Who Makes Markets?Do Dealers Provide or Take Liquidity?Joon Chae, Albert Wang*Sloan School of Management, MIT, Doctoral Program in Financial Economics,Cambridge, MA 02142, USAFirst Draft: 1.20.2003Current Draft: 8.24.2003ABSTRACTWe explore the role of dealers to determine whether they are liquidity-providingmarket makers or liquidity-taking information traders. Standard models of marketmaker trading imply a negative contemporaneous correlation between marketmaker order flow and stock returns. We test this relation with a unique datasetcontaining trades of all dealers in a well-developed, liquid market. The correlationis strongly positive, implying that dealers take liquidity. Furthermore, dealers earnsignificant excess returns, in aggregate driven by information returns rather thanmarket-making returns. Subgroup analysis reveals that dealer profits are driven byinformation in large-cap stocks and by market-making in small-cap stocks.JEL Classification: G14Keywords: Dealer, Information, Liquidity, Market makerWe gratefully acknowledge Charles Chang, Jonathan Lewellen, Andrew Lo, Robert Merton,Dimitri Vayanos, other MIT faculty and doctoral students, and participants at the MIT FinanceStudent Seminar. All errors are our own.*Corresponding author. Tel.: 1-617-253-3637; fax: 1-617-258-6855.E-mail address: [email protected] (A. Wang)

1. IntroductionDealers in financial markets are typically assumed to provide liquidity, and therefore they areoften afforded special trading privileges related to order flow and trade execution. Such privilegesinclude access to order flow and order flow information, direct connections to exchange tradingmechanisms, low transaction costs, and high transaction speeds. In return, they are often assumedto perform the social and market function of supplying liquidity, for example by absorbingtemporary order imbalances.If dealers actually trade as liquidity-providing market makers, then there will be a negativecontemporaneous correlation between their order flow and stock returns. This follows from bothinformation and inventory models of market maker trades, as typified in Kyle (1985) and Grossmanand Miller (1988). Kyle shows that market makers transact against net (informed plus uninformed)trade demand, with a price impact due to the potential information content embedded in net demand.Grossman and Miller show that in the absence of informational issues, market makers are willing toaccommodate temporary order imbalances if they can transact at advantageous prices. In bothmodels, when other participants buy (sell), they push the price up (down). Market makers tradingto accommodate the order imbalance must sell (buy). Thus, market maker order flow will benegative (positive) when stock returns are positive (negative), implying the negative relation.However, it is questionable whether dealers actually supply liquidity as described in eithermodel. While dealers may be meant to perform the socially beneficial function of liquidityprovision, the institutional advantages granted to them also give the ability to act as super-efficientproprietary traders if they choose to. Dealer activities such as focusing on order flow informationmay enable them to deduce pricing information and trade accordingly.1 Low transaction costs and1Chae (2003) finds that dealers increase their price-sensitivity prior to information-revealing events,implying that they can deduce information about the events.2

high transaction speeds may allow them to take advantage of opportunities that are not worthwhileto other market participants. Assuming that dealers want to maximize profits, their privileges mayvery well induce information as the primary motive for trade, rather than liquidity provision.In practice, dealer trading is based on a complex set of interactions with other marketparticipants in a variety of advantageous institutional setups. For example, New York StockExchange (NYSE) specialists are dealers that trade exclusively in a single stock. They aremandated to promote “stable and orderly” markets within an open-outcry system. In return, each isthe central access point for almost all market participants who trade in their stock. Like dealers inother markets, NYSE specialists have much discretion over which trades to participate in (i.e. theydo not blindly fill every single order immediately) and are granted special privileges, particularly intheir access to order flow. Due to a lack of detailed specialist trade data, we have little idea whetherthey increase liquidity provision, and we know few details about the empirical relations betweentheir order flow and stock prices.Contrary to theory and intuition, our main finding is that the contemporaneous correlationbetween weekly dealer order flow and stock returns is strongly positive. This implies that dealersdo not provide liquidity on a weekly frequency. Furthermore, using detailed intraweek transactionprice and quantity data, we find that dealers earn significant excess returns. These excess returnsare driven by information profits, rather than by market-making profits.2 This information-drivenprofitability reinforces the main result by showing that dealers do not provide liquidity within theweek. It also highlights the magnitude of the costs of allowing dealers institutional tradingadvantages. All the results strongly suggest that dealers are informed traders, since only informedtraders should have positive price effects and such high excess returns driven by information.2Definitions of information profits and market-making profits are detailed in Section II.3

These results have a few major policy and research implications. First, the common perceptionthat dealers trade primarily to provide liquidity should be closely re-examined. Second, themandate to provide liquidity (as NYSE, NASDAQ, and other dealers have) has a large shadow cost.For example, the NASDAQ requirement that dealers must always maintain two-sided quotes atreasonable depths is a costly restriction. Dealers with such constraints to provide liquidity may tryto strategically minimize this cost; in other words, dealers may provide liquidity only to the extentthat regulatory agencies require. Accordingly, future research and institutional policy aboutdealers should consider whether the advantages given impart the incentive to provide liquidity, andwhether the cost of inducing this social function is worthwhile.In most theory models of dealer trades, dealer roles and profits are analyzed assuming that theytrade primarily as market makers. Given this basic premise, these models show that dealers takeinto account asymmetric information and hold order imbalances as their own inventory forpotentially extended periods. In return, they are compensated with an amount related to half of thebid-ask spread3 for each trade. The short list of empirical research about dealer trades also takes asgiven that dealers are market makers and analyze the data as such. These studies, including Ho andMacris (1984), Madhavan and Smidt (1993), and a few others, typically focus on high-frequencydatasets and phenomena. They often have data for very few dealers (sometimes one) and very shorttime periods (sometimes a few weeks). Accordingly, these studies focus on short-horizon issues,such as determining components of the bid-ask spread, analyzing price-discreteness effects, anddisentangling high-frequency information and inventory effects. However, there are potentialproblems arising from the use of such specific data. Idiosyncrasies in inventory management3Glosten and Milgrom (1985) shows that bid-ask spreads may be caused by the information disadvantage ofdealers.4

strategies and information processing may dominate results when studying individual dealers, andtrading patterns during very short time periods may not accurately reflect typical patterns.To the best of our knowledge, ours is the first study to address the question of whetherinstitutional advantages granted to dealers give them the incentive to provide or take liquidity. Weuse a unique and comprehensive dataset of weekly dealer trades, transaction prices, and inventory,over a five-year horizon. We aggregate trades across all dealers in the market to lessen effects ofindividual dealer idiosyncrasies, and we use over 5 years of data to mitigate period-specificrelations. In addition, our use of weekly data rather than higher-frequency data helps to mitigatehigh-frequency microstructure effects in prices, such as bid-ask bounce.Section 2 describes the dataset and market in detail. Section 3 defines the hypotheses andcorresponding empirical tests. Section 4 documents the test results. Section 5 concludes with abrief summary, institutional implications, and directions for future research.2. Data and Markets2.1. DatabaseWe use the Taiwan Economic Journal (TEJ) database of equities traded on the Taiwan StockExchange (TSE) from January 1997 to January 2002. In particular, we use weekly price and dealertrade data. This is a comprehensive dataset of all individual dealer trades, including inventorylevels, gross buys (and sells), and average gross buy (and sell) prices. This unique data allow us toexplore dealer trading and profits in great detail. To better understand this data, we first list someTSE summary statistics in Table 1 and then describe the institutional setup of the TSE.44All information regarding the TSE, SFC, and financial system in Taiwan is sourced directly from the TSECwebsite at home.htm, the TSEC Fact Book (2002), and the TSEC AnnualReport (2002).5

INSERT TABLE 1Table 1 illustrates the clear pattern that large-cap stock returns were higher. The averagemarket cap of the largest quartile is roughly (in New Taiwan Dollars) NT 70 billion, or roughlyUS 2.1 billion. The average market cap of the smallest quartile is about 33 times smaller. Thisillustrates the magnitude and variation across equity capitalizations in this market. We also see thepattern that smaller stocks had larger autocorrelations at almost all lags. Given the time period of261 weeks (implying a standard error of autocorrelation estimates of roughly 0.062), there areseveral statistically significant values for lag 1 through lag 3 autocorrelations.Total Market Capitalization of stocks listed on the TSE in 2001 was NT 10.25 trillion (roughlyUS 316 billion) in 2001 for 614 listed stocks. Annual market volume was NT 18.35 trillion, sodollar turnover in 2001 was 179% (compared to roughly 100% dollar turnover on the NYSE in2001). While market capitalization and dollar volume obviously track market prices, share volumehas remained relatively stable around 600 billion shares per year since 1996. Taiwan’s equitymarket includes a wide range of participants, including local and international investmentcompanies, banks, and individuals. The TSE is a large, well-regulated, highly liquid market inwhich many traders participate. Therefore, it is not as susceptible to price manipulation ordominant informed trader effects as other emerging markets are.2.2. TSE Background and Institutional SetupThe Taiwan Stock Exchange Corporation (TSEC) was established in 1961 as a privateinstitution overseen by the government. The TSEC has operated the TSE, the sole centralized stockexchange for listed securities in Taiwan, since its founding. In 1985, the original open outcrytrading system was replaced by a computer-assisted limit order system; and finally the Fully6

Automated Securities Trading (FAST) system was implemented in 1993. FAST is a pure limitorder system with similar price/time priorities and trading rules as other limit order markets, suchas the Paris Bourse and Toronto Stock Exchange. Trades are processed through a series of callauctions executed every 30 seconds. The opening call auction is similar to that on the NYSE, withthe opening price determined chosen to maximize trading volume on the opening trade. There is noprice limit for the opening call auction, but over our data sample there was a 2-tick price changelimit on subsequent call auctions5 and a 7% limit on daily price fluctuations. It is worth notingagain that using weekly data mitigates many of the high-frequency microstructure issues associatedwith particular institutional setups (bid-ask bounce, discrete prices, etc.). Therefore, we do notexplicitly consider these issues in our analysis.Both listed (TSE stocks) and over-the-counter (OTC stocks) stocks are traded on the TSECplatform under the same trading rules.TSE stocks meet more stringent stability and sizerequirements, and the value-weighted performance of TSE stocks determines the TAIEX index.TSE stock trading is restricted to occur only on the TSEC platform, while OTC stocks may betraded off the system at prices negotiated between parties (the 7% daily price change limit stillholds, but 2-tick rule does not). However, in practice, most OTC stock trades take place on theTSEC platform. Our data do include all trades executed on the TSEC system (all TSE stock tradesand most OTC stock trades). Trade data are collected and recorded by the Securities and FuturesCommission (SFC) and reported to the TEJ, insuring completeness and reliability.2.3. TSEC Dealers5Tick sizes and stock prices are in NT , in format: Tick(stock price bounds) tick value. Tick(S 5) .01;Tick(5 S 15) .05; Tick(15 S 50) .1; Tick(50 S 150) .5; Tick(150 S 1000) 1; Tick(S 1000) 5.7

Only two types of institutions may submit trades directly to the TSEC trade execution system:TSEC Brokers and TSEC Dealers. All other individual and institutional trades must be submittedthrough TSEC Brokers. Brokers have access to the TSEC system purely to facilitate customertrades in exchange for commissions. They are not allowed to trade on their own accounts, and theirtrade data is not publicly available.TSEC Dealers are institutions that trade exclusively on their own accounts. The minimumcapital required to be a Dealer is NT 400 million (approximately US 12 million), and NT 10million (approximately US 350 thousand) must be left in an interest-bearing account as a securitydeposit. Dealer access to the TSEC is for proprietary trading purposes only, and their namesakeportrays the SFC’s desired role for them as liquidity-providing market makers. However, theyhave no explicit mandate to provide liquidity or price stability, (as NYSE Specialists andNASDAQ Dealers have) so they are almost unrestricted in their trades. Their only trade restrictionis that they cannot short sell securities. Since they are afforded access to the TSEC system and areexplicitly forbidden to trade on insider information, their trade data are readily available. See Table2 for summary statistics about TSEC Dealer trades and Figure 1 for Dealer trading dollar volumepercentiles.INSERT TABLE 2 and FIGURE 1As shown in Table 2, the number of dealers during our sample period ranged from 49 to 72.There is a noticeable 1-week autocorrelation in aggregate dealer net turnover6 , which decaysrapidly. Average weekly net turnover was -0.009%, implying that dealers generally sold a little6Turnover carries the implication of a standardized measure of unsigned trading volume. However, we use“net turnover” to indicate standardized dealer order flow, defined as [shares bought – shares sold] / [sharesoutstanding]. Similarly, “gross turnover” is defined as [shares bought shares sold] / [shares outstanding].8

over the sample period, while average weekly gross turnover was 0.226%. Average weekly netdollar volume was –NT 440 thousand, and average weekly gross dollar volume was NT 41.6million. TSEC Dealers accounted for roughly 2% of total share trading volume 7 . Figure 1illustrates the cross-sectional difference in dealer trading activity, plotting dollar volume at the 10%,50%, and 90% levels.TSEC Brokers and Dealers had strictly separated roles during our sample period, meaning aninstitution could only perform one of these functions.8 Therefore, we do not consider moral hazardor other effects of potential front-running by brokers.2.4. Dealer Transaction Speed and Cost AdvantagesOnly TSEC Brokers and Dealers have direct connections to the TSEC computer tradeexecution system, and they can enter trades as fast as they can key them in. They also receivedetailed transaction reports instantly upon trade execution. All other traders have to trade throughBrokers as an intermediary, and for most of the sample period internet trading was not widelyavailable. Hence, the actions required for a typical investor’s trade consisted of making a phonecall and describing the trade to a broking agent, the broking agent transmitting the trade to theBroker’s order-entry person, and the order entry person keying in the order. Confirmation of thetrade occurred after the TSEC trade sheet was sent from the trading room to the broking agent andthe broking agent had time to call back the customer. For many customers, trade confirmation didnot occur until the customer received the trade sheet in the mail. Clearly, TSEC Dealers had a largeadvantage in trade execution and confirmation speed before internet transactions, and even nowthey still enjoy a significant advantage over internet traders who must interact with brokers.7From 1997 to 2001, TSEC dealer trades accounted for only 1.37% to 1.94% of total dollar trading volume.Since 2002, there have been regulatory moves to relax this separation rule. However, searching by name,we did not find any broker-dealers at the end of 2002.89

Capital gains taxes have been exempted in Taiwan since 1990. Instead, stock sellers are levieda tax of 0.3% by the TSEC. Brokers can set their own commission rates up to a ceiling of 0.1425%of the value traded, and most set commissions very close to this rate. Since Dealers do not have totrade through brokers, they avoid this brokerage cost. Therefore, TSEC Dealers had not only anadvantage in trade execution speed, but they also had a discount of about 28.5 basis points onround-trip trading costs relative to other market participants.2.5. Order Flow InformationQuote from TSEC website in October 2002, “ the current order book is a black box where nounexecuted volume is disclosed (to the public). Starting July 1st of this year, the volume ofunexecuted orders at best bid and ask prices will be disclosed so that market participants can makean informed judgment when placing orders. Beginning 2003, the volume of unexecuted orders ofthe 5 best bid and ask prices will be disclosed as well.” Therefore, only very recently has any orderbook information been available to the public. The only order flow information available over oursample period was quotes of execution prices and aggregate daily volume. Other “ticker-style”order flow information was also available through fee-based terminals. It may be an interestingevent study to explore the structural changes in the market caused by the recent changes in markettransparency, but that is not within the scope of this paper.2.6. Similarities to Other MarketsTaiwan is just one of many countries whose major stock exchanges are pure limit order markets.Other examples include Canada, France, Germany, Korea, and others. Even the NYSE andNASDAQ have significant portions of their market that work as limit order aggregationmechanisms via ECNs. Though there may not always be explicitly called “dealers” in othermarkets, it is a common assumption that there are agents in every market looking to profit by10

accommodating order imbalances. These agents typically have no inside information and noexogenous need for immediate trade, regardless of the particular institutional setup or location. Weconjecture that the basic results of this paper should extend to such agents in other pure limit ordermarkets and “limit order market segments” (such as ECNs in US markets) of dealer/specialistmarkets.The basic results are also likely to extend to specialists and dealers in non-limit order markets,as long as the specialists or dealers have discretion over which trades to participate in and aregranted institutional trading advantages. The basic premise is that if specialists and dealers havediscretion over which trades to participate in, then they implicitly also have discretion whether tomake markets or to trade on information. We hope to confirm our results in other markets given theeventual availability of reliable data in other markets.3. Hypotheses3.1. Hypothesis 1: Dealers Trade as Liquidity-Providing Market MakersMost models of dealer trades imply that when dealers act as market makers to provide liquidity,there will be a negative relation between stock returns and aggregate dealer trades. Our primaryhypothesis is based on this implied relation. The intuition is the following: as overall demand fromall informed and uninformed traders increases (decreases), dealers providing liquidity istantamount to dealers selling to (buying from) the rest of the market. Insofar as other traders haveinformational or mechanical price impact, the stock price will increase (decrease) while dealers areselling (buying), creating a negative contemporaneous relation between the returns and dealer orderflow.Consider two of the seminal models of dealer trading, Kyle (1985) and Grossman and Miller(1988). Kyle explores the inference problem and trading demands of uninformed market makers11

and informed traders, with noise traders essentially adding uncertainty. In his model, informedtraders submit trades x in the direction of their information based on their trading aggressiveness,uninformed submit trades u for exogenous reasons, and market makers trade against the netdemand (if net demand is x u then market makers trade –x – u) with a price impact determined byinformed plus uninformed trader demand. This price impact will be in the direction of the netdemand x u, defined by the optimal (positive) market depth λ provided by the market maker. Thisprice impact exists regardless whether the net demand in a given period is driven by informed oruninformed traders. As long as the ex-ante price is fair, the contemporaneous return is negative(positive) when a Kyle market maker buys (sells) shares; i.e. market maker order flow and securityreturns are negatively correlated.Grossman and Miller (1988) consider dealer trading from an inventory risk perspective.Market makers are willing to provide liquidity when there is a net trade imbalance because they cantransact at a superior price. The greater the imbalance, the better the price they can transact at. Inreturn for holding a suboptimal inventory for a potentially extended period of time, they arerewarded with a premium that will be realized whenever the net trade imbalance returns to zero.Essentially, the model predicts that liquidity-providing market makers buy at lower than fair pricesand sell at higher than fair prices. For example, assume the ex-ante price is fair and no informationis revealed. Grossman and Miller market makers will only buy (sell) at a price below (above) fairvalue, so the price decreases (increases) when they buy (sell). Eventually, when the orderimbalance disappears and they sell (buy), they do so at the higher (lower), fair price. Thus, Kyle(1985) and Grossman and Miller (1988) show that both asymmetric information and inventorymodels of dealer trades imply the same negative contemporaneous relation between dealer tradesand security returns.Our primary test uses a slightly modified vector auto-regression (VAR) of dealer order flowand stock returns to isolate this contemporaneous relation and to give insights about other12

predictive relations. In a typical VAR, only lagged variables are included as independent variables,but we include the contemporaneous dealer order flow in the return regression (and vice versa)since this is precisely the relation we are interested in. By using the VAR, we can measure thecontemporaneous correlation while controlling for momentum or contrarian effects and potentialprice impact of dealer trades.9 Our basic VAR specification is shown in Equations (1) and (2),where r is stock return and x is dealer order flow. i 1j 0 k 0l 1⑴rt α Ai rt i B j x t j ε t⑵xt α Ck rt k Dl xt l ε tWe estimate this VAR with both raw returns and index-adjusted returns for robustness. Dealerorder flow is defined as aggregate dealer net turnover, or [net shares bought by dealers] / [sharesoutstanding]. Lo and Wang (2000) describe how this standardized measure of dealer order flowcontrols for shares outstanding and provides for cleaner interpretation of empirical results relativeto share or dollar order flow. Since individual dealers have unique considerations such as inventorymanagement and investment strategies, we aggregate order flow across dealers in each period toreduce the effects of dealer idiosyncrasies.As a practical matter for empirical applications of VARs, i, j, k, and l, are chosen as finite lags.There is no standard method to determine the “correct” number of lags to include in such aregression. We include six lags, which is enough to study and correct for predictive relations of upto six weeks. As previously mentioned, our modification is to include contemporaneous order flowand return (j 0 and k 0) because these are the primary coefficients we are interested in.9Hasbrouck (1991) offers a clear discussion of the general advantages and disadvantages of the VAR.Though there are significant differences in the tests between this paper and his, many of Hasbrouck’s generalarguments help justify our own implementation of the VAR. For robustness, we also run simple correlationsand univariate regressions of the contemporaneous relation between stock returns and dealer turnover.13

In the context of the regression specification, Hypothesis 1 can be restated as follows: B0 andD0 have negative sign. This would be consistent with dealers that provide liquidity.3.2. Hypothesis 2: Dealers Earn Excess ReturnsAfter establishing whether dealers provide liquidity, we test the profitability of dealers inaggregate to determine whether they earn excess returns. Any outcome from testing Hypotheses 1would be relatively benign if dealers are not any more profitable than the average marketparticipant. However, if dealers are making excess returns using their institutional advantages,then this is a direct social cost of providing them these advantages. Since we have detailedintraweek transaction price and quantity data, we are able to test exact dealer profits much moreaccurately than most previous studies. In particular, we can disentangle returns attributable toinformation and to market-making.In each period, we split dollar profits and returns into three components: information,market-making, and mixed. The dollar profits from each of these sources is calculated as describedbelow and then converted to returns. The base value for the return is defined in each period by thevalue of the inventory held at the beginning of the current period (or equivalently, at the end of theprevious period). The total dollar profit is the sum of the three dollar profit components, and thetotal return is the sum of the three return components.To calculate the dollar profit of each component in a particular time period, we first split thestocks into those in which aggregate dealer net trading was positive and negative. The formulas forprofit breakdown of a single stock in week t are shown below. [ InformationComponent ] [ MarketMakingComponent ] [ MixedComponent ] ( NetTrade ) [ INVt 1* rt ] [GrossSell * ( Psell Pbuy )] [ NetBuy * ( Pt Pbuy )]14⑶⑷

( NetTrade ) [ INV * r ] [GrossBuy * ( Pttsell Pbuy )] [ NetBuy * ( Pt 1 Psell )]⑸Equations (4) and (5) denote profits from the cases where dealer net trading is positive andnegative, respectively. INVt-1 (INVt) is the share inventory level at the beginning (end) of the periodt; r is the return on the stock; GrossSell (GrossBuy) is the gross shares sold (bought); Psell (Pbuy) isthe average sell (buy) price for the shares sold (bought); NetBuy is the net shares bought; and Pt (Pt-1)is the price at the end (beginning) of the period. The terms in brackets represent profits frominformation, market-making, and mixed, respectively.Information dollar profits are defined as the increase in value of the inventory held for theentire period. These profits can be attributed to information because dealers were committed tohold the inventory for an extended period (at least the entire week), indicating they believed thatsuch positions in the stocks might be profitable. If dealer order flow in a given stock is positive(negative), then the amount held for the entire period is the inventory from the beginning (end) ofthe period.10 The dollar profits on the inventory are calculated based on the return of the stock andthis definition of shares held for the entire week.Market-making profits are dollars earned from shares bought and sold in the same period.These profits are attributed to market-making because of the nature of providing short termliquidity. Providing short-term liquidity is tantamount to buying when there are too many sellers inthe market and selling when there are too many buyers. In each case, the goal is to trade at anadvantageous price due to the order imbalance and to undo the position when the imbalancedisappears. This is equivalent to the famous quote, “Buy low, sell high!” To the extent that dealersare able to first buy and then sell (or vice versa) shares of a security within the same week, they aretrying to do just that. If dealer net trading is positive (negative), then the relevant number of shares10Recall the short sale constraint prohibits negative positions.15

is the dealer gross sell (buy) amount. Since we have actual transaction prices for the gross buys andsells, we can calculate a very exact estimat

Total Market Capitalization of stocks listed on the TSE in 2001 was NT 10.25 trillion (roughly US 316 billion) in 2001 for 614 listed stocks. Annual market volume was NT 18.35 trillion, so dollar turnover in 200