How To Study Price Discovery In Canada

Transcription

1 Price Discovery without Trading: Evidence from Limit s* Jonathan Brogaard Terrence Hendershott Ryan Riordan First Draft: November 2014 Current Draft: September 2015 Abstract: Adverse selection in financial markets is traditionally measured by the correlation between the direction of market order trading and price movements. We show this relationship has weakened dramatically with limit orders playing a larger role in price discovery and with high-frequency traders (s) limit orders playing the largest role. s are responsible for 60 80% of price discovery, primarily through their limit orders. s limit orders have 50% larger price impact than s limit orders, and s submit limit orders 50% more frequently. s react more to activity by s than the reverse. s react more to orders both within and across stock exchanges. * We thank Helen Hogarth, Victoria Pinnington, and IIROC for providing data and comments. All errors are our own. We thank participants at the 2015 Cambridge Microstructure Theory and Application Workshop, Baruch College, Stockholm Business School, Hong Kong University, Chinese University of Hong Kong, University of Mannheim, Goethe University, Australia National University, and UC Santa Cruz Workshop for helpful comments Contact: Jonathan Brogaard, Foster School of Business, University of Washington, ( ) brogaard@uw.edu, (Tel) ; Terrence Hendershott, University of California Berkeley, Haas School of Business, ( ) hender@haas.berkeley.edu (Tel) ; and Ryan Riordan, Queen s School of Business, Queen s University ( ) ryan.riordan@queensu.ca (Tel)

2 I. Introduction This paper examines the nature of price discovery for stocks in modern fragmented markets. The fraction of information incorporated into prices through liquidity-demanding marketable orders has fallen dramatically. Using the standard Hasbrouck (1991b) approach Figure 1 shows this for a well-known stock, Royal Bank of Canada. 1 Between 2007 and 2011 trading in RBC fragments and the fraction of price discovery correlated with trading falls from close to 40 percent to roughly 10 percent. 2 INSERT FIGURE 1 ABOUT HERE Around the world traditional stock exchanges now face a range of competitors for order flow. In the U.S. exchanges compete with each other, electronic communications networks (ECNs), dark markets, and other execution venues. In Europe MiFID has led to a dramatic fragmentation in order flow. As shown in Figure 1 in Canada the previously dominant Toronto Stock Exchange lost market share to Chi-X, Alpha, and other smaller trading venues. Using Canadian regulatory data we study the contribution of high-frequency traders (s) and s trades and limit orders to price discovery overall and in each of the three largest stock exchanges in Canada. 3 While individual trades contribute more to price discovery than individual limit orders, trades represent roughly five percent of orders. Limit order placements and cancellations at the best prices represent almost 50 percent of orders. s limit orders are 1 This analysis uses the Thomson Reuters DataScope Tick History archive data provided by SIRCA. 2 See Hendershott, Jones, and Menkveld (2011) for related evidence on the changing nature of price discovery for New York Stock Exchange stocks and Riordan and Storkenmaier (2012) for Deutsche Borse stocks. Malinova, Park, and Riordan (2013) show algorithmic trading increasing on the primary market, the Toronto Stock Exchange, from 2006 through 2009, with a sharp increase in the second half of The sharp increase occurred as the new competitors, Chi-X and Alpha, entered with fast trading technology and lower fees (see Section III for details). 3 We use stock exchange and market interchangeably as Chi-X is technically an alternative trading system in Canada. Our sample period follows the Toronto Stock Exchange s (TSX) significant loss of market share to new entrants. In 2008 Chi-X and Alpha commenced trading and competed with the TSX by offering lower fees and new trading technology. The TSX improved its technology and reduced its fees in response. s are the algorithmic traders most sensitive to exchange technology and lower trading fees are thought to have increased their activities dramatically during this period. While the regulator data does not include the dates surrounding Chi-X s entry, we examine price discovery in the post-entry period where trade-related price discovery is greatly diminished and s are very active. 2

3 individually more informative and are roughly twice as prevalent as s limit orders. Therefore, s limit orders are the primary channel through which price discovery occurs. We examine price discovery using three approaches: i) a standard vector autoregression (VAR; Hasbrouck, 1991a) that incorporates limit order activity, ii) the frequency of order types that move prices (Biais, Hillion, and Spatt, 1995) by s and s; and iii) information shares (Hasbrouck, 1995) based on best limit orders by s and s. The VAR shows that s predominant role in price discovery stems from their new orders being more informative and from their greater activity at the best prices. The frequency of order type analysis provides an intuitive parametric confirmation of these results. For example, only eight percent of price changes occur following trades. The VAR and frequency results also show that s react more to trading by s than the reverse. The information shares show that s limit orders contribute more to price discovery than s limit orders. Because the information shares are estimated at a 1 second frequency, they also show that s role in price discovery is not primarily due to tiny differences in speed. Finally, we examine s activity and price discovery on each exchange and across exchanges. As in the market-wide results, s are the predominant channel of price discovery on each exchange through their limit orders. s react to information across exchanges more than s and faster than s. Significant price discovery occurs within the same second across exchanges with s limit orders being primarily responsible. II. Literature Review This paper contributes directly to the literature on high frequency trading. s are generally defined as technologically sophisticated short-horizon trades who hold low inventory, use many orders, and make frequent small trades throughout the day. Jones (2013), Biais and Woolley (2011), and Biais and Foucault (2014) provide an overview of this literature. Brogaard, Hendershott, and Riordan (2014) and others use data from NASDAQ that identifies 3

4 s trades and show these trades impound information into prices. 4 Our data allows us to identify the s ourselves without needing the exchanges to do so. We also show s trades are informative, but we find that s limit orders play an even greater role in price discovery. A number of theoretical papers examine how fast traders like s can adversely select slower traders. For example, Foucault, Hombert, and Rosu (2014), Biais, Foucault, and Moinas (2014), and Budish, Cramton, and Shim (2014) examine how some traders trading faster on public signals increases information asymmetry. Our results lend support to these concerns, but suggest they play a relatively small role in overall price discovery. Jovanovic and Menkveld (2015) show an equilibrium response to s ability to react faster and better process information is for other investors to not use limit orders. This leads to s incorporating most information through their limits order, which reduces measured information asymmetry associated with trading. However, s placing fewer limit orders can lead to excess intermediation by s. Empirically, technological changes have been used to examine how speed and fast trading impact markets. Hendershott, Jones, and Menkveld (2011) and Boehmer, Fong, and Wu (2012) show how algorithmic trading improves liquidity on the New York Stock Exchange and internationally. 5 These technological changes have enabled new trading venues to compete with existing exchanges. A growing body of literature analyzes the effects of market fragmentation on liquidity and price discovery. 6 4 Using the same NASDAQ Carrion (2013) also show that s trades are more informative than s trades. Hirschey (2012) uses more detailed data from NASDAQ that identifies trading by individual s and finds that s aggressive trades predicts subsequent s liquidity demand. 5 Gai, Yao, and Ye (2014) find that technological improvements at the NASDAQ are associated with decreasing depth. Menkveld and Zoican (2014) show that a new trading system introduced at NASDAQ OMX in 2010 increases spreads due to faster s picking off slower s. Brogaard, Hagströmer, Norden, and Riordan (2014) use a colocation upgrade at NASDAQ OMX Stockholm to find that s supplying liquidity are able to utilize the upgrade to improve liquidity. Malinova, Park, and Riordan (2013) use the introduction of a message fee on the Toronto Stock Exchange to show that s liquidity supplying orders are positively related to liquidity. Menkveld (2013) shows how the entry of one liquidity-supplying improves liquidity in Dutch stocks. 6 Papers examine competition in the U.S. from ECNs (see, for example, Barclay, Christie, Harris, Kandel, and van Ness (1999), Weston (2000), Huang (2002), and Barclay, Hendershott, and McCormick (2003)). In Europe fragmentation and competition followed the introduction of MiFID (see Hengelbrock and Theissen (2009), Degryse, de Jong, and van Kervel (2011), and others). 4

5 The question of how to ensure market integration when trading fragments remains an important research question. Battalio, Hatch, and Jennings (2004) analyze quote and execution quality of multiple listed U.S. equity options and conclude that competition between trading venues, improved technology, and the threat of increased regulation can integrate platforms without a formal linkage. 7 Stoll (2001) hypothesizes an informal linking of fragmented markets without formal regulated linkages. He envisions integration through the routing decisions of brokers and dealers. While the basic thrust of Stoll s analysis is prescient, the role of new types of market participants such as s was not explicitly anticipated. III. Data and Institutional Details Data is provided by the Investment Industry Regulatory Organization of Canada (IIROC). The data include every order submitted on recognized equity markets in Canada. The data include masked market IDs, masked participant IDs, security IDs, date and timestamp to the millisecond, order type, order volume, and a buy/sell indicator. 8 Importantly the data identifies activities across exchanges as the anonymous IDs remain constant across days, securities, and markets. a. Trading Landscape Canada has a number of equity markets upon which trading is organized. We identify 9 in total and present summary statistics on the 3 largest exchanges. 9 Trading on these 3 exchanges makes up more than 98% of the total trading volume. 7 In contrast, Foucault and Menkveld (2008) find that the lack of formal market linkages discourages liquidity supply. 8 The data are structured similar to the NASDAQ ITCH. They contain every message sent by each participant to the exchange. The messages include the initial order, cancels, and amendments to the order. As in the U.S. there are a number of different order types, such as hidden orders and immediate or cancel (IOC) orders, which are flagged in the data. We exclude hidden limit orders from the order book construction. IOC orders are included as an order and cancel if they are not executed, and a trade if they are filled. IIROC receive data with homogenized fields from each exchange in a format that allows for cross-platform integration. Specifically, exchange data must follow the Financial Information Exchange (FIX) protocol ( Any deviation from the FIX implementation must be approved by IIROC with a regulatory feed compliant solution. The data are timestamped by each exchange. The exchanges are required to synchronize their clocks with IIROC, which follows the National Research Council Cesium Clock. 9 See for an overview of marketplaces as of June 1st,

6 Markets in Canada are organized similarly to United States markets with electronic limit order books observing price-time-display priority. s in Canada are protected via order protections rules (OPR). 10 OPR apply to marketplaces that provide automated functionality. Automated functionality includes the ability to immediately and automatically accept incoming orders, execute those orders and cancel any unexecuted portion of those orders marked as immediate-or-cancel as well as automatically display and update the status of each participants orders. OPR only apply to visible orders and the visible parts of orders. The OPR requires marketplaces to implement rules to prevent trade-throughs or executing before, immediately accessible, visible, better-priced limit orders. In contrast to Regulation NMS in the United States Canadian Markets implement a full depth-of-book protection. This means that before an order is executed at marketplaces must ensure that all protected orders that are visible at better price levels have been executed. Canadian regulations also impose best execution obligations on brokers. These regulations require dealers and advisors to execute a trade on the most advantageous terms reasonably available under the circumstances when acting for a client. See Korajczyk and Murphy (2014) for additional institutional details. The Canadian market has seen a dramatic increase in competition for investor order-flow since In May of 2007 a consortium of Canada s largest banks announced a trading platform designed to compete with the TSX, called Alpha Trading Systems. Shortly thereafter in December of 2007 Chi-X announced their intention to commence trading in selected Canadian stocks on February 20 th, Both Alpha and Chi-X planned to offer new trading technology and lower trading fees than the incumbent TSX. In response TSX rolled out new trading technology to all TSX-listed stocks in a staggered fashion (TSX Quantum) in the first half of

7 Chi-X commenced trading in Canada as planned and Alpha commenced trading in December of Both Chi-X and Alpha introduced new trading technologies, innovative order types and new fee models. The increased competition lead to a dramatic decrease in the TSX market share falling from close to 100% in early 2008 to less than 65% in During the TSX transaction fees fell by 80% and they implemented an electronic liquidity supplier fee rebate program designed to attract US-Based s. 12 In 2012 TSX s parent company, the Maple Group, purchased Alpha and now operates Alpha as a separate exchange within the TMX group of exchanges. b. Sample We restrict our sample to the 15 securities that are part of the TSX 60, the primary Canadian equity index, at the end of 2014 that are not cross-listed in the United States. The other 45 stocks in the index are cross-listed. We exclude cross listed stocks as we are unable to observe as precisely measure trading occurring off Canadian exchanges. In addition, crosslisted stocks may have different market making properties (Bacidore and Sofianos, 2002). Table 1 reports descriptive statistics for the stocks in our sample: market capitalization, share price, trade size, number of trades, number of shares traded, dollar volume traded, national best bid and offer (NBBO) quoted half-spread, %, % demand, % supply and the standard deviation of prices. Market capitalization is the January 31, 2012 market capitalization collected from Datastream, all other variables are reported as stock day averages for the sample period, from 10/15/2012 to 06/28/2013, using the IIROC data. Table 1 includes activities from all exchanges, whereas the remaining tables only include observations from the three largest exchanges. 11 See page 17 of the TSX Annual Report 12 See for more details on fragmentation in Canada equity markets. 7

8 INSERT TABLE 1 ABOUT HERE The firms in our sample are relatively large. Market capitalization ranges from $1.95 Billion CAD to more than $28 Billion CAD. Share prices are relatively similar, between $20 and $72, with the exception of Bombardier with a price of $4.00. The stocks in our sample are actively traded with between $11.84 (million) and $70.97 (million) traded per stock day. The stocks in our sample are relatively liquid with quoted half-spreads between 1.38 and basis points. c. Identification We identify using the following criteria, using a similar methodology to Kirilenko et al. (2011) using CFTC data: (a) Make up more than.25% of trading volume; (b) Have an end of day inventory of less than 20% of their trading volume; and (c) Never hold more than 30% of their trading volume at one time within the trading day. In total we identify 61 IDs from a total 1706 IDs in the Canadian market. The average is more active in terms of quotes, trades, shares and volume traded, and has a higher order to trade ratio. Overall s hold less inventory throughout and at the end of the trading day. s hold considerably less inventory than their trading would imply. Their end-of-day inventory is roughly 10% of their traded volume versus 66% for s. Table 1 shows participation varies across the sample of stocks ranging from 12.1% to 30.1%. liquidity demand and supply participation are also not evenly distributed ranging from 11.4% and 23.0% for demand and 9.5% and 40.6% for supply. s generally supply more liquidity than they demand in our sample. Table 2 reports statistics on and participants. s on average submit 15 times as many orders (4,450 versus 290) and trade six times more often (239 versus 41) than do s. This leads to s having an order to trade ratio more than twice that of the 8

9 s (64 versus 27). s average trade size is less than one sixth the size of s trades ($5,664 versus $37,532). INSERT TABLE 2 ABOUT HERE We also compile inventory statistics to better understand how and manage their inventory. 13 We report the absolute value of the end-of-day inventory / dollar volume traded for that stock on that day, the absolute value of the maximum inventory observed / dollar volume traded for that stock on that day, and the number of days where the absolute value of the end-of-day inventory / dollar volume traded is below 3%. For all inventory statistics s hold less inventory, relative to their trading volume at the end-of-day 10.68% versus 69.82% and at their intraday maximum 17.99% versus 79.63%. s inventories are consistent with short-run speculators closely managing risk. To begin the examination of overall activities by s and s Table 3 reports the frequency of orders broken down by participant type and order type/aggressiveness. Limit orders aggressiveness is determined relative to the NBBO: marketable limit order (trade), limit order/cancel at the NBBO, limit order/cancel 1 tick behind the NBBO, and limit order/cancel > 1 tick behind the NBBO. On the average stock day 84,518 orders, including trades and cancelations, are placed. 14 Table 2 shows individual s being much more active, but the larger number of s results in s comprising about 53% of aggregate order activity. INSERT TABLE 3 ABOUT HERE 13 The inventory measure assumes each ID begins each day with a zero position. 14 amendments that increase quantity or improve the price, e.g., lower the price on a buy order, are considered new orders. order amendments, i.e., lower quantity or worse price, are counted as cancellations. There are 904 order amendments per stock day. 9

10 s trades are roughly 0.9% of all orders whereas s trades make up 3%. s new limit orders at the NBBO are the most numerous events making up 15.6% of all orders. s cancellations at the best prices are the next most numerous orders at 11.9%. s order submissions and cancellations are less numerous but exhibit a similar ratio of 10.0% and 6.2%. Non-s are more likely to submit and cancel less aggressive orders (> 1 tick behind the NBBO) than they are to engage in any other activity. IV. Market-wide Price Discovery Figure 1 demonstrates how price discovery and adverse selection change due to trading change over time on Canadian stock exchanges. The IIROC sample period follows the dramatic decline in trade-related adverse selection. To investigate how the different orders impact the evolution of price discovery shown in Figure 1 we use Fleming, Mizrach, and Nguyen s (2015) extension of Hasbrouck (1991a). The approaches incorporate trades and limit order activity, both submissions and cancelations, at various price levels into the standard price discovery VAR. We extend it further by separating s and s activity. The VAR model is: 5 1 1,1 r t = α i r t i + β i i=1 5 5 i=0 X 1 2 2,1 t = α i r t i + β i i=1 5 5 i=1 X 2 3 3,1 t = α i r t i + β i 5 1 X t i 1 X t i 1 X t i 5 + β i 1,2 i=0 5 + β i 2,2 i=1 5 + β i 3,2 2 X t i 2 X t i 2 X t i β i 1,14 i= β i 2,14 i= β i 3,14 X 14 1 t i + μ t X 14 2 t i + μ t X 14 3 t i + μ t i=1 i=1 i=1 i=1 = 5 X ,1 t = α i r t i + β i 5 1 X t i 5 + β i 15,2 2 X t i β i 15,14 X t i + μ t i=1 i=1 i=1 i=1 10

11 where α captures the coefficient on the midpoint return series, r, lagged 1 5 periods; β captures the coefficient on the 14 limit order and trade variables, X 1 X 14. Note the return equation includes the contemporaneous and five lag values of the limit order and trade variables, whereas the remaining VAR regressions only include the five lagged values. The VAR is in event time, t, with each order being an observation. X contains 14 variables:,,, 1 tick from NBBO, 1 tick from NBBO, > 1 tick from NBBO, > 1 tick from NBBO,,,, 1 tick from NBBO, 1 tick from NBBO, > 1 tick from NBBO, and > 1 tick from NBBO. named variables captures activity by firms, and named variables capture activity by firms. takes the value +1 for buy initiated trades, -1 for sell initiated trades, and 0 otherwise. takes +1 for bids placed at the NBB, -1 for offers placed at the NBO, and 0 otherwise. 1 tick from NBBO take +1 for bids placed at one cent from the NBB, -1 for offers placed at one cent from the NBO, and 0 otherwise. > 1 tick from NBBO take +1 for bids placed greater than one cent from the NBB, -1 for offers placed at greater than one cent from the NBO, and 0 otherwise. For cancels the analogous definition applies with the sign such that cancels at the bid take the value +1 and cancels at the offer take the value -1. We assume the trading process restarts each day, resetting all lagged values to zero. The observations include all displayed orders between 9:45 a.m. EST and 3:45 p.m. EST. To be included a stock-day must have at least 20 orders in each variable on each exchange. This eliminates 1,302 stock-days. The IRF is orthogonalized and order independent and reports the forecasted midpoint return, in basis points, after a +1 (buy event for orders and trades, sell event for cancels). The innovation is cumulative over 20 events. The model is calculated for each stock day and the average IRF reported. For and. a *, ** next to the coefficient represents that the coefficient differs from zero and is statistically significant at the 5 and 1% 11

12 level, respectively using standard errors clustered by stock and by day. In the Difference column *, ** next to the coefficient represents that the and coefficients differs from each other with statistical significance at the 5 and 1% level, respectively using standard errors clustered by stock and by day. To obtain the impulse response function we invert the VAR into its vector moving average representation. We calculate the impulse response functions (IRFs) for an unexpected buy order and measure its impact on returns. The IRFs measure all types of orders price impact, often referred to as the contribution to price discovery or information content. The VAR is estimated for each stock each day. The averages of these stock-day IRF estimates are reported in basis points. Throughout the paper statistical significance is clustered by stock and day to control for contemporaneous correlation across stocks and autocorrelation within stocks as in Thompson (2011). INSERT TABLE 4 ABOUT HERE Table 4 shows that an trade moves the efficient price basis points. Despite Table 2 showing that s trades are less than one sixth the size of s trades, trades move price by less than half as much, basis points. s new limit orders at the NBBO move price basis points and s move price basis points. The price impact of marketable limit orders is smaller than that of trades, but Table 3 shows there are five times as many marketable limit orders as trades. In addition, s limit orders are almost twice as numerous as s orders. Thus, limit order submissions primarily by s are the predominant source of innovations in the efficient price. ations of limit orders at the NBBO are also information with s and s cancels both moving price by about 0.17 basis points. This price impact is smaller than that of new limit orders at the NBBO. Limit order submissions and cancels at one or more price levels 12

13 away from the NBBO have much smaller price impacts that are often not statistically significantly different from zero. For ease of exposition and because orders outside the NBBO have little price impact the remainder of the paper focuses only on trades and orders/cancels at the NBBO. The VAR estimates used in Table 4 also generate IRFs for how innovations in buy and sell orders affect the direction of subsequent buy and sell orders within and across order type and participant type. These IRFs provide evidence on how s and s react to different market events with different order types. As in Table 4, Table 5 reports the average of stock-day IRF estimates. INSERT TABLE 5 ABOUT HERE The rows in Table 5 correspond to the order variable being shocked by one unit. The columns denote the responses of the subsequent order variables. Therefore, the trade row provides IRF estimates for how all the and order types response to an buy trade. The IRF for trade following an trade shows that more buy trades than sell trades follow an buy trade. Such positive autocorrelation can arise from s splitting their orders, from s copying other s, or from s reacting to common information. Similarly, the IRF for NBBO order following an trade shows that more buy limit orders than sell limit orders follow an buy trade. The IRF for NBBO cancel following an trade shows that more sell limit orders are cancelled than buy limit orders are cancelled following an buy trade. These two limit order results are consistent with models where trades have a price impact for either informational or inventory reasons. The price impact occurs through new orders in the same direction improving the NBBO and orders being cancelled in the opposite direction. 13

14 All the coefficients in Table 5 indicate positive auto and cross correlations in order direction. Both s and s respond more to activity by the same participant type than the other participant type, i.e., s respond to more s than to s and s respond more to s than to s. Summing the IRFs in each column provides a measure of how responsive that order type is to prior market activity. By this measure s respond more to activity than s, primarily through s new limit orders. Furthermore, s respond more to s than s respond to s. Overall, Table 5 shows s monitoring market activities and reacting to them. This is consistent with s deriving information from market data. The VAR/IRF results reported in Tables 4-5 model market dynamics in a straightforward linear way. O Hara (2015) argues that traders algorithms make the linear VAR structure inadequate in the modern market environment. However, absent linear theoretical models it is difficult to formulate a better specification. Therefore, we follow Biais, Hillion, and Spatt s (1995) straightforward parametric approach of simply examining the frequencies of different sequences of orders. Panel A of Table 6 presents stock-day averages of frequencies for trades, new orders, and cancelled orders. Similar to Hendershott and Riordan (2013) we separately analyze s and s. The frequencies represent the conditional probability of each order type (the column) conditional on the prior order type (the row). Thus, the frequencies within each row sum to 100 percent. Below the column frequencies are the unconditional probabilities of each order. Frequencies for buy and sell orders are symmetric in that buys following buys are similar to sells following sells and sells following buys is similar to buys following sells. Therefore, to reduce dimensionality the rows in Table 6 aggregate buy and sell orders. The columns in Table 6 are labeled same-side and opposite-side representing whether or not the second order is in the same or opposite direction as the preceding order. INSERT TABLE 6 ABOUT HERE 14

15 Consistent with Table 5, Table 6 shows a diagonal effect in that participants orders are likely to be followed by the same type of participant order. Table 5 focuses on the imbalance between buy and sell order occurrences after new orders. Table 6 provides the absolute magnitude of activity for each order type. After s trades 87% of the next orders are also by s with 78% being in the same direction as the trade. After an new limit order or cancel 74% of the next orders are by s. For s trades and orders about 60% of the next orders are by s. Consistent with the autocorrelation in the direction of orders in Table 5, new limit orders in the same direction by the same participant type are more likely than orders in the opposite direction, and cancels in the opposite direction are more likely. This is consistent with lagged price discovery as trades and orders lead to subsequent activity moving prices in the same direction. Panel B of Table 6 examines the subset of order sequences where the second order results in a NBBO change. About 12 percent of the order sequences in Panel A lead to NBBO changes. Panel B shows how order activity and sequences more directly related to price discovery differ from other activity. The unconditional probabilities show how often different orders cause NBBO changes. Consistent with price discovery correlated with trading being small as in Figure 1, only 8.1% of NBBO changes occur due to trades, corresponding to roughly 400 trades per stock-day. Consistent with the s trades and orders having a higher price impact in Table 4, comparing the unconditional frequencies between Panels A and B shows that a higher proportion of s trades and new orders lead to price changes than the proportion for s. The two most frequent conditional series of orders are and n trades followed by an order in the same direction, 45.1% and 38.6%, respectively. Tables 4-6 show that s new limit orders are the primary channel for price discovery. The analyses are in event time which does not account for possibly very small time differences in between when s and s incorporate new information. For example, if a public 15

16 information event, e.g., the S&P 500 futures contract price going, up leads to both s and s buy limit orders with s being slightly faster, event time analysis will attribute all price discovery to s. A standard approach to price discovery in calendar time is Hasbrouck s (1995) information shares. Mostly commonly this is done by different examining quotes in different markets for the same security, e.g., Huang (2000). This can be directly extended to the best quotes by different market participants. The information shares approach decomposes variation in the common efficient price into individual components attributable to specific markets or participants. The information share methodology focuses on innovation in different groups prices (quote midpoints). The information share of a group is measured as that group s contribution to the total variance of the common (random-walk) component. We calculate the price path for each group (/n or Exchange); denote the price vector p t representing the prevailing prices for each group i as p t i = m t + ε t i. The prices are assumed covariance stationary. The common efficient price path is the random walk process, m t = m t 1 u t where E(u t ) = 0, E(u 2 t) = σ u 2, and E(u t u s ) = 0. For t s. The price process vector can be modeled as a VMA: p t = t 1 t 1 2 t 2, where ε is a vector of innovations with a zero mean and a variance matrix of Ω. Ψ represents the polynomial in the lag operator. The information share is: Infoshare j = Ψ j 2 Ω j ΨΩΨ InfoShare j is interpreted as the fraction of price discovery attributable to participant j; the numerator is the variance of the efficient price attributable to participant j; and the denominator is total variance of the efficient price. As discussed in Hasbrouck (1995) when multiple series move at the same time the information share cannot be uniquely attributed to either series. In 16

17 our setting this occurs if both the and prices move at the same time. The information shares are typically estimated at a fixed sampling frequency. Higher sampling frequencies allow for price discovery to be more uniquely attributed to s and s, but also attribute price discovery occurring close together to the faster participant group, presumably s. To balance this tradeoff we estimate information shares at a one second frequency. Table 7 reports the information shares for s and s. As with the VARs in Tables 4-5, information shares are estimated for each stock each day. Table 7 provides the average maximum and minimum information shares. Table 7 reports whether the average minimum information share is statistically significantly greater than the average maximum information share. Testing the minimum against the maximum shows whether the limit order price discovery results in Tables 4-6 are due to s updating the quotes slightly faster than s. INSERT TABLE 7 ABOUT HERE Table 7 shows that the average 60 percent minimum information share for s is greater than the 40 percent maximum value for. This demonstrates that even with conservative timing assumptions limit order activity contributes more to price discovery than activity. The maximum information share attributes all the common innovations in price discovery within the same second to s and shows s providing 82 percent of price discovery. Thus, 22 percent of the price discovery occurs within the same second for s and s. 17

18 V. Activity and Price Discovery within and across Markets The market-wide price discovery analysis in Tables 4-7 shows that s limit orders are the primary channel through which price discovery occurs. Figure 1 shows that limit order activity becomes much more important after trading fragments away from the traditional primary exchange. This section studies whether s role in price discovery is concentrated on the new exchanges or is similar across the three largest exchanges. We extend Tables 4-7 to examine how s and respond to trading and orders within and across markets. Table 8 reports overall and by exchange statistics for each of the three exchanges: shares traded, dollar volume traded, % of dollar volume traded, % Demand, % supply, and %, the quoted half-spread, % of time at both NBBO, % of time at either NBB or NBO, % of time at neither NBBO, and % of time with no bid or offer. 15 Exchange 2 is the dominant exchange in terms of trading volume making up more than 63.0% in our sample stocks. Exchange 2 also quotes the lowest spreads, 4.06 basis points versus 8.29 and 7.83, and quotes at the NBBO more, 86.8% versus 38.5% and 40.3%. The fraction of trading by s varies across exchanges. s make up roughly 18.6% of the trading volume on exchange 2 and 28.8% and 29.9% on exchanges 1 and 3. s supply (demand) 45.3% (14.4%) of liquidity on exchange 3 and only 34.6% (23.1%) and 17.6% (19.6%) on exchanges 1 and 2. INSERT TABLE 8 ABOUT HERE Table 9 decomposes the order activity by participant type and order type relative to the NBBO in Table 4 for each exchange. Most activity is concentrated at the best bid and offer (BBO) across all three exchanges. Relative activity varies some across exchanges. are most active on exchange 3 and least active on the highest volume exchange, exchange 2. Non-s submit and cancel orders more behind the NBBO than they submit and cancel at the NBBO on 15 The overall numbers in Table 8 are slightly less than Table 1 because Table 1 includes all exchanges while Table 8 only includes the three largest exchanges. 18

19 all three exchanges. This is striking on exchange 2 where s are almost three times as likely to submit an order > 1 tick behind the NBBO than are s. Non- are more likely to submit an order more than 1 tick behind the best available prices than they are at the best and 1 tick behind together. Overall, while and activity differs across exchanges, the basic qualitative patterns of where activity occurs in Table 4 hold on each of the individual exchanges. INSERT TABLE 9 ABOUT HERE Table 10 extends the VAR in Tables 4-5 to allow order activity by s and s on the different exchanges to have different impacts. This is done by indexing all trade and order variables by exchange. The return variable continues to be the NBBO return. Thus, there are 7 order variables on each of the three exchanges for both s and s, making the system have 43 equations and variables. As in Table 4, Table 10 reports the IRFs of returns on shocks to the different order variables. As in Table 5 we only report the subset of coefficients capturing activity at the top of the limit order book: trades, orders submissions at the NBBO, and cancels at the NBBO. As before we estimate the VAR for each stock each day and report stock-day averages. INSERT TABLE 10 ABOUT HERE The return IRFs in Table 10 show that the basic patterns in Table 4 hold across all three exchanges. s trades and new orders have larger price impact than s on all three exchanges. The price impacts of cancels are similar on exchanges 1 and 2, but on exchange 3 s cancels have a larger price impact. The price impact of trades is noticeably lower on exchange 1. Among the largest marketplaces in Canada some cater to specific clientele, such as 19

20 retail traders. The lower price impacts on exchange 1 could be due to a more retail-focused clientele and the associated weaker trade-based price discovery of this group (Jones and Lipson 2005). The return IRFs in Table 10 suggest that the overall pattern of price discovery shown in Figure 1 is not directly driven by the new exchanges, as it appears consistent across exchanges. Table 11 provides the trade and order IRFs from the VAR used in Table 10. Thus, Table 11 relates to Table 10 in a similar way that Table 5 relates to Table 4. To avoid presenting all the possible cross exchange IRFs we group together IRFs for same and other exchanges. The same exchange IRFs are the across exchange average response of each variable to the variables on that exchange. For example, the same exchange trade response to an trade is the average of the IRFs for an trade on exchange 1 response to an trade on exchange 1, an trade on exchange 2 response to an trade on exchange 2, and an trade on exchange 3 response to an trade on exchange 3. The other exchange IRFs is the across exchange average response of each variable to the variables on other exchanges. For example, the other exchange trade response to an trade is the average of the IRFs for an trade on exchange 1 response to an trade on exchanges 2 and 3, an trade on exchange 2 response to an trade on exchanges 1 and 3, and an trade on exchange 3 response to an trade on exchanges 1 and 2. INSERT TABLE 11 ABOUT HERE The IRFs in Table 11 provide insight into whether the autocorrelations and cross-correlation patterns in Table 5 are driven by s and s responding to trades and order within each exchange or whether activity is integrated across exchanges. As in Table 5, all the coefficients in Table 11 indicate positive auto and cross correlations in order direction. In addition, Table 11 shows this is true both within and across exchanges. The other results from Table 5 hold both within and across exchanges: both s and s respond more to 20

21 activity by the same participant type than the other participant type; s respond more to activity than s, primarily through s new limit orders; s respond more to s than s respond to s. Table 11 shows that these responses are larger within exchange than on any other individual exchange. However, because the other exchange IRFs are the averages of twice as many IRFs as the within exchange IRFs, the aggregate other exchange responses are almost as large as the same exchange response. Overall, Table 11 suggests that s monitor and react to cross market activities more strongly than s. 16 Table 12 extends the aggregate information share price discovery analysis in Table 7 to incorporate prices from the different exchanges. Panel A calculates information shares by exchange. All exchanges contribute to price discovery. While exchange 1 has higher information shares, its minimum information is only The wide gap between the maximum and minimum exchange information shares suggests that price discovery is well integrated across markets as significant common price discovery occurs across exchanges within one second. INSERT TABLE 12 ABOUT HERE For each exchange separately Panel B of Table 12 calculates / information shares as in Table 7. On all exchanges the minimum information share for is greater the maximum information share than for and these differences are statistically significant. dominates price discovery more on the smaller exchanges where the minimum is more than three times the maximum. Not surprisingly, Table 8 shows that s supply liquidity more often on these smaller exchanges. Overall, while the new exchanges attract 16 Similar to Table 11 s decomposition of Table 5 s results, we decompose the frequencies of different sequences of orders from Panel A of Table 6 into sequences where both orders are on the same exchange, and sequences where the two order are on different exchanges. We report the results in the Internet Appendix. The results are consistent with the IRF by exchange. That is, more than 50 percent of orders sequences occur within exchange. 21

22 more trading and s provide more price discovery, the primary role of s limit orders in price discovery is not limited to the new, smaller exchanges. VI. Information Transmission Across Markets Table 12 shows that price discovery is integrated across markets and that s play a dominant role in price discovery on each market. Table 11 shows that s have stronger crossexchange activity than s. To link these results we next examine more directly if s cross-exchange activity leads to integrating price discovery across exchanges. Table 6 examines sequences where the NBBO changes. Table 13 extends this by examining order sequences where exchanges 2 and 3 start with the same best bid or offer price and both exchanges revise their bid/offer in the same direction. We focus on exchanges 2 and 3 in this analysis as these are the two largest exchanges by volume. 17 The initial order that causes the two exchanges to go from the same price to different prices is referred to as the first order and the subsequent order on the second exchange which causes its price to move is referred to as the second order. Table 13 provides insight into the sequence of orders that result in common price discovery across the two exchanges. Unlike Table 6 the first order does not need to immediately precede the second order. INSERT TABLE 13 ABOUT HERE Panel A of Table 13 reports the frequencies of order sequences where the price changes on the two exchanges do not occur simultaneously (not in the same millisecond). Panel B reports the subset of these sequences where a trader with the same ID is responsible for both the first and second order. Panel C examines the order sequences where both exchanges prices change in the same millisecond. Panel D provides the subset of these simultaneous price changes where 17 See the Internet Appendix for a similar analysis for exchanges 1 and 2. 22

23 the same trader ID is responsible for both the first and second order. Because Table 13 studies when both exchanges prices move in the same direction, not all order sequences are possible. For example, after a trade consumes all the liquidity available at the best price on one exchange the price on the same side of the market can only change in the same direction on the other exchange due to a trade or an order cancelation. Because there are fewer sequence possibilities Table 13 reports joint frequencies/probabilities as opposed to the conditional frequencies in Table 6. Consistent with our prior results showing the importance of limit orders in price discovery, Panel A of Table 13 shows that s trades and limit orders/cancels initiate the sequence of price changes 68 percent of the time. Initiating the price change sequence is naturally interpreted as the most significant component of price discovery across markets. The most frequent sequence is an order cancel followed by an order with 27.1%, followed by order cancel and then an order cancel with 17.7%. s also provide the second orders more often than s suggesting that they are responsible for both within and across market price discovery. Non-s are more likely to submit the second order than they are the first perhaps because they are slower to update their quotes than are s when both react to the common information. In this way order cancels may be related to slow adjustments to new information. Panels B-D of Table 13, which examine sequences by the same trader and simultaneous sequences, are consistent with Panel A. s open more of the price change sequences, primarily through their new limit order and order cancelations. Table 14 examines the duration of the simultaneous price change sequences in Table 13. Table 14 reports the mean number of seconds between the first and second orders. For the sequences in Panel A of Table 13, Panel A of Table 14 reports the length of time of all price sequences conditional on the trader type and order type that open and close. s close faster when the price sequence is opened by an. This is true overall and for each sequence pair, except cancel-cancel. 23

24 INSERT TABLE 14 ABOUT HERE Panel B examines the price sequences where both the open and close orders are from the same trader ID. Average durations for s are shorter for s. s generally close price sequences faster than s. When the same market participant is responsible on both markets the time between open and close is noticeably shorter. s are faster for trades and new orders. Somewhat surprisingly this is not the case for order cancellations, showing s can quickly cancel orders on multiple exchanges. The results in Tables 13 and 14 on price changes sequences across exchanges show s playing an important role in cross exchange price discovery. This is primarily through their new limit orders and cancelations of existing limit orders. Non-s cancelations are important in completing the price change sequence on the second exchange. This is consistent with s being slower than s. However, s access multiple exchanges at speeds within tens or hundreds milliseconds of the time s access multiple exchanges. Whether or not these small differential durations between s and s orders are meaningful is an important unanswered question. Given the much greater numbers of s limit orders and their larger importance in price discovery, the importance of speed differences appears less economically significant for trades. VII. Conclusion Adverse selection is traditionally measured by the correlation between trade direction and returns. This relationship has weakened as stock exchanges have upgraded their technology and trading has fragmented. Canadian regulatory data shows that s are responsible for between 60% and 80% of price discovery, primarily through their limit orders. s orders are both more informed and more numerous than s limit orders. s also react more to 24

25 activity by s than the reverse. s react more to orders both within and across stock exchanges [and ATSs]. Autocorrelated s trading and s reacting to market activity is consistent with s anticipating s trading. Critics of s and the modern market structure characterize this as s systematically front-running large orders that are sliced into smaller components. Because s react similarly to both s and s orders, our results provide little support for s specifically targeting s. However, s reacting to market activity in general can move prices against large s orders before completion. We find that s primarily react to market activity through their limit orders so s anticipation of s is better described as s market making activity fully incorporating public information than as s behaving as predatory traders and quasi-frontrunning s. The significant contemporaneous price discovery across markets and s reacting more to orders across multiple exchanges is consistent with s integrating markets. We find little evidence that s use superior technology to pick off slower s limit orders, often referred to as latency arbitrage. Our findings are most consistent with s dominating limit order book activity leading to more efficient prices and a virtually integrated limit order book. However, it is possible that s speed and information processing abilities discourage s from submitting limit orders. This could lead to less direct investor-to-investor trading and excess intermediation. 25

26 References Bacidore, J. M., and Sofianos, G. (2002). Liquidity provision and specialist trading in NYSElisted US stocks. Journal of Financial Economics, 63(1), Barclay, M., T. Hendershott, and D. McCormick (2003): Competition among trading venues: Information and trading on electronic communications networks, Journal of Finance, 58(6), Battalio, R., B. Hatch, and R. Jennings (2004): Toward a national market system for U.S. exchange-listed equity options, Journal of Finance, 59(2), Biais, B., T. Foucault, and S. Moinas, 2014, Equilibrium fast trading, Journal of Financial Economics, forthcoming. Biais, B., P. Hillion, and C. Spatt (1995): An empirical analysis of the limit order book and the order flow in the Paris Bourse, Journal of Finance, 50(5), Biais, B., and P. Woolley, 2011, High frequency trading, Working paper Toulouse University, IDEI. Boehmer, Ekkehart, Kingsley Fong, and Julie Wu, 2012, International evidence on algorithmic trading, Discussion paper, EDHEC. Brogaard, J., T. Hendershott, and R. Riordan. (2014). High frequency trading and Price discovery. Review of Financial Studies 27: Degryse, H., F. Jong de, and V. v. Kervel (2011): Equity market fragmentation and liquidity: The impact of MiFID, Working Paper. Fleming, M., B. Mizrach, and G. Nguyen (2015): The Microstructure of a U.S. Treasury ECN: The BrokerTec Platform, working paper. Foucault, T., J. Hombert, and I. Rosu, 2014, News trading and speed, Journal of Finance, forthcoming. Foucault, T., and A. J. Menkveld (2008): Competition for order flow and smart order routing systems, Journal of Finance, 63(1), Garriott, Corey, Pomeranets, A., Slive, J., & Thorn, T. "Fragmentation in Canadian equity Markets." Bank of Canada Review 2013.Autumn (2013): Hasbrouck, J. (1991a): Measuring the information content of stock trades, Journal of Finance, 46(1), Hasbrouck, J. (1991b): The summary informativeness of stock trades: An econometric analysis, Review of Financial Studies, 4(3),

27 Hasbrouck, J. (1995): One security, many markets: Determining the contributions to price discovery, Journal of Finance, 50(4), Hasbrouck, J., and G. Saar, 2011, Low latency trading, Working paper NYU Stern. Hendershott, T., and C. M. Jones (2005): Island goes dark: Transparency, frag- mentation, and regulation, Review of Financial Studies, 18(3), Hendershott, T., C. M. Jones, and A. J. Menkveld (2011): Does algorithmic trading improve liquidity?, Journal of Finance, 66(1), Hendershott, T., and P. C. Moulton (2011): Automation, speed, and stock mar- ket quality: The NYSE s Hybrid, Journal of Financial Markets 14, Hendershott, Terence, and Ryan Riordan, 2013, Algorithmic trading and the market for liquidity, Journal of Financial and Quantitative Analysis 48, Hengelbrock, J., and E. Theissen (2009): Fourteen at one blow: The market entry of Turquoise, Working Paper. Huang, R. D. (2002): The quality of ECN and Nasdaq market maker quotes, Journal of Finance, 57(3), Jovanovic, B., and A. Menkveld, 2015, Middlemen in limit-order markets, Working paper VU Amsterdam. Jones, Charles M., and Marc L. Lipson. "Are retail orders different." unpublished paper, Columbia University and University of Georgia (2004). Kirilenko, A., Kyle, A. S., Samadi, M., & Tuzun, T. (2011). The flash crash: The impact of high frequency trading on an electronic market. Manuscript, U of Maryland. Menkveld, A., 2013, High frequency trading and the new-market makers, Journal of Financial Markets 16(4), O Hara, M. (2015): High Frequency Market Microstructure, Journal of Financial Economic, forthcoming. O Hara, M., and M. Ye (2011): Is market fragmentation harming market quality?, Journal of Financial Economic, 100(3), Riordan and Storkenmaier (2012): Latency, liquidity and price discovery, Journal of Financial Markets 15 (4), Stoll, Hans R. "Market fragmentation." Financial Analysts Journal 57.4 (2001): Thompson, S. B. (2011): Simple formulas for standard errors that cluster by both firm and time, Journal of Financial Economics, 99(1),

28 Name Table 1: Descriptive Statistics. The table reports summary statistics for the 15 stocks used in this study. The 15 stocks are chosen as they are part of the TSX 60 and not cross-listed in the United States. The statistics are calculated using data from 10/15/ /28/2013. Name is the ticker. Mkt. Cap is the average market capitalization from Datastream, in billions of Canadian dollars. Share Price is the average traded stock price. Size is the average trade size. Number of s is the average number of trades, in thousands. Number of Shares d is the average number of shares traded, in thousands. Dollar Volume d is the number of shares traded multiplied by the stock price, in millions of dollars. NBBO Quoted Half-Spread is the calendar time weighted onehalf quoted difference between the national best bid and the national best ask price, in basis points. % is the double-sided dollarvolume % of trades by a high frequency trader (). % Demand is the dollar-volume percent of trades in which an is the liquidity taker. % Supply is the dollar-volume percent of trades in which an is the liquidity provider. Std. of Daily prices is the standard deviation of the end-of-day price for the stock. All statistics except Std. of Daily Returns is the standard deviation of the daily return, in percent. Mkt. Cap ($ billion) Share Price Size Number of s (thousand) Number of Shares d (thousand) Dollar Volume d ($ million) NBBO quoted half-spread (bps) % % Demand % Supply Std. Dev. of Returns (percent) ARX $8.00 $ , $ % 15.0% 28.1% 1.23 ATD.B $10.03 $ $ % 22.9% 12.0% 1.17 BBD.B $6.95 $ , $ % 19.5% 40.6% 2.00 COS $9.90 $ , $ % 14.4% 35.4% 1.27 CTC.A $6.77 $ $ % 19.1% 11.6% 1.51 FM $10.95 $ , $ % 14.2% 22.7% 2.71 FTS $6.96 $ $ % 13.9% 33.2% 0.77 HSE $28.74 $ , $ % 19.0% 29.8% 1.28 L $11.62 $ $ % 11.4% 12.8% 1.56 MRU $1.95 $ $ % 16.6% 15.6% 0.85 NA $12.46 $ $ % 18.7% 24.1% 0.62 POW $12.10 $ , $ % 14.4% 31.6% 0.92 SAP $9.61 $ $ % 23.0% 17.7% 0.97 SNC $6.45 $ $ % 16.4% 12.3% 1.54 WN $9.35 $ $ % 22.2% 9.5% 1.08 Average $10.12 $ , $ % 17.4% 22.5%

29 Table 2: r Type Statistics. The table reports summary statistics for trading, orders, and positions for individual and s. Column reports stock-day-trader averages for traders, column reports stock-day-trader for traders. Number of s is the average number of quotes, quote cancels, and quote amends a trader conducts. Number of s is the average number of trades conducted by a trader. Number of Shares d is the average number of shares traded by a trader. Dollar Volume (DV) d is the average number of shares traded by a trader multiplied by the share price. to Ratio is the number of orders deployed for each trade by a trader. DV d / Total DV d is the dollar volume traded by a trader scaled by the total dollar volume traded on that stockday. Abs(EoD Inv.) / DV d is the absolute value of a trader s end of day dollar-volume inventory scaled by that trader s dollar-volume traded. Abs(Max Intra. Inv.)/DV d is the trader s absolute value of the maximum intraday dollar-volume inventory position scaled by that trader s dollar-volume traded. % of days with Abs(EoD Inv) / DV d < 3% is the percent of stock-day-trader observations with the Abs(EoD Inv.) / DV d is less than 3%. Average DV Size is the average dollar volume size of a trade. Number of Participants is the number of traders on the average stock-day. Number of s (thousand) Number of s Number of Shares d (thousand) Dollar Volume (DV) d ($million) $1.03 $0.35 to ratio Number of s / Total s 2.81% 0.21% DV d / Total DV d 1.51% 0.46% Abs(EoD Inv) / DV d 10.68% 69.82% Abs(Max Intra. Inv.) / DV d 17.99% 79.63% % of days with Abs(EoD Inv)/DV d < 3% 50.21% 6.64% Average DV Size $5, $37, Number of Participants

30 Table 3: Distribution of Activities. The table reports the distribution order aggressiveness by s and s. Panel A considers all orders on the three exchanges. Total # of Observations, Overall is the number of observations on all three exchanges on the average stock-day. The point estimates report the percent of activity by and that are s, s, or s. captures the number of orders at that exchange s NBB or NBO. 1 tick from NBBO captures the number of orders one cent away from that exchange s NBB or NBO. > 1 tick from NBBO captures the number of orders more than one cent away from that exchange s NBB or NBO. For cancels the analogous definition applies. Total # of Observations, is the number of observations on the average stock-day. 0.9% 3.0% 15.3% 10.0% 11.9% 6.2% 1 tick from NBBO 5.1% 2.2% 1 tick from NBBO 6.1% 3.4% > 1 tick from NBBO 6.5% 11.0% > 1 tick from NBBO 7.4% 11.1% Total # of Observations 84,518 30

31 Table 4: Return Impulse Response Function. The table reports stock-day average Impulse Response Functions (IRF) from a Vector-autoregression (VAR) with 15 equations and 5 lags. One for each of the variables listed in the table, for each and, and the midpoint NBBO midpoint return. The VAR is in event time with each order being an observation. takes the value +1 for buy initiated trades, -1 for sell initiated trades, and 0 otherwise. take +1 for bids placed at the NBB, -1 for offers placed at the NBO, and 0 otherwise. 1 tick from NBBO take +1 for bids placed at one cent from the NBB, -1 for offers placed at one cent from the NBO, and 0 otherwise. > 1 tick from NBBO take +1 for bids placed greater than one cent from the NBB, -1 for offers placed at greater than one cent from the NBO, and 0 otherwise. For cancels the analogous definition applies with the sign such that cancels at the bid take the value +1 and cancels at the offer take the value -1. The observations include all displayed orders between 9:45 a.m. EST and 3:45 p.m. EST. To be included a stock-day must have at least 20 orders in each variable. The IRF is orthogonalized and order independent and reports the forecasted midpoint return, in basis points, after a +1 (buy event for orders and trades, sell event for cancels). The innovation is cumulative over 20 events. For and a *, ** next to the coefficient represents that the coefficient differs from zero and is statistical significance at the 5 and 1% level, respectively using standard errors clustered by stock and by day. For the Difference column *, ** next to the coefficient represents that the and coefficients differs from each other with statistical significance at the 5 and 1% level, respectively using standard errors clustered by stock and by day. Difference 0.855** 0.387** 0.468** 0.316** 0.220** 0.096** ** ** tick from NBBO 0.003* 0.016** ** 1 tick from NBBO * > 1 tick from NBBO ** ** > 1 tick from NBBO ** ** ** 31

32 Table 5: Impulse Response Function. The table reports stock-day average order Impulse Response Functions (IRF) from the same Vector-autoregression (VAR) used in Table 5. The rows represent the variable being shocked by one unit. The columns represent the variable being affected. *, ** next to the coefficient represents that the coefficient differs from zero and is statistical significance at the 5 and 1% level, respectively using standard errors clustered by stock and by day. Variable 0.114** 0.245** 0.091** 0.005** 0.026** ** 0.198** 0.079** 0.006** 0.051** 0.005** 0.003** 0.088** ** 0.003** 0.027** 0.023** 0.026** 0.208** 0.022** 0.119** 0.039** 0.020** 0.004** 0.069** 0.016** 0.010** 0.135** 0.060** 0.001** 0.035** 0.032** 0.014** 0.096** 0.114** 32

33 t-1 Table 6: Type Conditional on Past Type. The table reports stock-day average order sequence frequencies. The table reports the frequency in which the order type in the identified in the Row is followed by the order type in the Column. Each Row sums to 100%. Unconditional is the frequency in which the Column variable is observed in the data. The Columns represent probabilities of order following each other in the same direction (e.g. buy order followed by a buy order). Columns represent probabilities of orders following each other in the opposite direction (e.g. buy order followed by a sell order). Panel A includes all observations used in the IRFs in Tables 5 and 6. Panel B only includes observations where the NBBO changes in time t. Panel A: All t 31.1% 31.5% 15.2% 0.1% 3.1% 6.0% 2.1% 4.7% 1.8% 0.6% 2.3% 1.4% 1.4% 32.5% 20.7% 0.2% 6.7% 12.7% 2.5% 10.9% 2.9% 1.0% 4.6% 3.9% 0.9% 24.2% 24.0% 0.4% 16.7% 6.2% 1.5% 7.3% 7.4% 1.6% 6.8% 2.9% 2.2% 20.6% 4.6% 0.2% 3.8% 8.2% 35.9% 9.9% 7.2% 1.0% 4.0% 2.5% 0.6% 16.8% 7.9% 0.4% 5.3% 8.2% 3.6% 22.8% 18.4% 1.8% 8.0% 6.1% 0.3% 9.2% 11.5% 0.2% 9.5% 5.1% 5.2% 29.3% 14.4% 1.3% 9.8% 4.3% Unconditional 1.6% 23.3% 16.5% 0.3% 9.0% 8.7% 5.0% 14.8% 9.1% 1.4% 6.5% 4.0% Panel B: NBBO Changes in time t t-1 t 19.0% 45.1% 19.3% 0.2% 2.4% 4.6% 1.1% 2.4% 0.8% 0.8% 2.3% 1.9% 3.7% 31.4% 18.2% 0.5% 7.2% 11.4% 3.8% 10.8% 2.4% 0.9% 4.3% 5.5% 1.5% 31.0% 18.5% 0.6% 12.1% 4.8% 1.8% 9.7% 11.9% 1.1% 4.2% 2.9% 3.8% 38.6% 2.9% 0.4% 2.5% 4.6% 13.9% 9.7% 17.9% 1.1% 2.4% 2.2% 1.7% 21.7% 2.1% 1.3% 7.5% 7.8% 3.4% 21.6% 15.6% 1.4% 7.8% 8.2% 0.5% 17.3% 9.0% 0.5% 9.9% 4.3% 2.8% 28.0% 15.9% 1.0% 6.8% 3.9% Unconditional 2.7% 28.2% 12.5% 0.6% 8.5% 7.1% 3.7% 15.0% 10.8% 1.1% 5.2% 4.7% 33

34 Table 7: Information Shares. The table reports the average stock-day Hasbrouck (1995) minimum and maximum information shares for and orders/quotes. * and ** next to the -Min point estimate represents a statistically significant difference at the 5% and 1%, respectively, between the Min and the Max point estimates, using standard errors clustered by stock and by day. Min Max 0.60**

35 Table 8: Exchange Descriptive Statistics. The table reports summary statistics for the exchanges. The column Overall reports the statistics summed across the three largest exchanges in Canada. Columns Exchange 1, Exchange 2, and Exchange 3 report the statistics for each of the three exchanges separately. Shares d is the number of shares traded. Dollar Volume d is the number of shares traded multiplied by the share price. % of Dollar Volume d is the percent of the dollar volume trading on each exchange. % is the double-sided dollar-volume % of trades by a high frequency trader (). % Demand is the dollarvolume percent of trades in which an is the liquidity taker. % Supply is the dollarvolume percent of trades in which an is the liquidity provider. Quoted Half-Spread is the calendar time weighted one-half quoted difference between the best bid and the best ask price on each exchange. For Overall it is the calendar time weighted one-half quoted difference between the national best bid and the national best ask price on each exchange. % of time at both NBBO is the percent of the calendar time at which an exchange is quoting the nationally best bid and offer. % of time at either NBB or NBO is the percent of the calendar time at which an exchange is quoting either the national best bid or the national best offer, but not both. % of time at neither NBBO (while quoting) is the calendar time at which an exchange is quoting but neither its bid nor offer are at the national best. % of time with no B or O is the calendar time at which an exchange has no quotes on at least one side of the order book. Overall Exchange 1 Exchange 2 Exchange 3 Shares d (thousand) 1, , Dollar Volume d ($million) $32.70 $3.55 $20.61 $8.54 % of Dollar Volume d 10.9% 63.0% 26.1% % Demand 18.1% 23.1% 19.6% 14.4% % Supply 23.2% 34.6% 17.6% 45.3% % 20.7% 28.8% 18.6% 29.9% Quoted half-spread (bps) % of time at both NBBO 38.5% 86.8% 40.3% % of time at either NBB or NBO 40.9% 12.1% 38.3% % of time at neither NBBO (while quoting) 20.5% 1.1% 20.4% % of time with no B or O 0.1% 0.0% 1.1% 35

36 Table 9: Distribution of Activities by Exchange. The table reports the average stock-day distribution of orders by and on all three exchanges. The point estimates report the percent of activity by and that are s, s, or s. captures the number of orders at that exchange s NBB or NBO. 1 tick from NBBO captures the number of orders one cent away from that exchange s NBB or NBO. > 1 tick from NBBO captures the number of orders more than one cent away from that exchange s NBB or NBO. For cancels the analogous definition applies. Total # of Observations is the number of observations on that exchange on the average stock-day. Exchange 1 0.7% 2.1% 13.7% 10.8% 10.5% 6.8% 1 tick from NBBO 5.1% 1.9% 1 tick from NBBO 5.5% 3.2% > 1 tick from NBBO 4.8% 14.3% > 1 tick from NBBO 5.9% 14.7% Total # of Observations 17,670 Exchange 2 1.3% 4.2% 13.7% 13.0% 10.3% 7.7% 1 tick from NBBO 2.8% 2.9% 1 tick from NBBO 4.2% 4.3% > 1 tick from NBBO 3.9% 13.3% > 1 tick from NBBO 4.8% 13.5% Total # of Observations 39,316 Exchange 3 0.4% 1.8% 18.5% 5.2% 15.0% 3.7% 1 tick from NBBO 8.4% 1.4% 1 tick from NBBO 9.0% 2.1% > 1 tick from NBBO 11.4% 5.5% > 1 tick from NBBO 12.2% 5.5% Total # of Observations 27,532 36

37 Table 10: Return Impulse Response Function by Exchange. The table reports stock-day average Impulse Response Functions (IRF) from a 5-lag Vector-autoregression (VAR). The VAR is in event time with each order being an observation. takes the value +1 for buy initiated trades, -1 for sell initiated trades, and 0 otherwise. +1 for bids placed at the NBB, -1 for offers placed at the NBO, and 0 otherwise. 1 tick from NBBO take +1 for bids placed at one cent from the NBB, -1 for offers placed at one cent from the NBO, and 0 otherwise. > 1 tick from NBBO take +1 for bids placed greater than one cent from the NBB, -1 for offers placed at greater than one cent from the NBO, and 0 otherwise. For cancels the analogous definition applies with the sign such that cancels at the bid take the value +1 and cancels at the offer take the value -1. Only the,, and point estimates are reported. The observations include all displayed orders between 9:45 a.m. EST and 3:45 p.m. EST. To be included a stock-day must have at least 20 orders in each variable. The IRF is orthogonalized and order independent and reports the forecasted midpoint return, in basis points, after a +1 (buy event for orders and trades, sell event for cancels). The innovation is cumulative over 20 events. The first column reports the impulse response function (IRF), the second column the IRF, and the third column the difference between the and IRF. The observations are separated based on whether they occur on Exchange 1, 2, or 3 (there are 42 order variables, 7 order types on 3 exchanges for each and ). For and a *, ** next to the coefficient represents that the coefficient differs from zero and is statistical significance at the 5 and 1% level, respectively using standard errors clustered by stock and by day. For the Difference column *, ** next to the coefficient represents that the and coefficients differs from each other with statistical significance at the 5 and 1% level, respectively using standard errors clustered by stock and by day. Exchange 1 Difference 0.331** 0.179** 0.152** 0.255** 0.172** 0.084** ** ** Exchange ** 0.416** 0.542** 0.372** 0.239** 0.134** ** ** Exchange ** 0.427** 0.501** 0.294** 0.217** 0.077* ** * 37

38 Table 11: Impulse Response Function by Exchange. The table reports stock-day average Impulse Response Functions (IRF) from the same vector-autoregression (VAR) used in Table 5. The rows represent the variable being shocked by one unit. The columns represent the variable being affected. Panel A reports the average IRFs for the same exchange (e.g. the IRF of an innovation on Exchange 3 on on Exchange 3). Panel B reports the average IRFs for the other exchanges (e.g. the IRF of an innovation on Exchange 3 on on Exchanges 1 and 2). *, ** next to the coefficient represents that the coefficient differs from zero and is statistical significance at the 5 and 1% level, respectively using standard errors clustered by stock and by day. Panel A: Exchange Variable 0.056** 0.124** 0.070** ** ** 0.095** 0.057** ** 0.012** ** 0.001** 0.045** 0.074** 0.001** 0.004** 0.004** 0.011** 0.094** 0.013** 0.054** 0.012** 0.015** 0.001** 0.022** 0.002** 0.003** 0.061** 0.043** ** 0.009** 0.004** 0.056** 0.057** Panel B: Exchanges Variable 0.022** 0.053** 0.011** 0.004** 0.010** ** 0.048** 0.012** 0.004** 0.019** 0.004** 0.001** 0.017** 0.031** 0.001** 0.012** 0.010** 0.006** 0.043** ** 0.015** 0.002** 0.002** 0.022** ** 0.046** 0.008** 0.001** 0.011** 0.009** 0.005** 0.018** 0.033** 38

39 Table 12: Information Shares by Exchange. The table reports the average stock-day Hasbrouck (1995) minimum and maximum information shares for each exchange (Panel A) and for and orders/quotes on each exchange (Panel B). * and ** next to the -Min point estimates in Panel B represent a statistically significant difference at the 5% and 1%, respectively, between the Min and the Max point estimates, using standard errors clustered by stock and by day. Panel A: By Exchange Min Max Exchange Exchange Exchange Panel B: By Exchange Exchange 1 Min Max 0.75** Exchange ** Exchange **

40 Table 13: Cross Exchange Price Change Sequences. The table reports the frequencies of price change sequences on Exchange 2 and Exchange 3. The sequences capture all events in which Exchange 2 and Exchange 3 start with the same bid or ask price, and are followed by one of the exchanges best bid or ask price changing, and subsequently the other exchange updating its quote in the same direction. In Panel A and B the deviation and resolution must take at least one millisecond. In Panel C and D the deviation and resolution occur simultaneous. The statistics are calculated using data from 10/15/ /28/2013. Panel A evaluates the joint probability of a / and // opening a sequence and an / and // subsequently updating the second exchange s quotes. The average stock-day has 1009 sequences that last at least 1 millisecond. Panel B repeats the joint probability analysis for the average stock-day 164 sequences with both orders by the same trader. Panel C repeats the joint probability analysis for the average stock-day 1106 sequences that occur simultaneously. Panel D repeat the joint probability analysis for the average stock-day 170 sequences by the same trader that happen simultaneously. Panel A: Conditional on Open/Close Type and, > 1 Millisecond Second First 0.7% % 0.4% % 2.2% % % % 68.9% 2.0% % 3.8% % 30.6% 1.0% % 2.6% % 8.6% % % % 31.1% 0.5% % 1.5% % 10.6% % Close by r- 4.2% 33.9% 26.2% 8.3% 14.1% 13.3% % Close by r 64.3% 35.7% 40

41 Table 13 Continued Panel B: Conditional on Open/Close Type and by r, > 1 Millisecond Second % Open by r- % Open by r First 2.7% % % % % 64.5% 0.2% % % % % 15.3% % % 34.6% % % 13.6% % Close by r- 3.0% 45.7% 15.8% 13.6% 5.8% 15.3% % Close by r 64.5% 34.6% Panel C: Conditional on Open/Close Type and, Simultaneous Second % by r- % by r First 0.7% % 0.4% % 2.1% % % % 68.5% 2.4% % 4.5% % 31.4% 1.0% % 2.7% % 8.6% % % % 31.5% 0.6% % 1.7% % 10.9% % by r- 4.6% 33.1% 25.8% 9.2% 13.8% 13.5% % by r 63.5% 36.5% 41

42 Table 13 Continued Panel D: Conditional on Open/Close Type and by r, Simultaneous Second % by r- % by r First 2.7% % % % % 65.7% 1.1% % % % % 13.3% % % 34.3% % % 15.2% % by r- 3.7% 46.2% 15.8% 15.0% 5.8% 13.5% % by r 65.7% 34.3% 42

43 Table 14: Duration of Cross Exchange Price Change Sequences. The table reports the average length of time, in seconds, of price change sequences on Exchange 2 and Exchange 3. The sequences capture all events in which Exchange 2 and Exchange 3 start with the same bid or ask price, followed by one of the exchanges best bid or ask price changing, and subsequently the other exchange updating its quote in the same direction. The sequences must take at least one millisecond. The average stock-day has 1009 sequences. The average durations are calculated using data from 10/15/ /28/2013. Panel A evaluates the joint probability of a / and // opening a sequence and an / and // subsequently updating the second exchange s quotes. Panel B repeats the joint probability analysis for the 164 sequences by the same trader. Panel A: Conditional on Open/Close Type and Second First Panel B: Conditional on Open/Close Type and by r Second First

44 Figure 1: The Evolution of Price Discovery. The figure reports the trade correlated price discovery and the Herfindahl Index for trading in the Royal Bank of Canada (RBC) on the Toronto Stock Exchange, Alpha Trading System, and Chi-X. Data is provided by SIRCA on behalf of Thomson-Reuters from 2007 through