Information Friction in OTC Interdealer Markets

Finance and Economics Discussion Series Federal Reserve Board, Washington, D.C. ISSN 1936-2854 (Print) ISSN 2767-3898 (Online) Information Friction in OTC Interdealer Markets Benjamin Gardner and Yesol Huh 2024-040 Please cite this paper as: Gardner, Benjamin, and Yesol Huh (2024). “Information Friction in OTC Interdealer Markets,”FinanceandEconomicsDiscussionSeries2024-040. Washington: BoardofGovernors of the Federal Reserve System, https://doi.org/10.17016/FEDS.2024.040. 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.

Information Friction in OTC Interdealer Markets Benjamin Gardner and Yesol Huh∗ November 2, 2023 Abstract In over-the-counter (OTC) securities markets, interdealer markets are an important venue through which dealers can offload positions and share risk amongst themselves. Contrary to the popular conception that search frictions matter the most in OTC markets, we find that in the interdealer market for U.S. corporate bonds, information frictions are most relevant. Large dealers face large and informed customers and pay more than small dealers to transact in the interdealer market, despite on average providing liquidity to other dealers. Large dealers tend to trade through interdealer brokers (IDBs) to mitigate information leakage, but interdealer markets are still far from efficient. ∗BenjaminGardner: YaleUniversity, ben.gardner@yale.edu; YesolHuh: FederalReserveBoard, yesol.huh@frb.gov. Please direct comments and questions to yesol.huh@frb.gov. The analysis and conclusions set forth are those of the authors and do notindicateconcurrencebyothermembersofthestaff,bytheBoardofGovernors,orbytheFederalReserveSystem. 1

1 Introduction In over-the-counter (OTC) securities markets, interdealer markets are an important venue through which dealers can offload positions and share risk with one other. Dealers intermediate customer order flow and offload some of that order flow through the interdealer market. Much of the literature on OTC markets has focusedonsearchfrictionsandnetworkformationtoexplainpriceandtradingdynamics(Duffieetal.,2005; Lagos and Rocheteau, 2009; Wang, 2016). Moreover, the empirical literature on OTC interdealer markets has emphasized the core-periphery network in this market as an imperfect mechanism to mitigate search costs. In this paper, we study which frictions are most relevant in the OTC interdealer markets. We find that in the U.S. corporate bond market, contrary to popular conception, information frictions play a large role. Largedealersfacelargeandinformedcustomers,sotheyhaveamoredifficulttimeoffloadingcustomerorder in the interdealer market—large dealers pay more to transact in the interdealer market despite the fact that they on average provide liquidity in this market. Large dealers tend to trade through interdealer brokers (IDBs) to mitigate information leakage, but interdealer markets are still far from efficient. Wefirstdividedealersintosixcategories. Wegenerallythinkofdealersasengaginginsimilaractivities— intermediating customer order flow and offloading some of those order flow in the interdealer market—and differing mostly along the dimensions of size, search costs, or their position in the network. However, three categories—alternative trading systems (ATS), interdealer brokers (IDBs), and client brokers—are “special” types in that these types of dealers are somewhat different from the usual set of dealers that are typically discussed in the literature. ATS and IDBs predominantly engage in interdealer trades only. ATS are designated by FINRA and are mostly different types of electronic platforms.1 IDBs are brokers that match buyers with sellers in the interdealer market and account for 25% of interdealer volume. Client brokers mostly act as agents between customers and other dealers. The other three dealer categories are the “typical” dealers, which we divide into small, medium, and large by customer volume. Existing literature has mostly compared central dealers and peripheral dealers. Not surprisingly, large dealersaremorecentral,andsmalldealersareperipheral. Thus,some“centrality”effectsstudiedintheprior literature such as whether central dealers charge higher bid-ask spreads (Li and Schu¨rhoff, 2019; Hollifield et al., 2016; Di Maggio et al., 2017; Dick-Nielsen et al., 2020) may be driven by dealer balance sheet size or customervolume. Moreover,ourresultsindicatethatsomeoftheIDBsaccountforalargeshareofinterdealer volume and are quite central in the interdealer network. However, these IDBs behave quite differently from 1Seehttps://www.finra.org/filing-reporting/otc-transparency/finra-equity-ats-firms-listforthelistofATS. 2

large dealers. Thus, some of the centrality effects may be conflating IDBs and large dealers. For instance, the finding that central dealers charge higher bid-ask spreads in the interdealer market (Di Maggio et al., 2017; Dick-Nielsen et al., 2020) may be because IDBs charge higher bid-ask spreads and not because large dealers charge higher bid-ask spreads. We study who provides liquidity to whom and at what prices different categories of dealers trade in the interdealer market. Consistent with Li and Schu¨rhoff (2019), we find that large dealers tend to provide liquidity to smaller dealers. However, we also find that despite providing liquidity on average, large dealers actually pay a higher trading cost in the interdealer market compared to medium and small dealers. Additionally, interdealer trading costs have a U-shape in which the largest and the smallest dealers pay higher trading costs than medium sized dealers. We conjecture that for larger dealers (top 30-40 dealers), information asymmetry matter more, and as we go to smaller dealers, search frictions matter more. If so, given that top 30 dealers account for much more of the interdealer volume than small dealers, information asymmetry is the more dominant friction in the U.S. corporate bond interdealer market. Consistent with this conjecture, we find that large dealers absorb significantly more informed customer order flow. Thus, when they try to offload those order flow in the interdealer market, others would be reluctant to trade with them. Therefore, large dealers offload less and face higher trading costs in the interdealer market.2 Moreover,largedealersaremorelikelytotradethroughIDBsthansmallerdealersare. Ifsearchfrictions weremoreimportant,wewouldexpectsmalldealerstoutilizeIDBsthemostsincesmalldealershavehighest searchcosts. Ontheotherhand,informationasymmetrycanleadlargedealerstouseIDBs. Bilateralcontact can lead to information leakage even if a trade does not ultimately happen, because trading intent and identity are revealed to the counterparty. If this information leakage is costly, we would expect large dealers to trade through IDBs to keep their identity hidden and minimize information leakage. Furthermore, Glode and Opp (2016) argue that intermediation chains can help mitigate information asymmetry. If information asymmetry between the potential buyer and seller is high, trade may not happen despite the potential gains from trade. Trading through a moderately-informed intermediary can allow the trade to happen, leading to betterallocations. Consistentwiththesechannels,largedealerstendtooffloadlargerpositionsthroughIDBs andsmallerpositionsbilaterally. Moreover, whenlargedealerstradewithIDBs, theirultimatecounterparty is usually other large dealers that they already have a trading relationship with. Therefore, IDBs mostly 2Analternativebutnotmutuallyexclusiveexplanationforwhylargedealersoffloadlessisthatlargerdealersreceivemore customer order flow and potentially can more easily find offloading interest from customers, which will result in higher profit thanoffloadingintheinterdealermarket(U¨slu¨,2019). Whilethischannellikelyplaysarole,itcannotexplainwhylargedealers payhighertradingcostsintheinterdealermarket. 3

serve to mitigate information frictions rather than search frictions. Lastly, wemeasureinterdealermarketefficiency. Ifinformationfrictionsareimportantenough, potential gainsfromtradebetweenlargedealersmaybeforgone. Wefocusoncaseswithclearestpotentialgainsfrom trade, where one dealer had positive customer order flow and another dealer had negative customer order flowinthesamebondonthesameday. Wethentrackwhetherthetwodealerstradewitheachothertooffset theirpositionsinsubsequentdays, eitherdirectlythroughabilateraltradeorthroughachainoftrades. We find that such gains from trades are realized only less than 5% of the time through direct bilateral trading between the two dealers, and up to 23% of the time through a chain of trades, usually involving IDBs. Therefore, IDBs help mitigate information friction, but interdealer markets are still relatively inefficient. Overall, our results have a few implications on the effect of transparency and the search literature. Since information is contained in customer order flow, disseminating information about customer trades immediately would allow large dealers to more easily offload in interdealer markets but but make it harder for them to profit from the information. This is consistent with the results of Lewis and Schwert (2021). Moreover, our results indicate that even with post-trade transparency, the interdealer market, especially between large dealers, is inefficient. This implies that for large dealers, the risk of information leakage and information asymmetry are large compared to their inventory cost. OuranalysesalsohaveimplicationsfortheOTCsearchliterature. First,smalldealersandclientbrokers havetheroleofaggregatingandpassingonsmalluninformedcustomerorderflowtolargerdealers,andthese small dealers get somewhat higher but decent price because of two opposite frictions—higher search costs and lower information asymmetry. They offload a large share of their customer order flow in the interdealer market within a day, which is more consistent with active offloading than a search framework. Second, the length of intermediation chains is often used as a measure of the degree of search friction (Friewald and Nagler, 2019), but our results indicate that longer intermediation chains likely involve IDBs and may be driven by information frictions rather than search frictions. Li and Schu¨rhoff (2019) and Hollifield et al. (2016) document that dealers form trading networks with a core-periphery structure in OTC markets to mitigate search frictions.3 Subsequent papers have focused on the core-periphery structure of the interdealer segment of OTC markets and on whether customers pay a higher bid-ask spread to central dealers (“centrality premium”) or to peripheral dealers (“centrality discount”). In the municipal bond market, Li and Schu¨rhoff (2019) show that core dealers provide liquidity and immediacy to both customers and peripheral dealers and that there is a centrality premium. Di Maggio 3Hendershott et al. (2020b) document the importance of clients establishing trading relationships with dealers to mitigate searchfrictions. 4

etal.(2017)andDick-Nielsenetal.(2020)documentacentralitypremiuminthecorporatebondinterdealer market. Hollifield et al. (2016) show that there is a centrality discount in securitization markets. We add to this literature by showing that in the interdealer segment of OTC markets, information frictions matter greatly, and within large and medium dealers, more than search frictions. Given that top 30 dealers account for almost 90% of customer volume and that information frictions matter more for these dealers, decreasing information frictions would improve market efficiency more than decreasing interdealer marketsearchfrictions. Forthesmallretailtraderthattradeswithasmallperipheraldealer,searchfrictions mattermore. Thus,overall,thereisaU-shapepatterninthedegreeoffrictions,whichismissedbyprevious literature because they usually assume a linear effect on centrality (Dick-Nielsen et al., 2020). Also, because most papers have focused on completed intermediation chains, they do not look at the degree to and the speed of which various dealer types offload their customer order flows, and we fill that gap. We also show that there are in effect two types of dealers with high centrality—large traditional dealers and IDBs. These two types of dealers behave very differently, and simply considering a centrality dimension andputtingtheminthesamecategorymayleadtomisleadingconclusions. IDBsandtheroletheyplayhave not been studied much despite the large share of volume that they account for. An exception is De Roure et al. (2019), which document the extensive use of IDBs in the German sovereign bond interdealer market market. Their focus is on venue choice (exchange, bilateral, IDB) and argue that use of IDBs is driven by dealers’ desire to preserve an informational advantage and avoid front running. We document a similar extensive use of IDBs in the U.S. corporate bond market and show how that impacts network measures and risk sharing. A number of papers show that there is informed trading in the corporate bond market around default (Han and Zhou, 2014), acquisitions (Kedia and Zhou, 2014), and earnings announcements (Wei and Zhou, 2016). Hendershott et al. (2020a) show that short-sellers in the corporate bond market are informed. The focus in these papers are mostly to show the existence of informed trading and that customer order flow can predict future returns. Pinter et al. (2022) and Czech and Pint´er (2022) show that information asymmetry affects customer trading costs and dealer-customer connections. These papers mostly focus on the dealercustomer market and do not study the impact of informed trading in the interdealer market. Babus and Kondor (2018) models information percolation in an interdealer network, where dealers learn about their counterparties’ private information by trading. They find that in general, central dealers pay lower trading costs because their counterparties tend to be more connected. We show that dealers’ information primarily comes from their customer orders rather than through their trading relationships with other dealers. 5

2 Data We use the regulatory TRACE data for the sample period of August 2016 through July 2019. We apply standardcleaningsuchascleaningfortradecancellationsandcorrectionsanddeletetradeswithnon-FINRA affiliates. Becauseourfocusisoninterdealertrades,wekeepbothsidesofinterdealertradesaswellasadding the other side of trade for interdealer trades that are reported only once such as two-sided locked-in trades. We delete convertible bonds, MTNs, and 144A bonds as well as trades that happen in the first 30 days of issuance. Bond characteristics are from FISD Mergent. Similar to Choi et al. (2023), we aggregate the dealer identifiers (MPIDs) up to a high holder level because some dealers have multiple MPIDs or shift use of MPIDs over time. We also delete trades between MPIDs of the same high holder. We keep trades that are reported as principal trades only. Our end data has 11.8 million dealer-customer trades and 18.9 million interdealertradeobservations,withmostinterdealertradeappearingtwice,spanning11,510cusipsand1,069 dealers. WealsousetheFixedIncomeDataFeedfromICEDataPricing&ReferenceDatatocalculateinformation asymmetry in Section 3.2. The Fixed Income Data Feed contains end-of-day daily prices for most TRACE bonds over the sample period.4 3 Dealer Types and Information Asymmetry 3.1 Dealer classification We classify the dealers into six types—ATS, interdealer brokers (IDBs), client brokers, small, medium, and large. For each dealer with more than 2000 trades over the sample period, we calculate the share of the dealer’s trades, separately in terms of trade count and volume, that are interdealer trades. Also, for each dealer, we calculate the share of prearranged trades by volume and count.5 We then classify the dealers in the following way. • “ATS”: Of the dealers that are identified as ATS by FINRA, those that have more than 75% of their trades in interdealer trades by both volume and trade count basis or more than 90% of their trades by either volume or trade count basis 4Prices are “evaluated prices” by the data vendor (Intercontinental Exchange), which to our best of our knowledge, are calculatedfromdealerquotes,tradedprices,andmatrixpricingmodel. 5Prearrangedtradesareidentifiedastradesthatremaininthedealers’inventoryforlessthan15minutes,andtheconstruction followsChoietal.(2023). 6

• “Interdealerbrokers”(IDBs): Alldealersthathavemorethan75%oftheirtradesininterdealertrades by both volume and trade count basis or more than 90% of their trades by either volume or trade count basis that are not classified as ATS • “Client brokers” (CBs): Dealers that are not ATS and IDBs, and also have either prearranged share above 75% in both volume and trade count basis or above 90% in either volume or trade count basis • “Small,” “Medium,” and “Large”: Take the remaining dealers. For each year (Aug-Jul year), the top 10 by customer volume are classified as “large,” next 20 are classified as “medium,” rest are “small” Table1providessummarystatisticsondealergroupclassification. Panel(a)reportstheshareofcustomer volumeandtheshareofinterdealervolumethateachdealertypeisinvolvedin. Asshowninpreviouspapers, customer trades are concentrated, where the ten largest dealer account for almost 70% of customer volume, and the next 20 dealers (medium dealers) account for another 20%. There are a large number of small dealers that account for fairly little customer volume. This table also shows that there are a number of dealers that account for very little customer volume but a fairly large amount of interdealer volume. IDBs together account for more than 25% of interdealer volume, and ATS account for 8.6%, but both account for less than 1% of customer volume. Lastly, there are a large number of client brokers, which mostly act as an agent between customers and dealers. Panel (b) shows the share of trades that are DC-DC, DC-ID, ID-ID, or invt>15min trades. These classifications are from Choi et al. (2023). DC-DC trades are dealer-customer trade offloaded through another dealer-customer trade within 15 minutes, that is, the dealer prearranged offsetting customer trades. DC-ID trades are instances in which customer trades are prearranged with offsetting interdealer trades.6 Similarly, ID-ID trades are instances of prearranged offsetting interdealer trades. Lastly, invt>15min trades aretradestakenintodealers’inventories. ResultsinPanel(b)indicatethatIDBs,whicharenotrestrictedto having a high prearranged share, still prearrange almost 80% of their interdealer trades. Thus, these dealers mostly act as brokers between different dealers in interdealer trades rather than absorbing inventory, hence we named them “interdealer brokers.” ATS, by definition are platforms that dealers trade on, and thus are mostly ID-ID trades, and client brokers, by construction, contain a high share of DC-ID trades. The more “traditional” dealers take larger share of trades into inventory, but this share also varies with dealer size. Large dealers, compared to medium and small dealers, are more likely to take both customer trades and interdealertradesintoinventoryandtherebyprovideimmediacy. Thisresultondealersizeisconsistentwith 6BoththecustomertradesandtheinterdealertradesinthesepairsarereferredtoasDC-IDtrades. 7

Table 1: Summary statistics by dealer type: Panel (a) presents for each dealer type, the average number of dealers per year, share of interdealer trades in which the dealer type is a party to, share of dealer-customer trades in which the dealer type is a party to, and the share of dealer type’s trades that are dealer-customer trades. Panel (b) shows for each dealer type, the share of interdealer or dealer-customer trade volume that are DC-DC, DC-ID, ID-ID, or invt >15min trades. Trade type classifications are from Choi et al. (2023). In Panel (c), we present the centrality measures calculated from interdealer trades. deg, ev, and cl are degree centrality, eignenvector centrality, and closeless measures, respectively. deg vols and ev vols are degree centrality and eigenvector centrality using interdealer volume weights. We first calculate each centrality measure at the dealer-year level and present the average centrality measures, weighted by interdealervolume, foreachdealertype. Panel(d)presentssummarystatisticsonwhotradeswithwhomin the interdealer market. For each dealer type in each row, we present the share of their trade volumes with each counterparty types. (a) Dealer group summary stats dealer type # of dealers % of total ID volume % of total DC volume share DC large 10 32.01% 69.57% 81.65% medium 20 13.72% 19.42% 74.35% small 243.3 7.02% 4.22% 55.15% ATS 10 8.57% 0.35% 7.70% IDB 40 25.58% 0.66% 5.01% client broker 545 13.11% 5.79% 47.50% (b) Trade type by dealer group interdealer dealer-customer dealer type DC-ID ID-ID invt>15min DC-DC DC-ID invt>15min large 8.40% 0.67% 90.93% 12.03% 1.70% 86.27% medium 9.05% 1.70% 89.25% 15.51% 2.98% 81.51% small 20.15% 11.46% 68.39% 18.03% 16.20% 65.77% ATS 6.63% 91.74% 1.63% 14.61% 81.85% 3.54% IDB 2.15% 76.80% 21.05% 7.91% 41.44% 50.64% client broker 45.34% 39.16% 15.50% 26.63% 50.21% 23.15% (c) Dealer group centrality dealer type deg deg vols ev ev vols cl large 288.563 6.846 0.869 0.469 0.568 medium 225.482 1.848 0.76 0.105 0.542 small 160.911 0.463 0.581 0.026 0.508 ATS 110.383 2.376 0.423 0.114 0.482 IDB 120.198 7.556 0.499 0.503 0.493 client broker 135.6 5.655 0.511 0.258 0.492 8

(d) Who trades with whom: dealer type large medium small ATS IDB client broker large 9.18% 6.15% 6.08% 12.89% 49.11% 16.59% medium 16.68% 7.04% 7.33% 12.49% 32.50% 23.96% small 29.37% 14.48% 7.81% 8.59% 19.12% 20.63% ATS 48.95% 22.73% 7.53% 8.31% 12.48% IDB 61.53% 19.58% 5.70% 2.78% 1.85% 8.57% client broker 33.67% 24.70% 11.17% 7.25% 15.88% 7.32% Li and Schu¨rhoff (2019). Panel(c)presentstheaveragecentralitymeasuresforeachdealergroups. Manypapers(LiandSchu¨rhoff, 2019; Hollifieldetal.,2016)havedocumentedacore-peripherystructureinOTCinterdealermarkets. Looking at large, medium, and small dealer groups, dealers that are more central in the interdealer market also have more customer trades. It is also notable that IDBs have the highest centrality when volume-weighted centrality measures are used. Most of the literature misses ATS and IDBs that stand to intermediate between dealers. Because IDBs are central, papers that group dealers by centrality measures may group IDBs together with large dealers, which may confound the behavior of these two very different groups of dealers. Lastly, Panel (d) looks at who trades with whom in the interdealer market. Large dealers trade almost half of their interdealer volume with IDBs, which is quite surprising. If IDBs’ main function was to ease search frictions, smaller dealers should utilize IDBs significantly more than large dealers do. However, we find the exact opposite—large, medium, and small dealers trade about 49.1%, 32.5%, and 19.1% of their interdealer volume through IDBs, respectively. 3.2 Information Asymmetry In this subsection, we show that large dealers face the highest information asymmetry from their customers. We first calculate information asymmetry that each dealer faces from their customers at the dealer, year, and rating group (investment grade or high yield) level in the following way. If the dealer received order flow of v from customers for bond i on day t (positive v means that customers bought from the dealer, i,t i,t negative v means that customers sold to the dealer): i,t (cid:80) (cid:80) r |v | r |v | InfoAsym = vi,t>0 i,[t,t+τ] i,t − vi,t<0 i,[t,t+τ] i,t (1) (cid:80) (cid:80) |v | |v | vi,t>0 i,t vi,t<0 i,t 9

where r is the market-adjusted return of bond i between end of day t and t+τ where τ =5. Dealer i,[t,t+τ] and year subscripts are omitted in the equation. We get end-of-day bond prices from the Fixed Income Data Feed. To calculate market-adjusted return, we divide bonds into portfolios by rating (AAAs, AA+ through AA-, A+ through A-, BBB+ through BBB-, BB+ through BB-, B+ through B-, CCC and lower) and time-to-maturity. We then subtract the portfolio return from bond i return. We restrict the sample to bonds in which the price data (from Fixed Income Data Feed) is not stale by deleting bonds in which prices remain exactly same in consecutive days for more than 10% of the sample. This measure captures how much the prices move against the dealer within τ days after the customer trade. Because this measure doesn’t take into account the actual traded price, and therefore the bid-ask spread charged to the customer, a positive measure does not imply that the dealer loses money on the customertrade. Itrathersaysthatifthedealertradedwithacustomerondayt,thepricewillmoveagainst the dealer between end of day t and day t+τ. We do not calculate InfoAsym for ATS and interdealer brokers because these dealers do very little customer trades. Table 2 presents the summary statistics for InfoAsym by dealer group. Large dealers face highest information asymmetry—on average, after a customer buys a investment grade bond from a large dealer, market-adjusted prices increase by 10.9 bps, compared with after a customer sell. This number more than double for high-yield bonds, which also supports the idea that InfoAsym measures information asymmetry. For medium dealers, average InfoAsym is 6.9 bps and 11.3 bps for investment grade and high yield, respectively, and InfoAsym is much lower for small dealers and client brokers. Table 2: Information asymmetry summary stat: The table below presents the mean and median InfoAsym measure for each dealer group (excluding ATS and ID broker). investment grade high yield dealer type mean median mean median large 10.926 10.776 23.323 24.163 medium 6.882 6.655 11.265 14.059 small 0.99 0.188 0.231 1.486 client broker 0.001 0.353 4.957 3.68 We look at how information asymmetry varies with dealer size and ratings more formally by running the following regression: (cid:88) InfoAsym =α+ β 1[j in dlr group g]+ϵ (2) j,k,y g j,k,y g where InfoAsym is the information asymmetry measure for dealer j, ratings group k (either IG or j,k,y 10

HY), year y. We run the regression separately for investment grade bonds and high-yield bonds. Table 3 presents the results. Large dealers face the most informed customer order flow. For investment grade bonds, their information asymmetry is 4 bps higher than medium dealers and 10–11 bps higher than small dealers and client brokers, and these differences are statistically significant. Moreover, such difference in informationasymmetryislargerinhigh-yieldbonds;thedifferencewithmediumdealersis12bpsand18–23 bps withsmall dealers and client brokers. Table 3: Information asymmetry regression: The following table presents the results from regression (2). IG HY (1) (2) (3) (4) medium −4.043∗∗∗ −4.060∗∗∗ −12.058∗∗∗ −12.114∗∗∗ (0.736) (0.754) (3.438) (3.581) small −9.935∗∗∗ −9.943∗∗∗ −23.091∗∗∗ −23.012∗∗∗ (1.543) (1.546) (3.999) (4.028) client broker −10.925∗∗∗ −10.904∗∗∗ −18.366∗∗∗ −18.049∗∗∗ (1.112) (1.121) (3.499) (3.562) Constant 10.926∗∗∗ 11.151∗∗∗ 23.323∗∗∗ 26.573∗∗∗ (0.404) (1.254) (1.584) (3.708) Year f.e. No Yes No Yes Observations 1,571 1,571 980 980 Adjusted R2 0.002 0.001 0.001 0.004 Note: ∗p<0.1; ∗∗p<0.05; ∗∗∗p<0.01 4 Interdealer market Having established that larger dealers face more informed order flow, we show that information asymmetry affect trading and prices in the interdealer market. We provide further evidence of information asymmetry by looking more closely at interdealer brokers. Lastly, we measure the degree of efficiency in the interdealer market. 11

4.1 Who offloads to who? The main function of interdealer markets is for dealers to share risks and offload inventories that stem from customer order flow with other dealers (Ho and Stoll, 1983; Hansch et al., 1998; Viswanathan and Wang, 2004). Li and Schu¨rhoff (2019), using the municipal market data and creating intermediation chains, find that more central dealers provide liquidity to peripheral dealers. Central dealers providing more liquidity to peripheral dealers is consistent with multiple channels. First, usingasearchmodel,U¨slu¨(2019)showsthatcentraldealershaveloweraversiontoholdinginventorybecause theyhavemoreoffloadingopportunities,andthusendogenouslyariseasintermediationproviders. Relatedly, since large dealers have more customers (by construction) and dealers make higher profit by offloading to customers (Di Maggio et al., 2017), it may be optimal for large dealers to maximize profit by offloading less in the interdealer market, absorbing small dealers’ flows, and offloading to customers. Second, under information frictions, when large dealers try to offload, other dealers would be hesitant to take the other side, but would be happy to do so when small dealers try to offload. Last, large dealers may be associated with large dealer banks and have lower funding costs, which can also lead to the same result. In this section, we look at what share of customer order flow each dealer group offloads through the interdealer market, and to which dealer types that they offload to over what horizon. To the best of our knowledge, we are the first to quantify the degree of inventory offloading in the OTC interdealer market for different types of dealers. Because previous papers have focused on completed intermediation chains, the degree and the speed of risk offloading have not been measured directly. Our results overall, in addition to confirming a few other results that have already been shown, adds the following. First, smallest dealers offload a fairly large amount of customer order flow on the same day to other dealers, pointing to active inventory management and relative ease of offloading, which is inconsistent with low search intensity. One would need a fairly high search intensity for small dealers to get a large amount of same-day offloading. Second, largedealerstradewitheachotherthroughIDBs, whichisconsistentwithinformationfrictionsand inconsistent with search frictions. We first test, at the dealer type level, which dealer types on aggregate offload customer order flow to which dealer types. To do so, we run the following regression: (cid:88) IDG =0+ α DCG +ϵ , (3) i,g,t h i,h,t i,h h∈S where g and h denote dealer groups. DCG is the aggregate signed dealer-customer trade volume for i,g,t 12

dealer group g on bond i and day t, and IDG is the aggregate signed interdealer trade volume for i,g,t dealer group g on bond i and day t. A positive DCG implies that dealer group g on net bought from i,g,t customers, and a positive IDG implies that dealer group g on net bought from other dealers. We set i,g,t S = {large,medium,small,CB} and consider the same set of dealers for g. We do not include IDBs and ATS because these dealers do not take much aggregate daily net positions in both the dealer-customer and the interdealer segment. We use daily data for the estimations and run the regression separately by dealer group g. Because there is a very large number of observations for which all dependent and independent variables are all zeros, we set the intercept to be zero and omit those observations when estimating the regression. Table 4 report the regression results, which indicate the following. First, smaller dealers and client brokers offload larger share of their aggregate daily customer order flow through the interdealer market to other types of dealers. For instance, in high yield bonds (Panel b), small dealers offload 63.5% of their daily aggregate net customer order flow through the interdealer market, and 62% of client brokers do so. In contrast, only 4.6% and 20.6% of large and medium dealers, respectively, offload their daily aggregate net customer order flow to other dealer groups. Even if we delete trades that are prearranged, 38.1% of small dealers’ customer order flow is offloaded through the interdealer market on the same day (Table A.1 in the Appendix). This speed and ease for which small dealers offload in the interdealer market is consistent with other dealers being willing to take in those flows because these flows are uninformed. For a search model to incorporate such a fast inventory reversion, it would require a fairly high search intensity for small dealers. Second,asagroup,largedealersaremorelikelytoprovideliquiditytosmallerdealersandclientbrokers. For instance, in the high yield market, large dealers absorb 13.3%, 43.3%, and 37.0% of medium, small, and client brokers’ daily customer order flows, respectively. In contrast, medium dealer and small dealers only absorb1.3%and0.4%,respectivelyoflargedealers’aggregatecustomerflows. Aspreviouslymentioned,this empirical fact is consistent with a number of different economic channels. In Table 5, we take a further look at offloading at individual dealer level for large and medium dealers. We run the following regression separately for large dealers and medium dealers: (cid:88) ID =0+α DC +α DCG + β DCG +ϵ , (4) i,j,t 0 i,j,t 1 i,−j,t h i,h,t i,j,t h∈/g(j) where DC and ID are the signed customer order flow and the signed interdealer order flow for i,j,t i,j,t bond i, dealer j, day t. DCG is the aggregate customer order flow for bond i, dealer group h ∈ i,h,t 13

Table 4: Offloading regression: dealer group level: The following tables present the results from regression (3). Panel (a) presents the results for investment grade bonds, and panel (b) presents the results for high-yield bonds. (a) Investment grade large medium small client broker (1) (2) (3) (4) DCG −0.022∗∗∗ 0.010∗∗∗ 0.004∗∗∗ 0.002∗∗∗ large (0.001) (0.0004) (0.0002) (0.0003) DCG 0.059∗∗∗ −0.094∗∗∗ 0.012∗∗∗ 0.009∗∗∗ medium (0.003) (0.004) (0.001) (0.001) DCG 0.241∗∗∗ 0.111∗∗∗ −0.382∗∗∗ 0.011∗∗∗ small (0.008) (0.005) (0.010) (0.001) DCG 0.350∗∗∗ 0.185∗∗∗ 0.065∗∗∗ −0.646∗∗∗ CB (0.013) (0.010) (0.003) (0.011) Observations 3,006,140 3,006,140 3,006,140 3,006,140 Adjusted R2 0.109 0.078 0.247 0.560 Note: ∗p<0.1; ∗∗p<0.05; ∗∗∗p<0.01 (b) High yield large medium small client broker (1) (2) (3) (4) DCG −0.046∗∗∗ 0.013∗∗∗ 0.004∗∗∗ 0.006∗∗∗ large (0.003) (0.001) (0.0003) (0.001) DCG 0.133∗∗∗ −0.206∗∗∗ 0.014∗∗∗ 0.012∗∗∗ medium (0.007) (0.009) (0.002) (0.002) DCG 0.433∗∗∗ 0.082∗∗∗ −0.635∗∗∗ 0.019∗∗∗ small (0.052) (0.011) (0.035) (0.006) DCG 0.370∗∗∗ 0.113∗∗∗ 0.029∗∗∗ −0.626∗∗∗ CB (0.018) (0.008) (0.004) (0.013) Observations 749,991 749,991 749,991 749,991 Adjusted R2 0.146 0.123 0.492 0.520 Note: ∗p<0.1; ∗∗p<0.05; ∗∗∗p<0.01 14

S = {large,medium,small,CB}, day t. g(j) is the dealer group for which dealer j is a part of, and DCG =DCG −DC , or in other words, the aggregate customer volume of the group that j is i,−j,t i,g(j),t i,j,t in but excluding dealer j’s own order flow. Column (1), for instance, shows that in investment-grade bonds, a large dealer on average offloads 4.5% of its net customer order flow through the interdealer market. The average large dealer also absorbs 0.3% of other large dealers’ order flow and 0.6% of medium dealers’ order flow. In contrast, column (2) shows that theaveragemediumdealeroffloads11.4%ofitsnetcustomerorderflowthroughtheinterdealermarket; and absorbs 0.05% of large dealers’ order flow and 0.1% other medium dealers’ order flow. Overall, the results in this table show that large dealers do not share risk much amongst themselves but do provide liquidity to other types of dealers. Table 5: Offloading regression: Individual dealer level: The following table presents the results from regression (4). IG HY large medium large medium (1) (2) (3) (4) DCG 0.003∗∗∗ 0.0005∗∗∗ 0.007∗∗∗ 0.001∗∗∗ large (0.0001) (0.00002) (0.0003) (0.00005) DCG 0.006∗∗∗ 0.001∗∗∗ 0.013∗∗∗ 0.001∗∗∗ medium (0.0003) (0.00005) (0.001) (0.0002) DCG 0.024∗∗∗ 0.006∗∗∗ 0.043∗∗∗ 0.004∗∗∗ small (0.001) (0.0003) (0.008) (0.0004) DCG 0.035∗∗∗ 0.009∗∗∗ 0.037∗∗∗ 0.006∗∗∗ CB (0.002) (0.001) (0.003) (0.0004) DC −0.045∗∗∗ −0.114∗∗∗ −0.110∗∗∗ −0.228∗∗∗ j (0.001) (0.004) (0.004) (0.008) Observations 27,216,362 54,254,879 7,086,222 14,134,429 Adjusted R2 0.017 0.048 0.053 0.113 Note: ∗p<0.1; ∗∗p<0.05; ∗∗∗p<0.01 Next, we take a look at which counterparties large dealers offload customer order flow and provide liquidity through. We run the following regression separately by each counterparty group g: (cid:88) ID =0+α DC +α DCG + β DCG +ϵ (5) j,g,t 0 j,t 1 −j,t h h,t j,t h∈/g(j) 15

We suppress bond subscript i for readability. We consider large dealers only for j and all dealer types for thecounterpartyg. ID isthesumofthesignedinterdealerflowfordealerj inwhichthecounterpartyis j,g,t in group g. The sum of ID for all g would equal ID . The right hand side of the regression is exactly j,g,t j,t same as in (4). Table6: Offloadingregression: Individualdealerlevelforlargedealers: Thefollowingtablepresents the results from regression (5) for large dealers. large medium small ATS IDB client broker (1) (2) (3) (4) (5) (6) DCG 0.0004∗∗∗ −0.00004∗∗ 0.0001∗∗∗ 0.0004∗∗∗ 0.002∗∗∗ 0.0003∗∗∗ large (0.0001) (0.00001) (0.00002) (0.00002) (0.0001) (0.00003) DCG 0.00004 0.003∗∗∗ 0.0001∗∗∗ 0.0004∗∗∗ 0.002∗∗∗ 0.001∗∗∗ medium (0.00003) (0.0002) (0.00005) (0.00005) (0.0001) (0.0001) DCG 0.0003∗∗ 0.0002∗∗∗ 0.021∗∗∗ 0.001∗∗∗ 0.003∗∗∗ 0.001∗∗∗ small (0.0001) (0.0001) (0.001) (0.0001) (0.0002) (0.0001) DCG 0.0002∗ 0.0003∗∗∗ 0.001∗∗∗ 0.001∗∗∗ 0.002∗∗∗ 0.031∗∗∗ CB (0.0001) (0.0001) (0.0001) (0.0002) (0.0003) (0.001) DC −0.005∗∗∗ −0.003∗∗∗ −0.004∗∗∗ −0.006∗∗∗ −0.027∗∗∗ −0.008∗∗∗ j (0.001) (0.0002) (0.0003) (0.0002) (0.001) (0.001) Observations 34,302,584 34,302,584 34,302,584 34,302,584 34,302,584 34,302,584 Adjusted R2 0.001 0.003 0.014 0.002 0.009 0.018 Note: ∗p<0.1; ∗∗p<0.05; ∗∗∗p<0.01 Table 6 presents the results, and Table A.2 in the Appendix presents the results for medium dealers. We first consider how large dealers offload and absorb risk. Looking across all columns for the coefficient on DC , we find that when the customer order flow for a large dealer increases by 1, he offloads 0.027 through j,t IDB and smaller amounts through other dealers. Looking across the first row (coefficient for DCG ), −j,t when the customer order flow for all other large dealers increase by 1, a large dealer absorbs 0.0004 of it by trading directly with other large dealers (column 1) and 0.002 of it by trading with IDBs (column 4).7 This means that when a large dealer provides liquidity to other large dealers, they are five times as likely to do so through IDBs than bilaterally. Overall, the results point to large dealers offloading risk to each others mostly through IDBs. 7While 0.0004 and 0.002 may also seem minuscule, recall that the left hand side is for a single dealer–so overall, if the customer order flow increases by 1 for a large dealer, all other nine large dealers on aggregate would absorb about 0.0036 directlyand0.018throughIDBs. 16

When large dealers provide liquidity to other dealer types, they are more likely to do so bilaterally than through IDBs. For instance, comparing the coefficient for DCG between columns (3) and (5), we small,t find that small dealers are about nine times more likely offload to large dealers directly through bilateral trades than through IDBs. Medium dealers are somewhere in the middle in which large dealers are still twice as likely to absorb medium dealers’ customer order flow bilaterally than through IDBs (coefficient for DCG in columns 2 and 5). medium,t If search frictions were most relevant, large dealers finding each other would have the lowest possible costs, and thus they would have no reason to trade through IDBs. Moreover, smaller dealers will utilize IDBs more if search frictions were large. Our results are not consistent with search frictions but are rather consistent with large dealers contacting others through IDBs to minimize information leakage. Moreover, giventhat smalldealershave uninformedorlessinformedcustomerorderflow, theydonothavetheneedto conceal their trading intent through IDBs. Thus, use of IDBs in sharing risk overall points to information friction being most relevant in the interdealer markets. 4.2 Interdealer prices Next,welookatinterdealertradeprices. Searchfrictionandinformationfrictionwouldgivestarklydifferent predictions about interdealer prices. If search friction was most dominant, large dealers should receive best prices because they have lowest search cost and thus highest outside opportunity. Moreover, given that they are more likely to be providing liquidity,largedealersshouldreceivethebestprices. Incontrast,ifinformationfrictionsweremostdominant, large dealers would receive the worst prices when they are taking liquidity since they are most likely to be informed. However, because large dealers are also more likely to be providing liquidity, the predictions are lessclearcutwhennotconditionedonliquiditydemand. However,wecanconfidentlysaythatiflargedealers do not receive the best prices, search frictions likely are not the most dominant friction in the interdealer market. In the first regression, we compare prices of interdealer trades on the same bond-day by including bond times day fixed effects. We run the following regression: (cid:88) (cid:88) P =P∗ + β 1(seller(τ)=j)+ γ 1(buyer(τ)=k)+ϵ , (6) i,t,τ i,t j k i,t,τ j∈S k∈S where P is the traded price for bond i, day t, interdealer trade τ. P∗ is the “fundamental” value for i,t,τ i,t 17

Table 7: Interdealer price regression with bond-day interacted fixed effects.: Following table present results of regression (6). < 100k ≥ 100k IG HY IG HY (1) (2) (3) (4) seller = medium 3.185∗∗∗ 9.793∗∗∗ 2.984∗∗∗ 4.534∗∗∗ (0.134) (0.342) (0.097) (0.299) seller = small 2.373∗∗∗ 4.859∗∗∗ 3.149∗∗∗ 5.198∗∗∗ (0.222) (0.480) (0.129) (0.353) seller = ATS −1.087∗∗∗ −1.827∗∗∗ 2.959∗∗∗ 3.665∗∗∗ (0.137) (0.442) (0.145) (0.338) seller = IDB 5.550∗∗∗ 12.977∗∗∗ 2.772∗∗∗ 4.944∗∗∗ (0.319) (0.485) (0.133) (0.275) seller = client broker 1.125∗∗∗ −1.580∗∗∗ 2.409∗∗∗ 2.201∗∗∗ (0.166) (0.412) (0.121) (0.280) buyer = medium −3.637∗∗∗ −11.891∗∗∗ −1.021∗∗∗ −2.980∗∗∗ (0.185) (0.465) (0.082) (0.243) buyer = small −2.340∗∗∗ −5.365∗∗∗ −0.441∗∗∗ −2.380∗∗∗ (0.284) (0.432) (0.116) (0.363) buyer = ATS 5.279∗∗∗ 1.935∗∗∗ 0.507∗∗∗ −2.561∗∗∗ (0.133) (0.374) (0.148) (0.345) buyer = IDB 0.243 −10.577∗∗∗ −2.573∗∗∗ −5.172∗∗∗ (0.209) (0.544) (0.116) (0.265) buyer = client broker 0.985∗∗∗ 4.015∗∗∗ 0.354∗∗∗ −0.272 (0.125) (0.341) (0.117) (0.298) sell side = DC-ID −19.293∗∗∗ −25.115∗∗∗ −7.447∗∗∗ −11.086∗∗∗ (0.352) (0.620) (0.227) (0.556) buy side = DC-ID 12.838∗∗∗ 20.808∗∗∗ 7.119∗∗∗ 11.885∗∗∗ (0.211) (0.432) (0.135) (0.448) bond times day f.e. Yes Yes Yes Yes Observations 5,316,663 1,754,726 1,799,483 802,867 Adjusted R2 0.999 0.999 1.000 0.999 Note: ∗p<0.1; ∗∗p<0.05; ∗∗∗p<0.01 18

bond i on day t, which we proxy with bond-day interacted fixed effects. Large dealer type is the omitted category for both the buyer and the seller. We also include indicator variables for whether the trade was a DC-ID trade from the seller’s perspective and from the buyer’s perspective because the middle dealer in a DC-ID trade is taking liquidity in the interdealer trade and should get worse prices. Results are presented in Table 7. Results indicate that large dealers get worse prices than medium and small dealers. For instance, for investment-grade interdealer trades that are 100K or larger (column 3), medium sellers sell at 3 bps higher and small sellers sell at 3.1 bps higher than large sellers. Comparing columns(3)and(4), wecanseethatthedifferencesbetweenlargeandotherdealersaresomewhatlargerfor high-yieldbonds,whichisconsistentwiththenotionthathigh-yieldbondshavemoreinformationasymmetry. TradingcostshaveaU-shapeinwhichthelargestandthesmallestdealerspayhigherpricethanmedium dealers. For instance, in column (4), medium buyers pay less than both large buyers and small buyers. This would mean that when comparing smallest dealers to others, search costs may be more important. Or in other words, that up to top 30-40 dealers, information asymmetry friction matters more, and that beyond those, search frictions matter more. Given that trading tends to be dominated by the top 30-40 dealers, on a trade-by-trade basis, information would be more important. Also, because short positions are costly, dealers would have more incentive to offload customer buy trades regardless of information, so information frictions are likely weaker for buyers. This asymmetry may explain why the trading cost difference between large dealers and medium/small dealers are smaller for buyers and why trading costs are more of a U-shape for buyers. Client brokers, which tend to be on average smaller than small dealers, pay higher trading costs than small dealers, also adding to the U-shape. Lastly, the coefficient on IDBs indicate that large dealers pay about 2.5 bps (IG) and 5 bps (HY) more to trade through IDBs than bilaterally with each other and that this spread accrues to IDBs. We will look more closely into IDBs in Section 4.3. Because bond-days with multiple interdealer trades may be relatively scarce and different from other interdealer trades, we also estimate (6) in a different way. Instead of including bond-day fixed effects for P∗ , we take the first difference of (6) and proxy ∆P∗ , the change in fundamental price of bond i between i,t i,t 19

two consecutive trades, using market, default, and term factors: ret =α +α MKT +α DEF +α TERM i,t,τ 0 1 t(τ−1),t(τ) 2 t(τ−1),t(τ) 3 t(τ−1),t(τ) 6 (cid:88) + β [1(seller(τ)=j)−1(seller(τ −1)=j)] j j=1 6 (cid:88) + γ [1(buyer(τ)=j)−1(buyer(τ −1)=j)]+ϵ (7) j i,t,τ j=1 whereret isthereturnofbondibetween(τ−1)-thtradeandτ-thtrade, andt(τ)isthedateoftradeτ. i,t,τ MKT ,DEF ,TERM are the market, default, term factor calculated from index t(τ−1),t(τ) t(τ−1),t(τ) t(τ−1),t(τ) valuesatendofdayt(τ−1)(previoustradedate)andt(τ). ThesefactorsarecalculatedfromBAMLindices. Results for this first difference regression (7) are presented in Table 8. Overall, the results are both qualitatively and quantitatively similar to those in Table 7. 4.3 Who trades through interdealer brokers? In Section 4.1 and Table 6, we have established that large dealers are more likely to use IDBs and that this pattern is consistent with information friction than search friction. In this subsection, we look at the use of IDBs in more detail. InTable9,welookatwhotradeswithwhomthroughIDBs. About46%oftradingvolumethroughIDBs arelargedealerstradingwithotherlargedealers. Asarguedbefore, thesearepreciselythecaseswithlowest search cost and highest information friction. Thus, the use of IDBs is mostly motivated by information frictions. Another 30% is medium dealers trading with large dealers, which also has relatively low search costs. Putting the fact that large dealers trade with other large dealers bilaterally and IDBs for about 9% and 49%, respectively, of their interdealer volume (Table 1(d)) and the IDB trade composition together implies that large dealers trade a large share of their volume with other large dealers through IDBs rather than bilaterally. The fact that they do so despite the fact that large dealers pay more to trade with IDBs than bilaterallywithotherlargedealers(Table7)impliesthattherearebenefitsforlargedealerstotradethrough IDBs with each other despite the higher cost, such as lower risk of information leakage when trades fall through. Because results in Table 9 could in part be driven by the fact that large dealers trade more, we also study whether large dealers are more likely to trade with an IDB in a regression setting. Table 10 runs the 20

Table8: Interdealer price first difference regression: Followingtablepresentsresultsofregression(7). <100k ≥100k IG HY IG HY (1) (2) (3) (4) MKT 0.457∗∗∗ 1.107∗∗∗ 0.686∗∗∗ 1.092∗∗∗ (0.009) (0.173) (0.009) (0.052) TERM 0.267∗∗∗ −0.276 0.207∗∗∗ −0.037 (0.009) (0.185) (0.009) (0.055) DEF −0.150∗∗∗ 0.313∗∗∗ −0.207∗∗∗ 0.534∗∗∗ (0.004) (0.065) (0.004) (0.025) seller=medium 3.024∗∗∗ 9.759∗∗∗ 2.984∗∗∗ 4.697∗∗∗ (0.030) (0.206) (0.040) (0.204) seller=small 1.744∗∗∗ 5.949∗∗∗ 2.535∗∗∗ 5.402∗∗∗ (0.053) (0.279) (0.062) (0.349) seller=ATS −1.362∗∗∗ −3.176∗∗∗ 2.535∗∗∗ 3.953∗∗∗ (0.041) (0.242) (0.067) (0.323) seller=IDB 4.115∗∗∗ 13.928∗∗∗ 2.243∗∗∗ 4.904∗∗∗ (0.046) (0.199) (0.048) (0.177) seller=clientbroker 0.894∗∗∗ −2.304∗∗∗ 2.003∗∗∗ 2.496∗∗∗ (0.047) (0.235) (0.057) (0.233) buyer=medium −3.534∗∗∗ −13.050∗∗∗ −1.109∗∗∗ −3.339∗∗∗ (0.043) (0.225) (0.043) (0.211) buyer=small −2.268∗∗∗ −7.010∗∗∗ −0.212∗∗∗ −3.599∗∗∗ (0.060) (0.304) (0.057) (0.296) buyer=ATS 4.912∗∗∗ 1.377∗∗∗ 0.278∗∗∗ −3.127∗∗∗ (0.045) (0.237) (0.068) (0.296) buyer=IDB 0.112∗∗ −13.456∗∗∗ −2.307∗∗∗ −5.748∗∗∗ (0.050) (0.240) (0.048) (0.179) buyer=clientbroker 0.816∗∗∗ 5.521∗∗∗ 0.351∗∗∗ −0.115 (0.026) (0.152) (0.047) (0.213) sellside=DC-ID −18.415∗∗∗ −29.517∗∗∗ −7.354∗∗∗ −14.571∗∗∗ (0.050) (0.231) (0.072) (0.285) buyside=DC-ID 12.295∗∗∗ 22.889∗∗∗ 6.843∗∗∗ 13.547∗∗∗ (0.030) (0.166) (0.041) (0.217) Constant 0.560∗∗∗ 0.733∗∗∗ 0.239∗∗∗ 0.487∗∗∗ (0.014) (0.087) (0.018) (0.085) Observations 5,195,774 1,738,731 1,706,691 792,720 AdjustedR2 0.193 0.077 0.189 0.040 Note: ∗p<0.1;∗∗p<0.05;∗∗∗p<0.01 21

Table 9: Who trades with whom through IDBs: We focus on cases in which it is possible to identify the two end parties of trades that happen through IDBs. To do so, we match buy and sell trades of IDBs in which the two trades are in the same bond, within one minute apart, same IDB is the reporting party, but in opposite directions. In these cases, we say that the two counterparties of the two trades are the actual end parties that traded through the IDB. Within these trades, we pull the share (by volume) for the pairs. dealer type large medium small ATS IDB client broker large 46.15% 30.00% 6.71% 0.17% 1.41% 2.81% medium 4.44% 3.22% 0.30% 0.58% 0.83% small 0.63% 0.11% 0.26% 0.86% ATS 0.01% 0.06% 0.91% IDB 0.17% 0.30% client broker 0.06% following regression using trade-level data: (cid:88) 1(counterparty is IDB) =α+ β 1(j in dlr grp g)+γ log(size )+ϵ i,j,τ g 1 τ i,j,τ whereinterdealertradeτ isinbondiandreportingpartyisdealerj. Wealsoincludebondratingsandtrade size. Results indicate that large dealers are more likely to trade with IDBs, especially for large trades. The fact that results are stronger in institutional sized trades is consistent with information asymmetry being higher for large trades. Also, dealers are more likely to use IDBs for large trades and for lower-rated bonds, which is consistent with these trades having higher information asymmetry and IDB usage being driven by information frictions. In Table 11, we test whether IDBs allow dealers to trade with dealers that they normally do not trade with. We find that for large dealers, around 96% of their ultimate counterparties in IDB trades are those thattheyalreadytradewithbilaterally. Therefore,IDBsdonothelplargedealersreachnewcounterparties. For small dealers, IDBs do help in part with reaching new counterparties; about 47% of their trades with IDBs are with dealers that they do not trade bilaterally with. Overall,dealersuseIDBstomitigateinformationfriction,andIDBsareimportantpartoftheinterdealer market. Tothebestofourknowledge, inmostpapersstudyingOTCmarkets, andespeciallythosestudying the U.S. corporate bond markets, the role of IDBs has not been studied. Without isolating them separately, IDBs can be classified as central dealers together with large dealers, which can obscure some of the effects. 22

Table 10: IDB use regression: For interdealer trades in which the reporting party is not a ATS or a IDB, we look at the likelihood that the counterparty is a IDB. We run the following regression: (cid:88) 1(counterparty is IDB) =α+ β 1(j in dlr grp g)+γ log(size )+ϵ i,j,k g 1 k i,j,k where trade k is in bond i and rerporting party is dealer j. We also include bond rating group fixed effects. We report heterogeneity-consistent standard errors. all < 100k ≥ 100k (1) (2) (3) medium −0.040∗∗∗ −0.015∗∗∗ −0.073∗∗∗ (0.0002) (0.0002) (0.001) small −0.034∗∗∗ 0.025∗∗∗ −0.147∗∗∗ (0.0003) (0.0003) (0.001) client broker 0.041∗∗∗ 0.088∗∗∗ −0.074∗∗∗ (0.0002) (0.0002) (0.001) log(size) 0.047∗∗∗ 0.019∗∗∗ 0.087∗∗∗ (0.0001) (0.0001) (0.0002) rating BBB+:BBB- 0.001∗∗∗ 0.002∗∗∗ 0.004∗∗∗ (0.0002) (0.0002) (0.001) rating BB+:BB 0.064∗∗∗ 0.058∗∗∗ 0.081∗∗∗ (0.0004) (0.0004) (0.001) rating BB- or lower 0.104∗∗∗ 0.100∗∗∗ 0.095∗∗∗ (0.0003) (0.0003) (0.001) Constant −0.048∗∗∗ −0.003∗∗∗ −0.230∗∗∗ (0.0002) (0.0003) (0.001) Observations 13,525,402 10,047,278 3,478,124 R2 0.083 0.047 0.076 Adjusted R2 0.083 0.047 0.076 Note: ∗p<0.1; ∗∗p<0.05; ∗∗∗p<0.01 23

Table 11: IDB relationship: For IDB trades in which the end parties are identified (those used in Table 9), we look at whether the two end parties have a direct trading relationship outside of trading through IDBsorATS.“Currentmonth”columnshowstheshareofvolumeinwhichthetwoendpartieshaveatleast one trade with each other in the same month, “prior month” column shows the share in which the two end parties have at least one trade with each other in the prior month. dealer type current month prior month large 95.78% 93.92% medium 78.55% 76.56% small 52.51% 51.45% ATS 29.05% 28.94% ID broker 62.84% 61.61% client broker 59.45% 57.92% 4.4 Interdealer market efficiency How much do information frictions matter for market efficiency? The main function of interdealer markets is for dealers to share risk that arises from making markets for their clients (Viswanathan and Wang, 2004). Theclearestcaseofgainsfromtradeintheinterdealermarketwouldbewhenonedealerhadanetcustomer buy flow and another had a net customer sell flow in the same bond. In this case, the two dealers can trade to offload their inventory, and there would be a clear positive gains from trade. Using these clearly identifiable potential gains from trades between large dealers, we look at what share of potential gains are realized, either through direct (bilateral) trades between the two dealers or through indirect chains. We calculate the interdealer market efficiency in the following way. For bond i on day t, if dealer A has bought on net amount v from his customers, and if dealer B has sold on net amount v to her customers, A B the potential gains from trade between dealer A and B is v =min(v ,v ). If A sells bond i of amount v P A B D to B between day t and t+k, then min(vD,vP) of potential gains from trade is realized directly. If A sells vP bond i of amount v to B through dealer C between day t and t+k, then min(vD+vI,vP) of potential gains I vP from trade is realized either directly or indirectly. We only focus on cases in which both dealers A and B are large dealers, and use k =0,2,5. Table 12 present the share of potential gains that is realized. Only a very small share of potential gains arerealizedthroughbilateraltrades,andslightlymore(butstillrelativelysmall)isrealizedthroughindirect trades. For instance, only 1.54% of potential gains from trade between two large dealers in a high-yield bond is realized on the same day by direct trade between the two dealers. Another 9.8% is realized through indirect trades, mostly from trades through IDBs. Looking over multiple days increases gains from trade somewhat,butthenumbersarestillfairlylow. Thereislessdirectgainsbutoverallhighergainsfromtrades 24

in high-yield bonds, which is consistent with the notions that high-yield bonds have higher information asymmetry and also higher gains from trade due to greater risk reduction. Table12: Interdealermarketefficiencyforlargedealers: Followingtablepresentstheshareofpotential gains-from-trade between large dealers that are realized. Direct gains only include realized gains from two large dealers trading bilaterally, and all gains include both the direct gains and indirect gains from trading through one additional dealer. IG HY k direct all direct all 0 1.21% 6.16% 1.54% 11.34% 2 1.51% 9.82% 1.99% 17.21% 5 1.89% 13.33% 2.54% 22.52% Results overall indicate that information friction decreases interdealer market efficiency, especially betweenlargedealers. LargedealersuseIDBstogetaroundtheissue,butitdoesnotfullysolvetheinefficiency. 5 Conclusion Inthispaper,weshowthatinformationfrictionsmatterintheinterdealersegmentofOTCmarkets. Despite the focus on search frictions and network formation, we show that on average, information frictions seem to matter more in the interdealer market for U.S. corporate bonds. This is not to say that search frictions in thedealer-customersegmentdonotmatter. Theeaseoffindingcustomerbuyersandsellerslikelycontribute to the lower holding cost for large dealers. Since the adoption of Dodd-Frank and Volcker Rule in mid-2010’s, corporate bond liquidity has structurallyshiftedfromdealersprovidingimmediacyandinventoryspacetodealersactingmoreasmatch-makers because of higher inventory costs (Choi et al., 2023; Bessembinder et al., 2018). Our results indicate that there is room to improve market structure so that dealers, especially large dealers, can free up inventory space and provide more liquidity to customers. References Babus, A. and P. Kondor (2018): “Trading and information diffusion in over-the-counter markets,” Econometrica, 86, 1727–1769. 25

Bessembinder, H., S. E. Jacobsen, W. F. Maxwell, and K. Venkataraman (2018): “Capital Commitment and Illiquidity in Corporate Bonds,” The Journal of Finance, 73, 1615–1661. Choi, J., Y. Huh, and S. Seunghun Shin (2023): “Customer liquidity provision: Implications for corporate bond transaction costs,” Management Science. Czech, R. and G. Pinte´r (2022): “Informed trading and the dynamics of client-dealer connections in corporate bond markets,” . De Roure, C., E. Moench, L. Pelizzon, and M. Schneider (2019): “OTC discount,” Available at SSRN 3507679. Di Maggio, M., A. Kermani, and Z. Song (2017): “The value of trading relations in turbulent times,” Journal of Financial Economics, 124, 266–284. Dick-Nielsen, J., T. K. Poulsen, and O. Rehman (2020): “Dealer Networks and the Cost of Immediacy,” Available at SSRN. Duffie, D., N. Gaˆrleanu, and L. H. Pedersen (2005): “Over-the-Counter Markets,” Econometrica, 73, 1815–1847. Friewald, N. and F. Nagler (2019): “Over-the-counter market frictions and yield spread changes,” The Journal of Finance, 74, 3217–3257. Glode, V. and C. Opp(2016): “Asymmetricinformationandintermediationchains,”AmericanEconomic Review, 106, 2699–2721. Han, S. and X. Zhou (2014): “Informed bond trading, corporate yield spreads, and corporate default prediction,” Management Science, 60, 675–694. 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, 1623–1656. Hendershott, T., R. Kozhan, and V. Raman (2020a): “Short selling and price discovery in corporate bonds,” Journal of Financial and Quantitative Analysis, 55, 77–115. Hendershott, T., D. Li, D. Livdan, and N. Schu¨rhoff (2020b): “Relationship Trading in Over-the- Counter Markets,” The Journal of Finance, 75, 683–734. 26

Ho, T. S. and H. R. Stoll (1983): “The dynamics of dealer markets under competition,” The Journal of finance, 38, 1053–1074. Hollifield, B., A. Neklyudov, and C. Spatt (2016): “Bid-Ask Spreads, Trading Networks, and the Pricing of Securitizations,” The Review of Financial Studies. Kedia, S. and X. Zhou (2014): “Informed trading around acquisitions: Evidence from corporate bonds,” Journal of Financial Markets, 18, 182–205. Lagos, R. and G. Rocheteau (2009): “Liquidity in asset markets with search frictions,” Econometrica, 77, 403–426. Lewis, R. and M. Schwert (2021): “The effects of transparency on OTC market-making,” Jacobs Levy Equity Management Center for Quantitative Financial Research Paper. Li, D. and N. Schu¨rhoff (2019): “Dealer networks,” The Journal of Finance, 74, 91–144. Pinter, G., C. Wang, and J. Zou (2022): “Size discount and size penalty: Trading costs in bond markets,” . U¨slu¨, S. (2019): “Pricing and liquidity in decentralized asset markets,” Econometrica, 87, 2079–2140. Viswanathan, S. and J. J. Wang (2004): “Inter-dealer trading in financial markets,” The Journal of Business, 77, 987–1040. Wang, C. (2016): “Core-periphery trading networks,” Available at SSRN 2747117. Wei, J. and X. Zhou (2016): “Informed trading in corporate bonds prior to earnings announcements,” Financial Management, 45, 641–674. 27

A Additional Results Table A.1: Offloading regression: Dealer group level, unmatched: The following tables present the results from regression (3) but using unmatched trades only. Panel (a) presents the results for investment grade bonds, and panel (b) presents the results for high-yield bonds. (a) Investment grade large medium small client broker (1) (2) (3) (4) DCG −0.008∗∗∗ 0.006∗∗∗ 0.002∗∗∗ 0.0004∗∗∗ large (0.0005) (0.0003) (0.0002) (0.0001) DCG 0.026∗∗∗ −0.034∗∗∗ 0.005∗∗∗ 0.002∗∗∗ medium (0.002) (0.003) (0.0005) (0.0004) DCG 0.129∗∗∗ 0.048∗∗∗ −0.191∗∗∗ 0.004∗∗∗ small (0.008) (0.004) (0.008) (0.001) DCG 0.085∗∗∗ 0.041∗∗∗ 0.018∗∗∗ −0.152∗∗∗ CB (0.007) (0.004) (0.003) (0.007) Observations 2,961,361 2,961,361 2,961,361 2,961,361 Adjusted R2 0.020 0.013 0.096 0.063 Note: ∗p<0.1; ∗∗p<0.05; ∗∗∗p<0.01 (b) High yield large medium small client broker (1) (2) (3) (4) DCG −0.017∗∗∗ 0.009∗∗∗ 0.002∗∗∗ 0.002∗∗∗ large (0.002) (0.001) (0.0003) (0.0005) DCG 0.062∗∗∗ −0.095∗∗∗ 0.008∗∗∗ 0.002 medium (0.005) (0.007) (0.001) (0.001) DCG 0.260∗∗∗ 0.055∗∗∗ −0.381∗∗∗ 0.020∗∗∗ small (0.015) (0.006) (0.017) (0.007) DCG 0.104∗∗∗ 0.025∗∗∗ 0.011∗∗∗ −0.156∗∗∗ CB (0.011) (0.005) (0.004) (0.010) Observations 724,667 724,667 724,667 724,667 Adjusted R2 0.028 0.036 0.248 0.056 Note: ∗p<0.1; ∗∗p<0.05; ∗∗∗p<0.01 28

Table A.2: Offloading regression: Individual dealer level for medium dealers: The following table presents the results from regression (5) for medium dealers. large medium small ATS IDB client broker (1) (2) (3) (4) (5) (6) DCG 0.0001∗∗∗ −0.00000 0.00002∗∗∗ 0.00005∗∗∗ 0.0002∗∗∗ 0.0001∗∗∗ large (0.00001) (0.00000) (0.00000) (0.00000) (0.00001) (0.00001) DCG −0.00001 0.001∗∗∗ 0.00003∗∗∗ 0.0001∗∗∗ 0.0002∗∗∗ 0.0002∗∗∗ medium (0.00001) (0.00003) (0.00001) (0.00001) (0.00003) (0.00002) DCG −0.0001∗∗ −0.00001 0.004∗∗∗ 0.0004∗∗∗ 0.0003∗∗∗ 0.0004∗∗∗ small (0.00004) (0.00002) (0.0002) (0.0001) (0.0001) (0.0001) DCG −0.0002∗∗∗ −0.00001 0.0003∗∗∗ 0.0001∗∗∗ 0.0001∗ 0.008∗∗∗ CB (0.0001) (0.00002) (0.0001) (0.00003) (0.00005) (0.0005) DC −0.039∗∗∗ −0.012∗∗∗ −0.011∗∗∗ −0.007∗∗∗ −0.027∗∗∗ −0.026∗∗∗ j (0.003) (0.001) (0.001) (0.0005) (0.001) (0.002) Observations 68,389,308 68,389,308 68,389,308 68,389,308 68,389,308 68,389,308 Adjusted R2 0.027 0.007 0.008 0.002 0.007 0.015 Note: ∗p<0.1; ∗∗p<0.05; ∗∗∗p<0.01 29