Jennie Bai and Pierre Collin-Dufresne

The Determinants of the CDS-

Bond Basis During the Financial
Crisis of 2007-2009

DP 11/2011-124
 The Determinants of the CDS-Bond Basis During the
Financial Crisis of 2007-2009 ∗

Jennie Bai† Pierre Collin-Dufresne‡

First Draft: November 2009

This version: November 2011

Abstract
We investigate both the time-series and cross-sectional variation in the CDS-bond basis, which measures
the difference between the CDS spread and cash-bond implied credit spread, for a large sample of individual
firms during the financial crisis. We test several possible explanations for the violation of the arbitrage
relation between cash bond and CDS contract that would, in normal conditions, drive the basis to zero.
Our findings do not uncover a clear single explanatory factor for the anomaly. Rather they point towards
several drivers related to funding risk, counterparty risk and collateral quality that force the individual
bond basis into negative territory at different phases of the crisis.

JEL Classification: G12.

Keywords: limit of arbitrage; basis; credit default swaps; counterparty risk; liquidity.

∗
we welcome comments.
†
Economist, Capital Markets Function, Federal Reserve Bank of New York, 33 Liberty Street New York, NY 10045,
e-mail: [email protected].
‡
Department of Finance, Graduate School of Business, Columbia University, 3022 Broadway Street New York, NY
10027, e-mail: [email protected].

Electronic copy available at: http://ssrn.com/abstract=2024531

 Abstract

We investigate both the time-series and cross-sectional variation in the CDS-bond basis, which measures

the difference between the CDS spread and cash-bond implied credit spread, for a large sample of individual

firms during the financial crisis. We test several possible explanations for the violation of the arbitrage

relation between cash bond and CDS contract that would, in normal conditions, drive the basis to zero.

Our findings do not uncover a clear single explanatory factor for the anomaly. Rather they point towards

several drivers related to funding risk, counterparty risk and collateral quality that force the individual

bond basis into negative territory at different phases of the crisis.

Electronic copy available at: http://ssrn.com/abstract=2024531

1 Introduction

Financial markets experienced tremendous disruptions during the 2007-2009 financial crisis. Credit
spreads across all asset classes and rating categories widened to unprecedented levels.1 Perhaps even
more surprising, many relations that were considered to be text-book arbitrage before the crisis were
severely violated. For example, in currency markets, violations of covered interest rate parity occurred
for currency pairs involving the US dollar (Coffey, Hrung, Sarkar (2009)). In interest rate markets
the swap spread, which measures the difference between Treasury bond yields and libor swap rates,
turned negative. In Interbank markets, basis swaps that exchange different tenor LIBOR rates (e.g.,
3-month for 6-month) deviated from zero. In inflation markets, break-even inflation rates turned
negative implying an obvious arbitrage with inflation swaps (Fleckenstein, Longstaff, Lustig (2011)).
In credit markets, the CDS-bond basis which measures the difference between CDS and cash-bond
implied credit spreads turned negative.
These anomalies suggest that such relations are not, in fact, arbitrage opportunities in the tra-
ditional textbook sense. One possible explanation is that arbitrage relations broke down during the
crisis because of institutional or contractual features. For example, many of these relations involve a
fully funded (e.g., cash) instrument and one or more unfunded derivative positions. This raises the
possibility that counterparty risk of the derivative issuer made the ‘arbitrage’ risky. An alternative
hypothesis is that funding cost differentials between the cash instrument and derivative positions were
responsible for the deviations. In the former case, the payoff to the ‘arbitrage’ trade is not risk-free
for any investor, whereas in the latter case, the payoff to the arbitrage trade would still be risk-free
for an investor with very deep pockets. In the latter case, therefore, violations can only arise in con-
1
For example, investment-grade corporate credit spreads as measured by the CDX.IG index rose from 50bps in early
2007 to more than 250bps at the end of 2008. Even at the safest end of the spectrum the widening was dramatic.
AAA-rated synthetic debt products, that would have been deemed virtually risk-free before the crisis, saw their spreads
widen dramatically: CDX.IG super senior tranche widened from 5bps to 100bps, CMBX AAA “super duper” widened
from 2bps to 700bps, ABS-HEL AAA tranche price rose from 0 to 20% upfront plus 500bps running. These numbers
illustrate that it became much more expensive to insure AAA-rated debt across various markets (corporate, residential
and commercial real estate).

Electronic copy available at: http://ssrn.com/abstract=2024531

junction with ‘limits to arbitrage,’ such as the inability of arbitrageurs to raise capital quickly and/or
their unwillingness to take large positions in these ‘arbitrage’ trades because of mark-to-market risk.
Consequently, the dynamics of the arbitrage violations should be affected by the structure of funding
markets, the ability of investors to process information, and the ability to move capital across markets
(Duffie (2010)).
These apparent arbitrage violation thus potentially provide a unique opportunity to test several
of the ‘limits to arbitrage’ theories (as surveyed, for example, by Gromb and Vayanos (2010)).
In this paper, we focus on the CDS-bond basis, which measures the difference between the credit
default swap (CDS) spread of a specific company and the credit spread paid on a bond of the same
company. Figure 1 plots the time series of the average CDS-bond basis for investment-grade (IG)
and high yield (HY) bonds, where the funding cost is measured by the libor swap curve. The figures
show that the basis, which hovers usually around +5 bps for IG firms, fell to -250 bps and slashed
to -650 bps for HY firms. At first sight, a large negative basis smacks of arbitrage since it suggests
that an investor can purchase the bond, fund it at libor swap, and insure the default risk on the bond
by buying protection via the CDS contract. The resulting trade is ‘virtually’ risk-free and yet, as the
figures show it generates between 250 and 650 bps in guaranteed return per annum.
Studying the CDS-bond basis during the crisis is interesting for several reasons. First, early
studies of this basis found that the arbitrage relation between CDS and cash-bond spreads holds fairly
well (Blanco, Brennan, Marsh (2005), Hull, Predescu and White (2004), Nashikkar, Subrahmanyam
and Mahanti (2010)) during the pre-crisis period. In fact, if anything, these studies typically conclude
that the basis should be slightly positive. Indeed, the arbitrage is, in general, not perfect (Duffie
(1999)), and there are a few technical reasons (such as (i) the difficulty in short-selling bonds, (ii) the
cheapest-to-deliver option) that tend to push the basis into the positive domain (Blanco, Brennan,
Marsh (2005)). However, during the crisis the basis was tremendously negative, which suggests the
need for alternative explanations. Second, there is a large cross-sectional variation in the observed

2
 Figure 1: A. The CDS-bond Basis of IG Firms vs OIS-LIBOR spreads

200

100

−100

−200
bps

−300

−400

−500

−600

−700 Basis of Investment−Grade Firms

OIS − LIBOR
−800
1/06 4/06 7/06 10/06 1/07 4/07 7/07 10/07 1/08 4/08 7/08 10/08 1/09 4/09 7/09 10/09

B. The CDS-Bond Basis of HY Firms vs OIS-LIBOR spreads

200

100

−100

−200
bps

−300

−400

−500

−600

−700 Basis of High−Yield Firms

OIS − LIBOR
−800
1/06 4/06 7/06 10/06 1/07 4/07 7/07 10/07 1/08 4/08 7/08 10/08 1/09 4/09 7/09 10/09

3
basis across individual firms. One obvious dimension of variation is rating, as shown in the picture
above. There are several other sources of variation that offer the potential to test a wide range of
hypotheses about the determinants of the basis.2 We discuss these next.
There are several reasons why one might expect the basis to become negative during the financial
crisis of 2007-2009. Anecdotes for the negative basis claim that several major financial institutions,
pressed to free up their balance sheet and improve their cash balance, reduced their leverage by
selling-off bonds. Some evidence for this deleveraging is presented in Figure 2 below, which shows
primary dealers’ position in long-term (maturity larger than one year) corporate securities. This
exerted downward pressure on bond prices, and upward pressure on credit spreads relative to CDS
spreads that represent the ‘fair’ value of the default risk insurance. This however cannot be the whole
story, since in a perfect frictionless market, investors would simply borrow cash to buy the bonds, buy
protection and finance the position until maturity (or default). For deleveraging to have a persistent
impact on the basis, there must be some ‘limits to arbitrage’ (Shleifer and Vishny (1997)). In particular,
if risk-capital is limited then the (mark-to-market) basis trade becomes risky and investors will tend
to buy the bonds-basis packages that are (ex-ante) most attractive from a risk-return trade-off.
In this paper we analyze the risk-return trade-off in a basis trade for an investor with limited
capital. We find that the investor is exposed to (a) the basis becoming more negative, (b) increased
uncollateralized funding costs, (c) increased collateralized funding costs (repo rates), and (d) increased
hair-cuts. Further, the profitability of the trade per unit of capital is decreasing in the collateral that
must be posted to enter the basis trade (essentially the hair-cut). All else equal, this suggests that the
basis should be less negative for bonds with smaller hair-cuts (i.e., better collateral quality), and for
bonds with a basis that have a lower covariance with funding costs (i.e., lower funding cost risk). In
this explanation for the negative basis, the corporate yield spread is temporarily too high relative to
2
As we discuss in the next section, the cross-sectional variation in the basis is also useful to circumvent the potential
bias that arises when estimating the corporate bond credit spread from the fact that the risk-free rate is possibly
estimated with error.

4
 Figure 2: Primary Dealer Position in Long-term Corporate Securities (in Mil.$)
5
x 10
2.5

1.5

0.5

1/06 7/06 1/07 7/07 1/08 7/08 1/09 7/09 1/10

Source: Federal Reserve Bank of New York.

the fairly valued CDS, due to the lack of available ‘arbitrage risk capital.’
A different explanation for the negative basis focuses on the CDS side of the trade. In a basis
trade, the protection is typically bought from a broker dealer such as J.P. Morgan, Goldman Sachs,
or Lehman Brothers. Clearly, the counterparty risk of these protection sellers increased considerably
during the crisis. Figure 3 plots the value-weighted credit default spread of the primary broker-dealers
in the U.S. market. We see a striking widening in the default risk of the average broker-dealers. A direct
implication is that the insurance sold by these broker-dealers should be less valuable. So increasing
counterparty risk of the broker-dealers should directly lead to lower CDS spreads, and therefore could
explain the observed negative CDS-Bond basis.
Our empirical investigation seeks to identify what were the main drivers of the basis during the
crisis. First, we analyze the time-series determinants of the average (IG and HY) basis during the crisis
to uncover which factors have predictive impact on the basis. We find that important predictors are

5
 Figure 3: The Average Five-Year CDS Spread for the Primary Broker-Dealers

350

300

250

200
bps

150

100

50

0
1/06 4/06 7/06 10/06 1/07 4/07 7/07 10/07 1/08 4/08 7/08 10/08 1/09 4/09 7/09 10/09

Source: Markit Inc. and Federal Reserve Bank of New York.

related to funding costs (and counterparty risk) of financial intermediaries in the sense that an increase
in financial intermediaries funding costs (as measured by OIS-LIBOR or primary dealer CDS) leads
to a more negative average basis. But, interestingly, we find significant differences in the time series
behavior of IG and HY basis. Specifically, deleveraging as measured by the change in the bond position
of primary dealers (PD) mainly affected the HY basis (a lower PD-bond position lowers the HY-basis).
For the IG basis, a more important determinant was the collateral-quality-funding spread as measured
by the difference in repo rates using MBS versus Treasury as collateral. The wider the spread the more
negative the basis. This suggests that the level of the individual bond basis may depend in a complex
manner on individual characteristics such as risk, collateral quality, and aggregate market conditions
such as dealers’ financial health, availability of collateralized and uncollateralized funding, and selling
pressure.
We confirm this by investigating the impact of individual firm/bond characteristics on the cross-

6
sectional (within rating) variations of the CDS-bond basis. Specifically, we construct measures (i.e.,
‘betas’) of market risk, funding cost risk, flight to quality risk, collateral quality and counterparty risk
for each firm, and run Fama-McBeth style cross-sectional regressions of the individual firm basis on
these (funding, liquidity, collateral and counterparty) betas. We then plot the resulting time-series
of regression coefficients. Our main results are that unconditionally all risk measures are statistically
significant in explaining the cross-sectional variation in the basis, as one would expect if the marginal
investor were a leveraged hedge-fund trading off risk and return when allocating scarce risk-capital to
these different basis-investment opportunities. Overall, the empirical model is reasonably successful
at explaining cross-sectional variation in the basis during the post-Lehman phase. Interestingly, we
find that IG and HY basis behave quite differently, with the former more driven by our proxies for
counterparty risk and flight to quality and the latter more by counterparty risk funding costs and
collateral quality.
Finally, we conduct additional tests at the individual bond level to investigate the role of selling-
pressure. Specifically, we look at the average level and the range of the post-crisis individual bond
trading volume (relative to its pre-crisis level). We find that only for HY bonds is this measure
statistically significant in explaining the bond basis (confirming our time-series result for the average
basis indexes). We also investigate lead-lag effects between price-discovery in the CDS and bond
market, following Blanco, Brennan, Marsh (2005). One would expect that if deleveraging was a big
factor for the basis, then bond spreads would lead CDS, and more so for bonds with higher price
pressure, and more risk.3 Indeed, we find that the share of price-discovery occurring in the CDS
market falls significantly during the crisis. The drop is much more significant during the post-lehman
phase and more pronounced for HY firms. For the latter the share falls even below 50% indicating
that by that metric bond price innovations lead the CDS market and thereby lending some support
to the deleveraging story for the HY market.
3
Instead, if counterparty risk was driving the basis, then we would expect only the CDS component to be affected
with not much effect on price discovery across markets.

7
 The negative basis has been subject of considerable attention in the practitioner literature (DE
Shaw ‘The basis monster that ate wall street’ (2009), JP Morgan ‘The bond-CDS basis handbook’
(2009), Mitchell and Pulvino (2010)). These papers emphasize the role of financing risk in generating
the negative basis, as well as the deleveraging of key leveraged investors in generating downward
price pressure on cash-bonds. In the academic literature Garleanu and Pedersen (2011) provide a
theoretical model, where leverage constraints can generate a pricing difference between two otherwise
identical financial securities that differ in terms of their margin requirements or hair-cuts. Specifically,
their theory predicts that the difference between two basis should be related to the difference in
margin requirements (i.e., haircuts) times the difference between the collateralized and uncollateralized
borrowing rate. They find support for their model in explaining the average basis difference between
high grade and high-yield bonds by the average difference in hair-cuts times a proxy for the collateral
borrowing spread as proxied by LIBOR-OIS.4 Our study differs from these previous papers in that
we focus on the cross-sectional variation in individual firms’ basis (rather than on the average basis
level) during the crisis and try to relate it to firm, bond and CDS characteristics. As in Garleanu
and Pedersen (2011) we find that collateral quality (our proxy for hair-cut margins) is an important
determinant of the basis, especially in the post Lehman phase. However, we also find substantial
cross-sectional variation in the basis that is hard to attribute solely to differences in margins.
In section 2 we discuss some practical issues regarding an actual basis trade and isolate the
various sources of risk in such a trade. In section 3, we discuss our data sources, and various proxies
for liquidity risk, funding cost, price pressure, and collateral quality. In Section 4 we present evidence
from predictive regression on factors driving the average levels of IG and HY basis. In section 5 we
show evidence on cross-sectional risk-factors driving individual bond basis from Fama-McBeth style
regressions. In section 6 we present some evidence about lead-lag effects of the basis for CDS and
4
Below we argue that a better proxy for the collateral spread is the OIS-GC Repo spread. Indeed, LIBOR-OIS
contains a pure bank-credit-term-spread component since LIBOR is a 3-month rate and OIS is based on overnight
borrowing, which may somewhat muddle the pure collateral effect.

8
credit spread. We conclude in section 7.

2 The CDS-Bond Basis

A credit default swap is essentially an insurance contract against a credit event of a specific
reference entity. It is an over-the-counter transaction between two parties in which the protection
buyer makes periodic coupon payments to the protection seller until maturity or until some credit
event happens. When a credit event occurs,5 typically the protection buyer delivers a bond from a
pool of eligible bonds to the protection seller in exchange for its par value.6
The contract is designed so that the owner of a particular bond can hedge her credit risk exposure
to the issuer of that bond by buying CDS protection on that counterparty. As a result we would
expect CDS spreads to be similar to credit spreads observed on corporate bonds that are deliverable
into the CDS contract. In fact, under some conditions, an exact arbitrage relation exists which implies
that the CDS spread should equal the credit spread on the deliverable corporate bond.7 This leads to
the theoretical definition of the CDS-bond basis as the CDS spread minus the corporate bond credit
spread.
While the CDS spread is observable in the market, it is not obvious how to compute the appropriate
corporate bond spread. As discussed by Duffie (1999) the ideal corporate bond spread would be the
spread over libor of a floating rate note with the same maturity as the CDS referenced on the same firm.
In practice, this spread is often not observable as firms rarely issue floating rate notes. Instead, we have
to rely on other available fixed rate corporate bond prices. Several methodologies have been proposed
in the literature. Following Elisade, Doctor, and Saltuk (2009) we adopt the Par Equivalent CDS
(PECDS) methodology developed by J.P. Morgan. This method, which we present for completeness
5
In the 2003 definition, the International Swap and Derivative Association (ISDA) lists six items as credit events:
(1) bankruptcy, (2) failure to pay, (3) repudiation/moratorium, (4) obligation acceleration, (5) obligation default, and
(6) restructuring. For more detail, see “2003 ISDA Credit Derivatives Definitions,” released on 11 February 2003.
6
See Duffie and Singleton (2003) for a detailed description.
7
Duffie (1999) discusses the specific conditions and shows why this relation might not exactly hold in practice.

9
in the appendix, essentially amounts to extracting the default intensity consistent with the prices of
the corporate bonds observed in the market and using the libor swap curve as the risk-free benchmark
curve. Then one can calculate the fair CDS spread consistent with the bond implied default intensity
and the risk-free benchmark curve (given a standard recovery assumption). It is this theoretical bond-
implied CDS spread, called the PECDS spread, that we compare to the quoted CDS spread on the
same reference entity to define the CDS-bond basis:

Basisi (τ ) = CDSi (τ ) − P ECDSi (τ ), (1)

where τ is the maturity and i indicates the reference entity. This methodology has several advantages,
reviewed in Elisade, Doctor and Saltuk (2009). It has also been used by previous academic studies
such as Subramanyam et al. (2009).
Another important issue for the measurement of the basis is the funding or ‘risk-free’ rate bench-
mark (Hull, Pedrescu and White (2004)). Several authors have argued that the Treasury curve is
not the appropriate risk-free benchmark and, indeed, that it is lower than the typical funding cost
an investor can achieve via collateralized borrowing.8 In fact, Hull, Predescu and White (2004) use
the basis package (a portfolio long several corporate bonds and long CDS protection) to define a
risk-free asset available to any investor. They argue that since the average CDS-bond basis is zero
when measuring funding cost using swap rate minus 10 bps and the CDS-bond basis exhibits little
cross-sectional variation, this is evidence that the ‘true’ shadow risk-free rate for a typical investor
is around swap minus 10 bps (or approximately Treasury plus 50 bps). We note that the very large
cross-sectional variation in the basis (across rating categories) documented in Figure 1 allows us to
immediately dismiss the fact that mis-measurement of the risk-free rate benchmark is the explanation
for the puzzling behavior of the CDS-bond basis during the crisis. If we were simply mis-measuring
8
Studies that document the special status of the US Treasury curve, –presumably due to its greater liquidity– include
Longstaff (2004), Feldhutter and Lando (2008) among others.

10
the risk-free benchmark we would observe an approximately constant CDS-Bond basis across firms
reflecting the spread between our benchmark risk-free curve and the true (unobserved) risk-free curve.
Let us stress, however, that since we do not observe the true risk-free benchmark curve, it is important
to focus on the cross-sectional variation in the basis, rather than focusing on the average level, which
could be affected by the ‘flight-to-quality’ effects documented in the benchmark Treasury and Swap
yield curves.
When the basis is positive, the credit default swap spread is greater than the bond spread. An
investor could then short the bond and sell CDS protection to capture the basis. When the basis
is negative, the credit default swap spread is lower than the bond spread. By buying the bond and
buying protection, investors could “lock-in” a risk-free annuity equal to the (absolute value of the)
basis.
As discussed in the introduction, during normal times the CDS-bond basis tends to be very small
and, if anything, slightly positive when measured relative to the libor-swap benchmark. This has been
studied extensively by Blanco, Brennan, and Marsh (2005) and Subramanyam et al. (2009). However,
Figure 1, which shows the time-series pattern of the CDS-bond basis for the overall U.S. IG and HY
bonds over the past four years, reveals that the CDS-bond basis was significantly and persistently
negative during the recent financial crisis. Furthermore, there was substantial cross-sectional variation
in the negative basis as we can see from the conspicuous difference in basis between IG bonds in Figure
1A and HY bonds in Figure 1B.
While a positive basis can often be traced back to some inability to implement the ‘arbitrage’ trade
because either bonds are difficult to short, or there exists a cheapest-to-deliver option (see Blanco,
Brennan, and Marsh (2005)), a negative basis is harder to explain. Indeed, in the negative basis case,
the ‘arbitrage’ trade requires buying the bond, financing its purchase, and buying protection to hedge
against the default event. Figure 1 suggests that the return to the ‘negative basis’ trade would have
been between 250 bps and 650 bps for IG and HY bonds respectively. These seem like very high

11
arbitrage profits. So it is important to review the details of such a basis trade implementation to
better understand where the ‘limits to arbitrage’ may arise.

2.1 Negative Basis Trade

In practice, there are several reasons why a negative basis trade is not a pure arbitrage. These
risks are discussed in detail in Elisade, Doctor, Saltuk (2009) (see, in particular, their table 2 on page
23). The main issues when implementing a negative basis trade have to do with funding risk, sizing
the long CDS position, and counterparty risk.
Suppose we find a bond with negative basis that trades at a price B below its notional of N . A
negative basis trade requires buying the bond. The purchase is funded via the repo market where
investors face a haircut h. This effectively implies that arbitrageurs will have to provide hB dollars of
‘risk-capital’ funded at Libor + f where f is the funding spread over libor faced by the arbitrageur.
The repo contract is typically over-night (up to a few months at most) with an agreed upon repo
rate and needs to be rolled over repeatedly until the maturity of the basis trade which is the lesser of
default and maturity (e.g., 5 years).
At the same time, to offset the risk of default, the investor buys protection in the CDS market.
A question arises as to how to size the CDS position. A conservative approach from a point of view
of minimizing exposure to ‘jump to default’ is to buy protection on the full notional N of the bond.
Market participants typically prefer to buy less protection to improve the carry profile of the
trade (pay less in insurance premium). The justification is that the maximum capital at risk in the
transaction is the initial purchase price of B.9 In fact, a customary approach is to make an assumption
about recovery (for example, assume that in case of bankruptcy a fraction R of the notional of the
bond is recovered) and buy protection on a CDS notional of NCDS so as to cover the loss in capital,
i.e., such that B −N R = NCDS (1−R). This will increase the carry of the trade (since the CDS premia
9
For bonds that trade at a premium one may in fact buy more protection!

12
are now reduced), but expose the investor to a jump to default in case the recovery is smaller than
expected. An alternative approach is to choose the notional of the CDS position to match the spread
duration on the risky bond (this approach tries to minimize mark-to-market differences between the
bond and CDS position over the life of the bond as opposed to thinking about jump to default risk).
As explained in Duffie (1999) there is no perfect arbitrage when the underlying bond is not a floating
rate note with the same maturity as the CDS contract.
For illustration, suppose the investor buys protection on a notional NCDS . This requires a margin
payment of M and periodic mark-to-market margin calls. The margin has to be funded at Libor + f .
After one day the profit or loss (P&L) on the trade can be written as:

P &L(t + 1) = Bt+1 − Bt + NCDS DCDS ∗ (CDSt+1 − CDSt )

−Bt ∗ [h(libor + f ) + (1 − h) ∗ (repo)] − Mt (libor + f )

where DCDS is the duration of the CDS (such that the P&L on the CDS is the product of the
duration with the change in CDS rate; note that if CDS increases the long position makes money).
For illustration, suppose we size our position in the CDS to match the libor-spread duration on the
corporate bond, then we can rewrite the P&L as:

P &L(t + 1) = DB ∗ (Basist+1 − Basist )

−Bt ∗ [h(libor + f ) + (1 − h) ∗ (repo)] − Mt (libor + f )

Specifically, this relation shows that the typical basis trade, when rolled over repeatedly, is exposed
to:

• The basis becoming more negative,

• An increase in illiquidity as measured by the benchmark libor rate.

13
 • An increase in the arbitrageurs own credit risk, which would lead to a larger markup (f ). We
note that if the arbitrageur has a large position in basis trades then this could be tied to the
basis becoming more negative (i.e., the trade running away from him).

• A worsening of collateral quality of the bond (funding liquidity), which would lead to an increase
in the haircut (h) and the repo rate.

• An increase in the margin requirements on the CDS position (Mt ).

Last but not least, the trade is also affected by counterparty risk in the sense that if a default on
bond occurs at time τB , then the P&L will be:

P &L(τB ) = R N + NCDS (1 − R)1τC >τB

where τC denotes the default time of the counterparty selling protection. Specifically, if the
counterparty defaults (or has defaulted) when the underlying firm defaults then the CDS protection
expires worthless. This highlights the fact that from an ex ante perspective counterparty risk depends
on the correlation between the default risk of the underlying name and the counterparty selling the
protection, which is typically a large bank such as J.P. Morgan, Lehman Brothers, Bear Stearns,
or Goldman Sachs. Now, it is important to stress that, in general, counterparty risk is viewed as
likely to be small, since if the counterparty defaults prior to the default event (i.e., τC < τB ) and if
mark to market were perfect, then the investor could reopen a new position at no cost with another
counterparty. Thus, in theory, counterparty risk only affects the investor if the counterparty defaults
on the exact same day as the underlying bond (τC = τB ). In practice however, it is likely that the
failure of the counterparty, especially during an extraordinary period like the financial crisis, would
be associated with substantial costs and risks for the investor. These losses would typically be related
to the likely mark-to-marking loss in the position on the day of the counterparty default as well as
more technical considerations, which have to do with the specific bankruptcy provisions in the ISDA

14
covering the CDS trade (e.g., if the mark-to-market limits were insufficient, or if the collateral posted
with the counterparty was rehypothecated, or if the cash settlement done upon bankruptcy of the
counterparty is based on mid-market quotes).
Below we try to use the cross-sectional variation in individual bond basis to disentangle the effects
of various risks outlined above that affect the risk-return trade-off of a basis trade. Our working
hypothesis is that an arbitrageur having limited access to capital will try to exploit the basis trade
opportunities that offer the best expected return per unit of risk-capital. So he will choose basis trades
that have the most negative basis (highest expected return) but controlling for ex ante measures of
exposure to market and funding liquidity. All else equal he will prefer basis trades on bonds with low
haircuts, low exposure to funding cost (in the sense that for two bonds with equally negative basis,
the one which correlates more with funding costs is more attractive, since the basis trade converges
when funding costs rise), low counterparty risk (in the sense that the probability of the underlying
firm defaulting at the same time as the counterparty in the CDS is lower). If this hypothesis is correct
then we expect that the risk characteristics of the basis trade (counterparty risk, funding liquidity risk,
collateral quality) should be related to the level of the basis in the cross section.10 We first describe
the data sources and time-series behavior of the basis.

3 Data

The data used to study the CDS-bond basis come from several sources. We start with the universe
of firms whose single-name CDS is traded in the derivative market and transactions are recorded in
the Markit database. Then we collect these firms’ corporate bond information from the Mergent Fixed
Income datascope. Finally we match each firm’s credit default swap and bond spread to corresponding
equity returns in the Center for Research in Security Prices (CRSP). All data are in daily frequency
10
A more sophisticated analysis would be to solve the optimal capital allocation decision of the arbitrageur to the
available basis-trades and test his first order condition.

15
from January 1, 2006 through September 30, 2009. The whole sample is further partitioned into three
phases: Phase 1 is the period before the subprime credit crisis, named ‘Before Crisis’ (1/2/2006 -
6/30/2007);11 Phase 2 is the period between the subprime credit crisis and the bankruptcy of Lehman
Brothers, called ‘Crisis I’ (7/1/2007 - 8/31/2008); and Phase 3 is the period after Lehman Brothers’
failure, ‘Crisis II’ (9/1/2008 - 9/30/2009).

3.1 Credit Default Swap

We download single-name credit default swap data from Markit Inc. for U.S. firms. The prices
are quoted in basis points per annum for a notional value of $10 million and are based on the standard
ISDA contract for physical settlement. The original dataset provides daily market CDS prices in various
currencies and different types of restructuring documentation clauses. Following a conventional rule,
we choose the CDS price in US dollar and the documentation clause type as ‘Modified Restructuring’
(MR).12
The original dataset also provides a CDS spread term structure incorporating maturities of 3m,
6m, 1y, 2y, 3y, 4y, 5y, 7y, and 10y. We use all maturities in conjunction with matching interest rate
swaps to calculate a term structure of default probability, which is an integral component in deriving
the bond-implied CDS spread (PECDS) and hence the CDS-bond basis (see Appendix A). In the end
we focus on the CDS-bond basis with a maturity of five years because the 5-year CDS is by far the
most liquid in the credit derivative market and for the convenience of comparison, because the 5-year
CDS is widely used in the literature.
11
There is not a unanimously agreed day for the beginning of the subprime crisis. Popular opinion is that the subprime
crisis started in August 2007 for a series of credit crunch events. Here we take a conservative stance by starting the
crisis period in July.
12
Under the 2003 Credit Definitions by the International Swap and Derivative Association (ISDA), there are four
types of restructuring clauses: Cumulative Restructuring (CR), Modified Restructuring (MR), Modified-Modified Re-
structuring (MM), and No Restructuring (XR). ‘Modified Restructuring’ is used by most broker-dealers in the U.S.
market. This convention holds till April 8, 2009. Afterwards the U.S. market adopts the ‘No Restructuring’ convention.
For consistency, we choose the MR documentation clause throughout our sample.

16
3.2 Corporate Bond

We get corporate bond data from the Mergent Fixed Income Databases. This database contains
information on virtually all publicly-traded bonds issued in the United States. For each firm in the
Markit dataset, we search Mergent datascope for all bonds which have 3 to 7 years remaining to
maturity measured each day during the sample period. We find that quite a few firms such as Warren
Buffet’s Berkshire Hathaway Inc. don’t issue mid-term bonds less than 7 years in the sample period,
then we further expand the bond sample to include firms which issue bonds with 7 to 10 years left
to maturity also measured at each day during the sample period. In line with Blanco, Brennan and
Marsh (2005), we exclude floating-rate securities and all bonds that have embedded options, step-up
coupons, sinking funds, or any special feature that would result in differential pricing.
For each bond, we collect bond price, coupon rate, annual payment frequency, issuing date and
maturity date. We then apply the methodology described in Appendix A to calculate the bond price
implied CDS spread and further to calculate the CDS-bond basis. Since we target the 5-year CDS
contract, we prefer bonds with maturities as close to five years as possible for better maturity matching.
Therefore, if an underlying firm issues bonds with 3 ∼ 7 years left to maturity, we only use these mid-
term bonds to calculate the basis. If an underlying firm like Berkshire doesn’t have bonds with 3 ∼ 7
years remaining to maturity in the sample period, we then calculate the basis from the firm’s bonds
with 7 ∼ 10 years left to maturity.
The method we use to calculate the CDS-bond basis is quite different from Blanco, Brennan and
Marsh (2005). They choose bonds with 3 ∼ 5 year to maturity at the beginning of their sample period,
and use a linear interpolation method to estimate a 5-year bond yield to match the 5-year CDS spread.
In terms of CDS data, Blanco et al. (2005) only need the five-year CDS spread. Yet we need to use the
complete CDS term structure to get more accurate measures of default intensity and hence a better
fit of the bond-implied CDS spread.
Finally we match the combined Markit-Mergent data to CRSP to gather information on stock

17
prices and outstanding shares. All together we get 484 firms throughout the whole sample period.

3.3 Reference Rate

We use U.S. dollar interest rates swaps as reference rates. The reference rate is used to proxy the
risk-free interest rate when credit spreads are calculated. The natural choice is a government bond
yield. As Blanco, Brennan and Marsh (2005) point out, “government bonds are no longer an ideal
proxy for the unobservable risk-free rate” due to tax treatment, repo specialness, legal constraints,
and other factors. Importantly, the libor-swap rate represents a better indicator of the funding cost
for financial intermediaries and typical basis swap traders than the Treasury curve. Therefore we use
it as our benchmark funding curve for the basis calculations.13 As discussed above, we choose not to
focus on absolute levels for the CDS-bond basis to be more immune to the reference rate issue.

3.4 Firm Characteristics

To construct the risk factors introduced in Section 3, we download firm characteristics from Capital
IQ and Mergent. For each firm in the merged CDS-bond-Equity dataset, we collect and calculate the
following variables at quarterly frequency from 2006:Q1 to 2009:Q3: SIZE, the logarithm of total assets;
LEVERAGE, the ratio of total debt to market capitalization; TANGIBLE RATIO, the percentage of
tangible asset in total asset; RATING, the firm’s long-term credit rating provided by Standard &
Poor’s; and SECTOR, including eight subsectors for industrial firms and six subsectors for financial
firms.
13
See also the Swap-Treasury spread discussions in Hull, Predescu and White (2004), Collin-Dufresne and Solnik
(2001).

18
3.5 Interest Rate Benchmarks

Finally we download the libor rate, interest rate swap rates, repo rates, Treasury note rates from
the Federal Reserve Board, and download the overnight indexed swap (OIS) from Bloomberg. We
present in Figure 4 the time series of various interest rate proxies. As is apparent there are a number
of interesting spreads that can be constructed from these series. In particular, we will focus on the
following spreads:

• LIBOR-OIS: The difference between uncollateralized interbank 3-month borrowing and overnight
borrowing rates. We consider it a proxy for short-term bank credit risk as well as interbank
liquidity risk.

• OIS-RepoGC: The difference between an uncollateralized overnight rate and the rate obtained
for borrowing against general collateral (e.g., Treasury Bonds). This is the ‘collateral spread’

• RepoMBS-RepoTsy: the difference between the rate obtained for borrowing against collateral
using MBS versus against collateral using Treasury bonds. This is the ‘collateral quality spread.’

• RepoGC-Treasury: The difference between collateralized lending rate and the 3-month T-Bill
rate. This captures a flight-to-quality liquidity component in that a widening of the spread
captures episodes where investor would rather own a Treasury paying a lower yield than a loan
fully collateralized by that same Treasury!

We provide some evidence on the correlation matrix of our various interest rate spread measures
in Table 2.

3.6 Other Data Sources

We also collect the following variables:

19
 Figure 4: The Time-series Evolution of various Interest Rates

7
LIBOR
6 OIS
Repo MBS
Repo GC
5 Treasury

0
Jan-06 Jul-06 Jan-07 Jul-07 Jan-08 Jul-08 Jan-09 Jul-09

Note: All interest rates are in the three-month maturity.

• The Volatility index (VIX) from the Chicago Board Options Exchange (CBOE): This index
captures both expected volatility as well as a volatility risk-premium.

• Stock market index from the Center for Research in Security Prices (CRSP);

• The noise factor from Hu, Pan and Wang (2011): This measure captures the extent to which
Treasury yields deviate from a smooth curve, and can be interpreted as an indirect measure of
the availability of risk-capital, since in periods of easy access to capital, fixed income relative
value strategies will be less costly to implement, and therefore fixed income convergence trades
will be less abundant.

• Corporate bond trading volume from the Trade Reporting and Compliance Engine (TRACE).
We will use this to build a measure of selling pressure. We refine this measure, when available,
by making use of the ‘sell’ indicator, reflecting a seller initiated transaction.

20
 • Primary dealer position in long-term corporate securities from the Federal Reserve Bank of New
York. We use it as the proxy of deleveraging pressure in the bond market.

3.7 Summary Statistics

Table 1 presents summary statistics of the CDS-bond basis. The basis across all firms was slightly
negative before the crisis, -3bps on average between 1/2/2006 to 6/30/2007, but fell to -21bps in the
first phase of the financial crisis (7/1/2007 - 8/31/2008) and further fell to -171bps after the bankruptcy
of Lehman Brothers (9/1/2008 - 9/30/2009). Meanwhile the volatility of the basis kept increasing for
all types of firms, on average from 9bps before the crisis to 22bps and further to 46bps during the
turmoil of the financial crisis.
Panel A also shows that firms with the IG rating share the same pattern as firms overall, whose
bases become more and more negative and volatile as the financial crisis progresses. However, firms
with HY ratings have strikingly different basis dynamics. Before the crisis these HY firms have positive
bases, as high as 83 bps on average. These firms’ bases began to narrow at the start of the subprime
mortgage crisis in the summer of 2007, yet they were still positive on average and higher than the
bases of IG firms in the first phase of the crisis. Only after the collapse of Lehman Brothers, did HY
firms’ bases plunge to an average -322 bps. Further the bases of HY firms is always much more volatile
than those of IG firms.
Panel B provides additional evidence to the different basis patterns between IG and HY firms.
We refine the rating to subcategories from AAA, AA to CCC and NR (no rating). Here each rating
category includes both its + and - notch, for example, AA group contains firms with ratings of AA+,
AA and AA-. We find consistent patterns as in Panel A that firms with lower credit rating tend
to have more negative and more volatile bases, suggesting a monotonic relationship between a firm’s
rating and the discrepancy between the CDS and cash-bond credit spread of the same firm. However
such a relationship only holds in the second phase of the crisis, i.e. after the Lehman Brothers failure.

21
The basis displays a right-skewed ‘smile’ from AAA to CCC during the pre-Lehman period.
Figure 1A and 1B provide an illustration of the basis dynamics for IG and HY firms respectively.
The solid red line is the average CDS-bond basis for firms in each rating category, weighted by firms’
market capitalization. In addition to echoing numbers in Table 1, these plots suggest that credit
conditions for firms in both rating category, though improved, are still far below their pre-crisis levels.
By the end of September 2009, the IG firms still had an average -140 bps basis and the HY firms a
-170 bps basis. In contrast, the negativeness of libor-ois spread illustrated by the dotted blue line,
had already come back to its pre-crisis level (12 bps on September 30, 2009 compared with 9 bps on
January 3, 2006), indicating that the international bank financing system had significantly recovered
by then.
In Panel A we also notice that financial firms have more negative and volatile bases than non-
financial firms during the crisis. Such a pattern is well illustrated in Figure 3. Panel C reports detailed
results for 14 sectors. Unlike the rating-categorized results in Panel B, we cannot find clear patterns
across industry sectors except that (i) manufacturing firms tend to have positive CDS-bond basis before
the crisis, (ii) the credit/financing sector was the most hit in the crisis with the largest discrepancy
between CDS and the cash-bond spread, (iii) the leasing and manufacturing sectors have relatively
small discrepancies in the crisis. Though no clear pattern emerges, Panel C provides strong evidence
for the existence of heterogeneity across firms. To illustrate this further we select a few interesting
examples.

3.8 Evidence on the Cross-sectional Variation in the Basis.

Garleanu and Pedersen (2011) make the point that haircuts are typically around 25% for IG firms
(and very similar across firms rated from AAA to BBB) and of the order of 55% for HY bonds (rated
BB or lower). In their model, to the first order, the basis differential between IG and HY bonds should
be equal to the difference between haircut margins multiplied by the collateral funding spread (i.e.,

22
the difference between the collateralized rate and the uncollateralized funding rate). While indeed an
important plausible determinant of the basis, our data suggests that there may be additional important
factors. Indeed, as clearly apparent from Panel B in Table 1 there is tremendous amount of variation
in the basis within a credit-rating category, and certainly a lot of differences in the basis within the
IG and the HY category.14
To illustrate this point even more dramatically, we present the basis in the following table for
(twelve) firms which in our sample have positive basis for more than 100 days during post-Lehman
Crisis II (9/1/2008 - 9/30/2009, with 271 days) period. These firms have diverse credit ratings, ranging
from B (Las Vegas Sands Corp and Penn Natl Gaming Inc) to AAA (Berkshire Hathaway, and GE),
and belong to six separate industries! This is clearly at variance with a model that would have a
single factor, such as haircuts or margins, explaining the basis.15 Clearly, the haircuts and margin
requirement on Las Vegas Sands were much larger than for Berkshire bonds, and yet both display a
positive basis (when the bulk of both IG and HY bonds displayed strongly negative bases at the time).
We will focus more systematically on the cross-sectional variation in the CDS-bond basis below.
We begin by first investigating the time series determinants of the basis.

4 Time Series Determinants of the Average Basis

We seek to identify the main drivers of the basis during the crisis. We first analyze the time-series
determinants of the average (IG and HY) basis during the crisis to uncover which factors had predictive
impact on the basis. To that effect, we perform predictive regressions of the average (equity market
cap weighted) basis of all firms, and of only IG and HY firms respectively. The candidate explanatory
14
Overtime there is also a lot of variation in the basis in a way that cannot solely be explained by the variation in the
collateral funding spread, and as we argue below is unlikely to be explained solely by changes in haircuts, though the
latter are not directly observable.
15
Indeed, Garleanu and Pedersen (2011)’s general model predicts that other factors (such as the covariance of the
underlying cash-flows with aggregate consumption) in addition to the margin differential should predict the difference
in the basis. It is only for the specific application to the CDS basis that they focus on the margin difference. Our data
suggests it is important to look for additional factors.

23
 ShortName Crisis I Crisis II Credit Rating Industry

Newmont Mng Corp 286 250 BBB Basic Materials

Berkshire Hathaway 127 244 AAA Financials
Amern Tower Corp 237 226 BB Technology
Emc Corp 259 188 BBB Technology
MetLife Insurance Co 12 178 A Financials
Boyd Gaming Corp 253 163 BB Consumer Services
General Electric Co 89 154 AAA Industrials
Windstream Corp 54 131 BB Telecommunications
Penn Natl Gaming Inc 134 130 B Consumer Services
Mylan Inc 204 122 BB Health Care
AutoNation Inc 1 117 BB Consumer Services
Las Vegas Sands Corp 108 106 B Consumer Services

Note: Ratings are based on September 2008.

variables (Zt ) include

1. the 5-year CDS spread of the primary dealer index, where the constituent primary dealers are
banks and security broker-dealers that trade in the U.S. government securities with the Federal
Reserve System.

2. the stock return of the primary dealer index,16

3. the CBOE volatility index,

4. the primary dealers’ position in corporate bonds with remaining maturity larger than one year,

5. the 3-month libor-ois spread,

6. the 3-month repo-Tbill spread,

7. the 3-month repo spread between MBS collateral and Treasury collateral,

8. the 3-month ois-repo spread,

16
A current list of primary dealers can be found at the Federal Reserve Bank’s website:
http://www.newyorkfed.org/markets/primarydealers.html.

24
 9. the noise factor for illiquidity in Hu, Pan and Wang (2010).

We test the impact of contemporaneous and up to three biweekly lagged explanatory variables in
the following regression:

∆BasisJt = α0 ∆Zt + α1 ∆Zt−1 + α2 ∆Zt−2 + α2 ∆Zt−3 + εt , J ∈ {All, IG, HY }

The results are reported in Table 3. The t-statistics are reported in the parentheses and estimated
using Newey-West standard errors with a lag of two. The data are biweekly spanning from January
2006 to September 2009. We then investigate a more succinct multi-variate model by selecting different
predictors from our previous univariate regressions investigation, to see how robust these initial results
are. The results of the multi-variate regressions are given in Table 4.

4.1 Results

We find that important predictors are related to funding costs (and counterparty risk) of financial
intermediaries in the sense that an increase in financial intermediaries funding costs (as measured by
libor-ois or primary dealer CDS) leads to a more negative average basis. But, interestingly, we find
significant differences in the time series behavior of IG and HY basis. Specifically, deleveraging as
measured by the change in the bond position of primary dealers (PD) mainly affected the HY basis (a
lower PD-bond position lowers the HY-basis). For the IG basis, a more important determinant was
the collateral-quality-funding spread as measured by the difference in repo rates using MBS versus
Treasury as collateral. The wider the spread the more negative the basis. This suggests that the level
of the individual bond basis may depend in a complex manner on individual characteristics such as
risk, collateral quality, and aggregate market conditions such as dealers’ financial health, availability
of collateralized and uncollateralized funding, and selling pressure.

25
5 Cross Sectional Determinants of Individual Firm Basis

We now investigate the cross-sectional variation in the bond basis, by first constructing individual
measures of exposure to funding cost risk, collateral quality, market risk, and countparty risk. Then
we run cross-sectional Fama-McBeth style regression of the basis on these beta measures.

5.1 Constructing Measures of Risk Exposures

5.1.1 Counterparty Risk

Counterparty risk is the risk that the seller of protection, typically an investment bank, cannot
make good on its commitment to the protection buyer in case of default. Counterparty risk should
therefore make the insurance less valuable and lower the CDS spread, possibly contributing to the
negative basis. As explained previously, counterparty risk is bigger, the higher the correlation between
the default events of the underlying entity and the protection seller.17 The challenge is how to measure
the correlation between the default risk of the underlying name and the counterparty selling the
protection.
The CDS market is over-the-counter and the exact nature of counterparties is not known. Further,
the process of netting makes it difficult to establish an aggregate measure of counterparty risk for
individual reference entity.18 Instead, we construct a counterparty risk measure for a representative
CDS issuer using the list of primary dealers designated by the Federal Reserve Bank of New York. These
17
That counterparty risk is not irrelevant, can be seen from the Lehman Brothers experience. Suppose an investor
had bought protection on Washington Mutual from Lehman Brothers. Washington Mutual defaulted only a few days
after Lehman. Without marking to market, the investor would be a regular claimant in bankruptcy for the protection
purchased from Lehman, leading to at best a partial loss. Of course, if ISDA agreements were well enforced, and provided
the investor had negotiated full-two-way mark to market with Lehman, then the risk would be further mitigated.
However, in practice, it is likely that most funds would have ended with a at least some partial loss as a result of this
double default.
18
In September 2008 the bankruptcy of Lehman Brothers caused almost $400 billion to become payable to the buyers
of CDS protection referenced against the insolvent bank. However the net amount that changed hands was around
$7.2 billion This difference is due to the process of “netting”. Market participants cooperated so that CDS sellers were
allowed to deduct from their payouts the funds due to them from their hedging positions. Dealers generally attempt to
remain risk-neutral so that their losses and gains after big events will on the whole offset each other.

26
primary dealers are banks and security broker-dealers that trade in the U.S. government securities with
the Federal Reserve System. To become qualified as a primary dealer, a firm must be in compliance
with capital standards under the Basel Capital Accord, with at least $100 million of Tier I capital for a
bank or above $50 million of regulatory capital for a broker-dealer. As trading partners of the central
bank, these primary dealers often are the biggest and most competitive financial institutions who
happen to be dominant issuers of credit default swap contracts. As of September 2008, there were 19
primary dealers such as Citigroup, Goldman, J.P. Morgan Chase and Morgan Stanley. The list changes
over time since some primary dealers may fail to meet required capital standards. Accordingly, we
update the components of the primary dealer index. For example, the index includes Lehman Brothers’
Holdings before its bankruptcy on September 15, 2008, but exclude it afterwards and adds Nomura
Securities International, Inc. starting from July 27, 2009.
For the primary dealer index, we calculate its stock return (Rindex )and CDS spread (CDSindex )
weighted by each constituent’s market capitalization. Appendix B lists the component in our primary
dealer index as of September 30, 2009. We then measure an underlying entity’s counterparty risk as
the regression coefficient of its stock return or the change of CDS spread on the return or the CDS
spread change of the primary dealer index:

cov(Ri , (Rindex − Rmkt ))

βi,cp = (2)
var(Rindex − Rmkt )

or
cov(∆CDSi , ∆CDSindex )
βi,cp = (3)
var(∆CDSindex )

where Ri is company i0 s stock return, and Rindex is the stock return of the primary dealer index.
∆CDSi is the daily CDS spread change for firm i and ∆CDSindex is the daily CDS spread change of
the primary dealer index
The higher the βcp the larger the likelihood of a joint default and the less valuable we expect the

27
protection to be when purchased from that counterparty. So we expect a negative coefficient in the
cross-sectional regression of bases on counterparty betas.

5.1.2 Funding Cost Risk

For an arbitrageur entering a basis trade, the risk is that the basis becomes more negative at the
same time as her funding costs widen. So a measure of the funding cost risk should be obtained from
the regression coefficient of the change in the basis on a measure of the change in funding costs. We
consider several measures of funding costs. The libor-ois spread which proxies at least partially for the
uncollateralized funding cost of financial intermediaries (unfortunately it also contains a credit risk
component). Alternative measures of funding cost that we also considered is the ois-repoGC spread and
the repoMBS-repoTsy, which as we describe previously measures respectively the ‘collateral spreads’
and the ‘collateral quality spreads‘. We report the results with only libor-ois beta in the current paper,
since we found the others to be highly correlated. (see table 2 for the correlation).
We thus measure an underlying entity’s funding cost risk as the regression coefficient of credit
default spread change on the change in the libor-ois spread:19

cov(∆CDSi , ∆(libor − ois))

βfi = (4)
var(∆(libor − ois))

The higher the funding cost beta, the less aggressively an arbitrageur would invest in that basis
trade, as the basis will become more negative in trades where his funding cost increases. So we expect
a negative coefficient in the cross-sectional regression the bases on the funding cost betas.
A caveat on the libor-ois spread is that this spread contains not only liquidity risk but also
financial intermediary credit (and therefore counterparty) risk during the financial crisis of 2007-2009.
This effect will go in the same direction for the expected coefficient.
19
We also considered to construct the beta by regressing the change in the basis on the change in the libor-ois spread.
Since our betas are estimated from daily data, using the cleaner CDS data seems preferable. This is also partially
justified, by the fact that most price discovery occurs in the CDS market, as we document in the next section.

28
5.1.3 Liquidity Risk

We also include a measure of liquidity risk, the repo spread, which is calculated as the 3-month
general collateral repo rate minus the 3-month Treasury bill. The difference between these two rates
reflects a flight-to-quality liquidity component as we discussed above. Our estimate of the liquidity
risk beta is therefore:
cov(∆CDS i , ∆repospread)
βli = (5)
var(∆repospread)

The higher the liquidity beta, the less aggressively an arbitrageur would invest in that basis trade,
since the basis will become more negative in states where liquidity becomes more valuable. So we
expect a negative coefficient in the cross-sectional regression of the bases on liquidity betas.

5.1.4 Collateral Quality

A fourth risk factor that affects the CDS-bond basis is the quality of bonds issued by the reference
entity in a CDS contract. To do the negative basis trade, an arbitrageur needs to buy bonds which
are funded via the repo market using the same bonds as collateral. The haircut imposed on that
transaction reduces the amount of leverage available to the arbitrageur. Higher haircuts imply higher
funding costs and therefore less profitable basis trade. All else equal we expect bonds with higher
haircuts to have more negative basis. We also expect bonds whose haircuts are more likely to increase
in the future to be less attractive for a basis trade. Unfortunately, haircuts are not observed for
individual bonds. Instead, we build a collateral quality index, from firm characteristics that is likely
correlated (in cross-section) with lower contemporaneous and future expected haircuts.
Specifically, we expect a firm with more total assets, more tangible assets, higher rating, lower
leverage, lower credit default swap spread, lower CDS volatility, to have bonds with higher collateral
quality.20 Therefore, for each firm we construct a collateral index in the similar way as Altman’s
Z-score (Altman (1968)). For each phase (Before Crisis, Crisis I and Crisis II), we collect and calculate
20
See also Ashcraft and Santos (2009).

29
firms’ size(+), leverage(-), tangible ratio(+), rating(+), average CDS spread(-) and CDS volatility(-),
standardize them cross-sectionally and add up the values according to the sign in the parentheses.
The resulting value defines our collateral index which reflects the collateral quality of bonds for each
reference entity.
The lower the collateral quality, the less profitable a basis trade is, measured per unit of expected
risk-capital usage. Indeed, we view collateral quality as a measure of the current (and expected future)
level of the haircut, with higher collateral quality meaning lower current and expected future haircut.
Thus, the lower the collateral quality the more negative the basis to equalize expected returns on a
per unit of risk-capital basis. So we expect a positive coefficient in cross-sectional regressions of the
bases on collateral quality.

5.1.5 Additional Controls

We add a few other factors in the cross-sectional regression to control for other sources of risk that
might affect how an arbitrageur would allocate risk capital to the basis trade. First, we add a market
beta factor, a proxy for how the particular basis covaries with the arbitrageurs portfolio. We tried
two definitions of market beta: a standard beta defined with respect to the equity market and a beta
defined with respect to the CDS market (depending on whether the arbitrageur holds a diversified
portfolio or is specialized in fixed income markets either could be relevant). Other potentially useful
controls are beta with respect to market volatility such as VIX.

5.2 Empirical Results

We study the cross-sectional determination of the CDS-bond basis with the following regression:

Basisi = αi + γcp βcp

i
+ γf βfi + γl βli + γcoll Collaterali + γk βControls
i
+ i (6)

30
Table 6 shows the correlation between various estimated betas. Since some of them display fairly high
correlation, we present different specifications of the multivariate regression. This table is also to bear
in mind when we seek to interpret the results economically. Obviously, the correlation makes these
interpretations somewhat ambiguous.
In Table 7 we present several specifications of the multivariate cross-sectional regressions of in-
dividual firm bases on proxies for counterparty risk, funding cost risk, liquidity risk and collateral
quality.
The full sample regressions presented in Panel A show that for the whole sample, all coefficients
are statistically significant, and most have the expected signs: negative for counterparty risk, negative
for funding cost risk, and positive for collateral quality. Unexpectedly the liquidity beta comes in
with a positive sign. However, focusing on subsamples and subsets of investment-grade or high-yield
firms, we see that the picture is somewhat more complex. Before the crisis, and during the early
part of the crisis, the explanatory power of the regressions (measure by adjusted R2 ) is fairly low
(typically less than 5 percent). It becomes much higher during the post-Lehman Crisis II phase:
The R2 are around 20 percent for HY firms and 10 percent for IG firms. The main (statistically
and economically significant) drivers of IG firms basis during the Crisis II period are counterparty
risk and flight to quality risk, both entering with a negative sign, and also significant but smaller in
magnitude the collateral quality variable entering with the positive sign. Instead, funding cost is not
statistically significant. On the other hand for HY firms during the Crisis II period, we find that the
most important variables (statistically and economically) are counterparty risk, collateral quality, and
funding risk. Now, liquidity risk is not statistically significant (even though it enters with the expected
‘negative’ sign).
Figure 6 shows the time series of the dynamic γj coefficients in the Fama-Macbeth regressions.
We see that counterparty risk becomes significant in the second half of 2008 and in the first quarter of
2009. Interestingly these two periods correspond precisely to the bankruptcy filing for Lehman Brothers

31
Inc.(September 2008) and the stock market hitting bottom (March 2009) respectively. Outside these
periods, the cross-sectional regressions do not find significant coefficient loading on the counterparty
risk beta. Funding cost risk as measured by the libor-ois spread becomes statistically and economically
most significant during the post-Lehman period, achieving its largest negative loading in March 2009.
Similarly, collateral quality matters most during this post-Lehman period.
The graph refines our understanding of why the regression on flight-to-liquidity risk as measured
by the GCRepo-TBill spread is not of the expected ‘negative’ sign unconditionally. During the crisis
II phase it becomes statistically significant (negative) only during the two short periods of the second
half of 2008 and March 2009. However, before that (during 2007-2008) it tends to have a positive
loading.
Figure 7 shows a ‘naive’21 variance decomposition of the cross-sectional variance in the basis
explained by each beta. The picture is interesting because it provides again a richer perspective than
the unconditional results. In particular, we see that the flight-to-liquidity beta, while in general only a
minor component of the cross-sectional variation, became very important relative to the other factors
during the second half of 2008 at the same time when it became statistically significantly negative.
We also see that during the early part (Crisis I) the main explanatory power comes from funding cost
risk and counterparty risk, whereas in the post Lehman phase, funding cost risk and collateral quality
become dominant and the role of counterparty risk diminishes.
Of course, there is a general caveat with the interpretation we provide for our results, since it is
difficult to account for the endogenous correlation between the various risk-factors we have labeled
funding cost risk, market liquidity, counterparty risk and collateral quality. For example, what we
have interpreted as funding cost risk could be due to counterparty risk, as libor-ois also reflects bank
credit risk.22
21
We decompose the total explained variance into separate components explained by individual regressors, but ignoring
the correlation between them.
22
If unwilling to give an economic interpretation to the variables, then at a minimum, our results provide an interesting
statistical description of the cross-sectional determinants of the basis in terms of observable covariates.

32
 Overall, the empirical model is reasonably successful at explaining cross-sectional variation in the
basis during the post-Lehman phase. Our main results are that unconditionally all four types of risk
measures are statistically significant in explaining the cross-sectional variation in the basis, as one
would expect if the marginal investor were a leveraged hedge-fund trading off risk and return when
allocating scarce risk-capital to these different basis-investment opportunities. Further, we find that
these betas become economically more important after the Lehman bankruptcy. Interestingly, we
find that IG and HY basis behave quite differently, with the former more driven by our proxies for
counterparty risk and flight to quality and the latter more by counterparty risk, funding costs and
collateral quality.

6 Evidence on Deleveraging and Limits to Arbitrage.

As mentioned in the introduction, there was substantial discussion in the popular press and also
some empirical evidence (see Figure 2) that financial institutions were forced to off-load risky assets to
reduce their leverage, resulting in fire-sale prices for risky bonds. Combined with limits to arbitrage,
one would expect to see significant drops in bond prices as a result of the price pressure and only a slow
lagged reaction of the CDS spread as arbitrageurs step in and equilibrate markets. Our time-series
regression suggest that deleveraging may have played a role for the average basis of HY firms. To
further test this hypothesis we perform two additional experiments.
First, we look at the average level and the range of the post-crisis individual bond trading volume
(relative to its pre-crisis level). We find that only for HY bonds is this measure statistically significant
in explaining the bond basis (confirming our time-series result for the average basis). Second, we
investigate lead-lag effects between price-discovery in the CDS and bond market, following Blanco,
Brennan and Marsh (2005). One would expect that if deleveraging was a big factor for the basis, then
bond spreads would lead CDS, and more so for bonds with higher price pressure, and more risk.23
23
Instead, if counterparty risk was driving the basis, then we would expect only the CDS component to be affected

33
Indeed, we find that the share of price-discovery occurring in the CDS market falls significantly during
the crisis. The drop is much more significant during the post-Lehman phase and more pronounced
for HY firms. For the latter the share falls even below 50% indicating that by that metric bond price
innovations lead the CDS market and thereby lending some support to the deleveraging story for the
HY market.

6.1 Price Pressure in Bond Market

To test whether bond selling pressure is one determinant of the CDS-bond basis, we use bond
volume percentage change as a proxy of bond selling pressure. For each firm in the sample we first
calculate its average monthly dollar trading volume in the before-crisis period as the benchmark. We
then record the average monthly dollar volume in the post-Lehman period.24 The bond selling pressure
is then approximated by the percentage change in volume with respect to the pre-crisis level. Here we
use average monthly volume in order to reduce the noise in the daily bond volume measure. Further,
we construct another bond selling pressure measure by using the volatility of trading volume change.
We report our test results for the impact of deleveraging in Table 8. This table shows the influence
of the change of bond trading volume on the change of basis between the post-Lehman period and the
before-crisis period, using the following regression:

∆V ol J ∆(High − Low) J
∆BasisJi = β1 ( ) i + β2 ( )i + εJi , J ∈ [IG, HY, F in, N F ], (7)
V ol M ean

where ∆Basis ≡ Basis2 − Basis0 ; ∆V ol/V ol ≡ (V ol2 − V ol0 )/V ol0 is the percentage change of
corporate bond trading volume; High is the maximum monthly trading volume in a specific phase,
and Low is the minimum monthly trading volume in the same phase, Mean is the average monthly
with not much effect on price discovery across markets.
24
An accurate measure should use bond trading volume indicated as Sell. However, the buy/sell indicator only
becomes available since the late 2008 in the TRACE dataset. We do a robustness check using bond trading volume
indexed as sell for a much shorter sample period and find similar impact compared with the result using bond trading
volume without a sell indicator.

34
volume in the same phase. The second independent variable is a measure of the volatility of trading
volume change. Our intuition is that a very wide volume range may be a better proxy of price pressure
during the period than simply a change in average volume (though admittedly both are crude). The
subscript ‘2’ refers to the post-Lehman phase (9/1/2008 - 9/30/2009) and the subscript ‘0’ refers to
the before-crisis phase (1/2/2006 - 6/30/2007). We observe negative regression coefficients for the
volume percentage change consistently across all types of firms. However none of them is statistically
significant at the 10% critical value. We also observe consistent negative coefficients for trading volume
range, and find statistical significance for HY as well as non-financial firms. Overall, the explanatory
power of volume change and its volatility is marginally small in explaining the change of the basis.
The highest R-square is six percent.
Overall these results offer little evidence of the price pressure hypothesis driving the basis of
investmeng-grade or financial firms. There is marginal evidence that deleveraging might have con-
tributed to the negative basis for high-yield and non-financial firms, which is consistent with our
time-series regression results. We turn to lead-lag price-discovery for additional evidence next.

6.2 Price Discovery across Markets

Another channel to explore the determinant of CDS and cash-bond arbitrage is to evaluate the
information content in each market and to study which market provides more timely information over
various phases of the crisis. In line with Blanco, Brennan and Marsh (2005), we test the CDS market’s
contribution to price discovery in the following vector error-correction models:25

p
X p
X
∆CDSt = λ1 (CDSt−1 − P ECDSt−1 ) + β1 ∆CDSt−i + γ1 ∆P ECDSt−i + ε1,t
i=1 i=1
Xp Xp
∆P ECDSt = λ2 (CDSt−1 − P ECDSt−1 ) + β2 ∆CDSt−i + γ2 ∆P ECDSt−i + ε2,t (8)
i=1 i=1

25
Blanco, Brennan, and Marsh (2005) adopt a general form for the error term, allowing for flexible cointegration
relationship. In our case, PECDS is designed to match CDS and hence their difference, the basis, is a more direct and
accurate measure of the error term.

35
where P ECDS is the par-equivalent CDS calculated in section 2.

The contribution of the CDS market to the price discovery of common credit risk is defined by the

permanent factor in Gonzalo and Granger (1995). 26 The Gonzalo and Granger measure ignores the correlation

between the two markets and attributes superior price discovery to the market that adjusts the least to price

movements in the other market, defined as

λ2
GG = , (9)
λ2 − λ1

where the λ coefficients reveal which of the two market leads in price discovery.

We do the price discovery test for each firm and summarize the mean, median, and standard error in

Table 9 for all firms, investment-grade, high-yield, financial, and non-financial firms during each phase of the

financial crisis, – Pre-Crisis (1/2/2006 - 6/30/2007), Crisis I (7/1/2007 - 8/30/2008), and Crisis II (9/1/2008

- 9/30/2009). In addition to confirming the previous results in the literature that the CDS market leads the

bond market (indicated by a GG factor larger than 50%), we also find that the information content captured

by the CDS spread is almost monotonically declining as the financial crisis worsens. When averaged across

all firms, the CDS market contribution reduces from 92% before the crisis to 74% in the post-Lehman period.

Similarly, the CDS market contribution drops 5% for investment-grade firms, 80% for high-yield firms, 18%

for non-financial firms, and 16% for financial firms.

Summarizing, we find that the share of price-discovery occurring in the CDS market falls significantly

during the crisis. The drop is much more significant during the post-Lehman phase and more pronounced for

HY firms. For the latter the share falls even below 50% indicating that by that metric bond price innovations

lead the CDS market and thereby offer strong support to the deleveraging story for high-yield bonds.
26
We have also calculated alternative measure using Hasbrouck “information share.” Given that Hasbrouck’s approach
can only provide upper and lower bounds on the information shares of each market, it is not suitable to report the
evolution of information content over the phases of the crisis. The range of lower and upper bounds are not small and
overlapping across phases, making it hard to distinguish the nuance over time.

36
7 Conclusion
Focusing on the cross section of CDS-bond basis is interesting as it provides more general insights into the

‘limits to arbitrage’ literature. Our analysis is a preliminary step in testing the implications of the literature

that models the behavior of arbitrageurs with limited capital facing multiple ‘arbitrage’ opportunities (Gromb

and Vayanos (2010)).27

27
While we are not formally testing the first order condition of such an investor in the current paper, the CDS-bond
basis dataset clearly presents an interesting laboratory to test such a theoretical framework.

37
References
Altman, Edward, 1968, Financial ratios, discriminant analysis and the prediction of corporate bankruptcy,

Journal of Finance 23, 189 –209.

Ashcraft, Adam B., and Joo A.C. Santos, 2009, Has the cds market lowered the cost of corporate debt?,

Journal of Monetary Economics 56, 514–523.

Blanco, Roberto, Simon Brennan, and Ian W. Marsh, 2005, An empirical analysis of the dynamic relation

between investment-grade bonds and credit default swaps, Journal of Finance 60, 2255–2281.

Coffey, Niall, Warren B. Hrung, and Asani Sarkar, 2009, Capital constraints, counterparty risk, and deviations

from covered interest rate parity, Staff Reports 393, Federal Reserve Bank of New York.

Collin-Dufresne, Pierre, and Bruno Solnik, 2001, On the term structure of default premia in the swap and

libor markets, JF 56, 1095–1115.

Coval, Joshua D., Jackub W. Jurek, and Erik Stafford, 2009, Economic catastrophe bonds, American Eco-

nomic Review 99(3), 628–666.

D.E. Shaw Group, 2009, The basis monster that ate wall street, Market Insights March.

Duffee, Gregory R., 1999, Estimating the price of default risk, The Review of Financial Studies 12, 187–226.

Duffie, Darrell, 2010, Presidential address: Asset price dynamics with slow-moving capital, Journal of Finance

4, 1237 – 1267.

, and Ken Singleton, 2003, Credit Risk (Princeton University Press).

Elisade, Abel, Saul Doctor, and Yasemin Saltuk, 2009, Bond-cds basis handbook, J.P. Morgan Credit Deriva-

tives Research February 05.

Garleanu, Nicolae, and Lasse Heje Pedersen, 2011, Margin-based asset pricing and deviations from the law of

one price, The Review of Financial Studies 24, 1980 – 2022.

38
Gromb, Denis, and Dimitri Vayanos, 2010, Limits of arbitrage: The state of the theory, Annual Review of

Financial Economics 2, 251 – 275.

Hull, John, Mirela Predescu, and Alan White, 2004, The relationship between credit default swap spreads,

bond yields, and credit rating announcements, Journal of Banking and Finance 28, 2789–2811.

Hull, John, and Alan White, 1990, Pricing interest rate derivative securities, The Review of Financial Studies

3no4, 573–592.

Jack Bao, Jun Pan, and Jiang Wang, 2011, The illiquidity of corporate bonds, Journal of Finance 66, 911 –

946.

Longstaff, Francis A., 2004, The flight-to-liquidity premium in u.s. treasury bond prices, Journal of Business

77, 511–526.

Matthias Fleckenstein, Francis Longstaff, and Hanno Lustig, 2010, Why does the treasury issue tips? the

tips-treasury bond puzzle, working paper, University of California at Los Angeles.

McGinty, L, and R. Ahluwalia, 2004, Introducing base correlation, Research Paper JP Morgan.

Mitchell, Mark L., and Todd C. Pulvino, 2010, Arbitrage crashes and the speed of capital, working paper.

Shleifer, Andrei, and Robert W. Vishny, 1997, The limits of arbitrage, Journal of Finance 88, 35–55.

Stanton, Richard, and Nancy Wallace, 2009, The bear’s lair: Indexed credit default swaps and the subprime

mortgage crisis, working paper, University of California at Berkley.

Subrahmanyam, Marti G., Amrut J. Nashikkar, and Sriketan Mahanti, 2010, Limited arbitrage and liquidity

in the market for credit risk, working paper New York University.

Xing Hu, Jun Pan, and Jiang Wang, 2011, Noise as information for illiquidity, working paper, MIT.

39
A The Par-Equivalent CDS Methodology
We present the Par Equivalent CDS methodology developed by J.P. Morgan to calculate the CDS-bond

basis in Section 2. This survival-based valuation approach provides an apple-to-apple measure across the

cash-bond spread and the credit default swap spread.

The fair value of the coupon on a CDS is set so that the expected present value of the premium leg is

equal to the expected present value of the contingent payment (see Duffie (1999), Hull and White (2001)).

Assuming that we have a zero-coupon discount curve Z(t) extracted from swap spreads and assuming a

constant intensity survival probability S(t), the expected present value of the premium leg is given by:

n n
X X ti + ti−1
PVpremium (C) = Z(ti )S(ti ) C ∗ dt + Z( )[S(ti−1 ) − S(ti )] C ∗ dt/2, (10)
2
i=1 i=1

where the second component is the present value of the accrued interest upon default (assumed to occur

halfway between ti−1 and ti ). The expected present value of the contingent leg is:

n
X ti + ti−1
PVcontingent = (1 − R) Z( )[S(ti−1 ) − S(ti )], (11)
2
i=1

where R stands for the recovery rate. The fair credit default swap spread is the number C that sets

PVpremium (C) = PVcontingent (12)

The par-equivalent CDS uses the market price of a bond to calculate a spread based on CDS-implied

default probabilities. First, we need to get a CDS-implied default probability curve. We plug in the market

CDS price with 3-month maturity (C0.25 ) to get the 3-month survival probability, S0.25 . Then we plug in the

CDS price with 6-month maturity (C0.5 ) and the calculated 3-month survival probability (S0.25 ) to get the

6-month survival probability, S0.5 . Sequentially plugging in CDS spread with longer maturity from 1-year to

10-year, we can get a curve of the survival probability SCDS (ti ).

Second we need to get a bond-implied survival probability curve Sbond (ti ). Using the CDS-implied survival

probability as a prior, we calculate the bond-implied survival probability curve as the one that minimize the

40
pricing error between the market price and derived bond price:

Sbond (ti ) = SCDS (ti ) + ε. (13)

ε = arg min (P V (Sbond ) − Market Price of Bond)2 . (14)

s.t.

Then the bond-implied CDS spread term structure is defined by substituting the survival probability term

structure fitted from bond prices, Sbond (t), into the following equation for par equivalent CDS spreads, denoted

as PECDS:

Pn ti +ti−1
(1 − R) i=1 Z( 2 )[Sbond (ti−1 ) − Sbond (ti )]
P ECDS = P h i (15)
n ti +ti−1 dt
i=1 Z(ti )Sbond (ti ) ∗ dt + Z( 2 )[Sbond (ti−1 ) − Sbond (ti )] ∗ 2

B The Primary Dealers List

Effective since July 27, 2009. The list is downloaded from the website of the Federal Reserve Bank of

New York: http://www.newyorkfed.org/markets/pridealers_current.html.

41
BNP Paribas Securities Corp.

Banc of America Securities LLC

Barclays Capital Inc.

Cantor Fitzgerald & Co.

Citigroup Global Markets Inc.

Credit Suisse Securities (USA) LLC

Daiwa Securities America Inc.

Deutsche Bank Securities Inc.

Goldman, Sachs & Co.

HSBC Securities (USA) Inc.

Jefferies & Company, Inc.

J. P. Morgan Securities Inc.

Mizuho Securities USA Inc.

Morgan Stanley & Co. Incorporated

Nomura Securities International, Inc.

RBC Capital Markets Corporation

RBS Securities Inc.

UBS Securities LLC.

42
 Figure 5: A. The CDS-Bond Basis of Financial Firms vs OIS-LIBOR spreads

200

100

−100

−200
bps

−300

−400

−500

−600

−700 Basis of Financial Firms

OIS − LIBOR
−800
1/06 4/06 7/06 10/06 1/07 4/07 7/07 10/07 1/08 4/08 7/08 10/08 1/09 4/09 7/09 10/09

B. The CDS-Bond Basis of Non-Financial Firms vs OIS-LIBOR spreads

200

100

−100

−200
bps

−300

−400

−500

−600

−700 Basis of Non−Financial Firms

OIS − LIBOR
−800
1/06 4/06 7/06 10/06 1/07 4/07 7/07 10/07 1/08 4/08 7/08 10/08 1/09 4/09 7/09 10/09

43
 A. Counterparty Risk (γcp ) B. Funding Liquidity Risk (γl (repo))
4 8

2 6

0 4

−2 2

−4 0

−6 −2

−8 −4

−10 −6
1/06 7/06 1/07 7/07 1/08 7/08 1/09 7/09 1/10 1/06 7/06 1/07 7/07 1/08 7/08 1/09 7/09 1/10

C. Funding Liquidity Risk (γf (libor)) D. Collateral Quality (γcol )

2 1.4

44
1.2
0

−2
0.8

−4 0.6

0.4
−6

0.2

−8
0

−10 −0.2
1/06 7/06 1/07 7/07 1/08 7/08 1/09 7/09 1/10 1/06 7/06 1/07 7/07 1/08 7/08 1/09 7/09 1/10

i
Figure 6: Dynamic Coefficients in the Cross-Sectional Regression: Basisit = γt,cp βcp + γt,l (repo)βli +
i
γt,f (libor)βfi γt,coll Collaterali + γk βcontrol + εit , where βli is measured by the change of the repo spread (3m repo rate
with general collateral of Treasury bill minus the 3m Treasury bill rate) and βfi is measured by the change of the
libor-ois spread.
 0.7
Counterparty Risk
Liquidity (repo)
Liquidity (libor)
0.6 Collateral Quality
Market Risk

0.5

0.4

0.3

45
0.2

0.1

0
1/06 4/06 7/06 10/06 1/07 4/07 7/07 10/07 1/08 4/08 7/08 10/08 1/09 4/09 7/09 1

i
Figure 7: Dynamics of Variance Decomposition in the Cross-Sectional Regression: Basisit = γt,cp βcp + γt,l (repo)βli +
i
γt,f (libor)βfi +γt,coll Collaterali +γk βcontrol +εit , where βli is measured by the change of the repo spread (3m repo rate with
general collateral of Treasury bill minus the 3m Treasury bill rate) and βfi is measured by the change of the libor-ois
spread.
 Table 1: Summary Statistics of Discrepancies in CDS and Cash Bond Markets
This table provides descriptive statistics for the CDS-bond basis in three phases: Phase 1 is the period before the
subprime credit crisis, named “Before Crisis” (1/2/2006 - 6/30/2007), Phase 2 is the period between the subprime credit
crisis and the bankruptcy of Lehman Brothers, called “Crisis I” (7/1/2007 - 8/31/2008), and Phase 3 is the period after
Lehman Brothers’ failure, called “Crisis II” (9/1/2008 - 9/30/2009). The basis is calculated as the difference between
the CDS spread and the par-equivalent corporate bond spread, using the methodology in Appendix A. Panel A presents
the average results across investmeng-grade and high-yield firms, financial and non-financial firms; Panel B provides
further results according to firm credit rating; and Panel C reports the results by firm’s industry sector. All numbers
are in basis points.

Panel A

All IG HY Financial Non-Financial

Before Crisis Mean -3 -7 83 -17 4

(T=375 days) Std.Err 9 8 30 12 10
Min -22 -26 -9 -43 -23
Max 29 19 163 36 22

Crisis I Mean -21 -27 11 -2 -24

(T=295 days) Std.Err 22 19 53 28 20
Min -66 -69 -110 -67 -68
Max 20 6 136 47 15

Crisis II Mean -171 -165 -322 -206 -161

(T=271 days) Std.Err 46 44 136 73 43
Min -258 -252 -653 -352 -236
Max -17 -17 -28 11 -20

Panel B: By Credit Rating

IG HY

AAA AA A BBB BB B CCC NR

Before Crisis Mean -20 18 -7 -31 17 170 440 -17

(1/2/2006 - 6/30/2007) Std Err. 11 18 10 13 32 54 271 26
Min -54 -28 -49 -47 -27 -26 156 -58
Max 3 68 15 31 100 285 1211 105
Crisis I Mean 3 4 -35 -50 -22 36 436 32
(7/1/2007 - 8/31/2008) Std Err. 20 21 17 34 41 91 314 55
Min -58 -50 -80 -124 -110 -109 -334 -77
Max 44 43 -4 8 75 232 1232 135
Crisis II Mean 14 -83 -176 -267 -262 -431 -1315 -96
(9/1/2008 - 9/30/2009) Std Err. 47 29 53 76 116 238 1596 260
Min -87 -168 -263 -406 -754 -1244 -6099 -289
Max 188 -11 -1 -87 -41 -3 746 871

46
 Panel C: By Industry Sector

Before Crisis Crisis I Crisis II

Mean Std Min Max Mean Std Min Max Mean Std Min Max

Industrial Firms
Manufacturing 14 13 -20 35 -12 19 -56 23 -141 40 -214 71
Media/Commu 135 94 7 386 8 63 -142 100 -240 84 -443 31
Oil & Gas -22 13 -53 5 -50 17 -105 -12 -196 53 -280 -60
Railroad -50 17 -96 -19 -70 27 -134 5 -225 96 -494 -30
Retail -33 10 -70 -14 -52 16 -92 -22 -170 45 -281 -52
Service/Leisure -33 16 -52 44 -59 26 -112 25 -240 108 -473 -35
Transportation -13 14 -56 15 -66 35 -144 11 -295 89 -530 -99
Telephone -42 19 -69 9 -66 45 -160 38 96 652 -902 1104

Financial Firms
Banking -30 23 -60 80 26 45 -69 97 -227 84 -339 23
Credit/Financing -30 17 -63 14 -161 105 -311 108 -678 570 -2253 9

47
Financial Service 12 16 -29 114 31 36 -59 153 -269 104 -641 36
Insurance -34 7 -63 -16 -48 25 -114 36 -134 64 -356 26
Real Estate -24 13 -47 5 0 26 -55 76 -308 215 -1071 100
Leasing -54 19 -87 12 -32 77 -139 299 -770 970 -4141 64
 Table 2: Correlation Matrix of Interest Rates
This table shows the correlation values for five interest rates and their spreads. All interest rates have a three-month
maturity. The correlation is based on 942 daily observations spanning from January 2006 to September 2009.

Panel A: Level
Libor OIS RepoMBS RepoGC T-Bill

Libor 1.00
(0.00)
OIS 0.96 1.00
(0.00) (0.00)
RepoMBS 0.98 0.99 1.00
(0.00) (0.00) (0.00)
RepoGC 0.96 0.99 0.99 1.00
(0.00) (0.00) (0.00) (0.00)
T-Bill 0.94 0.99 0.97 0.99 1.00
(0.00) (0.00) (0.00) (0.00) (0.00)

Panel B: Spread
Libor - OIS OIS - RepoGC RepoMBS - RepoGC RepoGC - T-Bill

Libor - OIS 1.00

(0.00)
OIS - RepoGC -0.11 1.00
(0.00) (0.00)
RepoMBS - RepoGC 0.68 0.45 1.00
(0.00) (0.00) (0.00)
Repo GC - T-Bill 0.34 0.07 0.36 1.00
(0.00) (0.03) (0.00) (0.00)

Panel C: Change of Spread

∆Libor - OIS ∆OIS - RepoGC ∆RepoMBS - RepoGC ∆RepoGC - T-Bill

∆Libor - OIS 1.00

(0.00)
∆OIS - RepoGC -0.14 1.00
(0.30) (0.00)
∆RepoMBS - RepoGC 0.14 0.64 1.00
(0.00) (0.00) (0.00)
∆RepoGC - T-Bill 0.10 -0.76 -0.53 1.00
(0.00) (0.03) (0.00) (0.00)
48
 Table 3: The Time-series Determinant of the Average Basis: Univariate Regression
The basis is value-weighted by market capitalization across all firms, investment-grade firms, and high-yield firms. The candidate explanatory
variables include (1) the 5-year CDS spread of the primary dealer index, (2) the 3-month libor-ois spread, (3) the 3-month repo -Tbill spread,
(4) the change of primary dealers’ position (dollar volume) in long-term corporate bond, (5) the CBOE Volatility Index, (6) the 3-month
repo rate spread between MBS collateral and Treasury collateral, (7) the noise factor for illiquidity in Hu, Pan and Wang (2010), (8) the
stock return of the primary dealer index, and (9) the 3-month ois-repo spread. We test the impact of contemporaneous and up to three
biweekly lagged explanatory variables in the following regression:

∆BasisJt = α0 ∆Zt + α1 ∆Zt−1 + α2 ∆Zt−2 + α2 ∆Zt−3 + εt , J ∈ {All, IG, HY }

The t-statistics are reported in the parentheses using Newey-West standard errors with a lag of two. Due to the data availability of primary
dealer’s position on bond, we use biweekly data spanning from January 2006 to September 2009.

(1) (2) (3) (4) (5) (6) (7) (8) (9)

∆Basis ∆CDS P D ∆Libor − OIS ∆RepoSpread ∆P ositionP D ∆V IX ∆RepoM BS ∆N oise RP D ∆OIS − Repo
t 0.33 -0.02 0.17 0.01 0.04 -0.28 -0.03 -1.17 -0.02
(1.92) (-0.19) (1.38) (0.42) (0.48) (-1.47) (-0.76) (-0.93) (-0.27)
t-1 -0.53 -0.36 -0.24 -0.04 -0.24 -0.07 -0.15 2.17 -0.26
(-2.87) (-2.25) (-1.72) (-1.32) (-2.66) (-0.13) (-2.52) (0.87) (-2.37)
∆BasisAll t-2 -0.02 -0.19 -0.31 0.04 0.00 -0.47 0.06 0.53 0.19
(-0.08) (-1.32) (-1.43) (1.59) (0.05) (-1.59) (0.97) (0.27) (2.57)

49
t-3 -0.16 0.18 -0.27 0.06 0.09 -0.05 -0.04 -0.56 -0.05
(-0.86) (1.40) (-1.98) (1.52) (1.05) (-0.18) (-0.90) (-0.51) (-0.62)
2
Radj 0.21 0.10 0.05 0.00 0.17 0.04 0.17 0.03 0.34
t 0.16 -0.02 0.15 0.01 -0.01 -0.28 -0.03 -0.16 -0.04
(1.60) (-0.28) (1.56) (0.65) (-0.24) (-1.80) (-1.39) (-0.21) (-1.02)
t-1 -0.26 -0.40 -0.17 -0.01 -0.13 -0.39 -0.08 -0.12 -0.02
(-2.59) (-6.70) (-1.57) (-0.46) (-2.36) (-1.50) (-2.64) (-0.24) (-0.80)
∆BasisIG t-2 -0.31 -0.16 -0.38 0.01 -0.05 -0.21 -0.01 0.81 0.03
(-3.92) (-2.00) (-3.27) (0.76) (-1.73) (-3.10) (-0.47) (1.44) (1.27)
t-3 -0.09 0.04 -0.24 0.02 0.02 0.15 0.01 -0.26 -0.07
(-1.00) (0.58) (-2.73) (1.37) (0.60) (1.38) (0.29) (-0.57) (-1.56)
2
Radj 0.38 0.39 0.27 -0.02 0.19 0.19 0.17 -0.01 0.06
t -0.12 0.01 -0.03 0.01 -0.32 -0.41 -0.25 3.46 -0.29
(-0.26) (0.08) (-0.08) (0.42) (-2.42) (-0.80) (-3.15) (1.36) (-2.79)
t-1 -0.76 -1.31 -0.11 -0.02 -0.34 -1.12 -0.08 1.48 -0.04
(-2.31) (-5.44) (-0.35) (-0.28) (-3.11) (-1.54) (-1.04) (1.13) (-0.38)
∆BasisHY t-2 -0.53 0.22 -0.94 0.10 -0.14 -0.56 -0.03 4.11 0.04
(-2.15) (0.82) (-2.69) (2.68) (-1.63) (-1.13) (-0.35) (2.24) (0.32)
t-3 -0.02 -0.51 0.02 0.02 0.01 -0.21 -0.03 -3.90 -0.07
(-0.08) (-2.03) (0.08) (0.31) (0.12) (-0.87) (-0.45) (-2.23) (-0.44)
2
Radj 0.14 0.34 0.11 -0.01 0.25 0.12 0.21 0.18 0.10
 Table 4: The Time-series Determinant of The Average Basis: Multivariate Regressions
The basis is value-weighted by market capitalization across all firms, investment-grade firms, and high-yield firms. The candidate explanatory
variables include (1) the 5-year CDS spread of the primary dealer index, (2) the 3-month libor-ois spread, (3) the 3-month repo-Tbill spread,
(4) the change of primary dealers’ position (dollar volume) in long-term corporate bond, (5) the CBOE Volatility Index, (6) the 3-month
repo rate spread between MBS collateral and , (7) the noise factor for illiquidity in Hu, Pan and Wang (2010), (8) the stock return of the
primary dealer index and (9) the 3-month ois-repo spread. Based on the univariate results in Table 3, we choose the lag of explanatory
variables and use them in the multivariate regression. The t-statistics are reported in the parentheses using Newey-West standard errors
with a lag of two. Due to the data availability of primary dealer’s position on bond, we use biweekly data spanning from January 2006 to
September 2009.

50
∆CDS ∆Libor − OIS ∆V IX ∆N oise ∆OIS − Repo ∆RepoSpread ∆P ositionP D ∆RepoM BS RP D 2
Radj
(t-1) (t-1) (t-1) (t-1) (t-1) (t-2) (t-2) (t-2) (t-2)

∆BasisAll
t -0.45 -0.11 -0.05 -0.06 -0.20 -0.21 0.03 -0.58 0.57
(-2.85) (-0.82) (-1.44) (-2.14) (-1.84) (-1.92) (1.44) (-2.48) (1.05) 0.48

∆BasisIG
t -0.10 -0.25 -0.04 -0.04 0.02 -0.16 0.02 -0.25 0.58
(-1.98) (-2.59) (-1.76) (-2.55) (0.64) (-1.57) (1.24) (-4.09) 2.09 0.50

∆BasisHY
t -0.46 -0.67 -0.11 0.02 0.05 -0.54 0.10 -0.83 3.98
(-2.21) (-2.48) (-1.05) (0.44) (0.68) (-2.52) (2.69) (-1.52) (3.13) 0.45
 Table 5: Correlation Matrix of Time-Series Risk Factors
This table shows the correlation values for variables that explain the time-series variation of the average basis, including the CDS spread of
primary dealer index, the three-month libor-ois spread, the three-month repo spread, primary dealer’s position in corporate bond, volatility
index, the three-month repo rate spread between MBS collateral and Treasury collateral, the noise factor for illiquidity in Hu, Pan and
Wang (2010), the stock return of the primary dealer index and the ois-repo spread. The p-values are reported in the parentheses below the
correlation.

∆CDS P D ∆Libor-OIS ∆ RepoSpread∆P ositionP D ∆VIX ∆RepoMBS ∆Noise RP D ∆OIS-Repo

∆CDS P D 1.00
(0.00)
∆Libor-OIS 0.38 1.00
(0.00) (0.00)
∆RepoSpread 0.29 0.01 1.00

51
(0.00) (0.94) (0.00)
∆P ositionP D -0.03 0.00 -0.06 1.00
(0.78) (0.96) (0.58) (0.00)
∆VIX 0.34 0.38 0.01 0.13 1.00
(0.00) (0.00) (0.91) (0.21) (0.00)
∆RepoMBS 0.20 0.58 -0.33 -0.02 0.27 1.00
(0.05) (0.00) (0.00) (0.84) (0.01) (0.00)
∆Noise 0.13 0.23 0.18 -0.03 0.30 0.16 1.00
(0.21) (0.02) (0.07) (0.75) (0.00) (0.12) (0.00)
RP D -0.15 0.15 0.02 0.00 -0.37 0.02 -0.25 1.00
(0.13) (0.15) (0.81) (1.00) (0.00) (0.86) (0.01) (0.00)
∆OIS-Repo 0.13 -0.04 -0.05 0.01 0.90 -0.24 0.21 -0.15 1.00
(0.19) (0.72) (0.66) (0.96) (0.06) (0.02) (0.03) (0.15) (0.00)
 Table 6: Correlation Matrix of Cross-sectional Risk Factors
This table shows the correlation values for betas that capture counterparty risk, liquidity risk, funding cost risk, and
collateral quality, which potentially explain the cross-sectional variation of the basis. The correlations are calculated
using the whole sample, three subsamples of Before Crisis (1/2/2006 - 6/30/2007), Crisis I (7/1/2007 - 8/31/2008), and
Crisis II (9/1/2008 - 9/30/2009).

Full Sample (1/2/2006 - 9/30/2009)

βcp βl (repo) βf (libor) Collateral

βcp 1 - - -
βl (repo) 0.32 1 - -
βf (libor) 0.38 0.62 1 -
Collateral -0.16 -0.15 -0.33 1

Before Crisis (1/2/2006 - 6/31/2007)

βcp βl (repo) βf (libor) Collateral

βcp 1 - - -
βl (repo) -0.11 1 - -
βf (libor) 0.07 0.08 1 -
Collateral -0.01 0.23 -0.19 1

Crisis I (7/1/2007 - 8/31/2008)

βcp βl (repo) βf (libor) Collateral

βcp 1 - - -
βl (repo) 0.37 1 - -
βf (libor) 0.42 0.91 1 -
Collateral -0.25 -0.19 -0.19 1

Crisis II (9/1/2009 - 9/30/2009)

βcp βl (repo) βf (libor) Collateral

βcp 1 - - -
βl (repo) 0.42 1 - -
βf (libor) 0.48 0.89 1 -
Collateral -0.11 -0.07 -0.04 1

52
 Table 7: Cross-Sectional Regression of the CDS-Bond Basis on Risk Factors
This table shows the Fama-MacBeth cross-sectional regression results of the CDS-bond basis on the following variables:

BasisJi = α + γcp βi,cp + γl βi,lrepo + γf βi,flibor + γcoll Collaterali + γk βi,controls + εi , J ∈ [All, IG, HY ]

where
cov(Ri , (Rindex − Rmkt )
βi,cp = ,
var(Rindex − Rmkt )
cov(∆CDSi , ∆repospread)
βi,lrepo = ,
var(∆repospread)
cov(∆CDSi , ∆(libor − ois))
βi,flibor = .
var(∆(libor − ois))

Collateral is an index measuring the collateral quality of the bond issued by each reference entity, composed of firm size, leverage, rating,
tangible ratio, CDS level and CDS volatility. Each day we run the Fama-MacBeth regression using the betas calculated in corresponding
periods (Full sample, Before Crisis, Crisis I, and Crisis II). The table reports the average of the cross-sectional regression estimates, and
the standard deviations of the estimates in parentheses. The standard deviations are adjusted for the correlations among sampling errors.
Estimated coefficients are in bold script if the statistical significance is 5% or below.

53
Panel A: All Firms

The Whole Sample Before Crisis Crisis I Crisis II

(Jan 2006 - Sep 2009) (Jan 2006 - Jun 2007) (Jul 2007 - Aug 2008) (Sep 2008- Sep 2009)
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12)
γcp -0.86 -0.71 -1.49 0.13 0.20 0.12 -0.57 -0.56 -0.39 -2.55 -2.12 -4.96
(0.27) (0.25) (0.47) (0.02) (0.03) (0.02) (0.06) (0.06) (0.04) (0.77) (0.72) (1.21)
γlrepo 2.60 1.67 3.63 3.56 0.30 1.05 3.79 -0.26
(0.66) (0.64) (1.04) (1.02) (0.35) (0.38) (1.58) (1.47)
γflibor -1.09 -1.02 0.50 0.44 0.98 1.01 -5.64 -5.33
(0.51) (0.49) (0.10) (0.09) (0.18) (0.17) (1.79) (1.70)
γcoll 0.19 0.18 0.28 -0.20 -0.19 -0.19 -0.01 -0.01 -0.03 0.97 0.90 1.28
(0.08) (0.08) (0.11) (0.01) (0.01) (0.01) (0.02) (0.02) (0.02) (0.15) (0.15) (0.21)
Adj R2 0.11 0.09 0.08 0.09 0.06 0.09 0.03 0.03 0.02 0.21 0.19 0.12
 Table 7: Cross-Sectional Regression of the CDS-Bond Basis on Risk Factors (Cont’d)

Panel B: IG Firms

The Whole Sample Before Crisis Crisis I Crisis II

(Jan 2006 - Sep 2009) (Jan 2006 - Jun 2007) (Jul 2007 - Aug 2008) (Sep 2008- Sep 2009)
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12)
γcp -0.48 -0.52 -0.44 0.08 0.07 0.08 -0.28 -0.34 -0.21 -1.46 -1.52 -1.40
(0.14) (0.15) (0.15) (0.01) (0.01) (0.01) (0.05) (0.06) (0.04) (0.36) (0.38) (0.41)
γlrepo -2.49 -1.50 -2.99 -3.06 -3.16 -0.89 -1.09 -0.08
(0.41) (0.47) (0.36) (0.32) (1.09) (1.15) (0.43) (0.50)
γflibor 0.37 0.08 -0.18 -0.49 0.83 0.54 0.60 0.33
(0.15) (0.15) (0.10) (0.09) (0.15) (0.17) (0.42) (0.40)
γcoll 0.10 0.10 0.10 0.06 0.06 0.06 0.09 0.09 0.09 0.16 0.18 0.16
(0.01) (0.01) (0.01) (0.00) (0.00) (0.00) (0.01) (0.01) (0.01) (0.04) (0.04) (0.03)
Adj R2 0.04 0.03 0.03 0.01 0.01 0.03 0.03 0.03 0.03 0.09 0.08 0.06

Panel C: HY Firms

54
γcp -2.70 -2.60 -3.00 -0.48 -0.04 -0.60 -1.01 -0.97 -0.71 -7.57 -7.86 -8.80
(0.59) (0.62) (0.65) (0.05) (0.06) (0.06) (0.12) (0.11) (0.06) (1.27) (1.24) (1.19)
γlrepo 2.36 1.98 6.27 6.09 0.53 0.91 -0.89 -2.36
(0.93) (1.04) (1.38) (1.36) (0.31) (0.31) (2.05) (2.51)
γflibor 0.30 0.10 0.64 0.56 1.20 1.20 -1.15 -1.74
(0.26) (0.28) (0.14) (0.12) (0.28) (0.28) (0.65) (0.68)
γcoll -0.01 0.02 0.02 -0.60 -0.55 -0.61 -0.11 -0.11 -0.14 0.90 0.92 1.05
(0.12) (0.12) (0.14) (0.02) (0.02) (0.03) (0.03) (0.03) (0.03) (0.28) (0.28) (0.30)
Adj R2 0.12 0.09 0.11 0.10 0.04 0.01 0.02 0.03 0.01 0.25 0.21 0.22
 Table 8: Test of the Impact of Deleveraging
This table shows the influence of the change of bond trading volume on the the change of basis between the post-Lehman
period and the before-subprime period, using the following cross-sectional regression:

∆BasisJi = β1 (∆V ol/V ol)Ji + β2 (∆(High − Low)/M ean)Ji + εJi , J ∈ [IG, HY, F, N F ],

where ∆Basis ≡ Basis2 − Basis0 ; ∆V ol/V ol ≡ (V ol2 − V ol0 )/V ol0 is the percentage change of corporate bond trading
volume (source: TRACE); High is the maximum monthly trading volume in a specific phase, and Low is the minimum
monthly trading volume in that phase, Mean is the average monthly volume in that phase – this measure shows the
volatility of trading volume change. The subscript 2 refers to the phase of Crisis II (9/1/2008 - 9/30/2009) and the
subscript 0 refers to the phase of Before Crisis (1/2/2006 - 6/30/2007). The t-statistics are reported in the parentheses
below the regression coefficients.

∆Basis ∆V ol
V ol ∆ High−Low
M ean R2

IG -0.17 -0.53 0.02

(-1.12) (-1.58)
HY -0.35 -3.02 0.06
(-0.61) (-2.23)

Financials -2.04 -1.02 0.05

(-1.23) (-0.87)
NonFinancials -0.12 -0.08 0.03
(-0.71) (-2.55)

55
 Table 9: The CDS Market’s Contribution to Price Discovery
This table reports the summary statistics of the contribution to the credit price discovery made by the CDS market.
We do the price discovery test for each firm between its CDS spread and bond-implied CDS (PECDS) spread using the
daily data. We summarize the mean, median, and standard error for all firms, IG, HY, financial, and non-financial firms
during each phase of the financial crisis, – Pre-Crisis (1/2/2006 - 6/30/2007), Crisis I (7/1/2007 - 8/30/2008), Crisis II
(9/1/2008 - 9/30/2009). The test is based on the vector error correction model as
p
X p
X
∆CDSt = λ1 (CDSt−1 − P ECDSt−1 ) + β1 ∆CDSt−i + γ1 ∆P ECDSt−i + ε1,t
i=1 i=1
Xp Xp
∆P ECDSt = λ2 (CDSt−1 − P ECDSt−1 ) + β2 ∆CDSt−i + γ2 ∆P ECDSt−i + ε2,t
i=1 i=1

We use the Granger-Gonzalo measure λ2 /(λ2 − λ1 ), which indicates the price discovery contribution made by the CDS
market for each firm.

All
Pre-Crisis Crisis I Crisis II
Mean 0.92 0.84 0.74
Median 0.91 0.93 0.89
Std Err. 0.86 0.78 0.38

IG HY
Pre-Crisis Crisis I Crisis II Pre-Crisis Crisis I Crisis II
Mean 0.84 0.83 0.79 Mean 1.26 0.90 0.46
Median 0.91 0.94 0.92 Median 0.95 0.83 0.60
Std Err. 0.65 0.74 0.32 Std Err. 1.49 0.96 0.50

Non-Financial Financial
Pre-Crisis Crisis I Crisis II Pre-Crisis Crisis I Crisis II
Mean 0.91 0.83 0.73 Mean 0.92 0.94 0.76
Median 0.91 0.93 0.89 Median 0.91 0.93 0.89
Std Err. 0.95 0.84 0.39 Std Err. 0.20 0.34 0.28

56