Customer Liquidity Provision: Implications for Corporate Bond Transaction Costs Jaewon Choi† and Yesol Huh∗ †

University of Illinois at Urbana-Champaign *

Federal Reserve Board First draft: July 2016

Current draft: July 2017

Abstract Convention in calculating trading costs in corporate bond markets is to assume that dealers provide liquidity to non-dealers (customers), and calculate average bid-ask spreads that customers pay dealers. We show that customers often provide liquidity in corporate bond markets, and thus, average bid-ask spreads underestimate trading costs that customers demanding liquidity pay. Compared to periods before the 2008 financial crisis, substantial amounts of liquidity provision have moved from the dealer sector to the non-dealer sector, consistent with decreased dealer risk capacity due to new bank regulations or decreased dealer risk tolerance. Among trades where customers are demanding liquidity, we find that these trades pay 35–50% higher spreads than before the crisis. Our results indicate that liquidity decreased in corporate bond markets and can help explain why despite the decrease in dealers’ risk capacity, average bid-ask spread estimates remain low.

∗ Earlier

drafts were circulated under the title “Customer Liquidity Provision in Corporate Bond Markets.” We are grateful

to Scott Bauguess, Andrew Chen, Darrell Duffie, Terrence Hendershott, Stacey Jacobsen, Yoshio Nozawa, Clara Vega, Brian Weller, and the conference and seminar participants at Bank of Canada, the CFTC, the SEC, 2017 CICF, Workshop on Investor Behavior and Market Liquidity, and Women in Microstructure Meeting for their comments and discussions. The views expressed in this article are soley those of the authors and should not be interpreted as reflecting the views of the Federal Reserve Board or the Federal Reserve System. Please send comments to: [email protected]

1

1

Introduction

Whether corporate bond liquidity has deteriorated has been hotly debated in the recent years since the introduction of new bank regulations such as the Volcker Rule and more stringent capital regulations. On the one hand, market practitioners have argued that bond market liquidity has decreased, citing lower dealer inventory.1 Consistent with this argument, Bessembinder et al. (2016) finds that dealers commit less capital now than they used to, and that this decrease in capital commitment is due to bank-affiliated dealers. On the other hand, others refer to the fact that average bid-ask spreads that customers pay are lower than they were before the crisis as evidence that aggregate liquidity has not deteriorated.2 These competing findings seem puzzling; given that corporate bond markets are dealer-intermediated, we would expect market liquidity to worsen if dealers decrease their risk capacity. In this paper, we provide a potential channel that help reconcile these findings: an increase in liquidity provision by customers. In particular, we argue that bid-ask spreads are seemingly low because average bid-ask spreads that customers pay are a biased measure of the true cost of demanding liquidity. We first show that buy-side investors (‘customers’) often provide liquidity, and this customer liquidity provision causes average bid-ask spreads to underestimate the trading costs that customers demanding liquidity pay. Moreover, as dealers become less willing to take inventory risk, customer liquidity provision increases, which exacerbates the underestimation problem. Once we correct for this bias, the cost of demanding liquidity is substantially higher during the post-regulation period compared with both the pre-crisis period and the post-crisis/pre-regulation period. Thus, our results can help explain why average bid-ask spreads in the post-regulation periods are not higher despite decreased dealer liquidity provision. Liquidity provision by customers may cause observed bid-ask spreads to underestimate the true cost of liquidity demand. Say a customer (C1) wants to sell a bond and contacts a dealer. The dealer, due to limited risk-taking capacity, might be willing to take inventory risk only by charging substantial spreads (e.g., 30 bps). Suppose also that the dealer might search for someone who is willing to buy the bond. Instead of finding someone who has a liquidity need to buy the bond, the dealer might find a non-dealer (C2) who is willing to provide liquidity for a fee (i.e., buying at a lower price than the fundamental value of the bond) even without a strict liquidity need to buy the bond. The dealer then matches the two trades and profits from the difference in spreads. In this scenario, C1 is demanding liquidity and C2 is providing liquidity, while the dealer is not taking any inventory risk. C1, the liquidity-demanding customer, sells at a higher 1 For

example, see https://marketwatch.atavist.com/story/7571. example, see Trebbi and Xiao (2015) and http://libertystreeteconomics.newyorkfed.org/2015/10/ has-us-corporate-bond-market-liquidity-deteriorated.html. 2 For

2

price than she would without liquidity provision by C2, since the dealer is willing to charge a spread smaller than 30 bps as the dealer does not take on any inventory risk. C2 pays a even smaller spread (a negative spread in this scenario), implying that the average bid-ask spread paid by the two customers is much lower than the spread paid by the customer that demanded immediacy. For instance, if C1 paid the dealer 10 bps, and if the dealer paid 4 bps to C2 (i.e., dealer sells to C2 at 4 bps lower than the fundamental price), then the average bid-ask spread is (10 − 4)/2 = 3 bps, which is lower than 30 bps that a customer demanding immediacy will pay and also lower than 10 bps that C1 paid. Note, in this scenario, the transaction between the dealer and C2 occurs at a lower price than the fundamental value, indicating that C2 pays a negative spread.3 Also, the amounts of customer liquidity provision can be measured by the fraction of two matched trades between dealers and customers (DC-DC matched trades, henceforth). To the extent that customers exploit increased spreads quoted by dealers, customer liquidity provision will become stronger as dealers face increased inventory costs. Suppose that the dealer face increased inventory costs. The dealer increases his price for immediate execution to 40 bps. Then some customers that would have traded immediately with the dealer (and paid 30 bps) decides to wait for the dealer to find a counterparty, and some of the customers that would have traded immediately may decide to not trade altogether, since transaction costs would be too high. The dealer will also actively search for liquidity-providing customers, since she would be better off finding matching trades with a liquidity-providing customer than foregoing the transaction with C1, the liquidity-seeking seller. Thus, the fraction of matched trades will go up, and a higher fraction of liquidity provision is essentially done by non-dealers. This is consistent with market commentary that dealers provide less inventories in the post-regulation period, and large buy-side investors are stepping in.4 Moreover, despite the cost of immediacy going up, the average bid-ask spread paid by all customers may remain similar or even decrease due to the shift in the composition of liquidity provision. We show empirical evidence that is consistent with the above predictions. First, we show that customer trades that are matched with other customer trades (DC-DC matched trades) have lower average spreads (and even negative spreads) than customer trades that are not matched. Because we use corporate bond transaction data that have dealer identities, we can distinguish between matched trades and unmatched 3 It need not be the case that C2, the customer providing liquidity, always pays a negative spread. For instance, C1 may be looking to sell a particular high yield bond in the technology sector, and C2 may be a fund that has recently been investing in technology bonds. The dealer, knowing this, may quote a low spread (e.g., 2 bps) to entice C2 to buy the bond. Since C2 pays a much lower bid-ask spread than what he would have if he called the dealer to buy the bond, and since he had no strict liquidity need to buy, he may still be thought of as providing liquidity. 4 For example, BlackRock, a large institutional asset manager, commented that they are not only a price taker, but now also act as a “price maker,” who “expresses a price at which he or she is willing to buy (or sell) a particular security at a given time” (BlackRock (2015)). Also, a recent Wall Street Journal article mentions that “giant bond firms increasingly are taking on a price setting role in global debt markets, elbowing aside big banks facing tighter post-crisis regulation”. (https://www.wsj.com/articles/in-the-new-bond-market-bigger-is-better-1498046401)

3

trades with high precision. DC-DC matched trades measure the amount of customer liquidity provision, and customer trades that are not matched are empirical proxies for trades that dealers provide inventories and immediacy to customers. We show that DC-DC matched trades have 20–40% lower bid-ask spreads compared to trades where customers are demanding liquidity, and by extension, average bid-ask spreads underestimate trading costs for customers demanding liquidity. We also show that the degree of this underestimation problem depends greatly on how bid-ask spreads are measured. Next, we show that in the post-regulation period compared to the pre-crisis period, dealers match a higher fraction of trades and trading costs for unmatched trades have increased substantially. These results are consistent with liquidity being lower due to increased dealer inventory costs. The fractions of trades that are immediately offloaded from dealer inventories increased by 25–33% from the pre-crisis levels. When we restrict our sample to customer trades that are not matched, we find that trading costs in the post-regulation period is 7–11 bps higher than the trading costs in the pre-crisis period and that this estimated increase is 2 to 3 bps larger than when we use the full sample. Given that average bid-ask spreads for trades above $1 million are approximately 20 bps for the non-crisis periods, the 7–11 bps overall increases are economically substantial, and the 2–3 bps differences in the estimated change in trading costs are quite significant. Thus, customer liquidity provision increased post-regulation, and this increase leads the trading cost measures that are often used in the literature to underestimate the change in true trading costs for liquidity-demanding customers. This may explain why despite the decrease in dealers’ willingness to hold inventories, average trading cost estimates have not worsen. A few additional points should be made. Although we do not exactly pinpoint what caused the decrease in dealers’ risk taking and the increase in customer liquidity provision, our results also show that large dealers, who generally are bank-affiliated dealers affected by the Volcker Rule, drove the change, consistent with regulations having played a role. Moreover, the fact that dealers increased their offloadings to customers is consistent with their incentives to comply with the Volcker Rule. One of the key Volcker metrics that banks have to report is the fraction of trades that are conducted with customers. Hence, offloading inventories to customers is more advantageous in terms of compliance with the Volcker Rule than trading with other dealers, which can explain our findings. However, the results may also be due to changes in the risk bearing capacities of large dealer banks or risk management practices that may not be directly caused by regulations. Our main objective is to show that customer liquidity provision have increased and that this increase caused the average bid-ask spread to underestimate the change in liquidity. Also, given that the impact of regulations should matter most for large trades, for the most part of the paper, we focus our analysis to trades with

4

volume $1 million and above. Remainder of the paper is organized as follows. Section 2 reviews the relevant literature, and Section 3 describes the data and empirical measures. In Section 4, we show that customers often provide liquidity, and examine the impact of customer liquidity provision on various spread measures. In Section 5, we study how customer liquidity provision and trading costs for liquidity-demanding customers change after the bank regulations. Lastly, Section 6 concludes.

2

Literature Review

This paper is most closely related to a number of contemporaneous empirical papers that study the impact of regulations on corporate bond market liquidity.5 Overall, there are three main findings that emerge from these papers. First, dealers have decreased capital and inventory provision in the recent years, and the dealers that are more affected by various bank regulations have decreased more. Bessembinder et al. (2016) show that dealers commit less capital in the post-regulation period, and this reduction in committed capital is driven mainly by bank-affiliated dealers. Bao et al. (2016) find that dealers that are affected by the Volcker Rule provide less inventory service during downgrade events in the post-regulation period than they did before the regulation period. Adrian et al. (2017) find that bonds that are traded by dealers that are affected more by regulations have seen a larger decline in liquidity in recent years. Second, however, these studies find that broad price-based measures of liquidity such as average bid-ask spreads have not worsened. Trebbi and Xiao (2015) test whether there is a discontinuity in liquidity around the time when regulations were introduced and find no evidence of liquidity deterioration. Bessembinder et al. (2016) find that although dealers commit less capital, average trading costs remain largely similar to pre-crisis levels. Adrian et al. (2016) and Anderson and Stulz (2017) find similar results. Lastly, a number of papers look at specific market stress or liquidity events and find that liquidity during these times have worsen in the post-regulation period. Dick-Nielsen and Rossi (2016) study index exclusion events, and Bao et al. (2016) study bond downgrade events, and both find that liquidity during these events is worse than it was before the crisis. Anderson and Stulz (2017) find that liquidity is worse in the post-regulation period when the VIX spikes up but not during bond idiosyncratic events. Our paper adds to this literature in a few ways. We bridge the seemingly contrasting findings by showing that customer liquidity provision can help explain why average bid-ask spreads that customers pay may 5 A few theoretical papers also examine the effect of regulations on market liquidity. Cimon and Garriott (2016) argue that the Volcker Rule and capital regulations motivate dealers to switch to trading in an agency basis. Uslu (2016) finds that the welfare impact of the Volcker Rule is not clear.

5

seem low despite dealers’ committing less capital and providing less inventories. We also propose how to measure the costs for demanding liquidity without using specific market stress or liquidity events and show that liquidity has worsen broadly. There is a large literature that studies liquidity in corporate bond markets using bid-ask spreads. Trading costs in the corporate bond market is usually measured as average bid-ask spreads that customers pay. Hong and Warga (2000), Chakravarty and Sarkar (2003), and Adrian et al. (2015) compute daily bid-ask spreads for each bond as differences between average customer buy prices and sell prices. Feldh¨ utter (2012) use price differences between trades in the same bond and quantity within a short time window as dealer roundtrip costs. These trades are likely to be pre-arranged trades where a dealer is acting as an agent. On the other hand, Edwards et al. (2007) and Bessembinder et al. (2006) employ a regression-based methodology to estimate bid-ask spreads. We use differences between customer trading prices and interdealer prices to calculate bid-ask spreads; in Section 4.2, we will show how the selection bias that is built into bid-ask spread calculations affect the measurement of trading costs. Numerous other papers use lower frequency proxies such as price impact or daily autocorrelations. For a more complete review of various liquidity measures in corporate bond markets, see Schestag et al. (2016). This paper is also related to the literature studying the relationship between market liquidity and dealer inventory costs. If bank regulations increase dealer inventory costs, it would impact market liquidity. Brunnermeier and Pedersen (2009) model how funding liquidity impacts market liquidity, and Nagel (2012) shows that equity market liquidity moves with VIX, which proxies for market maker funding liquidity or risk aversion. Dick-Nielsen et al. (2012) and Randall (2015) find that the increase in corporate bond dealer inventory costs during the financial crisis decreased corporate bond liquidity. Paired trades where dealer effectively act as an agent is fairly common in corporate bond markets. Zitzewitz (2010), Ederington et al. (2014), and Randall (2015) study paired trades where a dealer-customer trade is paired with an interdealer trade, thus the dealer in the middle does not take any inventory but the second dealer provides liquidity. Zitzewitz (2010) show that this type of trade happens frequently in small trades and is much more common than a dealer-customer trade paired with another dealer-customer trade. In contrast, in the large trade size that we focus on, paired trades are more likely to be between dealer-customer trades, consistent with the results in Harris (2015), who study the relationship between paired trades and trade-throughs. Goldstein and Hotchkiss (2017) also show that dealers actively manage inventories by pre-arranging trades or offsetting trades during the same day.

6

3

Data

We use the regulatory Trade Reporting and Compliance Engine (TRACE) feed, which includes dealer identities for each trade. Customers are not identified and are marked as “C” only. This dataset is similar to one used in Goldstein and Hotchkiss (2012). Other than having dealer information, the rest of the data are similar to the Enhanced TRACE and include trade information such as trade date and time, volume, price, CUSIP, trading capacity (principal or agent), and trade direction. Because trades are intermediated and reported by dealers, trades are either between dealers (interdealer) or between a dealer and a customer (dealer-customer). Our sample period is January 2006 to June 2015. We start the sample period in 2006 as the trade information dissemination was introduced in multiple phases from 2002 to 2005, and we want to avoid the impact of increasing transparency in our empirical exercises (Goldstein et al. (2007)). We exclude MTNs, 144As, and exchangeable bonds and keep fixed coupon bonds only. We also exclude bond-days with less than 30 days since issuance. Dealers are identified by Market Participant Identifiers (MPIDs). Some dealers may have multiple MPIDs (they can be different subsidiaries) or shift MPIDs over time. Thus, we construct MPID2, in which MPIDs from the same dealers have the same MPID2. When two dealers merge, we attribute the acquired dealer’s MPID to acquiring dealer’s MPID2 after the merger. We delete trades between the same MPID2. To study the potential effects of bank regulations, we need to identify when regulations came into effect. Unfortunately, a clear start date most likely does not exist. For example, the Volcker Rule was first endorsed in 2010 (included in the Dodd-Frank Act) and was originally scheduled to go into effect in July 2012. However, the final rule did not get approved until 2014 and went into effect for most large banks in July 2015. Because there was a large time lag between its inception and the implementation deadline, most banks are considered to have made necessary changes in their business model before July 2015.6 Similarly, higher capital requirement such as the Basel III implementation and G-SIB surcharge were being implemented in phases over a long period of time. Thus, it is difficult to pinpoint the correct start date. Moreover, because the implementation periods were significantly long, it is likely that different banks adopted the regulations at different times, and the decision as to when to adopt them could have been an endogenous choice. Keeping the above caveats in mind, we follow Bessembinder et al. (2016) and use July 2012 as the regulation start date. We also follow their subperiod definitions for the most part and divide the sample period into four subperiods: January 2006 to June 2007 as the pre-crisis period, July 2007 to April 2009 6 http://www.wsj.com/articles/volcker-bank-risk-rule-set-to-start-with-little-fanfare-1437517061

7

as the financial crisis period, May 2009 to June 2012 as the post-crisis period, and July 2012 to June 2015 as the post-regulation period. Most of our empirical analyses will focus on comparing the post-regulation period with the pre-crisis and the post-crisis periods.

3.1

Dealer classification

If two trades by the same dealer for the same bond have the same quantities with opposite trade directions and occur less than one minute apart, then they are most certainly pre-arranged, where the dealer acted as an agent. As shown in Bessembinder et al. (2016) and Zitzewitz (2010), these trades are fairly common and are not always marked as agency trades (as opposed to principal trades) in TRACE. Additionally, there are dealers that almost exclusively engage in agency or riskless principal trades. For example, trading platforms, interdealer brokers, or certain introducing brokers may appear as such. Excluding these dealers are important when examining inventory holding periods, as they are somewhat different from typical dealers that perform a conventional role as market makers. To identify such dealers, we first calculate the fractions of the numbers of trades and trading volumes that are paired over the full sample period for each dealer. If more than half of trades and volumes are paired or if more than three-fourths of trades or volumes are paired for a dealer, then we classify that dealer as a one who engages primarily in riskless trading. We exclude trades where these dealers appear as a reporting party. Also, because we are interested in trading costs that clients face, we also exclude dealers that almost exclusively engage in interdealer trades. To define large dealers, we first rank dealers every month by trading volumes with customers. If a dealer is in the top ten for ten months or more, we define that dealer as a large dealer. There are 15 large dealers, and all other dealers (that are not excluded in the above) are classified as small dealers.

3.2

Inventory holding periods

We study two different dealer inventory holding periods: less than 15 minutes (short holding) and overnight. Overnight inventories proxy for the amounts of trades that dealers are willing to take on to their inventories for more than a day, and inventories held for less than 15 minutes proxy for the amounts of trades that dealers pre-arrange and do not take on risk for. We use 15 minutes because dealers are required to report trades to TRACE within 15 minutes. For each dealer-bond-day, we start the beginning of the day with an inventory of zero and accumulate the inventory throughout the day using the last-in-first-out (LIFO) method. Only trades conducted in principal 8

capacity are included for inventory calculation, although results remain qualitatively similar if we include all agency trades as having holding periods less than 15 minutes. We keep track of how long each trade remains in dealer inventories and which trades it is matched to. Table 1 gives a simple example of holding period calculation. Panel A provides fictitious trading data for a particular bond-dealer-day for illustrative purposes. Buy/sell indicators are from the perspective of the dealer. Panel B shows the result of LIFO inventory accumulation calculations for customer trades using the sample data. For example, the inventory of -200 accumulated from trade number 1 will leave when trade number 2 happens five seconds later. We set the ‘other side’ of trade 1 to be trade 2, and the inventory holding period to be 5 seconds. Trades may not always be exactly matched in terms of volume. 350 out of 500 in trade 4 is matched against trade 5 within 15 minutes, 100 is matched against trade 6 but with a 40 minute holding period, and 50 remains unmatched at the end of the day. A trade or a portion of a trade with a holding period less than 15 minutes is defined to have a short holding period, and the ones that remain in inventories at the end of the day are defined to be overnight inventories. For instance, all of trade 1 and a portion of trade 4 are defined to have short holding periods, and all of trade 3 and a portion of trade 7 are overnight inventories. For customer trades with short holding periods, we further classify them as DC-DC or DC-ID by whether they are matched with another customer trade (DC-DC) or they are interdealer trades (DC-ID). For instance, a portion of trade 7 that is matched with trade 8 is a DC-DC match and the portion matched with trade 9 is a DC-ID match. To calculate monthly (or daily) short holding and overnight inventories, we calculate the fractions of dealer-customer trading volumes that are short duration and the fractions that are overnight inventories. In this example, the daily short holding inventory and overnight inventory would be 1350/2050 and 500/2050, respectively. We only use customer trades for calculations, but the results using both customer trades and interdealer trades are similar. In Panel C, we assign a trade type at the trade level for all dealer-customer trades. For each trade, we calculate the volume-weighted average of DC-DC, DC-ID, and short-holding trades. If more than 50% of the trade is DC-DC match, for example, then the trade would be classified as a DC-DC trade. If 50% or less of the trade is short holding, then the trade would be classified as a invt>15min trade. We can also classify whether a given trade stays in inventories overnight in a similar fashion. Table 2 presents summary statistics by dealer size and trade size. On average, large dealers take on more inventory risk; they have higher overnight inventories and pre-arrange fewer trades. Large dealers also account for 75% or more of trading volumes with customers. Overall, DC-DC matched trading volumes are

9

higher than those of DC-ID matched trades, especially for large dealers. This is because for large trades, DC-DC matched trades have four times the volume of DC-ID matched trades, as shown in Panel B. For example, for high yield bonds, 33% of customer trades that are $1 million or larger are matched with other customer trades, while less than 4% of customer trades are matched with interdealer trades. For small trades, DC-ID trades have five times the volume of DC-DC matched trades. Given that trades that are $100K and smaller are much more frequent than trades that are $1 million and above, DC-ID matched trades are much more frequent than DC-DC trades in terms of trade count, consistent with the results in Zitzewitz (2010) and Harris (2015). A few comments are in order. First, overnight inventory and short holding measures might seem similar in concept to the capital commitment measure in Bessembinder et al. (2016). However, because our inventory holding periods are calculated for each bond separately, a buy trade in one bond does not net out a short position in a different bond. In Bessembinder et al. (2016), the capital commitment measures are calculated at a portfolio level and net out long and short positions in different bonds in the same portfolio, and thus capital commitment tends to be calculated as a much smaller fraction of the overall trading volumes than our overnight inventory measures. For example, their average capital commitment is less than 1% of trading volumes, while the average of overnight inventories is around 50%. While common risks may be hedged by having a long position in one bond and a short position in a similar bond, in terms of regulatory capital needed against the positions, having a long-short portfolio generally needs higher capital requirements compared with having no positions in either bonds. Second, our short-holding measures are similar to paired trade measures in Zitzewitz (2010), Randall (2015), and Harris (2015) and the ‘effectively agent’ trade measures in Bessembinder et al. (2016) and Bao et al. (2016), but we do not restrict trade volumes to match exactly, and we allow a trade to be matched with multiple opposite orders. We also study the DC-DC and DC-ID matched trades separately, which is important for our paper. Third, while we assume that inventories are wiped clean every morning, both the short holding and overnight inventory measures remain similar when we assume each trade is wiped from inventories 30 days after the trade if it has not left the inventories yet. Lastly, it is possible to calculate inventory levels by dealer or for each dealer-bond pair with our data by accumulating trades, but our preliminary results indicate that inventory calculations can be very sensitive to assumptions about initial levels, any data error and data cleaning, and issues such as how to deal with non-zero inventories at maturity. Thus, we focus on inventory holding periods instead.

10

3.3

Estimating bid-ask spread

We calculate bid-ask spreads for customer trades with par values greater than $1 million where dealers act in principal capacity. A reference price for a customer trade is calculated as the volume-weighted average of interdealer trade prices in that bond-week, using interdealer trades larger than $100,000. Reference prices are missing when there are no interdealer trades during the week for a given bond.7 For each trade, we then calculate the spread as: spread = 2Q ×

traded price − reference price reference price

(1)

where Q is +1 for customer buy and −1 for customer sell. We multiply by 2 to get the full spread. Note that we calculate bid-ask spreads only for customer trades (i.e., we do not calculate spreads for interdealer trades). Thus, for DC-ID pairs, we are only using the customer trades and not the interdealer trades to calculate average bid-ask spreads. We complement this spread measure using a dealer profit measure, which considers both buy and sell prices for dealers. Specifically, dealer profits are calculated as a difference between dealer sell and buy prices. Table 3 presents average bid-ask spreads and dealer profits by rating and trade types. Spreads are lowest for DC-DC trades, and highest for DC-ID trades. Dealer profits, on the other hand, are the lowest for DC-DC trades and highest for invt>15min trades.

3.4

Other data

We use the Mergent Fixed Income Database (FISD) to get corporate bond characteristics such as size, offering date, maturity, and rating. For index returns on investment grade and high-yield corporate bonds, we use the BofA Merrill Lynch indices.

4 4.1

Customer Liquidity Provision Is There Customer Liquidity Provision?

Do some clients provide liquidity and get compensated for it? And if so, to what extent do they provide liquidity? In this section, we provide empirical evidence that points to customer liquidity provision, and argue that DC-DC matched trades are likely driven by customer liquidity provision. 7 Interdealer trades tend to be smaller and are less frequent, hence we use the $100,000 cutoff instead of $1 million and use weekly average instead of daily average to increase the number of observations. Results are qualitatively similar if we use daily average interdealer price as the referece price.

11

To measure customer liquidity provision and study its impact on bid-ask spread estimates, we focus on short-holding trades. We hypothesize that DC-DC matched trades are largely instances where one customer demands liquidity and the other provides liquidity. In contrast, we hypothesize that DC-ID matched trades are generally driven by customer demanding liquidity and the second dealer providing liquidity, so these trades will have the lowest occurrences of customer liquidity provision. Although some unmatched trades can be associated with customer liquidity provision, dealers tend to use inventory capacity in such trades and thus unmatched trades are less likely driven by customer liquidity provision. Thus, DC-DC trades will involve the highest occurrences of customer liquidity provision, and DC-ID trades will have the lowest. As outlined in the introduction, when customers provide liquidity, they are compensated with a small or even negative spread (i.e., buying at a lower price or selling at a higher price than the fundamental value). Thus, trades with high customer liquidity provision will have low average bid-ask spreads. Our hypothesis predicts that DC-DC matched trades have the lowest average bid-ask spreads and DC-ID trades have the highest.8 An alternative hypothesis is that all matched trades are largely driven by matching two sides with opposite liquidity needs, and neither side is providing liquidity. In this case, DC-DC matched trades will not necessarily have more instances of negative spreads, since none of the customers needs to be compensated. Moreover, the average spread of DC-DC trades will not be any smaller than the average spread of DC-ID trades or invt>15min trades. In Figure 1, we plot the distribution of spreads for the three types of dealer-customer trades with par value greater than $100,000: invt>15min trades, DC-DC trades, and DC-ID trades. Panel A presents the distribution of dollar spreads. In Panel B, we divide the spreads by the weekly price dispersion in the bond. In both plots, DC-DC matched trades have the smallest spreads and have the highest fraction of negative spreads. In contrast, DC-ID matched trades have the largest spreads and have substantially smaller fraction of negative spreads. These results are consistent with our hypothesis that DC-DC trades are most likely to be driven by customer liquidity provision. In Table 4, we formally examine whether DC-DC trades have lower spreads relative to the other trade types in a regression setting. We run the following model: spreadi,j,t,k = β2 1(DC − DC)k + β3 1(invt > 15min)k + Ai + Bj + Ct

(2)

where spreadi,j,t,k is the spread, defined in (1), of trade k between dealer j and a customer for bond i on 8 We calculate bid-ask spreads for customer trades only, so for DC-ID trades, we are only using the customer trade and not the interdealer trades to calculate the average bid-ask spreads.

12

day t.

1(DC − DC)k and 1(invt > 15min)k are dummy variables indicating whether trade k is a DC-DC

matched trade or a long-holding-period trade. The dummy variable for DC-ID matched trades is omitted due to multicollinearity and thus form the base level. We restrict the sample to trades over $1 million in this regression. Table 4 provides the results for (2). First two columns for each rating category are for the full sample, and the last column for each rating category presents the results when data is restricted to large dealers. Consistent with our hypothesis, DC-DC matched trades have lowest spreads and DC-ID matched trades have highest spreads. For example, compared with DC-ID spreads, spreads for DC-DC trades are approximately 20–30bps lower, and spreads for invt>15min trades are 10–25 bps lower, for both investment grade and high yield samples. Given that the average spread for trades $1 million and above is approximately 25 bps, these coefficient estimates show a substantial difference in spreads between non-DC-DC trades and DC-DC trades. These results are consistent with our hypothesis that DC-DC matched trades are driven mostly by customer liquidity provision, whereas DC-ID trades are least likely associated with customer liquidity provision. We also find that our empirical results do not support the alternative hypothesis that dealers are simply matching two liquidity-seeking customers—note that a substantial fraction of customer buys (sells) in DCDC matched pairs occur at prices lower (higher) than reference prices ( a proxy for the fundamental value of a bond), as shown in Figure 1. Therefore, DC-DC trades have the highest instances of customer liquidity provision, and we will use DCDC trades to proxy for the degree of customer liquidity provision. Also, the higher the customer liquidity provision, the more likely average bid-ask spreads underestimate the cost of demanding liquidity.

4.2

Impact of Customer Liquidity Provision on Same-Day Spreads

An alternative method of calculating bid-ask spreads that is more frequently used is to take the difference between customer buy and customer sell prices instead of using interdealer prices as reference prices.9 We calculate bid-ask spreads using this method (we will call this “same-day spreads”) to compare with bid-ask spreads we calculate in Section 3.3. We will show that by construction, same-day spreads contain a higher fraction of DC-DC matched trades, leading to a more severe underestimation of trading costs for customers demanding liquidity. 9 For

example, see Adrian et al. (2015).

13

For bond i on day t, a same-day spread is measured as follows:

spread2i,t =

vwavg(customer buy)i,t − vwavg(customer sell)i,t (vwavg(customer buy)i,t + vwavg(customer sell)i,t )/2

(3)

where vwavg stands for volume-weighted average. To get daily averages, we take averages across bonds. To calculate same-day spreads by trade size groups (less than $100,000, $100,000–$1million, $1million and above), we simply restrict the customer trades used in (3) to those in the corresponding trade size group. This procedure reduces the number of trades that are included in spread calculations. Figure 2 plots the five day moving averages of same-day spreads by trade size and rating group. The figure shows that same-day spreads have peaked during the 2008 financial crisis and decreased to a pre-crisis level in the latter part of the sample (i.e., after 2014). In Panel D, for example, same-day spreads for high-yield bonds have reached almost to a historical low in mid-2014. Since the same-day spread calculation requires both customer buys and sells above $1million for a given bond for a day, the calculated spreads are likely to be based largely on trades where customers are providing liquidity (i.e., DC-DC matched trades will be overweighted in same-day spread calculations). Given that DC-DC trades have lower spreads due to customer liquidity provision as shown in Section 4.1, same-day spreads will underestimate true transaction costs of liquidity demanders. This may also affect the time series patterns if this issue becomes more aggravated over time. To gauge the severity of the underestimation problem, we examine three different subsamples, all based on dealer-customer trades $1 million or above: the full sample, the sample of trades that are used in calculating same-day spreads (the same-day sample), and the sample based on trades that are used in calculating the spreads in Section 3.3 (the spread sample). The full sample consists of all dealer-customer trades above $1million. The same-day sample consists of all trades that have both buys and sells from customers for a given day and bond. The spread sample consists of all customer trades with available interdealer trades during the same week. For each of the three sample, we calculate the fractions of DC-DC matched trades, DC-ID matched trades, and invt>15min trades. Table 5 presents the results. As expected, the same-day spread sample has a much higher fraction of DCDC match trades than the other two samples. For investment grade bonds, for example, 29.0% of same-day spread sample trades are DC-DC match trades, compared with 14.8% for the full sample, and for high yield bonds, the numbers are 39.1% and 27.7%, respectively. A back-of-the-envelope calculation in investment grade bonds yields that compared with the full sample, the same-day sample underestimates bid-ask spreads

14

by about 14%×11 bps=1.5 bps, which is roughly 7.5% of the bid-ask spreads for trades above $1 million. Perhaps not surprisingly, our spread calculation based on interdealer trades as a reference price (i.e., (1)) overweights the DC-ID match trades; however, the difference is fairly small, about 2 percentage points. Thus, same-day spreads underestimate liquidity costs the most as it overweights DC-DC matched trades. The underestimation may have increased over time, since customer liquidity provision has increased during the post-regulation period with dealer inventory costs increasing and assets under management in the customer sector (e.g., mutual funds) growing.10 Taking simple averages of our spread estimates based on (1) also underestimates the true spread paid by liquidity-demanding customers, but to a lesser degree.

5

Customer Liquidity Provision and Trading Costs Before and After the Banking Regulations

5.1

Overview of the Recent Banking Regulation Changes

In this section, we briefly discuss the recent post-crisis banking regulations and their impact on corporate bond market liquidity. Because bank-affiliated dealers intermediate majority of the trades in the U.S. corporate bond market, these regulations may affect dealer liquidity provision. We focus on the Volcker Rule and capital regulations (i.e., “Basel III”), as we believe these rules are the most relevant for the corporate bond market.11 The Volcker Rule prohibits banks from engaging in proprietary trading, except when making markets. As Duffie (2012) argues, however, the Rule will also discourage banks from market-making and providing liquidity, since distinguishing between market making and proprietary trading may be difficult. For example, the Volcker Rule specifies seven quantitative metrics to be reported for certain banks at the trading desk level. Both the inventory turnover and inventory aging metrics benefit from a quick turnover of the inventory. These metrics may discourage dealers from taking on inventory, especially for illiquid bonds. In addition, the ‘customer facing trade ratio’ measures the fraction of total trades that occur with a customer. Hence, all else equal, dealers would prefer to offload inventories to customers rather than in the interdealer market. This is consistent with the rise in customer liquidity provision that we document in this paper. Basel III overall establishes tougher capital rules through increases in the minimum captial ratio, among 10 Preliminary results indicate that using round trip trading costs as in Feldh¨ utter (2012) will pick up DC-DC and DC-ID trades almost exclusively and underestimate the spread even more. 11 Other rules such as the supplementary leverage ratio may also be relevant; see Adrian et al. (2017) for other post-crisis regulations.

15

many other changes. Stricter capital regulations increase the cost of financing inventories for dealers. This would then lead to banks’ decreasing liquidity provision, especially in trades that take up a lot of space on the inventories and are expected to remain longer. Since the 2008 financial crisis, corporate bond holdings of banks and broker-dealers have decreased. Instead, asset managers, especially mutual funds, have increasingly been holding a larger fraction of outstanding corporate bonds. This has prompted some concerns about outflow risk (Feroli et al. (2014)). On the other hand, market commentary indicates that these changes have pushed buy-side investors to increasingly be liquidity providers and price setters in the corporate bond market. For instance, a recent Bloomberg article documents the increasingly active role that buy-side investors play, and argues that they have gone from being price takers to price makers.12

5.2

Customer Liquidity Provision Over Time

In this section, we examine whether in the post-regulation period, customer liquidity provision increased, using the fraction of DC-DC matched trades as a measure of customer liquidity provision. An increase in the fraction of DC-DC matched trades in the sample will lead to more severe underestimation in liquidity costs in the corporate bond market. In Figure 3, we first examine the time series of the fractions of DC-DC and DC-ID matched trades. Figure 3 shows that the fractions of DC-DC matched trades indeed increased in the post-regulation period, while the fractions of DC-ID matched trades remained similar or decreased. The plots are consistent with the notion that there has been a shift from principal trading by dealers to a pre-arraged, search-and-match trading model, and with that change, increasing fractions of trades are associated with customer liquidity provision. While the time series plots hint at regulations playing a role, other changes in the market may also be driving the results. Thus, we examine how the fractions of DC-DC matched trades change differentially for large and small dealers. Most large dealers are affiliated with banks and are affected by various bank regulations, although the exact regulatory capital requirements might be different and the extent to which they have to alter their behavior due to these regulations might also be different. On the other hand, small dealers are a mix of bank-affiliated and non-bank dealers. In Figure 4, we plot short-holding trades separately for large and small dealers.13 Short-holding trades increase for large dealers, while there is no 12 See

https://www.bloomberg.com/news/features/2016-08-15/the-rise-of-the-buy-side we can plot the fraction of DC-DC trades. Since DC-DC volume are much higher than DC-ID volume in large trades and DC-DC trades have increased over time while DC-ID trades have not, most of the time series pattern in short-holding trades are dominated by that of DC-DC trades. Consistent with this, results are similar when we use DC-DC 13 Alternatively,

16

distinct pattern for small dealers. Results are similar for overnight inventories (presented in Appendix A). While these are not precise tests of causality, these results are consistent with regulations having an effect on dealers’ inventory holding periods and increasing customer liquidity provision. In Figure 5, we examine whether the increased short-holding trades are driven by outliers. Given that some of the large dealers have a substantial fraction of the total trading volume, the increase in short-holding trades can be due to one large dealer increasing the amounts of short-holding trades substantially. Figure 5 shows that both the average and the median across large dealers are higher post-regulation compared to the pre-crisis and the post-crisis levels.14 In Tables 6 and 7, we show the above results on short-holding trades more formally in a regression setting. For each rating group (investment grade, high yield), we run the following three regressions using daily data:

Aggregate level : yt = α +

4 X

βl 1(t ∈ Tl ) + t

(4)

l=2

Dealer group level : ym,t = α1 1(large)m + α2 1(small)m +

4 X

1(large)m βlarge,l 1(t ∈ Tl )

l=2

+

4 X

1(small)m βsmall,l 1(t ∈ Tl ) + m,t

(5)

l=2

Individual dealer level : yj,t =

X j

Dj +

4 X

βl 1(t ∈ Tl ) + j,t

(6)

l=2

where yt , ym,t , and yj,t are the fractions of short-holding trades at the aggregate, dealer group, and individual dealer levels, respectively. For the aggregate level regression, we also look at the fraction of DC-DC trades. Tl (l = 1, . . . , 4) are the four subperiods (pre-crisis, financial crisis, post-crisis, post-regulation). We also include the VIX and bond market volatility. For the dealer group level regression, we use the average fraction of short-holding trades calculated separately for large and small dealers, and the 15 large dealers are used for the individual dealer level regression. We also control for the VIX and bond index volatility and include dealer fixed effects for the dealer level regression. For the individual dealer regression, we run both OLS and median regressions. The median regression is used to rule out the possibility that the results are driven by one or a few large dealers. Table 6 presents the estimation results for the aggregate and the group level regressions. T1 , the precrisis period, is the omitted level, so the coefficients for the other subperiods measure the difference from trades instead of short-holding trades in Figure 4 and 5. 14 The post-regulation increase in short-holding trades is less acute for the median, which implies that there potentially are one or a few dealers with large increases in short-holding trades. These outliers seem to be mostly explained by dealers trading with their affiliates; see Appendix B for further explanation and related robustness checks.

17

the pre-crisis levels. β4 − β3 indicate the difference between the post-regulation and the post-crisis period coefficients. Consistent with Figures 4 and 5, the results indicate that customers provide more liquidity and dealers take on less risk in the post-regulation period compared with the pre-crisis and the post-crisis periods. These changes in customer liquidity provision and dealer risk taking are driven mainly by large dealers, consistent with the regulation effect story outlined above. For example, in investment grade bonds for the post-regulation period, short-holding trades increase by 4.4 percentage points compared with the pre-crisis value, and almost all of that increase is due to the increase in DC-DC trades. Column (3) and (6) show that large dealers increase short-holding trades while small dealers remain the same or decrease short-holding trades. The results for overnight inventories are qualitatively similar and are presented in the Appendix A. Table 7 shows the dealer level regression results. Both the OLS and median regressions indicate that there are more short duration trades in the post-regulation period compared with the pre-crisis and the post-crisis periods. Moreover, the estimated increase is larger in the OLS regression than in the median regression, consistent with Figure 5. For example, the OLS regression would indicate that the short duration trades increase by 7.5 percentage points compared with the pre-crisis period, while the median regression indicates a 4.3 percentage point increase. Overall, the results in this section show that customer liquidity provision increased in the post-regulation period compared with the pre-crisis and the post-crisis periods. This increase is driven mostly by large dealers, which is consistent with bank regulations having had an impact.

5.3

Trading Costs Over Time

So far, we have shown that the fractions of short-holding trades have increased post-regulation and that, because of customer liquidity provision, the usual bid-ask spread estimates in the literature underestimate what liquidity demanding customers would pay. Then, is it possible to calculate the spreads that liquiditydemanding customers would pay, and if so, are these spreads higher in the post-regulation period? Do the usual calculations of bid-ask spreads used in the literature underestimate this increase in trading costs for liquidity-demanding customers? To measure the spreads that liquidity-demanding customers pay, we need to distinguish between trades where customers are demanding liquidity from dealers and trades where customers are providing liquidity to the dealers. Then, spreads for trades where customers are taking liquidity would be the correct spreads to look at. As shown in Section 4.1, DC-DC matched trades have a high fraction of customer liquidity provision,

18

and invt>15min trades are largely trades where dealers provide immediacy to customers. Thus, focusing only on the invt>15min trades would alleviate the underestimation issue.15,16 Therefore, we examine whether spreads have widened more by looking at only invt>15min trades, and compare against two alternative measures. In Table 8, we examine whether bid-ask spread estimates based on invt>15min trades are indeed wider for the post-regulation period through regression analyses. In particular, we regress bid-ask spread estimates separately for the all-trade sample, the invt>15min trade sample, and the same-day spread sample. For the first two samples, we run the regressions using trade level data, and for the same-day spread sample, we use bond-day level data. Specifically, we estimate the following models:

trade level: spreadi,t,k = α +

4 X

βl 1(t ∈ Tl ) + β5 (bond i char) + β6 log(volk ) + β7 mktt + i,t,k

(7)

l=2

bond-day level: spreadi,t = α +

4 X

βl 1(t ∈ Tl ) + β5 (bond i char) + β6 log(voli,t ) + β7 mktt + i,t

(8)

l=2

where spreadi,t,k is the bid-ask spread for bond i, day t, trade k. bond i char are bond characteristics such as outstanding amounts, ratings (finer scale), age, and time-to-maturity. vol is the trade volume, and mktt are market controls such as the bond index volatility calculated as a 20-day rolling standard deviation and the VIX. Since same-day spreads are calculated at the bond-day level, regressions for the same-day spread sample are done at the bond-day level. Table 8 presents the regression results for the full sample and invt>15min trades. Same-day spread results are presented in Table 9. Compared with the pre-crisis period, spreads are higher in the post-regulation period across all samples and ratings groups. These results may partially be driven by sampling issues, as we are using trade level data, which inherently put more weight on frequently traded bonds. Another possibility is to create portfolios using bond characteristics, and using the average spread for each portfolio to run the regression at the portfolio-day level. However, to the extent that our focus is on comparing the results across samples, this potential sampling issue is not problematic. Most importantly, the estimated changes in trading costs are the largest in the invt>15min sample across all tests. For example, for investment grade bonds, spreads are 10.7 bps higher in the post-regulation period than in the pre-crisis period (when market volatilities are controlled for) for the invt>15min sample, 15 We

also exclude DC-ID matched trades to focus on trades were dealers provided immediacy. we argue in Section 4.1 that customer liquidity provision channel is stronger than dealers matching two customers with opposite inherent liquidity needs, it may be the case that the second scenario still matters in the time series. Even in this case, we would still focus on invt>15min trades to look at spreads paid by customers demanding immediacy. 16 Although

19

compared with 9.0 bps for the full sample and 6.4 bps for the same-day spread sample. For high yield bonds, spreads are higher by 6.8 bps (invt>15min), 3.9 bps (full sample), and 1.1 bps (same-day spread). When we compare the post-regulation period with the post-crisis period after controlling for market volatilities, the results remain qualitatively similar: for instance, in investment grade bonds, spreads are 5.2 bsp higher in the post-regulation period for invt>15min sample, 3.8 bps for all trade sample, and 2.1 bps for the same-day sample. Whether spreads are higher in the post-crisis or the post-regulation period depends on whether market controls such as VIX and bond index volatility are included. Without these market volatility controls, spreads are higher in the post-crisis period, but once they are controlled for, spreads are higher in the post-regulation period. Thus, some of the decrease in spreads after the crisis is due to a decrease in market volatility, which may hide the potential effect of regulations. The results that we obtain when volatilities are not controlled for are still consistent with rest of the results in that the differences are the least negative for the invt>15min sample and the most negative for the same-day spread sample. In sum, same-day spreads underestimate the changes in trading costs for those demanding liquidity by 3–6 bps and spread estimates using all trades underestimate the changes in trading costs by 2–3 bps. These results can help explain why despite an increase in dealer inventory costs, average bid-ask spread estimates remain low. When we restrict the sample to invt>15min sample, we find 7–11 bps increases in trading costs compared with the pre-crisis levels. Given that average bid-ask spreads during non-crisis times are around 20 bps, these increases are substantial. Moreover, since invt>15min trades do contain trades where customers are providing liquidity, these 7–11 bps increases are likely an underestimate as well. It is also notable that we are measuring this increase during non-crisis times.

6

Conclusion

We show that substantial amounts of liquidity are provided by the non-dealer sector, and this provision of liquidity by non-dealers causes the average bid-ask spreads to underestimate the true trading costs paid by liquidity-demanding customers. Decreases in dealers’ willingness or ability to provide inventories have pushed more liquidity provision to the non-dealer sector, which in turn made the bias more severe. We show that these mechanisms lead to an underestimation of the impact of regulations on liquidity, and once we reduce this bias, spreads paid by customers demanding liquidity have increased post-regulation. This increase in transaction costs may be due to bank regulations such as the Volcker Rule or more stringent

20

capital regulations. Alternatively, it may be due to changes in large dealer banks’ risk bearing capacity or risk management practices that may not directly be caused by regulations. Overall net effect of decreased dealer liquidity provision to customers may be ambiguous. Some buy-side investors who have enough liquidity and the expertise to provide liquidity may benefit from the reduced capabilities of the dealer sector in providing liquidity during the post-regulation period. For other customers that generally demand liquidity only, both trading costs and waiting time for demanding liquidity have increased in the post-regulation period. Also, it is unclear whether liquidity provision having partially moved to the non-dealer sector is healthy for the stability of financial markets. As Duffie (2012) notes, this may have potentially adverse consequences. Moreover, given that many non-dealers are likely buy-side participants subject to potential liquidity shocks from fund outflows, these shocks may have feedback effects. These potential negative consequences should be weighed against the potential positive effects of regulations. Lastly, it is important to note that we are not making any inference about whether the regulations have improved or decreased welfare. Even if regulations did cause the decrease in liquidity, to draw any conclusion about the overall welfare, we would need to weigh the liquidity deterioration against the potential positive impact that regulations have had on curbing systemic risk, which is outside the scope of this paper.

21

Figure 1: Distribution of spreads by match type The below graphs present the distribution of bid-ask spreads for customer trades larger than $100,000 by trade type. Panel A uses bid-ask spreads and is measured in basis points. Panel B uses standardized spreads, which we define as bid-ask spreads divided by the standard deviation of all traded prices for trades larger than $100,000 in that cusip-week. For standardized spreads, cusip-week pairs with less than 10 trades are excluded. We use kernel density estimation.

A. Bid−ask spread

B. Standardized spread

0.008 0.20

0.006

density

density

0.15

0.004

0.10

0.002

0.05

0.000

0.00 −400

0

400

800

−5.0

−2.5

spread

0.0

2.5

standardized spread DC−DC

DC−ID

22

invt>15min

5.0

Figure 2: Same-day spreads by trade size We plot the five-day rolling average same-day spreads by trade size. For a given cusip-day, same-day spread is measured as vwavg(customer buy)i,t − vwavg(customer sell)i,t spread2i,t = (vwavg(customer buy)i,t + vwavg(customer sell)i,t )/2 vwavg stands for value-weighted average, where volume is used for the weight. This drops trades where there is a customer buy but no customer sell trade in the particular bond-day-trade size category, or vice versa. Spreads are in basis points.

A. All

B. retail spread, in bp

spread, in bp

250 200 150 100

300 200 100

50 2006

2008

2010

2012

2014

2006

2012

C. 100K−1mil

D. above 1mil

100 50

2008

2010

date

150

2006

2008

date

spread, in bp

spread, in bp

400

2010

2012

2014

60 40 20 2006

2008

date

2010

date High yield

Investment grade

23

2014

2012

2014

Figure 3: Fractions of DC-DC and DC-ID Trades We plot the fraction of customer trades that are DC-DC and DC-ID matched trades over the sample period using monthly data. DC-DC and DC-ID matches are defined in Section 3.2.

A. Investment grade

B. High yield

0.25

0.4

Fraction of volume

Fraction of volume

0.20

0.15

0.10

0.3

0.2

0.1 0.05 2006

2008

2010

2012

2014

2006

date

2008

2010

date DC−DC

24

DC−ID

2012

2014

Figure 4: Short-holding trades by dealer size We plot the monthly fraction of customer trades that are short holding trades by dealer size. Dealer size classification is outlined in Section 3.1.

A. Investment grade

B. High yield 0.7

Fraction of volume

Fraction of volume

0.4

0.3

0.6

0.5

0.4

0.2 0.3

0.1 2006

2008

2010

2012

2014

2006

date

2008

2010

date large

25

small

2012

2014

Figure 5: Large dealers’ short holding trades We calculate the monthly fraction of customer trades that are short holding trades for each large dealer, and plot the average and median across dealers.

A. Investment grade

B. High yield

0.30

0.25

Fraction of volume

Fraction of volume

0.40

0.20

0.15

0.35

0.30

0.25 0.10 0.20 2006

2008

2010

2012

2014

2006

date

2008

2010

date average

26

median

2012

2014

Figure 6: Spreads for all and invt>15min trades We plot bid-ask spreads for two samples: all trades and invt>15min trades. Invt>15min sample excludes all short duration trades. We use trades that are 1million USD or larger only. Bid-ask spreads are calculated as traded price − reference price spread = 2Q × reference price where Q is 1 for customer buy and -1 for customer sell. We use weekly value-weighted average interdealer price as the reference price. Trades that we cannot find reference price for are dropped. Panel A and B plot the spreads for the full sample, and the last two panels plot the spreads for the post-crisis and the post-regulation periods only. Spreads are in basis points.

A. Investment grade

B. High yield

60

spread (bp)

spread (bp)

75

50

40

20

25

2006

2008

2010

2012

2014

2006

2008

2010

2012

2014

date

date

C. After crisis: Investment grade

D. After crisis: High yield 40

spread (bp)

spread (bp)

40

30

30

20

20 10 2010

2012

2014

2010

date

2012

date all trades

invt>15min

27

2014

Table 1: Example of inventory holding period construction Panel A: Sample (ficticious) trading data trade num 1 2 3 4 5 6 7 8 9

time

trade type

buy/sell

quantity

10:00:00 AM 10:00:05 AM 11:20:07 AM 11:50:00 AM 12:02:03 PM 12:30:00 PM 1:00:00 PM 1:00:03 PM 1:00:05 PM

DC DC DC DC ID DC DC DC ID

S B B B S S B S S

200 200 400 500 350 100 550 100 400

Panel B: Inventory holding period calculation trade num 1 2 3 4 4 4 6 7 7 7 8

other side

holding period

volume

short holding

volume ×short

2 1 NA 5 6 NA 4 8 9 NA 7

00:00:05 00:00:05 NA 00:12:03 00:40:00 NA 00:40:00 00:00:03 00:00:05 NA 00:00:03

200 200 400 350 100 50 100 100 400 50 100

1 1 0 1 0 0 0 1 1 0 1

200 200 0 350 0 0 0 100 400 0 100

sum

short type

overnight

DC-DC DC-DC

0 0 1 0 0 1 0 0 0 1 0

DC-ID

DC-DC DC-ID DC-DC

1350

2050

volume ×overnight 0 0 400 0 0 50 0 0 0 50 0 500

Panel C: Trade classification trade num 1 2 3 4 6 7 8

avg(short)

avg(DC-DC)

avg(DC-ID)

trade type

1 1 0 0.7 0 0.91 1

1 1 0 0 0 0.18 1

0 0 0 0.7 0 0.73 0

DC-DC DC-DC invt>15min DC-ID invt>15min DC-ID DC-DC

28

Table 2: Summary statistics Fraction of volume for each trade type category is presented in the table below. Panel A presents the results by dealer category, where large dealers are the 15 large dealers as defined in Section 3.1. Panel B presents the results by trade size. “Overnight” are the trades that stay in the dealers’ inventory overnight, and “short dur” are the short duration trades. Overnight inventory, short duration, DC-DC match, and DC-ID match are defined in Section 3.2. Volume is in par value million USD. Panel A: By large/small dealers rating IG IG HY HY

dealer size

overnight

short dur

DC-DC

DC-ID

vol

trade count

large small large small

60.99% 49.63% 46.71% 21.97%

19.76% 27.40% 32.34% 54.52%

15.82% 16.55% 28.95% 45.05%

3.94% 10.85% 3.39% 9.48%

9,344,740 2,938,289 5,994,429 1,206,477

9,162,931 6,051,187 4,443,710 2,210,241

Panel B: By trade size rating IG IG IG HY HY HY

tsize

overnight

short dur

DC-DC

DC-ID

vol

trade count

≤ 100K 100K-1mil ≥ 1mil ≤ 100K 100K-1mil ≥ 1mil

50.96% 68.32% 57.55% 46.29% 57.76% 41.51%

31.79% 18.35% 21.62% 31.26% 24.77% 36.87%

4.71% 8.95% 16.94% 5.78% 16.35% 33.05%

27.08% 9.39% 4.68% 25.48% 8.42% 3.82%

289,028 1,005,344 10,988,658 102,519 437,587 6,660,800

9,936,474 2,862,960 2,414,684 3,533,705 1,150,030 1,970,216

Table 3: Summary statistics for spread and dealer profit Average spread and dealer profit are presented by bond rating and trade type for trades 1million or above. Dealer profit is calculated only for trades that do not stay in dealers’ inventory overnight, and is calculated as dealer sell price minus buy price. Spread calculations are outlined in Section 3.3. Both spread and dealer profit are in basis points. rating IG IG IG HY HY HY

trade type

avg spread

avg profit

DC-DC DC-ID invt>15min DC-DC DC-ID invt>15min

9.08 38.5 25.88 17.95 49.02 23.26

11.92 16.06 19.15 18.73 19.08 21.19

29

Table 4: Regressions of Bid-Ask Spreads on Trade Types Following table presents the results from the regression spreadi,j,t,k = β2 1(DC − DC)k + β3 1(invt > 15min)k +

X

Ai +

X

Bj +

X

Ct + i,j,t,k

where spreadi,j,t,k is the spread for a trade in cusip i on day t between dealer j and a client. DC-ID matched trades are the omitted category. Ai is the bond fixed effect, Bj is the dealer fixed effect, and Ct is the date fixed effect. We restrict the sample to trades over $1million. Standard errors are double clustered by bond and date. IG

HY

all dealers (1)

all dealers (2)

large dealers (3)

all dealers (4)

all dealers (5)

large dealers (6)

1(DC-DC)

−21.571∗∗∗ (0.506)

−20.375∗∗∗ (0.501)

−18.602∗∗∗ (0.588)

−29.943∗∗∗ (0.867)

−28.725∗∗∗ (0.850)

−28.066∗∗∗ (1.052)

1(invt>15min)

−10.404∗∗∗ (0.503)

−10.024∗∗∗ (0.479)

−10.597∗∗∗ (0.649)

−25.309∗∗∗ (0.911)

−24.754∗∗∗ (0.891)

−25.268∗∗∗ (1.063)

log(outstanding)

−6.570∗∗∗ (0.381)

log(volume)

−0.878∗∗∗ (0.193)

1.413∗∗∗ (0.377)

1.120∗∗∗ (0.423)

rating

−0.408∗∗∗ (0.097)

0.073 (0.145)

age

0.045∗∗∗ (0.011)

0.109∗∗∗ (0.022)

log(age)

3.741∗∗∗ (0.339)

1.283∗∗ (0.625)

1.048 (0.677)

time-to-maturity

0.001 (0.004)

log(time-to-maturity)

9.590∗∗∗ (0.374)

7.607∗∗∗ (1.326)

6.467∗∗∗ (1.433)

-3.971∗∗∗ Yes Yes 1,355,386 0.006

-2.799∗∗∗ Yes Yes 1,113,068 0.005

IG volatility

−5.786∗∗∗ (0.520) −0.629∗∗∗ (0.186)

5.508∗∗∗ (0.384)

0.091 (0.211)

5.348∗∗∗ (0.438)

β2 − β3 dealer f.e. cusip & date f.e. Observations Adjusted R2

−0.547 (0.576) −0.036∗∗∗ (0.010)

9.687∗∗∗ (0.546)

9.323∗∗∗ (0.594)

5.175∗∗∗ (0.958)

161.109∗∗∗ (22.066) 136.347∗∗∗ (26.038)

HY volatility

VIX

1.620∗∗∗ (0.378)

0.900∗∗∗ (0.056) -11.167∗∗∗ Yes No 1,743,285 0.017

0.378∗∗∗ (0.100) -10.35∗∗∗ Yes Yes 1,743,285 0.023

-8.005∗∗∗ Yes Yes 1,239,969 0.018

-4.634∗∗∗ Yes No 1,355,386 0.004 ∗

Note:

30

p<0.1;

∗∗

p<0.05;

∗∗∗

p<0.01

Table 5: Fractions of Trade Types Across the Three Subsamples For three different dealer-customer trades larger than $1million samples—the full sample, same-day sample, and spread sample—we calculate the fraction of trades (in terms of trade count) that are DC-DC, DC-ID, agency, and invt>15min trades. Full sample refers to all dealer-customer trades. The same-day sample are the dealer-customer trades that goes into same-day spread calculation; in other words, a dealer-customer trade 1million or larger is included in the same-day sample if and only if there is another dealer-customer trade that is 1million or larger in the same cusip and day of the opposite direction. A dealer-customer trade that is 1million or larger is included in the spread sample if and only if there is at least one interdealer trade larger than 100K in that cusip-week. Trades are classified as DC-DC, DC-ID, or invt>15min trades, as defined in Section 3.2. IG DC-DC DC-ID invt>15min

HY

full sample

same day

spread sample

full sample

same day

spread sample

14.8% 6.2% 79.0%

29.0% 6.1% 64.9%

14.0% 8.5% 77.5%

27.7% 4.5% 67.8%

39.1% 4.4% 56.5%

24.5% 6.6% 69.0%

31

Table 6: Regressions of the Fractions of Short-Holding Period and DC-DC Matched Trades Following table presents the results from the aggregate level and dealer group level daily regressions Aggregate level : yt = α +

4 X

βl 1(t ∈ Tl ) + t

l=2

Dealer group level : ym,t = α1 1(large)m + α2 1(small)m +

4 X

1(large)m βlarge,l 1(t ∈ Tl )

l=2

+

4 X

1(small)m βsmall,l 1(t ∈ Tl ) + m,t

l=2

where yt is the fraction of short-holding trade or the fraction of DC-DC trades at the aggregate level. ym,t is the fraction of short-holding trades calculated separately for large and small dealers. Tl (l = 1, . . . , 4) are the four subperiod dummies (pre-crisis, financial crisis, post-crisis, post-regulation). Columns (1) and (4) present the aggregate level regressions for the fraction of short holding trades, and columns (2) and (5) present those for the fraction of DC-DC trades. Column (3) and (6) present the results for the dealer group level regressions. For the aggregate regressions, we use Newey-West standard errors with 20 lags, for the dealer group level regression, we cluster standard error by date.

32

IG

HY

short holding

DC-DC

group short

short holding

DC-DC

group short

(1)

(2)

(3)

(4)

(5)

(6)

crisis

∗∗∗

0.023 (0.008)

∗∗

0.014 (0.006)

∗∗

0.019 (0.008)

0.025 (0.007)

post-crisis

0.015∗∗∗ (0.005)

0.006 (0.004)

0.012∗ (0.007)

0.018∗∗∗ (0.007)

post-reg

0.044∗∗∗ (0.007)

0.043∗∗∗ (0.007)

0.073∗∗∗ (0.009)

0.081∗∗∗ (0.009)

VIX

0.001∗∗∗ (0.0003)

0.001∗∗∗ (0.0002)

0.002∗∗∗ (0.0002)

0.001∗ (0.001)

0.001∗ (0.001)

0.002∗∗∗ (0.0003)

IG volatility

0.474∗∗∗ (0.168)

0.445∗∗∗ (0.154)

0.481∗∗∗ (0.078) 0.240 (0.212)

0.235 (0.199)

0.165∗∗ (0.084)

HY volatility

∗∗∗

small

0.080∗∗∗ (0.004)

0.280∗∗∗ (0.005)

large × crisis

−0.010∗∗ (0.004)

0.003 (0.005)

small × crisis

0.042∗∗∗ (0.006)

0.018∗∗∗ (0.007)

large × post-crisis

−0.002 (0.003)

−0.006 (0.004)

small × post-crisis

0.010∗∗ (0.004)

−0.016∗∗∗ (0.006)

large × post-reg

0.046∗∗∗ (0.003)

0.089∗∗∗ (0.004)

small × post-reg

0.0002 (0.004)

−0.074∗∗∗ (0.005)

Constant

0.140∗∗∗ (0.006)

0.097∗∗∗ (0.005)

β4 − β3 βlarge,4 − βlarge,3 βsmall,4 − βsmall,3 Observations Adjusted R2

0.028∗∗∗

0.037∗∗∗

0.118∗∗∗ (0.003)

0.293∗∗∗ (0.007)

0.244∗∗∗ (0.006)

0.061∗∗∗

0.064∗∗∗

∗∗∗

2,301 0.261

2,301 0.225

0.048 -0.01∗∗∗ 4,602 0.462

2,301 0.230

2,301 0.255 ∗

Note:

33

p<0.1;

∗∗

p<0.05;

0.248∗∗∗ (0.004)

0.095∗∗∗ -0.058∗∗∗ 4,602 0.724 ∗∗∗

p<0.01

Table 7: Impact of Regulations on the Fractions of Short Holding Period Trades: Individual Dealer Level Regressions For the large dealers, we calculate the daily fraction of short duration trades for each dealer, and regress on time period dummies and dealer fixed effects. short durj,t =

X

Dj +

4 X

βl 1(t ∈ Tl )

l=2

where Tl (l = 1, . . . , 4) are the four subperiod dummies (pre-crisis, financial crisis, post-crisis, postregulation). short durj,t is the fraction of short duration for dealer j on day t, and Dj is the dealer fixed effect. The above equation is estimated via OLS and median regressions. T1 , the pre-crisis dummy, is the omitted variable. β4 − β3 indicate the difference between the post-regulation and the post-crisis coefficients. OLS standard errors are clustered by date, and median regression standard errors are based on 2,000 paired XY bootstaps. IG

HY

OLS

med

OLS

med

(1)

(2)

(3)

(4)

crisis

0.048∗∗∗ (0.003)

0.045∗∗∗ (0.003)

0.034∗∗∗ (0.004)

0.043∗∗∗ (0.004)

post-crisis

0.024∗∗∗ (0.002)

0.022∗∗∗ (0.002)

0.019∗∗∗ (0.003)

0.027∗∗∗ (0.003)

post-reg

0.075∗∗∗ (0.002)

0.043∗∗∗ (0.002)

0.069∗∗∗ (0.003)

0.057∗∗∗ (0.003)

VIX

0.001∗∗∗ (0.0001)

0.001∗∗∗ (0.0001)

0.001∗∗∗ (0.0002)

0.001∗∗∗ (0.0001)

IG volatility

0.387∗∗∗ (0.063)

0.226∗∗∗ (0.052)

HY volatility

0.298∗∗∗ (0.072) 0.068 (0.064)

Constant

0.032∗∗∗ (0.003)

0.035∗∗∗ (0.003)

0.200∗∗∗ (0.004)

0.195∗∗∗ (0.005)

dealer f.e. β4 − β3 Observations Adjusted R2

Yes 0.051∗∗∗ 28,839 0.765

Yes 0.021∗∗∗ 28,839

Yes 0.051∗∗∗ 28,499 0.614

Yes 0.029∗∗∗ 28,499

Note:

∗

34

p<0.1;

∗∗

p<0.05;

∗∗∗

p<0.01

Table 8: Impact of Regulations on Bid-Ask Spreads: invt>15min Subsamples This table presents the results for the trade-level regression spreadi,t,k = α +

4 X

Regressions for the All-Trade and

βl 1(t ∈ Tl ) + β5 (bond i char) + β6 log(volk ) + β7 mktt + i,t,k

l=2

where spreadi,t,k is the bid-ask spread for bond i, day t, trade k. Tl (l = 1, . . . , 4) are the subperiod dummies. bond i char are bond characteristics such as outstanding amount, ratings, age, and time-to-maturity. volk is the trade volume, and mktt is the market controls such as the bond index volatility and VIX. We look at two different sample; one with all dealer-customer trades and another with invt >15min trades only. β4 − β3 is the difference between the coefficient for post-regulation dummy and the coefficient for post-crisis dummy. Standard errors are double clustered by cusip and date. IG All trades

HY invt >15min

All trades

invt >15min

(1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

crisis

13.266∗∗∗ (1.266)

36.231∗∗∗ (1.514)

13.935∗∗∗ (1.483)

38.389∗∗∗ (1.716)

5.853∗∗∗ (1.646)

18.819∗∗∗ (1.602)

6.175∗∗∗ (2.142)

18.926∗∗∗ (1.996)

post-crisis

5.167∗∗∗ (0.758)

16.519∗∗∗ (0.661)

5.504∗∗∗ (0.894)

17.739∗∗∗ (0.746)

0.800 (1.272)

7.142∗∗∗ (0.879)

1.352 (1.674)

7.291∗∗∗ (1.125)

post-regulation

8.993∗∗∗ (0.598)

11.147∗∗∗ (0.577)

10.678∗∗∗ (0.685)

13.017∗∗∗ (0.666)

3.882∗∗∗ (0.842)

5.431∗∗∗ (0.809)

6.759∗∗∗ (1.087)

8.337∗∗∗ (1.052)

log(outstanding)

−8.213∗∗∗ (0.408)

−7.993∗∗∗ (0.410)

−8.823∗∗∗ (0.439)

−8.527∗∗∗ (0.442)

−6.487∗∗∗ (0.553)

−6.136∗∗∗ (0.549)

−7.082∗∗∗ (0.671)

−6.673∗∗∗ (0.668)

log(volume)

−2.429∗∗∗ (0.191)

−2.619∗∗∗ (0.196)

−1.355∗∗∗ (0.220)

−1.573∗∗∗ (0.225)

0.405 (0.358)

0.123 (0.362)

1.214∗∗ (0.481)

0.870∗ (0.486)

rating

−0.840∗∗∗ (0.103)

−0.859∗∗∗ (0.105)

−0.785∗∗∗ (0.114)

−0.823∗∗∗ (0.117)

0.069 (0.146)

0.015 (0.147)

−0.581∗∗∗ (0.179)

−0.630∗∗∗ (0.182)

age

0.064∗∗∗ (0.012)

0.062∗∗∗ (0.012)

0.061∗∗∗ (0.014)

0.060∗∗∗ (0.014)

0.128∗∗∗ (0.025)

0.123∗∗∗ (0.025)

0.134∗∗∗ (0.033)

0.129∗∗∗ (0.033)

log(age)

3.780∗∗∗ (0.363)

3.846∗∗∗ (0.364)

4.544∗∗∗ (0.405)

4.582∗∗∗ (0.408)

−0.799 (0.600)

−0.465 (0.603)

−0.843 (0.772)

−0.530 (0.776)

time-to-maturity

−0.007∗ (0.004)

−0.008∗ (0.004)

−0.004 (0.005)

−0.005 (0.005)

−0.034∗∗∗ (0.010)

−0.034∗∗∗ (0.009)

−0.037∗∗∗ (0.011)

−0.037∗∗∗ (0.011)

log(time-to-maturity)

10.392∗∗∗ (0.377)

10.367∗∗∗ (0.379)

11.348∗∗∗ (0.434)

11.311∗∗∗ (0.431)

4.715∗∗∗ (0.956)

4.354∗∗∗ (0.948)

5.075∗∗∗ (1.198)

4.740∗∗∗ (1.196)

IG volatility

125.271∗∗∗ (21.233)

152.104∗∗∗ (25.851)

HY volatility

135.377∗∗∗ (39.473)

0.416∗∗∗ (0.128)

0.336∗ (0.173)

VIX

0.868∗∗∗ (0.060)

Constant

76.245∗∗∗ (6.192)

90.163∗∗∗ (6.193)

67.412∗∗∗ (6.674)

82.190∗∗∗ (6.665)

73.778∗∗∗ (9.104)

79.774∗∗∗ (8.709)

83.198∗∗∗ (11.124)

88.135∗∗∗ (10.641)

β4 − β3 Observations Adjusted R2

3.826∗∗∗ 1,774,270 0.012

-5.372∗∗∗ 1,774,270 0.010

5.173∗∗∗ 1,376,991 0.014

-4.722∗∗∗ 1,376,991 0.011

3.082∗∗∗ 1,372,555 0.002

-1.711∗∗ 1,372,555 0.001

5.406∗∗∗ 946,257 0.002

1.046 946,257 0.001

Note:

0.913∗∗∗ (0.075)

105.817∗∗∗ (28.705)

35

∗

p<0.1;

∗∗

p<0.05;

∗∗∗

p<0.01

Table 9: Impact of Regulations on Same-Day Spreads This table presents the results for the regression spreadi,t = α +

4 X

βl 1(t ∈ Tl ) + β5 (bond i char) + β6 log(voli,t ) + β7 mktt + i,t

l=2

where spreadi,t is the same-day bid-ask spread for bond i and day t, where same-day spread calculations are defined in Section 4.2. Tl (l = 1, . . . , 4) are the four subperiod dummies. bond i char are bond characteristics such as outstanding amount, ratings, age, and time-to-maturity. volk is the trade volume, and mktt is the market controls such as the bond index volatility and VIX. β4 − β3 is the difference between the coefficient for post-regulation dummy and the coefficient for post-crisis dummy. Standard errors are double clustered by cusip and date. IG

HY

(1)

(2)

(3)

(4)

crisis

9.761∗∗∗ (0.679)

28.782∗∗∗ (0.937)

4.146∗∗∗ (0.628)

19.886∗∗∗ (0.901)

post-crisis

4.380∗∗∗ (0.405)

13.817∗∗∗ (0.404)

−1.596∗∗∗ (0.504)

6.682∗∗∗ (0.468)

post-regulation

6.431∗∗∗ (0.340)

8.198∗∗∗ (0.328)

1.059∗∗ (0.420)

2.673∗∗∗ (0.410)

log(outstanding)

−3.043∗∗∗ (0.272)

−2.921∗∗∗ (0.273)

−1.802∗∗∗ (0.281)

−1.551∗∗∗ (0.277)

log(volume)

−1.035∗∗∗ (0.097)

−1.130∗∗∗ (0.104)

−0.652∗∗∗ (0.116)

−0.821∗∗∗ (0.120)

rating

−0.228∗∗∗ (0.068)

−0.242∗∗∗ (0.069)

1.416∗∗∗ (0.089)

1.352∗∗∗ (0.089)

age

0.041∗∗∗ (0.005)

0.039∗∗∗ (0.005)

0.047∗∗∗ (0.009)

0.041∗∗∗ (0.009)

log(age)

1.154∗∗∗ (0.203)

1.251∗∗∗ (0.209)

−0.202 (0.255)

0.219 (0.269)

time-to-maturity

0.016∗∗∗ (0.002)

0.015∗∗∗ (0.003)

−0.018∗∗∗ (0.007)

−0.018∗∗∗ (0.007)

log(time-to-maturity)

6.617∗∗∗ (0.196)

6.693∗∗∗ (0.199)

4.256∗∗∗ (0.507)

3.965∗∗∗ (0.509)

IG volatility

71.103∗∗∗ (11.060) 53.727∗∗∗ (9.312)

HY volatility

VIX

0.766∗∗∗ (0.027)

Constant

14.846∗∗∗ (4.181)

25.834∗∗∗ (4.216)

0.401 (4.859)

9.323∗∗ (4.740)

2.05∗∗∗ 398,282 0.146

-5.619∗∗∗ 398,282 0.119

2.655∗∗∗ 386,972 0.082

-4.009∗∗∗ 386,972 0.051

β4 − β3 Observations Adjusted R2 Note:

0.682∗∗∗ (0.040)

∗

36

p<0.1;

∗∗

p<0.05;

∗∗∗

p<0.01

Appendix A

Overnight Inventory

In this section, we show further evidence that overnight inventory is lower in the post-regulation period compared to the pre-crisis period and the post-regulation period. Figures are constructed similar to those in Section 5.2, but using overnight inventory instead of short holding period. Figure 7 plots the overnight inventory separately for large and small dealers, and figure 8 plots the average and median overnight inventory for the 15 large dealers. Consistent with the results in Section 5.2, overnight inventory decreased post-regulation, and the change was driven by large dealers.

B

Zero profit trades

A surprisingly high fraction of short duration trades seem to yield zero profit to the dealer; that is, the dealer buys and sells the asset at the same price. While short duration trades are risk-free for the dealer, it is still surprising that the dealer does not extract any fees. Here, we explore the possibility that zero profit trades may be from intermediating for their affiliates, and test whether this may affect our results. When a dealer trades with its affiliate, it is more likely that the dealer does not demand a profit and only passes along whatever the cost of the trade is. Starting late 2015, dealers are required to note when their counterparty is a non-FINRA affiliate. This shows up as counterparty being ‘A’ (affiliate) instead of ‘C’ (customer). Before this change, affiliates that are not registered with FINRA is reported as ‘C’ and cannot be distinguished from non-affiliate customers. We first use 2016 data to test whether zero profit trades are generally due to dealers trading with their affiliates. We use the TRACE data from January 1 to October 15, 2016. We clean the data as before, and construct inventory holding periods and dealer profits. For trades with inventory life less than a day, we can see which trades it was matched with, and what the profit for the dealer was. For a given trade, if more than 50% of the trade had zero profit for the dealer, we classify that trade as having had zero profit. If the trade was with an affiliate, or if more than 50% of the trade volume was matched with a trade that was with an affiliate, we classify that trade as having 1(A) = 1 (0 otherwise). Since we match inventory coming in with inventory going out for all trades that do not stay in the inventory overnight, we can find 1(A) and whether it is a zero profit trade for all trades that do not stay in dealers’ inventory overnight.

37

Table 10 summarizes the average 1(A) by whether the trade is a short duration trade and whether it is a zero profit trade. A few results stand out. First, for short duration trades, almost all zero profit trades, both by trade count and volume, are either trades with affiliates or are matched with trades with affiliates. In contrast, only about 4% of short duration trades that are not zero profit trades have 1(A) = 1. Second, amongst non-short-holding trades, zero profit trades do have a higher fraction of trades with 1(A) = 1, but the differences between zero profit trades and others are much less stark. Since calculating dealer profit for trades that stay in the inventory longer is inherently noisier, it is impossible to tell whether this is because dealer profit is imperfectly measured or because invt>15min trades with affiliates actually often yield nonzero profit for the dealer. Lastly, amongst the short duration trades, about 35% of trades by trade count and about 50% of trades by volume yield zero profit for the dealers, and are trades with affiliates or are matched with trades with affiliates. This points to the need to be cautious when interpreting the short duration trades, or ‘agency’ or ‘riskless principal’ trades. Since almost half of short holding trades are with affiliates or are matched with trades with affiliates, it is important to test whether this affects our results. For instance, the increase in short-holding trades may be due to an increase in dealers trading with its affiliates, or due to higher fraction of trading with affiliates being matched. While these are not necessarily inconsistent with our hypothesis, interpretations may be different as affiliates may be different from the customers that we usually think about in this market. The ideal robustness check would be to rerun the tests after deleting all trades with

1(A) = 1. Since

we do not observe whether a trade is with an affiliate or not prior to 2015, we instead delete all zeroprofit short-holding trades. We do not delete invt>15min trades with zero profit, as majority of those have

1(A) = 0. Figure 9, 10, and 11 plot the fraction of short-holding trades on aggregate, by dealer size, and for the large dealers, respectively. As before, fraction of short duration trades is high during the crisis period, and increases again starting around 2011, and the levels are higher in the post-regulation period compared to the pre-crisis or the post-crisis periods. However, the actual shape is somewhat different from Figure 3, 4, and 5; in particular, the increases in 2014 or later are less pronounced. Interestingly, compared to Figure 5, the mean and medain values are much closer and move together more closely in Figure 11. This implies that some individual dealers that saw a much larger increase in short-holding trades may have been driven by an increase in trading with affiliates or an increase in matching trades with affiliates. Table 11 test whether short-holding trades increase post-regulation in a regression setting. Consistent with earlier results in Table 6, short-holding trades increase post-regulation, althrough the increase is slightly

38

smaller. Table 12 looks at how spread and profit are different for various trade types. Results are consistent with those in Table 4.

39

Figure 7: Overnight inventory by dealer size

A. Investment grade

B. High yield 0.6

0.7 0.5

Fraction of volume

Fraction of volume

0.6

0.5

0.4

0.4

0.3

0.2

0.3

0.1

2006

2008

2010

2012

2014

2006

2008

date

2010

2012

2014

date large

small

Figure 8: Large dealers’ overnight inventory

A. Investment grade

B. High yield 0.60

Fraction of volume

Fraction of volume

0.7

0.6

0.5

0.55

0.50

0.45

0.40

0.35 2006

2008

2010

2012

2014

2006

date

2008

2010

date average

40

median

2012

2014

Figure 9: Excluding affiliate trades—Short holding trades over time

A. Investment grade

B. High yield

0.20

Fraction of volume

Fraction of volume

0.3

0.15

0.10

0.2

0.1 0.05

2006

2008

2010

2012

2014

2006

2008

date

2010

2012

2014

date DC−DC

DC−ID

Figure 10: Excluding affiliate trades—Short holding trades by dealer size

A. Investment grade

B. High yield

0.4

Fraction of volume

Fraction of volume

0.6 0.3

0.2

0.4

0.1 0.2 2006

2008

2010

2012

2014

2006

date

2008

2010

date large

41

small

2012

2014

Figure 11: Excluing affiliate trades—Large dealers’ short holding trades

A. Investment grade

B. High yield 0.35

0.25

0.30

Fraction of volume

Fraction of volume

0.20

0.15

0.25

0.10 0.20

0.05 2006

2008

2010

2012

2014

2006

date

2008

2010

date average

42

median

2012

2014

Table 10: Zero profit trades and trades with affiliates The following table provides average and volume-weighted average (1(A)) by trade type and whether the trade is a zero profit trade for the dealer. We include all trades that are with customers or affiliates. Data period is Jan 1 to Oct 15, 2016. trade type short dur short dur invt>15min invt>15min

zero profit

avg(1(A))

vwavg(1(A))

trade count

volume (in million USD)

No Yes No Yes

4.06% 91.86% 10.51% 22.74%

4.44% 98.20% 9.08% 28.01%

562,767 303,068 1,777,026 6,645

459,180 459,305 1,676,295 11,358

43

Table 11: Excluding affiliate trades—Change in short-holding trades Following table presents the results from the aggregate level and dealer group level daily regressions Aggregate level : yt = α +

4 X

βl 1(t ∈ Tl ) + t

l=2

Dealer group level : ym,t = α1 1(large)m + α2 1(small)m +

4 X

1(large)m βlarge,l 1(t ∈ Tl )

l=2

+

4 X

1(small)m βsmall,l 1(t ∈ Tl ) + m,t

l=2

where Tl (l = 1, . . . , 4) are the subperiod dummies (pre-crisis, financial crisis, post-crisis, post-regulation). All dependent variables are calculated after deleting all trades with 1(A) = 1. For the aggregate level regressions, we look at two different dependent variables, fraction of short-holding trades and fraction of DC-DC matched trades. Both are calculated as a fraction of total dealer-customer trading volume. For the dealer group level regression, we calculate the fraction of dealer-customer trading volume that are shortholding trades separately for large and small dealers. Column (3) and (6) presents the dealer group level regressions. For the aggregate regressions, we use Newey-West standard errors with 20 lags, for the dealer group level regression, we cluster standard error by date.

44

IG

HY

short holding

DC-DC

group short

short holding

DC-DC

group short

(1)

(2)

(3)

(4)

(5)

(6)

crisis

∗∗∗

0.039 (0.007)

∗∗∗

0.028 (0.005)

∗∗∗

0.028 (0.007)

0.028 (0.007)

post-crisis

0.015∗∗∗ (0.005)

0.002 (0.004)

0.030∗∗∗ (0.007)

0.028∗∗∗ (0.006)

post-reg

0.035∗∗∗ (0.004)

0.028∗∗∗ (0.003)

0.061∗∗∗ (0.004)

0.063∗∗∗ (0.004)

VIX

0.002∗∗∗ (0.0003)

0.002∗∗∗ (0.0002)

0.003∗∗∗ (0.0002)

0.002∗∗∗ (0.0005)

0.002∗∗∗ (0.0005)

0.002∗∗∗ (0.0003)

IG volatility

0.322∗∗ (0.132)

0.257∗∗ (0.121)

0.398∗∗∗ (0.072) 0.058 (0.129)

0.054 (0.115)

0.046 (0.079)

HY volatility

∗∗∗

small

0.138∗∗∗ (0.003)

0.332∗∗∗ (0.004)

large × crisis

0.007∗∗ (0.004)

0.018∗∗∗ (0.004)

small × crisis

0.047∗∗∗ (0.006)

0.013∗ (0.007)

large × post-crisis

−0.008∗∗∗ (0.002)

0.020∗∗∗ (0.003)

small × post-crisis

0.008∗ (0.004)

−0.023∗∗∗ (0.006)

large × post-reg

0.025∗∗∗ (0.002)

0.073∗∗∗ (0.003)

small × post-reg

−0.001 (0.004)

−0.079∗∗∗ (0.005)

Constant

0.062∗∗∗ (0.005)

0.037∗∗∗ (0.004)

0.030∗∗∗ (0.003)

0.233∗∗∗ (0.006)

0.199∗∗∗ (0.006)

0.182∗∗∗ (0.003)

dealer f.e. β4 − β3 large ×(β4 − β3 ) small ×(β4 − β3 ) Observations Adjusted R2

No 0.02∗∗∗

No 0.026∗∗∗

Yes

No 0.031∗∗∗

No 0.035∗∗∗

Yes

2,301 0.492

2,301 0.402

0.033∗∗∗ -0.009∗∗∗ 4,602 0.725

2,301 0.273

2,301 0.262 ∗

Note:

45

p<0.1;

∗∗

p<0.05;

0.053∗∗∗ -0.056∗∗∗ 4,602 0.798 ∗∗∗

p<0.01

Table 12: Excluding affiliate trades—Bid-ask spreads and dealer profit Following regressions explore how bid-ask spreads and dealer profit vary by trade type, after deleting all trades with 1(A) = 1. We run X X X spreadi,j,t,k = β2 1(DC − DC)k + β3 1(invt > 15min)k + Ai + Bj + Ct + i,j,t,k X X X prof iti,j,t,k = β2 1(DC − DC)k + Ai + Bj + Ct + i,j,t,k where spreadi,j,t,k and prof iti,j,t,k are the bid-ask spread and the dealer profit, respectively, for a trade in cusip i on day t between dealer j and a client. Ai is the bond fixed effect, Bj is the dealer fixed effect, and Ct is the date fixed effect. We restrict the sample to trades over $1million and exclude trades with 1(A) = 1. DC-ID matched trades are the omitted category. For the profit regression, only short-holding trades are included. For the spread regression, β2 − β3 is the difference in coefficients for 1(DC-DC) and 1(invt>15min). Standard errors are double clustered by date and bond. spread IG

profit HY

(1)

IG

(2)

HY

(3)

(4)

1(DC-DC)

−29.195 (0.744)

−36.579 (0.988)

1(invt>15min)

−20.640∗∗∗ (0.543)

−35.030∗∗∗ (0.977)

log(volume)

−0.857∗∗∗ (0.202)

1.236∗∗∗ (0.403)

−2.706∗∗∗ (0.115)

−2.152∗∗∗ (0.086)

log(age)

5.977∗∗∗ (0.422)

1.551∗∗ (0.661)

2.062∗∗∗ (0.239)

0.736∗∗∗ (0.225)

log(time-to-maturity)

11.132∗∗∗ (0.619)

8.315∗∗∗ (1.413)

6.072∗∗∗ (0.405)

4.540∗∗∗ (0.739)

β2 − β3 dealer f.e. cusip f.e. date f.e. Observations Adjusted R2

-8.555∗∗∗ Yes Yes Yes 1,571,028 0.025

-1.549∗∗∗ Yes Yes Yes 1,252,653 0.007

Yes Yes Yes 219,240 0.321

Yes Yes Yes 317,860 0.179

∗∗∗

∗∗∗

∗

Note:

46

∗∗∗

−0.877 (0.210)

−0.858∗∗∗ (0.164)

p<0.1;

∗∗

p<0.05;

∗∗∗

p<0.01

References Adrian, T., Boyarchenko, N. and Shachar, O. (2017), ‘Dealer balance sheets and bond liquidity provision’, Journal of Monetary Economics . Adrian, T., Fleming, M. J., Shachar, O. and Vogt, E. (2016), ‘Market liquidity after the financial crisis’. Adrian, T., Fleming, M., Shachar, O. and Vogt, E. (2015), ‘Has U.S corporate bond market

liquidity

deteriorated?’,

http://libertystreeteconomics.newyorkfed.org/2015/10/

has-us-corporate-bond-market-liquidity-deteriorated.html. Anderson, M. and Stulz, R. M. (2017), ‘Is post-crisis bond liquidity lower?’. Bao, J., O’Hara, M. and Zhou, X. A. (2016), ‘The volcker rule and market-making in times of stress’, Available at SSRN . Bessembinder, H., Jacobsen, S. E., Maxwell, W. F. and Venkataraman, K. (2016), ‘Capital commitment and illiquidity in corporate bonds’, Available at SSRN 2752610 . Bessembinder, H., Maxwell, W. and Venkataraman, K. (2006), ‘Market transparency, liquidity externalities, and institutional trading costs in corporate bonds’, Journal of Financial Economics 82(2), 251–288. BlackRock (2015), ‘Addressing market liquidity’, BlackRock Viewpoint July. Brunnermeier, M. K. and Pedersen, L. H. (2009), ‘Market liquidity and funding liquidity’, Review of Financial studies 22(6), 2201–2238. Chakravarty, S. and Sarkar, A. (2003), ‘Trading costs in three us bond markets’, The Journal of Fixed Income 13(1), 39–48. Cimon, D. A. and Garriott, C. (2016), ‘Banking regulation and market making’, Working Paper . Dick-Nielsen, J., Feldh¨ utter, P. and Lando, D. (2012), ‘Corporate bond liquidity before and after the onset of the subprime crisis’, Journal of Financial Economics 103(3), 471–492. Dick-Nielsen, J. and Rossi, M. (2016), ‘The cost of immediacy for corporate bonds’. Duffie, D. (2012), ‘Market making under the proposed volcker rule’, Rock Center for Corporate Governance at Stanford University Working Paper (106).

47

Ederington, L. H., Guan, W. and Yadav, P. K. (2014), ‘Dealer spreads in the corporate bond market: Agent vs. market-making roles’. Edwards, A. K., Harris, L. E. and Piwowar, M. S. (2007), ‘Corporate bond market transaction costs and transparency’, The Journal of Finance 62(3), 1421–1451. Feldh¨ utter, P. (2012), ‘The same bond at different prices: identifying search frictions and selling pressures’, Review of Financial Studies 25(4), 1155–1206. Feroli, M., Kashyap, A. K., Schoenholtz, K. L. and Shin, H. S. (2014), ‘Market tantrums and monetary policy’, Paper for the U.S. Monetary Policy Forum, New York . Goldstein, M. A. and Hotchkiss, E. S. (2017), ‘Providing liquidity in an illiquid market: Dealer behavior in us corporate bonds’. Goldstein, M. A., Hotchkiss, E. S. and Sirri, E. R. (2007), ‘Transparency and liquidity: A controlled experiment on corporate bonds’, Review of Financial Studies 20(2), 235–273. Goldstein, M. and Hotchkiss, E. (2012), ‘Dealer behavior and the trading of newly issued corporate bonds’, Working Paper . Harris, L. (2015), ‘Transaction costs, trade throughs, and riskless principal trading in corporate bond markets’, Trade Throughs, and Riskless Principal Trading in Corporate Bond Markets . Hong, G. and Warga, A. (2000), ‘An empirical study of bond market transactions’, Financial Analysts Journal 56(2), 32–46. Nagel, S. (2012), ‘Evaporating liquidity’, Review of Financial Studies 25(7), 2005–2039. Randall, O. (2015), ‘How do inventory costs affect dealer behavior in the US corporate bond market?’, Available at SSRN 2590331 . Schestag, R., Schuster, P. and Uhrig-Homburg, M. (2016), ‘Measuring liquidity in bond markets’, Review of Financial Studies p. hhv132. Trebbi, F. and Xiao, K. (2015), ‘Regulation and market liquidity’. Uslu, S. (2016), ‘Pricing and liquidity in decentralized asset markets’, Working Paper . Zitzewitz, E. (2010), ‘Paired corporate bond trades’, Available at SSRN 1648994 . 48