You are on page 1of 54

THE JOURNAL OF FINANCE • VOL. LXXIV, NO.

1 • FEBRUARY 2019

Dealer Networks
DAN LI and NORMAN SCHÜRHOFF∗

ABSTRACT
Dealers in the over-the-counter municipal bond market form trading networks with
other dealers to mitigate search frictions. Regulatory data show that this network
has a core-periphery structure with 10 to 30 hubs and over 2,000 peripheral broker-
dealers in which bonds flow from periphery to core and partially back. Central dealers
charge investors up to double the round-trip markups compared to peripheral dealers.
In turn, central dealers provide immediacy by matching buyers with sellers more
directly and prearranging fewer trades, especially during stress times. Investors thus
face a trade-off between execution cost and speed, consistent with network models of
decentralized trade.

DEALERS PLAY A PIVOTAL ROLE in price formation and liquidity in over-the-


counter (OTC) financial markets. In frictionless markets, buyers and sellers
find each other immediately and realize gains from trade at competitive prices.
This is not the case, however, in decentralized markets. In particular, the
opaque nature of bilateral OTC transactions creates frictions in finding coun-
terparties that can reduce price and allocation efficiency (Duffie, Gârleanu,
and Pedersen (2005)). To address these search and matching frictions, broker-
dealers match buyers with sellers not only among their own clients, but also
warehouse securities and form trading networks with other dealers. The trad-
ing networks allow the dealers to find counterparties more quickly and real-
locate securities more efficiently. Given how crucial the dealer networks are
for liquidity provision in OTC markets, surprisingly little is known empirically
about them, as a lack of granular data so far has precluded serious empirical
study of the network structure and its impact on the terms of trade for investors.
∗ Dan Li is with the Board of Governors of the Federal Reserve System, Washington, DC. She has

no conflict of interest to disclose. Norman Schürhoff is with the Faculty of Business and Economics
and Swiss Finance Institute, University of Lausanne. Schürhoff is also Research Fellow of the
CEPR. He gratefully acknowledges research support from the Swiss Finance Institute and the
Swiss National Science Foundation under Sinergia project CRSII1_154445/1, “The Empirics of
Financial Stability” and has no conflict of interest to disclose. We thank the Editor, Bruno Biais,
an Associate Editor, three anonymous referees, and discussants Jeff Harris, Larry Harris, Terry
Hendershott, Christian Julliard, Bernd Mack, Tarun Ramadorai, and Elvira Sojli for extensive
comments. Seminar audiences at many conferences and invited seminars have provided valuable
feedback, particularly Andrew Ang, Jean-Edouard Colliard, Fany Declerck, Darrell Duffie, Amy
Edwards, Burton Hollifield, Laurence Lescourret, Craig Lewis, Semyon Malamud, Sophie Moinas,
Artem Neklyudov, Christine Parlour, Michael Rockinger, and Pierre-Olivier Weill. We are deeply
indebted to Rick Green for advice. We thank the Municipal Securities Rulemaking Board for
providing the transactions data with dealer identifiers.
DOI: 10.1111/jofi.12728

91
92 The Journal of FinanceR

In this paper, we employ regulatory audit trail data for the municipal bond
market to provide evidence on how dealer networks affect trade execution
in decentralized OTC markets during normal and stress times. In an effort
to improve market transparency, the Municipal Securities Rulemaking Board
(MSRB) has required broker-dealers to report all their trades since 1998. The
supervisory data set provides dealer identities for all 72 million municipal bond
trades between 1998 and 2012. The granularity of these trade data allows us
to trace the bonds through the dealer network, map the intermediation chains
and network structure, and compare the execution terms that investors receive
across different dealers.
We find that interconnectedness among dealers reduces search and allocation
frictions, but it also creates opportunities for dealers to employ market power in
privately negotiated transactions. The dealer network in municipal bonds has
a core-periphery structure with about 10 to 30 highly interconnected dealers at
the center and more than 2,000 broker-dealer firms sparsely connected at the
periphery.1 Use of the interdealer network is pervasive, with more than 20%
of round-trip chains connecting buyer with seller through two or more dealers.
The dealers with the most connections act as hubs in intermediation chains that
can involve up to seven dealers, consistent with models of intermediation chains
(Colliard and Demange (2014), Glode and Opp (2014), Hugonnier, Lester, and
Weill (2015), Neklyudov and Sambalaibat (2015)). Moreover, dealers’ trading
relations with each other are long-lived and centrality rankings of the dealers
are persistent over time. These findings suggest that a basic random search
model cannot easily match the data, and that network-based models of trade
are better suited to capture the interdealer market for municipal bonds (Babus
(2012), Chang and Zhang (2015), Wang (2016)).
We next examine whether central dealers pass on the time and cost savings
that result from their network connections in the form of smaller markups,
or whether these dealers charge larger markups for reducing search delays
and speeding up trade execution. We find that central dealers charge larger
markups to investors than their peripheral competitors. The centrality pre-
mium is sizable, with markups at central dealers up to twice those charged by
peripheral dealers. At the same time, central dealers handle the majority of
the order flow. The premium of 0.4% to 0.7% of par value that central dealers
charge relative to peripheral dealers is equivalent to several months of a bond’s
interest income. Investors in municipal bonds thus face a trade-off between ex-
ecution speed and cost. This result is important, as Lagos and Rocheteau (2007)
show that a sufficient degree of market power is required to incentivize dealer
entry and welfare-maximizing allocations.
The centrality premium and the fact that central dealers handle most of
the order flow in spite of being more expensive may seem surprising given the
presence of a large number of active broker-dealers in the market. We explore a

1 About 700 broker-dealer firms actively trade in municipals in a given month. Therefore, the

rolling dynamic dealer network that we use in the empirical analysis has, on average, about 700
dealers a month.
Dealer Networks 93

variety of potential explanations. The most obvious explanations do not appear


to play a role. In particular, central dealers do not systematically intermediate
riskier bonds and are not compensated for higher price risk, as they are less
likely to lose money in round-trip transactions than peripheral dealers.
Rather, consistent with central dealers providing more immediate execution
and getting compensated for doing so, we find that central dealers match buyers
with sellers more directly. After purchasing a bond from an investor, a central
dealer is more likely to sell it to the end-buyer than to off-load it to another
dealer. In addition, the length of the intermediation chain is shorter for round-
trips that start with a more central dealer: compared to sparsely connected
peripheral dealers, dealers that possess larger trading networks have better
access to clients and more information about which securities are available and
who wants to buy or sell, which results in shorter intermediation chains. When
we use detailed time stamps on each trade leg and associated dealer identifiers
to infer differences in investors’ execution delays based on the level of inventory
risk that dealers take, we find that, compared to peripheral dealers, central
dealers are more likely to offer immediacy by trading on a principal basis,
that is, by taking bonds into inventory, than to prearrange trades between a
buyer and a seller, which takes time to execute. Further, conditional on taking
a bond into inventory, central dealers tend to turn the bond around faster than
peripheral dealers, which suggests that central dealers also find buyers more
quickly. The higher propensity of central dealers to trade on a principal basis,
that is, to actively make a market, reduces execution delays and increases
trading gains for investors that value speed.
The municipal bond market provides a rich setting for studying search and
matching frictions. It is a typical OTC market in that trades are negotiated
bilaterally and search frictions are high, for a number of reasons. First, as-
set heterogeneity is vast, with more than 1.5 million different bonds and over
50,000 issuers, each with its own idiosyncrasies. Second, the natural holders
of municipals range from retail investors to sophisticated institutions such
as banks, mutual funds, and insurance companies. Because these are mostly
buy-and-hold investors, municipal bonds trade infrequently in the secondary
market (Green, Hollifield, and Schürhoff (2007)). When investors do trade,
it is typically for liquidity reasons, not private information about fundamen-
tals.2 Third, information about available bonds and suitable counterparties is
sparse. Indeed, it is known to be a buyer’s market in which identifying investors
willing to buy is a dealer’s most crucial task (Schultz (2012)). Locating bonds
and potential buyers thus requires that financial intermediaries have active
relationships with various types of investors as well as with other dealers. Fi-
nally, the municipal bond market is ideal for studying competition in liquidity
provision because of the large number of financial intermediaries—there are
over 2,000 registered broker-dealers—while the market’s segmentation along

2 Adverse selection risks are low in the municipal bond market since 75% of municipal bonds

are AAA-rated and historical default rates are just around 0.1% per year.
94 The Journal of FinanceR

state lines due to differential taxation of in-state and out-of-state bonds creates
submarkets that can curb dealer competition (Schultz (2013)).
Our empirical tests offer insights into which theories are more in line with the
trading and pricing in decentralized markets like the municipal bond market.
Several recent theories of decentralized trade in OTC markets yield predictions
on the structure of interdealer trading and the impact of dealer centrality on ex-
ecution quality. Yet, little is known about which of those models better describe
the data. Search-based models show that bilaterally negotiated prices depend
on dealer bargaining power and search efficiency (Duffie, Gârleanu, and Ped-
ersen (2005), Vayanos and Wang (2007), Lester, Rocheteau, and Weill (2014),
Hugonnier, Lester, and Weill (2015), Üslü (2015)) and that fast dealers are
central in intermediation chains (Neklyudov (2013), Weller (2013)). But inter-
dealer links are typically random and transitory in these models. By contrast,
network-based models show that trading costs depend on dealer centrality.
Predictions on a centrality premium or discount, however, are ambiguous in
both types of models (Chang and Zhang (2015), Hugonnier, Lester, and Weill
(2015), Neklyudov and Sambalaibat (2015), Üslü (2015)).
A stylized model of trade in OTC markets that we lay out in the Internet Ap-
pendix to this paper provides guidance on this issue.3 We derive conditions for
a centrality premium in terms of investors’ outside options or sophistication,
the relative bargaining power of dealer and investors, and the frequency and
size of investors’ liquidity shocks (i.e., their need for speed). We show that a
centrality, or speed, premium arises so long as investors have low negotiation
sophistication and a strong need for speed and dealers have sufficient bar-
gaining power as to reap profits from intermediation. Dealer market power is
thus a prerequisite for faster dealers to command a premium over their slower
competitors in bilaterally negotiated trades with investors.
We also provide evidence on additional predictions of the model. First, con-
sistent with variation in investor sophistication, we find that the centrality
premium is higher on smaller trades that likely originate from retail investors
than on larger trades that come from sophisticated institutions. Next, consis-
tent with the centrality premium being lower, and hence dealer competition
being more intense, in less segmented markets, we find that the centrality pre-
mium is lower in markets where the bond income of in-state and out-of-state
investors is taxed similarly—such as the markets for taxable municipals and
for tax-exempt Puerto Rico bonds. We also find support for the view that dealer
speed is especially valuable when investors’ liquidity shocks are large and
frequent, that is, during times of market stress. Specifically, large aggregate
mutual fund outflows, bond-specific mutual fund selling pressure, bond-rating
actions, and spikes in the volatility index (VIX) all predict the use of central
dealers. Weill (2007) and Lagos, Rocheteau, and Weill (2011) show that the type
of policy intervention needed in a crisis depends primarily on whether dealers
provide liquidity during stress times. Our evidence suggests that at least some

3 The Internet Appendix is available in the online version of the article on the Journal of Finance

website.
Dealer Networks 95

dealers do “lean against the wind,” that is, act as intermediaries of last resort,
which speaks to the efficiency of liquidity provision in the market.
To sharpen identification of what information is relevant for dealers’ search
process, we exploit differences in dealers’ centrality rankings when they inter-
mediate bonds from different states. Municipal bond investor clienteles differ
across states due to differences in the tax treatment of municipal interest
income in the state of the issuer and the state that the investor resides in.
Bonds from U.S. territories are traded in a national market, while bonds from
segmented states are traded locally. Dealers can therefore be central in the na-
tional market, for instance, but peripheral in a given state if they do not have
many trading connections for bonds from that state. We compute state-specific
measures of centrality for each dealer, state of issuance for the bonds the dealer
trades, and period. Consistent with the view that local information about coun-
terparties is crucial for trade matching, we show that state-specific centrality
has a larger effect on markups than aggregate centrality and is incremental to
dealer fixed effects.
We conduct a variety of tests that rule out alternative explanations. First, we
show that network centrality is not simply a proxy for dealer size, that is, our
results are not driven by scale effects. While several dealer characteristics—
including dealer asset size and indicators for primary dealers, dealer based in
New York City, and underwriters—are significant determinants of markups,
centrality is the single most important determinant and statistically signifi-
cant even after controlling for other dealer characteristics. Second, we show
that central dealers do not charge more simply because they intermediate
bonds that are a priori costlier or riskier to intermediate. For instance, central
dealers do not systematically intermediate bonds that are more illiquid based
on observable bond characteristics (e.g., unrated, lower rated, small issue size,
or exotic features such as a sinking fund or bank qualification). In addition,
central dealers are less likely to lose money on round-trip trades, which is con-
sistent with lower, not higher price risks. Third, we show that, while inventory
effects (Huang and Stoll (1997), Friewald and Nagler (2016)) are present, in the
sense that dealers decrease their markups when inventories are abnormally
high, higher inventory costs do not explain the centrality premium. Specifically,
we show that, conditional on taking bonds in inventory instead of prearranging,
central dealers have shorter inventory durations on principal trades, which is
consistent with central dealers having higher matching efficiency. Moreover,
while the dealer closest to the ultimate buyer earns the largest markup in an
intermediation chain, the last dealer does not hold the bond the longest.
To isolate the potential effect of adverse selection on markups, we test predic-
tions from Babus and Kondor (2018, BK hereafter). BK put forward a network
model of information diffusion in OTC markets in which market participants
possess private information. Their model predicts that better-connected dealers
are less exposed to adverse selection, as they observe more order flow, and thus
charge a smaller adverse-selection-related markup than peripheral dealers.
Consistent with BK, we find that the adverse selection component of markups
is lower for central than peripheral dealers, which suggests that central
96 The Journal of FinanceR

dealers are less exposed to adverse selection. This finding renders our centrality
premium result even more striking, as it implies that it is the noninformational
component of the markup that is higher at central dealers.
Our paper contributes to a rapidly growing empirical literature on
OTC markets. Bessembinder, Maxwell, and Venkataraman (2006), Edwards,
Harris, and Piwowar (2007), Green, Hollifield, and Schürhoff (2007), and
Green, Li, and Schürhoff (2010) study trading costs in OTC markets. Hender-
shott and Madhavan (2015) empirically explore the trade-off between bilateral
and electronic trading that Biais (1993) studies theoretically. Upper and Worms
(2004), Cocco, Gomes, and Martins (2009), Bech and Atalay (2010), Afonso and
Lagos (2015), Cohen-Cole, Kirilenko, and Patacchini (2012), and Craig and von
Peter (2014) explore the network topology of institutional money markets. Our
paper examines network topology in the municipal bond market that caters
to both retail and institutional investors, with a focus on the relation between
dealer centrality and execution quality. We extend the prior studies by docu-
menting the network structure in markets relevant for retail investors and the
impact on dealer markups and execution speed.
In more recent studies, Di Maggio, Kermani, and Song (2015) find support
for a centrality premium in the interdealer market for corporate bonds, while
Hollifield, Neklyudov, and Spatt (2017) find evidence of a centrality discount
for the largely institutional ABS/MBS market. Documenting the similarities
and differences between markets is important to better understand the fric-
tions that affect these markets. OTC financial markets, especially those for
municipal and corporate bonds and mortgage-backed securities, were hit hard
by the financial crisis and thus have been a focus of concern on the policy front.
Since OTC markets have no formal structure and bilateral trading technol-
ogy is arcane, slow, and expensive compared to centralized exchanges, dealers’
incentives to provide liquidity and the compensation they receive are of high
policy interest. Understanding the terms of trade different financial interme-
diaries offer, the sources of their market power, and the role of certain dealers
as liquidity providers of last resort during times of market stress enables us to
better assess the potential effect of various postcrisis regulatory changes such
as Basel III and the Volcker Rule, which have different implications for core
and peripheral dealers. Distinct from Di Maggio, Kermani, and Song (2015)
and Hollifield, Neklyudov, and Spatt (2017), however, we focus on the link be-
tween centrality and execution delays that many search-based models of OTC
trade have been concerned with.4
The remainder of our paper is organized as follows. Section I reviews the
theoretical literature and develops our hypotheses. Section II describes the
data and the procedure we use to identify intermediation chains. Section III
explores the dealer network. Section IV documents the centrality premium

4 Saunders, Srinivasan, and Walter (2002) analyze the behavior of a single dealer in the in-

terdealer corporate bond market. Hendershott et al. (2016) employ customer identifiers to study
optimal network size and the value of repeat relations between insurance companies and corporate
bond dealers, but they do not consider the interdealer market.
Dealer Networks 97

and Section V links dealer centrality with search efficiency. Section VI tests
additional model predictions and Section VII checks the robustness of our
results. Finally, Section VIII concludes.

I. Literature and Hypotheses


In a burgeoning theoretical literature on OTC markets, several recent theo-
ries of decentralized trade make predictions about the structure of interdealer
trading and the impact of dealer centrality on execution quality. Below, we
discuss the main theoretical predictions that motivate our empirical tests.
Existing theories of decentralized markets can be split into search models
and network models. Duffie, Gârleanu, and Pedersen (2005, DGP hereafter) an-
alyze the effect of decentralization on securities prices when investors search
randomly for counterparties and bargain over the terms of trade. DGP offer
insights as to why the law of one price can break down in the face of search
frictions and how bilaterally negotiated prices depend on traders’ bargaining
power.5 In DGP, dealers can off-load inventory at competitive prices in a fric-
tionless interdealer market. As a consequence, interdealer links are random
and transitory, intermediation chains and persistent interdealer network ef-
fects do not arise, and dealer centrality does not matter.
More recent theories of decentralized trade generate intermediation chains,
core-periphery network structures, and centrality premia or discounts. Colliard
and Demange (2014) and Glode and Opp (2014) develop liquidity and asym-
metric information theories, respectively, of intermediation chains. Neklyudov
(2013) extends the DGP random search setup to allow for fast and slow dealers
and a bilateral interdealer market (see also Weller (2013)). His model shows
that differences in search technologies across dealers lead to a core-periphery
structure that curbs dealer competition.6 In a generalization of DGP to a con-
tinuum of types with short-lived relations, Hugonnier, Lester, and Weill (2015,
HLW hereafter) show that intermediation chains can arise from valuation dif-
ferences among investors.7
A growing strand of literature models financial networks with long-lived
relations, focusing on equilibrium network structures and price formation in
networks. Babus (2012) and Wang (2016) predict core-periphery network struc-
tures. Their models highlight the trade-off between efficiency and stability in
network topology. Neklyudov and Sambalaibat (2015) study pricing in OTC

5 Bessembinder, Maxwell, and Venkataraman (2006), Edwards, Harris, and Piwowar (2007),
Green, Hollifield, and Schürhoff (2007), and Green, Li, and Schürhoff (2010) empirically study
prices in OTC markets and confirm several DGP predictions.
6 In Neklyudov (2013), dealers with higher meeting intensity provide a positive matching exter-

nality that limits slower dealers’ incentives to become more search efficient, which allows hetero-
geneity in search technology and dealer profits to persist.
7 Chang and Zhang (2015) also generate a core-periphery structure, albeit in a dynamic match-

ing model. HLW and Chang and Zhang (2015) draw no formal distinction between investor and
dealer. In their models, stable investor types endogenously become dealers and provide liquidity
to investors with larger and more frequent liquidity shocks.
98 The Journal of FinanceR

markets when dedicated dealers differ in search-and-matching efficiency be-


cause of different network connections. Their assumption that dealers search
for bonds among their own clients and connected dealers leads to the prediction
that bonds flow endogenously between fast central and slow peripheral dealers.
To differentiate between the two sets of models, we start by testing the
following set of predictions unique to the more recent theories of decentralized
trade that are not based on random trading relations.
HYPOTHESIS 1 (Intermediation chains and core-periphery structure): Deal-
ers intermediate bonds in chains, instead of always matching sellers with buy-
ers directly. Dealers form persistent and long-lived trading relations with other
dealers, so that the interdealer trading network has a core-periphery structure
and bonds flow from periphery dealers to the core dealers and back.
The two sets of models also differ in their predictions for bond pricing. Mod-
els with interdealer trading networks predict that prices and trading costs
depend on dealer centrality, while search models generate pricing differences
between core and periphery when dealers are heterogeneous in terms of search
efficiency or valuations. In both DGP and Neklyudov (2013), faster (future)
meetings with potentially the same dealer improve investors’ outside option,
which limits dealers’ bargaining position through “sequential competition” with
themselves and, in turn, their ability to internalize the benefit from faster ex-
ecution through larger markups. Trading costs therefore decline with dealers’
search efficiency in these models.
In HLW, a centrality premium arises in which central dealers charge larger
markups than their competitors, provided that investors are sufficiently het-
erogeneous and buyers have most of the bargaining power. A priori, these
conditions fit the municipal bond market, as it is geared toward both retail and
institutional investors and it is considered a buyer’s market (Schultz (2012)),
in which identifying investors willing to buy is dealers’ most crucial task. In
a related random search and bargaining model in which investors differ in
search intensity, Üslü (2015) shows that fast investors endogenously become
intermediaries and charge a speed, or centrality, premium. We test the pricing
prediction in Hypothesis 2.
HYPOTHESIS 2 (Centrality premium): Trading costs depend on dealer cen-
trality. Central dealers charge larger markups than peripheral dealers.
In Neklyudov (2013), Weller (2013), and Üslü (2015), the source of hetero-
geneity across dealers arises from dealers searching for counterparties with
different intensity, which in equilibrium leads to differences in their execution
speed and the markups they charge to investors. While the random search mod-
els of DGP and HLW do not directly address the trade-off between execution
cost and speed, it is at the core of several models of directed search.8 Most no-
tably, Vayanos and Wang (2007) develop a model of prearranged transactions

8 Demsetz (1968) was the first to discuss the idea that impatient investors can pay a markup to

speed up trading.
Dealer Networks 99

in two markets that requires a trade-off between execution price and speed
for a clientele equilibrium to exist. Investors in Vayanos and Wang (2007) are
characterized by their need for speed, as captured by the frequency with which
they receive liquidity shocks and their size, as well as their outside options or
sophistication. In Lester, Rocheteau, and Weill (2014), congestion externalities
generate an execution cost-speed trade-off in dealer-to-customer trades. The
drawback of these models is that trades occur only between customers or be-
tween customers and dealers. Intermediation chains or networks do not arise,
contrary to the network literature. We test predictions on the search efficiency
channel in Hypothesis 3.

HYPOTHESIS 3 (Execution cost-speed trade-off): Central dealers match buy-


ers with sellers more directly than peripheral dealers, leading to shorter interme-
diation chains. Central dealers offer immediacy by trading more on a principal
basis.

In the Internet Appendix to this paper, within a DGP-style model, we derive


conditions for a centrality premium in terms of investors’ outside option or so-
phistication, the relative bargaining power of the dealer and investor, and the
frequency and size of investors’ liquidity shocks, that is, their need for speed.
In the model, investors buy (sell) bonds from (to) dealers at prices that are set
through bilateral Nash bargaining when they meet. Investors are character-
ized by their sophistication and need for speed, as in Vayanos and Wang (2007).
Dealers are characterized by the intensity with which they search for counter-
parties, as in Neklyudov (2013), Weller (2013), and Üslü (2015). A centrality, or
speed, premium arises when investors have low negotiating sophistication and
a strong need for speed and dealers have sufficient bargaining power that they
can reap profits from intermediation. Dealer market power is thus a prerequi-
site for faster dealers to be able to command a premium over slower competitors
in bilaterally negotiated trades with investors.
Dealer market power is also important for allocations and welfare. Lagos
and Rocheteau (2007) show in a DGP setting with unlimited investor holdings
that some degree of market power is necessary to induce dealer entry and wel-
fare maximization. Weill (2007) shows that dealers with short-sale-constrained
inventory provide socially optimal liquidity during crises. By contrast, Lagos,
Rocheteau, and Weill (2011) find that dealer market power leads to socially
suboptimal liquidity provision, which provides a rationale for the government
to act as liquidity provider of last resort during stress times. Whether munic-
ipal bond dealers, or a subset thereof, provide sufficient liquidity at all times
is ultimately an empirical question. Given that central dealers seem to play a
more crucial role in liquidity provision than peripheral dealers, we predict that
central dealers act as liquidity providers of last resort during times of market
stress.

HYPOTHESIS 4 (Liquidity provider of last resort): Central dealers are liq-


uidity providers of last resort in that investors trade with more central dealers
during times of market stress.
100 The Journal of FinanceR

BK model information diffusion in OTC markets when market partici-


pants possess private information. Their network model predicts that better-
connected dealers are less exposed to adverse selection, as they observe more
order flow, and thus charge a smaller adverse-selection-related markup than
peripheral dealers. In addition to BK, Gofman (2011) and Malamud and Rostek
(2014) study pricing in networks under information frictions. We test predic-
tions from the information-based theory in BK in Section VII.

II. Data and Intermediation Chains


We use regulatory trade-by-trade data for the municipal bond market to test
the hypotheses developed in Section I. Below, we describe our data sources, the
filters used to cleanse the data of errors, and our procedure to map individual
OTC transaction legs to round-trip trades and intermediation chains.

A. Data Sources
Our main data source is the regulatory Transaction Reporting System au-
dit trail for municipal bonds collected by the MSRB. In an effort to improve
market transparency, the MSRB has required all broker-dealers to report their
transactions in municipal securities since 1998. The data are thus comprehen-
sive. Unlike the public version of historical municipal bond trades, our data
include identifiers for the broker-dealer firms intermediating each trade.9 For
customer trades, the data identify the dealer buying or selling the bonds. For
interdealer trades, the data identify the dealers on each side of the trade.10
We use these data to construct the trading networks of all broker-dealers each
day over the 15-year period from February 1998 to December 2012. The dealer
identifiers allow us to trace each bond as it flows from the selling customer
through one or several dealers to the ultimate buyer. The data also include
trade size, price, and time stamps for all trades up to the minute (in some cases
up to the second). Much of the analysis would not be possible without precise
time stamps. We use this information to trace the intermediation chains in
which the bonds flow through the network.
In addition to the comprehensive transactions data, we obtain reference in-
formation on all municipal bonds from Mergent. This information includes is-
suance date, maturity, coupon, taxable status, ratings, call features, issue size,

9 We use MSRB designations that aggregate subsidiaries with separate executing bro-

ker symbols (EBSs) of the same financial institution into a single MSRB identifier (see
http://www.msrb.org/msrb1/trsweb/dealers.asp). For example, Morgan Stanley Smith Barney LLC
has one MSRB ID for the three EBS IDs, DEAN, MSPW, and MSSB.
10 The data do not provide identifiers for the dealers’ customers. Afonso, Kovner, and Schoar

(2014), Hendershott and Madhavan (2015), and Hendershott et al. (2016) are recent studies em-
ploying customer IDs.
Dealer Networks 101

and issuer characteristics.11 We use several data sources to capture demand


pressure by investors and identify times of market stress. Weekly aggregate
municipal mutual fund flows come from the Investment Company Institute.
Data on investor flows to individual municipal bond funds come from the CRSP
Survivor-Bias-Free U.S. Mutual Fund database and security-level holdings in-
formation for municipal mutual funds is from Thomson Reuters eMaxx.12

B. Sample Construction
We filter the transactions to eliminate data errors and ensure data com-
pleteness. For a bond to be in our sample, it must have complete descriptive
data in Mergent and satisfy several trade-specific filters (time-stamped after
February 1998, four-letter alphabetic dealer identifier, no MSRB indicator for
away-from-market price, par value ≥$5,000) and bond-specific filters (fixed
or zero coupon, nonderivative, nonwarrant, not puttable, maturity ≥1 year,
$5,000 denomination). Appendix A provides a detailed description of our pro-
cedure and reports the number of observations affected at each step. We next
apply time- and price-based filters. Green, Hollifield, and Schürhoff (2007) and
Schultz (2012) document that trading and liquidity in newly issued bonds dif-
fer markedly from those for seasoned issues. Because the focus of our paper is
on the secondary market, we remove all trades during the first 90 days after
issuance and less than one year before maturity. We eliminate price outliers
by separately truncating the distribution at 0.5% and 99.5% for zero-coupon
bonds and all other bonds.
Our final sample consists of 72.2 million transactions in 1.28 million bonds
with distinct CUSIPs. Municipal bonds trade infrequently in the secondary
market. The median bond trades 12 times over our sample period, with 10% of
bonds trading just once or twice.13 A total of 2,238 broker-dealer firms (as iden-
tified by their MSRB designation) intermediate trades in our sample. Order
splitting is common. On average, each customer-to-dealer trade involves about
one interdealer trade and two dealer-to-customer trades. A total of 17.9 mil-
lion trades are customer(sell)-to-dealer(buy) trades, 20.2 million are between
dealers, and 34.1 million are dealer(sell)-to-customer(buy) trades.

C. Round-Trip Intermediation Chains


While municipal bonds trade infrequently, when they do trade, they do so in
round-trips. In a round-trip transaction, bond dealers intermediate by match-
ing buyers and sellers in a sequence of bilateral trades. A round-trip typically
starts with an investor who initially sells bonds to a dealer (CD leg), and then
11 We cross-checked with the Securities Data Company (SDC) Global Public Finance database.
While coverage is not the same, the results do not change materially if we switch to SDC reference
data.
12 See Jiang, Li, and Wang (2016) for more details about the two databases.
13 See Green, Hollifield, and Schürhoff (2007) for more discussion on secondary market trading

activity in municipals.
102 The Journal of FinanceR

the dealer sells the same bonds to another investor (DC leg) or other dealers
(DD leg(s)). The dealer may sell all of the bonds at once (we call this a Nonsplit
round-trip) or they may split the bonds into smaller lots (a Split round-trip).
The offsetting trades can occur through the same dealer, yielding a CDC round-
trip, or through different dealers after a sequence of interdealer trades. These
longer chains start with a tail dealer purchasing a bond from a customer, fol-
lowed by interdealer trades that move the bond to the tail dealer that sells the
bond to one or more customers. We call this a C(N)DC round-trip, where (N)
indicates that multiple dealers may be involved.14
Our round-trip match algorithm, described in Appendix B, builds on Green,
Hollifield, and Schürhoff (2007) and uses dealer identifiers to improve match
accuracy. The dealer identifiers allow us to trace the entire intermediation
chain by following each bond dealer-by-dealer through the network. By con-
trast, studies that do not know dealer identities are limited to focusing on the
two legs of trade that are between a dealer and customer (CD and DC). Our
round-trip concept explicitly follows bonds through the dealer network, which
not only improves matching accuracy but also allows us to record the com-
plexity of intermediation chains and compute both customer and interdealer
markups. We look for matching legs of CD, DD, and consecutive DC trades
(ordered by their time stamps) in the same CUSIP with matching dealer IDs
and same par size. For Split trades, we allow the original par size to be split
into smaller chunks along the intermediation chain, but order splitting may
occur only in the last leg of the chain—the DC leg—as otherwise the matching
becomes intractable.
Short selling is not very common in municipal bonds. To not rule out short
sales, and hence to capture them in the data, our algorithm allows for reverse
ordering in time of CD, DD, and DC trades for up to 10 days. While 2.1% of
all CDC round-trips are categorized as such, the time between the two legs is
short, with 0.84 (0.02) days on average (median). Our results are unchanged
when we eliminate these trades.
The round-trip matching is relatively straightforward, as trading in most
municipal bond issues is infrequent and clustered.15 The trade matching al-
gorithm yields a total of 11.4 million C(N)DC round-trip chains and 574,000
incomplete C(N)D chains.16 Dealer involvement varies across these chains.

14 We treat the classifications separately throughout the analysis. The baseline sample consists

of the CDC-Nonsplit round-trips for which we are the most certain that the bid- and ask-side trades
pair up exactly. The second sample allows for order splitting and thus includes all CDC round-trips.
This sample is more representative of a typical trade but may add some noise when split orders are
incorrectly assigned to be part of the same round-trip. The last sample comprises all the C(N)DC
trades, which include all matched transaction chains involving more than one dealer.
15 The algorithm matches a total of 82% of all CD trades to corresponding DD and DC trades.

The remaining trades are part of incomplete C(N)D chains in which we cannot unambiguously
assign each leg to a round-trip. C(N)D chains occur in issues with very frequent trading and in
very illiquid issues when the bond remains in the dealers’ inventory for more than a month. Our
results are not materially affected by including or excluding the C(N)D chains.
16 We eliminate trades between customers and dealers in which a dealer acts in the capacity of

an “agent,” as opposed to a principal (see Appendix B). Agency trades account for 6% of the sample,
Dealer Networks 103

The majority, or 8.8 million, are CDC trades in which a single dealer matches
a buyer with a seller. The simplest and most effective way in which a dealer
intermediates is a CDC-Nonsplit direct match in which the original bond lot
is not split into smaller order sizes. Of the total number of CDC trades, 6.3
million are CDC-Nonsplit and 2.5 million are CDC-Split.
Longer intermediation chains are common. Dealers involve other dealers
23% of the time in 2.6 million round-trips. We find up to seven dealers in the
sequence of trades, with two dealers in 1.5 million cases (CDDC), three dealers
in 0.87 million cases, and up to seven dealers in only 882 round-trips. We drop
the very few cases involving eight or more dealers.

III. Dealers’ Trading Networks


We start by documenting the structure of the dealer network in municipal
bonds. Two streams of theoretical work offer alternative predictions. An OTC
market with random search among homogeneous dealers and a competitive
interdealer market (e.g., DGP) results in a Poisson distribution for the de-
gree of connections between dealers, with these connections characterized as
transient. By contrast, a trading network with a core-periphery structure and
stable relations arises when the interdealer market is decentralized and bi-
lateral (Babus (2012), Chang and Zhang (2015), Neklyudov and Sambalaibat
(2015), Wang (2016)).
In this section, we test Hypothesis 1, that is, we examine whether the mu-
nicipal bond market has a stable core-peripheral structure. In particular, we
examine whether dealer connectivity appears random or is heavily skewed with
a fat right tail populated by a few core dealers, as well as whether connections
and centrality rankings are transient or persistent over time.

A. Core-Periphery Structure
Figure 1 illustrates the network of interdealer trades in municipal bonds.
Each node in the figure represents a broker-dealer firm, while each arrow cap-
tures directed order flow between two broker-dealers. In Panel A, two dealers
are connected if they traded more than 10,000 times over the sample period.
This threshold corresponds to about two or three trades per day and allows us
to isolate the most central dealers that form the core of the municipal bond
market.17 In Panel B, we plot the dealer network using all transactions.
There is tremendous heterogeneity in dealer connectivity. A small group of
about 10 to 30 dealers are highly interconnected, with more than 500 trading
partners, and trade heavily with one another. In contrast to these core dealers,

as these types of trades are compensated through commission and soft dollars, not markup. The
commissions are broker-specific and do not vary significantly over time. We have no access to
broker commissions data and therefore eliminate them from our study.
17 The plot is an artificial map generated by multidimensional scaling based on the criterion

that the more trade links that exist between two dealers, the closer their location.
104 The Journal of FinanceR

Panel A: Order flow among most active dealers


45431458 51138123
3442 5882

7361

33481975
6227
4569 2285 9469
1780
2121 2376 2213
1324 2140 2557
2266
5985 1394 1213
1450
1794
4538 4655 4736
7490 2007 1825 1066
1663
6731 8002
1476 3658
4661 2086
2167 1621
1875 4375
1971
4244 5253
2427 3241
2308 2057
3935
2482
1422
6588
6016 6429 1257 1571
2314 1471
1610 2004

6354
2001 1100
45855680 1712
2396
3077 7171

1483

Panel B: Order flow in entire network

jofi12728

Figure 1. Topology of the dealer network. The figure illustrates the network structure of
dealers in the municipal bond market in terms of the order flow between the dealers. Each node
represents a dealer firm. Each arrow represents directed order flow between a pair of dealers. In
Panel A, we impose the restriction that order flow between two dealers exceeds 10,000 transactions
over the sample period. In Panel B, we plot the dealer network using all transactions. The plots
are generated using multidimensional scaling based on the criterion that the more trade links
exist between two dealers, the closer is their location on the map. (Color figure can be viewed at
wileyonlinelibrary.com)
Dealer Networks 105

the remaining several hundred dealer firms interact with a limited number of
trading partners. The overall network is very sparse. Out of 5.1 million possible
directed links, only 124,000 (or 2.4%) exist over the 15-year period. Taken
together, the two panels in Figure 1 indicate that the municipal trading network
has a core-periphery structure.18 Ultimately, however, the overall structure is
complex and difficult to see from the figure.
To characterize the network topology more formally, in Panel A of Figure 2,
we compare the empirical distribution in the number of connections—that is,
the degree across dealers—to that of a random trading network with a Poisson
degree distribution. Using information from the direction of order flow, we
plot the empirical distribution for both in-degree, calculated from a dealer’s
purchases from other dealers, and out-degree, calculated from the dealer’s
sales to other dealers. Compared with a Poisson distribution that can arise in
a trading network formed through random search, the municipal interdealer
trading network has a higher level of heterogeneity among dealers in terms
of connectedness. The degree distributions, depicted on a log-log scale, follow
approximately power laws with exponential tails beyond about 100 dealers,
as the probability decreases almost linearly with connectedness up to this
point.19 The dealership network exhibits a heavy right tail, with 10% (5%) of
dealers having more than 100 (200) other dealers as trading partners. The
out-degree appears a bit more fat-tailed than the in-degree because of order
splitting.
The tax-induced segmentation of investors due to different tax treatment at
the state level of in-state and out-of-state bonds suggests that the municipal
trading network may be segmented as well, in which case it would not have
a single nationwide center but rather multiple local submarkets along state
lines. In Panel B of Figure 2, we explore this aspect of the dealer network by
documenting the local connectedness of dealers. One can infer the hierarchical
segmented structure of the dealer network by plotting the degree distribution
k across dealers (horizontal axis) against the clustering coefficient, or cliquish-
ness, cc of each dealer (vertical axis) (Ravasz et al. (2002)). The cliquishness is
a measure of the likelihood that two associates of a dealer are associates them-
selves. A higher value indicates greater cliquishness. The negative relation
between k and cc in the figure, again on a log-log scale, reveals that peripheral
dealers are part of local markets. We exploit this unique feature of the munici-
pal market when we explore the effect of dealers’ state-specific connectivity in
Section VII.D.

18 In much of the theoretical work on network formation, star-mesh or hub-and-spoke arise as


the Pareto-efficient structure. See Jackson (2005) and Babus (2012).
19 Scale-free networks that are often observed in the physical world are characterized by a

power-law degree distribution. Extending the random graph theory of Erdős and Rényi (1960),
Barabási and Albert (1999) show that preferential attachment leads to scale-free networks. The
municipal bond market shares features with such scale-free networks.
106 The Journal of FinanceR

Panel A: Market connectedness

Panel B: Local market structure

Figure 2. Core-periphery network structure. The figure documents the structure and re-
silience of the municipal bonds market. Panel A explores the market connectedness and the
(non-)randomness of trading relations between dealers. We show the inverse distribution function
for the degree across dealers in the network on a log-log scale. The dots correspond to out-degrees,
and the triangles represent in-degrees. We add for comparison the degree distribution of a random
trading network with the same average degree (dashed line). Panel B explores the local clustering
and hierarchical structure of the market. We plot the degree distribution across dealers in the net-
work (horizontal axis) against the clustering coefficient of each dealer (vertical axis) on a log-log
scale. (Color figure can be viewed at wileyonlinelibrary.com)
Dealer Networks 107

B. Dealer Centrality
We capture a dealer’s local connectivity and global importance using vari-
ous network statistics. At least four measures of centrality are widely used in
network analysis: degree centrality, eigenvector centrality, betweenness, and
closeness. The simplest measure is the dealer’s degree centrality, which cap-
tures the local connectivity of a dealer in the network by computing the fraction
of dealers with which the dealer trades directly. One can also make use of the
trading directions among dealers and calculated in/out degrees from the di-
rected network graph. In-degree counts the number or fraction of dealers that
sell to a given dealer, while out-degree counts the number or fraction of dealers
that buy from the dealer. Related to the degree is the k-core, which is the maxi-
mal subnetwork in which each dealer has degree of at least k. Degree centrality
considers only the local connectivities of a dealer (node). Eigenvector centrality
captures global importance by assigning scores to all dealers in the network
based on the idea that connections to high-scoring dealers contribute more to
the score of the dealer than equal connections to low-scoring dealers. Eigen-
vector centrality takes all direct and indirect trading partners into account.
Betweenness is similar in that it counts the number of the shortest interme-
diation chains linking any two dealers in the network that pass through the
dealer. Closeness is the inverse of the average number of steps that a dealer
needs to take within the network to reach or be reached by any other dealer,
based on prior trading relations. Appendix C provides a detailed description of
each measure and explains how we construct weighted and unweighted vari-
ants.
In Table I, Panel A, we provide summary statistics for the dealers’ network
centrality measures. To track the evolution of trading networks, we construct
the networks on a rolling basis each day in the sample period by aggregating all
trades between the dealers over the past 30 days. We lag the resulting measures
by one day. Over the 3,757 trading days, 2,099 dealers trade with other dealers
in at least one 30-day period, generating a nonmissing centrality measure in
at least one trading day, which yields 2.74 million dealer-day observations.
The statistics for out- and in-degree (normalized between 0 and 1) show that
dealers are connected to 1% of all dealers on average. At the 95% percentile,
the most connected dealers have links with 6% to 7% of all other dealers,
with a maximum of 30% for sales to other dealers (out-degree) and 25% for
purchases from other dealers. Betweenness peaks at 18%, closeness at 2%,
and the k-core at 30. For robustness, we construct equal- and value-weighted
centrality measures, where we weight each connection by the number of trades
or by the value of the order flow between the dealers. We find that all of the
centrality measures are highly skewed, with many peripheral and few central
dealers/dealer-days.
Due to the core-periphery structure, both the local and global measures are
highly correlated. Our results are therefore robust to the exact definition of
dealer centrality used. We employ an aggregate index in our analysis. Specifi-
cally, we aggregate the network variables into a single centrality index, denoted
108 The Journal of FinanceR

Table I
Summary Statistics for Dealer Centrality
Panel A reports descriptive statistics for the dealer centrality measure Net and the individual
centrality measures in the pooled dealer-day sample. Net (EW) is the equal-weighted and Net (VW)
the order flow-weighted centrality measure. Panel B reports correlations between the centrality
measure Net and the individual centrality measures in the pooled dealer-day sample. The number
of dealer-day observations is 2.74 million.

Panel A: Summary Statistics

Statistic Mean SD 25% 50% 75% 95% Max

Dealer centrality (equal-weighted measures):


Out-degree dgout 0.01 0.03 0.00 0.00 0.01 0.07 0.30
In-degree dgin 0.01 0.02 0.00 0.00 0.01 0.06 0.25
Eigenvector centrality ev 0.09 0.15 0.01 0.02 0.08 0.43 1.00
Betweenness bt 0.00 0.01 0.00 0.00 0.00 0.01 0.18
Closeness out clout 0.01 0.01 0.00 0.01 0.01 0.02 0.02
Closeness in clin 0.00 0.00 0.00 0.00 0.01 0.01 0.01
k-core out kcoreout 4.39 6.72 0.00 1.00 5.00 20.00 30.00
k-core in kcorein 5.49 6.86 1.00 2.00 8.00 21.00 30.00
Dealer centrality (value-weighted measures):
Out-degree weighted by no. of 0.11 0.89 0.00 0.00 0.02 0.34 49.86
trades dgoutwntrade
Out-degree weighted by par 0.14 1.03 0.00 0.00 0.02 0.44 79.73
flow dgoutwpar
In-degree weighted by no. of 0.11 0.83 0.00 0.01 0.03 0.31 46.58
trades dginwntrade
In-degree weighted by par 0.14 0.95 0.00 0.00 0.03 0.41 65.52
flow dginwpar
Eigenvector weighted by no. 0.01 0.06 0.00 0.00 0.00 0.03 1.00
of trades evwntrade
Eigenvector weighted by par 0.01 0.06 0.00 0.00 0.00 0.02 1.00
flow evwpar

Panel B: Correlation Matrix

Statistic Net (EW) Net (VW)

Dealer centrality Net (EW) 1.00 0.47


Out-degree dgout 0.94 0.54
In-degree dgin 0.94 0.49
Eigenvector centrality ev 0.98 0.50
Betweenness bt 0.73 0.42
Closeness out clout 0.32 0.08
Closeness in clin 0.13 0.03
k-core out kcoreout 0.88 0.33
k-core in kcorein 0.84 0.29
Dealer centrality Net (VW) 0.47 1.00
Out-degree weighted by no. of trades dgoutwntrade 0.41 0.88
Out-degree weighted by par flow dgoutwpar 0.39 0.88
In-degree weighted by no. of trades dginwntrade 0.40 0.84
In-degree weighted by par flow dginwpar 0.42 0.91
Eigenvector weighted by no. of trades evwntrade 0.43 0.79
Eigenvector weighted by par flow evwpar 0.39 0.81
Dealer Networks 109

by Net jt , for each dealer j = 0, . . . , J and day t by taking the first principle
component across all centrality measures. Appendix C provides details on this
procedure. To normalize centrality between 0 and 1 while keeping the origi-
nal ordering, we apply an empirical cumulative distribution function (ECDF)
transformation to each network variable.20 The ECDF transformation reduces
the effect of skewness and outliers and facilitates interpretation of economic
magnitudes. A one-unit change in centrality corresponds to moving from the
most peripheral dealer (Net = 0) to the most central dealer (Net = 1). In Panel
B of Table I, we report the correlation coefficients between each network char-
acteristic and Net for the equal- and value-weighted variants. As can be seen,
Net captures a sizable common component.

C. Stability in Dealer Relations


To document the persistence in interdealer trading relations, in Panel A
of Table II, we compare the trading relationship between any given pair of
dealers from one month to the next. As can be seen, if two dealers did not trade
last month, there is about a 99% chance that they will not trade with each
other this month. Conditional on a directional trade relation existing in the
prior month, trades in the same direction occur with 63% probability in the
next month. Ignoring trade directions, the probability that two dealers who
traded last month also trade this month is 66%. To put these numbers into
perspective, the probability would be 1.4% in a random network with the same
number of dealers and interdealer trades. In Panel B of Table II, we focus on the
persistence in dealer ranks, as measured by the ordering of dealers’ centrality
Net. Dealer ranks are also highly persistent from one month to the next. The
top 10 dealers remain in the top 10 ranking 93% of the time.

IV. Centrality Premium


Given that the dealer network in municipal bonds has a core-periphery struc-
ture (Section III), we now test Hypothesis 2, that is, we examine whether
trading costs depend on dealer centrality, as predicted by models of trading
networks (Neklyudov (2013), Hugonnier, Lester, and Weill (2015), Üslü (2015),
Babus and Kondor (2018)). We measure investor trading costs by the (gross and
net) markups charged by dealers
 on DC round-trips.21 Dealers’ total markup on
round-trip i (with size Pari = j Pari j ) is defined as the par-weighted selling
prices PiDC in the DC transactions, net of the purchase price PiC D from the

20 As an alternative, we use the share of interdealer trading volume as a weight in the ECDF

transformation so that, in this case, a centrality equal to Net means that dealers with 1−Net
market share are more central than the dealer. The results are not materially affected.
21 We thank a referee for pointing out that a round-trip cost is a good measure of the cost of

intermediation only in a market that is always efficient, that is, assuming that the final buyer has
the highest valuation for the bond. In a real OTC market, the round-trip costs that we measure
are a lower bound on the total costs of intermediation.
110 The Journal of FinanceR

Table II
Stability in Dealer Relations and Persistence in Dealer Ranks
Panel A reports the transition probability matrix for dealer relations from one month to the next.
The transition matrix is calculated separately for unconditional relations between dealers and
relations conditional on the direction of the order flow. The row headings indicate whether a pair
of dealers traded with each other in a given month or not. The column headings indicate whether
the same trade relation persists in the next month. Panel B documents the persistence on dealer
ranks over time. We report the transition matrix of dealer rank categories from one month to the
next. Dealer ranks are measured by the ordering of their centrality measure Net. To compute the
dealer rank in a given month, we use all interdealer trades during the past 30 trading days.

Panel A: Stability in Dealer Relations

Order Flow in Same


Order Flow Next Month Direction Next Month
Order Flow
This Month =0 >0 =0 >0

=0 0.98 0.02 0.99 0.01


>0 0.34 0.66 0.37 0.63

Panel B: Persistence in Dealer Ranks

Rank Month t + 1

Top 10 11 to 20 21 to 50 51 to 100 101 to 200 >200

Rank month t Top 10 0.93 0.07 0.00 0.00 0.00 0.00


11 to 20 0.07 0.78 0.14 0.00 0.00 0.00
21 to 50 0.00 0.05 0.81 0.14 0.00 0.00
51 to 100 0.00 0.00 0.08 0.79 0.13 0.00
101 to 200 0.00 0.00 0.00 0.06 0.79 0.15
>200 0.00 0.00 0.00 0.00 0.03 0.97

customer in the initial CD transaction, all normalized by the original purchase


price:22
1 
j Pari j Pi j − Pi
DC DC CD
Pari
Markup Mi = ∗ 100. (1)
PiC D
When multiple dealers intermediate in C(N)DC chains, we compute interdealer
markups using the dealer-to-dealer transaction prices PiDD to see how the
dealers split the total markup.

22 Net markup is gross markup minus the return on the Bond Buyer index. The net differs from

the gross markup by less than three basis points on average with a correlation of 0.9. They are
close because inventory times are short and interest rates do not move much from one day to the
next. We report results for gross markups. Results are unchanged when we compute markups net
of interest rate changes.
Dealer Networks 111

Table III
Dealer Markups on Round-Trip Transactions
The table reports descriptive statistics for dealer markups on round-trip transactions for different
types of trades. We document trading costs for different samples, trade sizes, and intermediation
chain lengths. The largest sample, C(N)DC, combines all round-trips that involve no more than
seven dealers. The CDC sample pools all round-trips without interdealer trading. The base sample
for our analysis, CDC-Nonsplit, comprises all round-trips without interdealer trading and order
splitting. Agency trades in which dealers act as customers’ agents are eliminated. All markups
are measured in percent of the head dealer’s purchase price from the customer, as defined in
expression (1).

Trade Type N Mean SD 5% 25% 50% 75% 95%

C(N)DC 11,404,321 2.02 1.54 0.10 0.90 1.93 2.92 4.50


CDC 8,808,119 1.92 1.39 0.08 0.84 1.90 2.84 4.17
CDC-Nonsplit 6,294,447 1.85 1.39 0.06 0.75 1.79 2.78 4.14
CDC-Split 2,513,672 2.08 1.39 0.11 1.08 2.08 2.95 4.24
C(N)DC, grouped by chain length:
CDC 8,808,119 1.92 1.39 0.08 0.84 1.90 2.84 4.17
CDDC 1,511,196 2.23 1.71 0.22 0.98 1.94 3.12 5.26
CDDDC 866,450 2.45 1.96 0.31 1.13 2.08 3.31 5.81
CDDDDC 173,579 3.12 2.52 0.41 1.46 2.65 4.13 7.39
CDDDDDC 37,229 3.44 2.83 0.45 1.64 2.85 4.54 8.29
CDDDDDDC 6,866 3.62 3.05 0.54 1.77 2.97 4.66 8.86
CDDDDDDDC 882 3.68 3.08 0.53 1.82 2.96 4.62 9.20
CDC-Nonsplit, grouped by trade size:
<25K 3,240,898 2.28 1.34 0.36 1.41 2.24 3.06 4.38
25K to 49K 1,398,791 1.82 1.28 0.19 0.85 1.69 2.62 4.00
50K to 99K 743,950 1.43 1.26 0.09 0.45 1.16 2.18 3.65
100K to 249K 525,517 0.97 1.12 0.00 0.18 0.57 1.48 3.10
250K to 499K 119,972 0.58 0.82 0.00 0.09 0.27 0.83 2.15
500K to 999K 68,967 0.42 0.70 0.00 0.05 0.18 0.56 1.75
≥1M 196,352 0.18 0.56 −0.04 0.03 0.09 0.22 0.80
Incomplete round-trip chains:
C(N)D 573,549 0.79 1.16 0.00 0.13 0.50 1.06 2.62

Table III reports descriptive statistics on trading costs.23 Trading costs on


municipal bonds are large on average, consistent with Harris and Piwowar
(2006) and Green, Hollifield, and Schürhoff (2007). Average round-trip trading
costs are 2.0% across all C(N)DC chains. Dealers earn 1.8% on nonsplit round-
trips and 2.1% on split round-trips. Total costs for investors are higher, the more
dealers are involved in a round-trip. Average markups increase monotonically
but at a declining rate with the number of dealers intermediating the chain,
consistent with evidence in Schultz (2012) on newly issued bonds, from 1.9%
on average when one dealer is involved to 3.7% with seven dealers involved.
The markups dealers charge other dealers in DD legs are substantially smaller
than the markups they charge investors. As documented in prior studies, av-
erage markups decline monotonically with transaction size, with retail trades

23 In reporting these numbers, we apply no data filters. For the regression analysis performed

later, we eliminate extreme values by truncating the distribution at 0.5% and 99.5%.
112 The Journal of FinanceR

between $25K and $50K costing 1.8% on average, whereas institutional trades
of $1M or more cost 18 basis points on average. The large standard deviations
show that dealer markups vary substantially across round-trips even if the
number of dealers involved and the size of the transaction are fixed.

A. Dealer Centrality and Markups


In Table III, we find that dealer markups vary substantially across round-
trips, even if the number of dealers involved and the size of the transaction are
fixed. We now show that markups are systematically related to the centrality
of the dealer intermediating the trade. We first confirm this relationship using
double-sorted univariate comparisons. We then show that this relationship
continues to hold in a multivariate regression setting that controls for a host
of explanatory variables.
In Table IV, we employ double-sorting across dealer tiers and trade-size
groups for different round-trip samples. We construct dealer tiers by sorting
dealers each day on their centrality Net jt . We assign the dealers with the
highest centrality and cumulative market share of 25% (top quartile) to the
central dealer tier (about 0.7% of all broker-dealer firms), the next quartile to
the midtier, and the remainder to peripheral dealer tier. We map trade size
into ranges that proxy for investor types. Small lots with a par size less than
$100K typically originate from retail investors. Medium institutional lots are
between $100K and $1M. Blocks of $1M or more likely come from taxable
institutions, regardless of whether the institution is a mutual fund or other
taxable institution.24
In Table IV, Panel A, we find that central dealers charge larger trading costs
than peripheral dealers across all trade sizes, despite their higher matching
efficiency. Average markups differ by up to 100%. For trades of $100K to $249K,
markups at central dealers are 1.5%, double the 0.7% at peripheral dealers. The
difference is smaller for larger trades but still positive and both economically
and statistically significant. For $1M+ trades, the centrality premium is 4 basis
points, or 20% over the markups charged by peripheral dealers. When we look at
the entire markup distribution, we find that not only do average markups differ
across dealers, but also the composition of trades. While peripheral dealers
intermediate many round-trips at markups that are positive but close to zero,
central dealers rarely do such trades. The body of the markup distribution is

24 These choices come out of quantitative analysis and conversations with practitioners. We

chose the size cutoff at $100K for retail trades compared with institutional trades for the following
reasons. Existing literature, such as Green, Hollifield, and Schürhoff (2007), uses $100K as a cutoff
for retail trades after talking to industry experts. We also explored the eMaxx data on municipal
mutual fund holdings and the NAIC data on insurance company trades to get some hard evidence.
Consistent with the notion that retail is active in par amounts less than $100K and mutual funds
above, we find that fewer than 15% of the liquidation trades by mutual funds are of sizes less
than $100K. Using NAIC data on insurance company trades in municipal bonds, we find that
fewer than 10% of trades are of sizes less than $100K. We chose the size cutoff of $1M for block
trades, consistent with the fact that MSRB considers trades greater than $1M as “large trades”
and therefore these trades are subject to size masking in reporting.
Dealer Networks 113

Table IV
Trading Costs and Dealer Centrality: Univariate Analysis
The table reports trading costs (Panel A) and loss probabilities (Panel B) by the centrality of
the dealer intermediating the trade. We report average markups on round-trip transactions and
the fraction of round-trip transactions with markdowns for different dealer tiers. We construct the
dealer tiers by sorting the dealers on the equal-weighted centrality measure Net. We assign the
top quartile of dealers to the central tier, the next quartile to the midtier, and the remainder to
peripheral dealers. The base sample, CDC-Nonsplit, comprises all round-trips without interdealer
trading and order splitting. The CDC sample pools all round-trips without interdealer trading.
The largest sample, C(N)DC, combines all round-trips that involve no more than seven dealers.
Agency trades in which dealers act as customers’ agents are eliminated. All markups are measured
in percent of the head dealer’s purchase price from the customer, as defined in expression (1).
Significance levels are indicated by * (10%), ** (5%), and *** (1%).

Panel A: Trading Costs by Dealer Centrality

Average Markup by the Dealer(s) (in Percent)

Peripheral Midtier Central Central


Trade Type Dealers Dealers Dealers Peripheral

C(N)DC 1.94 1.92 2.22 0.29***


CDC 1.79 1.87 2.21 0.42***
CDC-Nonsplit 1.73 1.81 2.15 0.42***
CDC-Split 1.96 2.01 2.36 0.39***
CDC-Nonsplit, grouped by trade size:
<25K 2.32 2.11 2.43 0.11***
25K to 49K 1.68 1.76 2.08 0.40***
50K to 99K 1.17 1.49 1.81 0.64***
100K to 249K 0.71 1.15 1.47 0.76***
250K to 499K 0.51 0.69 0.90 0.40***
500K to 999K 0.41 0.48 0.59 0.19***
≥1M 0.19 0.20 0.23 0.04***

Panel B: Trading Losses by Dealer Centrality

Average Loss Probability for the Dealer(s) (in Percent)

Peripheral Midtier Central Central


Trade Type Dealers Dealers Dealers Peripheral

C(N)DC 1.89 1.66 1.08 −0.81***


CDC 2.13 1.65 1.03 −1.10***
CDC-Nonsplit 1.73 1.46 0.82 −0.90***
CDC-Split 3.24 2.10 1.49 −1.75***
CDC-Nonsplit, grouped by trade size:
<25K 1.10 1.27 0.58 −0.52***
25K to 49K 1.38 1.25 0.82 −0.56***
50K to 99K 1.83 1.28 0.94 −0.89***
100K to 249K 2.81 1.74 1.29 −1.52***
250K to 499K 2.69 2.40 1.96 −0.73***
500K to 999K 3.17 2.90 2.88 −0.29
≥1M 6.21 7.63 5.84 −0.37***
114 The Journal of FinanceR

shifted to the right. Central dealers trade mostly at markups away from zero.
Thus, the centrality premium is not driven by outlier cases with excessively
large markups.
Table IV, Panel B reports the frequency with which dealers lose money on a
round-trip trade, grouped by the dealer’s network centrality. Overall, dealers
lose money in less than 2% of all trades. Importantly, central dealers are signif-
icantly less likely than peripheral dealers to lose on round-trips. It is therefore
not obviously the case that central dealers charge larger bid-ask spreads on av-
erage to compensate for the higher risk of losing money on a trade. We revisit
this issue in more detail in Section VII.B.
To account for other characteristics of the round-trips that may vary system-
atically across trade-size groups or dealer tiers, we run multivariate regressions
that explore how trading cost Mi in round-trip i is affected by dealer centrality
and other explanatory variables:

Markup Mi = αi + δ Neti + β  Xi + εi , (2)

where αi are state and month fixed effects, Neti is the centrality of the first
dealer in the chain, and Xi is a vector of control variables described in Ap-
pendix D that include information specific to the trade, bond, issuer, and dealer.
In particular, the trade characteristics are the par value of the trade interacted
with trade size group dummies. The bond and issuer characteristics include
credit quality, remaining maturity in years, age of the bond issue, issue size,
call feature, general obligation bond, sinking fund, bank qualification, tax sta-
tus, and whether the bond is subject to the alternative minimum tax (AMT).
The markups that dealers charge may include compensation for additional
services that broker-dealers provide, which could include pricing, underwrit-
ing, and research services. In addition, theory predicts and empirical evidence
suggests that inventory management affects bid-ask spreads (Friewald and
Nagler (2016)). To capture these effects, we include several dealer character-
istics: a dummy variable for whether the dealer is the lead underwriter of the
bond, indicators for primary dealer and New York City (NYC) headquarters,
dealer size, and the dealer’s abnormal aggregate inventory. We also allow for
state and month fixed effects and adjust standard errors for heteroskedasticity
as well as clustering by issuer and time.
The key variable of interest, Net, is computed on a 30-day rolling basis and
lagged by an extra day to not overlap with the round-trip (see Section III.B).
The coefficient δ measures the difference in trading costs between the most
peripheral dealer (Net = 0) and the most central dealer (Net = 1). Across
columns, we vary the centrality measure and the trade categories. In columns
(1), (3), and (5), the dealer centrality measure Net is defined as the first principle
component of the equal-weighted centrality proxies. In the remaining columns,
we use the order flow-weighted centrality proxies. The regression samples are
CDC-Nonsplit, CDC, and C(N)DC.
Table V summarizes the regression results, with alternate specifications re-
ported across columns. The estimates reveal that, consistent with Table IV,
Dealer Networks 115

Table V
Trading Costs and Dealer Centrality: Multivariate Analysis
The table reports the determinants of round-trip trading costs. The estimates are obtained from
panel regressions with state and month fixed effects. The determinants include the dealer char-
acteristics, trade characteristics, issue characteristics, and issuer characteristics described in Ap-
pendix D. The dealer centrality measure is the first principal component of the network variables
in Appendix C. The EW (VW) columns employ the equal-weighted (value-weighted) dealer cen-
trality measures. We vary the regression sample across columns, considering three types of trades
with varying dealer involvement. CDC-Nonsplits are round-trips intermediated by a single dealer
where the original bond lot is not split (N = 6,294,447). The CDC sample includes all round-trips
intermediated by a single dealer (N = 8,808,119). C(N)DC are round-trips intermediated by one or
several dealers (N = 11,404,321). Standard errors are adjusted for heteroskedasticity and double
clustered by issuer and time. Significance levels are indicated by * (10%), ** (5%), and *** (1%).

CDC-Nonsplit CDC C(N)DC

EW VW EW VW EW VW
Determinant (1) (2) (3) (4) (5) (6)

Dealer centrality 0.56*** 0.42*** 0.72*** 0.54*** 0.68*** 0.57***


Chain length 1.06*** 1.15***
Chain length*Centrality −0.78*** −0.87***
Primary dealer −0.02 −0.01 −0.09*** −0.08*** −0.09*** −0.09***
NYC dealer 0.01 0.01 0.09*** 0.09*** 0.04*** 0.04***
Underwriter 0.02 0.02 0.02 0.02 0.01 0.01
Dealer size −0.09*** −0.09*** −0.09*** −0.09*** −0.12*** −0.12***
Dealer inventory −0.01 −0.01 −0.02** −0.02** 0.00 0.00
log(Par)*Small −0.34*** −0.34*** −0.24*** −0.24*** −0.27*** −0.27***
log(Par)*Medium −0.37*** −0.37*** −0.26*** −0.26*** −0.29*** −0.29***
log(Par)*Large −0.35*** −0.36*** −0.30*** −0.30*** −0.31*** −0.31***
Maturity 0.66*** 0.66*** 0.75*** 0.75*** 0.79*** 0.79***
Seasoning −0.06*** −0.06*** −0.06*** −0.06*** −0.05*** −0.05***
Issue size −0.00** −0.00** −0.00*** −0.00*** −0.01*** −0.01***
Rating 0.07*** 0.07*** 0.07*** 0.07*** 0.08*** 0.08***
Junk rated −0.05 −0.05 −0.05 −0.05 0.04 0.04
Unrated 0.26*** 0.26*** 0.28*** 0.28*** 0.37*** 0.37***
Insured 0.08*** 0.08*** 0.08*** 0.08*** 0.05*** 0.05***
General obligation −0.01 −0.01 −0.01 −0.01 −0.04*** −0.04***
Callable −0.24*** −0.24*** −0.24*** −0.24*** −0.26*** −0.26***
Sinking fund −0.09*** −0.09*** −0.08*** −0.09*** −0.08*** −0.08***
Bank qualified 0.08*** 0.08*** 0.03*** 0.03*** 0.03*** 0.03***
Taxable bond 0.04 0.03 0.07* 0.06 0.11*** 0.10***
Subject to AMT 0.11*** 0.10*** 0.11*** 0.10*** 0.13*** 0.13***
R2 0.369 0.368 0.358 0.357 0.354 0.354

dealer characteristics—particularly their centrality—are important sources of


variation in trading costs. Holding trade size and bond characteristics constant,
trading costs are strongly positively related to dealer centrality. Highly inter-
connected dealers are able to charge customers 0.4% to 0.7% larger markups,
up to twice as large, on average.
Being lead underwriter on the bond issue has little effect on trading costs. As
one would expect, primary and larger dealers charge lower markups, but their
116 The Journal of FinanceR

effects are small. Size and centrality thus measure different aspects of a dealer.
We also find evidence of inventory effects. Specifically, dealers charge less the
larger their abnormal inventory. This result is consistent with Huang and Stoll
(1997) and Friewald and Nagler (2016), who show that dealers adjust their bid
and ask to revert inventory toward a target. Inventory effects alone, however,
do not explain the positive sign on centrality, as central dealers hold larger
inventories than peripheral dealers. The last two columns in Table V show
that, consistent with Table III, costs are larger the more dealers are involved
in the chain, but this effect is smaller the more central is the head dealer (we
define chain length in the regressions as the number of dealers minus one, so
it is zero for a CDC).

V. Dealers’ Network Centrality and Search Efficiency


Above we show that dealers vary substantially in their trading relationships
(Section III) and in the markups they charge to investors (Section IV). In this
section, we test Hypothesis 3, that is, we examine whether dealers differ in their
ability to locate trading counterparties, and ultimately to find buyers of bonds,
which would explain their ability to charge different markups and lead to an
execution cost-speed trade-off. To explore the link between dealer centrality
and search efficiency, we first document how dealers route bonds through the
network. We then show how the propensity of dealers to sell bonds to investors,
rather than to pass them on to other dealers, depends on centrality while
controlling for dealer size and other characteristics. Finally, we document how
dealer centrality affects the complexity and length of intermediation chains.

A. Order Flow Routing Occurs through Central Hubs


Table VI reports the average network centrality of each dealer in the inter-
mediation chain for C(N)DC trades. Recalling that Net = 0 (1) corresponds to
the least (most) central dealer, we find that bonds flow from periphery to center
and partially back. Dealers in the middle of a chain are more central than ei-
ther the dealer purchasing the bond from the customer or the dealer ultimately
selling the bond to the customer. Dealer centrality peaks with the second and
penultimate dealers for all types of C(N)DC trades. Central dealers thus act as
hubs by redistributing the order flow.
To more directly assess whether search and matching efficiency in finding
a buyer for a bond gives rise to the importance of central dealers as hubs, we
check how the complexity of intermediation, as measured by the number of
dealers in a chain, is affected by the network centrality of dealers in the trade
sequence. If central dealers are better able to locate investors, the centrality of
the (head) dealer intermediating the trade should be a significant determinant
of intermediation chain length. When a central dealer buys a bond from an
investor, the dealer is likely to know a counterparty or to find one more quickly.
Table VI
Order Flow Routing
The table reports the average network centrality for each dealer in round-trip chains of different length. We restrict the sample to round-trips that
involve no more than seven dealers. Agency trades, in which dealers act in the capacity of agent for an investor, are eliminated. We measure the
dealer centrality Net by the first principle component of the centrality proxies described in Appendix C, standardized by the empirical cdf. Net = 0
(1) corresponds to the most peripheral (central) dealer. Columns correspond to the position of the dealer in the round-trip chain. Rows correspond to
round-trip chains of given length. In parentheses, we report the difference in centrality between two consecutive dealers. Standard errors are adjusted
for heteroskedasticity and double clustered by issuer and time. Significance levels are indicated by * (10%), ** (5%), and *** (1%).

Dealer Centrality in Round-Trip Chain, #=Position in Chain

Trade Type N #1 #2 #3 #4 #5 #6 #7

CDC 8,808,119 0.96 . . . . . .


. . . . . .
CDDC 1,511,196 0.92 0.93 . . . . .
(0.01***) . . . . .
CDDDC 866,450 0.90 0.97 0.90 . . . .
Dealer Networks

(0.07***) (−0.07***) . . . .
CDDDDC 173,579 0.85 0.94 0.94 0.85 . . .
(0.09***) (−0.00) (−0.09***) . . .
CDDDDDC 37,229 0.86 0.95 0.89 0.95 0.83 . .
(0.10***) (−0.06***) (0.05***) (−0.11***) . .
CDDDDDDC 6,866 0.85 0.94 0.91 0.85 0.95 0.78 .
(0.09***) (−0.03***) (−0.06***) (0.09***) (−0.16***) .
CDDDDDDDC 882 0.84 0.94 0.87 0.90 0.88 0.94 0.79
(0.12***) (−0.08***) (0.03**) (−0.02*) (0.07***) (−0.15***)
117
118 The Journal of FinanceR

By contrast, when a peripheral dealer buys a bond, the dealer may pass the
bond on to a central dealer.25 We explore this issue in two ways.

B. Finding a Counterparty Is Easier for Central Dealers


We first examine the tendency of dealers to sell directly to investors in-
stead of passing bonds on to other dealers, and how this propensity is affected
by dealer centrality. Specifically, we estimate a panel Probit specification for
dealer-to-customer (DC Tradei = 1) versus interdealer (DC Tradei = 0) trades,
given that a trade occurred. Here, we allow for state and month fixed effects
and we adjust standard errors for heteroskedasticity as well as clustering by
issuer and time:

Pr(DC T radei | T radei ) = (αi + δ Neti + β  Xi + εi ), (3)

where  is the cumulative distribution of the standard normal distribution,


αi are state and month fixed effects, Neti is the lagged centrality of the dealer
in leg i selling the bond (where i denotes each trade in a round-trip chain),
and Xi is a vector of control variables described in Appendix D. These controls
include information specific to the trade, bond, issuer, and dealer. In particular,
the trade characteristics are the par value of the trade interacted with trade
size group dummies. The bond and issuer characteristics include credit quality,
remaining maturity in years, age of the bond issue, issue size, call feature, gen-
eral obligation bond, sinking fund, bank qualification, tax status, and whether
the bond is subject to the AMT.26 Dealer characteristics include dummies for
bond underwriter, primary dealer, and headquarters in NYC (dealers head-
quartered in NYC are likely to be national dealers, not regional), as well as the
total asset base of the dealer.
Table VII, columns (1) and (2), report results on how dealers route the order
flow. Central dealers are significantly more likely than peripheral dealers to
find an investor to buy the bond. The bond’s underwriter, primary dealers, small
dealers, non-NYC-headquartered dealers, and dealers with large abnormal
inventory are also more likely to sell to investors than other dealers, all else
being equal. The baseline probability of finding a buyer for an average bond
is 41%. The coefficient of 0.58 on Net corresponds to a marginal effect of 51%,
which raises this probability to about 60% for the most central dealers.

25 We thank the referee for pointing out that the chain length should be small if central dealers

are better at finding customers. But if, in equilibrium, trades go through the center of the network,
the length should be higher than in a purely bilateral market or in a market with CDC trades only.
It is ultimately an empirical question if centrality is a significant determinant of intermediation
chain length. In our empirical tests, we compare chain length across trades that arrive at the core
compared to the periphery.
26 In alternative specifications without calendar time controls, we include the VIX index, LIBOR

rate, and aggregate new issuance volume during the prior week. The results are very similar.
Dealer Networks 119

Table VII
Search Efficiency, Chain Length, and Dealer Centrality
The table documents dealers’ propensity to trade with customers and the length of intermediation
chains, and relates them to dealers’ network centrality. In the first two columns, we examine how
the frequency of trading with a customer as opposed to another dealer depends on the centrality
of the dealer. The estimates are obtained from Probit regressions with state and month fixed
effects. The number of observations is 27,749,874. In the last four columns, we examine how
the path length of round-trip chains is related to the network centrality of the head dealer. The
estimates are obtained from panel regressions and, alternatively, from Poisson regressions with
state and month fixed effects. The dealer centrality measure Net is the first principal component
of the network variables in Appendix C. The EW (VW) columns employ the equal-weighted (value-
weighted) dealer centrality measures. The sample consists of all C(N)DC round-trip transactions
(N = 11,404,321). Standard errors are adjusted for heteroskedasticity and double clustered by
issuer and time. Significance levels are indicated by * (10%), ** (5%), and *** (1%).

No. of Dealers in Round-Trip Chain

Pr(Sale to Customer
| Sale) OLS Poisson

EW VW EW VW EW VW
Determinant (1) (2) (3) (4) (5) (6)

Dealer centrality 0.58*** 0.50*** −1.23*** −0.95*** −0.46*** −0.36***


Primary dealer 0.04*** 0.05*** −0.07*** −0.09*** −0.03*** −0.04***
NYC dealer −0.03*** −0.03*** 0.08*** 0.07*** 0.03*** 0.03***
Underwriter 0.09*** 0.09*** −0.16*** −0.16*** −0.07*** −0.07***
Dealer size −0.08*** −0.08*** 0.21*** 0.20*** 0.07*** 0.07***
Dealer inventory 0.02*** 0.02*** −0.05*** −0.05*** −0.02*** −0.02***
log(Par)*Small −0.01*** −0.01*** 0.02*** 0.02*** 0.01*** 0.01***
log(Par)*Medium −0.01*** −0.01*** 0.00** 0.00*** 0.00** 0.00***
log(Par)*Large 0.01*** 0.01*** −0.02*** −0.02*** −0.01*** −0.01***
Maturity −0.00*** −0.00*** 0.00 0.00 0.00 0.00
Seasoning −0.01*** −0.01*** 0.02*** 0.02*** 0.01*** 0.01***
Issue size 0.00*** 0.00*** −0.00*** −0.00*** −0.00*** −0.00***
Rating −0.00*** −0.00*** 0.01*** 0.01*** 0.00*** 0.00***
Junk rated −0.03*** −0.03*** 0.03 0.03* 0.01 0.01*
Unrated −0.03*** −0.03*** 0.03*** 0.03*** 0.01*** 0.01***
Insured −0.00 −0.00 0.02*** 0.02*** 0.01*** 0.01***
General obligation 0.00 0.00 0.01*** 0.01*** 0.00*** 0.00***
Callable 0.01*** 0.01*** −0.02*** −0.02*** −0.01*** −0.01***
Sinking fund 0.01*** 0.01*** −0.03*** −0.03*** −0.01*** −0.01***
Bank qualified −0.04*** −0.04*** 0.03*** 0.04*** 0.01*** 0.02***
Taxable bond 0.05*** 0.04*** −0.13*** −0.10*** −0.05*** −0.04***
Subject to AMT 0.02*** 0.02*** −0.04*** −0.03*** −0.02*** −0.01***
R2 – – 0.115 0.096 – –
120 The Journal of FinanceR

C. Central Dealers Reduce Intermediation Chain Length


We now examine how the expected length of intermediation chains varies
with dealer centrality. We estimate the following equation for the total number
Ni of dealers involved in round-trip i:

Chain length Ni = αi + δ Neti + β  Xi + εi , (4)

where αi are state and month fixed effects, Neti is the centrality of the head
dealer in the chain, and Xi are the control variables compiled in Appendix D.
We again allow for state and month fixed effects and adjust standard errors for
heteroskedasticity and clustering at both issuer and time. In columns (3) and
(4) of Table VII we use panel OLS regressions, and in columns (5) and (6) we
report estimates from panel Poisson regressions for robustness.
The chain analysis complements the analysis of dealers’ propensity to trade
with customers versus dealers. All else being equal, it could be the case that
peripheral dealers pass on incoming bonds immediately to core dealers or that
they first search for a buyer and only occasionally pass on bonds in the in-
terdealer market. The expected chain length would be different in the two
cases. The dealer network is more relevant in the former than the latter case,
even though the network is core-periphery in both cases. In addition, even if
peripheral dealers passed on all bonds, central dealers may treat these trades
differently from bonds arriving directly at the center, which would impact chain
length.
We find that the complexity of intermediation is negatively related to the
centrality of the head dealer in all specifications. Intermediation chains are 0.9
to 1.2 dealers (columns (3) and (4)), or 36% to 46% (columns (5) and (6)) shorter
when the chain starts at the center compared with the periphery, while the
average chain length is about 1.5 with a minimum of one and a maximum of
seven. Similar effects hold for underwriters, primary dealers, and non-NYC-
headquartered dealers.

VI. Trading Fast or Slow in the Muni Market


The evidence in Sections III and V suggests that central dealers that are
better connected to other dealers can locate trade counterparties faster than
peripheral dealers. In this section, we provide direct evidence for differences
in execution delay, Hypothesis 3, and examine whether investors trade more
with central dealers when their need for speed is higher, Hypothesis 4.

A. Execution Delay: Principal versus Prearranged Trades


The models in Section I show that, in addition to trading cost, execution
speed is important for investors (see also Demsetz (1968)). While the evidence
in Section V on order flow routing through central dealer hubs suggests that
execution speed is higher when trading with central dealers, the evidence is still
indirect. Ideally, a direct measure of execution delay would evaluate the time
Dealer Networks 121

between the initial contact and trade consummation. However, in most OTC
markets, where a majority of business is conducted by phone, order routing
information is hard to come by, which makes it difficult to measure transaction
speed. A special feature of the municipal bond market allows us to construct
a proxy for whether execution occurs immediately or with delay, based on
transaction records only.
Bond dealers intermediate round-trip trades in one of two ways. Upon receiv-
ing a client call for a sell order, they face the choice between taking the bond into
inventory or asking the seller to wait until a matching buyer is found (the seller
may also offer to wait). The former choice, called a “principal trade,” provides
immediacy—that is, faster execution—to the seller, but it entails inventory risk
for the dealer. The latter choice, called a “prearranged” or “riskless principal”
trade, allows the dealer to act purely as a matchmaker, eliminating the need
for the dealer to commit capital, but the seller bears the cost of execution delay.
Prearranged and principal trades share features with limit orders and market
orders, respectively, in equity markets (in our model in the Internet Appendix,
a principle trade is tantamount to faster execution 1/λ for the investor).
While our transaction data do not flag prearranged trades directly, we can
infer them using the time stamps associated with the trade legs. We define
prearranged round-trips as those with buy and sell legs immediately offsetting
each other—they have the same time stamps, dealer identifiers, and par size
in the same CUSIP (immediate matches). As an alternative, we define pre-
arranged trades more broadly as those such that the buy and sell legs offset
each other within the day (same-day matches). Dealers face a capital charge
on bonds that they take into overnight inventory. Same-day matches, while
riskier for the dealer than immediate matches, are cheaper to intermediate
than overnight round-trips. Thus, while we cannot measure execution delay
precisely, we can use prearranged trades to proxy for trade delays.27
Prearranged trades occur often. Specifically, 11%, 23%, and 42% of all
C(N)DC round-trips are prearranged so that the time stamps are the same,
within 10 minutes, on the same day. We estimate the following Probit model
for the probability of a prearranged trade, assuming that the errors are nor-
mally distributed, allowing for state and month fixed effects αi , and adjusting
standard errors for heteroskedasticity and clustering by issuer and time:

Pr(T rade delayi | T radei ) = (αi + δ Neti + β  Xi + εi ). (5)

The dependent variable takes the value of 1 for a prearranged trade with the
same time stamp (Table VIII, columns (1) and (2)) or with the first two legs of
the trades on the same day (Table VIII, columns (3) and (4)), and 0 otherwise.
Table VIII relates the propensity of prearranging trades—and thus slower
trade execution—to the centrality of the intermediating dealer. In all specifi-
cations, we find that central dealers are significantly less likely to prearrange

27 Bessembinder et al. (2016) also examine the difference between principal trades and prear-

ranged trades.
122 The Journal of FinanceR

Table VIII
Prearranged Trades and Dealer Centrality
The table reports the determinants of prearranged trades (dependent variable = 1) versus prin-
cipal trades (dependent variable = 0). The estimates are obtained from panel regressions with
state and month fixed effects. The determinants include the dealer characteristics, trade char-
acteristics, issue characteristics, and issuer characteristics described in Appendix D. The dealer
centrality measure is the first principal component of the network variables in Appendix C. The
EW (VW) columns employ the equal-weighted (value-weighted) dealer centrality measures. The
sample consists of all CDC round-trip transactions (N = 8,808,119). Standard errors are adjusted
for heteroskedasticity and double clustered by issuer and time. Significance levels are indicated by
* (10%), ** (5%), and *** (1%).

Prearranged, Prearranged, Same-Day


Same-Minute Round-Trip Round-Trip

EW VW EW VW
Determinant (1) (2) (3) (4)

Dealer centrality −1.61*** −1.57*** −1.06*** −0.97***


Primary dealer 0.04** 0.03* −0.31*** −0.32***
NYC dealer −0.01 0.01 0.01 0.02
Underwriter −0.07 −0.06 −0.01 −0.00
Dealer size 0.09** 0.08** 0.07*** 0.06**
Dealer inventory −0.07*** −0.07*** −0.04*** −0.05***
log(Par)*Small 0.05*** 0.05*** −0.05*** −0.05***
log(Par)*Medium 0.04*** 0.04*** −0.01*** −0.01***
log(Par)*Large 0.05*** 0.05*** 0.06*** 0.06***
Maturity −0.07*** −0.07*** −0.12*** −0.12***
Seasoning −0.02*** −0.01*** 0.06*** 0.06***
Issue size 0.00*** 0.00*** 0.00*** 0.00***
Rating −0.01*** −0.01** −0.02*** −0.02***
Junk rated 0.16** 0.16** 0.04 0.04
Unrated −0.04* −0.04** −0.06*** −0.07***
Insured −0.17*** −0.16*** −0.05*** −0.05***
General obligation −0.08*** −0.08*** −0.03*** −0.04***
Callable 0.01 0.01 0.14*** 0.14***
Sinking fund −0.03** −0.02** 0.11*** 0.11***
Bank qualified −0.03** −0.03*** −0.01** −0.01**
Taxable bond 0.30*** 0.34*** 0.14*** 0.15***
Subject to AMT 0.04** 0.05*** 0.22*** 0.22***

trades than peripheral dealers.28 This difference suggests that central deal-
ers provide immediacy to investors, consistent with the model in the Internet
Appendix. Thus, central dealers more readily commit capital and take bonds
into their own inventory than peripheral dealers. In this regard, trading with
a central (peripheral) dealer is like submitting a market (limit) order.
Turning to the control variables, we find that, all else being equal, dealers
with larger balance sheets are more likely to prearrange trades. The effect of
centrality on the choice of trade type is thus distinct from a pure size effect,
as in Table V. The results are mixed for primary dealers and insignificant for

28 The results are robust to the round-trip sample that we use.


Dealer Networks 123

NYC-based dealers and bond underwriters. Finally, dealers with large inven-
tory tend to serve customers on a principal basis, as expected. The results
on bond characteristics are also generally consistent with riskier and smaller
bonds being less likely to be prearranged, suggesting that sellers are less
willing to face execution delay on such bonds.

B. Dealer of Last Resort: Trading in Stress Times


Having established that central dealers provide immediacy by trading more
on a principal than a riskless basis and absorbing bonds into inventory, we now
examine whether times of market stress affect their choice of central versus
peripheral dealers (as predicted by Proposition IA2 in the Internet Appendix).
We expect impatient investors with strong liquidity needs (high L in the model)
to trade with central dealers, especially if these shocks occur frequently (high
κ), for instance, during times of high marketwide illiquidity. Investors should
thus choose central dealers when they are particularly eager to sell quickly.
We also expect investors with strong outside options/sophistication (high σ ) to
trade more with central dealers.
In capturing investors’ demand for fast execution, we are handicapped by
data limitations. Specifically, demand for immediacy (captured by L, κ, and σ ) is
not directly observable. However, we can take advantage of market conditions
under which investors are particularly eager to trade fast. We look at the
following events: bond mutual fund fire sales, bond rating downgrades, and
market volatility spikes.
Mutual funds are generally considered sophisticated investors that, when
confronted with extreme outflows, are more likely to require fast execution in
liquidating holdings to meet redemptions. Bond mutual funds are particularly
prone to fire-sale pressure due to a lack of liquidity in their portfolio.29 We
study municipal bond fire sales at both the aggregate market and the indi-
vidual bond level. Following Coval and Stafford (2007), we proxy for fire-sale
pressure at the individual bond level by aggregating mutual fund trades under
extreme investor flows. We proxy for aggregate fire-sale pressure using total
weekly net outflows from municipal mutual funds. Next, bond rating changes
put pressure on institutional investors due to portfolio constraints, indexing
and benchmarking, and label effects. As a result, institutional investors are
eager to rebalance their portfolios when bond ratings change, especially due
to downgrades. A sharp rise in market volatility, as captured by the VIX in-
dex, also indicates stress in the market that can trigger panic-like selling by
municipal bond investors, who will be more willing to pay for a fast exit.
Table IX reports panel Probit regression results on the determinants of in-
vestors’ choice of whether to trade with a central or peripheral dealer. The
regression results are generally consistent with our predictions. We find that
fire-sale pressure due to extreme mutual fund outflows is associated with a

29 Coval and Stafford (2007) consider similar scenarios in the equity market. They find signifi-

cant price pressure associated with equity fire sales.


124 The Journal of FinanceR

Table IX
When Do Investors Trade with Central Dealers?
The table documents the determinants of investors’ choice to trade with a central versus periph-
eral dealer. The determinants include aggregate market conditions, bond-specific conditions, bond
characteristics, and trade characteristics as described in Appendix D. Column (1) reports estimates
from a Probit regression. In columns (2) and (3), we estimate panel Probit regressions with state
fixed effects and, respectively, state and month fixed effects. The sample consists of all C(N)DC
round-trip transactions (N = 11,033,647 after merging with explanatory variables). Standard er-
rors are adjusted for heteroskedasticity and robust to clustering by state and month. Significance
levels are indicated by * (10%), ** (5%), and *** (1%).

Determinant (1) (2) (3)

Aggregate events:
Net outflows from muni bond funds, 0.01*** 0.01***
current week
Net outflows from muni bond funds, last 0.01*** 0.01***
week
VIX 0.32*** 0.32***
End of month 0.00 0.00 0.01
Bond-specific events:
Sale pressure by mutual funds, current 0.16*** 0.10* 0.13**
quarter
Bond rating downgrade in notches, current 0.01*** 0.01 0.03***
quarter
log(Par)*Small 0.00 −0.01*** −0.01*
log(Par)*Medium −0.06*** −0.06*** −0.07***
log(Par)*Large −0.06*** −0.06*** −0.06***
Maturity 0.07*** 0.06*** 0.06***
Seasoning 0.03*** 0.02*** 0.02***
Issue size 0.00*** 0.00 0.00*
Junk rated −0.22*** −0.21*** −0.21***
Rating 0.00*** 0.00 0.00
Unrated 0.00*** 0.00 0.02
Insured 0.13*** 0.10*** 0.11***
General obligation 0.06*** 0.05*** 0.06***
Callable 0.03*** 0.02** 0.02*
Sinking fund −0.09*** −0.07*** −0.07***
Bank qualified −0.42*** −0.36*** −0.36***
Taxable bond −0.18*** −0.15 −0.18
Subject to AMT −0.06*** −0.05** −0.05**
Fixed effects – State State + month

significantly higher likelihood of investors trading with a central dealer when


they sell municipal bonds. Similarly, aggregate municipal bond fund outflows
in both the current week and the previous week are positively related to the
likelihood of choosing a central dealer. We further find that central dealers
are more likely to be chosen when there are downgrades and when market
volatility is high.
The remaining parameter estimates reveal that investors tend to choose cen-
tral dealers for more seasoned, longer maturity, larger issuance, higher-rated,
Dealer Networks 125

general obligation, callable, and bank-nonqualified bonds. These same charac-


teristics help explain trading costs and liquidity offerings. It is thus not the case
that central dealers systematically intermediate bonds that are a priori costlier
to intermediate, such as unrated or lower-rated bonds, smaller issue bonds, or
bonds with exotic features such as a sinking fund or bank qualification.
All of the evidence for fire sales, rating actions, and VIX suggest that sophis-
ticated investors choose to trade with central dealers that offer fast execution
when investors are impatient due to strong liquidity needs (Proposition IA2).

VII. Alternative Theories and Additional Tests


In this section, we check alternative explanations for the larger markups ob-
served at central dealers, and we conduct additional tests for specific channels.

A. Asymmetric Information
One alternative explanation for the larger markups observed at central deal-
ers is that trades with central dealers contain more private information, which
subjects central dealers to more adverse selection than peripheral dealers.
While this explanation stands in contrast to the predictions of recent theoretical
models, it is worth empirical investigation. BK provide a model of information
diffusion in OTC markets in which market participants possess private infor-
mation. Their model predicts that better connected dealers are less exposed
to adverse selection because they observe more order flow, and thus they can
charge a smaller adverse-selection-related markup than peripheral dealers for
the information component in the trade.
Using our trade-by-trade data on dealer identifiers, we take this prediction
to the data and empirically measure the portion of the bid-ask spread that is
associated with adverse selection and compare it between central dealers and
peripheral dealers. The results support the predictions of BK. At the same
time, we are able to rule out asymmetric information as a key factor explaining
the centrality premium. To the contrary, given that central dealers are less
subject to adverse selection, the centrality premium result is even more
striking.
Our empirical specification to isolate information effects is based on Bessem-
binder, Maxwell, and Venkataraman (2006) (see also Hollifield, Neklyudov, and
Spatt (2017)). Their model proceeds in two stages. In the first stage, order flows
are decomposed into expected and unexpected components. In the second stage,
the change in transaction prices from one trade to the next in a given bond is
regressed on the unexpected order flow, controlling for the expected order flow.
Importantly, to test the prediction in BK that dealer centrality correlates with
permanent price impact, we incorporate interaction terms between the two
types of order flow and the transacting dealers’ centrality.
Let
Pi = (Pi+1 − Pi )/Pi be the percentage change in bond prices from one
trade to the next. The variable Qi is a trade-size-weighted indicator equal to
1 (−1) if the trade is a customer buy from (sell to) a dealer at a par value less
126 The Journal of FinanceR

than $100K; 2 (−2) if the trade is a customer buy from (sell to) a dealer at a
par value of at least $100K but less than $1M; and 3 (−3) if the trade size is
$1M or higher (the exact cutoffs are immaterial). The change from one trade to
the next is given by
Qi .
The surprise in order flow is denoted by Qi∗ = Qi − Ei−1 [Qi ] and is estimated
in the first stage. We consider three different models for expected order flow
Ei−1 [Qi ]. The baseline model in column (1) of Table X is a simple AR(1) model,
where Qi = γ + ρ Qi−1 + i . The model in column (2) allows both the constant
γ and the AR coefficient ρ to interact with the size bucket (small, medium, or
large trade) of the last transaction, i − 1. The model in column (3) allows both
the constant γ and the AR coefficient ρ to interact with the centrality of the
dealer in trade i − 1.
With Qi∗ computed in three different ways across the three columns in
Table X, returns are decomposed following:


Pi = α + γ1 Qi∗ + γ2 Neti ∗ Qi∗ + δ1
Qi + δ2 Neti ∗
Qi +
Xi  β + εi , (6)

where
Xi is a vector of changes in control variables that include the Bond
Buyer 40 index; the spread between 10-year BBB muni bonds and yields of
10-year Treasuries; the slope of the Treasury yield curve; and the TED spread,
defined as the difference between the three-month LIBOR rate and the three-
month T-Bill rate. Again, Neti is the dealer’s centrality (normalized between
zero and one). In equation (6), γ1 measures the impact of adverse selection
on prices for the most peripheral dealers, γ2 measures the additional price
impact due to adverse selection for the most central dealers relative to the
most peripheral dealers, δ1 measures the bid-ask half-spread, and δ2 measures
the additional bid-ask half-spread associated with central dealers relative to
peripheral dealers.
Table X reports the results, with the estimates for the first-stage dependent
variable Qi in Panel A and the second-stage dependent variable
Pi in Panel
B. We find that, regardless of the model for order flow dynamics specified in
the first stage (across columns (1) to (3)), central dealers are associated with a
lower price impact due to adverse selection (γ2 < 0). The estimated coefficient
γ2 of −0.22 is economically large and virtually offsets γ1 = 0.22 in magnitude,
so that the price impact of unexpected order flow is close to zero for central
dealers. This finding is consistent with the prediction of BK that central dealers
should be less exposed to adverse selection. Indeed, central dealers charge a
higher bid-ask half-spread (δ2 = 0.18 > 0) even after controlling for asymmetric
information, consistent with our earlier results.
The evidence from the price impact regressions suggests that central dealers
are less subject to adverse selection, yet they charge a larger bid-ask spread.
Hence, the information story of BK, while at play in the data, does not explain
the centrality premium. It appears that alternative factors—most importantly,
faster trade execution—more than compensate for the difference in adverse
selection, and result in an overall positive relation between centrality and
execution costs.
Dealer Networks 127

Table X
Adverse Selection and Dealer Centrality
The table documents the relationship between the adverse selection component in dealer markups
and dealer centrality. The regression is conducted in two stages. The first stage in Panel A estimates
order flows dynamics. The variable Qi is a trade size-weighted indicator variable that equals 1 (−1)
if the trade is a customer buy from (sell to) a dealer at a par value smaller than $100K; 2 (−2) if
the trade is a customer buy from (sell to) a dealer at a par value of at least $100K but less than
$1M; and 3 (−3) if the trade size is $1M and above. The surprise in order flow, Qi∗ = Qi − Ei−1 [Qi ],
is estimated in the first stage. Columns (1) to (3) specify different models for expected order
flow Ei−1 [Qi ]. Column (1) is the baseline AR(1) model. Column (2) allows the AR coefficient to
interact with the size of trade i − 1. Column (3) allows the AR coefficient to interact with the
dealer centrality in trade i − 1. The second stage in Panel B estimates the impact of order flow
surprises and total order flow changes on transaction prices. The model is specified in equation (6).
The estimates are obtained from panel regressions with bond and date fixed effects. The sample
includes all trades in the municipal bond market between 1998 and 2012 (N = 63,014,690 in
Panel A, N = 59,254,786 in Panel B after merging with explanatory variables). Standard errors
are adjusted for heteroskedasticity and double clustered by bond and time. Significance levels are
indicated by * (10%), ** (5%), and *** (1%).

Determinant (1) (2) (3)

Panel A: First Stage Ei−1 [Qi ]

Qi−1 −0.13*** −0.04***


Qi−1 * Small trade −0.11***
Qi−1 * Medium trade −0.14***
Qi−1 * Large trade −0.33***
Qi−1 * Neti−1 −0.10***
Neti−1 0.21***
Medium trade −0.08***
Large trade −0.54***
Constant 0.32*** 0.34*** 0.13***
R2 0.055 0.065 0.056

Panel B: Second Stage


Pi = α + γ1 Qi∗ + γ2 Neti ∗ Qi∗ + δ1
Qi + δ2 Neti ∗
Qi +
Xi  β + εi

γ1 0.22*** 0.25*** 0.23***


γ2 −0.23*** −0.22*** −0.23***
δ1 0.47*** 0.45*** 0.46***
δ2 0.18*** 0.18*** 0.19***
β:
Bond Buyer index 0.46*** 0.46*** 0.46***
BBB 10-year spread 0.44*** 0.44*** 0.44***
Treasury yield slope 0.84*** 0.84*** 0.84***
TED spread 0.03*** 0.03*** 0.03***
R2 0.354 0.354 0.354

B. Do Central Dealers Intermediate Riskier Bonds and Trades?


Are the larger markups at central dealers compensation for taking on
noninformational risks? This can happen if central dealers intermediate
riskier bonds with less certain demand, lower liquidity, and higher price
volatility. We explicitly test this alternative explanation by looking at the
probability of dealers taking a trading loss on a round-trip. Panel B in Table IV
128 The Journal of FinanceR

Table XI
Dealer Losses and Centrality
The table reports the determinants of the probability that dealers take a loss on a round-trip
transaction. The estimates are obtained from panel regressions with state and month fixed effects.
The determinants include the dealer characteristics, trade characteristics, issue characteristics,
and issuer characteristics described in Appendix D. The dealer centrality measure is the first
principal component of the network variables in Appendix C. The EW (VW) columns employ
the equal-weighted (value-weighted) dealer centrality measures. For the C(N)DC sample, the
dealer centrality measure is defined as the centrality of the head dealer. We vary the regression
sample across columns, considering three types of trades with varying dealer involvement. CDC-
Nonsplits are round-trips intermediated by a single dealer where the original bond lot is not
split (N = 6,294,447). The CDC sample includes all round-trips intermediated by a single dealer
(N = 8,808,119). C(N)DC are round-trips intermediated by one or several dealers (N = 11,404,321).
Standard errors are adjusted for heteroskedasticity and double clustered by issuer and time.
Significance levels are indicated by * (10%), ** (5%), and *** (1%).

CDC-Nonsplit CDC C(N)DC

EW VW EW VW EW VW
Determinant (1) (2) (3) (4) (5) (6)

Dealer centrality −0.32*** −0.42*** −0.35*** −0.41*** −0.32*** −0.38***


Chain length −0.45*** −0.48***
Chain length*Centrality 0.46*** 0.49***
Primary dealer 0.02 0.02 0.06** 0.05** 0.08*** 0.08***
NYC dealer 0.17*** 0.18*** 0.07*** 0.08*** 0.05*** 0.06***
Underwriter −0.12*** −0.11*** −0.10*** −0.09*** −0.08*** −0.08***
Dealer size −0.21*** −0.21*** −0.17*** −0.17*** −0.14*** −0.13***
Dealer inventory −0.10*** −0.10*** −0.07** −0.07** −0.05** −0.05**
log(Par)*Small 0.02*** 0.02*** 0.03*** 0.03*** 0.01*** 0.01***
log(Par)*Medium 0.06*** 0.06*** 0.08*** 0.08*** 0.07*** 0.07***
log(Par)*Large 0.10*** 0.10*** 0.12*** 0.12*** 0.11*** 0.11***
Maturity −0.00 −0.00 −0.02** −0.02** −0.03** −0.02**
Seasoning −0.06*** −0.06*** −0.05*** −0.05*** −0.05*** −0.05***
Issue size 0.00 0.00 0.00*** 0.00*** 0.00*** 0.00***
Rating −0.01*** −0.01*** −0.01*** −0.01*** −0.01*** −0.01***
Junk rated 0.03 0.03 0.06* 0.06* 0.06** 0.06**
Unrated −0.03* −0.03* −0.06*** −0.06*** −0.07*** −0.07***
Insured −0.04*** −0.04*** −0.06*** −0.06*** −0.05*** −0.05***
General obligation −0.00 −0.00 −0.00 −0.00 0.00 0.00
Callable −0.17*** −0.17*** −0.20*** −0.20*** −0.20*** −0.20***
Sinking fund −0.03*** −0.03*** −0.04*** −0.04*** −0.05*** −0.05***
Bank qualified −0.03** −0.04** −0.07*** −0.07*** −0.06*** −0.07***
Taxable bond −0.08*** −0.08*** −0.07*** −0.07*** −0.08*** −0.08***
Subject to AMT −0.03* −0.03* −0.07*** −0.07*** −0.08*** −0.08***

already provides univariate results for the most common trade sizes. In each of
the three round-trip samples, dealers lose money in less than 2% of all trades.
Still, the loss probability depends strongly on dealers’ relative position in the
network, but with the opposite sign predicted by the risk-return trade-off.
Table XI extends this analysis to a multivariate setting. We document the
determinants of trading losses using a panel Probit model that controls for
Dealer Networks 129

various dealer, trade, and bond characteristics, includes state and month fixed
effects, and adjusts standard errors for heteroskedasticity and clustering by
issuer and time. The results are in line with those in Table IV: central dealers
are significantly less likely than peripheral dealers to lose on round-trips. Thus,
the profits of more connected dealers are on average larger and less risky.
Moreover, the larger markups that central dealers charge are not obviously
compensation for price risks.

C. Inventory Risk Channel


The explanation that central dealers are able to charge a premium due to
their higher search efficiency and willingness to trade on a principal basis is in-
herently related to their ability to carry inventory (consistent with Üslü (2015)
and others). One simple alternative story that may explain a centrality pre-
mium is that it is more expensive for central dealers to hold inventory, and thus
they need to charge higher markups to compensate for the higher cost. This
story is problematic in a number of ways. First, it is a partial equilibrium argu-
ment in that investors can choose not to trade with high-cost intermediaries.
Second, we show in Section VI that central dealers trade more on a principal
basis by agreeing to take bonds into inventory than on a prearranged trade
basis. In this sense, the centrality premium compensates for taking bonds into
inventory, but it is valuable because of the immediacy it provides to clients,
not because inventory is costlier to carry for central than peripheral dealers.
Our findings suggest that central dealers are more willing to take inventory
risks, likely because it is easier and cheaper—not more expensive—for them to
do so.
A subtler hypothesis is that the centrality premium compensates for a longer
duration of inventory risk that central dealers face. This story is consistent
with our finding that central dealers trade on a principal basis. However, longer
inventory risk exposure alone does not seem to explain the centrality premium.
Several pieces of evidence speak against this explanation for the centrality
premium. First, we show in Section IV that dealers charge less the larger their
abnormal inventory. This suggests that additional inventory reduces dealers’
ability to charge more, yet central dealers hold larger inventory. Second, the
inventory risk compensation explanation predicts that the dealer earning the
largest markup, or the largest share of the total markup, should be the one
taking the biggest inventory risk, and the additional markup that the dealer
earns compensates for the risk of value changes in the bond holdings. We check
this alternative explanation by comparing dealer markups to holding periods
in the round-trip chain.
Table XII, Panel A, reports the average markup of each dealer in the chain.
One can see that, regardless of how many dealers are involved in an interme-
diation chain, the last dealer (the dealer closest to the ultimate buyer) earns
a markup of about 1.1%, which corresponds to the largest share of the overall
markup.
130 The Journal of FinanceR

Table XII
Dealer Markups and Holding Periods
Panel A reports average markups per dealer on round-trip transactions with varying degrees
of dealer involvement. Total dealer markups are broken down by the number of dealers (across
rows) and by each dealer (across columns) in the sequence of dealers intermediating the round-
trip transaction. Markups are measured in percentage of the first dealer’s purchase price from
the customer. We restrict the sample to nonsplits. No additional data filters are applied. Panel B
reports inventory time in calendar days for each dealer along the round-trip chain. We report the
average time, measured in business days, each dealer in the round-trip chain holds the bond. We
restrict the sample to nonsplits. No additional data filters are applied.

Panel A: Markup by Dealer in Intermediation Chains

Dealer Markup, # = Position in Chain

Trade Type Total Markup CD #2 #3 #4 #5 #6 DC

CDC 1.85 . . . . . . .
CDDC 1.94 0.84 . . . . . 1.10
CDDDC 2.26 0.66 0.52 . . . . 1.08
CDDDDC 2.92 0.64 0.60 0.55 . . . 1.13
CDDDDDC 3.26 0.63 0.30 0.82 0.40 . . 1.11
CDDDDDDC 3.57 0.60 0.27 0.45 0.85 0.27 . 1.14
CDDDDDDDC 3.71 0.62 0.23 0.43 0.53 0.47 0.29 1.15

Panel B: Holding Period by Dealer in Intermediation Chains

Time in Dealer Inventory, # = Position in Chain

Trade Type Total Time CD #2 #3 #4 #5 #6 DC

CDC 1.70 . . . . . . .
CDDC 2.64 1.85 . . . . . 0.79
CDDDC 2.64 0.89 1.32 . . . . 0.43
CDDDDC 3.18 0.42 1.30 1.03 . . . 0.43
CDDDDDC 3.66 0.25 0.47 1.92 0.77 . . 0.25
CDDDDDDC 4.24 0.35 0.30 0.83 2.07 0.44 . 0.25
CDDDDDDDC 6.06 0.13 0.67 1.01 1.68 1.69 0.72 0.16

Table XII, Panel B, reports the average dealer holding period as measured
by the number of business days for each dealer along the chain. We see that the
dealer closest to the buyer is not the one that holds the bond the longest, even
though the last dealer earns the largest markup. The results on markup and
holding period together suggest that the source of market power for dealers
appears to come from their ability to identify buyers, not inventory risk-taking
alone.
Third, we examine whether central dealers keep bonds in inventory longer
when they conduct a principal trade and get compensated for doing so. We
measure the time it takes a dealer to unwind a position by the inventory
duration of the bond. The inventory duration is the number of days, hours,
and minutes in a principal round-trip trade between the initial purchase and
the ultimate sale by the dealer. When bond lots are split, a bond’s inventory
Dealer Networks 131

Table XIII
Times to Unwind Principal Trades and Dealer Centrality
The table reports the determinants of inventory duration. The estimates are obtained from panel
Tobit regressions (lower limit = 0) with state and month fixed effects. The determinants include
the dealer characteristics, trade characteristics, issue characteristics, and issuer characteristics
described in Appendix D. The dealer centrality measure is the first principal component of the net-
work variables in Appendix C. The EW (VW) columns employ the equal-weighted (value-weighted)
dealer centrality measures. For the C(N)DC sample, the dealer centrality measure is defined as
the centrality of the head dealer. We vary the regression sample across columns, considering three
types of trades with varying dealer involvement. Short sales are excluded from the sample. CDC-
Nonsplits are round-trips intermediated by a single dealer where the original bond lot is not split
(N = 5,731,489). The CDC sample includes all round-trips intermediated by a single dealer (N
= 8,215,616). C(N)DC are round-trips intermediated by one or several dealers (N = 9,862,339).
Standard errors are adjusted for heteroskedasticity and double clustered by issuer and time. Sig-
nificance levels are indicated by * (10%), ** (5%), and *** (1%).

CDC-Nonsplit CDC C(N)DC

EW VW EW VW EW VW
Determinant (1) (2) (3) (4) (5) (6)

Dealer centrality −0.76*** −1.31*** −0.32 −1.05*** −0.13 −0.80***


Chain length 0.41*** 0.42***
Chain length*Centrality 0.73*** 0.70***
Primary dealer 0.29*** 0.30*** 0.17** 0.18*** 0.09 0.10*
NYC dealer 0.41*** 0.44*** 0.46*** 0.49*** 0.42*** 0.45***
Underwriter −0.00 −0.00 0.02 0.02 0.01 0.01
Dealer size −0.56*** −0.58*** −0.63*** −0.65*** −0.50*** −0.50***
Dealer inventory 0.04** 0.04** 0.02 0.02 0.02 0.02
log(Par)*Small −0.42*** −0.42*** −0.02* −0.03** −0.09*** −0.09***
log(Par)*Medium −0.48*** −0.48*** −0.02* −0.02** −0.10*** −0.10***
log(Par)*Large −0.42*** −0.42*** −0.12*** −0.12*** −0.15*** −0.15***
Maturity 0.20*** 0.20*** 0.35*** 0.36*** 0.31*** 0.32***
Seasoning −0.19*** −0.19*** −0.26*** −0.26*** −0.27*** −0.27***
Issue size −0.01*** −0.01*** −0.01*** −0.01*** −0.01*** −0.01***
Rating 0.03*** 0.03*** 0.05*** 0.05*** 0.04*** 0.04***
Junk rated −0.07 −0.09 0.05 0.03 0.02 0.01
Unrated 0.10** 0.09** 0.18*** 0.18*** 0.14*** 0.13***
Insured −0.19*** −0.19*** −0.22*** −0.21*** −0.24*** −0.24***
General obligation −0.03 −0.03 −0.09*** −0.09*** −0.09*** −0.09***
Callable −0.49*** −0.49*** −0.63*** −0.64*** −0.64*** −0.65***
Sinking fund −0.49*** −0.49*** −0.54*** −0.54*** −0.56*** −0.56***
Bank qualified 0.34*** 0.31*** 0.18*** 0.15*** 0.20*** 0.18***
Taxable bond 0.23** 0.22** 0.35*** 0.33*** 0.29*** 0.27**
Subject to AMT −0.42*** −0.41*** −0.43*** −0.43*** −0.48*** −0.48***

duration is the average time, weighted by the size of the split orders, that it
takes to resell the entire bond lot. We eliminate short sales from the sample.
Univariate analysis shows that the average (median) time it takes a dealer to
unwind a position is 3.5 (1.1) days.
Table XIII provides evidence on the determinants of inventory times across
principal trades. We estimate the following equation in a panel Tobit regression
132 The Journal of FinanceR

with state and month fixed effects αi , where the lower limit is set to zero since
inventory time cannot be negative:30

Inventory time
Ti = αi + δ Neti + β  Xi + εi . (7)

The coefficients δ and β capture the sensitivity of inventory duration to dealer


centrality and, respectively, to characteristics of the bond and trade. The
estimates in Table XIII reveal that central dealers hold bonds in their in-
ventory for less time than peripheral dealers, controlling for other charac-
teristics. The difference in average inventory time between the most cen-
tral and the most peripheral dealers is up to 1.3 days depending on the
specification.31
Overall, liquidity provision by central dealers is tied to taking inventory
risks, and by doing so, they provide immediacy to investors who seek to trade
fast. But higher inventory holding costs alone do not seem to explain the cen-
trality premium, as central dealers more willingly take inventory and turn
around inventory faster than peripheral dealers.

D. Local Information: Evidence from Tax-Based State Segmentation


The municipal bond market has some unique characteristics that are in-
teresting to explore and that help us sharpen the identification of our results
on network centrality. Municipal bond investors exhibit a local bias due to
the exemptions they receive on their income tax, that is, they prefer bonds
issued by their own states and municipalities due to preferential tax treat-
ment. This state-based market segmentation allows us to address a number of
questions.
One concern is that, despite our efforts to control for observable dealer char-
acteristics, unobservable characteristics that drive the cross-sectional results
on dealer markups and centrality may exist. To mitigate this concern, we modify
our baseline specification along two dimensions. First, we construct the trading
network separately for each state and use state-specific centrality measures in
the regression model. If dealers’ search efficiency matters, then state-specific
centrality should matter beyond aggregate centrality. In this case, our results
do not simply capture a correlated unobservable dealer characteristic. Second,
we allow for dealer fixed effects that absorb all dealer characteristics so long

30 Results from a panel OLS model with state and month fixed effects are similar and omitted.

Coefficient estimates in the OLS model are larger and more significant.
31 In an alternate specification, we include dealer centrality Net and its square, since univariate
i
analysis revealed a nonmonotonocity:
Ti = αi + δ1 Neti + δ2 (Neti )2 + β  Xi + εi . The estimates con-
firm that central dealers hold bonds in their inventory for a shorter time than peripheral dealers,
controlling for other characteristics. The difference in average inventory time between the most
central and the most peripheral dealers is now between 0.9 and 1.3 days depending on the specifi-
cation. We find significant coefficients on squared centrality for the equal-weighted measures. This
suggests that the relation between centrality and inventory duration is nonmonotonic. Peripheral
dealers hold bonds the longest when they conduct a principal trade, followed by core dealers and
midtier dealers.
Dealer Networks 133

as they are time-invariant. Any network effect on markups would then come
from time-series variation in dealer centrality.
State-specific dealer centralities are calculated using a rolling window of
interdealer trades on bonds issued by each state. Each dealer has a different
centrality measure, which we denote by State-Net, for each state and time.
This state-specific measure captures the search efficiency of each dealer in
each state.
Table XIV summarizes the results from the specifications that address
whether centrality proxies for capital base and other dealer characteristics.
Columns (1) and (2) augment the centrality measure Net by dealers’ state-
specific centrality State-Net. In columns (3) and (4), we add dealer fixed effects
to the specification, so that the coefficient on Net captures the impact of time-
series variation in dealer centrality.
When the two centrality measures are included, both are significant, sug-
gesting that dealers’ search efficiency at both the aggregate market level and
the state level matters. Regardless of whether aggregate dealer centrality or
state-specific centrality is used, dealer centrality is positively and significantly
related to markups, except if both value-weighted centralities and dealer fixed
effects are included. Overall, the results suggest that dealers’ network position
itself, beyond any unobservable dealer characteristics, drives the dispersion in
trading costs.

E. Cross-Sectional Tests: Taxables and Puerto Rico Bonds


Another natural question that arises is whether our findings on the cen-
trality premium are special to the tax-exempt municipal bond market or are
they generalizable to other OTC markets. To address this issue, we perform
subsample analysis in which we target specific bond types.32
The general market structure of the municipal bond market is similar to
other OTC markets for debt securities, including corporate bonds and struc-
tured products. In all of these markets, a majority of trading is conducted by
phone or electronic messaging through broker-dealers, the number of unique
bonds is high (rendering search for the physical asset and short-selling difficult
and risky), the majority of investors are buy-and-hold (including mutual funds,
insurance companies, pension funds, and households), and there are a large
number of broker-dealers. Given these similarities, we expect our results to
generalize to these markets as well.33

32 We thank our referee for suggesting that we explore the taxable municipal bond market.
33 Using supervisory TRACE data on corporate bond transactions, we were able to confirm that
our results of a centrality premium hold for the corporate bond market. Results are available upon
request. In follow-on work, Di Maggio, Kermani, and Song (2015) confirm the centrality premium
in the OTC market for corporate bonds that has a similar structure to the municipal bond market.
Whether our findings apply to other OTC markets with a different market structure—such as the
markets for credit default swaps and tri-party repos, in which a smaller number of broker-dealers
are active and some of the trades are centrally cleared, is a question for future research.
134 The Journal of FinanceR

Table XIV
Local Information: Trading Costs and State-Specific Dealer
Centrality
The table reports the determinants of round-trip trading costs. The estimates are obtained from
panel regressions with state and month fixed effects, or indicated otherwise. In columns (1) and (2),
we add dealer fixed effects to the specification so that the coefficient on Net captures the impact of
time-series variation in dealer centrality. In columns (3) and (4), we add to the aggregate centrality
measure Net dealers’ state-specific centrality State-Net in the state of issuance of the bond being
traded. The determinants include the dealer characteristics, trade characteristics, issue charac-
teristics, and issuer characteristics described in Appendix D. The dealer centrality measure is the
first principal component of the network variables in Appendix C. The EW (VW) columns employ
the equal-weighted (value-weighted) dealer centrality measures. The sample consists of all CDC
round-trip chains (N = 8,808,119). Standard errors are adjusted for heteroskedasticity and double
clustered by issuer and time. Significance levels are indicated by * (10%), ** (5%), and *** (1%).

State-Specific Centrality + Dealer Fixed


State-Specific Centrality Effects

EW VW EW VW
Determinant (1) (2) (3) (4)

State-specific 0.50*** 0.49*** 0.21*** 0.19***


dealer centrality
Aggregate dealer 0.28*** 0.13*** 0.03 −0.16**
centrality
Primary dealer −0.09*** −0.09*** 0.00 0.00
NYC dealer 0.08*** 0.08*** 0.06** 0.07***
Underwriter 0.01 0.01 0.06*** 0.06***
Dealer size −0.11*** −0.11*** 0.00 0.00
Dealer inventory −0.02** −0.02** −0.02*** −0.02***
log(Par)*Small −0.24*** −0.24*** −0.20*** −0.20***
log(Par)*Medium −0.26*** −0.26*** −0.22*** −0.22***
log(Par)*Large −0.29*** −0.30*** −0.25*** −0.25***
Maturity 0.75*** 0.75*** 0.68*** 0.68***
Seasoning −0.06*** −0.06*** −0.07*** −0.07***
Issue size −0.00*** −0.00*** −0.00 −0.00
Rating 0.07*** 0.07*** 0.06*** 0.06***
Junk rated −0.04 −0.05 −0.07* −0.07*
Unrated 0.27*** 0.27*** 0.19*** 0.19***
Insured 0.08*** 0.08*** 0.08*** 0.08***
General obligation −0.01 −0.01 −0.01 −0.01
Callable −0.25*** −0.25*** −0.26*** −0.26***
Sinking fund −0.09*** −0.09*** −0.10*** −0.10***
Bank qualified 0.03*** 0.02*** 0.02** 0.02**
Taxable bond 0.07* 0.06 0.09*** 0.09***
Subject to AMT 0.11*** 0.11*** 0.10*** 0.10***
Fixed effects State+Month State+Month Dealer+State+Month Dealer+State+Month
R2 0.360 0.359 0.425 0.424
Dealer Networks 135

Table XV
Centrality Premium in Taxable and Puerto Rico Bonds
The table reports the determinants of round-trip trading costs for subsamples of taxable bonds and
Puerto Rico bonds. Taxable bonds are municipal bonds for which the interest income is not exempt
from federal income tax. Interest income from bonds issued by the Puerto Rico state and local
municipalities are exempt from federal and state taxes regardless of what state the bond investors
reside in. The estimates are obtained from panel regressions with state and month fixed effects
for taxable bonds, and with month fixed effects for Puerto Rico bonds. The determinants include
the dealer characteristics, trade characteristics, issue characteristics, and issuer characteristics
described in Appendix D. The dealer centrality measure is the first principal component of the net-
work variables in Appendix C. The EW (VW) columns employ the equal-weighted (value-weighted)
dealer centrality measures. The sample consists of all CDC round-trip chains (N = 190,578 for
taxables, N=198,855 for Puerto Rico bonds). Standard errors are adjusted for heteroskedasticity
and double clustered by issuer and time. Significance levels are indicated by * (10%), ** (5%), and
*** (1%).

Taxable Bonds Puerto Rico Bonds

EW VW EW VW
Determinant (1) (2) (3) (4)

Dealer Centrality 0.31*** 0.26*** 0.50*** 0.26***


Primary dealer −0.39*** −0.39*** −0.06*** −0.05***
NYC dealer 0.34*** 0.34*** −0.01 0.02
Underwriter −0.06 −0.07 −0.29*** −0.30***
Dealer Size −0.40** −0.41** 0.14*** 0.16***
Dealer inventory 0.04 0.04 0.01 0.00
log(Par)*Small −0.13*** −0.13*** −0.19*** −0.19***
log(Par)*Medium −0.19*** −0.19*** −0.22*** −0.23***
log(Par)*Large −0.23*** −0.23*** −0.28*** −0.28***
Maturity 0.73*** 0.73*** 0.55*** 0.55***
Seasoning 0.05** 0.05** −0.07*** −0.07***
Issue size −0.02** −0.02** −0.01*** −0.01***
Rating 0.06*** 0.06*** 0.07*** 0.07***
Junk rated −0.06 −0.06 −0.00 −0.01
Unrated 0.19*** 0.19*** 0.16*** 0.16***
Insured 0.14** 0.14** 0.03*** 0.04***
General obligation 0.01 0.01 −0.03*** −0.03**
Callable −0.19*** −0.19*** −0.32*** −0.32***
Sinking fund −0.17*** −0.17*** −0.19*** −0.19***
Bank Qualified 0.00 0.00 0.00 0.00
Taxable 0.00 0.00 −0.28*** −0.33***
Subject to AMT 0.00 0.00 0.41*** 0.42***
R2 0.408 0.407 0.330 0.327

Even within the municipal bond market, there exists variation that we can
exploit. Some municipal bonds are not subject to state-based segmentation.
They are either taxable, that is, interest income from these bonds is not exempt
from federal and state taxes, or they are tax-exempt for investors from all
states. The most recent case of widely issued taxable bonds is Build America
Bonds (Cestau, Green, and Schürhoff (2013)). The most prominent nationwide
136 The Journal of FinanceR

tax-exempt bonds (by federal law) are issued by Puerto Rico. We use these
bonds to show that state-based segmentation is not what drives our results.
In Table XV, we test for a centrality premium for these two subsamples of
municipal bonds. Results on taxable municipal bonds are reported in columns
(1) and (2), while estimates for Puerto Rico bonds are shown in columns (3) and
(4). In both subsamples, investors’ trading costs increase with dealer centrality,
confirming our full-sample results. Therefore, our findings do not seem to be
driven by the geographical segmentation of the municipal bond market but
rather reflect more general trade-offs.

VIII. Conclusion
In decentralized markets, dealers are pivotal in liquidity provision and price
discovery. Dealers match buyers with sellers among their own clients, ware-
house securities, and form trading networks with other dealers to locate coun-
terparties and reallocate securities.
Using proprietary audit trail data from the municipal bond market, we show
that a dealer’s centrality in the trading network is an important determinant of
execution quality for investors. The interdealer trading network in municipal
bonds exhibits a stable core-periphery structure with 10 to 30 highly intercon-
nected core dealers and thousands of peripheral broker-dealer firms. Relative
to peripheral dealers, central dealers use their superior ability in locating
counterparties to provide faster execution. They match buyers with sellers
more directly, which shortens intermediation chains, and they trade more on
a principal basis than their peripheral competitors. Investors, in turn, trade
with core dealers when they need immediate execution, for instance, during
times of market stress, which highlights the systemic role of central dealers as
liquidity providers of last resort.
Central dealers are compensated by larger markups for the immediacy they
provide. At 0.4% to 0.7% of par value, the centrality premium is a sizable com-
ponent of trading costs. Investors thus face a nonnegligible trade-off between
execution cost and speed when choosing which dealer to trade with, consistent
with recent theories of trade in OTC markets.

Initial submission: October 22, 2014; Accepted: February 2, 2018


Editors: Bruno Biais, Michael R. Roberts, and Kenneth J. Singleton

Appendix A: Data Filters


This appendix describes our data filters. We apply several data filters to
clean the raw MSRB data from erroneous entries and missing information.
The filters check the trades sequentially based on information about the trade,
bond, price, and time. In the following table, we provide a description of each
filter and report the number of trades remaining in the sample after applying
each filter.
Dealer Networks 137

No. of Individual
Description Trades (in Millions)

1. Trade-specific filters
Keep trades after February 1998 (only interdealer trades are
reported before 1998)
Keep trades for which the dealer identifier is a four-letter
alphabetic symbol
Drop trades with MSRB indicator for away-from-market
prices
Keep trades with par value of at least $5K 123.20
2. Bond-specific filters
Keep bonds that have fixed or zero coupon (based on Mergent
data)
Keep bonds that are nonderivative and nonwarrant, and not
puttable
Keep bonds that have at least one year to maturity at the
time of issuance
Keep bonds that have a denomination face value of $5K (large 100.65
majority of bonds)
3. Price- and time-based filters
Keep only trades at least 90 days after issuance (seasoned 75.49
bonds)
Drop bonds with less than a year to maturity (maturing 72.93
bonds)
Eliminate price outliers by truncating the distribution at 72.20
0.5% and 99.5%, separately for zero-coupon and other bonds

We next search the filtered MSRB sample for trade sequences that corre-
spond to round-trip chains using the algorithm documented in Appendix B. The
algorithm matches 82% (14.6 million) of all customer-to-dealer trades to corre-
sponding dealer-to-dealer and dealer-to-customer trades. Of these, 13.3 million
trades are associated with a complete round-trip chain, which ends in a dealer
selling the bonds to a customer. The remainder are incomplete intermediation
chains, in which case a dealer keeps the bond in inventory or matching is in-
complete. The potential reasons are as follows. First, dealers may aggregate
purchases from multiple customers and then sell to a single customer. In this
case, the dealer sale is in a par size larger than the original dealer purchase.
Second, dealers may split orders at the interdealer stage. Third, any of the con-
secutive legs in a round-trip may be more than 30 days apart. Fourth, dealers
may report their IDs erroneously.
We apply additional filters. We drop round-trip chains where the head or
tail dealer acts as an agent for the investor (agent trades), where the first
trade happened in March 1998 (incomplete data), and where the first or
last dealer does not have an MSRBID assigned or it is unknown. The final
data set consists of 11.4 million round-trips, which constitutes our C(N)DC
sample.
138 The Journal of FinanceR

Appendix B: Matching Algorithm for Round-Trip Chains


This appendix describes the algorithm used to match customer-to-dealer and
consecutive dealer-to-dealer and dealer-to-customer transactions into round-
trip chains. Our round-trip matching algorithm is an extension of the algorithm
first used in Green, Hollifield, and Schürhoff (2007). The differences are due to
the information we have on the identities of the dealers involved in each trade.
Our data allow us to trace the order flow of bonds through the dealer network
with higher accuracy.
A typical trade in seasoned municipal bonds starts with an existing holder,
subject to a liquidity shock, wanting to liquidate the position. The investor will
contact a broker-dealer to sell the bonds to a buyer with intermediation by
the broker-dealer. We search for sequences of trades in each bond issue that
resemble such bond orders flowing through the dealer network. The reverse
chain of trades is more unlikely since short-selling of municipal bonds is very
difficult and costly. Our round-trip chains therefore start with a time-stamped
trade from a customer to a dealer, followed by either a trade from a dealer to
a customer or to another dealer. The round-trip chain ends with a sale to a
customer.
As described in Algorithm 1, for each CUSIP and each customer-to-dealer
trade, we sequentially look for one of two types of matches, namely, non-
split chains and split chains. For nonsplit round-trips, we restrict attention
to matching trades of the same par size. We look for trades by the initial dealer
within a calendar-day window up to 30 days after and 10 days before the ini-
tial customer-to-dealer trade. The latter criterion ensures that we capture any
short-selling activity. If there is a trade of the same dealer selling the same
amount of bonds to customers, this identifies a CDC round-trip. The round-trip
chain is then removed from the trading data for future lookups. If no such
direct match exists, we look for interdealer trades in which the initial dealer is
a seller. If such an interdealer trade exists, we have a chain of CDD trades.
We recursively look for matching dealer-to-customer trades or interdealer
trades where each leg is no more than 30 calendar days in the future or 10
days in the past. We allow for up to seven dealers in a round-trip (very few
matches occur with larger thresholds) and we require that each leg is at most
10 trades away in the original sequence of execution-time-sorted MSRB trades.
These restrictions are intended to ensure accuracy in trade matching. The last
restriction allows us to implement the lookup loop more efficiently and reduce
incorrect matching of unrelated trades. We look for order splitting at the end
of the chain if no matching dealer-to-customer trade can be found. For these
split matches, we require that the total par amount of the dealer-to-customer
trades is not greater than the original par size. We take the weighted average
of the sale prices to compute a price for the last leg, and the weighted-average
trade date is the trade date for the last leg.
Dealer Networks 139

Algorithm 1. Procedure to find round-trip chains.

foreach CUSIP do
while not the last trade do
foreach Dealer buys from customer (Dealer = A, P ar = X) do
Run Algorithm 2 to find matching trades (Dealer = A, P ar = X)
if Procedure to find matching trades returns finished chain, Finished=1 then
Record round-trip, type=‘‘C(N)DC,” remove from data
else
if Find splitting dealer sells to customer (Dealer =Last dealer in the chain,
P ar < X) then
Record round-trip, type=‘‘C(N)DC-Split,” remove from data
else
Record unfinished chain, type=‘‘C(N)D,” remove from data

Algorithm 2. Procedure to find matching trades (Dealer = A, Par = X).

return linked list of round-trip chain while not the last trade do
if Dealer A sells to customer, P ar = X then
Add to the round-trip chain, Finished=1, Exit
else
if Dealer A sells to dealer B, P ar = Y ≤ X then
if Existing chain has more than 7 dealers then
Finished=0, Exit
else
Add to round-trip chain, Run procedure to find matching trades (Dealer = B,
P ar = Y )
else
Next trade within [-10,+30]-day window

Appendix C: Measures of Dealer Centrality


Several measures of centrality are widely used in network analysis. They
describe either the local connectivity of a dealer or its global importance. The
following provides a brief description of each measure and its variants, and
explains how we aggregate them into a single index.

r Degree dg: A measure of the local connectivity of a dealer. The degree of


a dealer is computed as the sum of all direct relations that a dealer has
with other dealers in the network, divided by the total number of dealers
in the network. Using information on the direction of order flow, one can
calculate in-degree dgin from the dealer’s purchases from other dealers and
out-degree dgout from the dealer’s sales to other dealers. Using information
on the volume of order flow, one can calculate several weighted variants:

– dgoutwntrade = Out-degree, weighted by number of trades.


– dginwntrade = In-degree, weighted by number of trades.
140 The Journal of FinanceR

– dgoutwpar = Out-degree, weighted by total par amount.


– dginwpar = In-degree, weighted by total par amount.
r Related to the degree is the k-core. The k-core is defined as the maxi-
mal subnetwork in which each dealer has degree of at least k. For di-
rected graphs, one differentiates between kcoreout and kcorein, the largest
k-cores the dealer belongs to, counting only out-links or in-links.
r Eigenvector centrality ev: A measure of the overall importance of a dealer
firm in the network. It assigns relative scores to all dealers in the network
based on the idea that connections to high-scoring dealers contribute more
to the score of the dealer firm than equal connections to low-scoring dealers.
For weighted graphs, one can calculate several weighted variants:
– evwntrade = Eigenvector centrality, weighted by number of trades.
– evwpar = Eigenvector centrality, weighted by total par amount.
r Betweenness bt: A measure of the absolute position of a dealer in the net-
work. The betweenness of a dealer is computed as the number of shortest
paths linking two dealers in the network that pass through the dealer firm.
Betweenness measures the connections beyond the first neighbors, to take
into account the connections of second- and higher-order neighbors. We
use the directed version of betweenness.
r Closeness cl: A measure of influence with respect to centrality. The close-
ness of a dealer is computed as the inverse of the average number of steps
that a dealer needs to take within the network to reach or be reached
by any other dealer firm. It captures the connection to highly influential
dealers. For directed graphs, clout (clin) is based on out-links (in-links)
only.
We define an aggregate centrality measure Net for each dealer and time
period by aggregating the above network measures. We compute Net as the
first principle component of the above individual centrality measures, in equal-
and value-weighted variants:
r Net (EW): First principle component of the equal-weighted network mea-
sures dgout, dgin, kcoreout, kcorein, bt, clout, clin, and ev.
r Net (VW): First principle component of the value-weighted network mea-
sures dgoutwntrade, dginwntrade, dgoutwpar, dginwpar, evwntrade, and
evwpar.

Appendix D: Explanatory Variables


Dealer characteristics:
r Primary dealer: Dummy variable for primary dealers (mean = 0.15 [stan-
dard deviation = 0.36]).
r NYC dealer: Dummy variable for NYC-headquartered dealers (0.42 [0.49]).
r Underwriter: Dummy variable for the dealer being the underwriter of the
bond (0.09 [0.28]).
Dealer Networks 141

r Dealer size: Natural logarithm of the size of the dealer firm in terms of
assets (6.00 [0.59] in billions).
r Dealer inventory: Aggregate dealer inventory on the day prior to the trade,
standardized by subtracting the average dealer inventory and dividing by
its standard deviation (−0.14 [0.98]). In more detail, to proxy for the de-
sired target level of inventory, we compute for each dealer the moving av-
erage level of the end-of-day inventory over the past 30 days. The moving-
average measure allows for a dealer-specific and time-varying target. It
is also robust to errors in trade reporting that can bias the calculation of
inventory levels due to the fact that we reconstruct dealer inventories by
cumulating past trades. To account for the fact that municipal bond dealers
are different in their levels of tolerance to inventory risk, we standardize
the abnormal inventory measure for each dealer by the standard deviation
of the dealer’s inventory.

Trade characteristics:

r Natural logarithm of the par value of the trade (3.53 [1.39]), interacted
with a trade size dummy to capture potential nonmonotonicity. Trade size
dummies are for Par <$100K, $100K≤ Par <$1M, and Par ≥$1M.

Bond characteristics:

r Maturity: Natural logarithm of the time until the bond matures, expressed
in years (2.27 [0.83]).
r Seasoning: Natural logarithm of the time since the bond was issued (1.46
[0.83]).
r Issue size: Natural logarithm of the bond’s issue size (7.87 [4.16]).
r Credit quality: Indicator variable for rating category (1 = AAA, 2 = AA+,
etc.; mean = 2.65 [SD = 2.68]); dummy for high-yield rated bonds (0.01
[0.08]); dummy for unrated bonds (0.33 [0.47]); dummy for insured bond
(0.58 [0.49]).
r GO bonds: Dummy variable for general obligation bond (0.34 [0.48]). A GO
bond is a municipal bond backed by the credit and taxing power of the
issuing jurisdiction rather than the revenue from a given project.
r Callable: Dummy variable for a bond with a call feature (0.71 [0.45]).
r Sinkable: Dummy variable for a bond with a sinking fund feature (0.34
[0.47]).
r Bank qualified: Dummy variable for a bank-qualified bond that commercial
banks can purchase with tax benefits (0.05 [0.21]).
r Taxable: Dummy variable for a taxable municipal bond, the interest in-
come from which is subject to federal and state taxation (0.02 [0.15]).
r AMT: Dummy variable for a bond subject to the AMT (0.06 [0.23]). An
AMT bond is a private-activity municipal bond whose interest is treated
as a preference item for the purpose of computing the AMT imposed on
individuals and corporations.
142 The Journal of FinanceR

Liquidity conditions:

r Aggregate net outflows from mutual bond funds: Signed aggregate weekly
municipal bond mutual fund net flow (positive numbers correspond to net
inflows, negative numbers to net outflows). Numbers are in billion dollars.
r VIX: Level of the volatility index VIX, in decimals.
r Calendar time controls: End-of-month dummy that equals 1 for the last
three trading days of the month, and 0 otherwise.
r Bond-specific sale pressure by mutual funds: Sale component of the
pressure measure of Coval and Stafford (2007), that is, the sales made
by mutual funds that experienced extreme outflows. Quarterly fund
flows and fund holdings data are from Lipper eMaxx. Sale pressure =
 j (max(0,−
Holdings j,i,t )|Flow j,t <10th percentile)
.
r IssueSize
Bond rating downgrade in notches, current quarter: Number of notches
across the three Nationally Recognized Statistical Rating Organizations
(Moody’s, S&P, Fitch) by which the average rating of the bond has been
downgraded during the current quarter, and zero if the average rating has
been upgraded or has remained unchanged.

REFERENCES
Afonso, Gara, Anna Kovner, and Antoinette Schoar, 2014, Trading partners in the interbank
lending market, Working paper, Federal Reserve Bank of New York.
Afonso, Gara, and Ricardo Lagos, 2015, Trade dynamics in the market for federal funds, Econo-
metrica 83, 263–313.
Babus, Ana, 2012, Endogenous intermediation in over-the-counter markets, Working paper, Fed-
eral Reserve Bank of Chicago.
Babus, Ana, and Peter Kondor, 2018, Trading and information diffusion in over-the-counter mar-
kets, Econometrica 86, 1727–1769.
Barabási, Albert-László, and Réka Albert, 1999, Emergence of scaling in random networks, Science
286, 509–512.
Bech, Morten L., and Enghin Atalay, 2010, The topology of the federal funds market, Physica A:
Statistical Mechanics and its Applications 389, 5223–5246.
Bessembinder, Hendrik, Stacey Jacobsen, William Maxwell, and Kumar Venkataraman, 2016,
Capital commitment and illiquidity in corporate bonds, Working paper, Arizona State Univer-
sity and Southern Methodist University.
Bessembinder, Hendrik, William Maxwell, and Kumar Venkataraman, 2006, Market transparency,
liquidity externalities, and institutional trading costs in corporate bonds, Journal of Financial
Economics 82, 251–288.
Biais, Bruno, 1993, Price formation and equilibrium liquidity in fragmented and centralized mar-
kets, Journal of Finance 48, 157–185.
Cestau, Dario, Richard C. Green, and Norman Schürhoff, 2013, Tax-subsidized underpricing: The
market for Build America bonds, Journal of Monetary Economics 60, 593–608.
Chang, Briana, and Shengxing Zhang, 2015, Endogenous market making and network formation,
Working paper, University of Wisconsin at Madison.
Cocco, Joao F., Francisco J. Gomes, and Nuno C. Martins, 2009, Lending relationships in the
interbank market, Journal of Financial Intermediation 18, 24–48.
Cohen-Cole, Ethan, Andrei A. Kirilenko, and Eleonora Patacchini, 2012, How your counterparty
matters: Using transaction networks to explain returns in CCP marketplaces, Working paper,
University of Maryland.
Dealer Networks 143

Colliard, Jean-Edouard, and Gabrielle Demange, 2014, Cash providers: Asset dissemination over
intermediation chains, Working paper, Paris School of Economics.
Coval, Joshua D., and Erik Stafford, 2007, Asset fire sales (and purchases) in equity markets,
Journal of Financial Economics 86, 479–512.
Craig, Ben, and Goetz von Peter, 2014, Interbank tiering and money center banks, Journal of
Financial Intermediation 23, 322–347.
Demsetz, Harold, 1968, The cost of transacting, Quarterly Journal of Economics 82, 33–53.
Di Maggio, Marco, Amir Kermani, and Zhaogang Song, 2015, The value of trading relationships in
turbulent times, Working paper, Columbia Business School Research Paper No. 15-65.
Duffie, Darrell, Nicolae Gârleanu, and Lasse Heje Pedersen, 2005, Over-the-counter markets,
Econometrica 73, 1815–1847.
Edwards, Amy K., Lawrence E. Harris, and Michael S. Piwowar, 2007, Corporate bond market
transaction costs and transparency, Journal of Finance 62, 1421–1451.
Erdős, Paul, and Alfréd Rényi, 1960, On the evolution of random graphs, Publication of the Math-
ematical Institute of the Hungarian Academy of Sciences 5, 17–61.
Friewald, Nils, and Florian Nagler, 2016, Dealer inventory and the cross-section of corporate bond
returns, Working paper, Norwegian School of Economics and Bocconi University.
Glode, Vincent, and Christian Opp, 2014, Adverse selection and intermediation chains, Working
paper, University of Pennsylvania, The Wharton School.
Gofman, Michael, 2011, A network-based analysis of over-the-counter markets, Working paper,
University of Wisconsin, Madison.
Green, Richard C., Burton Hollifield, and Norman Schürhoff, 2007, Financial intermediation and
the costs of trading in an opaque market, Review of Financial Studies 20, 275–314.
Green, Richard C., Dan Li, and Norman Schürhoff, 2010, Price discovery in illiquid markets: Do
financial asset prices rise faster than they fall?, Journal of Finance 65, 1669–1702.
Harris, Lawrence E., and Michael S. Piwowar, 2006, Secondary trading costs in the municipal
bond market, Journal of Finance 61, 1361–1397.
Hendershott, Terrence, Dan Li, Dmitry Livdan, and Norman Schürhoff, 2016, Relationship trading
in OTC markets, Working paper, Swiss Finance Institute.
Hendershott, Terrence, and Ananth Madhavan, 2015, Click or call? Auction versus search in the
over-the-counter market, Journal of Finance 70, 419–447.
Hollifield, Burton, Artem Neklyudov, and Chester Spatt, 2017, Bid-ask spreads, trading networks,
and the pricing of securitizations, Review of Financial Studies 30, 3048–3085.
Huang, Roger D., and Hans R. Stoll, 1997, The components of the bid-ask spread: A general
approach, Review of Financial Studies 10, 995–1034.
Hugonnier, Julien, Benjamin Lester, and Pierre-Olivier Weill, 2015, Heterogeneity in decentralized
asset markets, Working paper, Swiss Finance Institute.
Jackson, Matthew O., 2005, A survey of models of network formation: Stability and efficiency,
in Gabrielle Demange and Myrna Wooders, eds.: Group Formation in Economics: Networks,
Clubs, and Coalitions (Cambridge University Press: Cambridge).
Jiang, Hao, Dan Li, and Ashley Wang, 2016, Dynamic liquidity management by corporate bond
mutual funds, Working paper, Federal Reserve Board.
Lagos, Ricardo, and Guillaume Rocheteau, 2007, Search in asset markets: Market structure, liq-
uidity, and welfare, American Economic Review 97, 198–202.
Lagos, Ricardo, Guillaume Rocheteau, and Pierre-Olivier Weill, 2011, Crises and liquidity in over-
the-counter markets, Journal of Economic Theory 146, 2169–2205.
Lester, Benjamin R., Guillaume Rocheteau, and Pierre-Olivier Weill, 2014, Competing for order
flow in OTC markets, Working paper, Federal Reserve Bank of Philadelphia.
Malamud, Semyon, and Marzena Rostek, 2014, Decentralized exchange, Working paper, Swiss
Finance Institute.
Neklyudov, Artem, 2013, Bid-ask spreads and the over-the-counter interdealer markets: Core and
peripheral dealers, Working paper, Carnegie Mellon University.
Neklyudov, Artem, and Batchimeg Sambalaibat, 2015, Endogenous specialization and dealer net-
works, Working paper, University of Lausanne and Indiana University.
144 The Journal of FinanceR

Ravasz, Erzsébet, Anna Lisa Somera, Dale A. Mongru, Zoltán N. Oltvai, and A.-L. Barabási, 2002,
Hierarchical organization of modularity in metabolic networks, Science 297, 1551–1555.
Saunders, Anthony, Anand Srinivasan, and Ingo Walter, 2002, Price formation in the OTC corpo-
rate bond markets: A field study of the inter-dealer market, Journal of Economics and Business
54, 95–113.
Schultz, Paul, 2012, The market for new issues of municipal bonds: The roles of transparency and
limited access to retail investors, Journal of Financial Economics 106, 492–512.
Schultz, Paul, 2013, State taxes, limits to arbitrage, and differences in municipal bond yields across
states, Working paper, Notre Dame University.
Upper, Christian, and Andreas Worms, 2004, Estimating bilateral exposures in the German inter-
bank market: Is there a danger of contagion?, European Economic Review 48, 827–849.
Üslü, Semih, 2015, Pricing and liquidity in decentralized asset markets, Working paper, UCLA.
Vayanos, Dimitri, and Tan Wang, 2007, Search and endogenous concentration of liquidity in asset
markets, Journal of Economic Theory 136, 66–104.
Wang, Chaojun, 2016, Core-periphery trading networks, Working paper, Stanford University.
Weill, Pierre-Olivier, 2007, Leaning against the wind, Review of Economic Studies 74, 1329–1354.
Weller, Brian M., 2013, Fast and slow liquidity, Working paper, Northwestern University.

Supporting Information
Additional Supporting Information may be found in the online version of this
article at the publisher’s website:
Appendix S1: Internet Appendix.

You might also like