Exposing The Exposed: Intermediation Capacity in the

Credit Default Swap Market

Or Shachar ∗

Federal Reserve Bank of New York †

August 2012

Abstract

I examine the role of liquidity provision by dealers in the credit default swap (CDS) mar-

ket, using traded volume of a sample of CDS contracts on North American financial firms

during 2007 – mid 2009, a period that notably includes the financial crisis. Consistent with the

standard microstructure literature, I find that order imbalances of end-users cause significant

price impact, which depends on their direction relative to the direction of dealers’ inventory.

In addition, I establish that counterparty risk, measured by the level of exposure in the inter-

dealer market, limits the willingness of dealers to provide liquidity. The more “congested” the

interdealer market becomes from a build-up of bilateral credit exposure, the more averse deal-

ers become towards inventory risk. Dealers’ desire to hedge counterparty risk leads to limited

intermediation in the CDS market.

∗I am grateful to my advisers Viral V. Acharya, Robert Engle, Joel Hasbrouck, and Itamar Drechsler for fruitfuldiscussions and generous guidance. I thank Patrick Augustin, Joshua Coval (WFA discussant), Thomas Philippon,Oliver Randall, Patrik Sandas (discussant at the Conference on Current Topics in Financial Regulation), MartiSubrahmanyam, Raghu Sundaram, Michelle Zemel, and seminar participants at Arizona State University, FederalReserve Bank of Chicago, Federal Reserve Bank of New York, Federal Reserve Board’s Division of Banking Supervisionand Regulation, Hofstra University, NYU Stern, Ohio State University, Princeton University, University of Illinoisat Urbana-Champaign, University of Maryland, and U.S. Commodity Futures Trading Commission, Conference onCurrent Topics in Financial Regulation, and WFA 2012 for helpful comments. I thank Peter Axilrod, Marisol Collazo,Michael Kopcak and Gary Kotlyar from The Depository Trust & Clearing Corporation for providing the data andfor explaining its features. All errors are mine. The views expressed here are the author’s and are not representativeof the views of the Federal Reserve Bank of New York or of the Federal Reserve System.†Send correspondence to Federal Reserve Bank of New York, Capital Markets Function, 33 Liberty Street, New

York, NY 10045-0001. Email: or.shachar@ny.frb.org.

1 Introduction

A central question concerning dealer markets is understanding the role which dealers play in

liquidity provision. The existing microstructure literature addresses this question in equity,

bond, option and foreign exchange (FX) markets using information-based (e.g., Glosten and

Milgrom, 1985, Easley and O’Hara, 1987), inventory control (e.g., Garman, 1976, Stoll, 1978,

Amihud and Mendelson, 1980, Ho and Stoll, 1983), and search-and-bargaining (e.g., Duffie,

Garleanu, and Pedersen, 2005, Lagos and Rocheteau, 2009, Lagos, Rocheteau, and Weill, 2011)

models. In this paper, I attempt to answer this key question using a comprehensive proprietary

dataset of Credit Default Swap (CDS) transactions. My empirical study shows that, to an

extent, standard microstructure models can be used to understand the functioning of the

CDS market in normal times, but that counterparty risk is also empirically relevant in the

determination of the intermediation capacity of dealers and in the resilience of the market

during times of stress.

In order to understand the effect of the counterparty risk friction on intermediation, it is

necessary to understand how CDS contracts trade. Unlike equities or bonds, CDS contracts

do not trade by means of transfer of ownership. Instead, they involve a bilateral agreement

that the protection buyer will pay the protection seller regular premium payments until either

the contract matures or the default of the associated credit instrument (for example, a bond

issued by a firm), whichever occurs first. In return, if the associated credit instrument defaults,

the seller of protection pays the buyer for the loss, and the buyer ceases paying premia to the

seller. As such, either party of a CDS contract is exposed to loss both through the performance

of the underlying asset and through the potential default of the counterparty. The possibility

that the other party of the contract will default before meeting all of his obligations is often

referred to as counterparty risk.1

As a liquidity provider, a dealer firm enters into a transaction with a client even if it does

not wish to retain the exposure. To balance its exposures, the dealer firm enters into offsetting

hedge transactions.2 Over time, the dealer may engage in a large number of interdealer and

other hedge transactions to adjust its positions and limit exposure to market movements.

Unlike cash markets, where the passing of inventory does not leave any residual obligations to

the original seller or buyer, these hedge transactions in the CDS market might eliminate or at

least reduce credit risk, but they do not cancel any previous contract. As such, as long as the

contract is in place, the counterparty risk remains, and builds up as these transactions involve

many counterparties. If a counterparty defaults, the dealer’s hedged exposure will change

1See Deutsche Bank, 2009, “Credit Default Swaps: Heading Towards a More Stable System” for further detailson how CDS contracts work.

2There are many ways in which a dealer firm might choose to offset its exposure. It might choose the most basichedging strategy of entering into offsetting transactions in the same reference entity within a short time period sur-rounding end-users trades; alternatively, the exposure can be hedged in the bond market. In this paper I concentrateonly on the CDS market.

1

and the dealer will consequently find itself exposed to unanticipated market movements. To

neutralize the exposures, the dealer will need to adjust the portfolio to bring it back into

balance, by replacing the defaulted transactions, unwinding the offsetting hedge transactions,

or both. However, if the default of the counterparty is a harbinger of additional credit problems,

non-defaulting parties would have incentives either to cut back or terminate transactions with

troubled counterparties, and the dealer will lose its ability to engage in offsetting trades. At

the aggregate level, this in turn will affect the ability of other dealers to supply liquidity.

This ongoing hedging and re-balancing process by dealers, coupled with the empirical fact

that the daily market turnover of interdealer trading accounts for more than 60% of total

contract turnover, and that each non-dealer customer trade is passed between three dealers on

average, imply that the concentration of exposures in the interdealer market is a good proxy

for the potential consequences of the exit of a major market participant if all the transactions

along the chain of offsetting positions are to be untangled. I refer to this accumulation of credit

risk within the dealing community as “congestion”. Therefore, to determine what affects the

dealers’ capacity to supply liquidity, I explore the interaction between the exposure level among

dealers and the ability of dealers to provide liquidity. I expect that when the interdealer market

is congested, dealers find it harder to mitigate their inventory risk and become more averse to

supplying liquidity to their clients.

It is important to note that this issue has been identified by market participants and reg-

ulators as problematic to the functioning of the market. Since 2008, procedures to reduce the

redundant positions have been developed. The procedure that appears to be most efficient for

reducing the value of outstanding positions is referred to as “compression”. Compression nets

the positions across dealers, taking into account the limits established by dealers on counter-

party risk of other dealers and identifying contracts that can be eliminated or replaced with

new, lower notional value contracts, with the aim of keeping the same risk profile. Compres-

sion, therefore, has resulted in a great reduction of the gross notional value of CDS positions,

but has not resolved the issue completely.

The economic importance of the mechanism of congestion in the interdealer market that

lies at the heart of this paper is reflected in the size of the CDS market. At the end of 2010, the

gross notional value of CDS positions registered at the Depository Trust Clearing Corporation

(DTCC) amounted to $26.3 trillion. Of this total, $15 trillion was single name CDSs and $11.3

trillion was index or tranche CDSs. But these gross notional amounts include a large number

of offsetting trades that can be netted down across investors. For single-name CDS, the $15

trillion gross notional amount nets down to a $1.2 trillion net notional amount (8% of the gross

notional value), while for index and tranche CDS the $11.3 trillion gross notional amount nets

down to a $1.1 trillion net notional amount (approximately 10% of the gross notional value).

Despite the size of the CDS market and its importance, as the financial crisis has illustrated,

there are virtually no empirical studies using transaction flows and dealer inventories. Lack of

2

data on trading volume figures has been the primary impediment to such research. This paper

is, to the best of my knowledge, the first to study the role of dealers in the CDS market. The

empirical study of CDS dealers is particularly interesting since low regulation has allowed the

patterns of liquidity provision to evolve endogenously. I analyze a unique dataset of transactions

between dealers and their customers spanning from February 2007 to June 2009 for CDS

contracts on 35 North American financial firms. This sample is obtained from the DTCC, the

only central trade registry. DTCC claims to cover about 90% of the notional traded in the

CDS market, hence, my sample reflects the vast majority of market activity in the 35 reference

entities.

The data contain an anonymized identification of the parties of each transaction. Each

individual market player has a consistent identifier throughout the dataset and a classification

of its type (dealer vs end-user). Hence, I am able to construct the daily order imbalances

of dealers and end-users and study the impact of order imbalances. I refer to a non-dealer

customer as an end-user. End-users in the sample include hedge-funds, pension funds, asset

managers, banks, and non-financial companies. The dealers in the sample include the G14

dealers, who are the major 14 over-the-counter (OTC) derivatives dealers, as well as some

smaller dealer firms.

Using this sample, I first establish benchmark results that are drawn from the standard

microstructure models. For each reference entity I examine (i) the association between the

change in the CDS price and the imbalance between buyer- and seller-initiated orders arriving

at the market; and, (ii) how dealers’ inventory levels affect this price impact. Since order flow

might be correlated with fundamental information, I control for contemporaneous and lagged

equity returns (following Acharya and Johnson, 2007). Both Collin-Dufresne et al. (2001) and

Schaefer and Strebulaev (2008) show that equity price changes seem to be adequate to capture

all fundamental information related to default risk, so the unexplained component of CDS price

changes beyond that can be attributed to microstructure frictions.

I find that there are different price reactions to buying versus selling by end-users. The mar-

ket imputes an information content to the trade, both for buying and selling by end-users. Yet,

the contemporaneous price impact of buying is much larger than the immediate price impact of

selling. Specifically, I find that a one standard-deviation increase in net notional amount bought

by end-users is associated with a 0.3% contemporaneous increase in CDS spreads, whereas a

one standard-deviation increase in net notional amount sold by end-users predicts a contempo-

raneous decrease of 0.06%, which is statistically insignificant, and about one-fifth the impact

of buying. The result that the market reacts differently to buy and sell orders resembles the

results that emerge from the empirical research on block transactions and institutional trades

in equity markets (for a survey of these results see Saar, 2001). The empirical results also

suggest that the price impact is permanent, as the effect persists for at least five days. The

asymmetric impact of buys and sells suggests that end-users face a more illiquid market when

3

they are buying default protection compared to when they are selling protection to the dealers.

The results are consistent with dealers being naturally net long credit risk, and, therefore,

being less willing to sell protection, as it will further increase their exposure to the financial

sector.

As for the effect of dealers’ inventory on price impact, I find evidence consistent with the

microstructure models, such as, Madhavan and Smidt (1991), Madhavan and Smidt (1993), and

Hasbrouck and Sofianos (1993) which predict a lesser price response for order flow that moves

dealer inventory to a preferred position, and that order flow which increases dealer inventory

exposure is associated with a larger price impact. The intuition behind this result follows from

the following example. If dealers hold a long position of CDS protection (i.e., they are short

credit risk) at the end of day t−1, net-selling by end-users during day t should increase dealers’

long position further. Therefore, to deter additional potential sellers, or conversely to attract

potential buyers, dealers will lower their quoted prices, intensifying the price impact of the

selling pressure. If, however, during day t, end-users are net-buyers, this should help dealers

revert to their target inventory level (assuming that the quantity demanded by end-users is

smaller than the level of dealers’ inventory).

The benchmark specification, however, is an incomplete description of the CDS trading

architecture as it excludes a key feature – counterparty risk. To this end, I construct a measure

of exposure in the interdealer market based on pairwise exposures between dealers. Using this

measure, I demonstrate that when dealers are more exposed to each other their willingness

to intermediate markets is reduced and they become less predisposed to absorb supply and

demand shocks. Furthermore, the congestion in the interdealer market has a secondary impact

by discouraging dealers from holding buffer inventories. I find that these two effects increase

the price impact of order imbalances, and that dealers’ inventories matter even more when

significant counterparty risk is built up amongst the dealers. Though the key component of

the price impact in “normal” times is a result of the quantity the end-users demand, the

interdealer exposure effect kicks in when the inventory-absorption capacity of the dealers in

the CDS market is strained.

To gauge the economic magnitude of each of the effects, I decompose the total price impact

of trading into the key components found in this paper for: (a) a benchmark case that is

based on mean values of equity and CDS returns, dealers’ inventory level and end-users’ net

buying imbalance, and for (b) an extreme case when the realizations of all variables are at their

maximum. Using the mean level of the variables, the “basic” price impact accounts for 28%

and 7% of the total price impact for buys and sells, respectively, and an incoming order-flow

that will increase dealers’ inventory results in 16% and 23% of the price impact for buys and

sells, respectively. While the congestion effect plays a minor role in normal times, it accounts

for as much as 23% and 37% of the total price impact, for buys and sells, respectively, if all

independent variables are taken at their maximum values. In either normal times or crisis

4

times, the combined microstructure effect of end-users’ order imbalance, dealers’ inventory

and interdealer exposure congestion (including the interactions between the three) is about

the same magnitude as the impact of information flow from the stock market and lagged CDS

spreads. These microstructure effects are not “informationless” and they actually generate new

fundamental information over and above the information generated in the equity markets. I also

find that both end-users’ order imbalances in the CDS market and the information innovations

from the CDS market predict stock returns 1-day ahead.

The results for the single-name contracts are contrasted with the trading in the CDX index,

a basket product. Apart from the information advantage one might have when trading a single-

name contract, the trading mechanism of the index is inherently different. Rather than using

offsetting trades to unwind positions, the party that wishes to unwind his position “assigns”

the trade, i.e. transfers his side of the trade, to a new party. Therefore, I repeat the analysis

using the transactions of the CDX index and I find results that are consistent with the use of

the index for hedging purposes by end-users. In the CDX index, dealers do not seem to play a

significant role in liquidity provision.

Finally, to support the main results in the paper, I also provide evidence that dealers’

inventories mean-revert in a non-linear fashion. Underlying the inventory result lies an implicit

assumption that dealers have a desired level of inventory. As they buy and sell, their inventory

deviates from the target and, therefore, they require compensation. Rather than just assume

this, I test a prediction of Ho and Stoll (1983) about competitive dealer markets pertaining to

mean reversion in dealers’ relative inventories, using the methodology in Hansch et al. (1998).

Similar to the results found by Hansch et al. (1998) for the London Stock Exchange dealers,

I also find evidence that dealers with extreme inventory positions trade larger quantities that

move their inventories toward desired levels, and that the intensity of mean reversion increases

with the deviation of the inventory from its desired level. Nevertheless, the time-scale of mean

reversion is not days, but weeks or months. The half-life of a dealer’s inventory that is one

standard-deviation away from the median level of all dealers is about 160 days, compared to 11

days for an inventory level that is five standard-deviations from the median. Dealers’ persistent

positions are also consistent with the result that the price impact it permanent.

The novel contribution of this paper – the “congested interdealer market” hypothesis – can

be extended to other markets trading bilateral agreements, such as, FX derivatives, interest

rate swaps, and OTC equity derivatives, where counterparty risk is present, and to markets

that are linked to them through an arbitrage relationship. For example, in related work, Choi

and Shachar (2011) use this hypothesis to shed light on the persistent “violation” of the close

relation between CDS and corporate bond spreads – the CDS-Bond Basis – during the crisis.3

3Bai and Collin-Dufresne (2010) and Fontana (2011) also study the deviation of the CDS-bond basis from parityduring the financial crisis. These papers document association with the following risk factors: funding cost risk,counterparty risk, and collateral quality. Choi and Shachar (2011), on the other hand, focus on counterparty risk andthe transmission mechanism between dealers’ liquidity provision and hedge funds’ trading.

5

My results have a number of important implications. From a modeling perspective, my

findings suggest that counterparty exposure in the interdealer market is a critical determinant

of liquidity provision by dealers. It follows that existing OTC trading models can benefit

from a better understanding of the structure of the interdealer network and how exposures

are spread in this network. The microstructure frictions found in this paper also help to shed

some light on the possible use of CDS prices for regulation purposes, e.g. Hart and Zingales

(2009). The economic magnitude of these frictions should be estimated and accounted for before

extrapolating the credit risk component from CDS prices. Finally, my results give insight

regarding the potential effects of a central clearing counterparty (CCP) in the CDS market

that is currently being discussed by policy makers (Duffie and Zhu, 2009, Acharya, Shachar,

and Subrahmanyam, 2010). Given that heightened levels of credit exposure in the interdealer

market inhibit liquidity provision in times of stress, it seems that CCPs will primarily create

value by reducing counterparty risk during crisis periods, conditional on them not succumbing

to aggregate crisis.

Related Literature

This paper is closely related to the inventory control models in dealer markets. In particular, the

key finding of this paper is anchored in a trading mechanism that was identified in the Foreign

Exchange (FX) market. Lyons (1997) studies the repeated passing of inventory imbalances

between dealers, i.e. “hot-potato trading,” in the FX market. He shows that hot-potato

trading reduces the information content of interdealer trades. Since the FX market is also an

OTC derivative market like the CDS market, it is not surprising that I find evidence that such

hot-potato trading also happens in the CDS market. In the context of the CDS market, I find

that this trading mechanism imposes a negative externality as interdealer exposure builds up.

My paper is related to several themes within the CDS literature. Acharya and Johnson

(2007) study the flow of information from the CDS market to the equity market during 2001-

2004. Berndt and Ostrovnaya (2007) conduct a similar analysis of the flow of information across

the CDS, option and equity markets. My work shows that there is also an additional information

in the microstructure effects that flows from the CDS market to the equity market. Arora et al.

(2010) use a proprietary dataset of CDS transaction prices and quotes to identify how dealers’

credit risk affects the prices of CDS contracts. They find that counterparty risk is barely

priced in the CDS market, which is consistent with a market structure in which participants

require collateralization of swap liabilities by counterparties. Whether counterparty risk is

priced directly in the CDS price or not, I show that it imposes a significant cost on trading

in stress times through the liquidity channel. My analysis suggests that the quantities traded

between dealers demonstrate the importance of counterparty risk. Chen, Fleming, Jackson, Li,

and Sarkar (2011) analyze a similar dataset to the one analyzed in this paper. Their dataset

encompasses CDS transactions from DTCC for single-name and index contracts for a period of

6

three months. They focus on the trading frequency in the CDS market and the implications for

post-trading public reporting. Oehmke and Zawadowski (2012) investigate the determinants

of weekly CDS net notionals. Their findings support the view that the CDS market serves as

an alternative trading venue when the corporate bond trading for the reference entity becomes

illiquid.

My paper is also tangential to the literature on OTC markets, and limited risk-bearing ca-

pacity in market-making. The theoretical OTC market literature that was recently reignited by

Duffie, Garleanu, and Pedersen (2005, 2007) identifies counterparty search and bargaining over

the terms of trade as important frictions to trading in this type of market. In these search mod-

els there is no role of liquidity provision by dealers in the face of temporary demand pressures

due to the assumptions that investors hold 0 or 1 unit of the asset and dealers are restricted

from holding asset inventories. Yet, as my empirical results show in the context of the CDS

market, dealers’ inventories appear to be an important determinant of their ability to provide

liquidity and the price of that liquidity to end-users. Lagos, Rocheteau, and Weill (2011) relax

the restriction on dealers’ and investors’ inventory, to allow them to hold any positive amount.

They find that when frictions are severe, even well-capitalized, profit-maximizing dealers may

not find it optimal to “lean against the wind,” that is, to accumulate inventories during the

crisis and to unload these assets when the economy recovers. Though the theoretical predic-

tion of reduced liquidity provision during the crisis seems to be consistent with my empirical

findings in the CDS market, this prediction is driven by a shock to investors’ risk-aversion, not

dealers’. Dealers, however, are able to trade continuously with each other. My main insight

suggests that the lack of liquidity provision by dealers comes from the search of dealers for

other dealers with sufficiently low counterparty risk. A future theoretical model that desires to

explain liquidity provision in the CDS market should incorporate such friction.

In the broader context of intermediated markets, trading capacity constraints result in limits

to arbitrage. Liquidity suppliers profit from providing immediacy to investors. They have,

however, limited risk-bearing, and thus inventory-carrying, capacity. The limits to arbitrage

arguments rely on the idea that these liquidity suppliers will take the role of arbitrageurs and

accommodate buying or selling pressure, only if they are compensated by favorable subsequent

price movements. Different papers identify different sources for what limits arbitrage. Shleifer

and Vishny (1997) show that capital constraints due to agency problems hinder dealer firms

exploiting mispricing. Deviation from fundamental value might be also a result of the time it

takes for capital to come to the market. Grossman and Miller (1988), Mitchell et al. (2007), and

Duffie (2010), to name a few, document alternative reasons for capital dislocations. Gromb and

Vayanos (2002) and Brunnermeier and Pedersen (2009) study how leverage constraints affect

the ability of dealers to supply liquidity to outside clients. These frictions result in some part

of the risk in derivatives becoming unhedgeable, leading to an upward sloping supply curve

(Bollen and Whaley, 2004, Cetin et al., 2006, and Garleanu et al., 2009).

7

2 Hypotheses Development

Trading in the CDS market integrates essentially two parallel trading environments: a trading

environment where end-users trade directly with dealers, and an interdealer trading environ-

ment where dealers are able to share their inventory risks across the dealers community rather

than waiting for the arrival of offsetting customer orders. Hence, the effect of each trading

arena on the other should be carefully assessed. Although I draw from the existing microstruc-

ture literature, the inherent differences of the CDS market architecture pose some questions

regrading the validity of the use of the existing theoretical models that are applied to cash

markets. I try to bridge this gap by supplying empirical evidence on what should be added to

any theoretical model that would describe the CDS market.

Theoretical motivation for the empirical investigation of the order imbalances I undertake

comes from the standard market microstructure models which suggest that trades are associated

with price movements resulting from the presence of trading frictions – such as inventory

costs, information asymmetry, funding constraints, and search costs. In the inventory control

paradigm (e.g., Garman, 1976, Stoll, 1978, Amihud and Mendelson, 1980, Ho and Stoll, 1983,

among others) dealers accommodate buying and selling by outside investors and adjust their

quoted prices to restore their inventories to some desired level. The implication of these models

is that price may depart from expectations of value if the dealer is long or short relative to

its desired inventory, giving rise to transitory price movements during the day and possibly

over longer periods. In the asymmetric information models (e.g., Glosten and Milgrom, 1985,

Easley and O’Hara, 1987, among many others), a subset of the market participants have private

information about the asset’s (expected) value, so order imbalances sometimes signal private

information, which should reduce liquidity at least temporarily, and could also move the market

price permanently, as also suggested by the Kyle (1985) model of price formation.

Hypothesis 1. End-users’ net buying (selling) positively (negatively) affects price changes.

The microstructure literature has extensively explored the relation between trading activity

and returns. Both theory and empirical papers conclude that order imbalances should have an

important effect on asset returns and liquidity. Order imbalance can sometimes signal private

information, which should reduce liquidity at least temporarily and could also move the market

price permanently, as also suggested by the well-known Kyle (1985) model of price formation.

His model predicts that the coefficient on order flow will be positive. Order imbalance can also

be caused by a random large order. Then, it might exacerbate the dealer’s inventory problem

only temporarily until it is reversed by other traders.

Hypothesis 2. The impact of end-users’ demand depends on whether their demand will im-

prove the aggregated level of inventory held by dealers or move it away from an optimal level.

Stoll (1978) and O’Hara and Oldfield (1986) use inventory-control models to show how

risk-averse dealers charge premia to accommodate order imbalances. Since dealers use quotes

8

to elicit trades to balance their inventory positions, dealers’ inventory is expected to affect the

price response to order flow (Madhavan and Smidt, 1991, and Madhavan and Smidt, 1993).

These models predict that market depth is greater (lower price response) for order flow which

moves dealers’ inventory to a preferred position, and that order flow which increases dealers’

inventory exposure, in either direction, is associated with larger price impacts.

Hypothesis 3. Dealers’ inventories matter more when significant counterparty risk has built

up within the interdealer market.

Since there are no natural sellers of protection, dealers stand ready to absorb temporary

imbalances between demand and supply. The interdealer market then serves as a vital venue for

dealers to offload their inventory risk and share it with other dealers. In the process, new trades

are created until the credit risk finds a willing home. The data shows that an end-user trade is

passed between three dealers, on average. Each link in the chain involves a different exposure

to counterparty risk. As long as the contract is in effect, the counterparty risk remains. The

interdealer activity is concentrated among five major dealers, so such circular trading might

increase the overall counterparty exposure among them.

If the level of counterparty risk exposure amongst dealers increases, they become less able

or predisposed to absorb supply and demand shocks, and to hold buffer inventories of assets.

Under extreme conditions, dealers might even exit the market, which in turn might leave fewer

dealers to make most of the market. The inventory levels of these active dealers is expected to

build up. Then, the price impact of a trade must be greater for them to absorb the demand of

end-users.

Hypothesis 4. The inventory level of an individual dealer relative to its peers determines the

direction and size of their trades. The force of mean reversion in a dealer’s relative inventory

position should be increasing in the relative inventory level.

In the theoretical inventory models of dealer markets, the likelihood of execution depends

on the degree of competitiveness of the dealer’s quotes, which in turn depends on his relative

inventory position. When a dealer has an extreme inventory position, he is able to post com-

petitive quotes on one side and stands a better chance of executing the end-users’ order flow

in the desired direction. This results in a relatively quick reduction of his inventory imbalance.

On the other hand, when a dealer’s inventory is close to the median, he is not able to post

competitive prices and therefore stands a poor chance of executing the end-users order flow.

As a result, his inventory takes a longer time to revert to the desired level. This implies that

in competitive dealer markets, the relative inventories of the dealers should be mean reverting

and the strength of mean reversion should be increasing in the relative inventory level.

9

3 Data Description & Sample Construction

I was fortunate to be given access to an extensive CDS transactions data by DTCC, who

provides clearing, settlement and information services for OTC derivatives. In November 2006,

DTCC established its automated Trade Information Warehouse as the electronic central registry

for CDS contracts. Since that time, the vast majority of CDS contracts traded have been

registered in the Warehouse. In addition, all of the major global CDS dealers have registered

in the Warehouse many of the contracts that were executed among each other before that date.

The data covers all transacted CDS contracts of 35 financial firms between February 2007

and June 2009. Each transaction contains the following information: the name of the refer-

ence entity, the trade date and the effective date, the (expected) maturity of the contract,

anonymized identities of the participating counterparties including the type (dealer or end-

user), and the executed notional amount. The dealers, who generally provide liquidity and

trade on both sides, include the “G14 dealers”: Bank of America-Merrill Lynch, Barclays

Capital, BNP Paribas, Citibank, Credit Suisse, Deutsche Bank, Goldman Sachs, HSBC, J.P.

Morgan, Morgan Stanley, The Royal Bank of Scotland, Societe Generale, UBS, and Wachovia

Bank, as well as smaller dealers. The data also contain a finer classification of the end-users

to: asset managers, banks, financial services, hedge-funds, insurance firms, and “other” (the

last group accounts for less than 3% of the notional amount traded by end-users). Information

about the identities of parties trading through DTCC is privileged, and their agreements with

the participating institutions do not allow them to reveal it. Yet, each counterparty is identified

by a unique number over time, and over underlyings, which allows me to follow counterparties’

positions over-time and across names.

Most transactions datasets, even of equity markets, do not identify buyers and sellers,

forcing researchers to classify buyer-initiated or seller-initiated categories using some algorithm.

The Lee and Ready (1991) algorithm is often used and it classifies a trade as buyer-initiated

if it is closer to the ask of the prevailing quote and as seller-initiated if it is closer to the

bid. Since transaction prices are unavailable for the entire time-series, using this algorithm is

infeasible. Nevertheless, the DTCC dataset does include a classification of the buyer and the

seller. Hence, for the purpose of analyzing the price formation process, I assume that end-users

are the ones who initiate trades in this market, rather than dealers. Given that assumption,

I am able to infer the trade direction and to classify transactions as buyer-initiated or seller-

initiated. The signing of transactions allows me to borrow from the common techniques used

in the microstructure literature of the equity market. I believe that this assumption is not so

far from what happens in practice, as incoming trades are generally initiated by customers for

which the dealer will always be the supplier of liquidity. Note that transactions are submitted

post-execution and are not time-stamped for the actual “execution” time, so the price-impact

analysis can be performed only at the daily level.

Another important feature of the dataset is the specification of the type of the transaction,

10

which determines the traded notional amounts and the exposures. Specifically, a transaction

can be a new trade, an assignment of an already existing trade, or a termination of an existing

trade. When an investor wants to enter into a CDS contract he can either enter into a new

contract or find a counterparty that wishes to assign his position to him. When an investor

wants to unwind an existing contract, he faces three alternatives: enter into a new offsetting

transaction, assign the contract to a new counterparty, or terminate the transaction with

the original CDS counterparty. Each alternative carries different implications for the level of

counterparty risk exposure in the market. For further details on each of these three mechanisms

see Appendix A.

The transaction-level dataset is augmented with quotes of CDS prices from Markit Group,

a financial data provider, specialized in security and derivatives pricing. One of its services is

to gather, validate and distribute end-of-day composite CDS spreads provided by major deal-

ers. These reported prices are averaged for each CDS contract after eliminating outliers. The

maturities of Markit CDS contracts range between 6 months and 30 years. In the benchmark

analysis, I use the prices of the most popular and liquid 5-year CDS contracts on senior un-

secured obligations of North American reference entities with modified restructuring clauses.

I choose to use the modified restructuring since the default trading style in North America

moved from trading with restructuring to trading without restructuring only towards the end

of the sample in April 2009 along with the introduction of the Big Bang Protocol and “Stan-

dard” trading. CDS spreads of other contract maturities ranging between 1- and 10-year are

relatively liquid and are used for robustness checks. I observe that the vast majority of quotes

lie between 0 and 300 basis points. However, quotes occasionally exceed 3,000 bps and even

4,000 bps.4

After the initial filtration of the DTCC data and the merge with Markit and CRSP datasets,

some reference entities are rarely traded and would not provide reliable observations. Therefore,

to be included in the sample, I require that a CDS contract on a specific reference entity is

traded at least 10% of the trading days in its sample. I also exclude CDS contracts that

their underlyings are Real Estate Investment Trusts (REITs) as their trading characteristics

might differ from ordinary single-name contracts. The final sample consists of 35 individual,

financial-related reference entities financial firms. Table 1 lists the different firms. Of the

reference entities, 5 are classified as security and commodity brokers, dealers, and services, 7

are depository institutions, 5 are non-depository credit institutions, 23 are insurance related,

1 is real-estate related, and 1 is an asset management firm.

To investigate the role of supply and demand in the CDS market I assess the relation

4Such high spreads were observed in the sample especially during September 2008 and May 2009 in the follow-ing reference entities: Ambac Financial Group, American International Group Inc, Avis Budget Group, GenworthFinancial Inc, MBIA, Radian Group Inc. These levels are associated with a very high probability of default in theshort-term horizon. For example, if it were known with certainty that an entity would default in one year and thatthere would be no recovery, then the loss would be 100% on year from now and to cover this cost it would be necessaryto charge a CDS spread of about 10,000 bps per year.

11

between net buying pressure of end-users and dealers and the change in CDS spreads. I define

the change in inventory in a specific reference entity of each agent as the difference between

the notional amount bought and sold by him during the trading day, regardless of the contract

maturity,5 where the direction of the transaction is determined based on the assumption that

the end-user initiates the trade. Specifically, I construct the time series Qu,d,t, the net notional

amount bought by (Qu,d,t > 0), or sold to (Qu,d,t < 0), dealer d in reference entity u on day

t, and Iu,d,t = Iu,d,0 +∑t

s=1Qu,d,s, the inventory of dealer d in the CDS contract on reference

entity u accumulated by day t, where Iu,d,0 is the inventory at the start of the sample period.

In my sample, u = 1, 2, . . . , 35 and d = 1, 2, . . . , Du, where Du is the number of dealers in CDS

u, which varies between 4 and 17 dealers depending on the reference entity. I further construct

Qu,d2d,t =∑Du

d=1Qu,d,t, the daily aggregated net notional amount bought by, or sold to, dealers,

and Iu,t =∑Du

d=1

(Iu,d,0 +

∑ts=1Qu,d,s

), the daily aggregated inventory of the dealers as a whole.

Different contracts have different spreads, trading intensities, volume levels and volatility,

and, therefore, cannot be compared directly. To facilitate comparison across securities and

across dealers, I standardize the inventories series using Hansch et al. (1998) methodology for

equities. I scale the inventory changes both at the reference entity level, and at the dealer level.

More precisely, for each of the inventory series, Iu, I compute the sample time-series mean (Iu)

and the sample standard-deviation (σu), and define Iu,t, the standardized aggregated inventory

of dealers in CDS u at day t, as:

Iu,t :=Iu,t − Iuσu

=Iu,0 +

∑tτ=1Qu,τ −

1T

∑Tτ=1 (Iu,0 +

∑τr=1Qu,r)

σu(1)

=

∑tτ=1Qu,τ −

1T

∑Tτ=1

∑τr=1Qu,r

σu

This equation implies that, although the initial inventory level is unobserved in the data, the

changes in inventories and the standardized inventories that I use in the empirical tests are

independent of that initial level, Iu,0, although notice that Iu,0 is usually non-zero.

Similarly, I standardized the inventory series of individual dealers as follows:

Iu,d,t :=Iu,d,t − Iu,d

σu,d=Iu,d,0 +

∑tτ=1Qu,d,τ −

1T

∑Tτ=1 (Iu,d,0 +

∑τr=1Qu,d,r)

σu,d(2)

=

∑tτ=1Qu,d,τ −

1T

∑Tτ=1

∑τr=1Qu,d,r

σu,d

where Iu,d is the sample time-series mean of the inventory that dealer d holds in CDS u and σu,d

is the sample standard-deviation of that inventory time-series. This standardization captures

the notion that different dealers perceive the inventory risk in a similar way when their inventory

is measured in terms of the distance (in standard-deviations) of their standardized inventory

5This definition treats all maturities as essentially the same security, as price movements on different maturitiesare strongly correlated.

12

from the respective sample mean.

In order to be consistent with the standardization process, I also measure the daily net

positions in units of the sample standard-deviation of dealers’ inventory in that reference entity

(σu). I define Qu,eu,t as the net change in the inventory of end-users as a result of engaging in

Nu,eu,t trades on day t:

Qu,eu,t :=1

σu

Nu,eu,t∑n=1

qu,eu,t(n) (3)

where qu,eu,t(n) is the signed quantity of the nth trade, where qu,eu,t(n) < 0 if an end-user sold

to a dealer and qu,eu,t(n) > 0 if an end-user bought from a dealer.

Table 2 details the summary statistics for the key variables that are used in the empirical

analysis. Tables 3–5 provide summary statistics of different aspects of the trading activity of

the reference entities. They present the trading activity in each of the reference entities in the

sample in terms of days when there was a transaction and in terms of traded contracts, re-

spectively. These summary statistics compare the activity of end-user/dealer and dealer/dealer

activity.

Table 3 details the number of transactions each group executed over the sample period.

The table also breaks down the type of transaction. The table suggests that asset managers,

banks and hedge funds are the most active traders among end-users. Dealers engage in more

assignments than end-users. While end-users do use assignments to unwind positions, they

are particularly active as transferors. That is, they seldom take a new position through an

assignment. Investors, and especially hedge funds, tend to prefer to unwind a position through

an assignment rather than through an offset because they are reluctant to incur additional

counterparty exposure and potentially additional margin requirements on the offsetting swap.

They generally prefer assignment to termination because termination forces them to accept the

price proposed by the original counterparty, and can provide insights into trading strategies

to the counterparty. The table indicates that the share of assignments is roughly 25%, which

is consistent with the results of the March 2007 Report of the Committee on Payment and

Settlement Systems, a survey of dealers in the OTC derivatives market.

The last row in Table 3 breaks down the market share of each of the aforementioned class.

It might seem from this table that end-users trade on both sides of the market, but this is not

the case. The reported traded notional takes into account the round-trip of the investors, that

is, it counts the initial buy and then the sell. When the traded notional amounts are calculated

only for new trades, most end-user groups are more active as buyers of protection rather than

as sellers.

Table 4 suggests that there is some relation between the activity in the interdealer market

and the activity in the client-dealer market. On most trading days there is activity in the client-

13

dealer front as well as in the interdealer market. There is no trading activity only in a small

fraction of the business days, and the least traded contract trades on more than 70% of the

business days. Table 5 further supports the links between interdealer trading and client-dealer

trading. It breaks down the activity in terms of number of traded contracts.

Additional support for the importance of the interdealer market can be found in the con-

centration of the trading activity among the G14 dealers. Using the Herfindahl-Hirschman

Index, I find that the activity is moderately concentrated among these dealers. The five most

active participants made up almost 50% of the buys and about 50% of the sells for all trades

in terms of notional amount traded.

In contrast to the single-name contracts, the summary statistics for the CDX index indicate

that the trading mechanism for the CDX index, a basket product, is quite different. Equivalent

to Table 3, Table 6 details the summary statistics for the CDX index. It is interesting to note

that the majority of the trading in the CDX index is done through assignments rather than

through trades. Among the end-users, the most active participants in assignments are hedge

funds and asset managers. This observation is key in interpreting some of the empirical results.

Mainly, the common use of unwinding a position in the CDX index by assigning it to another

party relaxes some of the counterparty risks that raise from creating another offsetting new

trade, as common in the trading of the single name contracts.

4 The Ability of Dealers to Absorb End-Users’ De-

mand

As a first step to assess the role of liquidity provision by dealers in the CDS market, I explore

how price changes relate to net (pooled) order flow in the CDS market. Methodologically, my

analysis of CDS transactions data draws from the microstructure literature on equity and fixed-

income markets that considers the price impact in the presence of trading frictions – such as,

information asymmetry, funding constraints, search and inventory costs – and their interplay

leads to “price impact”, in which purchases (sales) of assets typically increase (depress) the

price. Given the price impact of end-users’ order imbalances, I then examine the ability of

dealers to absorb that demand, conditioning on their inventory level.

To estimate end-users’ impact on market depth (Hypothesis 1), I regress CDS returns

on the contemporaneous imbalance of end-users’ order-flow and lags of order imbalance. I

also account for the flow of information originating from the equity market by considering

the contemporaneous effect of stock returns on CDS returns (following Acharya and Johnson,

2007), and lagged log-returns of the CDS contracts. I use these controls because order flow

might be correlated with fundamental information. Collin-Dufresne et al. (2001) and Schaefer

and Strebulaev (2008) show that the simple Merton (1974) model seems to be adequate to

capture all fundamental information, as far as default risk goes. Therefore, after controlling

14

for equity returns, the unexplained component of the CDS returns might be specific to the

microstructure of the CDS market. Specifically, I use the following specification:

∆ logSu,t = α+

K∑k=0

γk

∣∣∣Qu,eu,t−k∣∣∣ 1>0 +

K∑k=0

δk

∣∣∣Qu,eu,t−k∣∣∣ 1<0

+5∑

n=0

ζn∆ logPu,t−n +5∑

n=1

ηn∆ logSu,t−n + εu,t (4)

To account for the asymmetries of buying and selling pressure, I use∣∣∣Qu,eu,t−k∣∣∣ 1>0 :=

max{Qu,eu,t−k, 0

}and

∣∣∣Qu,eu,t−k∣∣∣ 1<0 := max{−Qu,eu,t−k, 0

}. Su,t denotes the 5-year CDS

quoted spread for reference entity u at the end of day t; and, Pu,t denotes the equity price of

the same firm at the end of day t. If no price changes are recorded during the trading day, price

change is designated as zero. Also, since outliers might coincide with high illiquidity, I run the

regression for log-difference: ∆ logSu,t := logSu,t− logSu,t−1, ∆ logPu,t := logPu,t− logPu,t−1,

∆Su,t := Su,t − Su,t−1, and ∆Pu,t := Pu,t − Pu,t−1, as it is better in terms of not having

undue influence of outliers. This is the baseline structure of the regression specifications. Since

I perform a pooled regressions (all reference entities and all trading days), I must account

for correlation across underlyings and across time. To address this, I compare the White

standard errors and clustered standard errors (when I cluster by time, by firm, and by both

firm and time), and I find weak firm- and time-effects in the panel data. Hence, I conduct

the empirical analysis by adjusting for heteroscedasticity and controlling for the time effect by

adding daily time dummies. Because of the presence of time dummies, I do not include any

other macroeconomic variables in the regression analysis.

Table 7 reports the results of the pooled price-impact regressions that are based on the

standardized traded notionals. I find that when end-users buy more contracts, CDS returns

tend to increase, and when they sell more contracts the returns diminish. The price reaction to

buying versus selling by end-users seems to be asymmetric. The contemporaneous effect of net

buying by end-users is significant, while the net selling effect is insignificant in the majority of

the specifications. This is not surprising, considering the composition of end-users in the CDS

market, who mostly buy CDS contracts to hedge their core businesses. To address the issue

that a termination of a CDS contract might not be as informative as a new CDS contract in

which an end-user is the seller, I re-run this regression with a restricted sample with only new

trades. I get the same qualitative asymmetry between buys and sells.

The regression results also suggest that the trend in the price impact is not temporary, not

reversing even five trading days later. This might lend support to an information-based story,

where end-users as a whole are more informed than dealers, while the last take the other side

of the trade as part of their role to provide liquidity in the market.

I have checked for robustness using a number of different variations of Equation (4). The

15

results are not significantly affected by: (i) the inclusion or exclusion of lagged CDS returns; (ii)

the use of unscaled imbalance measures; (iii) the use of the number of transactions rather than

the dollar volume; and, (iv) use of new trades only. The asymmetry response to net-buying

and net selling could have been a by-product of the type of the transaction. In other words, it

is possible that taking the selling side in termination carries different implication in terms of

information than taking the selling side in a new trade. Therefore, I re-estimated the results

for new trades alone and the results are substantively similar to the results presented in Table

7. These robustness checks are discussed in detail in Section 7.

Hypothesis 2 suggests that the effect of end-users’ order-imbalances depends on the dealers’

inventory level. To test this hypothesis, I estimate the following regression:

∆ logSu,t = α +∣∣∣Qu,eu,t∣∣∣ 1>0

(β1 + δ11|Iu,t−1+Qu,eu,t|↗

)+∣∣∣Qu,eu,t∣∣∣ 1<0

(β2 + δ21|Iu,t−1+Qu,eu,t|↗

)+ γ1|Iu,t−1+Qu,eu,t|↗ +

5∑n=0

ζn∆ logPu,t−n +

5∑n=1

ηn∆ logSu,t−n + εu,t (5)

where the dummy variable captures whether the demand of end-users during day t increases

dealers’ t− 1 absolute aggregate inventory of underlying u. Specifically,

1|Iu,t−1+Qu,eu,t|↗ :=

1∣∣∣Iu,t−1∣∣∣ > ∣∣∣Iu,t∣∣∣

0 otherwise(6)

All microstructure models predict that the coefficient on order flow will be positive. The

inventory management models suggest that the coefficient on the dummied order flow will also

be positive. Intuition for the prediction can be developed by considering the case of a dealer

with long inventory. End-user sales (negative order flow) to the dealer lead to inference that

“true” price is lower, as in Kyle (1985), and this is captured in the term βQu,eu,t. However,

selling to a dealer who is long also increases the dealer’s inventory exposure, which according

to the theory, further lowers the price at which the market is willing to make a purchase.

The results of this regression are presented in Table 8. I find that aggregated inventory

level of dealers indeed amplifies the price impact of end-users if the incoming order flow further

increases the absolute value of the inventory. The inventory effect seems to be symmetric for

net-buying and net-selling, although the definition of the dummy variable might conceal any

asymmetry. Furthermore, a more dynamic definition of the indicator variable for the increase

of dealers’ aggregated inventory takes into account the speed of the mean reversion of dealers’

inventories (as I later show in Section 6). Using the mean reversion coefficients, one can

calculate the expected inventory level, and then determine whether there was a greater than

expected increase in the aggregated level of dealers’ inventory.

Column (4) of Table 8 details the regression results for the CDX index. I present these

results to contrast the importance of end-users’ information and dealers’ inventory in single-

16

name contracts versus a basket product such as the CDS index. The results suggest that

end-users’ net-selling is associated with an increase in the CDX index price, whereas end-users’

net-buying is associated with a decrease in price. The lags of the order-imbalance is statistically

insignificant, so the two results combined imply that the microstructure of the trading in the

CDX index is very different from the single-name contracts. Information does not seem to play

a role, and the results are consistent with the intuition that end-users use the CDX index for

hedging purposes.

Columns (5) and (6) show the results of the regression if we condition on the period before

and after the crisis (taking the break point as Lehman’s failure in September 15, 2008). It

emerges from the table that the inventory effect is statistically significant when end-users were

net-buyers before the crisis and when end-users were net-sellers during the crisis.

The results in this section are consistent with the standard inventory models of dealer

markets. Continuing to build upon this line of research, one would expect interdealer trading

to occur when the inventories of the dealers diverge a lot and when the dealers choose between

the uncertain probability of an arrival of an end-user trade and the certainty of interdealer

trading (Ho and Stoll, 1983). Thus dealers have a greater willingness to trade in the interdealer

market when they have divergent inventories. In other words, the interdealer market facilitates

the management of inventory risks and allows dealers to take large inventory position that they

would be unwilling to take in market-types where they can only unwind their inventory positions

against the end-users order flow. Yet, due to counterparty risk, “redundant” interdealer trading,

in terms of transferring credit risk, might restrict the level of liquidity provision by dealers.

This implication of counterparty risk on trading in the CDS market is explored in the next

section.

5 The Effect of Interdealer Exposure on Dealers’

Liquidity Provision

The objective of a CDS dealer, along with making a two-way market, is to earn profits by

managing the risk of its inventory. When a dealer takes on risk from a client, the dealer

typically hedges the risk but might choose to leave some of the risk uncovered. The willingness

of dealers to leave some risks uncovered – that is, to speculate – adds liquidity to the market

but requires close management. Typically, however, dealers remain “flat” by hedging their

risks in some manner. Most simply, a dealer might offset the risk of a new deal against that of

other clients. Further, the dealer might hedge the risk of a deal in the underlying market, that

is, the cash bond market. Finally, a dealer might choose to offset risks by means of offsetting

transactions with other dealers; this is a primary function of the interdealer market.

In this section I focus on the interdealer market and explore whether it facilitates or impedes

liquidity provision. Reiss and Werner (1998) and Bjønnes and Rime (2005) show that in the

17

equity and FX market, respectively, interdealer markets are mainly used by dealers to share

inventory risks. Yet, in the CDS market along with the market risks of their positions and the

accompanying basis risks, dealers also have to manage the counterparty risk associated with

their positions. With each trade in the interdealer market, a new exposure between a pair of

dealers is being created in addition to the credit exposure to the underlying. Since I observe

that an end-user’s trade may pass between three dealers on average, an interesting mechanism

arises. As dealers pass the end-user’s trade, the credit exposure to the underlying is unchanged,

but the counterparty exposure between dealers is tripled. Although there are various ways to

manage counterparty risk, such as collateralization of net exposures, the passing-on of inventory

among dealers creates a cascade of new positions that entail new risks. That is, dealers’ desire

to hedge counterparty risk effectively creates increased exposure among dealers.

The figure below demonstrates the potential implications of repeated passes among dealers.

First, consider Case 1 where an end-user A buys credit protection from dealer 1, which hedges

its exposures with dealer 2, while dealer 2 passes on the risk to monoline insurer, end-user

B. In this case three individual contracts have been written, yet, in economic terms, only

one party at the very end of the risk transfer chain bears the risk that the reference entity

defaults. At the same time, aggregate gross notional value has been inflated to three times

the aggregate net exposure. By contrast, consider Case 2, where there are two contracts, but

neither of the protection selling parties passes on its risk to another company. Under these

circumstances, net notional amount equals gross notional amount. Both dealer 2 and end-user

B bear the economic risk that the reference entity defaults. Case 1 illustrates the importance

of the exposure measure. Even though dealers might not bear any credit risk associated with

the “open” positions, they still have to hedge the counterparty risk that is associated with the

existing contractual ties, until the unwinding of the trades.

Case 1

Case 2

Protection BuyerEnd-User A

Protection SellerDealer 1

Protection BuyerDealer 1

Protection SellerDealer 2

Protection BuyerDealer 2

Protection SellerEnd-User B

Protection BuyerEnd-User A

Protection SellerDealer 1

Protection BuyerDealer 2

Protection SellerEnd-User B

18

If the exposure level to counterparty risk amongst dealers increases, they become less able

or predisposed to absorb supply and demand shocks, and to hold buffer inventories of assets.

Under extreme conditions, dealers might even exit the market, which in turn might leave fewer

dealers to make most of the market. The inventory levels of these active dealers is expected to

build up. Then, the price impact of a trade must be greater for them to absorb the demand of

end-users.

To test this hypothesis, I construct a measure of exposure in the interdealer market. First,

I calculate the daily bilateral exposure of each pair of dealers at the reference entity level,

and I denote it by Ei,j,u,t. Then, these exposure quantities can be netted within and across

underlyings, and can also be netted across dealers. In reality, however, other than compres-

sions that net across dealers, day-to-day operations include only netting of bilateral positions

(multilateral positions cannot be observed in the OTC market). To keep the interpretation

of the results simple, I chose to net the exposures of dealers only at the pair level, without

netting across other reference entities in the sample. Note that for each dealer that buys a CDS

contract there is a dealer that sells a CDS contract. Hence, to calculate the overall exposure

in the interdealer market, I sum the (long) exposures over all dealers.

As reported in Section 3, the trading activity is moderately concentrated among the G14

dealers. The measure of exposure, however, captures another angle of concentration, namely,

the concentration of counterparty risk that remains in the system. Figure 3(a) depicts the

network of exposures among the top five dealers over the entire sample period, whereas Figure

3(b) represents the changes that occurred after Lehman Brothers’ collapse. The two figures

suggest that the total average exposure among the top five dealers increased in the period after

Lehman, although the trading activity remained at the same level.

Using this measure for the exposure levels in the interdealer market, I test Hypothesis 3.

Specifically, I run the following regression:

∆ logSu,t = α+ Qu,eu,t

(β1 + β21|Iu,t−1+Qu,eu,t|↗ + β3Eu,d2d,t + β41|Iu,t−1+Qu,eu,t|↗ × Eu,d2d,t

)

+β51|Iu,t−1+Qu,eu,t|↗ + β6Eu,d2d,t + β71|Iu,t−1+Qu,eu,t|↗ × Eu,d2d,t + εu,t (7)

where Eu,d2d,t is the dealers’ long exposure to other dealers.

The regression results are detailed in Table 9. The results in Panel A of Table 9 show

that dealer-to-dealer exposure intensifies the effect that dealers’ inventory has on the price

impact of end-users. Panel B of Table 9 refines the picture by separating the price impacts

of net-buying and net-selling of end-users. It is evident that the “triple” interaction term

Qu,eu,t × 1|Iu,t−1+Qu,eu,t|↗ × Dealer-to-Dealer Exposure is positive for net buying demand and

negative for net selling pressure, while the magnitude for net selling is almost twice the impact

when clients want to buy. Also, the results suggest that the exposure effect is prominent in

19

extreme market scenarios.

Figures 4–7 summarize the economic magnitude of each of the effects. I decompose the total

price impact of trading into the key components found in this paper for: (a) a benchmark case

that is based on mean values of equity and CDS returns, dealers’ inventory level and end-users’

net buying imbalance, and for (b) an extreme case when the realization of all variables is at

their maximum. Using the mean level of the variables, the “basic” price impact accounts for

28% and 7% of the total price impact for buys and sells, respectively, and an incoming order-

flow that will increase dealers’ inventory results in 16% and 23% of the price impact for buys

and sells, respectively. While the congestion effect plays a minor role in normal times, it does

account for as much as 23% and 37% of the total price impact, for buys and sells, respectively,

if all independent variables are taken at their maximum values. In either normal times or crisis

times, the combined microstructure effect of end-users’ order imbalance, dealers’ inventory

and interdealer exposure congestion (including the interactions between the three) is about

the same magnitude as the impact of information flow from the stock market and lagged CDS

spreads. These microstructure effects are not “informationless” and they actually generate new

fundamental information over and above the information generated in the equity markets. I also

find that both end-users’ order imbalances in the CDS market and the information innovations

from the CDS market predict stock returns 1-day ahead.

6 How Do Dealers Manage Their Inventory?

The analysis in previous sections assumes the existence of a preferred inventory position for

dealers. In this section, I test Ho and Stoll (1983) model pertaining to mean reversion in

dealers’ relative inventories, using the methodology in Hansch et al. (1998). Similar to the

results found by Hansch et al. (1998) for the London Stock Exchange dealers, I also find

evidence that dealers with extreme inventory positions trade larger quantities that move their

inventories toward desired levels, and that the intensity of mean reversion increases with the

deviation of the inventory from its desired level.

To test the relation between a dealer’s relative inventory position and the force of mean

reversion I define Imedianu,t to be the median inventory level of reference entity u at the end of

day t across dealers, and RIu,d,t = Iu,d,t−Imedianu,t to be dealer d’s inventory position in reference

entity u relative to the median inventory at time t. Then, I estimate the following piecewise

linear regression:

∆RIu,d,t = α+

5∑n=1

βnDnRIu,d,t−1 + εu,d,t (8)

20

where

Dn =

1 if (n− 1) ≤ |RIu,d,t−1| < n and n ∈ {1, 2, 3, 4}

1 if (n− 1) ≤ |RIu,d,t−1| and n = 5

0 otherwise

The Dn are indicator variables that allow for differences in the degree of mean reversion as a

function of different bands of relative inventory levels. For example, β1 captures the intensity of

mean reversion when the relative inventory level lies between zero and one standard-deviation,

β2 captures the intensity of mean reversion when the relative inventory level is greater than or

equal to one but less than two standard-deviations, and so on. I compute the inventory level

dependent mean reversion coefficients (βn’s) from this piecewise linear regression for all dealers

in all underlyings.

The findings in Table 10 show that the variation in the degree of mean reversion is strongly

related to the level of inventory. In particular, the speed of mean reversion increases as the

inventory divergence increases, evidence consistent with the predictions of the inventory models

of dealer markets. If I constrain all β’s to be equal to each other, I find a mean reversion

coefficient of −0.0064, which implies an inventory half-life of 108.66 days.6 This result is

obviously considerably different from the inventory-level dependent half-lives that range from

11.15 days starting from 5σ to 159.85 days starting from 1σ (see Table 10). I see that the speed

of mean reversion reduces as the inventory divergence increases. I repeat the same analysis

using standardized inventories and find that they also exhibit considerable level-dependent

variation.

A more refined picture of the mechanism of dealers for controlling their inventory is obtained

from the analysis of whether dealers actively manage their inventories by increasing trading

among themselves as their inventories diverge, or do they passively wait for a random incoming

end-users’ flow. For this purpose I decompose the change in the inventory of a dealer into two

components, one arising out of his trading with other dealers and the other arising out of his

trading with end-users. Similarly to the definition of QIDu,t and QEU

u,t in Section 3, I define QIDu,d,t

and QEUu,d,t as the net change in the inventory of dealer d as a result of engaging in N ID

u,d,t and

NEUu,d,t interdealer and end-users trades on day t, respectively:

QIDu,d,t =

1

σu,d

N IDu,d,t∑n=1

qIDu,d,t(n) and QEUu,d,t =

1

σu,d

NEUu,d,t∑n=1

qEUu,d,t(n) (9)

where qu,d,t(n) is the size of the nth trade and σu,d is the standard-deviation of dealer d’s

inventory in that CDS contract.

6The implied half-life of inventories is calculated as − ln 2ln(1+β) , which follows from the solution of It ≡ It−1 + ∆It =

(1 + β) It−1 = (1 + β)t × I0.

21

I express the change in the inventory of dealer d in day t, Qu,d,t, as the sum of the change

due to his trading with end-users and the change due to his trading with other dealers; that

is, Qu,d,t = QIDu,d,t + QEU

u,d,t. In order to investigate whether the relative proportion of end-user

and interdealer trading depends on the level of inventory, I run the following two regressions:

QIDu,d,t = αID +

5∑n=1

γIDn DnIu,d,t−1 + εIDu,d,t (10)

QEUu,d,t = αEU +

5∑n=1

γEUn DnIu,d,t−1 + εEUu,d,t (11)

where

Dn =

1 if (n− 1) ≤

∣∣∣Iu,d,t−1∣∣∣ < n and n ∈ {1, 2, 3, 4}

1 if (n− 1) ≤∣∣∣Iu,d,t−1∣∣∣ and n = 5

0 otherwise

and where d denotes the dealer, and Iu,d,t represents the standardized inventory of dealer d at

the end of day t.

Equations (10) and (11) divide βn’s, the mean reversion coefficients for inventories obtained

in Table 10, into two parts: γIDn captures the mean reversion in inventories due to interdealer

trading, and γEUn captures the mean reversion in inventories due to trading with end-users.

Thus, γIDn + γEUn equals βn. This decomposition enables me to evaluate the extent to which

interdealer trading contributes to mean reversion in inventories across different inventory levels.

Table 11 summarizes the findings. I observe that a large proportion in the mean reversion

in inventories occurs due to interdealer trading. The more the inventory level diverges from

the mean, the greater the importance of the interdealer trading.

7 Robustness of Results

The reporting problems of the DTCC dataset coupled with the trading features of the CDS

market pose several problems to the empirical methodology used. For example, trade frequency

in the CDS market is relatively low, so there are many days with zero order-flow for some

reference entities. In the price impact tests I use quoted prices from Markit rather than

transaction prices. This combination results in a change of price that is not associated with

end-users’ order flow. Such a scenario can indeed happen due to information percolation from

other asset markets, such as, the equity, bond and/or option markets. Therefore, I not only

control for information flow from the equity market, which is considered the most liquid, but

I also check the relation between the quoted and transaction prices. In a robustness check, I

22

look at new trades first for which I can estimate the transaction price.7 I find that end-of-day

quoted prices are closely related to the intra-day average of transaction prices, and I choose to

use the quotes as they cover a longer time-series. Also, I repeat the regressions where I consider

days with zero order-flow as missing observations. The empirical results are also qualitatively

robust to the use of arithmetic differences rather than log-differences and to the differentiation

of contracts by maturity rather than pooling them together and considering them as a 5-year

maturity contract.

The rest of this section details the different specifications and their results.

7.1 Different Sources for CDS Prices

In my main empirical analysis I use the composite prices provided by Markit Group. Markit

collects CDS prices from about 23 dealers to produce its end-of-day prices. The quotes are

subject to filtering that removes outliers and stale observations. Markit then computes a daily

composite spread only if it has two or more contributors. I opt to use Markit as the data source

for the main analysis since it has a longer time series and is one of the most widely employed

in the literature related to CDS.

Other possible sources for CDS prices include the Credit Market Analytics DataVision

(CMA) and Fitch. CMA gets prices by parsing e-mails from dealers to 36 investment firms,

aggregating data where possible to produce continuous prices. It relies on its relationships with

buyside investors, including major global investment banks, hedge funds, and asset managers,

and any prices it gets from dealers come only from second-tier market makers for comparison

purposes. Fitch uses prices from 20 dealers, who are tier-one sellside market makers.

Additional robustness check that I undertake uses the transaction prices from the DTCC

database (see Figure 9). I am able to use only the prices that are associated with new trades

without upfront payment. Due to reporting standards I am unable to interpret the prices

that are associated with assignments, while new trades with an upfront payment require some

conversion to basis points that its accuracy depends on the assumptions made. As part of

the Big Bang Protocol, commenced on April 8, 2009, North American corporate names are

traded with a fixed coupon. The coupon is either 100 bps or 500 bps and upfront payments

are exchanged. Therefore, to be able to compare all transaction prices to the quoted prices, I

need to calculate the equivalent (full) spread to the sum of the fixed coupon and the upfront

payment for these observations.

To convert the fixed coupon and the upfront payment to a CDS price, I replicate the ISDA

CDS pricing model that is being used as the standard conversion model. For the conversion, I

7Many of the trades were traded for a fixed spread and an upfront payment even before the Big Bang Protocol inApril 2009. To be able to compare prices across different transactions, I need to convert the upfront payment, whichis usually in USD terms, to basis-points. The conversion involves some assumptions about the recovery rate and therisky term-structure that could have been different from trade to trade. Thus, using the same set of assumptions forall trades will not necessarily deliver the true transaction price. Section 7.1 fleshes out the conversion details.

23

use the following formula:

K∑k=1

(1−R)× 1

2[Z(0, tk−1) + Z(0, tk)]× [Q(0, tk−1)−Q(0, tk)]

= Upfront Payment + S01

2

K∑k=1

∆(tk−1, tk)Z(0, tk) [Q(0, tk−1) +Q(0, tk)]︸ ︷︷ ︸Risky PV01

(12)

where S0 is a fixed payment (%, annualized) that the buyer of the protection pays for the

duration of the protection period, or up to a default event; Q(t, T ) is the time t survival

probability of the reference entity to time T ; Z(t, T ) is the Libor discount curve which is

anchored to the CDS effective date; R is the expected recovery rate as a percentage of par;

and, ∆(tk−1, tk) is the day count fraction between dates tk−1 and tk in the appropriate day

count conversion, typically Actual 360.

I assume a constant recovery as a fraction of par (40%), a piecewise constant risk neutral

hazard rates, and independent default events of changes in the default free yield curve. The raw

transaction prices and the converted prices are presented in Figures 9–10, where the conversion

was applied for AIG contracts as an example.

7.2 The Distribution of CDS Contracts’ Maturity

In the construction of the time-series for volume, inventory and exposure variables, I aggregate

the position regardless of the maturity of the contract. This leads to a maturity mismatch

problem. This aggregation at the maturity level might therefore mask the true ramifications

on dealers’ portfolios. The alternatives I face include:

1. Constrain the sample to a specific maturity, say CDS contracts with 5 year maturity, as

it is the most liquid

2. Run separate regressions for different maturities

3. Pool together different contracts with different maturities

On the one hand, there is a difference in counterparty risk if one holds a 5-year contract vs 1-

year contract. On the other hand, a dealer might buy a 5-year contract with the expectation to

hold it only for one year because the 5-year contract is more liquid. Using only contracts with

5-year maturity (most liquid) leaves some extra information in other maturities unexploited.

Since the majority of the traded transactions are for 5-year maturity (see Figure 8), I choose

to use contracts with all maturities in my benchmark analysis and to run robustness check in

the form of different regressions for different maturities, using the CDS return for this specific

maturity.

24

7.3 Using Arithmetic Differences

In all the results presented thus far, I measure CDS returns by logarithmic differences in

CDS spreads (log CDStCDSt−1

), i.e., the percentage changes in CDS spreads. An alternative way of

measuring CDS return is to calculate the arithmetic difference in CDS spreads (CDSt−CDSt−1).

While both these methods are legitimate, the arithmetic difference in CDS spreads has a certain

intuitive appeal for measuring CDS return because it approximately equals the return generated

from buying an underlying bond of the CDS entity at t− 1 and selling it at t. Table 12 reports

the re-estimation of Equation (4) using CDS return from arithmetic differences. The empirical

results remain unchanged. The re-estimation of the other regressions using arithmetic difference

delivers qualitatively similar results, not reported here.

8 Concluding Remarks

In this paper I examine the intermediation capacity of dealers in the CDS market by exploiting

a proprietary dataset of transactions in 35 single-name reference entities. I find three important

results. First, the traditional inventory-based market microstructure models can be used in the

setting of the seemingly more complex and opaque market, such as the CDS market. Order

imbalances of end-users appear to impact CDS price changes, and dealers’ aggregated inventory

position amplifies this effect. Second, the force of mean reversion of dealers’ inventories to the

target level is stronger as the deviation from the target increases, and in turn, the price impact of

end-users’ order flow is greater at such points in time. Finally, counterparty risk among dealers

affects liquidity in the CDS market, especially in stress conditions. It would be interesting in

future research to examine how this effect is being translated in other derivative markets, such

as FX derivatives, interest rate swaps, and OTC equity derivatives, where counterparty risk is

also present.

Further, my empirical analysis provides a factual basis on the trading activity and the

provision of liquidity in the CDS market about which little was known until this point. These

results might lend some guidance to the effect of the anticipated migration of trading from an

OTC setting to one intermediated by central clearing counterparty (CCP). The CCP would

stand between the two original counterparties, acting as the seller to the original buyer, and as

the buyer to the original seller. This will allow a CCP to net the positions across investors, and

should mitigate the key impediment of counterparty risk to liquidity provision by dealers raised

in this paper. It seems that the establishment of a CCP can lead to a reduction in risk and a

substantial improvement in allocational efficiency, as long as it will be able to avoid becoming

systemically risky itself (see Acharya, Shachar, and Subrahmanyam, 2010 and Duffie and Zhu,

2009 for further details).

Especially in the midst of the transition of the CDS market from an OTC structure to

central clearing, CDS transactions data, and order imbalance data in particular, open avenues

25

of research beyond those that are explored in this paper. Ever since Lehman’s collapse, the

DTCC has published weekly data on individual reference entities, gross and net volumes of

CDSs outstanding. This data is cruder, but it allows forming a broader picture of the CDS

market activity and to address issues at the cross-section. Nevertheless, the dataset that I

exploit in this paper is much more detailed and includes a unique identifier for the identity of

investors, which is consistent across firms and over time. This information will allow me to

examine the role of bargaining power and repeated interactions in the terms of trade in future

papers.

26

References

Acharya, V. V. and T. C. Johnson (2007). Insider Trading in Credit Derivatives. Journal of

Financial Economics 84 (1), 110 – 141.

Acharya, V. V., O. Shachar, and M. Subrahmanyam (2010). Regulating Wall Street: The Dodd-

Frank Act and the New Architecture of Global Finance. Viral V. Acharya, Thomas Cooley,

Matt Richardson, Ingo Walter, co-editors, John Wiley and Sons.

Amihud, Y. and H. Mendelson (1980). Dealership Market: Market-Making with Inventory.

Journal of Financial Economics 8 (1), 31 – 53.

Arora, N., P. Gandhi, and F. A. Longstaff (2010, January). Counterparty Credit Risk and the

Credit Default Swap Market. UCLA Working Paper .

Bai, J. and P. Collin-Dufresne (2010, November). The Determinants of the CDS-Bond Basis

During the Financial Crisis of 2007-2009.

Berndt, A. and A. Ostrovnaya (2007). Information Flow between Credit Default Swap, Option

and Equity Markets. working paper .

Bjønnes, G. H. and D. Rime (2005). Dealer Behavior and Trading Systems in Foreign Exchange

Markets. Journal of Financial Economics 75 (3), 571 – 605.

Bollen, N. P. B. and R. E. Whaley (2004). Does Net Buying Pressure Affect the Shape of

Implied Volatility Functions? The Journal of Finance 59 (2), 711–753.

Brunnermeier, M. K. and L. H. Pedersen (2009). Market Liquidity and Funding Liquidity.

Review of Financial Studies 22 (6), 2201–2238.

Cetin, U., R. Jarrow, P. Protter, and M. Warachka (2006). Pricing Options in an Extended

Black Scholes Economy with Illiquidity: Theory and Empirical Evidence. Review of Financial

Studies 19 (2), 493–529.

Chen, K., M. Fleming, J. Jackson, A. Li, and A. Sarkar (2011). An Anlysis of CDS Transactions:

Implication for Public Reporting. Working Paper, FRBNY .

Choi, J. and O. Shachar (2011). Back to Basis: When Liquidity Providers Become Liquidity

Seekers. working paper .

Collin-Dufresne, P., R. S. Goldstein, and J. S. Martin (2001). The Determinants of Credit

Spread Changes. The Journal of Finance 56 (6), 2177–2207.

Duffie, D. (2010). Presidential Address: Asset Price Dynamics with Slow-Moving Capital. The

Journal of Finance 65 (4), 1237–1267.

27

Duffie, D., N. Garleanu, and L. H. Pedersen (2005). Over-the-Counter Markets. Economet-

rica 73 (6), 1815–1847.

Duffie, D., N. Garleanu, and L. H. Pedersen (2007). Valuation in Over-the-Counter Markets.

Review of Financial Studies 20 (6), 1865–1900.

Duffie, D. and H. Zhu (2009). Does a Central Clearing Counterparty Reduce Counterparty

Risk? Working Paper, Stanford University .

Easley, D. and M. O’Hara (1987). Price, trade size, and information in securities markets.

Journal of Financial Economics 19 (1), 69 – 90.

Fontana, A. (2011, September). The Negative CDS-Bond Basis and Convergence Trading

during the 2007/09 Financial Crisis.

Garleanu, N., L. H. Pedersen, and A. M. Poteshman (2009). Demand-Based Option Pricing.

The Review of Financial Studies 22 (10), 4259–4299.

Garman, M. B. (1976). Market Microstructure. Journal of Financial Economics 3 (3), 257 –

275.

Glosten, L. R. and P. R. Milgrom (1985). Bid, Ask and Transaction Prices in a Specialist

Market with Heterogeneously Informed Traders. Journal of Financial Economics 14 (1), 71

– 100.

Gromb, D. and D. Vayanos (2002). Equilibrium and Welfare in Markets with Financially

Constrained Arbitrageurs. Journal of Financial Economics 66 (2-3), 361 – 407.

Grossman, S. J. and M. H. Miller (1988). Liquidity and Market Structure. The Journal of

Finance 43 (3), pp. 617–633.

Hansch, O., N. Y. Naik, and S. Viswanathan (1998). Do Inventories Matter in Dealership

Markets? Evidence from the London Stock Exchange. The Journal of Finance 53 (5), 1623–

1656.

Hart, O. and L. Zingales (2009). A New Capital Regulation For Large Financial Institutions.

working paper .

Hasbrouck, J. and G. Sofianos (1993). The Trades of Market Makers: An Empirical Analysis

of NYSE Specialists. The Journal of Finance 48 (5), pp. 1565–1593.

Ho, T. S. Y. and H. R. Stoll (1983). The Dynamics of Dealer Markets Under Competition.

The Journal of Finance 38 (4), pp. 1053–1074.

Kyle, A. S. (1985). Continuous Auctions and Insider Trading. Econometrica 53 (6), pp. 1315–

1335.

28

Lagos, R. and G. Rocheteau (2009). Liquidity in Asset Markets With Search Frictions. Econo-

metrica 2, 403–426.

Lagos, R., G. Rocheteau, and P.-O. Weill (2011). Crises and Liquidity in Over-The-Counter

Markets. Journal of Economic Theory .

Lee, C. M. C. and M. J. Ready (1991). Inferring Trade Direction from Intraday Data. The

Journal of Finance 46 (2), pp. 733–746.

Lyons, R. K. (1997). A Simultaneous Trade Model of the Foreign Exchange Hot Potato. Journal

of International Economics 42 (3-4), 275 – 298.

Madhavan, A. and S. Smidt (1991). A Bayesian Model of Intraday Specialist Pricing. Journal

of Financial Economics 30 (1), 99 – 134.

Madhavan, A. and S. Smidt (1993). An Analysis of Changes in Specialist Inventories and

Quotations. The Journal of Finance 48 (5), pp. 1595–1628.

Merton, R. C. (1974). On the Pricing of Corporate Debt: The Risk Structure of Interest Rates.

The Journal of Finance 29 (2), pp. 449–470.

Mitchell, M., L. H. Pedersen, and T. Pulvino (2007). Slow Moving Capital. NBER working

papers.

Oehmke, M. and A. Zawadowski (2012). The Anatomy of the CDS Market. Working Paper,

Columbia University and Boston University .

O’Hara, M. and G. S. Oldfield (1986). The Microeconomics of Market Making. The Journal

of Financial and Quantitative Analysis 21 (4), pp. 361–376.

Reiss, P. C. and I. M. Werner (1998). Does risk sharing motivate interdealer trading? The

Journal of Finance 53 (5), 1657–1703.

Saar, G. (2001). Price Impact Asymmetry of Block Trades: An Institutional Trading Expla-

nation. Review of Financial Studies 14 (4), 1153–1181.

Schaefer, S. M. and I. A. Strebulaev (2008). Structural models of credit risk are useful: Evidence

from hedge ratios on corporate bonds. Journal of Financial Economics 90 (1), 1 – 19.

Shleifer, A. and R. Vishny (1997). The Limits of Arbitrage. Journal of Finance 52 (1), 28–55.

Stoll, H. R. (1978). The Supply of Dealer Services in Securities Markets. The Journal of

Finance 33 (4), pp. 1133–1151.

29

Table 1The Reference Entities in The Sample

This table details the 35 reference entities that compose the sample. All 35 reference entities are North American andare associated with the financial industry. The specific sub-industry of each firm is also detailed in the table below.

Reference Entity Industrical Classfication

AMBAC Financial Group Financial ServiceAmerican Express Company Personal Credit InstitutionsAmerican International Group Fire, Marine & Casualty InsuranceAon Corp Accident & Health InsuranceAvis Budget Group Real Estate Agents & Managers (For Others)Bank Of America Corp National Commercial BanksBerkshire Hathaway Fire, Marine & Casualty InsuranceCapital One Financial Corp Personal Credit InstitutionsCIT Group Miscellaneous Business Credit InstitutionCitigroup National Commercial BanksCNA Financial Corp Fire, Marine & Casualty InsuranceCredit Suisse State Commercial BanksFairfax Financial Holdings Limited Fire, Marine & Casualty InsuranceGenworth Financial Life InsuranceGoldman Sachs Group Security Brokers, Dealers & Flotation CompaniesHartford Financial Services Group Insurance Agents, Brokers & ServiceHSBC Finance Corp Commercial Banks, NECJPMorgan Chase & Co National Commercial BanksLehman Brothers Holdings Investment AdviceLincoln National Corp Life InsuranceLoews Corp Fire, Marine & Casualty InsuranceMarsh & Mclennan Companies Surety InsuranceMBIA Surety InsuranceMerrill Lynch & Co Security Brokers, Dealers & Flotation CompaniesMetlife Insurance Agents, Brokers & ServiceMorgan Stanley Security Brokers, Dealers & Flotation CompaniesPMI Group Surety InsurancePrudential Financial Life InsuranceRadian Group Surety InsuranceSLM Corp Federal & Federally Sponsored Credit AgenciesThe Chubb Corp Fire, Marine & Casualty InsuranceThe Travelers Companies Fire, Marine & Casualty InsuranceUNUM Group Accident & Health InsuranceWells Fargo & Company National Commercial BanksWestern Union Company Functions Related To Depository Banking, NEC

30

Tab

le2

Su

mm

ary

Sta

tist

ics

Th

isT

able

det

ails

the

sum

mary

stat

isti

csof

the

key

vari

ab

les

inth

esa

mp

le:

end

-use

rs’

net

-bu

yin

g,

end

-use

rs’

net

-sel

lin

g,

an

dd

eale

rs’

inve

nto

ry.

Ials

ore

por

tth

esu

mm

ary

stat

isti

csfo

rth

esc

aled

vari

ab

les

as

det

ail

edin

Sec

tion

3.

Note

that

the

un

itof

ob

serv

ati

on

isen

tity

-day

.P

an

elA

incl

ud

esall

35re

fere

nce

enti

ties

,w

her

eas

Pan

elB

excl

ud

es9

refe

ren

ceen

titi

esth

at

are

als

od

eale

rth

emse

lves

.T

he

9d

eale

rsin

clud

e7

dea

lers

wh

oare

curr

entl

yco

nsi

der

edas

G14

dea

lers

,as

wel

las

Leh

man

Bro

ther

san

dM

erri

llL

yn

ch.

Late

r,I

use

the

two

sam

ple

sto

show

that

the

corr

elati

on

bet

wee

nth

eu

nd

erly

ings

and

the

sup

pli

ers

ofli

qu

idit

yd

oes

not

aff

ect

any

of

the

resu

lts.

Pan

el

A:

Enti

reS

am

ple

En

d-U

sers

’E

nd

-Use

rs’

Dea

lers

’In

vento

ryS

cale

dE

nd

-Use

rs’

Sca

led

En

d-U

sers

’S

cale

dD

eale

rs’

Net

Bu

yin

gN

etSel

lin

gN

etB

uyin

gN

etS

elli

ng

Inve

nto

ry

N18

,102

18,1

02

18,1

02

18,1

02

18,1

02

18,1

02

Mea

n9,

238,

767

9,76

1,1

44

-740,6

37,7

35

0.0

30.0

30.0

0M

edia

n0

0-4

55,0

95,8

81

0.0

00.0

0-0

.07

Std

Dev

24,9

29,1

8337

,025,3

06

1,2

56,5

86,8

27

0.0

70.0

71.0

0M

inim

um

00

-5,0

89,9

08,4

34

0.0

00.0

0-2

.90

Max

imu

m85

5,45

0,00

02,

254,

051,4

16

3,4

01,5

15,2

06

1.5

42.3

44.5

9

Exc

lude

Zer

oO

bser

vati

on

sN

6,87

77,4

12

18,1

02

6,8

77

7,4

12

18,1

02

Mea

n24

,318

,768

23,8

39,2

10

-740,6

37,7

35

0.0

70.0

60.0

0M

edia

n12

,699

,075

11,0

00,0

00

-455,0

95,8

81

0.0

40.0

4-0

.07

Std

Dev

35,6

26,0

8154

,887,4

73

1,2

56,5

86,8

27

0.1

00.1

01.0

0M

inim

um

011,2

08

-5,0

89,9

08,4

34

0.0

00.0

0-2

.90

Max

imu

m85

5,45

0,00

02,

254,

051,4

16

3,4

01,5

15,2

06

1.5

42.3

44.5

9

Pan

el

B:

Exclu

din

gR

efe

ren

ce

Enti

ties

that

are

Deale

rF

irm

s

En

d-U

sers

’E

nd

-Use

rs’

Dea

lers

’In

vento

ryS

cale

dE

nd

-Use

rs’

Sca

led

En

d-U

sers

’S

cale

dD

eale

rs’

Net

Bu

yin

gN

etSel

lin

gN

etB

uyin

gN

etS

elli

ng

Inve

nto

ry

N13

,396

13,3

96

13,3

96

13,3

96

13,3

96

13,3

96

Mea

n6,

447,

799

6,55

8,4

25

-326,8

40,6

03

0.0

30.0

30.0

0M

edia

n0

0-3

35,0

24,0

20

0.0

00.0

0-0

.02

Std

Dev

17,1

18,9

1316

,559,4

99

857,8

23,2

62

0.0

70.0

71.0

0M

inim

um

00

-2,8

66,3

93,0

59

0.0

00.0

0-2

.90

Max

imu

m58

2,17

1,00

031

3,07

7,5

46

3,4

01,5

15,2

06

1.5

41.2

84.5

9

Exc

lude

Zer

oO

bser

vati

on

sN

4,77

85,2

49

13,3

96

4,7

78

5,2

49

13,3

96

Mea

n18

,077

,587

16,7

37,7

91

-326,8

40,6

03

0.0

80.0

70.0

0M

edia

n10

,000

,000

10,0

00,0

00

-335,0

24,0

20

0.0

40.0

4-0

.02

Std

Dev

24,7

27,8

7423

,010,8

50

857,8

23,2

62

0.1

10.0

91.0

0M

inim

um

011,2

08

-2,8

66,3

93,0

59

0.0

00.0

0-2

.90

Max

imu

m58

2,17

1,00

031

3,07

7,5

46

3,4

01,5

15,2

06

1.5

41.2

84.5

9

31

Tab

le3

Bre

akd

ow

nof

Nu

mb

er

of

Tra

nsa

cti

on

by

Invest

ors

Gro

up

an

dT

ran

sacti

on

Typ

e:

Th

e35

Sin

gle

Nam

eC

ontr

acts

Th

ista

ble

bre

aks

dow

nth

etr

ansa

ctio

ns

for

each

gro

up

of

inve

stors

an

dtr

an

sact

ion

typ

e.P

an

elA

pre

sents

nu

mb

erof

tran

sact

ion

san

dP

an

elB

pre

sents

the

frac

tion

.

Pan

el

A

Ass

etB

anks

Fin

an

cial

Hed

ge

Insu

ran

ceP

ensi

on

Oth

erT

ota

lD

eale

rsM

anag

ers

Ser

vic

esF

un

ds

Fir

ms

Fu

nd

sE

nd

-Use

rsE

nd

-Use

rs

New

Tra

de

As

ab

uye

r13

,548

8,27

31,0

29

22,3

11

1,0

36

169

873

47,2

39

133,4

32

As

ase

ller

10,9

706,

283

1,4

07

9,7

77

473

57

738

29,7

05

150,9

66

Ass

ign

men

tS

tep-I

nA

sa

bu

yer

586

432

134

1,1

86

50

22,3

45

52,0

63

As

ase

ller

251

326

143

404

40

81,1

36

53,2

72

Ass

ign

men

tS

tep-O

ut

As

ab

uye

r4,

124

1,71

3326

5,2

63

258

14

300

11,9

98

35,2

96

As

ase

ller

6,90

32,

457

2,1

59

341

308

56

12,2

24

35,0

74

Matu

red

As

ab

uye

r1,

428

285

72

205

37

15

29

2,0

71

7,7

71

As

ase

ller

8725

824

271

63

30

706

9,1

36

Pan

el

B

Ass

etB

anks

Fin

an

cial

Hed

ge

Insu

ran

ceP

ensi

on

Oth

erT

ota

lD

eale

rsM

anag

ers

Ser

vic

esF

un

ds

Fir

ms

Fu

nd

sE

nd

-Use

rsE

nd

-Use

rs

New

Tra

de

As

ab

uye

r55

%57

%42

%70

%69

%75

%54

%61

%47

%A

sa

sell

er45

%43

%58

%30

%31

%25

%46

%39

%53

%A

ssig

nm

ent

Ste

p-I

nA

sa

bu

yer

70%

57%

48

%75

%56

%20

%67

%49

%A

sa

sell

er30

%43

%52

%25

%44

%80

%33

%51

%A

ssig

nm

ent

Ste

p-O

ut

As

ab

uye

r37

%41

%13

%94

%46

%100

%84

%50

%50

%A

sa

sell

er63

%59

%87

%6

%54

%16

%50

%50

%M

atu

red

As

ab

uye

r94

%52

%75

%43

%37

%83

%100

%75

%46

%A

sa

sell

er6

%48

%25

%57

%63

%17

%0

%25

%54

%

Tot

alM

arke

tS

har

e(o

ut

ofth

eto

tal

not

ion

al

bou

ght

an

dso

ld)

As

ab

uye

r21

%21

%2

%49

%3

%11%

3%

19

%81

%A

sa

sell

er24

%19

%3

%49

%2

%1%

3%

19

%81

%

32

Tab

le4

Non

-Zero

Tra

din

gD

ays

For

Each

Refe

ren

ce

Enti

ty

Th

ista

ble

conta

ins

ad

etai

led

des

crip

tion

ofth

etr

ad

ing

act

ivit

yin

term

sof

trad

ing

day

sw

hen

ther

ew

as

at

least

1tr

ade

(i.e

.p

osi

tive

volu

me)

for

each

ofth

ere

fere

nce

enti

ties

inth

esa

mp

le,

and

pre

sents

the

rela

tion

bet

wee

nin

terd

eale

ract

ivit

yan

dd

eale

r-cl

ient

act

ivit

y.If

ther

eis

act

ivit

yam

on

gd

eale

rsin

asp

ecifi

cd

ayth

enQD2D6=

0.If

end

-use

rse

lls

toa

dea

ler

(QE2D

)or

ad

eale

rse

lls

toan

end

-use

r(Q

D2E

),th

enQE2D

+QD2E6=

0,

oth

erw

ise

QE2D

+QD2E

=0.

Not

eth

atth

esu

mQE2D

+QD2E

den

ote

sth

egro

ssn

oti

on

al

trad

ed,

not

the

net

.

Fir

m#

of

#of

Day

sw

ith

QD2D6=

0an

dQD2D6=

0an

dQD2D

=0

an

dT

rad

ing

Day

sN

oA

ctiv

ity

QE2D

+QD2E6=

0QE2D

+QD2E

=0

QE2D

+QD2E6=

0

AM

BA

CF

inan

cial

Gro

up

574

1466

89

18

Am

eric

anE

xp

ress

Com

pan

y585

1484

75

25

Am

eric

anIn

tern

atio

nal

Gro

up

587

2522

52

11

Aon

Cor

p446

9240

154

43

Avis

Bu

dge

tG

rou

p542

2362

151

27

Ban

kO

fA

mer

ica

Cor

p599

1553

36

9B

erksh

ire

Hat

haw

ay553

2372

156

23

Cap

ital

On

eF

inanci

alC

orp

583

3457

99

24

CIT

Gro

up

592

0568

22

2C

itig

rou

p594

0538

43

13

CN

AF

inan

cial

Cor

p440

11

208

185

36

Cre

dit

Su

isse

282

496

109

73

Fai

rfax

Fin

anci

alH

old

ings

Lim

ited

394

25

157

181

31

Gen

wor

thF

inan

cial

524

2311

169

42

Gol

dm

anS

achs

Gro

up

595

1585

81

Har

tfor

dF

inan

cial

Ser

vic

esG

rou

p564

4433

91

36

HS

BC

Fin

ance

Cor

p571

5464

90

12

JP

Mor

gan

Ch

ase

&C

o599

1557

30

11

Leh

man

Bro

ther

sH

old

ings

404

0399

41

Lin

coln

Nat

ion

alC

orp

370

6154

150

60

Loew

sC

orp

527

10

343

129

45

Mar

sh&

Mcl

enn

anC

omp

anie

s560

4416

111

29

MB

IA590

1526

53

10

Mer

rill

Lyn

ch&

Co

475

1468

42

Met

life

558

5421

104

28

Mor

gan

Sta

nle

y599

0584

10

5P

MI

Gro

up

588

0544

36

8P

rud

enti

alF

inan

cial

499

6313

134

46

Rad

ian

Gro

up

590

0532

48

10

SL

MC

orp

560

0529

23

8T

he

Chu

bb

Cor

p519

11

335

131

42

Th

eT

rave

lers

Com

pan

ies

360

14

169

137

40

UN

UM

Gro

up

411

14

216

149

32

Wel

lsF

argo

&C

omp

any

587

2541

36

8W

este

rnU

nio

nC

omp

any

281

1132

119

29

33

Tab

le5

Nu

mb

er

of

Tra

ded

Contr

acts

For

Each

Refe

ren

ce

Enti

ty

Th

ista

ble

conta

ins

ad

etai

led

des

crip

tion

ofth

ed

ail

ytr

ad

ing

act

ivit

yin

term

sof

nu

mb

erof

CD

Sco

ntr

act

sfo

rea

chof

the

refe

ren

ceen

titi

esin

the

sam

ple

.T

he

nu

mb

erof

contr

acts

sold

toen

d-u

sers

rep

rese

nts

the

nu

mb

erof

tran

sact

ion

sin

wh

ich

end

-use

rsw

ere

the

bu

yers

of

pro

tect

ion

;th

enu

mb

erof

contr

acts

bou

ght

by

end

-use

rsre

pre

sents

the

num

ber

of

tran

sact

ion

sin

whic

hen

d-u

sers

wer

eth

ese

ller

of

pro

tect

ion

.In

calc

ula

tin

gth

enu

mb

erof

tran

sact

ion

s,te

rmin

atio

ns

wer

eco

nsi

der

edas

sell

s.T

her

efore

,it

mig

ht

seem

that

the

diff

eren

ceb

etw

een

end

-use

rs’

bu

yin

gan

dse

llin

gis

small

erth

an

wh

enon

lyn

ewtr

ades

are

con

sid

ered

.

Fir

m#

Con

trac

tsB

ou

ght

by

En

dU

sers

#C

ontr

act

sS

old

toE

nd

Use

rs#

Contr

act

sT

rad

edA

mon

gD

eale

rsM

ean

Med

ian

S.D

.T

ota

lM

ean

Med

ian

S.D

.T

ota

lM

ean

Med

ian

S.D

.T

ota

l

AM

BA

CF

inan

cial

Gro

up

53

62,1

69

63

62,2

58

11

713

6,1

22

Am

eric

anE

xp

ress

Com

pan

y4

24

1,4

17

42

51,6

69

86

10

4,5

90

Am

eric

anIn

tern

atio

nal

Gro

up

63

82,7

49

74

10

3,2

13

14

918

8,1

39

Aon

Cor

p3

16

491

21

3441

43

71,7

37

Avis

Bu

dge

tG

rou

p3

24

805

32

3816

74

12

3,5

95

Ban

kO

fA

mer

ica

Cor

p8

410

3,9

47

85

10

4,0

45

17

10

21

9,7

19

Ber

ksh

ire

Hat

haw

ay2

13

616

32

4956

53

82,5

48

Cap

ital

On

eF

inan

cial

Cor

p4

36

1,5

93

43

61,6

86

96

10

4,9

22

CIT

Gro

up

64

52,7

60

64

62,9

70

18

14

17

10,4

28

Cit

igro

up

95

11

4,2

83

95

11

4,6

66

16

11

18

9,3

53

CN

AF

inan

cial

Cor

p2

13

392

21

3417

42

91,5

79

Cre

dit

Su

isse

21

2188

21

2204

21

2432

Fai

rfax

Fin

anci

alH

old

ings

Lim

ited

21

2245

21

2238

32

41,1

17

Gen

wor

thF

inan

cial

32

4641

32

3765

53

52,2

22

Gol

dm

anS

achs

Gro

up

106

14

5,8

07

11

715

6,1

80

21

15

22

12,2

24

Har

tfor

dF

inan

cial

Ser

vic

esG

rou

p4

26

1,4

45

42

51,5

06

74

10

3,8

28

HS

BC

Fin

ance

Cor

p4

35

1,4

03

42

51,5

69

85

94,1

70

JP

Mor

gan

Ch

ase

&C

o7

411

3,3

66

74

10

3,6

21

14

917

7,9

66

Leh

man

Bro

ther

sH

old

ings

149

18

5,6

43

17

925

6,6

47

28

19

36

11,1

36

Lin

coln

Nat

ion

alC

orp

31

3312

31

4370

32

41,0

32

Loew

sC

orp

22

3557

21

3597

43

51,9

17

Mar

sh&

Mcl

enn

anC

omp

anie

s3

25

1,1

80

32

41,0

40

74

93,4

53

MB

IA7

48

3,2

98

74

83,2

86

14

916

8,3

01

Mer

rill

Lyn

ch&

Co

138

13

5,6

18

14

915

6,3

47

28

20

49

13,3

45

Met

life

32

51,0

04

42

51,2

26

75

10

3,7

07

Mor

gan

Sta

nle

y11

521

5,9

30

11

622

6,4

05

21

14

24

12,2

87

PM

IG

rou

p6

46

2,9

03

64

63,0

33

14

10

16

8,1

92

Pru

den

tial

Fin

anci

al3

24

777

32

4932

43

61,9

28

Rad

ian

Gro

up

54

72,5

91

64

72,7

87

16

11

18

9,2

44

SL

MC

orp

85

13

3,6

85

10

521

4,9

74

18

11

28

9,7

23

Th

eC

hu

bb

Cor

p3

24

716

32

5800

43

71,9

80

Th

eT

rave

lers

Com

pan

ies

32

4309

21

3355

32

71,0

27

UN

UM

Gro

up

21

4370

21

4400

42

61,3

40

Wel

lsF

argo

&C

omp

any

64

62,8

20

74

83,2

56

13

10

14

7,6

41

Wes

tern

Un

ion

Com

pan

y2

12

197

21

2214

32

3842

All

Fir

ms

63

10

72,2

26

63

11

79,8

87

11

618

191,7

85

34

Tab

le6

Bre

akd

ow

nof

Nu

mb

er

of

Tra

nsa

cti

on

by

Invest

ors

Gro

up

an

dT

ransa

cti

on

Typ

e:

Th

eC

DX

Ind

ex

Th

ista

ble

bre

aks

dow

nth

etr

ansa

ctio

ns

for

each

grou

pofin

vest

ors

an

dtr

an

sact

ion

typ

e.A

ssig

nm

ent

Ste

p-I

nis

the

new

leg

ofth

etr

ad

e,w

hil

eA

ssig

nm

ent

Ste

p-O

ut

isth

eex

itin

gle

gof

the

trad

e.P

anel

Ap

rese

nts

the

nu

mb

erof

tran

sact

ion

san

dP

an

elB

pre

sents

the

fract

ion

.

Pan

el

A

Ass

etB

anks

Fin

an

cial

Hed

ge

Insu

ran

ceP

ensi

on

Oth

erT

ota

lD

eale

rsM

anag

ers

Ser

vic

esF

un

ds

Fir

ms

Fu

nd

sE

nd

-Use

rsE

nd

-Use

rs

New

Tra

de

As

ab

uye

r17

634

9161

62

41

447

1,2

25

As

ase

ller

190

43

17

114

27

10

1402

1,2

70

Ass

ign

ment

Ste

p-I

nA

sa

bu

yer

7410

04

638

13

0820

64,7

10

As

ase

ller

285

115

2698

10

415

1,1

29

64,4

07

Ass

ign

ment

Ste

p-O

ut

As

ab

uye

r8,

729

1,62

7338

8,9

36

1,6

11

126

107

21,4

74

36,7

76

As

ase

ller

9,61

11,

683

250

9,7

05

1,7

02

128

78

23,1

57

35,0

86

Matu

red

As

ab

uye

r0

10

40

00

5555

As

ase

ller

23

14

01

011

549

Pan

el

B

Ass

etB

anks

Fin

an

cial

Hed

ge

Insu

ran

ceP

ensi

on

Oth

erT

ota

lD

eale

rsM

anag

ers

Ser

vic

esF

un

ds

Fir

ms

Fu

nd

sE

nd

-Use

rsE

nd

-Use

rs

New

Tra

de

As

ab

uye

r48

%44

%35

%59

%70

%29

%50

%53

%49

%A

sa

sell

er52

%56

%65

%41

%30

%71

%50

%47

%51

%A

ssig

nm

ent

Ste

p-I

nA

sa

bu

yer

21%

47%

67

%48

%9

%42

%50

%A

sa

sell

er79

%53

%33

%52

%91

%100

%58

%50

%A

ssig

nm

ent

Ste

p-O

ut

As

ab

uye

r48

%49

%57

%48

%49

%50

%58

%48

%51

%A

sa

sell

er52

%51

%43

%52

%51

%42

%52

%49

%M

atu

red

As

ab

uye

r25

%50

%31

%50

%A

sa

sell

er10

0%

75%

100

%50

%100

%69

%50

%

35

Tab

le7

Th

eE

ffect

of

En

d-U

sers

’D

aily

Tra

din

gA

cti

vit

yon

CD

SS

pre

ad

s

Inth

ista

ble

Ist

ud

yth

eas

ym

met

ric

imp

act

ofen

d-u

sers

’ord

erim

bala

nce

son

log-r

etu

rns

of

CD

Ssp

read

s.T

he

tab

led

etail

sth

eco

effici

ent

of

the

foll

owin

gp

ool

ed-r

egre

ssio

n:

∆lo

gSu,t

=α

+

K ∑ k=0

γk

∣ ∣ ∣Q u,eu,t−k

∣ ∣ ∣1 >0+

K ∑ k=0

δ k

∣ ∣ ∣Q u,eu,t−k

∣ ∣ ∣1 <0+

5 ∑ n=0

ζ n∆

logPu,t−n

+

5 ∑ n=1

η n∆

logSu,t−n

+ε u,t

wh

ere∣ ∣ ∣Q u,e

u,t−k

∣ ∣ ∣×1>0

isth

est

and

ard

ized

net

dail

yn

oti

on

al

am

ou

nt

bou

ght

by

end

-use

rs.

Sim

ilarl

y,∣ ∣ ∣Q u,e

u,t−k

∣ ∣ ∣×1<0

isth

est

an

dard

ized

net

dail

y

not

ion

alam

ount

sold

toen

d-u

sers

.B

usi

nes

sd

ays

wit

hou

tany

trad

ing

act

ivit

yare

incl

ud

edin

the

sam

ple

wit

hQu,eu,t

as

zero

,b

efore

the

stand

ard

izati

on

.S

tan

dar

der

rors

that

app

ear

inp

aren

thes

esar

eh

eter

osc

edast

ic-r

ob

ust

stan

dard

erro

rs.∗∗∗,∗∗

,an

d∗

den

ote

ssi

gn

ifica

nce

at

1%

,5%

,an

d10%

leve

l,re

spec

tive

ly.

All

regr

essi

ons

also

incl

ud

ean

inte

rcep

t.

Dep

end

ent

Var

iab

le:

All

Ref

eren

ceE

nti

ties

Ref

.E

nt.

that

are

Excl

ud

ing

Ref

.E

nt.

that

are

Dea

lers

Th

emse

lves

Dea

lers

Th

emse

lves

∆lo

gSu,t

K=

0K

=1

K=

3K

=5

K=

5K

=5

γ0×

103

31.5∗∗∗

29.5∗∗∗

29.

9∗∗∗

29.7∗∗∗

15.

0∗∗

31.0∗∗∗

(5.2

)(5.4

)(5.4

)(5.5

)(6.2

)(6.3

)∑ K k

=1γk×

103

14.8∗∗∗

20.

3∗∗∗

23.2∗∗∗

35.

6∗∗∗

16.8∗∗∗

(3.3

)(4.9

)(5.6

)(1

1.9)

(5.9

)δ 0×

103

−3.

6−

3.5

−2.

9−

2.3

−10.1

−1.

8(4.6

)(4.6

)(4.7

)(4.7

)(1

0.4)

(4.3

)∑ K k

=1δ k×

103

−10.1∗∗∗

−22.9∗∗∗

−30.

5∗∗∗

−36.6∗∗

−23.9∗∗∗

(3.8

)(5.7

)(6.7

)(1

4.3)

(7.1

)ζ n×

103

−13

9.9∗∗∗

−13

9.8∗∗∗

−139.5∗∗∗

−139.2∗∗∗

−64.2

−161.6∗∗∗

(34.

7)(3

4.7

)(3

4.6

)(3

4.5

)(7

6.3)

(30.6

)∑ 5 n

=1ζ n×

103

−70.2∗

−68.

1−

67.0

−66.

744.

0−

48.7

(41.

9)(4

1.7

)(4

1.7

)(4

1.7

)(1

08.

4)

(38.6

)∑ 5 n

=1ζ n×

103

44.9∗

46.3∗

45.

2∗41.8

−135.4∗∗

108.1∗∗∗

(25.

4)(2

5.3

)(2

5.5

)(2

5.7

)(5

3.6)

(29.2

)

Coeffi

cien

tes

tim

ates

OL

SO

LS

OL

SO

LS

OL

SO

LS

Sta

nd

ard

erro

rsW

hit

eW

hit

eW

hit

eW

hit

eW

hit

eW

hit

eF

ixed

-Eff

ects

Tim

e(D

ay)

Tim

e(D

ay)

Tim

e(D

ay)

Tim

e(D

ay)

Tim

e(D

ay)

Tim

e(D

ay)

Ad

just

edR

-squ

ared

0.43

0.4

30.4

30.4

30.

72

0.42

Nu

mb

erof

Ob

serv

atio

n18,1

5018,1

17

18,0

52

17,9

89

4,5

90

13,3

99

36

Tab

le8

Deale

rs’

Abil

ity

toA

bso

rbE

nd

-Use

rs’

Dem

an

d

Th

ista

ble

show

sth

ere

sult

sof

the

foll

owin

gfo

ur

regre

ssio

ns:

∆lo

gSu,t

=α

+∣ ∣ ∣Q u,e

u,t

∣ ∣ ∣1 >0(β

1+δ 1×

Du

mm

y)

+∣ ∣ ∣Q u,e

u,t

∣ ∣ ∣1 <0(β

2+δ 2×

Du

mm

y)

+γ×

Du

mm

y+

Contr

ols

+ε u,t

wh

ere

the

du

mm

yva

riab

leis

defi

ned

as:

1|I u

,t−

1+Q

u,e

u,t|↗

:=

{ 1if∣ ∣ ∣I u,t∣ ∣ ∣>∣ ∣ ∣

I u,t−1

∣ ∣ ∣0

oth

erw

ise

and

wh

ereI d,t−1

isth

eag

greg

ated

inve

nto

ryof

all

dea

lers

at

tim

et−

1sc

ale

dby

the

stan

dard

-dev

iati

on

of

dea

lers

’in

vento

ryin

refe

ren

ceen

tityu

,σu,

and

1 |Iu,t

−1+Q

u,e

u,t|↗

isan

ind

icat

orva

riab

leth

ateq

uals

1w

hen

end

-use

rs’

net

bu

yin

g/se

llin

gat

tim

et

incr

ease

sd

eale

rs’t−

1aggre

gate

inven

tory

,an

d

0ot

her

wis

e.A

llre

gres

sion

sal

soin

clu

de

anin

terc

ept,

five

lags

of

the

dep

end

ent

vari

ab

le,

the

conte

mp

ora

neo

us

retu

rnof

the

stock

ass

oci

ate

dw

ith

the

CD

Sco

ntr

act’

sre

fere

nce

enti

ty,

and

five

lags

ofth

est

ock

retu

rn.

Sta

ndard

erro

rsare

inp

are

nth

eses

.∗∗∗,∗∗

,an

d∗

den

ote

ssi

gn

ifica

nce

at

1%

,5%

,an

d10

%le

vel,

resp

ecti

vely

.

Dep

end

ent

Var

iab

le:

(1)

(2)

(3)

(4)

(5)

(6)

∆lo

gSu,t

All

Ref

.E

nt.

that

Exc.

Ref

.E

nt.

that

are

CD

XB

efore

Aft

erR

ef.

Ent

Dea

lers

Th

emse

lves

Dea

lers

Th

emse

lves

Ind

exL

ehm

an

Leh

man

(“C

risi

s”)

β1×

103

25.3∗∗∗

20.9∗∗∗

24.

9∗∗∗

−61.6∗∗∗

26.

9∗∗

23.5∗∗∗

(5.4

)(8.0

)(6.1

)(1

0.4

)(1

3.0)

(5.5

)δ 1×

103

17.4

−4.0

20.5

19.

446.

2∗8.

0(1

0.7)

(11.

8)

(12.6

)(2

4.0

)(2

7.8)

(7.8

)β2×

103

6.8

14.

73.7

21.7∗∗

15.

03.

0(5.7

)(1

1.0)

(6.1

)(1

0.7

)(1

2.6)

(6.2

)δ 2×

103

−24.4∗∗

−57.

3∗∗

−15.9∗

5.8

−33.6

−18.

6∗∗

(10.

4)(2

7.7)

(8.4

)(1

9.2

)(2

1.2)

(9.1

)γ×

103

−0.

50.

7−

0.4

−1.

1−

1.2

−0.2

(0.6

)(0.9

)(0.6

)(1.9

)(1.1

)(0.5

)ζ n×

103

−13

8.8∗∗∗

−63.

7−

162.0∗∗∗

−101.1∗∗∗

−270.3∗∗∗

(34.

7)(7

4.9)

(30.9

)(3

6.1

)(2

7.8

)∑ 5 n

=1ζ n×

103

−67.5

44.

7−

46.7

−54.4

−125.3∗∗

(42.

1)(1

08.

7)

(39.1

)(5

1.1

)(5

0.0

)∑ 5 n

=1ζ n×

103

38.7

−148.3∗∗∗

105.1∗∗∗

164.6

8.4

60.2∗

(25.

8)(5

2.8)

(29.4

)(1

07.6

)(4

2.9)

(30.9

)

Coeffi

cien

tes

tim

ates

OL

SO

LS

OL

SO

LS

OL

SO

LS

Sta

nd

ard

erro

rsW

hit

eW

hit

eW

hit

eW

hit

eW

hit

eW

hit

eF

ixed

-Eff

ects

Tim

e(D

ay)

Tim

e(D

ay)

Tim

e(D

ay)

Non

eT

ime

(Day

)T

ime

(Day

)

Ad

just

edR

-squ

ared

0.43

0.72

0.4

20.

14

0.40

0.47

Nu

mb

erof

Ob

serv

atio

n17,6

644,6

24

13,0

40

439

5,546

12,1

18

37

Table 9The Relation between Inventory and Interdealer Exposures

Panel A of this table details the results of the following regression:

∆ logSu,t = α+ Qu,eu,t

(β1 + β21|Iu,t−1+Qu,eu,t|↗ + β3Eu,d2d,t + β41|Iu,t−1+Qu,eu,t|↗ × Eu,d2d,t

)+β51|Iu,t−1+Qu,eu,t|↗ + β6Eu,d2d,t + β71|Iu,t−1+Qu,eu,t|↗ × Eu,d2d,t + εu,t

where Eu,d2d,t is the dealers’ long exposure to other dealers, and 1|Iu,t−1+Qu,eu,t|↗ equals 1 if∣∣∣Iu,t∣∣∣ > ∣∣∣Iu,t−1∣∣∣, and 0

otherwise. Panel B of this regression refines the regression above and separate the price impacts of net-buying andthe net-selling by end-users.

Panel A

Dependent Variable: 10, 000∆ logSu,t

Qu,eu,t 157.27∗∗

(69.16)

Qu,eu,t × 1|Iu,t−1+Qu,eu,t|↗ 508.18∗∗∗

(99.94)

Qu,eu,t ×Dealer-to-Dealer Exposure 27.19(83.96)

Qu,eu,t × 1|Iu,t−1+Qu,eu,t|↗ ×Dealer-to-Dealer Exposure 353.61∗∗

(165.70)1|Iu,t−1+Qu,eu,t|↗ ×Dealer-to-Dealer Exposure 5.47

(17.73)1|Iu,t−1+Qu,eu,t|↗ −10.75

(10.03)Dealer-to-Dealer Exposure 26.22∗∗

(10.45)

Panel B

Dependent Variable: 10, 000∆ logSi,t

Qu,eu,t × 1>0 402.22∗∗∗

(113.62)

Qu,eu,t × 1>0 × 1|Iu,t−1+Qu,eu,t|↗ 459.74∗∗∗

(159.69)

Qu,eu,t × 1>0 ×Dealer-to-Dealer Exposure 233.09∗

(119.93)

Qu,eu,t × 1>0 × 1|Iu,t−1+Qu,eu,t|↗ ×Dealer-to-Dealer Exposure 212.52(215.37)

Qu,eu,t × 1<0 80.86(96.00)

Qu,eu,t × 1<0 × 1|Iu,t−1+Qu,eu,t|↗ −539.57∗∗∗

(146.64)

Qu,eu,t × 1<0 ×Dealer-to-Dealer Exposure 228.34∗

(138.66)

Qu,eu,t × 1<0 × 1|Iu,t−1+Qu,eu,t|↗ ×Dealer-to-Dealer Exposure −497.24∗

(262.85)

38

Table 10Test of Mean Reversion in Dealers’ Inventories

The table shows the average mean reversion coefficients of the following piecewise linear regression:

(RIu,d,t −RIu,d,t−1) = α+

5∑n=1

βnDnRIu,d,t−1 + εu,d,t

where d denotes the dealer, and RIu,d,t is the inventory of dealer d relative to the median dealer inventory at the endof day t. The Dn are indicators variables that allow for differences in the degree of mean reversion as a function ofdifferent bands of relative inventory levels, and are defined as follows:

Dn =

1 if (n− 1) ≤ |RIu,d,t−1| < n and n ∈ {1, 2, 3, 4}1 if (n− 1) ≤ |RIu,d,t−1| and n = 5

0 otherwise

I run this regression for each dealer in each reference entity. I also calculate implied half-lives of inventories. Theregression is repeated for the standardized inventories.

Dependent Variable: 10, 000×∆ logSu,t(1) (2)

Dummy: 1|Id,t−1+Qeu,t|↗ 1exceedsE(increase)∣∣∣Qu,eu,t∣∣∣× 1>0 459.6∗∗∗ 572.9∗∗∗

(81.9) (93.6)∣∣∣Qu,eu,t∣∣∣× 1<0 8.6 177.3

(90.6) (1979.3)Dummy −4.9 −20.5∗∗∗

(10.9) (7.7)∣∣∣Qu,eu,t∣∣∣× 1>0 ×Dummy 376.5∗ 3310.5∗∗

(203.2) (1332.9)∣∣∣Qu,eu,t∣∣∣× 1<0 ×Dummy −357.1∗∗ −254.5

(156.1) (1980.8)

Coefficient estimates OLS OLSStandard errors White WhiteFixed-Effects Time (d) Time (d)

R-squared 0.32 0.32Number of Observation 25, 406 25, 367

39

Table 11Decomposition of the Mean Reversion Coefficients

The table shows the average coefficient of the following piecewise linear regressions:

QIDu,d,t = αID +

5∑n=1

γIDn DnIu,d,t−1 + εIDu,d,t

QEUu,d,t = αEU +

5∑n=1

γEUn DnIu,d,t−1 + εEU

u,d,t

where d denotes the dealer, Qu,d,t is the net amount traded and the superscript ID refers to the interdealer trading,

and the superscript EU refers to dealers’ trading directly with end-users, and Iu,d,t is the standardized inventory ofdealer d at the end of day t. The Dn are indicator variables that allow for differences in the degree of mean reversionas a function of different bands of inventory levels, and are defined as follows:

Dn =

1 if (n− 1) ≤

∣∣∣Iu,d,t−1∣∣∣ < n and n ∈ 1, 2, 3, 4

1 if (n− 1) ≤∣∣∣Iu,d,t−1∣∣∣ and n = 5

0 otherwise

I run this regression for each dealer in each reference entity. The sum of the end-users and interdealer coefficients asshown in the Total column is exactly equal to the corresponding mean reversion coefficients of standardized inventoriesin Table 10.

Interdealer Trading Trading with End-Users Total Interdealer Trading Fraction

γ1 −0.004 −0.001 −0.005 87%

γ2 −0.007 −0.001 −0.008 86%

γ3 −0.013 −0.003 −0.015 83%

γ4 −0.027 −0.003 −0.030 90%

γ5 −0.038 −0.004 −0.042 90%

40

Interdealer63%

BoughtbyEnd‐Users

SoldbyEnd‐Users19%

63%End‐Users18%

(a) In Terms of Notional Amount

Interdealer56%

B ht b

SoldbyEnd‐Users23%

BoughtbyEnd‐Users21%

(b) In Terms of Number of Trades

Figure 1: Market shareThis figure depicts the market share of interdealer trading and of end-users. Figure 1(a) shows the market share ofthe traders in the CDS market as a fraction of the total notional amount traded, and Figure 1(b) details the marketshares in terms of the total number of transactions.

41

Fig

ure

2:

Invento

ryP

osi

tion

sO

fK

ey

Mark

et

Part

icip

ants

Over

Tim

eT

his

figu

retr

acks

the

inve

nto

ryev

olu

tion

of

CD

Sco

ntr

act

son

all

35

refe

ren

ceen

titi

esh

eld

by

dea

lers

(blu

eli

ne)

,h

edge

-fu

nd

s(r

edli

ne)

,as

set

man

ager

s(g

reen

lin

e)an

db

an

ks

(ora

nge

lin

e).

Th

esh

ad

edare

as

rep

rese

nt

key

stre

ssev

ents

:1)

the

mon

eym

arke

tfr

eeze

on

Au

gu

st9,

2007;

2)

Bea

rS

tearn

sco

llap

seon

Marc

h14,

2008;

3)

Leh

man

Bro

ther

sd

ebac

leon

Sep

tem

ber

15,

2008;

4)

the

ap

pro

val

of

the

Tro

ub

led

Ass

etR

elie

fP

rogra

mon

Oct

ob

er2008;

and

,5)

the

init

iati

onof

the

stre

sste

sts

on

Feb

ruary

24,

2009.

42

(a) Over All the Sample Period

(b) After Lehman’s Failure on September 15, 2008

Figure 3: Top Five Dealers’ Exposures NetworkFigure 3(a) depicts the network of bilateral exposures among the top five dealers over the entire sample period. Thereported numbers on the edges are the daily averages over the entire sample period of the net amount traded acrossall reference entities in million of USD. Figure 3(b) represents the changes that occurred after Lehman Brothers’collapse on September 15, 2008. The two figures suggest that the total average exposure among the top five dealersincreased in the period after Lehman.

43

Net Pos Demand of EU x Inventory, 8.83

(16%)Net Pos Demand of EU x 

Exposure, 0.58(1%)

Net Pos Demand of EU x Exposure x Inventory, 0.27

(0%)

Exposure, 0.30(1%)

Contemporaneous Equity Return, 9.10

(17%)

Lagged Equity Returns, 6.69 (12%)

Lagged CDS Returns, 8.13 (15%)

20

30

40

50

Net Pos Demand of EU, 15.28(28%)

( )

Inventory, ‐5.27(10%)

‐10

0

10

Lagged CDS Returns

Lagged Equity Returns

Contemporaneous Equity Return

Exposure

Inventory

Net Pos Demand of EU x Exposure x Inventory

Net Pos Demand of EU x Exposure

Net Pos Demand of EU x Inventory

Net Pos Demand of EU

Figure 4: The Mean Impact of Net Buying by End-UsersThis figure breaks-down the price impact in response to net buying of end-users. I use the mean values of equity andCDS returns, end-users’ net buying quantities, and dealers’ inventory level, and calculate the contribution of each ofthe variables using its estimated regression coefficient.

44

Lagged CDS Returns, 1,176(5%)

15,000

Net Pos Demand of EU x Exposure x Inventory, 5,394

(23%)10,000

Net Pos Demand of EU x Exposure, 5,9165,000

Lagged CDS Returns

Lagged Equity Returns

Contemporaneous Equity Return

Net Pos Demand of EU x Inventory, 1,060

(5%)

(26%) Contemporaneous Equity Return

Exposure

Inventory

Net Pos Demand of EU x Exposure x Inventory

Net Pos Demand of EU x ExposureNet Pos Demand of EU, 928

(4%)

Contemporaneous Equity

0Net Pos Demand of EU x Exposure

Net Pos Demand of EU x Inventory

Net Pos Demand of EU

Contemporaneous Equity Return, ‐7,132

(31%)‐5,000

Lagged Equity Returns, ‐1,325(6%)

‐10,000

Figure 5: The Maximum Impact of Net Buying by End-UsersThis figure breaks-down the price impact in response to net buying of end-users. I use the maximum values of equityand CDS returns, end-users’ net buying quantities, and dealers’ inventory level, and calculate the contribution ofeach of the variables using its estimated regression coefficient.

45

Net Neg Demand of EU, 3.02( %)

Net Neg Demand of EU x Exposure, 0.56

(1%)

Exposure, 0.30(1%)

Contemporaneous Equity Return, 9.10

(21%)

Lagged Equity Returns, 6.69(15%)

Lagged CDS Returns, 8.13(19%)

5

10

15

20

25

30

Lagged CDS Returns

Lagged Equity Returns

Contemporaneous Equity Return

Exposure

Inventory

Net Neg Demand of EU x Exposure x Inventory

Net Neg Demand of EU x Exposure(7%)

Net Neg Demand of EU x Inventory, ‐10.18

(23%)

(1%)

Net Neg Demand of EU x Exposure x Inventory, ‐0.62

(1%) Inventory, ‐5.27(12%)

‐20

‐15

‐10

‐5

0Net Neg Demand of EU x Exposure

Net Neg Demand of EU x Inventory

Net Neg Demand of EU

Figure 6: The Mean Impact of Net Selling by End-UsersThis figure breaks-down the price impact in response to net selling of end-users. I use the mean values of equity andCDS returns, end-users’ net buying quantities, and dealers’ inventory level, and calculate the contribution of each ofthe variables using its estimated regression coefficient.

46

Net Neg Demand of EU x

Exposure, 50(0%)

Lagged CDS Returns, 1,176(5%)

4,000

6,000

Net Neg Demand of EU, 129(1%)

Net Neg Demand of EU x I 862

Net Neg Demand of EU x Exposure, 4,014

(17%)

0

2,000

Inventory, ‐862(4%)

Net Neg Demand of EU x Exposure x Inventory 8741 74

‐4,000

‐2,000

Lagged CDS Returns

Lagged Equity Returns

Contemporaneous Equity ReturnExposure x Inventory, ‐8741.74(37%)

Inventory, ‐10.43

‐8,000

‐6,000Contemporaneous Equity Return

Exposure

Inventory

Net Neg Demand of EU x Exposure x Inventory

Net Neg Demand of EU x ExposureInventory,  10.43(0%)

Contemporaneous Equity Return 7 132

‐12,000

‐10,000Net Neg Demand of EU x Exposure

Net Neg Demand of EU x Inventory

Net Neg Demand of EU

Lagged Equity Returns ‐1 325

Return, ‐7,132(30%)

‐16,000

‐14,000

Lagged Equity Returns,  1,325(6%)

‐20,000

‐18,000

Figure 7: The Maximum Impact of Net Selling by End-UsersThis figure breaks-down the price impact in response to net selling of end-users. I use the maximum values of equityand CDS returns, end-users’ net buying quantities, and dealers’ inventory level, and calculate the contribution ofeach of the variables using its estimated regression coefficient.

47

Figure 8: Distribution of the Contract MaturitiesThis figure plots the number of contracts that were traded in each maturity. The transactions include new tradesand assignments. There were a few transactions with maturity greater than 10 years that were not included in thisfigure.

48

Fig

ure

9:

Exam

ple

of

Tra

nsa

cti

on

Pri

ces

vs

Qu

ote

dP

rices:

No

Convers

ion

Th

isfi

gure

plo

tsth

epri

ces

for

5-Y

matu

rity

contr

act

son

AIG

.T

he

figu

reco

mp

are

sth

een

d-o

f-d

aym

id-q

uote

sup

pli

edby

Mar

kit

and

the

tran

sact

ion

pri

ces

as

rep

ort

edby

DT

CC

.A

lmost

all

tran

sact

ion

pri

ces

that

are

not

ali

gn

edw

ith

the

qu

oted

pri

ces

star

tin

gA

ugu

st2008

are

ass

oci

ate

dw

ith

an

up

front

pay

men

t.D

ue

toth

eon

goin

gd

evel

op

men

tsin

the

CD

Sm

arket

,si

ngl

e-n

ame

CD

Sco

ntr

act

sin

Nort

hA

mer

ica

start

edtr

ad

ing

wit

ha

fixed

cou

pon

of

eith

er100

or50

0b

ps

and

anu

pfr

ont

pay

men

taft

erth

eB

igB

an

gw

as

imp

lem

ente

don

Ap

ril

8,

2009.

Som

em

ark

etp

art

icip

ants

star

ted

trad

ing

inth

isfo

rmev

enb

efore

this

chan

ge

bec

am

eoffi

cial.

Th

etr

an

sact

ion

pri

ces

incl

ud

ein

terd

eale

rp

rice

s(b

lue

dot

)an

dd

eale

r-cl

ient

pri

ces

(red

dots

repre

sent

trad

esth

at

the

clie

nt

sold

toth

ed

eale

r;gre

end

ots

rep

rese

nt

trad

eth

ecl

ient

bou

ght

pro

tect

ion

from

the

dea

ler)

.

49

Fig

ure

10:

Exam

ple

of

Tra

nsa

cti

on

Pri

ces

vs

Qu

ote

dP

rices:

Ap

ply

ing

Convers

ion

Th

isfi

gure

plo

tsth

epri

ces

for

5-Y

matu

rity

contr

act

son

AIG

.T

he

figu

reco

mp

are

sth

een

d-o

f-d

aym

id-q

uote

sup

pli

edby

Mar

kit

and

the

tran

sact

ion

pri

ces

as

rep

ort

edby

DT

CC

.P

rice

sof

trad

esth

at

wer

ere

port

edas

an

up

front

pay

men

tan

dfi

xed

spre

adw

ere

conve

rted

usi

ng

Equ

ati

on

12.

50

Appendices

A Unwinding Mechanisms

When an investor wants to unwind an existing contract, he faces three alternatives, each of

which carries different implications for the level of counterparty risk exposure in the market.

The investor may:

1. Enter into a new offsetting transaction: The investor takes the opposite protection po-

sition. This offsetting transaction should have virtually the same contract terms as the

original CDS regarding the reference entity, the maturity date, etc., except for (i) the

premium, which is dictated by the prevailing market level; and (ii) the tenor, which is the

remaining time to maturity of the original CDS.

Strictly speaking, this is hedging rather than unwinding, as it leaves the original position

open and creates a new one. It also involves more documentation and additional legal

and counterparty risk. An investor must ensure that any legal or other basis risk between

the two transactions is minimized.

2. Assign the contract to a new counterparty: The investor can transfer the contract to a

new counterparty, and settle the mark-to-market gain or loss with this new counterparty.

This assignment, or “novation” as it is called in the ISDA Definitions, involves the original

counterparties, the “Transferor” and the “Remaining Party”, and the new party, the

“Transferee”, all agreeing to the transfer of all the Transferor’s rights and obligations to

the Transferee. Part of this is obviously that the Remaining Party must agree to take

on the counterparty risk of the Transferee (the ISDA 2005 Novation Protocol entails the

technical details). The Transferor thereby ends his involvement in the transaction. Notice

that assignments do not change the aggregate exposure levels in the market, but they do

increase the notional amount traded.

At the inception of the market, assignments were relatively infrequent; the usual method

of exiting a transaction was an offsetting transaction. But as hedge funds have become

more active in trading CDS, assignments have become increasingly common. Figure

11 depicts the use of assignments (in terms of notional amount traded and number of

assignments) over-time for the 35 reference entities in the sample.

3. Terminate the transaction with the original CDS counterparty: The investor can settle the

mark-to-market gain or loss with the original counterparty and terminate the contract,

conditioned on the latter’s consent. There will be no residual cash flows or exposure to

the counterparty’s credit risk.

51

Figure 11: Assignments Over TimeThis figure plots the daily notional amount traded through assignments as well as the daily number of assignmentsover all 35 reference entities.

B The List of Variables

Qu,eu,t is the net change in the inventory of end-users as a result of all trades on day t.

All transactions are bilateral, between an end-user and a dealer. Qu,eu,t < 0 if end-users

are net sellers for the day, and Qu,eu,t > 0 if end-users are net buyers for the day.

Qu,eu,t is the net change in the inventory of end-users divided by the standard-deviation

of dealers’ inventories.

Qu,d2d,t is the gross notional traded among dealers on day t, excluding double-counting

of both sides of the contract.

Qu,d2d,t is the gross notional traded among dealers divided by the standard-deviation of

dealers’ inventories.

Iu,t the cumulative position of all dealers in reference entity u on day t.

52

C Robustness Tables

Table 12Robustness: The Effect of End-Users’ Daily Trading Activity on CDS Spreads Using

Arithmetic Differences

In this table I study the asymmetric impact of end-users’ order imbalances on first-difference of CDS prices. Thetable details the coefficient of the following pooled-regression:

∆Su,t = α+

K∑k=0

γk

∣∣∣Qu,eu,t−k∣∣∣ 1>0 +

K∑k=0

δk

∣∣∣Qu,eu,t−k∣∣∣ 1<0 +

5∑n=0

ζn∆Pu,t−n +

5∑n=1

ηn∆Su,t−n + εu,t

where∣∣∣Qu,eu,t−k∣∣∣×1>0 is the standardized net daily notional amount bought by end-users. Similarly,

∣∣∣Qu,eu,t−k∣∣∣×1<0

is the standardized net daily notional amount sold to end-users. Business days without any trading activity areincluded in the sample with Qu,eu,t as zero, before the standardization. Standard errors that appear in parenthesesare heteroscedastic-robust standard errors. ∗∗∗, ∗∗, and ∗ denotes significance at 1%, 5%, and 10% level, respectively.All regressions also include an intercept.

Dependent Variable: ∆Su,tK = 0 K = 1 K = 3 K = 5

γ0 18.8∗∗∗ 16.0∗∗∗ 16.4∗∗∗ 15.7∗∗

(6.2) (6.1) (6.2) (6.3)∑Kk=1 γk 17.1∗∗∗ 22.8∗∗∗ 29.9∗∗∗

(6.1) (7.4) (9.5)δ0 −10.0 −11.8 −10.5 −11.2

(7.5) (7.8) (7.8) (7.8)∑Kk=1 δk 5.0 −16.4 −11.2

(11.4) (15.0) (16.0)ζn × 104 17.3∗∗∗ 17.2∗∗∗ 17.3∗∗∗ 17.3∗∗∗

(6.0) (6.0) (5.9) (5.9)∑5n=1 ζn × 104 19.7 19.9 19.3 19.4

(17.9) (17.9) (17.9) (17.9)∑5n=1 ηn × 10 1.7∗∗ 1.7∗∗ 1.7∗∗ 1.7∗∗

(0.8) (0.8) (0.8) (0.8)

Coefficient estimates OLS OLS OLS OLSStandard errors White White White WhiteFixed-Effects Time (Day) Time (Day) Time (Day) Time (Day)

R-squared 0.15 0.15 0.15 0.15Number of Observation 18, 813 18, 778 18, 709 18, 639

53