out of our data for two reasons. Either the rm is a traded reference
entity in the CDS market but the gross notional amount of CDS outstanding is too small to make it
into the top 1,000 traded reference entities. Alternatively, the company may not be a traded reference
entity at all (i.e., no CDS market exists for this reference entity). This means that the coecients
resulting from specication (1) are driven both by the existence of a CDS market for a given rm,
and its size conditional on existence.
18
To disentangle these eects, we then run two separate regressions, one investigating the existence
of a CDS market, the other investigating the determinants of the size of net CDS positions conditional
on existence of a CDS market. To investigate the existence of CDS market we run a probit regression
to investigate which rm characteristics are determinants of the emergence of a CDS market for a
rm. Existence of a CDS market is dened as being in the DTCC dataset of the largest 1,000 reference
entities and/or having at least on CDS quote on Bloomberg in a given month. Coecients from this
regression should isolate, as much as possible, the determinants of CDS market existence. Saretto and
Tookes (2010) run a similar probit regression
We then run a censored regression conditional on existence of a CDS market to investigate the
determinants of the size of net CDS positions conditional on existence. Here we again run a censored
regression, but only using companies for which we know that a CDS market exists. In other words,
companies for which we know that a market exists because they are traded reference entities on
Bloomberg but for which we do not observe net notionals outstanding because they are not among
the largest 1,000 traded reference entities, are treated as censored observations. The coecients from
this specication should capture the determinants of the amount of net notional outstanding for a
reference entity, conditional on a CDS market existing.
In all of our regressions we control for time xed eects. This is important, because our sample
period contains parts of the nancial crisis, which means that some months, for example the months
after Lehmans failure, may be somewhat special. Time xed eects should correct for this. In
addition, we also control for rst-digit industry xed eects.
4.4 Regression Analysis
4.4.1 Baseline Regression
Table 2 reports our baseline specication, based on the DTCC data and Compustat. In this baseline
regression we broadly focus on three potential determinants of CDS market existence and position
sizes in the CDS market: (i) insurable interest (how many loans, bonds and other obligations does
a rm have outstanding?), (ii) credit quality (what is the credit quality of the rm?), and (iii) bond
market liquidity (how easy is it to trade a rms credit risk directly in the underlying bond?).
19
Table 2 is structured as follows. Column (1) presents the results of a censored regression using
all rated rms. As discussed above, coecients from this regression indicated both whether a certain
right hand side variable is associated with existence of a CDS market on a given reference entity, and
whether conditional on existence that right hand side variable is associated with larger net notional
CDS positions. Columns (2) and (3) disentangle this eect. Column (2) presents the results of a
censored regression conditional on a CDS market existing. In other words, we only use reference
entities for which we know that a CDS market exists, either because the reference entity is part of the
DTCC sample, or because CDS quotes are available on Bloomberg. Column (3) presents the results
of a probit regression that investigates the existence of CDS markets. Coecients from this regression
show which right hand side variables are associated with a CDS market existing for a given reference
entity. Columns (4)-(6) repeat the analysis in columns (1)-(3) for US rms with at least one bond
issue in Mergent FISD. Here we also include a number of additional regressors describing the bond
and stock market. Columns (7)-(9) repeat the analysis in columns (4)-(6) using a more conservative
subsample that excludes all companies which have bonds issued by or CDS written on subsidiaries.
Size and Insurable Interest. Let us start by investigating the censored regression that uses
all the data, shown in column (1). The regression output shows that rms with more assets tend to
have more net notional CDS outstanding: the coecient on log(assets) is positive and statistically
signicant when looking at the entire sample. Second, we see that rms with more debt outstanding
(even controlling for size) have more net notional CDS outstanding, as illustrated by the positive and
signicant coecient on log(debt). Both of these eects seem to be slightly concave, as illustrated by
negative coecients on the squared terms.
However, from column (1) we cannot tell whether these eects arise because log(assets) and
log(debt) are associated with more CDS outstanding, or whether they make it more likely that a
CDS market on a given reference entity exists. Columns (2) and (3) decompose these two eects.
This reveals that while log(assets) is a signicant determinant in the existence regression, once a CDS
market exists on a given reference entity, a higher value for log(assets) is not a statistically signi-
cant determinant of the net notional amounts outstanding. log(debt), on the other hand, is both a
signicant determinant of a CDS market existing and, conditional on the CDS market existing, it is
associated with higher net notional amounts outstanding in the CDS market.
20
One interpretation of the positive coecients on log(debt) in both specications is that insurable
interest (i.e., the amount of debt outstanding on a given reference entity) is a signicant determinant
both of whether a CDS market emerges on a given reference entity, and of the net notional amount
of credit risk traded in the CDS market if a CDS market exists. This is consistent with the view that
(at least some) debtholders use the CDS market to hedge their exposure.
Credit Risk. The next set of explanatory variables explores the eect of credit risk, as proxied
by credit ratings, on the net notional amount CDS outstanding. To investigate credit quality, we
include dummies for ratings classes, such as AA, A, BBB, etc. We use dummies for ratings buckets
rather than a numerical rating scale to allow for non-linear and non-monotonic eects of credit quality.
The regression does not include a dummy for BB, which is the benchmark case. We choose BB as
a benchmark since it is well populated and proves to be a useful benchmark. Second, we include
two regressors to capture changes in credit quality. rating change last year indicates the change in a
reference entitys rating over the preceding year in number of notches. For example, a value of one
indicates that the reference entity has been downgraded by one notch in the preceding year. Negative
values correspond to upgrades. In addition, lost inv. grad in last 5 years is a dummy variable that
takes value one for reference entities that are not currently investment grade, but were investment
grade at some point over the preceding ve years. We use ve years because it is the most typical
maturity of a CDS contract and it also allows us to have enough downgrades over time to produce
meaningful statistics.
The analysis for credit quality reveals two main eects. First, we see that rms tend to have more
CDS outstanding when they are closer to the investment grade/non investment grade cuto. The
coecients on the ratings dummies should be interpreted as eects relative to a BB benchmark rm.
Across the dierent specications in Table 2, the coecients on the ratings dummies are generally
increasing when we move from AA or higher towards the investment grade cuto BBB. The positive
coecient on BBB indicates that reference entities that are just above the investment grade cuto
tend to have more CDS outstanding than reference entities that are BB (the highest credit quality
that is not investment grade). Reference entities that are BB, B or CCC and CC or lower, on the other
hand, have on average fewer CDS outstanding than those just above the investment grade cuto, but
more than BB bonds: The coecients are generally positive, but smaller than the coecient on BBB.
21
Moreover, most of the ratings dummies below investment grade, while positive, are not statistically
signicant.
In addition to the level of a reference entitys rating, the change in the rating also matters. First,
the S&P rating change last year has a positive and signicant coecient both in the overall sample in
column (1) and (particularly so) conditional on a CDS market existing in column (2). This indicates
that a downgrade over the last year is associated with more net CDS outstanding. Second, the
positive and signicant coecient on lost inv. grad in last 5 years across all specications indicates
that reference entities that were investment grade at some point, but have lost their investment grade
rating, tend to have more CDS outstanding.
The positive coecient on S&P rating change last year and lost inv. grad in last 5 years may be
due either to investors adding credit protection in response to downgrades or, alternatively, investors
adding CDS protection in anticipation of downgrades. Table 8 investigates this by running the same
regression as before, but including both the rating change over the last 12 months and the rating
change over the coming 12 months. While for the overall sample (column (2)) the eect seems to be
driven by past rating changes, for rms in FISD, future downgrades are also associated with more
CDS outsanding.
Bonds Outstanding and Bond Market Liquidity. Specications (4)-(6) in Table 2 include a
number of additional control variables. First, we investigate whether, controlling for total debt out-
standing, the amount of bonds outstanding on a reference entity, log(bonds outstanding) is associated
with more net CDS outstanding. The positive and signicant coecients show that this is indeed the
case, and comparing columns (4) and (6), we see that the is driven both by the fact that CDS markets
are more likely to exist for companies that have more bonds outstanding and by larger CDS positions
conditional on existence.
In addition to the amount of bonds outstanding, we also see that bond turnover (last 12 months) is
associated with a higher likelihood that a company is a traded entity in the CDS market. This result
is consistent with the probit regression reported by Saretto and Tookes (2010). Their interpretation
is that higher bond turnover proxies for higher interest in trading the credit risk of a certain reference
entity. Finally, analyst disagreement: std/price and equity turnover (monthly) are associated with
more net CDS outstanding for traded reference entities, but are not (or only marginally) signicant in
22
the probit regression. Given that a CDS market exists, the signicant coecients on equity turnover
(monthly) and analyst disagreement: std/price are likely to be driven by trading demand. For example,
disagreement about a rms prospect generates both trading in the equity market and volume in the
CDS market. A similar reason is likely to drive the strongly positive coecient of equity turnover
(monthly) conditional on the existence of a market.
Next, we investigate the eect of how fragmented a rms bond issues are. We measure bond market
fragmentation at the issuer level by calculating a Herndahl index of bond market fragmentation. Of
course, rms that have more bonds outstanding are also likely to have a larger number of separate bond
issues. We thus adjust our Herndahl measure for the amount of bonds that a rm has outstanding
by taking the residuals of a regression of the log of the Herndahl index of bond issues on log total
bonds that this company has outstanding.
6
The adjusted log(bond Herndahl) in specications (4)-
(9) thus measures the fragmentation of a companys outstanding bonds, controlling for the overall
dollar amount of bonds issued and is a proxy for how easy it is to trade a rms bonds controlling for
the amount of bonds outstanding.
7
Table 3 documents that the adjusted Herndahl index is indeed a
reasonable measure of liquidity in the sense that it is a signicant determinant of bond turnover after
controlling for bonds outstanding and a number of other issuer characteristics.
The adjusted log(bond Herndahl) is signicant in the overall regression, as well as the regression
conditional on existence of a CDS market and the probit regression for existence. The strongly
negative coecient in the existence regression suggests that controlling for the amount of bonds that
a rm has outstanding, CDS markets are more likely to emerge when a rms bond issues are very
fragmented (more negative adjusted log(bond Herndahl)) and thus likely hard to trade. In addition,
the regression conditional on existence of a CDS market suggests that rms that are traded reference
entities in the CDS market tend to have more CDS outstanding if their bonds are very fragmented.
This nding supports the argument in Stulz (2009), who points out that rms have all sort of dierent
bonds whose prices are aected by call provisions, covenants, coupon, maturity, liquidity, and so on;
in contrast, CDS are like standardized bonds. Hence, the more fragmented and diverse a companys
6
We exclude companies with only a single bond issue from the adjustment regression, since having one bond issues might
reect a corner solution (lower bound on the number of bond issuances).
7
This measure is consistent with empirical evidence in Longsta et al. (2005) and Mahanti et al. (2008), who document
that bond issue size tends to be correlated with secondary market liquidity.
23
bond (which corresponds to a lower value for the rms adjusted log(bond Herndahl)), the more
attractive the CDS market may be as hedging or speculation instruments. This is also consistent with
the theoretical model of Vayanos and Wang (2007) which suggests that if there are multiple markets
with equivalent payos, liquidity and trading tend to concentrate in one. Clearly the CDS market with
its unifying feature across many bond issuances is an ideal candidate to become the liquid market,
where credit risk transfer takes place, instead of the relatively more segmented bond markets.
Industry Dummies. Finally, Table 2 investigates a number of industry dummies, two of which
are interesting. First, there is a large amount of net notional CDS protection written on companies
that provide credit enhancement. These include monoline insurers and other insurance companies.
8
Most likely, there are more CDS outstanding on companies that provide credit enhancement because
investors who rely on insurance from monoline insurance companies and other providers of credit
enhancement purchase CDS in order to eliminate their counterparty risk (for a model where CDS are
used to insure counterparty risk, see, e.g., Zawadowski (2011)). In these cases, the protection provided
by credit enhancement rms represents an insurable interest that purchasers of this insurance may
want to hedge in the CDS market. Anecdotal evidence for such behavior is given by the report of the
Financial Crisis Inquiry Commission:
9
in 2007-2008 Goldman Sachs reportedly bought CDS protection
on AIG after buying substantial amounts of under-collateralized OTC derivatives on subprime housing
from AIG. Once we move to the US sample of FISD rms, the credit enhancement dummy is dropped
because only a few rms providing credit enhancement have traded stock, most are subsidiaries.
Second, the negative coecient on the nancial industry dummy indicates that, on average, there
are less CDS outstanding on nancial companies than on non-nancial companies. Mostly, this eect
seems to be driven by smaller, non-systemic nancial institutions. This can be seen by the positive
coecient on the systemic nancial institution dummy which basically cancels out the negative eect
of the nance dummy. This dummy takes value one if a certain reference entity belongs to the top 30
systemic nancial institutions as ranked by NYU-Sterns mean expected shortfall measure. However,
8
The list of companies we categorized as providing credit enhancement are: AMBAC, MBIA, Primus Guaranty, Triad
Guaranty, Assured Guaranty, XL Group, Radian Group, ACE, Berkshire Hathaway, PMI Group, AIG.
9
See the supporting document of The Financial Crisis Inquiry Report compiled by the FCIC describing the timeline
of Goldmans hedges with AIG downloadable from http://fcic-static.law.stanford.edu/cdn media/fcic-testimony/2010-0701-
Goldman-AIG-Collateral-Call-timeline.pdf
24
conditional on a CDS market existing, the systemic dummy is not signicant.
10
Finally, looking at
the entire sample we nd that service companies are less likely to be traded reference entities in the
CDS market, as indicated by the negative coecient on the service dummy in the probit regression,
and have fewer CDS outsanding conditional on existence.
4.4.2 Hedging, Speculation and Liquidity
In this Section we investigate the uses of CDS. In particular, we are interested in whether CDS are
used for hedging purposes, speculation, or both. As pointed out in Section 2, on the one hand investors
who hold a reference entitys bonds or are otherwise exposed to the reference entitys credit risk may
use CDS to partially or fully hedge this exposure. Alternatively, investors may use CDS to take bets.
They may purchase naked CDS protection on a reference entity to prot from future deterioration
in the reference entitys credit quality, or they may sell CDS protection to other investors such that
they prot from future improvements in the reference entitys credit quality. We expect these types of
speculative trades to be particularly pronounced when there is more disagreement about the reference
entities credit prospects.
However, both hedging and speculation need not necessarily involve trading in CDSs. A trader
that wants to rid himself of the exposure to the credit risk inherent in a particular bond can also simply
sell the bond. A speculator who wants to bet that a reference entitys credit is going to deteriorate
can simply take a short position in the bond. The eects of hedging and speculation on CDS markets
should thus be related to liquidity in the reference entitys bond(s).
In fact, the analysis in the previous section suggested that both hedging needs and speculation
are signicant determinants of CDS positions. For example, the signication coecients on variables
that proxy for insurable interest point at hedging needs as a determinant of CDS positions. The
nding that disagreement about earnings potential (and thus credit prospects) is associated with
higher net notional CDS positions outstanding in the CDS market, on the other hand, suggests that
CDS positions to some extent be driven by speculation or taking views on default probabilities.
However, because it is possible to take views directly in the underlying or, alternatively, through
10
We use the list as of September 2008, the month before our DTCC sample starts. For details, see
http://vlab.stern.nyu.edu/analysis/RISK.WORLDFIN-MR.GMES.
25
the CDS market, one would expect that the choice of the instrument that is used for speculation
depends on how expensive it is to trade in the bond market relative to the CDS market. In Table ??,
we investigate this potential eect of the liquidity of the underlying bond on the incentives to use the
CDS market in order to hedge or speculate.
Liquidity Measures. In order to investigate the impact of bond market illiquidity, we use three
alternative measures for bond market liquidity to group companies into three liquidity buckets. Our
rst liquidity measure is the number of trades in a companys bond in a given year. Given that the
number of trades is relatively low for many companies, this measure is similar to that of the number
of zero return days measure used in other studies, such as Chen et al. (2007) and references therein.
Our second liquidity measure is the annual turnover in the rms bonds, which is constructed using
the TRACE bond market data. Our third liquidity measure is the adjusted log(bond Herndahl)
discussed in the previous section. This proxy is based on the hypothesis that, controlling for the
amount of bonds outstanding, a more fractured bond market is less liquid. We calculate these three
bond liquidity measures for each US rm with at least one bond issue in the Mergent FISD dataset
and then sort rms into three terciles according to the liquidity of their bonds.
Speculation and Liquidity. Table 4 investigates speculation and liquidity. The results indicate
that disagreement, our proxy for speculation, is a stronger determinant of the net notional amount
of outstanding CDS for rms whose bonds are not particularly liquid. Specically, column (2) shows
that conditional on existence of a CDS market, increased analyst forecast dispersion is associated with
larger net notional amounts of CDS when the underlying bond has low or medium liquidity than when
it has high liquidity. This is indicated by the statistically signicant and positive coecients on low
liquidity * disagree or medium liquidity * disagree relative to the benchmark case of high liquidity.
The same broad pattern emerges for all three liquidity measures (although when using the Herndahl
index, the eect is strongest for the medium liquidity bucket).
This nding conrms the intuition, that more disagreement should go hand in hand with more bets
being taken in the CDS market when trading in the underlying bond is costly. This is consistent with
the interpretation that when a rms bond market is liquid, investors are more likely to express their
views directly through the bond market, while when the bonds are illiquid and trade infrequently,
increased disagreement among investors manifests itself as bets in the CDS market. This corroborates
26
our result from the previous section, which suggested that CDS markets are more likely to emerge
and have larger positions when the bonds of the underlying reference entity are hard to trade.
Hedging and Liquidity. Table investigates how the CDS markets role as a hedging tool corre-
lates with the liquidity of the reference entitys bonds. In order to proxy for the eect of hedging, we
look at reference entities that lose investment grade status. The reason is that as a companys debt
becomes speculative grade, investors may want, or, in the case of some institutional investors, even
be required to hedge or ooad their exposure.
Recall that in the previous section we found that rms that have lost investment grade status
have more CDS outstanding. This is indicative of investors reducing their exposure to those rms by
purchasing CDS protection, either in anticipation of the downgrade or in response to the downgrade.
We can now use our liquidity proxies to investigate how this eect depends on how easy it is for
investors to trade the underlying bond. As with the eects of disagreement on CDS outstanding, we
nd that this eect is again concentrated in rms with less liquid bonds.
In particular, while losing investment grade is generally associated with more CDS outstanding
(the coecient on lost inv. grade in last 5 years is positive and signicant across all specications),
the eect of losing investment grade is stronger for rms with less liquid bonds. This is illustrated
by positive and signicant coecient on the the interaction term low liquidity * 5yr lost IG across all
dierent liquidity proxies. This nding is consistent with the view that investors who wish to reduce
their exposure after a downgrade tend to sell their bonds when the market for the bond is liquid,
while they tend to use CDS to hedge their exposure when the market for the downgraded rms bond
is illiquid.
This nding sharpens our results on credit quality from the previous section. Consider Figure 6,
which shows the dummy coecients for dierent rating categories and the additional eect of having
lost investment grade status. The net notional amount of CDS outstanding increases as the credit
rating deteriorates up to the investment grade cuto. Once below the investment grade cuto, on the
other hand, it is clearly lower than right above the investment grade cuto. This echoes the results
on credit quality in 2.
However, if a rm used to be investment grade status and has been downgraded to non-investment
grade, it tends to have more CDS outstanding. Figure 6 illustrates this additional eect for each
27
non-investment grade ratings category by adding the interaction term between the underlying bonds
liquidity and lost investment grade status to the unconditional ratings dummy. This shows that
conditional on having lost investment grade status, rms with more liquid bond markets continue to
have a higher level of CDS, comparable to that of investment grade BBB rms. Hence, rms with
liquid bonds that used to be investment grade continue to have more CDS outstanding even after
losing investment grade status, but there is no clear increase in CDS because of having lost investment
grade. On the other hand, for rms with illiquid bonds, losing investment grade is associated with an
increase in the amount of CDS outstanding relative to comparable rms right above the investment
grade cuto. This may be the case as for rms with illiquid bonds investors choose to hedge their
positions in the CDS market when a rm loses investment grade status, resulting in an increase in the
net notional amount of CDS outstanding.
Tables 4 and 5 also conrm our earlier nding that the liquidity of the underlying bonds have an
unconditional eect on the amount of CDS outstanding. This can be seen by inspecting the dummy
coecients on medium liquidity and low liquidity for our three liquidity measures. What is interesting
is that the coecients on the low liquidity dummy are generally positive for the adjusted log(bond
Herndahl), while they are negative when using bond turnover (last 12 months) or the number of
bond trades (last 12 months). The negative coecient on low liquidity as measured by bond turnover
seemingly contradicts the hypothesis that rms with less liquid bond markets should unconditionally
have more net CDS. One possible explanation is that bond turnover is aected by two factors: bond
liquidity and trading demand. If bond turnover also proxies for trading demand in the bond market,
this might spill over to the CDS market as suggested by Saretto and Tookes (2010). Hence, higher
bond turnover may be associated with higher trading demand in the CDS market, explaining the
negative coecient for the low liquidity dummy when using bond turnover as a liquidity proxy. The
coecient on the adjusted log(bond Herndahl) liquidity measure supports this interpretation: the
adjusted log(bond Herndahl) is unlikely to be aected by trading demand and should thus be a purer
measure of bond liquidity. Indeed, looking at this measure we nd a positive coecients on medium
liquidity and an even higher coecient on low liquidity. Hence, when measuring liquidity by the
adjusted log(bond Herndahl), companies with less liquid bonds tend to have more CDS outstanding,
suggesting that the CDS acts as a unied bond market for those companies. Finally, Tables 4 and 5
28
also conrm the eect log(bonds outstanding) as a signicant determinant of the existence of a CDS
market and the amount of CDS outstanding conditional on existence.
4.4.3 Which Type of Debt Matters?
Table 6 investigates whether the mix of debt outstanding aects the net notional amount of CDS
outstanding, using detailed balance sheet data from Capital IQ. This data allows us to split a rms
debt obligations into ner categorizations such as bonds, commercial paper, capital leases, revolving
credit, term loans, trust preferred borrowing, etc. Additionally we use accounts payable data from
Compustat. Because balance sheet information from Capital IQ is only available for a fraction of the
companies in our dataset, we can only run this analysis using a smaller sample. Also, the data is
annual, which means that we only have two time series observations: 2008 and 2009.
Table 6 uses two dierent specications. The rst specication uses a log-log setup, meaning that
we regress the log of the net notional amount of CDS outstanding on the log of the dierent balance
sheet items. In the second specication we regress the net notional amount of CDS outstanding
normalized by total assets on the dierent balance sheet times normalized by assets. This second
specication has the advantage that coecients are more easily interpretable as the dollar amount of
net CDS for every additional dollar of borrowing of a certain type. As we can see from the summary
statistics in Table 1, the three main sources of borrowing for rms are term loans, accounts payable,
and bonds.
The analysis points to three main eects on the amount of outstanding CDS. First, a higher ratio
of bonds relative to assets is associated with more CDS outstanding. This conrms the positive
eect of log(bonds outstanding) in Table 2. First, this eect is present when looking at all rms, as
documented in columns (1). Moreover, the eect seems to be driven both by larger CDS positions
conditional on existence of a CDS market (see column (2)), and a higher likelihood that the issuing
form is a traded reference entity in the CDS market (see column (3)).
Second, rms with higher accounts payable relative to assets have more CDS outstanding. The
eect of accounts payable on net CDS outstanding allows for two possible interpretations. First, it
may be driven by counterparty risk (e.g., suppliers to rms may be hedging their exposure in accounts
payable by purchasing CDS on the rm). This is consistent with the notion that CDS are used to
29
hedge non-bond receivables discussed in Stulz (2010). Another interpretation is that accounts payable
may, at least in some instances, be a proxy for nancial distress: accounts payable may be high
for rms that are struggling to make payments to suppliers and other counterparties, and may thus
signal dire straits. In this second interpretation, the positive relation between accounts payable and
net CDS results from investors who are hedging exposure to, or speculating against, distressed rms.
Note however, that we control for rm rating in the regression so this second interpretation holds only
if assets payable reects distress not incorporated in the credit rating.
Third, we nd that rms with more term loans relative to assets on average have fewer CDS
outstanding. In both specication, this eect arises because more term loans are associated with a
smaller probability that a CDS market exists in the rst place. However, conditional on existence
of a CDS market, term loans are not a signicant determinant of CDS positions. This is somewhat
counterintuitive as one would expect term loans, such as bank lending, to behave in exactly the same
way as other types of insurable interest, as argued in Stulz (2010). First, Minton et al. (2009) also
nd that few banks use CDS to insure loans and even if they do so, they insure only a small fraction:
2 cents for each dollar of lending. Second, the negative coecient might be due to rms with term
loans being safer, e.g. because they face less rollover risk and thus there is less demand for CDS.
4.4.4 The CDS-Bond Basis
One quantity that has received considerable attention since the nancial crisis is the CDS-bond basis.
The CDS-bond basis is dened as the CDS spread minus the yield above Treasury (or risk-free rate)
of the underlying corporate bond (also known as the Z-spread). No arbitrage implies that the CDS-
bond basis should be approximately zero. The reason is that a portfolio consisting of a long bond
position and a CDS that insures the default risk of the bond should yield the risk-free rate. While the
CDS-bond basis should be exactly equal to zero only if certain assumptions hold (see Due, 1999),
absent limits-to-arbitrage frictions it should be approximately zero in practice.
11
During the recent nancial crisis, the CDS-bond basis became signicantly negative for many
reference entities as documented, for example, by Bai and Collin-Dufresne (2010) and Fontana (2011).
11
In practice, the CDS-bond basis has historically been slightly positive for technical reasons, such as imperfections in the
repo market and the cheapest-to-deliver option (see JPMorgan, 2006).
30
A negative CDS-bond basis means that the CDS spread is lower than the spread over the risk-free
rate on the underlying bond. Intuitively speaking, this implies that one can earn a higher spread on
the bond than it costs to insure the default risk of the bond in the CDS market. This gives rise to
the so-called negative basis trade, in which a trader who seeks to prot from a negative basis buys
the underlying bond and purchases credit protection on the bond in the CDS market. Because the
arbitrage trade involves a long position in the CDS, if arbitrageurs seek to prot from a negative
CDS-bond basis, such a negative basis should be associated with larger net notional CDS positions
outstanding.
We investigate the role of the CDS-bond basis on net notional amounts outstanding in the CDS
market in Table 7. In order to do this, we run a similar regression as before, but include the CDS-bond
basis as a right hand side variable. We use the CDS-bond basis as calculated by Bloomberg. We are
careful to eliminate bonds with embedded options (puttable, callable, redeemable), and end up with
data on the CDS-bond basis data for 78 companies.
The results in Table 7 show that a negative basis is associated with more net notional in CDS
outstanding, as predicted by the arbitrage trade required to prot from a negative basis. Specication
(1) examines the eect of the CDS-bond basis using our baseline regression (comparable to columns (1)-
(3) in Table 2). The signicant negative coecient on the CDS-bond basis means that, unconditionally,
a negative basis is associated with larger net notional CDS amounts outstanding.
On the one hand, this result provides evidence that a more negative CDS-bond basis is associated
with larger positions in the CDS market. However, note that this result is not obvious: also a positive
CDS-bond basis gives rise to a trading opportunity that involves trading in the CDS and should, in
theory, thus also be associated with larger net notional CDS positions. The negative coecient on the
CDS-bond basis in column (1) thus suggests that positions in the CDS market react more strongly to
a negative basis than to a positive basis. We explore this in columns (2) and (3), where we investigate
negative and positive CDS bases separately. Column (3) also includes a number of additional controls,
such as the adjusted bond Herndahl, bonds outstanding, and equity turnover.
The results in Column (2) and (3) indicate that the eect of the CDS-bond basis is indeed asym-
metric. While a negative CDS-bond basis is associated with a statistically signicant increase in net
notional CDS outstanding, the coecient on the positive CDS-bond basis is essentially zero. While
31
this could partially be driven by lack of data (during our sample period, signicant positive CDS-bond
bases are rare), the result is suggestive of an asymmetry between negative and positive CDS-bond
bases. In particular, proting from a positive CDS-bond basis requires short-selling the bond, which
is often dicult and costly. Trading against a negative CDS-bond basis, on the other hand, does
not require short-selling the bond. This may explain why arbitrageurs trade less aggressively against
a positive CDS-bond basis. Consistent with this interpretation, Blanco et al. (2005) argue that the
diculty of shorting bonds may be one of the reasons why during normal times (i.e., prior to the
nancial crisis) the CDS-bond basis has usually been slightly positive.
To the extent that the signicant coecient on the negative CDS-bond basis reects arbitrage
activity, this points at a potential economic eect of CDS markets. If via the negative basis trade
the presence of CDS can help to (partially) eliminate mispricings in the bond market and potentially
compress spreads for bond issuers, this may improve rms access to nancing. This interpretation
echoes the arguments in Saretto and Tookes (2010), who maintain that the presence of CDS allows
rms to borrow more and at longer maturities.
5 Robustness
We have performed a number of robustness checks to corroborate our main results. For example,
we replicated our results using alternative measures for disagreement. Specically, we replace our
main proxy for disagreement (the dispersion of analyst forecasts scaled by price) with two alternative
measures (the dispersion of analyst forecasts scaled by their mean, equity turnover). Our results remain
qualitatively unchanged. We also rerun our regressions explicitly controlling for membership of certain
reference entities in the major main CDS indices, specically the CDX.NA.IG (the investment grade
CDX index) and the CDX.NA.HY (the high yield CDX index). Also in this case, the results remain
qualitatively the same and remain signicant. This specication also shows that reference entities that
are in the CDX have more net notional amounts in CDS outstanding than reference entities that are
not in the CDX. This is to be expected for two reasons. First, reference entities with more outstanding
CDS may be more likely to be selected into the CDX. Second, once part of the CDX, reference entities
may become more heavily traded in the CDS market. Finally, we also check that our main results
32
also carry through when we look directly at the net notional amount of CDS as our left-hand side
variable, as opposed to log(net CDS), which we used in our main specication. For brevity, we do not
report these robustness checks in the paper.
6 Conclusion
This paper investigates the determinants of the amount credit protection bought (or equivalently
sold) in the market for credit default swaps (CDS). Combining data on net notional CDS positions
outstanding from the Depository Trust & Clearing Corporation (DTCC) with a number of other
data sources, our results broadly suggest that CDS markets function as alternative trading venues for
both hedging and speculation on the underlying bond. They are alternative trading venues in the
sense that investor could often make the same economic trade directly in the underlying bond, but
choose to use the CDS market. This interpretation of CDS markets as alternative trading venues
is supported by the nding that CDS markets more likely to emerge when the underlying bonds of
the rm are illiquid and thus hard or expensive to trade. Similarly, conditional on a CDS market
existing for a reference entity, the positions taken in the CDS market are larger when the underlying
bond is illiquid. We also document that CDS positions are increasing in insurable interest (a proxy
for hedging needs) and disagreement (a proxy for speculation). These eects are stronger when the
underlying bond is hard to trade. Finally, rms which have a more negative CDS-bond basis (i.e., the
bond is undervalued relative to the CDS) have more CDS outstanding, suggesting that arbitrageurs
lean against the negative CDS-bond basis. Through the negative basis trade the presence of CDS may
help eliminate mispricings in the bond market and compress spreads for bond issuers, thus improving
rms access to nancing.
References
Acharya, Viral V. and Timothy C. Johnson, Insider Trading in Credit Derivatives, Journal
of Financial Economics, 2007, 84 (1), 110141.
33
Ashcraft, Adam B. and Joao A. C. Santos, Has the CDS Market Lowered the Cost of Corporate
Debt, Journal of Monetary Economics, 2009, 56 (4), 514523.
Bai, Jennie and Pierre Collin-Dufresne, The Determinants of the CDS-Bond Basis During the
Financial Crisis of 2007-2009, 2010. Working Paper, Columbia University.
Blanco, Roberto, Simon Brennan, and Ian W. Marsh, An Empirical Analysis of the Dynamic
Relation between Investment-Grade Bonds and Credit Default Swaps, The Journal of Finance,
2005, 60 (5), pp. 22552281.
Bolton, Patrick and Martin Oehmke, Credit Default Swaps and the Empty Creditor Problem,
Review of Financial Studies, 2011, 24 (8), 26172655.
Che, Yeon-Koo and Rajiv Sethi, Credit Derivatives and the Cost of Capital, 2011. Working
Paper, Columbia University.
Chen, Long, David A. Lesmond, and Jason Wei, Corporate Yield Spreads and Bond Liquidity,
The Journal of Finance, 2007, 62 (1), 119149.
Choi, Jaewon, Dirk Hackbarth, and Josef Zechner, Granularity of Corporate Debt: Theory
and Tests, 2012. Working Paper, University of Illinois.
DTCC, Deriv/SERV Today, October 2009.
Duee, Gregory R. and Chunsheng Zhou, Credit derivatives in banking: Useful tools for
managing risk?, Journal of Monetary Economics, 2001, 48 (1), 25 54.
Due, Darrell, Credit Swap Valuation, Financial Analyst Journal, 1999, 55 (1), 7387.
Fontana, Alessandro, The Negative CDS-bond Basis and Convergence Trading during the 2007/09
Financial Crisis, 2011. Working Paper, Geneva Finance Research Institute.
Geanakoplos, John and Ana Fostel, Tranching, CDS and Asset Prices: How Financial Innovation
can Cause Bubbles and Crashes, 2011. Working Paper, Yale University.
34
Hirtle, Beverly, Credit Derivatives and Bank Credit Supply, Journal of Financial Intermediation,
2009, 18 (2), 125150.
Hu, Henry T. C. and Bernard Black, Hedge Funds, Insiders, and the Decoupling of Economic
and Voting Ownership: Empty Voting and Hidden (Morphable) Ownership, Journal of Corporate
Finance, 2007, 13, 343367.
JPMorgan, Credit Derivatives Handbook, 2006.
Lee, Dongyoup, The Information in Credit Default Swap Volume, 2011. Working Paper, Columbia
University.
Longsta, Francis A., Sanjay Mithal, and Eric Neis, Corporate Yield Spreads: Default Risk
or Liquidity? New Evidence from the Credit Default Swap Market, Journal of Finance, 2005, 60
(5), pp. 22132253.
Mahanti, Sriketan, Amrut Nashikkar, Marti Subrahmanyam, George Chacko, and Gau-
rav Mallik, Latent liquidity: A new measure of liquidity, with an application to corporate bonds,
Journal of Financial Economics, 2008, 88 (2), 272 298.
Minton, Bernadette, Rene M. Stulz, and Rohan Williamson, How Much Do Banks Use
Credit Derivatives to Hedge Loans, Journal of Financial Services Research, 2009, 35 (1), 131.
Nashikkar, Amrut, Marti G. Subrahmanyam, and Sriketan Mahanti, Liquidity and Ar-
bitrage in the Market for Credit Risk, Journal of Financial and Quantitative Analysis, 2010.
(forthcoming).
Parlour, Christine A. and Andrew Winton, Laying o Credit Risk: Loan Sales versus Credit
Default Swaps, 2008. Working Paper, UC Berkeley.
and Guillaume Plantin, Loan Sales and Relationship Banking, Journal of Finance, 2008, 63
(3), 12911314.
Qiu, Jiaping and Fan Yu, Endogenous liquidity in credit derivatives, Journal of Financial Eco-
nomics, 2012, 103 (3), 611 631.
35
Saretto, Alessio and Heather Tookes, Corporate Leverage, Debt Maturity and Credit Default
Swaps: The Role of Credit Supply, 2010. Working Paper, Yale University.
Shachar, Or, Exposing the Exposed: Intermediation Capacity in the Credit Default Swap Market,
2011. Working Paper, NYU Stern.
Stulz, Rene M., Credit Default Swaps and the Credit Crisis, 2009. NBER Working Paper 15384.
, Credit Default Swaps and the Credit Crisis, Journal of Economic Perspectives, 2010, 24 (1),
7392.
Thompson, James R., Credit Risk Transfer: To Sell or to Insure, 2007. Working Paper, University
of Waterloo.
Vayanos, Dimitri and Tan Wang, Search and endogenous concentration of liquidity in asset
markets, Journal of Economic Theory, 2007, 136 (1), 66 104.
Wooldridge, Jerey M., Econometric Analysis of Cross Section and Panel Data (2nd edition),
MIT Press, 2010.
Zawadowski, Adam, Entangled Financial Systems, 2011. Working Paper, Boston University.
36
Example (a): Gross and net notional positions
A
Example(a):Grossandnetnotionalpositions
B C
10m
B
Gross CDS bought Gross CDS sold Net CDS
C
GrossCDSbought GrossCDSsold NetCDS
A 0 10 (10)
B 10 0 10
C 0 0 0
Total Gross Notional
Bought=10
GrossNotional
Sold =10
NetNotional
Bought/Sold=10
Example (b): Gross and net notional positions
A
Example(b):Grossandnetnotionalpositions
B C
10m
B
Gross CDS bought Gross CDS sold Net CDS
C
10m
GrossCDSbought GrossCDSsold NetCDS
A 0 10 (10)
B 10 10 0
C 10 0 10
Total Gross Notional
Bought=20
GrossNotional
Sold =20
NetNotional
Bought/Sold=10
Example (c): Gross and net notional positions
A
Example(c):Grossandnetnotionalpositions
B C
10m 10m
B
Gross CDS bought Gross CDS sold Net CDS
C
10m
GrossCDSbought GrossCDSsold NetCDS
A 10 10 0
B 10 10 0
C 10 10 0
Total Gross Notional
Bought=30
GrossNotional
Sold =30
NetNotional
Bought/Sold=0
Figure 1: Gross Notional vs. Net Notional Amounts
The gure illustrates the dierence between gross notional and net notional amounts in the DTCC data. In Example (a), B has purchased $10m
in protection from A. Both the gross notional and the net notional amount outstanding are $10m. In Example (b), B osets the initial trade by
selling $10m in protection to C. This raises the gross notional amount to $20m. The net notional amount remains at $10m. In Example (c), B
sells $10m in protection to A, such that all three parties have a net zero position. The gross notional is now $30m, but because all net positions
are zero, the net notional is $0.
37
0
4
0
0
8
0
0
1
2
0
0
1
6
0
0
2008m7 2009m1 2009m7 2010m1 2010m7 2011m1
time
total single name net CDS (billions)
top 1000 single name net CDS (billions)
single name net CDS in sample (billions)
Figure 2: Single-name CDS net notional outstanding over time
The top solid line plots the total amount of net CDS outstanding on all single-name reference entities, as reported by DTCC. It thus captures the
net notional outstanding in the entire single-name CDS market. The dashed line below the solid line is the net total net notional in CDS protection
written on the top 1,000 single name entities. Comparing this line to the total single-name CDS market demonstrates that the top 1,000 reference
entities make up almost the entire single-name CDS market when measured in terms of net notional outstanding. The dotted line plots the total
net notional amounts of outstanding CDSs for the reference entities that we managed to match with Compustat.
5
1
0
1
5
2
0
2008m7 2009m1 2009m7 2010m1 2010m7 2011m1
time
gross/net CDS (median) gross/net CDS (10th percentile)
gross/net CDS (90th percentile)
Figure 3: Gross to net CDS over time
The gure plots the ratio of gross to net amounts in CDS outstanding over time. While overall positions in CDS markets decline over our sample,
the ratio of gross to net increases. This is true for the median ratio of gross to net, but also for the 10th and 90th percentile. Hence, net positions
have been decreasing faster than gross positions.
38
-
5
-
4
-
3
-
2
-
1
0
1
2
l
o
g
(
n
e
t
)
-2 0 2 4 6 8
log(assets)
in DTCC and Bloomberg censored, no Bloomberg
in DTCC, no Bloomberg censored, in Bloomberg
Figure 4: Net CDS and rm size
This gure illustrates the censoring and market existence issues in our data for Compustat rms rated by S&P for December 2009. The shape of
the marker indicates whether we can locate a certain reference entity (i) as one of the 1,000 largest reference entities as provided by the DTCC
and (ii) as a reference entity for which we can nd at least one price quote in Bloomberg. For reference entities that we observe in the DTCC data
set and/or for which we can nd Bloomberg quotes we know that a CDS market exists. If a reference entity is in Bloomberg but not in DTCC we
know that it is censored, i.e., a CDS market exists but we do not observe the net notional outstanding. Reference entities that we cannot locate in
Bloomberg or DTCC are censored and we do not know whether a CDS market exists for those reference entities. Censored reference entities have
been set to the log of the minimum contract size of $10m, i.e., log(0.01) 4.61
0
0
05
5
510
1
0
10 15
1
5
15 Density
D
e
n
s
i
t
y
Density 0
0
0.1
.1
.1 .2
.2
.2 .3
.3
.3 .4
.4
.4 .5
.5
.5 net CDS / gross CDS
net CDS / gross CDS
net CDS / gross CDS
Figure 5: Gross Notional vs. Net Notional Amounts
This gure plots the empirical density of the ratio of the net notional amount outstanding to the gross notional amounts outstanding for single-name
CDS contracts in the DTCC data.
39
-0.2
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
AAorhigher A BBB BB B CCCorlower
didnotloseIG
lostIG,highliquidity
lostIG,medliquidity
lostIG,lowliquidity
Figure 6: Eect of ratings and losing investment grade status
The gure depicts the dummy variables on ratings for the regression in column (8) of Table 5. The black bars are the unconditional dummy
coecients. The other three bars are the coecients conditionally on having lost investment grade split by the liquidity of the bond market based
on the adjusted Herndahl measure.
40
Table 1: Summary Statistics
This table presents summary statistics for monthly data from October 2008 to December 2010 for all Compustat companies above 100 million
of assets and rated by S&P. assets is total assets and debt is total long and short-term debt from Compustat. Quarterly data from Compustat
is converted into monthly. net CDS, gross CDS, and number of CDS are the net and gross notional amount, and number of CDS outstanding
respectively as reported by the DTCC. The last weekly DTCC observation each month is used. accounts payable is from Compustat and set
to zero for nancials. S&P rating (notch) captures a rms S&P rating; it takes value 1 for AAA, 2 for AA+ etc. The maximum value, 22,
indicates that the bond has defaulted. S&P rating change is dened as the number of notches by which a bond was upgraded (negative values) or
downgraded (positive values) over the last 12 months. When S&P rating takes a value between 1 and 10, the corresponding bond is investment
grade. lost investment grade last 5 years is an indicator variable that takes value one if a company lost its investment grade rating in the last 5
year. number of trades (last 12 months) is total number of trades registered in Trace for all bonds in Mergent FISD of a specic company over
the preceding 12 months. annual bond turnover (last 12 months) is total bond trading volume from Trace for all bonds of a company in Mergent
FISD divided by total bonds outstanding over the preceding 12 months. analyst disagreement: std/price is dened as the standard deviation
of analyst earnings forecast from IBES normalized by stock price. analyst disagreement: std/abs(mean) is dened as the standard deviation of
analysts earnings forecasts normalized by the mean analyst forecast. commercial paper, other borrowing, capital lease, revolving credit, term loans
and trust preferred borrowing are annual data (2008 and 2009) from the detailed balance sheet data from Capital IQ and normalized by total
assets from Compustat. issuer bonds outstanding is the total amount of bonds outstanding of the issuing entity in from Mergent FISD, while
consolidated bonds outstanding is the total amount of bonds issued by all entities belonging to the same ultimate parent. All dollar amounts in
billions, ratios winsorized at the 1% level.
All Compustat rms rated by S&P:
VARIABLES N mean std p10 p50 p90
net CDS (USD billions) 14,784 1.192 1.123 0.334 0.891 2.285
gross CDS (USD billions) 14,786 14.63 14.23 2.831 10.20 30.33
number of CDS 14,786 2,185 1,580 551 1,840 4,257
net CDS / assets 14,784 0.0799 0.118 0.00695 0.0378 0.199
net CDS / debt 14,680 0.286 0.486 0.0268 0.133 0.646
assets (USD billions) 48,634 40.03 190.4 0.809 5.133 50.45
book leverage 48,634 0.354 0.251 0.0818 0.321 0.636
accounts payable / assets 48,634 0.0591 0.0693 0 0.0390 0.143
S&P rating (notch) 48,634 10.50 3.713 6 10 15
S&P rating change last year 48,634 0.206 1.106 0 0 1
investment grade 48,634 0.553 0.497 0 1 1
lost inv. grade in last 5 years 48,634 0.0640 0.245 0 0 0
disagree: analyst std/price 30,607 0.0180 0.0338 0.00204 0.00780 0.0392
disagree: analyst std/abs(mean) 29,125 0.190 0.320 0.0263 0.0902 0.404
US rms conditional on having at least one bond in Mergent FISD:
VARIABLES N mean std p10 p50 p90
net CDS (USD billions) 9,535 1.167 1.199 0.336 0.856 2.180
net CDS / assets 9,535 0.0945 0.125 0.0107 0.0470 0.242
net CDS / issuer bonds (FISD) 9,291 0.662 1.162 0.0805 0.296 1.350
net CDS / consold bonds (FISD) 9,535 0.427 0.661 0.0524 0.190 1.053
5y CDS spread (bps) 12,060 297.1 566.3 51.73 136.2 633.6
CDS-bond basis (mly avg, %) 1,756 -0.902 1.552 -2.708 -0.733 0.527
assets (USD billions) 29,554 27.26 132.4 0.901 4.908 38.53
issuer bonds outstanding / assets 29,554 0.219 0.186 0.0268 0.188 0.435
consd bonds out. / assets 29,554 0.562 2.450 0.0634 0.250 0.824
S&P rating (notch) 29,554 10.71 3.577 6 10 15
S&P rating change last year 29,554 0.233 1.101 0 0 1
investment grade 29,554 0.540 0.498 0 1 1
lost inv. grade in last 5 years 29,554 0.0740 0.262 0 0 0
bond turnover (monthly) 29,554 0.0337 0.0372 0 0.0240 0.0791
bond turnover (last 12 months) 27,838 0.399 0.340 0.0205 0.327 0.840
number of trades (last 12 months) 27,838 5,422 24,979 14 657 8,978
issuer bond Herndahl 29,554 0.477 0.347 0.0956 0.360 1
equity turnover (monthly) 23,741 2.904 2.285 0.974 2.245 5.591
disagree: analyst std/price 22,015 0.0186 0.0362 0.00199 0.00753 0.0408
Firms in Capital IQ (annual observations):
VARIABLES N mean std p10 p50 p90
net CDS (USD billions) 774 1.333 1.137 0.385 1.032 2.463
net CDS / bonds 720 0.258 0.453 0.0362 0.153 0.518
assets (USD billions) 1,189 112.6 322.2 11.57 27.51 202.4
bonds / assets 1,189 0.167 0.149 0 0.136 0.379
accounts payable / assets 1,189 0.0595 0.0808 0 0.0356 0.162
term loans / assets 1,189 0.0389 0.0890 0 0 0.135
commercial paper / assets 1,189 0.00444 0.0141 0 0 0.0146
other borrowing / assets 1,189 0.0178 0.0478 0 0 0.0508
capital lease / assets 1,189 0.00276 0.0118 0 0 0.00496
revolving credit / assets 1,189 0.00964 0.0320 0 0 0.0275
trust preferred / assets 1,189 0.00117 0.00366 0 0 0.00329
41
Table 2: Baseline Results
This table presents the determinants of the size and existence of the CDS market using Compustat companies with S&P ratings. Columns (1)-(3)
use all rated companies; columns (3)-(6) only US companies with at least one bond issue in Mergent FISD; (7)-(9) further restricts the sample
to parent companies that do not have subsidiaries issuing bonds or with CDS written on them. Within every sample the rst column uses all
companies, the left hand side variable is log net notional CDS and applies a censored regression approach as described in Section 4.2; the second
column restricts the sample to those companies with a CDS market (in the DTCC database or with a Bloomberg quote) and also applies a
censored regression approach; the third column runs a probit regression on a dummy variable indicating the existence of a CDS market. credit
enhancement is an indicator dummy that is one for rms that provide credit enhancement. systemic is a dummy for the top 30 most systemic
nancial institutions. All other variables are dened in Table 1. The number of observations refers to rm-month observations. Industry eects
controlled for using rst digit SIC codes. Time xed eects are included. Clustered standard errors given in parentheses. ***, **, and * denote
signicance at the 1%, 5% and 10% level, respectively.
Baseline: US and international US rms in FISD US in FISD (conservative subs.)
(1) (2) (3) (4) (5) (6) (7) (8) (9)
rated if exist probit all if exist probit all if exist probit
log(assets) 0.767

0.208 0.586

0.684

0.154 0.903

1.314

0.292 1.210

(4.86) (1.64) (6.65) (3.87) (0.98) (5.45) (4.17) (0.98) (4.82)

log(assets)
2
-0.0624

-0.00203 -0.0650

-0.0763

-0.00573 -0.123

-0.183

-0.00750 -0.197

(-2.45) (-0.12) (-4.09) (-2.63) (-0.26) (-4.31) (-2.57) (-0.12) (-3.49)

log(debt) 0.683

0.401

0.248

0.0146 0.138 -0.179

-0.101 0.180 -0.256

(7.69) (5.58) (5.61) (0.12) (1.20) (-1.68) (-0.56) (1.13) (-1.92)

log(debt)
2
-0.0410

-0.0180 0.0210

0.0216 0.0131 0.0281 0.0123 -0.00635 0.0150

(-1.86) (-1.12) (1.93) (0.86) (0.65) (1.19) (0.20) (-0.13) (0.38)
log(bonds outstanding) 0.290

0.153

0.252

0.737

0.195 0.699

(3.64) (2.91) (3.43) (5.23) (1.58) (5.83)

log(consold bonds outstanding) 0.309

0.0650 0.362

(3.06) (0.82) (3.51)

bond turnover (last 12 months) 0.269

0.0271 0.402

0.258 -0.167 0.412

(1.87) (0.24) (2.68) (1.13) (-0.80) (2.16)

adj. log(bond Herndahl) -0.494

-0.182

-0.552

-0.760

-0.453

-0.514

(-4.44) (-2.17) (-4.76) (-4.04) (-2.75) (-3.05)

disagree: analyst std/price 2.204

1.561

0.803 3.893

2.376

2.157

(2.37) (2.22) (0.89) (3.15) (2.54) (1.83)

equity turnover (monthly) 0.0964

0.0733

0.0411

0.119

0.112

0.0339
(4.72) (5.21) (1.88) (3.63) (4.67) (1.19)
AA or higher rating 0.0904 -0.432

0.406

0.563

0.257 0.509 0.483 0.0459 0.521

(0.40) (-2.67) (2.10) (2.05) (1.32) (1.64) (1.04) (0.14) (1.39)
A rating 0.598

-0.123 0.856

0.817

0.414

0.783

1.041

0.512

0.899

(3.58) (-1.00) (6.93) (3.77) (2.74) (3.75) (3.18) (2.09) (3.37)

BBB rating 0.826

0.167 0.828

1.086

0.578

1.025

1.320

0.696

1.079

(5.68) (1.50) (8.14) (6.21) (4.70) (6.04) (5.44) (3.99) (4.89)

B rating -0.0689 -0.269

0.0895 0.0881 -0.114 0.322

0.397

0.0435 0.287
(-0.42) (-2.29) (0.80) (0.49) (-0.85) (1.88) (1.76) (0.25) (1.40)
CCC or lower rating 0.112 -0.0441 0.0749 0.661

0.129 1.003

1.117

0.619

0.697

(0.39) (-0.22) (0.36) (2.00) (0.48) (3.18) (3.54) (2.23) (1.92)

S&P rating change last year 0.0474

0.0537

0.0151 0.0210 0.0862

-0.0964

-0.0301 0.0476 -0.0832

(1.69) (2.68) (0.66) (0.56) (3.60) (-2.29) (-0.61) (1.63) (-1.43)
lost inv. grade in last 5 years 1.631

0.774

1.057

1.856

0.964

1.745

2.293

1.184

1.817

(8.67) (6.19) (7.36) (9.21) (6.72) (7.62) (8.48) (6.44) (6.26)

credit enhancement (dummy) 2.582

1.393

1.902

(5.39) (4.98) (3.64)

systemic 1.465

0.274 perfect +
(3.15) (0.80) (.)
industry: nance -1.596

-0.581

-1.134

-1.072

-0.493

-0.958

-1.659

-1.108

-0.851

(-10.85) (-5.55) (-9.75) (-5.95) (-3.61) (-5.44) (-5.77) (-4.39) (-3.53)

industry: agricultural 0.499 0.101 0.173 0.135 0.462

-0.333 0.295 0.505

-0.224
(1.17) (1.09) (0.34) (0.17) (3.15) (-0.47) (0.37) (2.52) (-0.32)
industry: service -0.617

-0.419

-0.248

-0.157 -0.0355 -0.244

-0.0369 0.172 -0.269

(-5.84) (-5.17) (-3.12) (-1.22) (-0.37) (-1.82) (-0.19) (1.13) (-1.55)
industry: cons. & mining 0.0510 -0.113 0.0977 -0.162 -0.215

0.0251 -0.0426 -0.275 0.252

(0.32) (-0.98) (0.75) (-0.93) (-1.73) (0.13) (-0.16) (-1.47) (1.11)
Constant -3.956

-1.383

-1.451

-3.833

-1.804

-1.721

-4.963

-2.361

-2.050

(-18.03) (-6.99) (-11.41) (-13.45) (-7.00) (-6.85) (-10.60) (-5.49) (-5.84)

time xed eects Yes Yes Yes Yes Yes Yes Yes Yes Yes
Observations 47988 19275 47536 19914 9758 19914 12306 4703 12306
42
Table 3: Determinants of Bond and Equity Turnover
This table presents the results of a tobit regression of bond turnover (left column in each pair) and equity turnover (right column in pair) using
S&P rated US companies in Compustat for which we have at least one bond issue in Mergent FISD. A tobit approach is used to account for the fact
that turnover is bounded from below by zero. Columns (1)-(2) use disagreement: std/price as a measure of disagreement; (3)-(4) use disagreement:
std/abs(mean) as a measure of disagreement; (5)-(6) use disagreement: std (unscaled) as a measure of disagreement; while (7) uses equity turnover
as a measure of disagreement. All other variables are dened in Table 1. The number of observations refers to rm-month observations. Industry
eects controlled for using rst digit SIC codes. Time xed eects are included. Clustered standard errors given in parentheses. ***, **, and *
denote signicance at the 1%, 5% and 10% level, respectively.
disagree: analyst std/price disagree: analyst std/abs(mean) disagree: equity turnover
(1) (2) (3) (4) (5)
bonds equity bonds equity bonds
log(bonds outstanding) 0.0136

0.00702 0.0142

0.00291 0.0134

(6.60) (0.07) (6.98) (0.03) (6.86)

log(issuer bond Herndahl) 0.0152

-0.128 0.0157

-0.0632 0.0159

(6.82) (-1.01) (7.00) (-0.50) (7.27)

disagree: analyst std/price 0.00640 13.65

(0.33) (6.65)
disagree: analyst std/abs(mean) 0.000969 1.177

(0.44) (7.16)
equity turnover (monthly) 0.00358

(8.96)
log(assets) 0.00443 0.678

0.00546

0.639

0.00156
(1.60) (4.17) (1.89) (3.80) (0.61)
log(assets)
2
0.000141 -0.0821

-0.000121 -0.0784

0.000485
(0.28) (-2.59) (-0.23) (-2.31) (1.09)
log(debt) -0.000945 -0.0161 -0.000990 0.0297 -0.000682
(-0.46) (-0.16) (-0.46) (0.31) (-0.35)
log(debt)
2
-0.000748 0.0282 -0.000607 0.0303 -0.000914

(-1.62) (1.03) (-1.32) (1.07) (-2.08)

AA or higher rating -0.0105

-2.305

-0.0112

-2.252

0.000289
(-2.39) (-9.08) (-2.53) (-8.56) (0.07)
A rating -0.0104

-1.615

-0.0113

-1.521

-0.00293
(-3.00) (-8.56) (-3.12) (-7.98) (-0.88)
BBB rating -0.00371 -0.924

-0.00490

-0.899

0.000583
(-1.41) (-5.97) (-1.78) (-5.66) (0.23)
B rating 0.00651

0.171 0.00762

0.261 0.00480

(2.24) (0.91) (2.49) (1.35) (1.70)

CCC or lower rating 0.0152

1.668

0.0258

1.898

0.00650
(2.18) (2.85) (2.52) (3.59) (1.12)
Constant 0.0420

4.005

0.0423

4.104

0.0273

(9.00) (15.93) (8.98) (16.35) (6.32)

time xed eects Yes Yes Yes Yes Yes
industry xed eects Yes Yes Yes Yes Yes
Observations 21235 21235 19789 19753 22808
43
Table 4: Disagreement and Bond Market Liquidity
This table presents the eect of the interaction of disagreement and bond market liquidity on the size and existence of the CDS market using
US Compustat companies with S&P ratings that have at least bond issue according to Mergent FISD. low liquidity, medium liquidity, and high
liquidity are assigned based on the liquidity terciles in each given month. Columns (1)-(3) uses number of bond trades to measure liquidity,
columns (4)-(6) use bond turnover (last 12 months), while columns (7)-(9) use adjusted log(bond Herndahl). Within every liquidity measure,
the rst column uses all companies, the left hand side variable is log net notional CDS and applies a censored regression approach as described in
Section 4.2; the second column restricts the sample to those companies with a CDS market (in the DTCC database or with a Bloomberg quote)
and also applies a censored regression approach; the third column runs a probit regression on a dummy variable indicating the existence of a
CDS market. The coecients on the interaction terms medium liquidity * disagree and low liquidity * disagree measure the additional eect
of disagreement on medium and low liquidity companies beyond the unconditional eect measured by analyst disagreement: std/price. In the
interaction terms disagree refers to analyst disagreement: std/price. All other variables are dened in Table 1. The number of observations refers
to rm-month observations. Industry eects controlled for using rst digit SIC codes. Time xed eects are included. Clustered standard errors
given in parentheses. ***, **, and * denote signicance at the 1%, 5% and 10% level, respectively.
liquidity: # of bond trades liquidity: bond turnover liquidity: adj. bond Herndahl
(1) (2) (3) (4) (5) (6) (7) (8) (9)
all if exist probit all if exist probit all if exist probit
disagree: analyst std/price 2.872

2.173

1.151 3.838

2.740

1.283 0.405 0.656 -0.524

(2.84) (3.21) (0.98) (3.66) (3.55) (1.22) (0.20) (0.48) (-0.28)
medium liquidity * disagree 2.465 3.547

-0.412 -0.828 1.103 -1.477 4.781

4.777

1.344
(1.28) (2.53) (-0.26) (-0.49) (0.93) (-1.02) (1.78) (2.72) (0.60)
low liquidity * disagree 5.144

6.057

1.449 5.473

4.227

3.200 3.777

1.967 2.614
(1.80) (2.45) (0.63) (2.29) (2.79) (1.35) (1.76) (1.27) (1.09)
lost inv. grade in last 5 years 2.039

1.021

1.675

2.001

1.004

1.650

1.861

0.951

1.551

(9.70) (6.95) (8.13) (9.18) (6.56) (7.83) (9.07) (6.63) (7.48)

medium liquidity (dummy) -0.436

-0.325

-0.191

0.166

-0.00684 0.244

0.162 -0.0236 0.249

(-3.33) (-2.94) (-1.83) (1.83) (-0.10) (2.52) (1.18) (-0.21) (2.16)

low liquidity (dummy) -1.028

-0.686

-0.527

-0.491

-0.363

-0.244

0.512

0.171 0.583

(-5.39) (-3.73) (-3.49) (-3.42) (-3.09) (-1.99) (3.44) (1.54) (4.25)

log(assets) 0.501

0.0625 0.645

0.568

0.0965 0.694

0.687

0.123 0.802

(2.73) (0.40) (3.93) (3.10) (0.62) (4.23) (3.86) (0.79) (4.74)

log(assets)
2
-0.0503 0.00605 -0.0838

-0.0569

0.00269 -0.0898

-0.0644

0.000896 -0.0953

(-1.62) (0.27) (-2.89) (-1.81) (0.12) (-3.08) (-2.15) (0.04) (-3.16)

log(debt) 0.251

0.150 0.0964 0.293

0.195

0.109 0.266

0.219

0.0635
(1.92) (1.45) (0.94) (2.14) (1.88) (1.03) (2.03) (2.10) (0.63)
log(debt)
2
0.0113 0.0192 0.00896 0.00158 0.0101 0.00721 0.00394 0.00704 0.0114
(0.43) (0.99) (0.36) (0.06) (0.51) (0.29) (0.15) (0.36) (0.45)
log(bonds outstanding) 0.334

0.117

0.335

0.433

0.168

0.390

0.391

0.159

0.356

(3.42) (2.31) (3.84) (4.47) (3.24) (4.60) (4.16) (2.93) (4.60)

AA or higher rating 0.375 0.0837 0.456 0.319 0.0450 0.429 0.158 -0.0164 0.257
(1.41) (0.47) (1.56) (1.17) (0.25) (1.45) (0.57) (-0.09) (0.88)
A rating 0.873

0.316

0.918

0.806

0.278

0.879

0.610

0.210 0.676

(3.99) (2.18) (4.61) (3.56) (1.88) (4.37) (2.71) (1.38) (3.35)

BBB rating 1.166

0.533

1.070

1.111

0.486

1.040

0.945

0.425

0.888

(6.52) (4.27) (6.92) (6.10) (3.88) (6.76) (5.30) (3.38) (5.71)

B rating 0.0386 -0.170 0.273

0.0826 -0.133 0.296

0.0869 -0.135 0.292

(0.22) (-1.26) (1.76) (0.45) (-0.95) (1.90) (0.47) (-0.96) (1.81)

CCC or lower rating 0.887

0.290 1.068

0.987

0.349 1.125

0.940

0.341 0.998

(2.58) (1.13) (3.27) (2.94) (1.36) (3.49) (2.82) (1.28) (3.52)

Constant -2.964

-1.096

-1.162

-3.329

-1.275

-1.401

-3.640

-1.415

-1.670

(-11.72) (-4.64) (-5.42) (-13.29) (-5.44) (-6.53) (-13.59) (-5.84) (-7.14)

time xed eects Yes Yes Yes Yes Yes Yes Yes Yes Yes
industry xed eects Yes Yes Yes Yes Yes Yes Yes Yes Yes
Observations 21235 9914 21235 21235 9914 21235 21235 9914 21235
44
Table 5: Downgrades and Bond Market Liquidity
This table presents the eect of the interaction of losing investment grade status and bond market liquidity on the size and existence of the CDS
market using US Compustat companies with S&P ratings that have at least bond issue according to Mergent FISD. low liquidity, medium liquidity,
and high liquidity are assigned based on the liquidity terciles in each given month. Columns (1)-(3) uses number of bond trades to measure liquidity,
columns (4)-(6) use bond turnover (last 12 months), while columns (7)-(9) use adjusted log(bond Herndahl). Within every liquidity measure the
rst column uses all companies, the left hand side variable is log net notional CDS and applies a censored regression approach as described in
Section 4.2; the second column restricts the sample to those companies with a CDS market (in the DTCC database or with a Bloomberg quote)
and also applies a censored regression approach; the third column runs a probit regression on a dummy variable indicating the existence of a CDS
market. The coecients on the interaction terms medium liquidity * lost IG and low liquidity * lost IG measure the additional eect of losing
investment grade on medium and low liquidity companies beyond the unconditional eect measured by lost inv. grade in last 5 years. All other
variables are dened in Table 1. The number of observations refers to rm-month observations. Industry eects controlled for using rst digit SIC
codes. Time xed eects are included. Clustered standard errors given in parentheses. ***, **, and * denote signicance at the 1%, 5% and 10%
level, respectively.
liquidity: # of bond trades liquidity: bond turnover liquidity: adj. bond Herndahl
(1) (2) (3) (4) (5) (6) (7) (8) (9)
all if exist probit all if exist probit all if exist probit
lost inv. grade in last 5 years 1.634

0.851

1.490

1.417

0.658

1.247

1.546

0.663

1.403

(7.48) (5.15) (6.27) (5.73) (3.27) (5.32) (5.52) (3.30) (5.14)

medium liquidity * 5yr lost IG 0.313 0.0833 0.121 0.407 0.281 0.384 0.261 0.297 -0.0319
(1.02) (0.33) (0.42) (1.61) (1.37) (1.39) (0.73) (1.09) (-0.09)
low liquidity * 5yr lost IG 1.393

1.024

0.420 1.364

0.858

0.769

0.652

0.496

0.902

(3.10) (3.13) (1.05) (3.65) (3.09) (2.18) (1.82) (1.97) (1.75)

disagree: analyst std/price 4.239

3.350

1.284 4.436

3.497

1.468 3.928

3.354

0.921
(4.69) (4.86) (1.41) (4.83) (4.97) (1.57) (4.03) (4.58) (0.98)
medium liquidity (dummy) -0.415

-0.265

-0.205

0.137 0.00160 0.198

0.220

0.0208 0.276

(-3.31) (-2.48) (-2.01) (1.56) (0.03) (2.13) (1.69) (0.20) (2.51)

low liquidity (dummy) -1.179

-0.820

-0.548

-0.582

-0.424

-0.268

0.508

0.138 0.575

(-5.99) (-4.42) (-3.70) (-3.99) (-3.66) (-2.24) (3.58) (1.29) (4.33)

log(assets) 0.519

0.0934 0.644

0.589

0.125 0.704

0.697

0.155 0.804

(2.81) (0.60) (3.92) (3.22) (0.80) (4.29) (3.88) (0.97) (4.76)

log(assets)
2
-0.0552

0.00107 -0.0843

-0.0611

-0.00185 -0.0917

-0.0651

-0.00279 -0.0947

(-1.78) (0.05) (-2.90) (-1.96) (-0.08) (-3.14) (-2.15) (-0.12) (-3.14)

log(debt) 0.259

0.141 0.102 0.305

0.197

0.113 0.246

0.188

0.0490
(2.02) (1.39) (0.99) (2.26) (1.93) (1.08) (1.89) (1.80) (0.50)
log(debt)
2
0.0127 0.0216 0.00835 0.00243 0.0121 0.00779 0.00596 0.0111 0.0113
(0.49) (1.11) (0.34) (0.09) (0.62) (0.31) (0.23) (0.56) (0.45)
log(bonds outstanding) 0.325

0.115

0.332

0.416

0.160

0.382

0.396

0.166

0.372

(3.63) (2.63) (3.86) (4.46) (3.29) (4.60) (4.40) (3.24) (5.09)

AA or higher rating 0.402 0.121 0.462 0.331 0.0529 0.439 0.176 0.00675 0.253
(1.50) (0.67) (1.58) (1.21) (0.29) (1.48) (0.63) (0.04) (0.87)
A rating 0.894

0.344

0.926

0.822

0.287

0.892

0.632

0.235 0.674

(4.07) (2.40) (4.64) (3.65) (1.96) (4.44) (2.81) (1.54) (3.35)

BBB rating 1.193

0.569

1.078

1.125

0.501

1.052

0.961

0.448

0.888

(6.64) (4.59) (6.95) (6.18) (4.01) (6.81) (5.39) (3.54) (5.72)

B rating 0.0655 -0.127 0.276

0.0877 -0.122 0.298

0.0956 -0.106 0.276

(0.36) (-0.95) (1.77) (0.48) (-0.90) (1.90) (0.52) (-0.76) (1.70)

CCC or lower rating 0.960

0.422 1.078

0.959

0.401 1.096

0.946

0.402 0.991

(2.86) (1.64) (3.27) (2.70) (1.45) (3.40) (2.85) (1.53) (3.38)

Constant -3.006

-1.172

-1.162

-3.355

-1.320

-1.411

-3.686

-1.469

-1.677

(-11.66) (-5.00) (-5.42) (-13.36) (-5.69) (-6.59) (-13.62) (-6.02) (-7.22)

time xed eects Yes Yes Yes Yes Yes Yes Yes Yes Yes
industry xed eects Yes Yes Yes Yes Yes Yes Yes Yes Yes
Observations 21235 9914 21235 21235 9914 21235 21235 9914 21235
45
Table 6: Detailed Debt Structure
This table presents how dierent types of debt aects the size and existence of the CDS market using Compustat companies with S&P ratings for
which we have detailed balance sheet information from Capital IQ using annual data for 2008 and 2009. In columns (1)-(3) we take logs of all
components of debt (after adding 10 million USD). Column (1) uses all companies, the left hand side variable is log net notional CDS and applies
a censored regression approach as described in Section 4.2; the column (2) restricts the sample to those companies with a CDS market (in the
DTCC database or with a Bloomberg quote) and also applies a censored regression approach; the column (3) runs a probit regression on a dummy
variable indicating the existence of a CDS market. In columns (4)-(6) we normalize all the explanatory variables (including xed eects dummies)
by assets and allow log standard errors allowed to vary with log(assets) to accommodate heteroskedasticity in the censoring. Column (4) uses all
companies, the left hand side variable is net notional CDS / assets and applies a censored regression approach as described in Section 4.2; the
column (5) restricts the sample to those companies with a CDS market (in the DTCC database or with a Bloomberg quote) and also applies a
censored regression approach; the column (6) runs a probit regression on a dummy variable indicating the existence of a CDS market. The number
of observations refer to rm-year observations. Industry eects controlled for using rst digit SIC codes. Time xed eects are included. Clustered
standard errors given in parentheses. ***, **, and * denote signicance at the 1%, 5% and 10% level, respectively.
log(net CDS) net CDS / assets
(1) (2) (3) (4) (5) (6)
all if exist probit all if exist probit
log(bonds+0.01) 0.150

0.0598

0.146

(5.63) (3.15) (6.01)

log(accounts payable+0.01) 0.0879

0.0829

0.127

(1.66) (2.16) (1.87)

log(term loans+0.01) -0.0631

-0.0139 -0.0774

(-3.61) (-1.06) (-3.56)

log(comm. paper+0.01) 0.0296 -0.0164 0.123

(1.39) (-1.05) (3.24)

log(other borrowing+0.01) -0.0110 0.0454

-0.0917

(-0.57) (3.32) (-3.48)

log(capital lease+0.01) 0.00778 -0.0139 0.0347
(0.26) (-0.64) (0.74)
log(revolving credit+0.01) 0.00863 0.0197 -0.0236
(0.33) (1.15) (-0.69)
log(trust preferred+0.01) -0.128

-0.00610 -0.153

(-3.08) (-0.17) (-3.83)

bonds / assets 0.0165

0.0123

3.222

(3.42) (2.59) (5.76)

accounts payable / assets 0.0388

0.0302

3.743

(3.13) (2.23) (3.33)

term loans / assets -0.00778 -0.00336 -1.101

(-1.43) (-0.62) (-1.76)

commercial paper / assets -0.0145 -0.00931 16.05

(-0.99) (-0.75) (2.74)

other borrowing / assets -0.00237 -0.00115 -2.210

(-0.50) (-0.27) (-1.79)

capital lease / assets -0.102 -0.0720 -3.476
(-0.86) (-0.73) (-0.72)
revolving credit / assets 0.00395 0.0355 -1.497
(0.08) (1.06) (-0.70)
trust preferred / assets 0.0283 0.0638

-70.58

(0.70) (1.74) (-4.60)

log(assets) -0.391 -0.130 -0.348 -0.00540 -0.00420 -0.498
(-1.56) (-0.66) (-1.14) (-0.99) (-0.75) (-0.67)
log(assets)
2
0.112

0.0456

0.104

0.000296 0.000224 0.0722

(4.21) (2.12) (3.06) (0.75) (0.56) (1.09)
Constant/assets 0.261 0.522

1.775
(1.44) (2.92) (0.22)
Constant -1.391

-0.533 0.608 0.0251 0.0197 1.307

(-2.39) (-1.19) (0.85) (1.32) (0.99) (0.64)
time xed eects Yes Yes Yes Yes Yes Yes
industry xed eects Yes Yes Yes Yes Yes Yes
rating xed eects Yes Yes Yes Yes Yes Yes
Observations 1180 847 1177 1180 847 1177
46
Table 7: The CDS-bond basis
This table presents the results of a regression with the dependent variable of log net notional CDS using US companies for which we have S&P
ratings and there is at least one bond issue in FISD-Mergent. Column (1) constrains the coecient on positive and negative basis to be equal, while
columns (2) and (3) allow them to dier. Column (3) includes additional control variables. The CDS-bond basis (mly avg, bps) from Bloomberg
is the dierence between the CDS spread and the Z-spread of the most liquid xed-coupon uncollateralized senior bond of the given company that
does not have any embedded options. All other variables are dened in Table 1. The number of observations refers to rm-month observations.
The CDS basis is winsorized at the 5% and the 95% level to avoid outliers driving the result. Industry eects controlled for using rst digit SIC
codes. Time xed eects are included. Clustered standard errors given in parentheses. ***, **, and * denote signicance at the 1%, 5% and 10%
level, respectively.
(1) (2) (3)
pooled split split w/ controls
CDS-bond basis (mly avg, %) -0.0982

(-3.69)
negative CDS-bond basis (mly avg, %) -0.145

-0.0884

(-3.69) (-2.30)
positive CDS-bond basis (mly avg, %) 0.0102 0.0198
(0.18) (0.31)
log(assets) 0.530

0.516

0.463
(1.76) (1.74) (1.50)
log(assets)
2
-0.0512 -0.0501 -0.0447
(-1.49) (-1.50) (-1.35)
log(debt) -0.00353 0.0118 -0.238
(-0.02) (0.06) (-1.17)
log(debt)
2
0.0606

0.0585

0.0716

(1.94) (1.96) (2.66)

adj. log(bond Herndahl) 0.0595
(0.42)
log(bonds outstanding) 0.241

(2.25)
equity turnover (monthly) -0.00910
(-0.35)
AA or higher rating 0.928

1.033

1.019

(2.61) (2.78) (2.44)

A rating 0.735

0.850

0.934

(2.28) (2.54) (2.25)

BBB rating 1.092

1.166

1.244

(3.32) (3.47) (2.98)

BB rating 0.474 0.535 0.906

(1.38) (1.54) (2.25)

B rating 0.338 0.375 0.665

(1.13) (1.24) (1.72)

lost inv. grade in last 5 years 0.723

0.714

0.430
(2.81) (2.83) (1.42)
Constant -2.174

-2.357

-2.083

(-3.74) (-4.12) (-3.25)

time xed eects Yes Yes Yes
industry xed eects Yes Yes Yes
Observations 1850 1850 1515
47
Table 8: Baseline Results: Anticipating future downgrade
This table presents the results of a regression with the dependent variable of log net notional CDS using companies for which we have Compustat
data and S&P ratings. Columns (1)-(3) use all rated companies; columns (3)-(6) only US companies with at least one bond issue in Mergent FISD;
(7)-(9) further restricts the sample to parent companies that do not have subsidiaries issuing bonds or with CDS written on them. Within every
sample the rst column uses all companies and applies a censored regression approach as described in Section 4.2; the second column restricts
the sample to those companies with a CDS market (in the DTCC database or with a Bloomberg quote) and also applies a censored regression
approach; the third column runs a probit regression on a dummy variable indicating the existence of a CDS market. S&P rating change next year
is the realized rating change over the consecutive 12 months. credit enhancement is an indicator dummy that is one for rms that provide credit
enhancement. systemic is a dummy for the top 30 most systemic nancial institutions. All other variables are dened in Table 1. The number of
observations refers to rm-month observations. Industry eects controlled for using rst digit SIC codes. Time xed eects are included. Clustered
standard errors given in parentheses. ***, **, and * denote signicance at the 1%, 5% and 10% level, respectively.
Baseline: US and international US rms in FISD US in FISD (conservative subs.)
(1) (2) (3) (4) (5) (6) (7) (8) (9)
rated if exist probit all if exist probit all if exist probit
S&P rating change last year 0.0465 0.0545

0.0171 0.0263 0.0863

-0.0846

-0.0239 0.0502

-0.0791
(1.56) (2.65) (0.69) (0.70) (3.70) (-1.96) (-0.48) (1.71) (-1.32)
S&P rating change next year 0.00809 0.0219 -0.0110 0.0293 0.0943

-0.0948

0.0283 0.0773 -0.102

(0.26) (0.85) (-0.49) (0.69) (3.12) (-2.27) (0.45) (1.57) (-1.75)

log(assets) 0.745

0.199 0.604

0.690

0.135 0.969

1.343

0.305 1.221

(4.66) (1.53) (6.58) (3.88) (0.85) (5.67) (4.26) (1.03) (4.80)

log(assets)
2
-0.0601

-0.000459 -0.0686

-0.0776

-0.00270 -0.134

-0.186

-0.00683 -0.200

(-2.34) (-0.03) (-4.17) (-2.69) (-0.12) (-4.56) (-2.60) (-0.11) (-3.50)

log(debt) 0.687

0.391

0.252

-0.0190 0.113 -0.201

-0.127 0.163 -0.251

(7.55) (5.31) (5.54) (-0.15) (0.96) (-1.78) (-0.70) (1.00) (-1.86)

log(debt)
2
-0.0421

-0.0174 0.0220

0.0234 0.0128 0.0357 0.0163 -0.00535 0.0169

(-1.89) (-1.07) (1.96) (0.94) (0.63) (1.47) (0.26) (-0.11) (0.43)
log(bonds outstanding) 0.270

0.135

0.254

0.740

0.200 0.690

(3.48) (2.60) (3.34) (5.24) (1.60) (5.72)

log(consold bonds outstanding) 0.348

0.105 0.361

(3.18) (1.28) (3.11)

bond turnover (last 12 months) 0.307

0.0325 0.447

0.270 -0.181 0.436

(2.11) (0.30) (2.90) (1.16) (-0.85) (2.26)

adj. log(bond Herndahl) -0.501

-0.181

-0.572

-0.762

-0.453

-0.520

(-4.55) (-2.18) (-4.85) (-4.03) (-2.75) (-3.05)

disagree: analyst std/price 1.897

0.530 1.710

3.650

1.529 3.200

(1.92) (0.74) (1.66) (2.52) (1.46) (2.38)

equity turnover (monthly) 0.101

0.0770

0.0417

0.117

0.116

0.0277
(4.90) (5.30) (1.84) (3.55) (4.76) (0.96)
AA or higher rating 0.0908 -0.434

0.409

0.560

0.239 0.509 0.454 0.0173 0.553

(0.40) (-2.65) (2.09) (2.03) (1.23) (1.61) (0.97) (0.05) (1.48)
A rating 0.606

-0.118 0.862

0.803

0.396

0.774

1.011

0.489

0.915

(3.59) (-0.95) (6.85) (3.66) (2.59) (3.64) (3.02) (1.93) (3.40)

BBB rating 0.820

0.160 0.832

1.061

0.560

1.017

1.288

0.685

1.081

(5.57) (1.41) (8.00) (5.98) (4.53) (5.86) (5.26) (3.82) (4.87)

B rating -0.0799 -0.270

0.0655 0.104 -0.0398 0.254 0.410

0.110 0.234
(-0.48) (-2.27) (0.57) (0.58) (-0.30) (1.45) (1.78) (0.62) (1.12)
CCC or lower rating 0.223 0.0298 0.148 0.693

0.271 0.952

1.245

0.750

0.706

(0.71) (0.13) (0.64) (2.03) (0.98) (2.69) (3.55) (2.36) (1.83)

lost inv. grade in last 5 years 1.652

0.783

1.077

1.848

0.966

1.733

2.257

1.166

1.801

(8.59) (6.18) (7.15) (9.10) (6.85) (7.27) (8.33) (6.36) (6.10)

credit enhancement (dummy) 2.714

1.423

2.259

(5.58) (5.01) (4.66)

systemic 1.462

0.281 perfect +
(3.11) (0.82) (.)
industry: nance -1.597

-0.600

-1.139

-1.073

-0.520

-0.941

-1.653

-1.118

-0.832

(-10.78) (-5.70) (-9.61) (-5.96) (-3.83) (-5.29) (-5.70) (-4.39) (-3.46)

industry: agricultural 0.481 0.0762 0.184 0.0812 0.379

-0.347 0.252 0.450

-0.243
(1.13) (0.85) (0.36) (0.10) (2.62) (-0.49) (0.32) (2.19) (-0.34)
industry: service -0.611

-0.414

-0.261

-0.150 -0.0330 -0.255

-0.0204 0.178 -0.258

(-5.73) (-5.07) (-3.19) (-1.17) (-0.34) (-1.88) (-0.10) (1.16) (-1.48)
industry: cons. & mining 0.0526 -0.110 0.0941 -0.149 -0.199 0.0196 -0.0352 -0.269 0.256
(0.32) (-0.94) (0.70) (-0.85) (-1.61) (0.09) (-0.13) (-1.42) (1.12)
Constant -3.908

-1.355

-1.462

-3.841

-1.797

-1.762

-4.961

-2.404

-2.011

(-17.38) (-6.67) (-10.99) (-13.40) (-6.93) (-6.79) (-10.62) (-5.55) (-5.64)

time xed eects Yes Yes Yes Yes Yes Yes Yes Yes Yes
Observations 44067 18059 43634 18530 9158 18530 11579 4486 11579
48