Counterparty Risk and Counterparty Choice in the Credit Default Swap Market

Finance and Economics Discussion Series Divisions of Research & Statistics and Monetary Affairs Federal Reserve Board, Washington, D.C. Counterparty Risk and Counterparty Choice in the Credit Default Swap Market Wenxin Du, Salil Gadgil, Michael B. Gordy, and Clara Vega 2016-087 Please cite this paper as: Du, Wenxin, Salil Gadgil, Michael B. Gordy, and Clara Vega (2016). “Counterparty Risk and Counterparty Choice in the Credit Default Swap Market,” Finance and Economics DiscussionSeries2016-087. Washington: BoardofGovernorsoftheFederalReserveSystem, https://doi.org/10.17016/FEDS.2016.087. NOTE: Staff working papers in the Finance and Economics Discussion Series (FEDS) are preliminary materials circulated to stimulate discussion and critical comment. The analysis and conclusions set forth are those of the authors and do not indicate concurrence by other members of the research staff or the Board of Governors. References in publications to the Finance and Economics Discussion Series (other than acknowledgement) should be cleared with the author(s) to protect the tentative character of these papers.

Counterparty Risk and Counterparty Choice in the Credit Default Swap Market∗ Wenxin Du Salil Gadgil Michael B. Gordy Clara Vega† September 8, 2016 Abstract We investigate how market participants price and manage counterparty risk in the post-crisis period using confidential trade repository data on single-name credit default swap (CDS) transactions. We find that counterparty risk has a modest impact on the pricing of CDS contracts, but a large impact on the choice of counterparties. We show that market participants are significantly less likely to trade with counterparties whose credit risk is highly correlated with the credit risk of the reference entities and with counterparties whose credit quality is relatively low. Furthermore, we examine the impact of central clearing on CDS pricing. Contrary to the previous literature, but consistent with our main findings on pricing, we find no evidence that central clearing increases transaction spreads. Keywords: Counterparty credit risk, credit default swaps, central clearing. JEL Classifications: G12, G13, G24 ∗The authors are with the Federal Reserve Board, 20th and C Streets NW, Washington, DC 20551, USA. The authors can be reached via email at wenxin.du@frb.gov, salil.u.gadgil@frb.gov, michael.gordy@frb.gov, and clara.vega@frb.gov. We are grateful to DTCC for providing the data. We have benefitted from helpful comments from Aaron Brown, Sean Campbell, Eduardo Canabarro, Jorge Cruz Lopez, Michael Gibson, Erik Heitfield, Greg Hopper, Michael Imerman, Charles Jones, Pete Kyle, Francis Longstaff, Batchimeg Sambalaibat, Martin Scheicher, Emil Siriwardane, Fabrice Tourre, Ingrid Werner, and participants in seminars at Johns Hopkins University and Lehigh University and in the American Finance Association meetings, Mont Tremblant Risk Management conference, Women in Microstructure conference, Early-Career Women in Finance conference, and IFSID Conference on Structured Products and Derivatives. The opinions expressed here are our own, and do not reflect the views of the Board of Governors or its staff. †Vega: Corresponding author.

1 Introduction Counterparty risk in over-the-counter (OTC) derivative markets played an important role in the propagationoftheglobalfinancialcrisisin2008. TheinabilityofBearStearnsandLehmanBrothers to find counterparties willing to trade, as their troubles became apparent, hastened their descent into insolvency (Duffie, 2010). Senior policymakers justified government assistance in the sale of Bear Stearns to JP Morgan Chase, in large part, by the need to avoid the further dislocations in OTC derivative markets that would have ensued in a rush to liquidate Bear Sterns’ collateral and to replicate positions with new counterparties. The bailout of AIG was motivated by the fear of a cascade of counterparty defaults in credit default swap (CDS) markets.1 Structural reforms introduced by Title VII of the Dodd-Frank Act in the United States and similar measures in the European Union were intended to reduce dramatically the scope for counterparty risk in derivative markets to generate systemic crises. In this paper, we investigate how market participants manage and price counterparty risk in the changing regulatory environment of the post-crisis period. We use four years (2010–13) of confidential transaction level data from the CDS trade repository maintained by the Depository Trust & Clearing Corporation (DTCC) to estimate the effects of counterparty risk on the choice of counterparties and CDS prices. Theliteraturehasthusfarbeenfocusedonthepriceeffectsofcounterpartyrisk. FollowingArora, Gandhi, and Longstaff (2012), we begin by investigating whether transaction spreads decrease with the credit risk of the seller and increase with the credit risk of the buyer. We separately analyze client-facing (i.e., between dealer and non-dealer) transactions and interdealer transactions. A priori, it is not clear which of these sets of transactions will show larger price effects of counterparty risk. On the one hand, many client-facing transactions have unilateral collateral terms in favor of the dealer, i.e., clients may be required to post initial margin in the dealer’s account, while dealers may be exempt from posting any collateral to the client. In contrast, interdealer transactions have bilateral collateral terms. Given this asymmetry in market practice, prices of client-facing transactions should be more sensitive to dealer’s credit spreads than interdealer transactions. On the other hand, in a market where dealers have some monopoly power, clients may be less able to 1TheFinancialCrisisInquiryCommission(2011)reportprovidesadetailednarrativebasedonprimarydocuments and testimony of senior policymakers and industry leaders. See especially pp. 287, 291, 329, and 347. 1

extract compensation for bearing a dealer’s counterparty risk relative to other dealers (Siriwardane, 2015). In a sample of client-facing transactions, we find statistically significant but economically modest effects of counterparty credit spreads on transaction spreads. We find even smaller pricing impacts, if any at all, arising from counterparty credit risk for interdealer transactions. These findings are consistent with asymmetric collateral terms for clients, and substantively consistent with Arora, Gandhi, and Longstaff (2012), who find that the effect of dealer counterparty risk on CDS quotes is statistically significant but very small in magnitude for a single institutional investor. Wenextinvestigatewhethernon-dealersmanagecounterpartyexposurebychoosingdealercounterpartiesbasedondealerriskprofiles,which,sofarasweareaware,hasnotpreviouslybeenstudied. We estimate a multinomial logit model for the non-dealer buyer’s choice of dealer counterparty, and find that non-dealers are less likely to buy protection from dealers whose credit quality is relatively low. This is consistent with the view that investors manage counterparty risk via counterparty choice. We also find that market participants are less likely to buy CDS protection from a dealer whose credit risk is highly correlated with the credit risk of the reference entity, i.e., buyers of protection avoid wrong-way risk. We estimate a similar model for the seller’s choice of dealer counterparty. Similarbutsomewhatweakereffectsareobservedinthenon-dealerseller’schoiceofdealer counterparty. The credit-rationing implied by our results is consistent with the findings of Afonso, Kovner, and Schoar (2011), who show that weak banks are rationed in the U.S. interbank market even in normal times. We also find evidence of stickiness in non-dealer choice of dealer counterparty, i.e., a nondealer is more likely to choose a dealer with whom it has transacted in the recent past. Previous literature has highlighted the persistence of customer-dealer relationships in equity dealer markets (e.g.,Bernhardt,Dvoracek,Hughson,andWerner,2005)andinotherover-the-countermarkets(e.g., Afonso, Kovner, and Schoar, 2014; Hendershott, Li, Livdan, and Schürhoff, 2016). This finding is consistent with economies of scale or scope in client trading relationships such as efficiencies in collateral usage. As an extension, we explore whether counterparty choice depends on the investment horizon of the non-dealer. The longer a position is to be held, the greater the exposure to the risk of the counterparty. Consistent with this intuition, we find that non-dealers with a short investment horizon are less sensitive as buyers of protection to the default risk of the dealer, relative to longer 2

horizon buyers. A consequence is that short horizon non-dealers may be more affected than others by a sudden deterioration in the solvency of a dealer. Lastly, weinvestigatewhethercentralclearinghashadanimpactonhowparticipantspriceCDS contracts. Clearinghouses have strict collateral and margin requirements for clearing members and maintain additional default funds to cover capital shortfall in the event of counterparty default, and thereby greatly reduce counterparty risk. Loon and Zhong (2014) hypothesize that centrally cleared trades should have higher spreads than uncleared trades, and report evidence in support. Contrary to their findings, we find that transaction spreads on centrally cleared trades are significantly lower relativetospreadsoncontemporaneousunclearedtransactions. Inthetime-series,thedatadoesnot support the hypothesis that central clearing eligibility increases transaction spreads. These results are consistent with our view that counterparty risk has minimal effect on pricing. Our paper adds to the literature on counterparty risk management in OTC derivative markets. Theoretical treatments of counterparty risk valuation include Cooper and Mello (1991), Duffie and Huang (1996), Jarrow and Yu (2001), and Hull and White (2001). Pykhtin (2011) provides a succinct introduction to the effects of netting, collateral, thresholds, margin call frequency and grace periods on the dynamics of counterparty exposure. For a comprehensive treatment, see Gregory (2010). Turnbull (2014) reviews the construction of credit valuation adjustments used by dealers in marking OTC derivative portfolios. Empirical research testing these models, however, is sparse. Besides Arora, Gandhi, and Longstaff (2012), Giglio (2014) infers counterparty risk from thecorporatebond-CDSbasis. Shachar(2012)usestheDTCCdataduringtheglobalfinancialcrisis period and shows that liquidity deteriorates as counterparty exposures between dealers accumulate. Other papers using the DTCC CDS data include Oehmke and Zawadowski (2013) and Siriwardane (2015), although the focus of these papers is not on counterparty risk. For a broad survey of the literature on the CDS market, see Augustin, Subrahmanyam, Tang, and Wang (2014). We also contribute to the literature on the impact of central clearing, which includes recent studies by Duffie and Zhu (2011), Bernstein, Hughson, and Weidenmier (2014), Loon and Zhong (2014) and Duffie, Scheicher, and Vuillemey (2015). Campbell and Heitfield (2014) describe postcrisis reforms aimed at encouraging central clearing. We proceed as follows. In Section 2, we provide background on counterparty risk in the CDS market and describe the DTCC data. We re-visit the evidence in Arora, Gandhi, and Longstaff 3

(2012) and test the effect of counterparty credit risk on CDS pricing in Section 3. In Section 4, we estimate the multinomial choice model for buyers and sellers of protection. In Section 5, we examine the effects of central clearing on cross-sectional and time series variations in CDS pricing. Section 6 concludes. 2 Background and Data Description 2.1 Background on pricing and managing counterparty risk Market participants respond to counterparty risk either by managing the risk or by demanding compensation for bearing the risk. Broadly, four mechanisms have evolved for risk-management. First,counterpartiesarrangefornettingofoffsettingbilateralpositionsandcollateralizetradesunder the terms of a credit support annex (CSA). In the aftermath of the financial crisis, interdealer CSAs have required the daily exchange of variation margin equal to the change in the market value of the bilateral portfolio. In practice, variation margin mitigates but does not eliminate counterparty risk. A counterparty in distress can exploit valuation disputes and grace periods to delay delivery of collateral, and the failure of a dealer is likely to coincide with unusual market volatility and reduced liquidity. Client CSAs are subject to negotiation and are therefore more varied in terms. Exchange of collateral takes place when the unsecured exposure exceeds an agreed threshold, which essentially serves as a limit on a line of credit. For the largest and most-favored clients, the thresholds are typically symmetric between dealer and client, and often tied to an agency credit rating of the counterparty.2 As witnessed in the case of AIG during the financial crisis, ratings-based thresholds may prove ineffective, as the event of downgrade of a large financial institution may trigger the immediate default of that institution (see Financial Crisis Inquiry Commission, 2011, Chapter 19). Furthermore, for smaller or more-risky clients, the CSA may require the posting of initial margin by clienttodealerandmayexemptthedealerfrompostinganycollateraltotheclient(i.e., thedealer’s threshold is infinite). Such clients are therefore most exposed to the credit risk of the dealer. 2Dealer banks all are rated by the major rating agencies. For an unrated client, the threshold may depend on an internal credit rating assigned by the dealer. 4

As with other studies in this literature, we are unable to observe bilateral CSA agreements, exchange of collateral between counterparties, and bilateral counterparty exposures in other derivative classes that are likely to be in the same netting set (e.g., interest rate derivatives). Thus, we cannot address the effects of collateralization and netting in mitigating counterparty risk, and simply maintain the assumption that these mitigants do not fully eliminate counterparty risk. Second, regulatory reform has mandated central clearing of trades on most standardized and liquid OTC contracts. Central counterparties impose standardized margining rules and effectively mutualize counterparty risk. In the CDS market, recent series of the most heavily-traded indices are eligible for clearing, as are the constituent single-name swaps. While central clearing of many North American indices became mandatory recently, central clearing of single-name swaps remains voluntary. In our sample period, we find that central clearing of interdealer single-name swaps is common, but clearing on behalf of clients is quite rare (we observe two instances in a sample of over one thousand transactions on eligible single-name reference entities involving a non-dealer counterparty).3 We proceed under the assumption that central clearing is not yet a viable riskmitigating option for non-dealers engaged in trading single-name CDS. Third, market participants can hedge counterparty risk by purchasing CDS protection on their dealercounterpartiesasreferenceentities. Suchhedgingwouldbedifficulttoexecuterigorouslydue to the stochastic size of the exposure, but market participants might pursue approximate strategies. Gündüz (2015) shows that financial institutions buy more protection on a dealer as reference entity when exposed to that dealer through counterparty risk. However, non-dealers hedge in this manner at lower frequency than do the dealer banks. Fourth, marketparticipantscanmitigatecounterpartyrisksimplybytradingpreferentiallywith counterpartiesthatarelessriskyorlesscorrelatedwiththeunderlyingreferenceentity. Forexample, if dealer ABC were to become too risky, participants might preferentially trade with ABC when a contract offsets existing bilateral exposure, but otherwise preferentially trade with other dealers. In addition, market participants may simply avoid buying protection from counterparties whose credit risk is highly correlated with credit risk of the reference entities. For example, a buyer of CDS protection on French banks might avoid transacting with a French dealer. 3Thefinancialpresshasreportedonrecentindustryinitiativestolowerbarrierstocentralclearingofsingle-name CDS for non-dealers. See, for example, “Credit Suisse, Goldman Said Mulling Plan to Promote CDS Clearing,” Bloomberg Business, November 9, 2015. 5

Finally, counterparty risk may be reflected in transaction prices of derivative contracts. The credit valuation adjustment (CVA) measures the difference in values between a derivative portfolio and a hypothetical equivalent portfolio that is free of counterparty risk. Intuitively, it represents the cost of hedging counterparty risk in the bilateral portfolio. To the extent that this cost can be imposed on the counterparty through the terms of trade, we will observe the price of a contract varying with the credit risk of the counterparties.4 It is important to recognize that adjustments to pricing do not mitigate counterparty risk, but rather serve as compensation for bearing the risk. The CVA is the net present value of future losses, so in normal circumstances it will be orders of magnitude smaller than the potential losses that could result from counterparty default. Whether managed or priced, counterparty risk in the CDS market has a natural asymmetry betweenbuyerandsellerofprotection. Ifthesellerofprotectiondefaultspriortothereferenceentity, losstothebuyercanbeaslargeasthenotionalvalueofthecontract. Ifthebuyerdefaults,theseller’s lossisboundedabovebythediscountedpresentvalueoftheremainingstreamofpremiumpayments, which is typically one or two orders of magnitude smaller than the notional amount. Furthermore, because financial firms (especially dealer banks) are more likely to default when prevailing credit losses are high, wrong-way risk is invariably borne by the buyer of protection. Thus, we expect the buyer of credit protection to be more sensitive to the credit risk of the seller than the seller is to the credit risk of the buyer. 2.2 DTCC CDS transaction data DTCC maintains a trade repository of nearly all bilateral CDS transactions worldwide. Each transaction record specifies transaction type, transaction time, contract terms, counterparty names and transaction price. We access the data via the regulatory portal of the Federal Reserve Board (FRB) into DTCC servers. The portal truncates the DTCC data in accordance with so-called entitlement rules (Committee on Payment and Settlement Systems, 2013, S3.2.4). As a prudential supervisor, the FRB is entitled to view transactions for which (i) at least one counterparty is an institution regulated by the FRB, or 4Inpractice,compensationforCVAmaybelimitedbythebilateralnatureofcounterpartyrisk. Iftwoequallyrisky counterparties with symmetric collateral terms enter a trade in which return distributions are roughly symmetric, then each demands similar compensation from the other. If the trade is to be executed, it will be executed near the hypothetical CVA-free price, so neither party will be compensated. 6

(ii) the reference entity is an institution regulated by the FRB. In particular, the largest dealer banks in the U.S. (Bank of America, Citibank, Goldman Sachs, JP Morgan Chase and Morgan Stanley) are FRB-regulated institutions, so we observe all trades by those major dealers and all trades on those dealer banks as reference entities. We refer to these major US dealer-banks as the “US5.” To avoid selection bias in our analyses, we will construct samples that align with one of the two entitlement windows. Our sample period is January 2010 through December 2013.5 Throughout our analyses, we consider only new, price-forming trades. Specifically, we drop novations, terminations, intra-family housekeeping transactions, and records resulting from trade compression. For a very small number of observations, the seller of the transaction is also the reference entity. Such contracts pose an extreme form of wrong-way risk (known as specific wrong-way risk in Basel capital rules), so it is somewhat puzzling that this is ever observed. We drop these few observations. To avoid undue influence of a single market participant on the results, we exclude trades involvingthelargestnon-dealerparticipantintheCDSmarketduringoursampleperiod. Thisnon-dealer appears as seller of protection in over 25% of client-facing (i.e., between a non-dealer and dealer) transactionsinourbaselinesample(describedbelow).6 ResultsforourpricingregressionsinSection 3) are robust to including this non-dealer in the sample. However, for a subset of our counterparty choice regressions in Section 4, the magnitude (but not the sign) of the coefficient changes materially.7 To maintain confidentiality of the data, we cannot report results both with and without this non-dealer in sample, as this would reveal trading behavior specific to this firm. Wenowdescribeconstructionofourbaselinesample. Toestimateamodelofcounterpartychoice, we must observe all transactions on a given reference entity for a given non-dealer. Therefore, we restrictourbaselinesampletotransactionswheretheunderlyingreferenceentityisregulatedbythe 5Our window has no overlap with the period of March 2008 to January 2009 studied by Arora, Gandhi, and Longstaff (2012), and overlaps only partially with the period of 2009–11 studied by Loon and Zhong (2014). 6Excluding this market participant reduces the Herfindahl index of concentration among sellers in notional value transacted from 500 to 389. The effect on concentration among buyers is much smaller. 7To estimate the counterparty choice regression, we assume that non-dealers chose which dealer to trade with, and not vice-versa. This assumption is not necessary in our pricing regression. Thus, it may not be surprising that our counterparty choice regression changes materially when we include this large non-dealer player, because this identifying assumption may not hold for this large non-dealer. 7

FRB. To mitigate any bias associated with illiquidity, we drop reference entities that are traded less than once per month on average. These restrictions leave us transactions on 12 reference entities.8 Similar to Arora, Gandhi, and Longstaff (2012), we restrict our analyses to trades involving at least one of the 14 largest CDS dealers: Bank of America Merrill Lynch, Barclays, BNP Paribas, Citibank,CreditSuisse,DeutscheBank,GoldmanSachs,HSBC,JPMorganChase,MorganStanley, RBS Group, Société Générale, UBS, and Nomura.9 These 14 dealers account for 99.8% of trades in our sample of liquid, FRB-regulated reference entities. To facilitate interpretation of the economic magnitude of our empirical estimates, we convert upfront points to par spreads using the standard ISDA conversion formula.10 The calculated par spreads can then be compared to Markit’s end-of-day par spread quotes. We drop observations if a valid par spread cannot be constructed or the underlying cannot be matched to a Markit spread for the same terms on the same date. Further, to ensure the comparison between spreads is valid,wedroptradesthatdonotadheretostandardmarketconventionsestablishedbyISDA.These conventions specify reporting protocols, coupon rates, credit event settlement procedures, and other administrative details. To ensure our results are not driven by large outliers, we drop observations for which the absolute difference between the logs of the Markit and DTCC spreads is greater than 0.3. This cutoff corresponds to the 98.5 percentile of absolute differences. After imposing the full set of restrictions, we are left with a baseline sample of 83,401 transactions. Summary statistics for the difference between DTCC and Markit par spreads are given in Panel A of Table 1. The median difference is 0.8 basis points and the 95th percentile of the difference is 14 basis points in the baseline sample. Inrobustnessexercises, wemakeuseofalargersampleconsistingofalltransactionsforwhichat least one counterparty is a US5 dealer. We do not require that these transactions have a reference entity regulated by the FRB, but all other restrictions described above are imposed. This sample contains some highly distressed reference entities which may be subject to significant intraday 8The included entities are Ally Financial, American Express, Bank of America, Capital One Bank, Capital One Financial Corporation, CIT Group, CitiGroup, JPMorgan Chase, Metlife, Morgan Stanley, Goldman Sachs Group, and Wells Fargo. 9Relative to the list of 14 dealers appearing in the sample of Arora, Gandhi, and Longstaff (2012), Lehman is dropped (as it no longer exists), Bank of America and Merrill Lynch are merged, and Nomura Holdings and Société Générale are added. 10We use initial payment, total notional amount and the ISDA convention interest accrual convention to compute the implied upfront points associated with each transaction. We then apply the program provided by the ISDA for conversion of upfront points into par spreads. 8

volatility. Therefore, we drop transactions for which the Markit par spread on the contract on that date is above 1000 basis points. This sample, which we refer to as the US5 counterparty sample, is comprised of over a million transactions on roughly 1600 single name reference entities. Summary statistics are provided in Table 2. We next define several variables used in our analyses. We measure the risk of dealer default by the dealers’s five year CDS spread quoted at the end of the previous trading day. For observation s date t, the lagged spread is denoted cds when the dealer is acting as selling of protection, and t−1 b cds when the dealer is acting as buyer. In robustness exercises, we consider an alternative t−1 measure for the risk of dealer default based on the bankruptcy hazard rate model of Chava and Jarrow (2004). In the baseline sample of entitled reference entities, we measure wrong-way risk, WWRs, as a i dummy variable equal to one if both the seller of protection is a US5 dealer and the reference entity is either a US5 dealer or Wells Fargo.11 For the case of a dealer buying protection, we construct our measure of right-way risk RWRb in exactly the same way, i.e., as a dummy variable equal to i one if both the buyer of protection is a US5 dealer and the reference entity is either a US5 dealer or Wells Fargo. We modify the variable name (from WWR to RWR) only to highlight that the economic interpretation is reversed. In the US5 counterparty sample, we measure WWRs (RWRb) t t as a dummy equal to one if the log CDS spreads on the reference entity and on the selling (buying) dealer have a correlation over 70%.12 Perhaps to achieve operational efficiencies, trading relationships in OTC markets often persist through time. In the case of buyer of protection b and seller of protection s, Relationss,b is t−1 defined as the share of notional value that market participant b traded with dealer s in the recent past. We measure this share using the past 28 business days prior to the transaction if there were more than 28 transactions in the last month, otherwise we estimate the share using the past 28 transactions, requiring a minimum of 10 transactions. To express Relationss,b in share terms, t−1 we divide the total notional value transacted between b and s by the total notional value that 11WellsFargoisgroupedwiththeUS5dealersforthispurposeduetosimilarityinsizeandnationalscaleofbanking operations. It is a dealer as well, but not a significant one in the CDS market. 12Tobeprecise,weestimatethecorrelationdailyusingarollingfive-yearwindow,whichimpliesthatthemeasure is dynamic. However, it is so strongly persistent that it is substantively correct to describe the variable as static. OurresultsfortheentitledreferenceentitiesarequantitativelysimilarifweinsteadconstructstaticWWRandRWR dummies for correlation above 70% over the entire sample period. 9

market participant b traded. When the analysis is done from the protection seller’s perspective, Relationsb,s is expressed as a fraction of the total notional value that market participant s traded. t−1 3 Effects of Counterparty Credit Spreads on CDS Pricing In this section, we study the effects of buyer and seller credit risk on CDS pricing. Under the maintainedassumptionthatOTCCDStradesareimperfectlycollateralized,protectionsoldbyhighrisk counterparties should be less valued than protection sold by low-risk counterparties. Whether thisdifferenceaffectsmarketprices, however, isanempiricalquestion. Ifitdoes, then, holdingfixed the buyer and contract, we expect sellers’ CDS spreads to be negatively correlated with transaction spreads. Similarly, as high-risk buyers of protection are less likely to fulfill their premium leg obligations than low-risk buyers, we expect buyers’ CDS spreads to be positively correlated with transaction spreads, holding fixed the seller and contract. We perform fixed effect panel regressions to test the hypotheses. Panel B of Table 1 summarizes characteristics of transactions on the same reference entity with the same tenor, tier, currency, restructuring or non-restructuring clause and fixed coupon rate, traded on the same date. We restrict these summary statistics to the subsample in which there are at least ten trades on the identical contracts during the same day, which is about 28 percent of our baseline sample. We find significant pricing dispersion within the day on the same contract, with a median within-day standard deviation of 1.4 percent. In terms of counterparty choice, we see that on average a buyer (seller) trades with more than one seller (buyer) on the same contract and the same day. Observing multiple counterparties for the same party and the same contract serves to identify whether cross-sectional pricing dispersion in transaction spreads varies with cross-sectional dispersion in counterparty credit spreads. 3.1 Effect of seller credit spreads on CDS pricing We investigate whether counterparty risk is priced in the CDS market from the protection buyer’s perspective. OurbenchmarkspecificationissimilarinspirittothatofArora,Gandhi,andLongstaff (2012). We compare the transaction spreads on the same contract, traded on the same date, bought by the same buyer, but sold by different sellers that vary in their credit risk. Identification comes 10

from pricing dispersion within the same day. Our benchmark specification is log(cdss,b)−log(cds ) = αb +βlog(cds s )+ηWWRs+λRelationss,b +δlog(size)+(cid:15)s,b, (1) i,t i,t i,t t−1 t t−1 i,t where log(cdss,b) is the log par spread on CDS transaction i on time t. Superscripts s and b denote i,t the seller and buyer of credit protection, respectively. We denote by cds the par spread quoted i,t by Markit on reference entity i on date t. The dependent variable measures the difference between a specific transaction spread and the Markit quote on the same reference entity at time t. s Independent variables of primary interest are the log of the seller’s quoted CDS spread (cds ), t−1 thewrong-wayriskindicator(WWRs),andthemeasureofpastbuyer-sellerrelationship(Relationss,b ). t−1 The fixed effectαb refersto the buyer-contract-timeinteractivefixed effect. Thelog ofthe notional i,t value of the traded contract, log(size ), is included in the regression to allow for the contract size i,t tohavesomepotentialimpactontransactionspreads. Ifsellercounterpartyriskispriced, weexpect β < 0 and η < 0. One potential concern with the benchmark specification is that the seller’s credit spread could be correlated with other unobserved characteristics of the sellers which also affect pricing of the contract. To mitigate this concern, we also add seller fixed effects αs to control for the impact of seller’s time-invariant characteristics on pricing in the following regression: log(cdss,b)−log(cds ) = αb +αs+βlog(cds s )+ηWWRs+λRelationss,b +δlog(size)+(cid:15)s,b. (2) i,t i,t i,t t−1 t t−1 i,t Estimates are reported in Table 3. We present results for equations (1) and (2) in Panels A and B, respectively, depending on the fixed effect specification. In all specifications, we restrict the sellers of protection to be the 14 largest dealers. We report the number of effective observations for which one buyer faces at least two different sellers in each fixed effect group. Our benchmark results are presented in Column 1. In this column, we examine the effect of seller’s credit spreads on transaction spreads for non-dealers as buyers of protection. In Panel A, the coefficient on the seller’s credit spread is negative and statistically significant, but the economic magnitude of the coefficient value -0.007 is very small. A 100 percent increase in the seller’s log spread translates to only a 0.7 percent decrease in the transaction spread. To translate the change 11

from percentages to levels, we note that the mean level of transaction spreads is about 195 basis points in the estimated sample and the mean dealer spreads is about 173 basis points, and hence a 100 basis point increase in the seller’s credit spreads translate into about 0.6 basis point reduction (cid:16) (cid:104) (cid:105)(cid:17) in the transaction spread = 195× (cid:0)173+100(cid:1)−0.007 −1 . The average notional value of trades 173 in Column 1 is about $7.4 million, so the 0.6 basis point pricing impact on transaction spreads translates into about $450 difference in the total cost of an average-sized trade. In Panel B, the magnitudeofthecoefficientontheseller’screditspreadincreasesto0.018withadditionalsellerfixed effects, but remains modest. Our finding that the impact of seller credit spread is significant, but smallineconomicmagnitude,isbroadlyconsistentwiththefindinginArora,Gandhi,andLongstaff (2012) that a 100 basis point increase in dealer spreads translates to 0.15 basis point reduction in the quoted CDS spread. Furthermore, we note that the WWR variable enters slightly positive, the sign opposite to that predicted by the counterparty risk hypothesis, and past relationship between the buyer and the seller does not enter significantly. In Column 2, we restrict the sample to the set of reference entities which are not eligible for central clearing, and obtain coefficient estimates very similar to those in Column 1. This suggests that clearing eligibility does not significantly affect client-dealer pricing. In Column 3, we repeat the same regressions for interdealer transactions. We obtain smaller negativecoefficientsontheseller’sCDSspreads,marginallysignificant. Forinterdealertransactions, an increase of 100 basis point in the seller’s spread translates into about 0.2 to 0.3 basis point reduction in the transaction spread. The coefficient on WWR is slightly negative, consistent with the counterparty risk hypothesis. The coefficient on past relationship is marginally negative, i.e., buyers obtain slightly more favorable prices from dealers with whom they traded more in the past. In Column 4, we examine interdealer pricing in the US5 counterparty sample. A trade appears in this sample if one or both of the counterparties is a major US dealer-bank. We find that the coefficient on the seller’s CDS spread is slightly positive and statistically insignificant across both fixed effect specifications. The coefficient on WWR is slightly negative. The coefficient on past relationship is slightly positive without the seller fixed effects, but loses significance after including seller fixed effects. In summary, we find significant but economically small effects for non-dealers as protection buyers, and either even smaller or insignificant effects of seller’s credit spreads on transaction spreads 12

for dealers as protection buyers from other dealers. These results are consistent with the anecdotal evidence that CSA provisions are symmetric between large dealers, but are more likely to be unilateral in favor of dealers for client-facing transactions. Neither WWR nor past relationship affects transaction spreads in a robust manner. 3.2 Effect of buyer credit spreads on CDS pricing Parallel to the previous subsection, we consider the effect of the buyer’s CDS spread on transaction spreads. First, we hold the seller, contract and trade date fixed, and examine the impact of buyer’s credit spreads on transaction spreads using the following regression: log(cdss,b)−log(cds ) = αs +γlog(cds b )+ηRWRb+λRelationsb,s +δlog(size)+(cid:15)s,b, (3) i,t i,t i,t t−1 t−1 i,t wherecds b denotesthebuyer’sCDSspreadandαs istheseller-contract-timefixedeffect. Wealso t−1 i,t include variables for the right-way risk (RWRb) and buyer-seller past relationship (Relationsb,s ). t−1 Asintheprevioussection,wealsoestimatearegressionwithadditionalbuyerfixedeffectstocontrol for time-invariant buyer characteristics: log(cdss,b)−log(cds ) = αs +αb+γlog(cds b )+ηRWRb+λRelationsb,s +δlog(size)+(cid:15)s,b. (4) i,t i,t i,t t−1 t−1 i,t If the buyer’s counterparty risk is priced, we should expect γ > 0 and η > 0. Table 4 reports our estimation results. For non-dealers as protection sellers (Columns 1 and 2), the coefficient on the buyer’s CDS log-spread are positive and significant in the fixed effect specification (Panel B), but somewhat smaller in magnitude than the corresponding coefficients in Table 3. For interdealer transactions reported in Columns 3 and 4, the buyer’s log-spread is insignificant. Theright-wayriskvariableentersasslightlypositive,consistentwiththecounterparty risk hypothesis, though not always statistically significant. Past relationship does not have a robust significant effect on transaction spreads. 13

4 Analysis of Counterparty Choice In this section, we show that market participants actively manage counterparty risk by choosing counterparties of better credit quality and less subject to wrong-way risk. As in Shachar (2012), we assume that trades in the OTC market are initiated by the non-dealer, and that the dealer supplies liquidity upon demand.13 An immediate implication is that the matching of counterparties in a transaction is determined exclusively by the choice of the non-dealer as client. Summary statistics are reported in Table 5. Our baseline sample contains 196 non-dealer buyers ofprotectionin11,948transactions,ofwhich7,927referenceUS5dealersandtheremainderreference otherFRB-regulatedinstitutions. Wehavefewerobservationsinoursellerchoicemodel: thesample contains 169 non-dealer sellers of protection in 7,767 transactions, of which 5,693 reference US5 dealers. Table 6 provides preliminary evidence of aversion to wrong-way risk in our baseline sample. We dividethe14dealersbydomicile(USvs.foreign), andalsosorttheFRB-regulatedreferenceentities into two groups by severity of WWR when the seller is a US dealer. A group of reference entities composed of the US5 dealers and Wells Fargo is deemed “high WWR,” and the remaining FRBregulated reference entities are deemed “low WWR.” We then calculate for each group of dealers the aggregate share of protection sold (by notional value) on the high and low WWR groups of reference entities. For US5 dealers, the trading share is 44 percent when the US5 dealer is selling protection on one of the US5 or Wells Fargo compared to a trading share of 54 percent for other reference entities. This suggests that non-dealers are less likely to buy CDS protection on the US5 dealers or Wells Fargo from one of these five US dealers than from a foreign dealer. We estimate McFadden’s (1974) multinomial conditional logit model for the choice made by the buyer of protection among the 14 dealers. The probability of choosing dealer s conditional on characteristics xs is specified as t exp(xs β) Pr(yb = s|xs ) = i,t , s = 1,...,N . (5) i,t i,t (cid:80)Ni exp(xsˆ β) i sˆ=1 i,t 13AsdetailedinSection2.2,wedropanon-dealerthataccountsformorethanone-quarterofCDSprotectionsales. For this institution, the assumption that non-dealers initiate the trade may not hold. 14

wherethechoicesethascardinalityN = 13whenthereferenceentityiisaUS5dealerandN = 14 i i otherwise. In total, our multinomial model has 159,345 (= 7,927×13+(11,948−7,927)×14) observations. The independent regressors are: credit risk of the seller, proxied as before by the CDS spread on the seller of protection quoted on Markit on date t−1; an indicator variable for general wrong-way risk (WWRs); Relationss,b , to allow for “stickiness” in buyer-dealer relationships; a set i t−1 of dummy variables for the N dealers, to allow for baseline differences in market share; interactions between seller dummy variables and the spread on the five-year CDX.NA.IG index, to allow for the possibility that buyers may gravitate towards particular sellers when market-wide spreads are high. Results are reported in Table 7. The coefficients on seller dummy variables are omitted to respect the confidentiality of the data. InColumn1wereportcoefficientsestimatedonthebaselinesample. Aspredicted,thecoefficient onseller’sCDSisnegativeandstatisticallysignificant,i.e.,customersarelesslikelytobuyprotection from a dealer whose own CDS spread is high relative to other dealers. The coefficient on WWR is large, negative and statistically significant, which shows that buyers avoid wrong-way risk in their choice of dealer. Finally, the coefficient on past relationship is large, positive and statistically significant, which is indicative of persistence in trading relationships. In Column 2 we report coefficients estimated on the subsample of FRB-regulated reference entities that are not eligible for clearing. The coefficients on the seller’s CDS and past relationships are similar to those in Column 1, which suggest that clearing eligibility has not had an impact on how non-dealers respond to our measure of dealer credit risk. The coefficient on WWR is not statistically significant, but this is not surprising because the reference entities that are eligible for clearing are also those that suffer least from wrong-way risk, so that dropping these observations makes it difficult to identify the impact of WWR. In Column 3 we report coefficients estimated on a subsample that excludes “captive” buyers; we designate a buyer as captive if at least 60 percent of their transactions as buyer face a single dealer as counterparty. To the extent that such buyers are truly captive to a single dealer, they may be unresponsive to our measures of dealer credit risk. Nonetheless, our coefficient estimates are similar to those in Column 1. MarginaleffectsforthisregressionarereportedinTable8. Weseparatelyreportmarginaleffects at sample means for the large dealers (those with unconditional transaction shares of 7–13%) and small dealers (those with unconditional transaction shares of 1–6%). We find that a 100 basis point 15

increase in a large dealer’s CDS spread is associated with an average decline in the likelihood of buyingprotectionfromthatdealerof2.6percentagepoints. Wrong-wayriskreducestheprobability by2percentagepoints. Aonestandarddeviationincreaseinpast-monthtransactioncountincreases the probability of selection by 4 percentage points. Relative to unconditional transaction shares of 7–13 percentage points, these effects are all of large economic magnitude. We next consider the possibility that the determinants of the buyer’s choice of seller might depend on the investment horizon of the buyer. A buyer intending to hold a CDS position for a short period of time should be less sensitive to counterparty credit risk than a buyer intending to hold a position to maturity. We define a dummy variable SHORTb equal to one if buyer b closes at least fifty percent of its new trades within 28 days of the original trade date. We re-estimate equation (5) including interactions of the independent variables with SHORTb, and report results in Table 9. Consistent with our prediction, we find that short-horizon buyers of protection are less than half as sensitive to the default risk of the dealer, relative to long-horizon buyers. We repeat these analyses from the perspective of non-dealer sellers of protection and report the resultsinTable10. Consistentwiththeasymmetrybetweenbuyerandsellercounterpartyexposure, we find that the seller’s choice of dealer counterparty is slightly less sensitive to the dealer’s credit risk. The marginal effects for this regression, reported in Table 11, indicate that a 100 basis point increase in a large dealer’s CDS spread is associated with an average decline in the likelihood of a non-dealer selling protection to that dealer of 2.2 percentage points compared to 2.6 percentage points in the buyer-choice model. In Table 12 we report results of interactions of the independent variables with SHORTs. We find no difference between short-horizon and long-horizon sellers of protection in sensitivity to the default risk of the dealer. As a robustness check, we replicate the fixed effect specification of Section 3 in a linearized choice model. As shown in Table 13, our results are robust to the inclusion of these fixed effects. As further robustness checks (not tabulated), we repeat our main analyses excluding from our sample the height of the European debt crisis.14 We obtain slightly higher coefficients on WWR, RWR, and dealer’s CDS when we exclude this period. Thus, our results are not driven by the European debt crisis period. 14Specifically, we exclude observations from October 4, 2011, when the Belgian government announced Dexia’s bailout, to July 26, 2012, when Mario Draghi announced that “the ECB is ready to do whatever it takes to preserve the euro. And believe me, it will be enough.” 16

In addition, we consider an alternative measure for the risk of dealer default, the dealer’s five year probability of default (PD) as estimated by Kamakura Corporation using the hazard rate model of Chava and Jarrow (2004). Whereas the CDS represents an assessment of credit risk under the pricing measure Q, the PD is an assessment under the empirical measure P. First, we repeat these analyses replacing the dealer’s five year CDS spread by the dealer’s five year PD. Second, we include both measures of the risk of dealer default in our multinomial logit model. We find that both measures of risk of dealer default affect the buyer’s choice of seller and the seller’s choice of buyer, but the dealer’s five year CDS spread has the bigger impact on the non-dealer’s choice of seller and buyer. Finally, we include as a control variable a proxy for the dealer’s pricing aggressiveness. This is intended to address a concern that the risk of dealer default is somehow related to how aggressively the dealer pursues CDS trade flow from clients. We measure a dealer’s pricing aggressiveness at date t as the average over the last 28 business days prior to date t of the difference between the log par spread the dealer charges its non-dealer counterparties and the log Markit par spread on the same reference entity.15 As expected, we find that in the buyer’s choice model the coefficient on dealer’s price aggressiveness is negative and statistically significant, while in the seller’s choice model the coefficient is positive and statistically significant. Yet, our results are robust to including thiscontrolvariablebecausethecoefficientsonthedealer’sfive-yearCDSspread, WWR,andRWR, are virtually unchanged. 5 Central Clearing Loon and Zhong (2014) find that central clearing increases CDS spreads, and attribute this to the mitigation of counterparty risk. Their finding could be seen as inconsistent with our result in Section 3 that counterparty risk has a modest effect on pricing. Our data does not support the Loon and Zhong (2014) result. In the cross-section, we find that transaction spreads from centrally cleared trades are associated with lower spreads than uncleared trades. In the time series, we find 15If there were fewer than 28 transactions in the last month, then we estimate the average using the past 28 transactions, requiring a minimum of 10 transactions. This measure is constructed on our baseline sample of FRBregulated reference entities. 17

no robust evidence that a reference entity’s transaction spreads increase around the commencement of eligibility for central clearing. Clearing was introduced first for CDX.NA.IG, an index composed of investment grade North American corporates, and iTraxx Europe, its European counterpart. Select single-name reference entities became eligible for clearing by Intercontinental Exchange (ICE) in waves beginning in December 2009. By the end of our sample period, most index constituents had been made eligible to clear. The single-names index constituents that remain ineligible are primarily the European dealer banks listed in iTraxx Europe. US dealer banks are excluded from the CDX.NA.IG index, and also remain ineligible for clearing. Within an investment grade index, reference entities have been made eligible for central clearing in cohorts, which typically consist of several firms in the same sector. Table 14 summarizes the introductionofclearingeligibilityduringoursampleperiodbysectorforCDX.NA.IGconstituents.16 Cohorts are typically small, and selection driven by tenure within the index. As an example, five firms in the Technology sector were made eligible for clearing on 19 April 2010, two more on 28 March 2011, and a final two on 13 June 2011. The first cohort was composed of reference entities that had been continuously listed in the CDX.NA.IG since at least September 2008. The second cohort consisted of reference entities added to the index by March 2009, while the last cohort was composed of reference entities added by March 2011. Our analysis exploits the staged introduction ofclearingforCDX.NA.IGconstituentstostudytimeserieseffectsofcentralclearingontransaction spreads. 5.1 Cross-sectional comparison for pricing In our sample period, there were two methods by which market participants could engage in cleared trades. Under the first method, known as backload clearing, the parties initially transact bilaterally in the OTC market, and later submit the trade to a central counterparty (CCP) for clearing, typically on the Friday following the trade. Our assumption is that the backloaded trades were designated for clearing by the counterparties at the time of the bilateral transaction. Under the 16Toavoidtheinfluenceoftechnicalevents,suchastheissuanceofnewREDidentifiersaftercorporaterenamings, weexcludereferenceentitieswithclear-eligiblepredecessorsfromouranalysis. SectorinformationcomesfromMarkit. On February 15, 2012, Markit modified its classification scheme. To avoid inconsistencies, we categorize reference entities by their sector under the old scheme. For the small number of reference entities that did not exist before February 15, 2012, we use the sector of their most recent predecessor classified under the old scheme. 18

second method, the trade is cleared on the same day as the initial trading date. These same-day clearingtradesareoftenclearedatinceptionandexecutedonaswapexecutionfacility(SEF),which matches buyer and seller anonymously. A same-day clearing trade appears in the repository data as two simultaneous transactions with a CCP as buyer on one leg and as seller on the other. As discussed in Section 2, non-dealers rarely clear trades during our sample period, so we assume that all cleared transactions in our sample are interdealer trades.17 Table 15 presents results on how transaction characteristics affect CDS pricing. We categorize transactions into four types: (i) same-day clearing trade; (ii) backload clearing trade; (iii) uncleared OTC client-facing trade; and (iv) uncleared OTC interdealer trade. The fourth type is the omitted category in the regressions. We use the subsample of transactions on clearable reference entities in the union of the baseline and US5 counterparty samples. Of the 509,248 transactions on clearable reference entities in which either the buyer of the seller is one of the 14 largest dealers, we have 368,402 transactions in which thebuyerisone ofthe14 largest dealers and409,053transactionsin whichthe seller isoneof the14 largest dealers. In Column 1, we estimate the effect of seller characteristics when the buyer is one of the 14 largest dealers. Holding contract, time and the buyer fixed, we find that same-day clearing tradesareassociatedwithsignificantlylowerspreadsthanOTCinterdealertrades,withamagnitude around 0.4 percent. Non-dealer sellers sell to dealers at spreads about 0.6 percent lower than on comparable OTC interdealer transactions. Backloaded clearing trades have marginally significantly lowerspreadsthantheinterdealerspreadsatabout0.1percent. InColumn2, weestimatetheeffect of buyer characteristics when the seller is one of the 14 largest dealers. Holding contract, time and seller fixed, we again find that same-day clearing trades are associated with lower spreads, with a magnitude around 0.2 percent. Furthermore, non-dealers buyers of protection in OTC transactions pay dealers about 0.4 percent more than dealers pay in comparable OTC interdealer transactions. Backloaded clearing trades do not enter significantly. The dealer’s pricing advantage we document is consistent with Siriwardane (2015), who shows that the market is dominated by a handful of buyers and sellers of protection, the majority of which are dealers. Siriwardane (2015) finds that a reduction in the capital of these dealers has an impact on CDS prices. 17Of the 210,680 instances of backload clearing trades in our data, only eight involve non-dealers. There are no instances of non-dealers engaging in same-day clearing on our entitled reference entities. Our entitlement rules preclude us from observing same-day clearing by non-US5 firms on non-entitled reference entities. 19

In Column 3, we fix contract and time only and allow both buyer and seller’s characteristics to enter simultaneously (while either the buyer or the seller is one of the largest 14 dealers). We confirmtheeffectsdocumentedinColumns1and2,i.e.,thatsame-dayclearingtradesareassociated with significantly lower transaction spreads, with a magnitude around 0.3 percent. Furthermore, dealers extract rents from non-dealer clients while buying and selling CDS protection. Estimated dealer rents in this specification are larger than in Columns 1 and 2, with magnitudes around 0.9 percent. Finally, mean spreads on backloaded clearing trades are slightly lower than on comparable interdealer OTC trades by about 0.2 percent. As an aside, we note that dealers’ pricing advantage over non-dealers is present before and after the introduction of central clearing. In Figure 1, we plot the average log-difference between transactionspreadsinDTCCdataandMarkitspreadsagainstthenumberofdayssincethereference entity became eligible for clearing. The top panel shows that transaction spreads are larger when dealers sell protection to non-dealers than when the dealers buy protection, and the gap does not appear to diminish following the introduction of clearing. In the bottom panel, we see that sameday clearing trades have lower spreads than OTC trades for which the non-dealer is the buyer, and higher spreads than OTC trades for which the non-dealer is the seller. Dealers’ pricing advantage over non-dealers prompts us to focus on interdealer transactions in the following subsection when we examine the time series effect of clearing eligibility. In summary, we find centrally cleared trades are associated with lower spreads compared with bilateral interdealer trades and client-facing trades when the client is a buyer, as opposed to higher spreads as a result of counterparty risk mitigation. 5.2 Time series comparison for pricing We estimate the time series effect of clearing eligibility on DTCC transaction and Markit quotes spreadsusingadifference-in-differences(DID)approach. OurDTCCsampleconsistsoftransactions in the union of the baseline and US5 counterparty samples. To distinguish clearly from possible time-variation in the magnitude of pricing advantage over non-dealers, we focus on interdealer and cleared transactions only. We exploit the fact that clearing eligibility was introduced by sector in smallcohortsforCDX.NA.IGconstituents,andusetheDIDspecificationtoestimatethetimeseries 20

effect of clearing: log(cdss,b) = α +βTreatment +γTreatment ×After +(cid:15) , (6) i,t sector,t i i i,t i,t where α denotes sector and date interactive fixed effects. The variable Treatment is a sector,t i dummyindicatingwhetherthereferenceentityiisinthetreatmentgroup, andAfter isadummy i,t indicating whether the reference entity i is eligible at time t. The coefficient γ represents the timeserieseffectofclearingeligibility. Weusetreatmentandcontrolgroupsinthesamesectortomitigate the impact of common macroeconomic and sectoral shocks on our estimates. We estimate three DID specifications. In the first specification, we use DTCC transactions or Markit quotes on reference entities cleared in the first cohort for each sector as the treatment group, and pre-eligibility DTCC transactions or Markit quotes on reference entities cleared in later cohorts as the control group. In the second specification, we use post-eligibility cleared transactions or quotes on the first cohort of clearable reference entities as the control group, and all transactions or quotes on reference entities in later cohorts in the same sector as the treatment group. We ensure that entities in both treatment and control groups of the first two DID specifications were already CDX index members before clearing was introduced, so that the effects of clearing eligibility can be separately identified from the effects of index inclusion. In the third specification, we use all transactions or quotes on all reference entities that become eligible during the sample period as the treatment group, and transactions on other North American investment grade reference entities as the control group. We construct our event windows in two different ways. First, we use the window 50 days before and after the introduction of clearing eligibility. Second, we follow Loon and Zhong (2014) and use the window between 250 days before the introduction of clearing to 20 days after the introduction of clearing. The effects on four intervals of [-20, -10), [-10, 0), [0, 10) and [10, 20) days around the introduction of clearing eligibility are separately examined. As a robustness check, we also perform our DID analysis using daily CDS quotes from Markit as dependent variable. We note that the numbers of observations in the DTCC and Markit samples can differ significantly in each specification. This is because daily Markit quotes are observed on 21

almost all business days for each reference entity in our sample. However, multiple transactions or no transactions might occur on these referee entities in our DTCC sample. DID results for transaction spreads and Markit quotes are depicted in in Figure 2. In all three specifications, mean transaction spreads and Markit quotes in the treatment group co-move with those in the control group very closely before and after commencement of eligibility. There is no discrete jump in treatment group around the commencement of clearing. Regression estimates confirm this visual observation. In Panel A of Table 16, we report DID estimates for the three specifications using a 50-day event window before and after clearing eligibility. The coefficient γ on the Treatment×After interaction is negligibly small and statistically insignificant for the first five columns and is only marginally significant for the last column. In Panel B, we use a window for the DID sample from 250 days before the introduction of clearing eligibility to 20 days afterwards. Across all three DID specifications for both DTCC transaction spreads and Markit quotes, none of the coefficients on the interactions between treatment and event intervals are significantly positive. Therefore, our evidence does not support the hypothesis that central clearing eligibility increases Markit quotes or transaction spreads. 6 Conclusion In this paper, we study how market participants price and manage counterparty credit risk in the CDS market. Our results demonstrate that participants in the credit default swap market manage counterparty risk by buying protection preferentially from counterparties of lower credit risk and lesser “wrong-way” correlation with the reference entity. These results are qualitatively consistent withtheexperienceofLehmanandBearStearnsin2008inwhichtheycouldnotfindcounterparties to trade with, and suggest that OTC dealers remain vulnerable to a run of counterparties in the post-crisis environment. Wefindlittleevidenceofmaterialpricingimpactsarisingfromcounterpartycreditriskforinterdealer transactions. For client-facing transactions, we find statistically significant but economically small effects of dealer credit spreads on transaction spreads. Furthermore, we find that centrally cleared trades are not associated with higher spreads, which is counter to the intuition that central clearing should increase the value of credit protection by reducing counterparty risk. 22

Therefore, we conclude that compensation for counterparty risk in prices plays a less important role in CDS markets than the management of counterparty risk by counterparty choice. 23

References Afonso, G., A. Kovner, and A. Schoar (2011): “Stressed, Not Frozen: The Federal Funds Market in the Financial Crisis,” The Journal of Finance, 66(4), 1109–1139. (2014): “Trading Partners in the Interbank Lending Market,” FRBNY Staff Report 620, Federal Reserve Bank of New York. Arora, N., P. Gandhi, and F. A. Longstaff(2012): “CounterpartyCreditRiskandtheCredit Default Swap Market,” Journal of Financial Economics, 103(5), 280–293. Augustin, P., M. G. Subrahmanyam, D. Y. Tang, and S. Q. Wang (2014): “Credit Default Swaps: A Survey,” Foundations and Trends in Finance, 9(1–2), 1–196. Bernhardt, D., V. Dvoracek, E. Hughson, and I. M. Werner (2005): “Why Do Large Orders Receive Discounts on the London Stock Exchange?,” Review of Financial Studies, 18(4), 1343–1368. Bernstein, A., E. Hughson, and M. Weidenmier (2014): “Counterparty Risk and the Establishment of the NYSE Clearinghouse,” Working Paper, MIT and Claremont McKenna College. Campbell, S., and E. Heitfield (2014): “The Regulation of Counterparty Risk in Over-The- Counter Derivatives Markets,” in Counterparty Risk Management, ed. by E. Canabarro, and M. Pykhtin. Risk Books, London. Chava, S., and R. Jarrow (2004): “Bankruptcy Prediction with Industry Effects,” Review of Finance, 8(4), 537–569. CommitteeonPaymentandSettlementSystems(2013): “Authorities’accesstotraderepository data,” Publication No. 110, Bank for International Settlements. Cooper, I. A., and A. S. Mello (1991): “The Default Risk of Swaps,” The Journal of Finance, 46(2), 597–620. Duffie, D. (2010): “The Failure Mechanics of Dealer Banks,” Journal of Economic Perspectives, 24(1), 51–72. Duffie, D., and M. Huang (1996): “Swap Rates and Credit Quality,” The Journal of Finance, 51(3), 921–949. Duffie, D., M. Scheicher, and G. Vuillemey (2015): “Central Clearing and Collateral Demand,” Journal of Financial Economics, 116(2), 237–256. Duffie, D., and H. Zhu (2011): “Does a Central Clearing Counterparty Reduce Counterparty Risk?,” Review of Asset Pricing Studies, 1(1), 74–95. Financial Crisis Inquiry Commission (2011): “The Financial Crisis Inquiry Report: Final Report of the National Commission on the Causes of the Financial and Economic Crisis in the United States,” Commission report, United States Government. Giglio, S. (2014): “Credit Default Swap Spreads and Systemic Financial Risk,” Working paper, University of Chicago. 24

Gregory, J. (2010): Counterparty Credit Risk: The New Challenge for Global Financial Markets. Wiley. Gündüz, Y. (2015): “Mitigating Counterparty Risk,” Working paper, Deutsche Bundesbank. Hendershott, T., D. Li, D. Livdan, and N. Schürhoff (2016): “Relationship Trading in OTC Markets,” Working paper, University of California, Berkeley. Hull, J. C., and A. D. White (2001): “Valuing Credit Default Swaps II: Modeling Default Correlations,” The Journal of Derivatives, 8(3), 12–21. Jarrow, R., and F. Yu (2001): “Counterparty Risk and the Pricing of Defaultable Securities,” The Journal of Finance, 56, 1765–1799. Loon, Y. C., and Z. K. Zhong (2014): “The Impact of Central Clearing on Counterparty Risk, Liquidity and Trading: Evidence from Credit Default Swap Market,” Journal of Financial Economics, 112(1), 91–115. McFadden, D.(1974): “TheMeasurementofUrbanTravelDemand,” JournalofPublicEconomics, pp. 303–328. Oehmke, M., and A. Zawadowski (2013): “The Anatomy of the CDS Market,” . Pykhtin, M. (2011): “Counterparty Risk Management and Valuation,” in Credit Risk Frontiers, ed. by T. Bielecki, D. Brigo, and F. Patras, chap. 16. John Wiley & Sons, Hoboken. Shachar, O. (2012): “Exposing the Exposed: Intermediation Capacity in the Credit Default Swap Market,” Working paper, NYU. Siriwardane, E. (2015): “Concentrated Capital Losses and the Pricing of Corporate Credit Risk,” Working paper, NYU. Turnbull, S. M.(2014): “CounterpartyRisk: AReview,” Annual Review of Financial Economics, 6, 241–258. 25

Figure 1: Effects of Transaction Type on log(cdss,b)−log(cds ) t t Notes: The top figure plots mean transaction spreads on transactions between dealers and non-dealers before and after clearing. The bottom figure adds mean transaction spreads on same-day cleared transactions. 26

Figure 2: Time Series Effects of Clearing Eligibility on Transaction Spreads (a) DID Specification I (b) DID Specification II (c) DID Specification III Notes: The left panels plot results based on DTCC transaction spreads and the right panels plot results based on Markit quotes. In Figure (a), the treatment group consists of transactions or quotes on reference entities cleared in the first cohort for each sector of the CDX.NA.IG index before and after clearing. The control group consists of pre-clearing transactions or quotes of reference entities in CDX.NA.IG that are cleared in later cohorts. In Figure (b), the control group consists of post-clearing reference entities cleared in the first cohort for each sector of the CDX.NA.IGindex. Thetreatmentgroupconsistsoftransactionsorquotesofreferenceentitiesinthesamesectorof CDX.NA.IG that are cleared in later cohorts. In Figure (c), the treatment group consists of transactions or quotes onallclearablereferenceentitiesinCDX.NA.IGbeforeandafterclearing. Thecontrolgroupconsistsoftransactions or quotes on all other non-clearable investment grade reference entities in North America. Only interdealer and same-day cleared transactions are used in computing the means. 27

elpmaS enilesaB sisylanA gnicirP rof scitsitatS yrammuS :1 elbaT )104,38= N( setouq tikraM dna sdaerps noitcasnart lla rof scitsitatS yrammuS )A( 99p 59p 09p 57p 05p 52p 01p 5p 1p elitnecreP 00.92 76.31 41.9 32.4 97.0 90.2- 47.6- 65.11- 75.82sdc−CCTDsdc 49.73 19.81 77.21 38.6 52.3 03.1 84.0 42.0 50.0 |sdc−CCTDsdc| 70.0 70.0 70.0 30.0 10.0 10.0- 50.0- 90.0- 81.0- )sdc(gol−)CCTDsdc(gol 42.0 41.0 01.0 50.0 20.0 10.0 00.0 00.0 00.0 |)sdc(gol−)CCTDsdc(gol| )902,32= N( yad rep sedart net tsael ta htiw stcartnoc rof scitsitatS yrammuS )B( 99p 59p 09p 57p 05p 52p 01p 5p 1p elitnecreP 0880.0 8740.0 1630.0 3220.0 3410.0 8800.0 3400.0 0000.0 0000.0 ])sdc(gol−)CCTDsdc(gol[ σ t 0.65 0.13 0.52 0.81 0.31 0.11 0.01 0.01 0.01 yad rep snoitcasnart fo .oN 4.4 3.3 9.2 4.2 9.1 5.1 2.1 0.1 0.1 reyub rep srelles fo .oN egarevA 6.4 4.3 0.3 4.2 9.1 5.1 2.1 0.1 0.1 relles rep sreyub fo .oN egarevA ehT . sdc ,setouq tikraM dna ,CCTDsdc ,sdaerps noitcasnart CCTD neewteb secnereffid troper ew ,A lenaP nI .3102 ot 0102 morf si doirep elpmas ruO :setoN t t rof noitaived dradnats fo noitubirtsid troper ew ,B lenaP nI .detroper osla era sdaerps owt eht ni secnereffid gol ehT .stniop sisab ni desserpxe era sdaerps eht troper ew ,noitidda nI .tcartnoc emas eht no yad rep sedart net tsael ta era ereht taht nevig ,yad emas eht nihtiw tcartnoc emas eht no sdaerps noitcasnart .yad emas eht no )relles( reyub emas eht rof )sreyub( srelles fo rebmun eht fo noitubirtsid 28

elpmaS ytrapretnuoC 5SU – sisylanA gnicirP rof scitsitatS yrammuS :2 elbaT )014,915,1= N( setouq tikraM dna sdaerps noitcasnart lla rof scitsitatS yrammuS )A( 99p 59p 09p 57p 05p 52p 01p 5p 1p elitnecreP 00.36 06.52 31.51 77.5 58.0 56.2- 89.8- 25.51- 43.93sdc−CCTDsdc 92.17 03.23 96.02 48.9 81.4 75.1 65.0 72.0 60.0 |sdc−CCTDsdc| 42.0 41.0 90.0 40.0 10.0 20.0- 70.0- 11.0- 12.0- )sdc(gol−)CCTDsdc(gol 62.0 71.0 21.0 60.0 30.0 10.0 00.0 00.0 00.0 |)sdc(gol−)CCTDsdc(gol| )205,281= N( yad rep sedart net tsael ta htiw stcartnoc rof scitsitatS yrammuS )B( 99p 59p 09p 57p 05p 52p 01p 5p 1p elitnecreP 3190.0 8540.0 6330.0 4020.0 1210.0 7600.0 0000.0 0000.0 0000.0 ])sdc(gol−)CCTDsdc(gol[ σ t 0.06 0.43 0.62 0.81 0.31 0.11 0.01 0.01 0.01 yad rep snoitcasnart fo .oN 4.4 3.3 8.2 2.2 7.1 2.1 0.1 0.1 0.1 reyub rep srelles fo .oN egarevA 2.5 7.3 1.3 4.2 8.1 3.1 0.1 0.1 0.1 relles rep sreyub fo .oN egarevA ehT . sdc ,setouq tikraM dna ,CCTDsdc ,sdaerps noitcasnart CCTD neewteb secnereffid troper ew ,A lenaP nI .3102 ot 0102 morf si doirep elpmas ruO :setoN t t rof noitaived dradnats fo noitubirtsid troper ew ,B lenaP nI .detroper osla era sdaerps owt eht ni secnereffid gol ehT .stniop sisab ni desserpxe era sdaerps eht troper ew ,noitidda nI .tcartnoc emas eht no yad rep sedart net tsael ta era ereht taht nevig ,yad emas eht nihtiw tcartnoc emas eht no sdaerps noitcasnart .yad emas eht no )relles( reyub emas eht rof )sreyub( srelles fo rebmun eht fo noitubirtsid 29

Table 3: Effects of Seller CDS Spreads on Log Spread Differentials (1) (2) (3) (4) Buyer Type Non-dealer Non-dealer 14 Dealers 14 Dealers (A) Log Seller CDS -0.00686** -0.00785** -0.00164* 0.000170 (0.00349) (0.00357) (0.000865) (0.000342) Wrong-Way Risk 0.00569*** 0.00589*** -0.00122** -0.00230*** (0.00211) (0.00213) (0.000561) (0.000565) Past Relationship -0.000627 0.00136 -0.00551* 0.00577*** (0.00409) (0.00431) (0.00322) (0.00202) No. of effective obs. 3,403 2,909 17,477 172,097 No. of transactions 11,948 10,311 52,750 1,012,029 Fixed effects Contract × Buyer × Date (B) Log Seller CDS -0.0176*** -0.0186*** -0.00338* -0.000758 (0.00549) (0.00642) (0.00188) (0.000757) Wrong-Way Risk 0.0124*** 0.00434 -0.00491*** -0.00271*** (0.00406) (0.00350) (0.00155) (0.000606) Past Relationship 0.000111 0.00211 -0.00646* 0.000924 (0.00425) (0.00451) (0.00341) (0.00276) No. of effective obs. 3,403 2,909 17,477 172,097 No. of transactions 11,948 10,311 52,750 1,012,029 Fixed effects Contract × Buyer × Date, Seller Reference entities Entitled Entitled Entitled Non-Entitled Clearability Unrestricted Non-Clearable Non-Clearable Non-Clearable Notes: In all columns, we consider only transactions where the seller is one of the 14 largest dealers. We use the log of the notional value of the contract as a control. In Column 1, we use the sample of transactions on all entitled reference entities in which the buyer is a non-dealer and the seller is one of the 14 largest dealers. In Column 2, we use the sample of transactions on all non-clearable entitled reference entities in which the buyer is a non-dealer and the seller is one of the 14 largest dealers. In Column 3, we use the sample of transactions on non-clearable entitled reference entities in which both the buyer and the seller are one of the 14 largest dealers. In Column 4, we use the sample of interdealer transactions on non-clearable non-entitled reference entities in which at least one of the counterparties is a US5 bank. The two panels differ by the fixed effect specifications. In Panel A, we include contract-buyer-date fixed effects. In Panel B, we include contract-buyer-date fixed effects and additional seller fixed effects. The number of effective observations is reported for buyers facing at least two different sellers for the same contract on a single trading date. Robust standard errors are in parentheses. *** p<0.01, ** p<0.05, * p<0.1. 30

Table 4: Effects of Buyer CDS Spreads on Log Spread Differentials (1) (2) (3) (4) Seller Type Non-dealer Non-dealer 14 Dealers 14 Dealers (A) Log Buyer CDS 0.000810 0.0115** -0.000700 -0.000516 (0.00265) (0.00474) (0.000838) (0.000341) Right-Way Risk 0.00158 0.00770** 0.00108* 0.000974* (0.00165) (0.00327) (0.000583) (0.000535) Past Relationship 0.00647 0.00488 -0.00726* -0.00177 (0.00504) (0.00545) (0.00400) (0.00207) No. of effective obs. 2,158 1,946 16,691 162,792 No. of transactions 7,767 6,955 52,750 1,012,029 Fixed effects Contract × Seller × Date (B) Log Buyer CDS 0.0115** 0.0109** -0.00152 0.000328 (0.00474) (0.00461) (0.00210) (0.000818) Right-Way Risk 0.00770** 0.00305 0.00201 0.000566 (0.00327) (0.00391) (0.00126) (0.000638) Past Relationship 0.00488 0.00520 -0.00716* 0.000112 (0.00545) (0.00568) (0.00400) (0.00272) No. of effective obs. 2,158 1,946 16,691 162,792 No. of transactions 7,767 6,955 52,750 1,012,029 Fixed effects Contract × Seller × Date, Buyer Reference entities Entitled Entitled Entitled Non-Entitled Clearability Unrestricted Non-Clearable Non-Clearable Non-Clearable Notes: In all columns, we consider only transactions where the buyer is one of the 14 largest dealers. We use the log of the notional value of the contract as a control. In Column 1, we use the sample of transactions on all entitled reference entities in which the seller is a non-dealer and the buyer is one of the 14 largest dealers. In Column 2, we use the sample of transactions on all non-clearable entitled reference entities in which the seller is a non-dealer and the buyer is one of the 14 largest dealers. In Column 3, we use the sample of transactions on non-clearable entitled reference entities in which both the buyer and the seller are one of the 14 largest dealers. In Column 4, we use the sample of interdealer transactions on non-clearable non-entitled reference entities in which at least one of the counterparties is a US5 bank. The two panels differ by the fixed effect specifications. In Panel A, we include contract-seller-date fixed effects. In Panel B, we include contract-seller-date fixed effects and additional buyer fixed effects. The number of effective observations is reported for sellers facing at least two different buyers for the same contract on a single trading date. Robust standard errors are in parentheses. *** p<0.01, ** p<0.05, * p<0.1. 31

sisylanA eciohC ytrapretnuoC rof noitpircseD elpmaS :5 elbaT ledoM eciohC relleS )2( ledoM eciohC reyuB )1( 21 21 seititne ecnerefer fo rebmuN 4 4 gniraelc rof elbigile seititne ecnerefer fo rebmuN 961 691 srelaed-non fo rebmuN 767,7 849,11 srelaed tsegral 41 eht fo eno dna srelaed-non neewteb snoitcasnart fo rebmuN ...hcihw fO 396,5 729,7 ytitne ecnerefer eht sa raeppa relaed 5SU A 559,6 113,01 gniraelc rof elbigile ton si ytitne ecnerefer ehT detaluger BRF na no noitcetorp gniyub si relaed-non a erehw snoitcasnart redisnoc ylno ew )1( nmuloC nI .3102 ot 0102 morf si doirep elpmas ruO :setoN .ytitne ecnerefer detaluger BRF na no noitcetorp gnilles si relaed-non a erehw snoitcasnart redisnoc ylno ew )2( nmuloC nI .ytitne ecnerefer ksiR yaW-gnorW :6 elbaT srelleS ngieroF srelleS SU serahS gnidarT 65.0 44.0 ograF slleW ro relaed 5SU a si ytitne ecnerefeR 64.0 45.0 ograF slleW ron relaed 5SU a ton si ytitne ecnerefeR ro relaed 5SU a si ytitne ecnerefer eht( ksir tiderc s’relles .S.U htiw detalerroc ylhgih era taht seititne ecnerefer rof serahs gnidart setalubat elbat sihT :setoN tcirtsereW .)ograFslleWronrelaed5SUatonsiytitneecnerefereht(ksirtidercs’relles.S.Uhtiwdetalerrocsseleratahtseititneecnereferrofdna)ograFslleW .seititne ecnerefer detaluger BRF ot seititne ecnerefer eht 32

srelaeD 41 morf noitcetorP gniyuB srelaed-noN fo eciohc ytrapretnuoC :7 elbaT )3( )2( )1( sreyub evitpac oN gniraelc rof elbigile toN elpmas enilesaB ***762.0- ***862.0- ***972.0- SDC relleS )1140.0( )5240.0( )7930.0( ***813.0- 960.0- ***712.0ksiR yaW-gnorW )1840.0( )2960.0( )3740.0( ***660.3 ***956.3 ***574.3 pihsnoitaleR tsaP )8950.0( )2450.0( )7940.0( 344,641 724,631 543,951 snoitavresbo fo rebmuN 979,01 113,01 849,11 snoitcasnart fo rebmuN 551 591 691 sreyub fo rebmuN 953.0 983.0 383.0 derauqs-R oduesP dradnats tsuboR .slortnoc sa relles hcae rof elbairav rotacidni na htiw detcaretni xedni GI.AN.XDC eht fo daerps eht dna stceffe dexfi relles edulcni eW :setoN .1.0<p * ,50.0<p ** ,10.0<p *** .sesehtnerap ni era srorre ledoM eciohC reyuB eht fo stceffE lanigraM :8 elbaT )3( )2( )1( .dts 1 a nevig ytilibaborp ni egnahc egarevA ytilibaborp ni egnahc egarevA a nevig ytilibaborp ni egnahc egarevA pihsnoitaler tsap ni egnahc .ved RWW ot RWW on morf SDC relles ni egnahc pb 001 eciohc fo ytilibaborp fo egnaR elpmas enilesaB 240.0 020.0- 620.0- 70.0-31.0 610.0 700.0- 010.0- 10.0-60.0 gniraelc rof elbigile toN 340.0 600.0- 520.0- 70.0-31.0 710.0 300.0- 010.0- 10.0-60.0 sreyub evitpac oN 430.0 920.0- 520.0- 70.0-31.0 310.0 110.0- 010.0- 10.0-60.0 .7 elbaT ni detroper stneicffieoc no desab era stceffe lanigraM :setoN 33

sreyuB noziroH trohS :srelaeD 41 morf noitcetorP gniyuB srelaed-noN fo eciohC ytrapretnuoC :9 elbaT )3( )2( )1( sreyub evitpac oN gniraelc rof elbigile toN elpmas enilesaB ***073.0- ***273.0- ***263.0- SDC relleS )6440.0( )7540.0( )3240.0( ***932.0 ***452.0 ***312.0 reyuB noziroH trohS x SDC relleS )5930.0( )7040.0( )9730.0( ***953.0- 740.0- ***502.0ksiR yaW-gnorW )0250.0( )7170.0( )5050.0( 460.0 250.0- 340.0reyuB noziroH trohS x ksiR yaW-gnorW )2840.0( )0740.0( )7640.0( ***797.2 ***265.3 ***163.3 pihsnoitaleR tsaP )0560.0( )9650.0( )3350.0( ***831.1 ***084.0 ***806.0 reyuB noziroH trohS x pihsnoitaleR tsaP )131.0( )821.0( )121.0( 344,641 724,631 543,951 snoitavresbo fo rebmuN 979,01 113,01 849,11 snoitcasnart fo rebmuN 551 591 691 sreyub fo rebmuN 163.0 983.0 483.0 derauqs-R oduesP dradnats tsuboR .slortnoc sa relles hcae rof elbairav rotacidni na htiw detcaretni xedni GI.AN.XDC eht fo daerps eht dna stceffe dexfi relles edulcni eW :setoN .1.0<p * ,50.0<p ** ,10.0<p *** .sesehtnerap ni era srorre 34

srelaeD 41 ot noitcetorP gnilleS srelaed-noN fo eciohC ytrapretnuoC :01 elbaT )3( )2( )1( srelles evitpac oN gniraelc rof elbigile toN elpmas enilesaB ***552.0- ***681.0- ***052.0- SDC reyuB )2350.0( )8550.0( )8250.0( 860.0- ***082.0 080.0ksiR yaW-thgiR )8950.0( )9570.0( )3850.0( ***971.3 ***573.3 ***293.3 pihsnoitaleR tsaP )0470.0( )6170.0( )7560.0( 733,79 776,19 540,301 snoitavresbo fo rebmuN 043,7 559,6 767,7 snoitcasnart fo rebmuN 321 561 961 srelles fo rebmuN 863.0 283.0 383.0 derauqs-R oduesP dradnatstsuboR .slortnocsareyubhcaerofelbairavrotacidninahtiwdetcaretnixedniGI.AN.XDCehtnodaerpsehtdnastceffedexfireyubedulcnieW :setoN .1.0<p * ,50.0<p ** ,10.0<p *** .sesehtnerap ni era srorre ledoM eciohC relleS eht fo stceffE lanigraM :11 elbaT )3( )2( )1( .dts 1 a nevig ytilibaborp ni egnahc egarevA ytilibaborp ni egnahc egarevA a nevig ytilibaborp ni egnahc egarevA pihsnoitaler tsap ni egnahc .ved RWR ot RWR on morf SDC reyub ni egnahc pb 001 eciohc fo ytilibaborp fo egnaR elpmas enilesaB 730.0 700.0- 220.0- 70.0-61.0 010.0 200.0- 600.0- 10.0-60.0 gniraelc rof elbigile toN 430.0 420.0 610.0- 70.0-61.0 110.0 800.0 500.0- 10.0-60.0 srelles evitpac oN 430.0 600.0- 220.0- 70.0-61.0 010.0 200.0- 600.0- 10.0-60.0 .01 elbaT ni detroper stneicffieoc no desab era stceffe lanigraM :setoN 35

srelleS noziroH trohS :srelaeD 41 ot noitcetorP gnilleS srelaed-noN fo eciohC ytrapretnuoC :21 elbaT )3( )2( )1( srelles evitpac oN gniraelc rof elbigile toN elpmas enilesaB ***662.0- ***571.0- ***462.0- SDC reyuB )7750.0( )3160.0( )2750.0( 5920.0 4910.0- 7330.0 relleS noziroH trohS x SDC reyuB )9050.0( )1450.0( )2050.0( 7080.0- ***932.0 2790.0ksiR yaW-thgiR )2660.0( )5080.0( )5460.0( 5640.0 3280.0 4850.0 relleS noziroH trohS x ksiR yaW-thgiR )7950.0( )6950.0( )8850.0( ***377.2 ***790.3 ***151.3 pihsnoitaleR tsaP )1680.0( )8970.0( )8270.0( ***482.1 ***089.0 ***698.0 relleS noziroH trohS x pihsnoitaleR tsaP )541.0( )441.0( )631.0( 733,79 776,19 540,301 snoitavresbo fo rebmuN 043,7 559,6 767,7 snoitcasnart fo rebmuN 321 561 961 srelles fo rebmuN 073.0 383.0 483.0 derauqs-R oduesP dradnatstsuboR .slortnocsareyubhcaerofelbairavrotacidninahtiwdetcaretnixedniGI.AN.XDCehtnodaerpsehtdnastceffedexfireyubedulcnieW :setoN .1.0<p * ,50.0<p ** ,10.0<p *** .sesehtnerap ni era srorre 36

stceffE dexiF htiw ledoM ytilibaborP raeniL :31 elbaT srelaeD 41 morf noitcetorP gniyuB srelaed-noN fo eciohC ytrapretnuoC )A( )3( )2( )1( sreyub evitpac oN gniraelc rof elbigile toN elpmas enilesaB ***9210.0- ***3110.0- ***4110.0- SDC relleS )93300.0( )13300.0( )22300.0( ***0120.0- 672000.0- 0110.0ksiR yaW-gnorW )31700.0( )80800.0( )39600.0( ***724.0 ***275.0 ***635.0 pihsnoitaleR tsaP )3120.0( )4710.0( )7810.0( 344,641 724,631 543,951 snoitavresbo fo rebmuN 979,01 113,01 849,11 snoitcasnart fo rebmuN 551 591 691 sreyub fo rebmuN 060.0 590.0 880.0 derauqs-R srelaeD 41 ot noitcetorP gnilleS srelaed-noN fo eciohC ytrapretnuoC )B( )3( )2( )1( srelles evitpac oN gniraelc rof elbigile toN elpmas enilesaB ***6210.0- ***1010.0- ***2110.0- SDC reyuB )77300.0( )87300.0( )76300.0( 76300.0- *1020.0 71300.0ksiR yaW-thgiR )9010.0( )3110.0( )1010.0( ***934.0 ***005.0 ***405.0 pihsnoitaleR tsaP )3520.0( )4120.0( )0320.0( 053,79 096,19 850,301 snoitavresbo fo rebmuN 043,7 559,6 767,7 snoitcasnart fo rebmuN 321 561 961 srelles fo rebmuN 660.0 180.0 380.0 derauqs-R dexfi relaed-etad-tcartnoc dna ,relaed hcae rof elbairav rotacidni na htiw detcaretni xedni GI.AN.XDC eht no daerps eht ,stceffe dexfi relaed edulcni eW :setoN .1.0<p * ,50.0<p ** ,10.0<p *** .sesehtnerap ni era srorre dradnats tsuboR .slortnoc sa stceffe 37

stneutitsnoC GI.AN.XDC rof setaD gniraelC :41 elbaT seitilitU moceleT ygolonhceT saG/liO slairtsudnI erachtlaeH slaicnaniF svrS .noC sdooG .noC slairetaM 3 0 0 0 0 0 0 0 0 0 9002/12/21 3 0 0 0 0 0 0 0 0 0 0102/11/1 0 3 0 0 0 0 0 0 0 0 0102/1/2 0 0 0 0 51 0 0 0 0 0 0102/51/2 0 0 0 5 0 0 0 0 4 0 0102/8/3 0 0 0 0 0 0 0 41 0 0 0102/92/3 0 0 5 0 0 0 0 0 0 5 0102/91/4 0 0 0 0 0 4 6 1 0 0 0102/01/5 0 0 0 0 0 1 1 0 0 0 0102/7/6 0 0 0 0 0 0 0 1 0 0 0102/82/6 0 0 0 2 0 0 0 0 8 0 0102/9/8 0 0 0 0 0 0 0 8 0 0 0102/03/8 0 0 2 0 0 0 0 4 2 2 1102/82/3 0 0 0 0 0 0 01 0 0 0 1102/11/4 0 0 0 2 3 3 0 0 0 0 1102/2/5 0 0 2 0 0 0 5 3 1 0 1102/31/6 0 0 0 1 0 1 0 1 1 0 1102/41/11 0 0 0 0 0 0 5 0 0 0 2102/9/01 0 0 0 6 0 0 0 0 0 0 2102/22/01 2 0 0 0 1 0 1 1 2 1 2102/5/11 0 0 0 0 0 0 1 0 0 0 2102/91/11 0 0 0 0 2 0 2 2 2 0 3102/03/9 8 3 9 61 12 9 13 53 02 8 latoT egnahcxE latnenitnocretnI ta gniraelc lartnec rof elbigile emoceb taht GI.AN.XDC ni seititne ecnerefer rof rotces yb setad gniraelc stroper elbat sihT :setoN .ECI :ecruoS .srossecederp deraelc a dah taht seititne ecnerefer elbaraelc edulcxe eW .)ECI( 38

Table 15: Effects of Clearing and Counterparty Characteristics on Log Spread Differentials (1) (2) (3) Seller Buyer Pair Seller CCP -0.00363*** -0.00332*** (0.000871) (0.000437) Buyer CCP -0.00217** -0.00290*** (0.000891) (0.000440) Backload clear -0.00128* -0.000462 -0.00160*** (0.000734) (0.000732) (0.000380) Seller non-dealer -0.00555*** -0.00939*** (0.000939) (0.000533) Buyer non-dealer 0.00402*** 0.00896*** (0.000810) (0.000426) Number of observations 368,402 409,053 509,248 Fixed effects Contract × Date × Buyer Contract × Date × Seller Contract × Date Notes: This table shows effects of counterparty characteristics and clearing on transaction spreads. In Column 1 we holdtimeandthebuyerfixedandestimateeffectsofseller’scharacteristics. Thebuyerisoneofthe14largestdealers. InColumn2weholdtimeandthesellerfixedandestimateeffectsofbuyer’scharacteristics. Thesellerisoneofthe 14largestdealers. InColumns3weholdtimefixedandjointlyestimateeffectsofbuyer’sandseller’scharacteristics. Eitherthebuyerofthesellerisoneofthe14largestdealers. Robuststandarderrorsareinparentheses. ***p<0.01, ** p<0.05, * p<0.1. 39

ytilibigilE gniraelC fo stceffE seireS emiT :61 elbaT III noitacfiicepS II noitacfiicepS I noitacfiicepS )6( )5( )4( )3( )2( )1( tikraM CCTD tikraM CCTD tikraM CCTD *1910.0 7840.0 63500.0 0770.0- 3920.0- 1410.0retfA X tnemtaerT )A( )3110.0( )1930.0( )5020.0( )4450.0( )1910.0( )5850.0( 622000.0 ***644.0- ***251.0 **1490.0 ***1070.0- **011.0tnemtaerT )31800.0( )6030.0( )1510.0( )5240.0( )7110.0( )4840.0( 404,039 220,132 193,47 682,921 586,27 196,44 snoitavresbO 12800.0 ***512.0- ***661.0- **851.0- 6810.0 0910.0- )01- ,02-[ X tnemtaerT )B( )7710.0( )1250.0( )4430.0( )9960.0( )1520.0( )2990.0( 66600.0 511.0 858000.0 3380.0- 971000.0- 1710.0- )0 ,01-[ X tnemtaerT )6320.0( )7770.0( )5740.0( )201.0( )3630.0( )2590.0( 06800.0 711.0 91100.0 611.0- 21100.0- 0830.0 )01 ,0[ X tnemtaerT )8320.0( )8870.0( )9740.0( )011.0( )8930.0( )111.0( 3220.0 ***922.0- ***951.0- ***644.0- 4810.0 6840.0- )02 ,01[ X tnemtaerT )0810.0( )5940.0( )7530.0( )411.0( )2430.0( )9960.0( ***3020.0- ***592.0- ***762.0 ***372.0 ***0780.0- ***821.0tnemtaerT )66400.0( )9810.0( )04700.0( )3020.0( )30600.0( )8840.0( 394,475 795,942 054,68 958,541 863,28 503,04 snoitavresbO etaD-rotceS etaD-rotceS etaD-rotceS etaD-rotceS etaD-rotceS etaD-rotceS EF dexiF .setouqtikraMdnasdaerpsnoitcasnartCCTDnoytilibigilegniraelcfostceffeseiresemitehtrofsnoitacfiicepsDIDeerhtrofstluserswohselbatsihT :setoN noitcudortni eht retfa dna erofeb syad 05 morf wodniw tneve eht sesu A lenaP .snoitacfiiceps lla ni desu era setad edart dna rotces fo noitcaretni eht rof stceffe tnemtaert eht ,I noitacfiicepS nI .sdrawretfa syad 02 ot ytilibigile gniraelc fo noitcudortni eht erofeb syad 052 morf tneve eht sesu B lenaP .ytilibigile gniraelc fo ehT .gniraelc retfa dna erofeb xedni GI.AN.XDC eht fo rotces hcae rof trohoc tsrfi eht ni deraelc seititne ecnerefer no setouq ro snoitcasnart fo stsisnoc puorg lortnoc eht ,II noitacfiicepS nI .strohoc retal ni deraelc era taht GI.AN.XDC ni seititne ecnerefer fo setouq ro snoitcasnart gniraelc-erp fo stsisnoc puorg lortnoc snoitcasnartfostsisnocpuorgtnemtaertehT .xedniGI.AN.XDCehtforotceshcaeroftrohoctsrfiehtnideraelcseititneecnerefergniraelc-tsopfostsisnocpuorg snoitcasnartfostsisnocpuorgtnemtaerteht,IIInoitacfiicepSnI .strohocretalnideraelceratahtGI.AN.XDCforotcesemasehtniseititneecnereferfosetouqro elbaraelc-nonrehtollanosetouqrosnoitcasnartfostsisnocpuorglortnocehT .gniraelcretfadnaerofebGI.AN.XDCniseititneecnereferelbaraelcllanosetouqro .1.0<p * ,50.0<p ** ,10.0<p *** .sesehtnerap ni era srorre dradnats tsuboR .aciremA htroN ni seititne ecnerefer edarg tnemtsevni 40