Department of Economics and Finance Working Paper: September 2011

Transcription

1 Department of Economics and Finance Working Paper: September 2011 Price Discovery in the U.S. Treasury Market: Automation versus Intermediation Kasing Man, Junbo Wang and Chunchi Wu Abstract This paper examines the contribution to price discovery by electronic and voice-based trading systems in the U.S. Treasury market. Evidence shows that the preponderance of price discovery in the Treasury market takes place in the electronic trading system. However, individual trades facilitated by voice brokers contain a significant amount of information per trade. The relative contribution of a trading system to price discovery depends on liquidity, volatility, trade size, and order imbalance. Although overall the electronic trading system has the most price discovery, the voice-based trading system contributes more to price discovery in the segment of large-size trades. Moreover, prices generated from the voice-based trading system are less noisy and volatile. These findings provide important implications for the design of electronic fixed-income markets by Junbo Wang. All rights reserved. Short sections of text, not to exceed two paragraphs, may be quoted without explicit permission provided that full credit, including notice, is given to the source.

2 Price Discovery in the U.S. Treasury Market: Automation versus Intermediation Kasing Man, Junbo Wang and Chunchi Wu * September 5, 2011 Abstract This paper examines the contribution to price discovery by electronic and voice-based trading systems in the U.S. Treasury market. Evidence shows that the preponderance of price discovery in the Treasury market takes place in the electronic trading system. However, individual trades facilitated by voice brokers contain a significant amount of information per trade. The relative contribution of a trading system to price discovery depends on liquidity, volatility, trade size, and order imbalance. Although overall the electronic trading system has the most price discovery, the voice-based trading system contributes more to price discovery in the segment of large-size trades. Moreover, prices generated from the voice-based trading system are less noisy and volatile. These findings provide important implications for the design of electronic fixed-income markets. JEL classification: G12; G14 Keywords: Price discovery; electronic trading; information share; liquidity; error correction; variance ratios * Kasing Man is at Western Illinois University, Junbo Wang is at City University of Hong Kong, and Chunchi Wu is at State University of New York at Buffalo. Junbo Wang and Chunchi Wu acknowledge the RGC research infrastructure grant by Hong Kong Special Administration Region, China. We thank Wei Xiong, Department Editor, an Associate Editor, two anonymous referees, Bin Chen, Sergei Davydenko, Craig Doidge, Esther Eiling, David Goldreich, Raymond Kan, Lisa Kramer, Peter Lerner, Bill Lesser, Jan Mahrt-Smith, Tom McCurdy, Romans Panco, Bill Schwert, Loren Tauer, Kevin Wang, Jason Wei, Alan White, Nese Yildiz and seminar participants at Cornell University, Risk Management Institute of National University of Singapore, University of Rochester, and University of Toronto for helpful comments. Address correspondence to Chunchi Wu, School of Management, State University of New York at Buffalo, Buffalo, NY ; chunchiw@buffalo.edu; Tel: ; Fax:

3 Technological innovations are dramatically transforming the organization and operations of markets. Markets that have traditionally relied upon collocation of traders and products are steadily transitioning to electronic forums. Information technology (IT) has affected both the way markets operate and their performance. Recently, the rapid development and widespread availability of information technology have spurred considerable interest in research on the economic consequences of automation in trading and changes in the design of markets (see Banker and Kauffman, 2004; Hendershott and Moulton, 2011; Hendershott, Jones and Menkveld, 2011). A vast literature has investigated the impacts of IT on market performance. Most studies have shown that electronic trading lowers trading costs and price dispersion and increases the speed of information transmission and transactions, transparency, and access to the market (see, for example, Brynjolfsson and Smith, 2000; Garicano and Kaplan, 2001; Barclay, Hendershott and Kotz, 2006; Easley, Hendershott and Ramadorai, 2009; Garvey and Wu, 2010). The adoption of electronic forums is shown to reduce costs of search, presentation, and transactions, particularly for rare and heterogeneous products (Bakos, 1997; Kambil and van Heck, 1998; Brynjolfsson, Hu and Smith, 2003). By allowing for remote access to broaden market base, electronic trading channels also enhance liquidity and market efficiency. Although electronic trading offers benefits such as low cost and high speed of trading to market participants, it may create other problems. Overby and Jap (2009) show that the use of electronic trading channels can increase trading risk. 1 Dewan and Hsu (2004) find adverse selection in ebay auctions, which incurs significant adverse selection cost and price discount. Hendershott and Moulton (2009) find that trading automation in the NYSE increases adverse selection. Electronic trading may also increase volatility and decrease market depth (Levecq and Weber, 2002; Hendershott, Jones and Menkveld, 2011). An important issue in the IT literature is the choice of trading channels. Quite often separate electronic markets have been launched and in some cases new electronic trading channels have been added to the physical market to create a hybrid structure. An important question is what determines traders choice between the two trading venues. Overby and Jap (2009) show that the choice of electronic and physical channels in a market depends on product quality risk and transaction costs. They find that transactions with low quality uncertainty and high search cost occur in the electronic channels, and transactions with high quality uncertainty and low search cost occur in the physical channels. Hendershott (2003) suggests that the adoption of electronic trading depends on types of securities and market environments. Weber (2006) finds that firm-specific factors and effects of network and liquidity 1 Electronic trading channels are capable of accurately transmitting digital attributes of products such as size, weight, catalog number, and description but not the non-digital attributes like product quality and trader reliability. Inability to transmit non-digital attributes induces noise, misrepresentation and possibly, fraud, which can incur information asymmetry and adverse selection costs. While these problems are common in traditional physical markets, electronic markets are particularly vulnerable since information asymmetry tends to higher when buyers and sellers are separated by time and space. 1

4 externality influence the willingness to adopt an electronic trading venue. While information technology has affected virtually every aspect of the economy, its impact on financial markets is particularly strong. Financial industry is highly dependent on information technology, and is one of the most information-intensive sectors of the US economy. On average, a financial company allocates about 20% of its capital outlays to information systems and half of the revenue of trading desks goes to information technology (Dewan and Mendelson, 1998, 2001). The importance of information technology infrastructure is manifested by the proliferation of sophisticated trading strategies facilitated by algorithm trading and complex financial securities in the market. In the global financial market, trading depends heavily on information technology and the speed and accessibility of information is paramount. 2 Innovations in computing and communication technology enable global electronic routing and broad dissemination of quote and trade information. Like in the physical product market, information technology has dramatically transformed the financial market. It has increased the speed of information processing and efficiency of trading and settlement, lowered operation and trading costs and latency, and allowed remote access to the trading venue. Although information technology lowers costs and improves operational efficiency and liquidity in financial markets, it is less clear whether it can increase pricing efficiency and reduce information risk (Cardella and Hao, 2010). The securities market is an ideal forum for evaluating the impacts of information technology on market quality. The securities market is more efficient than any product markets and data generated from the securities market are much more complete and of highest quality. Long-span data series of low measurement errors facilitates powerful tests. Moreover, securities traded in the exchanges are standardized instruments, which allow researchers to have a better control for the effect of characteristics of goods and provide more accurate assessment on the impact of changes in information system technology. Past studies have shown that the design of information systems can affect market quality (Bard and Bejjano, 1991; Levecq and Weber, 2002). Clemons and Weber (1997) show that an information system offering an effective signaling mechanism is essential for communicating information about the nature of the trade and reduces information uncertainty. Dewan and Hsu (2004) suggest that a proper design in the trading mechanism to signal the quality of trades can reduce adverse selection cost in electronic markets. Weber (2006) shows that a trading system designed to integrate high trading speed with information and services has a much better chance to succeed. Advanced information and communication technology have enabled faster information propagation and arbitrage at lighting speeds, which have potential to reduce pricing errors, and enhance risk sharing, hedging, securities allocation, and liquidity in the financial market. This could ultimately reduce cost of capital and increase efficiency of the financial market and resource allocation. However, technology has 2 See Artend (1991). 2

5 complex effects which are not always desirable and empirical evidence has shown mixed effects of automation in the trading process. Hendershott and Riordan (2009) find that algorithmic trading contributes more to discovery of the efficient price than human trading. Hendershott and Moulton (2011) show that increasing automation and trading speed after the NYSE s introduction of its Hybrid Market reduces the noise in prices, making prices more efficient, but raises the cost of immediacy. Hendershott, Jones and Menkveld (2011) find that algorithmic trading has narrowed spreads, reduces adverse selection but reduces trade-related price discovery. Their results suggest that algorithmic trading improves liquidity and enhances the informativeness of quotes. In a separate vein, Hasbrouck and Saar (2009) find that fleeting orders induced by improved technology can impair market liquidity. A fleeting order refers to the case that a trader submits and quickly cancels a hidden limit order, or uses an immediate-or-cancel order. This type of limited order consumes liquidity rather than provides liquidity. Electronic trading can also increase market vulnerability and volatility as experienced by several flash crashes occurred recently, and high frequency trading could be used to manipulate the market (Nishimura, 2010). This paper examines the impact of fully automated trading on price discovery of the U.S. Treasury market. In the secondary U.S. Treasury market, virtually all trades between dealers are handled by interdealer brokers (IDBs). There are two types of interdealer trading systems: electronic and voice-based. In the former, orders are matched electronically whereas in the latter, orders are placed and trades are executed over the telephone. By interacting with dealers over the phone, a voice broker is able to collect more information and offer more services to participating dealers. A voice broker can search and negotiate with a compatible trading partner on behalf of the client to obtain a better price. These services are valuable when order size is large and volume is thin. Trading through voice brokers takes up more time but orders may be executed with price improvement. By contrast, electronic brokers offer speedy trading and anonymity, which may encourage informed trading and enhance information efficiency. It remains unclear to what extent each trading system contributes to price discovery in the Treasury market as a whole and in different segments of this market. The interdealer broker market of U.S Treasuries provides an excellent laboratory for studying the roles of automation versus human intermediaries in the price discovery process. First, Treasury securities traded in the electronic and voice-based platforms are virtually identical instruments. Their payoff structure is similar and there is no complicated firm-specific information involved as in the trading of stocks or equity derivatives. Second, Treasury securities are traded by dealers and interdealer brokers using the same trading and settlement procedures except that one venue is automated and the other is voice-based. This enables us to have better control for effects associated with differences in security characteristics, trading mechanism and market structure, which was not feasible in previous studies on price discovery involving different trading systems. For example, past studies on price discovery 3

6 contribution of regional exchanges, Nasdaq and ECNs for the NYSE-listed stocks were unable to isolate the effects of differences in market structure and the quote-setting process (see Benveniste, Marcus, and Wilhelm, 1992; Heidle and Huang, 2002). 3 Similar problems are encountered by studies on information efficiency of spot versus derivatives markets (see Easley, O Hara and Srinivas, 1998). As a consequence, it is unclear whether differences in price discovery contributions between these venues are due to different market structures/trading mechanisms, security characteristics, trading platforms, or something else. By effectively isolating the differences in market structures, trading mechanisms and security characteristics, we provide cleaner evidence on the contribution of the electronic trading system to price discovery relative to human intermediaries. Barclay, Hendershott and Kotz (2006) are the first to study the choice of the electronic versus voicebased trading system by dealers in the U.S. Treasury securities to determine what services provided by human intermediaries are difficult to replicate in a fully automated trading system. They analyze the choice of the trading venue by dealers when an on-the-run Treasury security goes off the run. By assuming that there is no significant difference in information asymmetry between the on-the-run note and the newly off-the-run note, they emphasize the aspect of liquidity provision and the matching function of the two trading systems. 4 They find that human intermediaries can uncover hidden liquidity and facilitate better matching of less liquid off-the-run Treasury securities. While Barclay et al. (2006) provide great insights into the matching function performed by human intermediaries, they have not explored voice brokers role in the price discovery process of the U.S. Treasury market. This issue is important because recent studies have uncovered significant information asymmetry in the Treasury market, contrary to the perception that there is little information asymmetry in active Treasury securities. Using interdealer transaction data, Brandt and Kavajecz (2004) and Green (2004) have documented convincing evidence of information asymmetry in the Treasury market even though securities traded in this market are relatively homogeneous with simple known payoffs. In particular, they find that informed trading concentrates on the on-the-run Treasuries. Similarly, using the same transaction dataset, Li, Wang, Wu and He (2009) find strong evidence of higher informed trading intensity for the on-the-run note than for the off-the-run note and that information risk is priced. These findings have significant implications for price discovery via different trading venues. If human intermediaries are more capable of discerning information-based trades, informed traders will prefer not to trade in the voice-based system because revealing their identity will place them at a disadvantage. To the extent that price discovery depends on where informed traders trade, the human-assisted trading 3 These studies find that market structure affects the intensity of information-based trading. In addition, Barclay, Hendershott and McCormick (2003) show that differences in the trading process can affect the amount of information revealed and the timing of that revelation. 4 See Barclay et al. (2006, p. 2413). 4

7 system would tend to have less price discovery. On the other hand, human intermediaries provide more information to traders, which may result in better order execution and less noisy prices. In this paper we evaluate the trade-off between these two functions. Our work complements the Barclay et al. (2006) study by exploring the informational roles of electronic and voice interdealer brokers in the price discovery process under information asymmetry. It contributes to the literature by providing direct evidence on the relative contribution of electronic and voice-based trading to price discovery in the U.S. Treasury market as well as in different segments of this market. Understanding the price discovery process in the U.S. Treasury market is important for academicians, practitioners and policy makers. Price discovery the efficient and timely incorporation of new information into security price is arguably the most important function of securities markets (see Lehmann, 2002; Huang, 2002). Financial economists are interested in information efficiency and price discovery efficacy of a trading venue. Measuring information efficiency and tracking the price discovery process are of particular interest to policy makers who monitor market conditions regularly and have a great deal of concern about market quality. A question of fundamental importance is whether automation in trading improves market quality. The Treasury market is an important market with marketable issues of about 9.1 trillion. 5 From the investment perspective, it is essential to ascertain how price is formed in this market. More importantly, Treasury price affects the performance of almost every facet of financial markets as riskfree rates of varying maturities are benchmarks for pricing other financial assets. Understanding Treasury price formation is thus an integral part of any unified asset pricing theory. Recently, there is increasing interest in the issue of price discovery in the U.S. Treasury market (see, for example, Balduzzi, Elton and Green, 2001; Huang, Cai and Wang, 2002; Green, 2004; Brandt and Kavajecz, 2004; Vega, 2006; Brandt, Kavajecz and Underwood, 2007; Pasquariello and Vega, 2007, 2009; Li et al., 2009). These studies examine issues related to price discovery of the Treasury market using the data from GovPX, a firm that consolidates quote and trade information from several voice interdealer brokers. Unlike these studies, we assess the contribution to price discovery by voice-based and electronic trading systems of the interdealer broker market using both GovPX data and an electronic transaction dataset made available by BrokerTec. Boni and Leach (2002) examine depth discovery using only GovPX data. Mizrach and Neely (2006) use espeed data to examine the effect of trade automation on bid-ask spreads, trading volume and price impacts. Our paper differs from these studies by comparing price discovery across trading venues using a comprehensive transaction data set and systematic price discovery measures. Our paper makes several unique contributions to the literature. First, using the long-span transaction data from GovPX and BrokerTec, we document the first empirical evidence on price discovery across 5 See Treasury Bulletin, Federal Debt, June Of the total Treasury debt outstanding (14.3 trillion), the amount of nonmarketable debt is 5.2 trillion. 5

8 trading venues in the interdealer brokerage market of US Treasuries. We find that the electronic trading system has the most price discovery. The electronic platform assumes a more dominating role in price discovery in the Treasury market than that documented for the equity market (see Huang, 2002; Theissen, 2002). This discrepancy could be due to less quality uncertainty and relative homogeneity in the Treasury securities, which make it easier to trade in the electronic market. Second, although overall the electronic trading system has the most price discovery, individual trades facilitated by voice brokers contain significant information on a per-trade basis. Given the low market share (volume) of the voice-based trading system, the amount of price discovery on this system is relatively high. In addition, the voice-based trading system contributes more to price discovery in the segment of large-size trades. As large-size trades account for a significant portion of the total on-the-run trading volume, voice brokers services are important even for active, liquid on-the-run securities. Third, prices generated from the voice-based trading system are less noisy and less volatile. Order executions by voice brokers are of better quality, which is attributable to information search by voice brokers. Fourth, the contribution to price discovery by each trading system varies over time. On days when trading is less active and liquidity is low, when trades are larger, when it is more costly to supply liquidity by placing a firm limit order, and when a dealer has a larger imbalance between buying and selling, the contribution of the electronic trading system to price discovery is lower. This finding implies that price discovery will be quite different for off-the-run securities. To the extent that voice brokers provide more liquidity to off-therun securities, they should play a more important role in price discovery of the off-the-run securities market. This argument is consistent with the finding that GovPX has a much larger market share in offthe-run securities (see Barclay et al., 2006). Our results also suggest that human services will be more important during turbulent periods when liquidity is scarce. Lastly, the information share of the electronic trading system increases significantly surrounding the macroeconomic news announcement. Results strongly suggest that traders prefers the electronic trading venue that offers speedier order execution in anticipating that prices will react quickly to the news announcement. Our empirical findings also contribute to the broad literature of information systems and communication technology. The design and implementation of electronic markets involves complicated computational, economic and behavior issues and resolving these issues requires interdisciplinary efforts. Although electronic trading has come to play a very important role in equity markets, the adoption of the fully automated trading system is slow in the fixed-income market. Understanding why electronic platforms are less successful in the fixed-income market provides valuable information for developers of new trading platforms. Our study extends our understanding of what determine a successful adoption of an electronic trading system in the fixed-income market. We show that electronic and human-assisted trading systems perform different functions and contribute to price discovery of securities segments with 6

9 different size and liquidity in the Treasury market. The introduction of the electronic trading system in the interdealer brokerage market of US Treasuries has not caused traders to completely abandon the humanassisted trading system. Previous studies suggest that human intermediation survive because there is quality and information uncertainty in traded products (see Koppius, van Heck and Wolters, 2004; Overby and Jap, 2009). Our study shows that there remains an important role for human intermediation in relatively active securities such as Treasury securities with little quality uncertainty and simple payoff structures. This finding implies that information uncertainty and quality are not the only issues in the implementation of the electronic trading system. Rather, the characteristics of traders and instruments are likely to be more important factors in the design of electronic fixed-income markets. Our empirical evidence suggests a combination of electronic and human trading systems should improve the welfare of market participants. In addition, our results have practical implications for the design of new trading platforms. The automation of the trading process has long been an important dimension of the financial market design. Our results suggest that information system designers must take into consideration of the features of products and trading environment in the fixed-income market dominated by institutions and block trading. Evidence shows that the electronic trading platform is more suitable for more homogeneous securities with a simple payoff structure and smaller trade size. Our finding suggest that when deciding the appropriate mixture of electronic and human-assisted channels or adding communication devices to assist trading by participants, it is important to consider characteristics of the traded instrument, trade size and market environments. The remainder of this paper is organized as follows. Section I reviews the related literature and Section II describes empirical methodology and the procedure for estimating the amount of price discovery. Section III discusses the data and presents empirical estimates. Section IV examines the determinants of information shares. Section V reports additional tests for the amount of new information incorporated into prices of transactions with different sizes, noisiness of price signals, and macroeconomic news announcement effects. Finally, Section VI summarizes our major findings and concludes the paper. I. Related Literature This paper is related to a large literature of interdisciplinary research for the design of electronic markets (Anandalingam and Raghavan, 2005; Viswanathan, 2005) and the effects of information and communication technology on market quality (Bakos, 1997; Kambil and van Heck, 1998; Brynjolfsson, Hu and Smith; 2003; Dewan and Hsu, 2004; Banker and Kauffman, 2004; Overby and Jap, 2009). Our work is distinguished from these IT studies by focusing on the effect of electronic trading on Treasury market quality. Because this paper examines the roles of electronic and brokered markets, it is also closely related to studies on the choice of trading channels in the product market (Brynjolfsson and Smith, 7

10 2000; Kazumori and McMillan, 2005; Overby and Jap, 2009). While these papers examine electronic and physical trading channels for heterogeneous physical products, we focus on the choice of electronic and voice-based trading venues in the US Treasury market with relatively homogenous securities. In comparing different trading systems, it is important to understand what values human intermediaries can bring to market participants. The existing literature suggests that human intermediaries serve two major functions in security trading. First, human intermediaries are more efficient in consummating customer orders during the time when trading volume is low, trade size is large, or a large imbalance in buying and selling is present (see Barclay, Hendershott and Kotz, 2006). Second, human intermediaries are able to detect informed trades when there is information asymmetry, and offer better protection and prices to their customers through interactions with market participants (see Seppi, 1990). These important functions explain why securities with higher information asymmetry and lower liquidity tend to trade through human intermediaries such as specialists and multiple dealer systems. For example, Nasdaq market makers have a greater share in smaller, less frequently traded stocks with higher information asymmetry whereas electronic communications networks (ECNs) have a bigger market share in the largest, most frequently traded Nasdaq stocks (see Barclay, Hendershott, and McCormick, 2003). Similarly, NYSE specialists have higher participation rates in smaller, less liquid stock trading (see Madhavan and Sofianos, 1998). Barclay et al. (2006) examine the choice of trading venues by dealers in the U.S. Treasury securities market. 6 They find that there is an important role for human intermediaries in relatively inactive Treasury securities such as off-the-run issues. Voice brokers services are most valuable when orders are difficult to match, i.e., when trades are large and there is a large imbalance between buying and selling. The difference between this paper and ours is that we focus on the price discovery function of electronic and voice-based trading systems, rather than the matching function and liquidity provision. Along this line, we document evidence that human intermediaries are important for active on-the-run Treasuries, not just in liquidity provision but also in price discovery. Our study on the roles of electronic and voice-based trading venues is tied to studies on multimarket trading. Huang (2002) compares the quality of quotes posted by electronic communication networks (ECNs) and Nasdaq market makers for actively traded stocks. He finds that ECNs contribute significantly to price discovery and ECN quoted spreads are much smaller than those posted by Nasdaq market makers. Barclay et al. (2003) find that ECNs attract more informed trades, and these trades are more likely to occur when trading volume, volatility, and information asymmetry are high. Their finding contrasts with Easley, Kiefer and O Hara (1996) and Bessembinder and Kaufman (1997) who claim that the secondary market such as regional exchanges skim less-informed order flow from the primary 6 Barclay et al. (2006) employ data provided by the Fixed Income Clearing Corporation (FICC) from January 2001 through November

11 markets, resulting in higher trading costs and lower depth in the NYSE. Grünbichler, Longstaff and Schwartz (1994) find that prices of screen-traded futures contracts lead spot prices of the German stock index, whose component stocks are floor traded. Seppi (1990), Grossman (1992), Keim and Madhavan (1996), Booth, Lin, Martikainen and Tse (2002), and Bessembinder and Venkataraman (2004) examine off-exchange brokered markets. These studies show evidence that upstairs brokers are able to uncover hidden liquidity and match block orders more efficiently in the equity market. Venkataraman (2001) examines execution costs on the Paris and New York Stock Exchanges and finds that bid-ask spreads are wider on the electronic exchange. Jain (2005) examines the impact of automation and provides evidence that electronic trading improves liquidity and informativeness of stock markets and lowers the equity premium. Mizrach and Neely (2006) review the transition to electronic trading in the secondary Treasury market and show that the electronic trading system has greater volume, smaller spreads and lower trade impacts. Unlike these studies, we examine issues related to the price discovery function of the electronic and voice-based trading systems in the Treasury market. Our study on the price leadership of different trading venues is motivated by several recent studies on price discovery in the U.S. Treasury market. Green (2004) examines the informational role of trading in the U.S. Treasury market surrounding macroeconomic news releases and finds that the level of information asymmetry significantly increases after the news announcement. Brandt and Kavajecz (2004) find that order flow imbalances explain a substantial portion of day-to-day variations in government bond yields, even on days without major macroeconomic announcements. They show that the effect of order flow on yields is permanent and is due to private information rather than liquidity or inventory effects. Pasquariello and Vega (2007) examine the role of private and public information in the process of price formation in the U.S. Treasury market by considering the effects of information heterogeneity and imperfect competition among insiders. They find that daily bond price dynamics are strongly related to fundamentals and agents beliefs. Li et al. (2009) show evidence that information risk significantly affects long-term expected returns of Treasuries. These papers employ the voice-based transaction data to study price discovery. By contrast, we investigate the channels of price discovery in the Treasury market via electronic and voice-based trading systems based on the trading data from both platforms. A related literature examines the issue of price discovery contribution by a single market when homogeneous or closely related securities are traded in multiple markets. Several papers have attempted to determine where price discovery is being produced by examining stocks listed on the New York Stock Exchange but traded on other venues (see Garbade and Silber, 1979; Harris, McInish, Shoesmith and Wood, 1995; Hasbrouck, 1995). Theissen (2002) examines price discovery in floor-based and electronic trading systems of the German stock market and finds both trading systems contribute to price discovery almost evenly. Another strand of research looks into spot and derivatives markets to determine where new 9

12 information is first incorporated in prices (see Garbade and Silber, 1982; Kawaller, Koch and Koch, 1987; Bhattacharya, 1987; Stephan and Whaley, 1990; Stoll and Whaley, 1990; Chan, 1992; Easley, O Hara and Srinivas, 1998; Mizrach and Neely, 2008). Our paper differs from these studies in that we examine price discovery in the trading of the same Treasury securities via two distinct venues in the interdealer brokerage market. Lastly, our empirical methodology is linked to several studies on the measurement issue of price discovery (De Jong, 2002; Harris, McInish and Wood, 2002; Lehmann, 2002; Theissen, 2002). De Jong (2002) and Lehmann (2002) compare the Hasbrouck and Gonzalo-Granger (GG) measures and recommend using both approaches in empirical research. Baillie, Booth, Tse and Zabotina (2002) provide evidence on differences and similarities in empirical results based on these two measures using equity market data. In empirical investigation, we construct both price discovery measures for comparative purposes. II. Trading Venues and Price Discovery When securities are traded in a single centralized market, price discovery is solely produced by that market. Conversely, when trades of a security take place in multiple trading venues, questions naturally arise as to what is the relative contribution of each venue to price discovery of the whole market. The answer to this question to a large extent depends on the functions that each trading system performs and how valuable these functions are for market participants. In the interdealer Treasury market electronic and voice-based systems perform several similar functions. Limited order information is continuously displayed in both trading systems and dealers can trade aggressively by hitting an existing limit order or trade passively by submitting a limit order through either venue. Trade prices and volume are made publicly available immediately after transactions through the Internet and data vendors. More recently, the electronic trading platform has added a number of features aimed to mimic voice brokers services to traders with large orders to fill, i.e., auto-refresh to replace a limit order once it is executed, a reserve size feature to indicate likely additional demand, and a negotiation feature to directly negotiate large trades. Despites these efforts, there remain unique human services which cannot be replicated by the electronic trading platform, particularly when there is complex qualitative information that cannot be easily conveyed through the electric trading system. An important function of voice brokers is bringing together counterparties with substantial size to trade for liquidity reasons or trades with complicated terms which are difficult to match. When the order is large or terms are complicated, a bilateral search for the most favorable bids (or asks) by an individual dealer can be cumbersome. Acting as the focal point for buy and sell orders, voice brokers reduce the cost of information search and waiting time for such trade execution. Voice brokers play a subtle role in matching buyers and sellers with larger orders. In negotiating with the other party, the voice broker 10

13 attempts to conceal information about his client and transaction amount. This is particularly desirable for a large trade because revealing the order information will place a trader at a disadvantage when trading requires extensive search for the counterparty. The broker will only reveal the buyer s information to a natural counterparty, i.e., another compatible trading partner having similar size and needs to trade. Through repeated interactions with buyers and sellers, the voice broker may uncover hidden liquidity and detect informed trades to protect their customers against adverse selection. Orders matched by voice brokers carry higher cost and so a higher commission is charged, which results in higher marginal cost of liquidity supply. By contrast, because orders are matched without human intervention, the electronic trading platform incurs lower cost per trade and charges a lower commission. Lower cost of trading and speedy order execution attracts traders to participate in the electronic system. These are main reasons that the market share of the electronic trading system has steadily increased over time since its inception. Market microstructure theory suggests that informed traders choose to trade in a venue with higher liquidity and greater depth. Also, informed traders like to trade in a venue where their activity is less likely to be detected, fearing that revealing their presence will drive liquidity traders out of the market place. This means that informed traders would prefer a trading venue that offers liquidity, anonymity and speedy execution. The electronic platform seems to have most of these features that suit informed traders. On the other hand, the electronic trading platform has no brokers to facilitate trades. Customer orders are directly crossed with one another. While speedy matching and lower trading costs are quite appealing to liquidity traders, without brokers assistance their orders may be executed with prices substantially deviated from the full-information or fundamental value. Lack of information coupled with high speed of trading may generate larger or more frequent temporary price distortions. As such, transaction prices or quotes in the electronic trading system may be prone to errors. Thus, it is not clear whether the electronic trading system would always prevail in price discovery. In this section, we construct empirical measures to determine which trading venue contributes more to price discovery in the interdealer Treasury market. We construct price discovery measures using the methods suggested by Hasbrouck (1995) and Gonzalo and Granger (1995) to gauge the contribution to price discovery by electronic and voice-based IDB trading systems. These are two widely accepted price discovery measures for security trading in multiple markets in microstructure literature (see Ballie et al., 2002; Huang, 2002; Lehmann, 2002). We outline the estimation procedure here. Constructing these measures requires a specification of price dynamics. Consider two trading systems with prices represented by two cointegrated I(1) series P t ( t t x, y )', which share a common implicit efficient (equilibrium) price. The error-correction term of the cointegrated I(1) price series is ( 2 zt ' Pt xt 2 y and the (normalized) cointegration vector is 1, )'. The vector error-correction model (VECM) can be expressed as t 11

14 Pt ' Pt 1 A1 Pt 1 A2 Pt 2... Ar 1 Pt r 1 et (1) where 1, )' and 1 B is the first difference operator. The error term e t is a zero-mean vector ( 2 of serially uncorrelated innovations with a covariance matrix ij where and are the innovation variances, and is the covariance with being the correlation of innovations in the two markets. 7 Prices for the same security traded in different markets would tend to converge to a common efficient price in the long run but might deviate from each other in the short run due to imperfect market conditions. The Hasbrouck measure is defined as the proportional contribution of a market s innovation to the innovation in the common efficient price. This price discovery measure essentially captures the extent of the efficient price variation explained by the innovation in each market. By contrast, the Gonzalo- Granger method decomposes the price into the permanent and transitory components, and associates the permanent component with the long run price. The weight given to price discovery is defined as the change in the permanent component with respect to the information shock. For convenience, we refer to the common factor weight in the Gonzalo-Granger model as the GG measure, and use the term information share to represent the contribution of a trading system to price discovery of the Treasury market for both measures in the remaining analysis. The Gonzalo-Granger measure for x t and GG x 2, and 1 2 y t are defined as 1 GG y. (2) 1 2 As indicated, the GG measure is primarily based on the speed (the α coefficient) of the error-correction term, or how price in each market changes in response to the preceding price disequilibrium in the two markets. 8 By contrast, the Hasbrouck measure accounts for the innovation e t in the two markets, in addition to the speed of adjustment. The Hasbrouck measures for x t and y t are given as follows: Given the ordering of P t H x ( ) ( U ) 2 ( ) ( ) ( ) H y ( L). (3) ( ) ( 1 ) ( t t x, y )', the above measures generate the maximum value (upper bound) for the first price series x t and the minimum value (lower bound) for the second price series y t. Reversing 7 The VECM setup is consistent with the microstructure model suggested by Glosten (1987). See also Hasbrouck (1996) and Booth et al. (2002). 8 It has been shown in the literature that there is no guarantee that the coefficients are always between 0 and 1 in empirical estimation, and so the GG measure for the market can be less than zero or above one. 12 2,

15 the order and re-estimate the VECM for ( t y t, x )' gives the upper bound H y (U ) for y t and the lower bound H x (L) for x t. The upper and lower bounds can be averaged to give a single measure for each market. When the correlation coefficient is small, the two bounds are tight and their average is an effective summary measure. Both Hasbrouck and GG measures have merits and drawbacks, and there is no consensus in the literature as to which one is superior. Lehmann (2002) and the references therein give a thorough comparison between these two measures. Appendix B provides some further discussion of these measures. In our empirical investigation, we report both price discovery measures for comparison. For the analysis to be meaningful, it is important to check if the I(1) cointegration assumption is satisfied for the two price series representing the electronic and voice-based trading systems, respectively. Our empirical analysis begins with performing the augmented Dickey-Fuller (ADF) unit root test for the price series. After confirming that unit root exists in each price series and that the two series are cointegrated, we set up a VAR model for the bivariate price series and use BIC to determine the AR order in the model. This in turn determines the AR order used in the VECM model. Based on this model, we use Johansen's trace statistic to determine the number of cointegration vectors in the price series. As shown later, our empirical evidence points to one cointegration vector and hence the ' specification for the lagged disequilibrium error in (1) is appropriate. 9 Based on the estimated VECM model, we compute the Hasbrouck (H) and Gonzalo-Granger (GG) measures. The Hasbrouck's upper and lower bounds are calculated and their averages are reported. In addition, we examine the dynamic relations between the two price series and analyze the pattern of impulse response to a unit shock in each trading system. We now turn to empirical estimation. III. Data and Estimation of Price Discovery Measures Trades and quotes of U.S. Treasury bonds are from GovPX and BrokerTec datasets. The GovPX, Inc. was set up under the guidance of the Public Securities Association as a joint venture by primary dealers and several interdealer brokers in 1991 to increase public access to U.S. Treasury security prices. GovPX consolidates trade and quote information from voice interdealer brokers. Historically, the interdealer brokerage market was operated primarily through a voice-based system. Later, several interdealers developed platforms for electronic trading. Acquired by ICAP in May 2003, BrokerTec provides access to a pool of liquidity that includes major participants in both the U.S. and European fixed-income markets. 10 Publicly accessible trade and quote data for the electronic trading platform are recently made available by BrokerTec. 9 Details for the implementation of these methodologies are given in Zivot and Wang (2006), which we follow in our computations. 10 Fourteen primary dealers formed an electronic trading platform BrokerTec for interdealer trading. For the discussion of the evolution of the electronic trading system, see Mizrach and Neely (2006). 13

16 The GovPX dataset contains detailed individual security information such as cusip, coupon, maturity date, and an indicator of whether the security is trading when issued, on-the-run, or off-the-run. The GovPX data date back to 1991 (June 17). The BrokerTec dataset also includes individual Treasury security information, price and quantity but only for on-the-run issues. We have BrokerTec data from October The BrokerTec data are continuously available on the daily basis only for the 2-, 5-, and 10-year on-the-run Treasuries though the dataset covers other maturities. Therefore, we focus on the 2-, 5- and 10-year on-the-run notes. 11 Because our empirical analysis requires both GovPX and BrokerTec data, our study period starts from October The sample period ends in Both datasets contain quote and trade information and trade side (buy or sell). The quote data include the best bid and ask prices, associated yields, the time of each bid and offer, and the respective bid and ask size (depths). The trade data include the time of trade, trade size, price and yield. In empirical investigation, we use both quotes and transaction prices to obtain price discovery measures. The BrokerTec dataset includes a variable for identifying each individual trade while the GovPX reports accumulated trading volume instead. Trades in the GovPX dataset can be identified by changes in accumulated volume. However, GovPX stopped reporting accumulated volume after March 2001 and so we need to use other information to determine trades. In this study, we identify trades in the GovPX dataset based on the information for changes in trade sign (hit or take), price and size, and work-up information. We describe this trade identification procedure and evaluate its accuracy in Appendix A. The basic principle behind this procedure is that if one line (record) of observations in the dataset has updates on any of three important items: trade sign, price or size, we would treat it as a new trade. To check the reliability of this trade identification algorithm, we use the GovPX data from 1992 to 1999, a period when accumulated trading volume is reported such that trades can be identified accurately and a large number of transactions are covered, to evaluate the precision of our procedure. Overall, we find our procedure can identify trades with about 96% accuracy, suggesting that our trade identification procedure is quite reasonable. Another issue is related to the reporting process of workup trades. Both electronic and voice brokers adopt the expandable limit order protocol, which grants a trader whose order has been executed the rightof-refusal to trade additional volume at the same price. The workup process refers to this quantity negotiation period. Although strict anonymity is enforced throughout the workup process in each trading system, there is a difference in the reporting procedure of workup trades by electronic and voice brokers. The GovPX posts the completed transaction after all trading interest is exhausted at the initially-agreed price. By contrast, the BrokerTec reports each individual component (increment) as a separate trade. This reporting discrepancy causes downward bias in trade size and upward bias in the number of transactions 11 For example, the data for 3- and 30-year on-the-run Treasuries are not continuous partly because these securities are not auctioned regularly. 14

17 on the BrokerTec system. To provide consistent comparison for the transactions of the two trading systems and empirical results based on transaction data, we need to cumulate the components of workup trades on BrokerTec into transactions comparable to GovPX trades. BrokerTec provides timestamp and detailed information such as the name and Cusip of the security, price, workup state and whether an order is aggressive or passive. These important data allow us to identify the individual components associated with a particular workup trade precisely and to sum all individual components of a workup trade from BrokerTec into a completed trade to provide a consistent comparison with the GovPX trade. Table I reports the summary statistics of average daily trading price (p), volume ($billion), trading frequency, number of quotes, bid-ask spread, trade size, volatility and depth. The quoted bid-ask spread (in hundredths of a percent of price) is the ask price minus the bid price divided by the midquote. Bond price volatility ( ) is calculated each day using the range-based method suggested by Alizadeh, Brandt and Diebold (2002), where t = exp[ln(max p t min p t ) 0.43]. Depth is measured by the average of bid and ask quote size. Panel A reports results before the adjustment for BrokerTec workup trades. Bid-ask (quoted) spreads are lower, and volume, trading frequency, depth and price volatility are higher for the electronic trading system. Average trade size is much larger for the voice-based trading system. Panel B reports results after adjusting BrokerTec workup trades to give a consistent comparison with GovPX data. As expected, the adjustment reduces the number of trades and raises the size of trades for the BrokerTec. But the effect on average price and volatility is quite modest. Despite the adjustment, results continue to show that trade size is much larger for GovPX, suggesting that larger orders are more likely to be handled by voice brokers. We use the data adjusted for BrokerTec workup trades in all empirical analyses based on transaction data. [Insert Table I here] Figure 1 shows the time trend in GovPX s market share in terms of trading volume. The market share of the voice-based trading system has declined since This confirms that the electronic trading system has gained the market share over time in all segments of the interdealer broker market of on-therun U.S. Treasuries. [Insert Figure 1 here] Price series are constructed from GovPX and BrokerTec datasets at 5-minute intervals from 7:30-17: If a price is not available at the end of an interval, we linearly interpolate the prices right before and after the end of that interval. If there is no transaction at all within an interval, that interval is regarded as having missing data. The return is calculated for each 5-minute interval and trading volume is the sum 12 Although the Treasury market is a 24-hour round-the-clock market, the number of transactions drops significantly after 17:00ET. We constrain the trading period to 7:30 to 17:00 each day. This period mainly covers the trading activity in New York and excludes the overnight returns from estimation. 15

18 of all trades over each interval. A. Unit Root Test, Number of Cointegration Vectors, and VECM Estimation We employ both midquotes and transaction prices in our empirical estimation. Hasbrouck (1995) indicates that results based on transaction prices are vulnerable to the problem of autocorrelation induced by infrequent trading and stale prices. By contrast, quotes are updated more frequently and therefore, more informative. Using quote data also alleviates the burden to identify trades. Following Hasbrouck (1995), our inference is mainly based on quote data. However, as transaction prices contain additional information, we also report the results using transaction prices to provide additional evidence. We first examine temporal properties of 2-, 5- and 10-year notes over the sample period Time-series plots of BrokerTec and GovPX prices do not show an obvious time trend over the entire sample period and the standard t-test fails to reject the null hypothesis that the differenced price series have a zero mean. We then perform the augmented Dickey-Fuller (ADF) unit root test for each price series. The ADF test regression takes the following form: p t p t q 1 p (4) j 1 j t j t where p t represents the midquote or transaction price for a trading system at time t. The null hypothesis is that there is a unit root, and the test aims at examining if is zero. For the lag order, we first try q = 6 and repeat with 12 lags. 13 We found that these two choices of lag order q produce essentially the same test conclusion, suggesting that results are robust. The ADF tests for each bond group using both midquotes and transaction prices all fail to reject the null hypothesis that there is a unit root. The p-values of the tests are all higher than 0.5. We apply Johansen's trace test to the VECM to determine the number of cointegrated vectors. In each bond group, the hypothesis of no cointegration vector is rejected at the one percent significance level, and the hypothesis that there is one cointegration vector cannot be rejected. Furthermore, the maximum likelihood estimates of the cointegration vector 2 in all the VECM models are essentially -1. This implies the error-correction term is z t x y, or simply the price difference between the electronic t t (BrokerTec) and voice-based (GovPX) trading systems. This is a sensible and intuitive result because in the long-run equilibrium the price difference between the two trading systems should approach zero. In sum, our results show that the price series in each respective market is an I(1) process while their price difference is I(0) or stationary, and there exists a long-term equilibrium relationship between the electronic and voice-based trading systems. Table II reports the results of VECM estimation over the whole sample period. Panel A reports the 13 We did not use common order selection criteria in conducting the ADF test since it does not necessarily optimize the test performance (see, for example, Ng and Perron, 1995 for lag length selection). 16

19 model estimate using midquotes and Panel B reports the result using transaction prices. The results in these two panels show similar patterns. The estimate sheds light on how BrokerTec and GovPX prices adjust to correct the disequilibrium in the preceding period. As shown, 1 estimates are all negative, which are quite small but significant (in all cases except one). By contrast, 2 estimates are all positive, much larger than 1 and are highly significant. This renders the following interpretations. Suppose prices in the two markets are in disequilibrium, say the BrokerTec price is higher (lower) than the GovPX price in the preceding period. The model suggests that the GovPX price will increase (decrease) in the next period, thus correcting the positive (negative) disequilibrium error. For the BrokerTec price, it will decrease (increase) to correct the disequilibrium error, but the magnitude of changes is much smaller. [Insert Table II here] The pattern of estimates suggests that overall the electronic trading system generates more price discovery than the voice-based trading system for all bond groups. In addition, it is more often that the price in the voice-based trading system adjusts to the price in the electronic trading system. This finding implies that voice brokers use the price information on the electronic platform in their pricing. Based on the VECM estimates, one can construct the Hasbrouck and Gonzalo-Granger measures to assess the contribution to price discovery by each trading system. Several interesting observations can be made before constructing these measures. First, as shown in Table II, the innovation variance in the electronic trading system ( 1 ) is higher than that in the voice-based trading system ( 2 ) for both midquote and transaction price series, suggesting more information shocks in the former. Second, the correlation of the innovations between the two markets is small, ranging from 0.1 to 0.3. As a result, Hasbrouck s upper and lower bounds are expected to be fairly tight, and the midpoint is a reasonably good summary measure of the information share. 14 Finally, while results suggest that the electronic trading system has more price discovery, it is important to note that due to the significance of the 1 term, price discovery does not entirely come from this venue. The dynamic relation between the two trading systems can be inferred from the AR coefficients in the (s) s A ij VECM model. Let A ( ) where (s) A ij is the (i,j)-th element of the sth AR coefficient A s in (1) and i, j = 1 (electronic) and 2 (voice) are indicators of the trading systems. The off-diagonal elements (s) A ij, i j, capture how the price change in one system is affected by the preceding price changes in the other system. The relationships between two time series can be classified as uni-directional, bi-directional or no feedback relations. For the uni-directional relation, the coefficient matrix (s) ij A s is either upper or lower ( s) triangular, that is, for i j, all A = 0 and A 0 such that one trading venue affects the other but not ji 14 There is only one case in which is around 0.3 but still the two Hasbrouck bounds are not far apart. 17

20 the other way around. For the bi-directional relation, trading venues affect each other, that is, (s) ij A 0 ( s) ji and A 0. For no feedback relation between two venues, we have (s) ij ( s) ji A = 0 and A 0 for all s. (1) 1 ij The AR coefficients reported in Table II are for the first lag A ( A ). We observe that the diagonal elements (1) A11 and (1) A 22 are negative and significant whereas the off-diagonal elements (1) A12 and (1) A21 are positive and significant. For the rest of the AR coefficients at higher lags, the estimates are naturally different but their sign and significance remain broadly the same and are omitted for brevity. The results suggest that a bi-directional relation exists between the electronic and voice-based trading systems and that the preceding price change in one system affects the other. If the price in one trading system increases, its own price will tend to decrease in the next period ( (1) A ii < 0), or revert to its equilibrium value. Conversely, the price in the other trading system will increase in the next period ( (1) A ij > 0), or follow suit to the increment in the other trading system. B. Estimation of Price Discovery Measures From the VECM parameter estimates, we can obtain price discovery measures for the two trading systems. The lower part of each panel in Table II reports the estimates of the Hasbrouck and GG measures for different Treasury issues. The lower part of Panel A reports the results using dealers quotes. Results show that price discovery mainly takes place in the electronic trading system. The Hasbrouck midpoint measure ranges from for the 10-year note to for the 5-year note. 15 The range of the upper and lower bounds (in parentheses) of the Hasbrouck measure is fairly tight. On the other hand, the GG measure varies from 0.78 for the 10-year note to 0.96 for the 5-year note. Although the GG measure is smaller than the Hasbrouck measure, the general patterns of both measures are similar. The lower part of Panel B shows the price discovery measures estimated from transaction prices. The Hasbrouck midpoint measure ranges from for the 2-year note to for the 10-year note. The upper and lower bounds of the Hasbrouck measure are reported in parentheses. The GG measure is again smaller than the Hasbrouck measure, ranging from for the 2-year note to for the 10-year note. The discrepancy between the information shares estimated from quote and trade data is mainly due to differences in quoting and trading activities between the two trading systems. The overall result again indicates that the contribution to price discovery by the electronic trading system is much larger. To get a sense of the temporal variation of information shares, we estimate these measures for each year. Table III tabulates the Hasbrouck midpoint and GG measures from 2001 to 2005 based on midquotes. These measures are derived from the parameters of the VECM model estimated for each year. For ease of comparison, we replicate the full-period estimates (from Table II) of the information share at 15 Hasbrouck midpoint is the average of upper and lower bounds. 18