The Information Content of Hidden Liquidity in the Limit Order Book

Transcription

1 The Information Content of Hidden Liquidity in the Limit Order Book John Ritter January 2015 Abstract Despite the prevalence of hidden liquidity on today s exchanges, we still do not have a good understanding whether the traders who use hidden orders are informed. This paper empirically examines if there is information contained in hidden or displayed liquidity. The results support the hypothesis that informed traders tend to conceal their information advantage by using hidden orders when they supply liquidity. I find that imbalances in hidden liquidity in the limit order book predict future returns, while imbalances in displayed liquidity do not. Buying a portfolio of stocks with excess hidden liquidity on the bid side of the limit order book and shorting a portfolio of stocks with excess hidden liquidity on the offer side earns statistically significant midpoint returns of 4.3 bp after a holding period of 5 minutes and 10.2 bp after 30 minutes. A similarly constructed portfolio based on displayed imbalances produces negative, insignificant returns. I explore whether these results are explained by past returns or order flow, but find that the relationship between hidden limit order book imbalances and future returns continues to hold after controlling for past returns and order flow in a VAR framework and in portfolios double sorted based on either past returns or order flow and imbalances in hidden shares in the limit order book. I also find that hidden liquidity supplied by non-high frequency traders (NHFTs) contains more information than hidden liquidity supplied by high-frequency traders (HFTs). Longshort portfolios constructed from NHFT hidden imbalances earn returns of 7.4 bp and 18.5 bp over holding periods of 10 minutes and 2 hours, while similar portfolios for constructed from HFT hidden imbalances earn insignificant, negative returns. These results are consistent with the hypothesis that NHFTs who possess long-lived information supply liquidity using hidden orders to prevent information leakage. JEL Classification: G10, G12, G14 I would like to thank Gideon Saar, Ron Goettler, Ron Kaniel, Jerold Warner, Mihail Velikov and seminar participants at the Simon Business School for their comments and suggestions. I would also like to thank NASDAQ OMX for supplying the data used in this study and the Center for Integrated Research Computing (CIRC) at the University of Rochester for providing the computer resources used in this study. All remaining errors are my own. William E. Simon Graduate School of Business Administration, University of Rochester, Rochester, NY 14627; john.ritter@simon.rochester.edu Address: Carol Simon Hall, Rochester, NY 14627, United States of America 1

2 Keywords: Market Microstructure, Hidden Liquidity, Pre-trade Transparency, Limit order book, Informed Trading, Market Impact, High Frequency Trading 2

3 1 Introduction The ability to conceal an order is a common feature of today s electronic limit order markets. The SEC reports that in the first quarter of 2013, an average of 12.5% 1 of stock volume traded on U.S. public exchanges executed against hidden liquidity. 2. In the NASDAQ sample utilized in this study, 18% of the shares traded execute against hidden orders and 25% of the depth is the limit order book is concealed. Despite the prevalence of hidden liquidity on today s exchanges, we still do not have a good understanding of who uses hidden orders and whether or not they contain information. This paper attempts to shed some light on these questions. The answers can have important implications for market design and market quality in today s exchanges. On the one hand, if hidden liquidity could be used by informed traders to conceal their trading intentions, it could lead to increased adverse selection and reduced price discovery. On the other hand, liquidity could be improved if the ability to hide orders attracts liquidity providers who would not normally trade in a fully displayed market. My results support the hypothesis that informed traders tend to conceal their information advantage by using hidden orders when they supply liquidity. I find that imbalances in hidden liquidity in the limit order book predict future returns, while imbalances in displayed liquidity do not. Specifically, I find that buying a portfolio of stocks with excess hidden liquidity on the bid side of the limit order book and shorting a portfolio of stocks with excess hidden liquidity on the offer side earns statistically significant midpoint returns of 4.3 bp after a holding period of 5 minutes and 10.2 bp after 30 minutes. A similar portfolio constructed based on displayed imbalances in the limit order book earns insignificant returns of -0.2 bp and -1.2 bp over 5 minute and 30 minute holding periods, respectively. I explore whether these results are explained by past returns or order flow 3, but find that the relationship between hidden limit order book imbalances and future returns continues to hold after controlling for past returns and order flow in a VAR framework and in portfolios double sorted based on either past returns or order flow and imbalances in hidden shares in the limit order book. Usually public exchanges that allow the use of hidden orders fall into one of two types. 4 The first type of market only allows traders to post iceberg or reserve orders, which require the trader to display a minimum portion of their order size. Exchanges outside the U.S., such as the London Stock Exchange, 1. An average of 11.3% of traded volume executed against hidden volume for stocks in the smallest marketcap decile, while 16.0% of traded volume executed against hidden volume for stocks in the largest market cap decile. 2. Source: Hidden Volume Ratios on SEC website ( highlight html#.uqag8uka5_a) 3. I define order flow as the shares bought in market buy orders minus the shares sold in market sell orders. 4. A third type of hidden trading system is dark pools or crossing networks, which are usually private exchanges used by institutional investors that are not available to the public. Usually all orders in a dark pool are hidden until after the transactions are completed. 1

4 Euronext, the Australian Stock Exchange, and the Toronto Stock Exchange, fall into this category. In these types of exchanges all price quotes are displayed on the limit order books, but the displayed depth at each price might not be the total depth available for immediate execution. The second type of exchange allows the posting of iceberg orders as well as completely hidden orders, which do not show up on the limit order book. U.S. exchanges such as NYSE, NASDAQ, BATS, and Direct Edge fall under this category. On these types of exchanges, not only may some available depth not be displayed, but some available prices may not be displayed if there is no displayed depth at that price quote. If the either of the best price quotes is completely hidden, then the displayed spread will not reflect the true spread on the exchange. The existing theoretical literature documents a number of advantages to concealing an order. Harris (1997) and Butti and Rindi (2013) show that traders can use hidden orders to reduce competition when supplying liquidity. By not exposing their limit order, they reduce the chances of their order being undercut or front-run by parasitic traders. Hidden orders can also be of use to informed traders when they supply liquidity. Harris (1997) and Moinas (2010) show that informed traders can use hidden orders when supplying liquidity to reduce information leakage and the informational impact of displaying a large limit order. Although there are a number of advantages to using a concealed order, there are also a number of disadvantages. The main disadvantage involved in concealing an order is that the hidden part of a limit order loses execution priority to any displayed order on the limit order book at the same price level, even if the displayed order was submitted at a later time than the hidden order. Another disadvantage, discussed in Harris (1997) and modeled in Cebiroglu, Hautsch, and Horst (2013), is that hidden orders also have a lower chance of executing because they fail to attract reactive traders, who monitor the market but only trade when a trading opportunity presents itself. 5 The theoretical and empirical literature have not reached definitive conclusions about who uses hidden liquidity and the affect that it has on market quality. A number of empirical studies such as Aitken, Berkman, and Mak (2001), Pardo and Pascual (2011), Bessembinder, Panayides, and Venkatamaran (2009), and De Winne and D Hondt (2007), have found that hidden orders tend to be used by uninformed traders. While other empirical studies including Tuttle (2006), Belter (2007b), Kumar, Thirumalai, and Yadav (2010), and Frey and Sandas (2009) have found that hidden orders are used by informed trades, since they can forecast future price movements. 5. A disadvantage of hidden orders on US public exchanges under REG NMS, is that they are not included in the NBBO (National Bid and Best Offer) and are not eligible to attract market orders from other exchanges when they are the best posted quotes. Another possible disadvantage is that the rebate earned for supplying liquidity is smaller for supplying hidden liquidity than for supplying displayed liquidity. The rebate for adding displayed liquidity on NASDAQ ranges from 0.2 to cents per share executed. The rebate for adding non-displayed liquidity on NASDAQ ranges from 0.05 to 0.18 cents per share executed (Source: 2

5 One possible explanation for the inconclusive empirical results is that traders vary their hidden order strategies based on the trading environment. Experimental evidence from Bloomfield, O Hara, and Saar (2014) shows that changing the market opacity from allowing partially to fully hidden orders alters the trading strategies of both informed and uninformed traders. Informed and uninformed traders tend to supply more liquidity using hidden orders in the fully hidden regime than in the partially hidden one, with informed traders supplying more liquidity using hidden orders than uninformed traders in both markets. While the majority of prior empirical studies were conducted on markets that only allowed the use of iceberg orders, i.e. partially hidden, the data used in this study comes from a market, NASDAQ, that allows the submission of completely hidden orders. This could be one reason why my results conflict with the results from some previous studies which found hidden orders were not used by informed traders. My data does not let me directly observe the submission of hidden orders, instead it contains snapshots of the prices and aggregated depths at each price of the NASDAQ limit order book at the beginning of each minute. The depths are broken down based on whether they are hidden or displayed and whether they are submitted by a high frequency trader (HFT) or non-high frequency trader (NHFT). In order to examine the information content of hidden liquidity, I sort stocks into quintile portfolios based on imbalances in hidden liquidity between the bid and offer sides of the limit order book. I find that hidden imbalances in favor of the bid side of the order book can forecast positive midpoint returns of 5.9 bp 30 minutes after portfolio formation. Hidden imbalances in favor of the offer side of the book also forecast negative midpoint returns of -4.4 bp half an hour after portfolio formation. Similarly constructed portfolios based on displayed imbalances in the limit order book do not possess the ability to forecast returns. The ability of hidden imbalances to forecast returns appears to be greatest in small cap stocks, where information symmetry is likely to be higher. Although the ability to predict future returns can indicate that hidden orders are informed, there could be other possible explanations for why hidden imbalances predict future returns. One possibility is that the hidden order imbalance is a result of an imbalance in the order flow, which is known to predict future returns. If there was an excess number of market buy orders, they would decrease the depth on the offer side of the book, generating positive returns and leaving a balance in favor of the bid side of the limit order book. I investigate this alternative and find that the order flow is in the opposite direction of the future returns and of both the hidden and displayed order book imbalances, i.e. the traders on the side with the imbalance are providing liquidity to the excess order flow. It is also possible that hidden imbalances are correlated with the order flow of a subset of traders,hirschey (2013) shows that HFT order flow can predict future returns over short horizons. I separate out the HFT and NHFT net order flow and find that they are both in the opposite direction of the hidden imbalances and future returns. 3

6 Another explanation that fits with the observed direction of order flow is that risk averse investors use hidden orders to provide liquidity to traders who require immediacy. As discussed in Kaniel, Saar, and Titman (2008), the risk averse traders demand price concessions when supplying liquidity, which results in positive future returns. I investigate this alternative using portfolios double sorted based on lagged order flow and hidden or displayed limit order book imbalances. I find that the relationship between hidden orders and future returns still exists when the order flow is in the same direction as the hidden imbalance, indicating that the returns are not likely to be caused by risk-averse traders supplying liquidity using hidden orders. Another alternative is that the returns are a result of the hidden orders creating a mismatch between the supply and demand for liquidity, as modeled in Cebiroglu, Hautsch, and Horst (2013). Eventually the unfilled hidden orders are canceled and submitted as market orders, resulting in future positive returns. However, this alternative is not likely, since the order flow observed after the formation of the hidden portfolios is in the opposite direction of what the model predicts. I also find that the relationship is not a result of autocorrelation in the return process, since the relation between hidden imbalances and future returns still persists after controlling for past returns and HFT and NHFT order flow in a VAR framework. Instead, the results seem to support the hypothesis that informed liquidity suppliers use hidden orders. One advantage of my dataset is that it allows me to identify the type of trader who submitted the hidden order as a high frequency trader (HFT) or non-high frequency trader (NHFT). I examine which types of traders using hidden liquidity possess information and find that information appears to be concentrated in NHFTs. Long-short portfolios associated with NHFT hidden imbalances earn returns of 7.4 bp and 18.5 bp over periods of 10 minutes and 2 hours, while similar portfolios for non-hft displayed imbalances do not earn significant returns. I find that there does not appear to be a difference in information between the displayed and hidden liquidity supplied by HFTs, both portfolios of visible and hidden HFT imbalances earn negative returns after a period of 5 minutes. The ability of NHFTs hidden liquidity to predict future returns is consistent with the hypothesis that NHFTs possess long-lived information and use hidden orders to prevent information leakage when supplying liquidity. Overall, my findings lend support to the theory that informed traders use hidden orders. My results may be of interest to both regulators and academics as they continue to debate the effects pre-trade transparency has on price discovery and market quality. The paper is organized as follows. Section 2 reviews the literature on hidden liquidity and informed traders using limit orders. Section 3 describes the data and discusses how the hidden and displayed limit order book imbalance measures are constructed. Section 4 investigates if informed traders supply liquidity using hidden or displayed orders by examining the ability of hidden and displayed limit order book 4

7 imbalances to predict future returns. Section 5 looks at alternatives explanations for why hidden limit order imbalances forecast future returns. Section 6 examines the profitability of trading strategies based on hidden limit order imbalances. Section 7 studies which locations in the limit order book contain the most information about future returns. Section 8 examines whether hidden liquidity supplied by HFTs or NHFTs is more likely to be informed. Section 9 concludes. 2 Literature Review This study primarily relates to the literature on the use of concealed orders in limit order book markets. The theoretical literature has mainly focused on how much a liquidity supplier in an iceberg market should conceal their order. It has documented a number of costs and benefits to using hidden orders, but has not reached a consensus about whether informed or uninformed traders use them. The main cost of using a hidden limit order is a decreased probability that the order will be executed. Hidden orders lose priority to displayed orders at the same price, even if the displayed order was submitted at a later time than the hidden order. In addition to lower priority, Harris (1997) mentions that hidden orders also have a lower chance of executing because they fail to attract reactive traders, who monitor the market but only trade when a trading opportunity presents itself. The main benefit to hiding an order is reducing the order exposure costs associated with limit orders. Harris (1997) proposes that traders can use hidden orders to reduce competition when supplying liquidity. By not exposing their limit order, they reduce the chances of their order being undercut or front-run by parasitic traders. Butti and Rindi (2013) examine order exposure costs in a model in which an uninformed trader competes to supply liquidity by simultaneously deciding on an order s price, size, and the amount of the order that is visible. They show that large traders use hidden orders to reduce exposure costs and avoid being undercut by competing liquidity suppliers. Cebiroglu, Hautsch, and Horst (2013) also model an uninformed trader s decision to expose their order when competing to supply liquidity in the limit order book while also competing with an off-exchange upstairs market for order flow. They find that requiring orders to be displayed in the limit order book attracts order flow from the upstairs market and improves market quality by helping coordinate the supply and demand of liquidity. Allowing hidden orders in the limit order book generates excess volatility and higher trading costs due to a lack of coordination between liquidity demand and supply. One interesting finding of their model is that uninformed traders generate positive returns when supplying hidden liquidity, but not when supplying displayed liquidity. In their model, when an uninformed trader supplies liquidity using a hidden bid order, it fails to attract upstairs order flow. The impatient liquidity supplier then converts their hidden limit order into a market buy order, which generates a positive return. If the liquidity supplier submits 5

8 a displayed order, it attracts upstairs order flow and the liquidity supplier does not replace their limit order with a market order. One drawback of these models is that they do not include informed traders and cannot account for the role that adverse selection plays when deciding whether to use hidden liquidity. Harris (1996) and (1997) propose that uniformed traders can reduce the chances of being adversely selected by informed liquidity demanders by using hidden orders to mitigate the option value of their standing limit orders. Hidden orders can also be of use to informed traders when they supply liquidity. Harris (1997) proposes that informed traders can use hidden orders when supplying liquidity to reduce information leakage and the informational impact of displaying a large limit order. Moinas (2010) theoretically models a liquidity supplier s decision to hide an order. In her model, informed liquidity suppliers can cause uninformed liquidity demanders to retreat from the market if they post a large displayed order. Moinas finds that both informed and uninformed liquidity suppliers who submit large orders use hidden orders in order to decrease their permanent price impact and not scare away uninformed liquidity demanders. Boulatov and George (2013) model the choice of informed traders to supply or demand liquidity in hidden and displayed markets. They find informed traders use hidden orders to supply liquidity in an opaque market, but demand liquidity in fully displayed markets. This results in improved market quality in the hidden market, because the informed traders compete to supply liquidity. Bloomfield, O Hara, and Saar (2014) is one of the few studies to examine hidden order usage in a market structure that allows the trader to completely conceal their order. Using an experimental setting to study how market transparency affects market quality and trader behavior, they examine three different opacity regimes in which informed and uninformed traders can submit: only displayed orders, displayed orders and iceberg orders, or displayed orders, iceberg orders, and fully hidden orders. They find that order strategies are greatly affected by the opacity of the market, but market liquidity and informational efficiency do not appear to vary between the markets. They find that both informed and uninformed traders alter their order strategies and use hidden orders to supply liquidity when the option is available, but informed traders tend to use hidden orders more than uninformed traders. They also find that as opacity increases, so does the usage of hidden orders. Gozluklu (2012) uses an experimental setting to compare fully displayed markets with markets that permit the use of iceberg orders. He finds that both informed traders and large liquidity traders use more iceberg orders compared to noise and small liquidity traders. He also finds that when informed traders use iceberg orders, they tend to be in large quantities and at less aggressive prices (i.e. deeper in the limit order book). Gozluklu concludes that market transparency does not affect market quality when there are informed traders. Most empirical studies investigating hidden liquidity in the limit order book examine markets that 6

9 only allow orders to be partially hidden through the use of iceberg orders. They draw differing conclusions as to whether hidden orders are used by informed or uninformed traders. In one of the first empirical studies of hidden orders, Aitken, Berkman, and Mak (2001) examine hidden orders on the Australian Stock Exchange. They find that the price impact of hidden orders is similar to that of visible limit orders and conclude that hidden orders are used by uninformed traders to manage their exposure risk. Pardo and Pascual (2011) investigate iceberg order transactions on the Spanish Stock Exchange. They find that the detection of hidden volume does not impact prices and traders on the other side of the market submit more aggressive orders after hidden liquidity is detected. Bessembinder, Panayides, and Venkatamaran (2009) examine the use of iceberg orders on Euronext Paris. They find that iceberg orders have lower execution probabilities and take longer for the order to fill, but also experience lower execution costs in the form of lower implementation shortfall costs. They conclude these orders are used by uninformed traders to balance their exposure and executions risks. De Winne and D Hondt (2007) also examine hidden orders on Euronext. They find that uninformed traders use hidden orders to manage their picking off risk. They also find that after hidden depth is detected, order aggressiveness increases on the opposite side of the limit order book. Fleming and Mizrach (2009) document the use of iceberg orders on the BrokerTec U.S. Treasuries ECN. They find that the use of iceberg orders increases when adverse selection and volatility increase. Anand and Weaver (2004) look at the removal and reintroduction of iceberg orders on the Toronto Stock Exchange. They find that spreads, volume, and quoted depth are unaffected but total depth increases when hidden orders are allowed. They also find evidence that traders use hidden orders to manage their exposure risk and reduce their price impact. A number of studies have found that iceberg orders can predict future returns. Tuttle (2006) looks at hidden depth in the NASDAQ market-maker SuperSOES system. She finds that hidden depth supplied by investment banks and wirehouses predicts later day price changes, while displayed depth does not. Belter (2007b) and Belter (2007a) investigate iceberg orders using a sample of 18 stocks from the Copenhagen Stock Exchange. They find that imbalances in hidden order book depth forecast future intraday returns, but displayed imbalances do not. However, a trading strategy cannot be constructed that earns positive returns. Kumar, Thirumalai, and Yadav (2010) examine iceberg orders on the Indian Stock Exchange. They find that trader types with higher levels of information are more likely to submit hidden orders. They also find that these trader types earn higher profits when they submit hidden orders than when they submit displayed orders. Frey and Sandas (2009) look at the use of iceberg orders on Deutsche Borse s XETRA platform. They find that the use of iceberg orders impacts prices and their execution attracts market orders, but conclude that the price impact is associated with liquidity rather than informed trading, because the price impact decreases as a greater portion of the iceberg order is executed. 7

10 More recent empirical studies have looked at markets that allow for the submission of completely hidden orders. Using data from INET, Hasbrouck and Saar (2009) examine fleeting orders, limit orders that are canceled within a few seconds of being submitted. They find that one use of fleeting orders is the detection of completely hidden orders. Hautsch and Huang (2012) use order data from NASDAQ to examine executions against hidden orders. Since they are only able to observe visible orders that execute against hidden orders, they use a censored ordered probit model to estimate the presence of hidden liquidity. They find that traders use hidden liquidity to compete for the provision of liquidity and to avoid being adversely selected or having their orders front-run. Cebiroglu, Hautsch, and Horst (2013) supplement their model by empirically examining hidden liquidity on 13 stocks on NASDAQ. They find that an increase in hidden depth on the bid side of the limit order book generates positive returns, but an increase in displayed depth does not. They do not attribute the returns to hidden orders containing information, but instead to a mismatch between the supply and demand of liquidity. They find that an increase in bid side displayed depth causes an increase in buy-side executions. While an increase in bid side hidden depths causes an increase in sell-side executions and an increase in returns. They associate the sell-side executions with the traders supplying hidden limit orders on the bid side of the limit order book canceling their limit buy orders and replacing them with marketable buy orders. This study is also related to the literature on the use of limit orders by informed traders. Kaniel and Liu (2006) present a model in which informed traders with long lived information prefer to trade using limit orders. They use TORQ data to provide empirical support that limit orders are more informative than market orders. Goettler, Parlour, and Rajan (2009) solve a model of a limit order market with asymmetric information. They find that speculators, traders with no intrinsic motive to trade, are the most likely to submit limit orders and are also willing to pay the most for information. Rosu (2014) presents a model of a limit order market in which informed traders can submit market or limit orders. He finds that informed traders submit market orders when they possess an extreme information advantage and limit orders when they possess a moderate information advantage. In an experimental setting, Bloomfield, O Hara, and Saar (2005) observe that informed traders prefer using limit orders instead of market orders. Empirically, Harris and Panchapagesan (2005) show that the NYSE limit order book provides information about future prices and that specialists use this information at the expense of limit-order traders. Cao, Hansch, and Wang (2009) use data from the Australian Stock Exchange and find that limit order book quotes above the inside quotes are significantly related to future short-term price movements in the same direction as the imbalance. This study is also related to the recent literature on HFTs supplying liquidity. Theoretically, a number of papers examine market makers who supply liquidity when they posses a speed advantage in processing 8

11 information. Jovanovic and Menkveld (2012) present a model in which HFTs act as informed middlemen in a limit order market. They find that HFTs improve welfare by reducing adverse selection if informed traders are liquidity demanders, but reduce welfare if the liquidity demander is uninformed. Hoffman (2013) models a limit order market with fast and slow traders. He finds the speed advantage of fast traders reduces their risk of being picked off and allows them to submit more aggressive limit orders. However, this induces slow trader s to submit limit orders with lower execution probabilities. Aït-Sahalia and Saglam (2014) model an HFTs trading decisions when they receive a signal about future order flow. They find that HFTs quotes and cancellations increase when they receive faster signals, and that they provide less liquidity when volatility increases. Menkveld and Zoican (2014) examine a model with HFT market makers and liquidity demanders. They find that the adverse selection HFT market-makers face from high-frequency liquidity demanders increases as latency decreases, which causes quoted spreads to increase. A number of empirical studies have also looked at high-frequency traders. Brogaard, Hendershott, and Riordan (2014) use the same NASDAQ transaction dataset used in this study, they find that when HFTs demand liquidity they trade in the same direction as the permanent price impact, which aids in price discovery. They also find that when HFTs supply liquidity they trade in the opposite direction of the permanent price impact, which makes them subject to adverse selection. Hirschey (2013) looks at HFT order flow and finds that they anticipate and trade ahead of the order flow of NHFTs. Hagströmer and Nordén (2013) look at different types of HFTs on the NASDAQ-OMX Stockholm exchange, they find that HFT market makers compromise the majority of HFT trading volume and help reduce intraday volatility. Brogaard et al. (2014) use data on changes in the speed of trader colocation subscriptions on NASDAQ OMX Stockholm to examine how trading speed affects trading activity and market quality. They find that traders with faster colocation connections face lower adverse selection costs, improve their inventory management, and increase the share of liquidity they supply to the market. Hagströmer, Nordén, and Zhang (2014) examine the aggressiveness of HFT order submissions. They find that HFTs submit more aggressive orders when same side depth is large and supply liquidity when the spread is wide. Hasbrouck and Saar (2013) study how HFT activity affects market quality using a proxy for HFT activity based on NASDAQ message data. They find that increased HFT activity reduces spreads, increases displayed depth, and lowers short-term volatility. Gai, Yao, and Ye (2014) examine how trading speed effects market quality. They find that although trading speed increases after technology upgrades at NASDAQ, market quality measures such as spreads, trading volume, and price efficiency remain the same, cancellation/execution ratios and short term volatility increase, and displayed depth decreases. Kirilenko et al. (2014) examine HFT activity in the S&P index futures market during the flash crash 9

12 and find that HFT trading increased market volatility. Malinova, Park, and Riordan (2013) examine how retail trading costs are affected by algorithmic trading using an increase in the fees charged for message traffic in Canada. They find that the increased fees reduced message traffic of HFT market makers and increased bid-ask spreads. Trading costs increased for institutions but not for retail traders. Menkveld (2013) looks at the trading activity of an HFT firm on NYSE-Euronext and Chi-X. He finds that the HFT supplies liquidity on 80% of its trades, earns money on the bid-ask spread, but loses money due to its inventory positions. 3 Data 3.1 Data Sources and Sample Description This study utilizes a number of data sources made available by NASDAQ to investigate hidden liquidity in the limit order book. The primary data sets used in this study are the same ones used in Brogaard, Hendershott, and Riordan (2014) to analyze the impact high frequency traders (HFTs) have on price discovery. 6 The sample provided by NASDAQ consists of 120 stocks that were randomly chosen by selecting 40 firms from three different market cap categories. The stocks were selected so that 20 firms in each category have a primary listing on NASDAQ and 20 firms have a primary listing on NYSE. The Large cap stocks were selected from the largest market capitalization, the medium cap stocks were selected from stocks around the 1,000th largest stock in the Russell 3000, and the small cap stocks were selected from stocks around the 2,000th largest stock in the Russell Two of the stocks in the dataset were excluded from this study, because TAQ data was not available for these stocks during the entire sample period. The first dataset contains data on all transactions that occurred on NASDAQ during regular trading hours in 2008, 2009, and the week of 02/22/2010-2/26/ The trades are timestamped to the millisecond and include data on the price and size of the trade, whether the trade was buyer or seller initiated, and whether the liquidity demander and supplier were high frequency traders (HFTs) or a non-high frequency trader (NHFTs). 8 NASDAQ classifies 26 firms as HFTs based on an analysis of their trading activity, including how long their orders last, how long they hold their inventory, and their order to trade ratio. Unfortunately, the data cannot classify all HFT activity. Firms that engage in brokerage services and also run proprietary trading desks would not be identified as HFTs in the analysis, nor would HFTs who 6. NASDAQ makes the HFT data available for research to academics who sign a nondisclosure agreement. 7. The trade data does not include trades that occur during the opening, closing, or intraday crossings sessions. 8. The HFT indicator is an aggregate indicator identifying if a trade or limit order book depth belongs to one of the 26 firms that NASDAQ identifies as high-frequency traders in this study; it does not provide a firm level id, so it is not possible to identify the transactions or limit orders for individual HFT firms. 10

13 route their orders through large brokerage firms. The data set can be thought of as indicating a lower bound on the activity of HFTs. The second dataset contains snapshots of the NASDAQ limit order book at the beginning of each minute during regular trading hours. The dataset contains data for the first full week of the first month of each quarter in 2008 and 2009, the week of 09/15/ /19/2008 during the financial crisis, and the week of 2/22/2010-2/26/2010. Each snapshot includes the ten best bid and offer prices on Nasdaq s order book along with the aggregate depth at each price, broken down based on whether the depth is hidden or displayed and whether it is supplied by an HFT or an NHFT. Since the short-sale ban imposed by the SEC during September and October 2008 might have altered liquidity provision and trading behavior, I exclude data from these months from my sample. I also exclude observations that occur less than five minutes after the open or less than five minutes before the close, leaving me with 380 snapshots for each of 40 trading days for each stock in the sample. NASDAQ also provided data from NASDAQ TotalVieW ITCH, which contains information on displayed limit orders that are canceled or added to the NASDAQ limit order book and market orders on NASDAQ that execute against displayed and hidden liquidity. Quote data from TAQ was used to calculate NBBO spreads, depths, and midpoint prices. CRSP is used to calculate daily measures of total dollar trading volume, market capitalization, share price, daily returns, and daily proportional trade range 9. Table 1 provides descriptive statistics for the pooled time-series of the whole sample and each of the market capitalization groups. The average market capitalization for the whole sample is $16.9 billion. The different market cap groups vary greatly in size, with the large cap group having an average marketcap of $48.0 billion, the medium cap group an average of $1.6 billion, and the small cap group an average of $397 million. Prices tend to be higher and daily volatility 10 tends to be lower in the large cap stocks. On average, the stocks in the sample experienced negative daily returns of -9 bp, with a standard deviation of 3.1%. The average daily trading volume is $205 million, and exhibits a wide range from an average of $580 million in large cap stocks to an average of $3.6 million in small cap stocks. Trading on NASDAQ accounts for an average of 30% of daily traded volume, but experiences a wide range from a minimum of 3.4% to a maximum of 94.8%. The wide range is caused by the sample be constructed from a mix of firms whose primary listings are on NASDAQ and NYSE. The distribution of the percent of trading on NASDAQ appears to be similar across the different market cap groups. On average HFTs account for 28.9% of the NASDAQ daily trading volume within the sample. They 9. Defined as the difference between the daily high and low price divided by the closing price 10. Daily volatility is computed as the difference between the daily high and low trading prices divided by the midpoint closing price 11

14 demand liquidity for 33.2% of NASDAQ traded volume and supply liquidity to 24.6% of traded volume. HFT liquidity supply differs greatly across market cap groups, with HFTs supply liquidity to 40.8% of traded volume in large cap stocks and 13.1% of traded volume in small cap stocks. Similarly, HFTs demand liquidity for 41.9% of traded volume in large caps and 21.6% of traded volume in small caps stocks. These patterns support the findings in other studies, that HFTs tend to concentrate more of their trading activity in large, liquid stocks. Hidden orders constitute 18.1% of the supplied liquidity that market orders trade against on NAS- DAQ, with the 25% and 75% percentile distributions accounting for 10.4% and 23.8% of supplied liquidity. Hidden liquidity tends to be slightly greater in smaller stocks, representing on average 20.7%, 18.6%, and 16.1% of the liquidity supplied in trades in small, medium, and large cap stocks. 3.2 Limit Order Book Aggregation In order to make the limit order book data easier to analyze, I aggregate the depth offered at different price quotes together into groups based on where they are in relation to the displayed prices in the limit order book. For each side of the limit order book, shares are aggregated together in each group based on whether they are hidden or displayed and if they are supplied by an HFT/NHFT. Hidden shares at prices inside the NASDAQ best visible quotes 11 (BVQ hereafter) are grouped together (Inside BVQ). Shares at the BVQ price are grouped together (BVQ). Shares at prices above 12 the BVQ up to and including shares in the 5th best visible quote (5VQ hereafter) are grouped together (BVQ to 5VQ). Shares at prices above the 5VQ are grouped together (Above 5VQ). Table 2 provides a summary of the NBBO spread 13 as well as both the displayed and true spreads on NASDAQ for each market cap group. 14 Data in the table is reported from the cross-sectional distribution constructed from the time series means for each firm in the sample. As expected, stocks in the large market cap group have the lowest spreads, with median NBBO spreads of 1.19 cents and 4.1 bp. Spreads vary greatly over the different market cap groups, with medium and small cap median NBBO spreads being 2.3 and 6.3 times greater, respectively, when measures in basis points. Both the NASDAQ displayed and true spreads are larger than the NBBO spreads across all market cap groups, with the median true and displayed spreads being 5.3 bp and 5.8 bp (1.3 and 1.4 times greater) for large cap stocks, 15.1 bp and 17.7 bp (1.3 and 1.5 times greater) for medium cap stocks, and 37.6 bp and 43.8 bp (1.5 and The best visible quotes represent the highest displayed bid quote and the lowest displayed offer quote on NASDAQ. 12. Offer prices higher than the lowest displayed offer quote and bid prices less than the highest displayed bid quote are referred to as being above the best quotes 13. NBBO stands for national best bid and offer, which is the best displayed bid and offer price across all public exchanges in which the stock trades in the U.S. National Market System 14. The true spread is the difference between the best bid and offer quotes among both hidden and displayed limit orders on NASDAQ. The displayed spread is the difference between the best bid and offer quotes among only displayed limit orders on NASDAQ. 12

15 times greater) for small cap stocks. The true spread is less than the displayed spread 11.7%, 34.7%, and 34.9% of the time for the median large, medium, and small cap firm, indicating hidden orders inside the visible spread are more common in small and medium cap stocks than in large cap stocks. Table 2 also summarizes the aggregated depths from the bid side of the limit order book. 15 Total bid depth is greatest in large cap stocks. With a value of 23,100 shares ($1,575,000), the median total bid depth for the large cap sample is almost 4 times larger than the median values of the medium and small cap samples, 5,185 shares ($316,000) and 5,841 shares ($164,000), respectively. The median depth at the BVQ is 1,750 shares ($142,000) for large cap stocks, which represents 7.6% of the median total depth for the 10 best price levels. HFT liquidity supply appears to vary across the market cap groups. HFTs in the large cap group focus their liquidity supply near the BVQ, accounting for 62% of quoted depth inside the BVQ and 52.6% at the BVQ, but tend to be less active at quotes above the BVQ, accounting for only 35% of quoted depth. HFTs do not concentrate their supplied liquidity near the top of the limit order book as much in medium and small cap stocks, instead they tend to distribute it more evenly throughout the book. The use of hidden orders varies greatly across the different market cap groups and tends to be inversely related to size, with hidden orders accounting for 13.7% of the quoted depth in large cap stocks and 37.7% of the quoted depth in small cap stocks. The use of hidden liquidity also varies with the location of the quoted price. Hidden liquidity inside the BVQ accounts for 3% of the median total depth in large cap stocks and 6.9% of total depth in small cap stocks. Hidden depth accounts for 15% of the depth at the BVQ, with a similar distributed across the different market cap groups. The supply of hidden liquidity deeper in the limit order book varies across different market cap groups, with hidden depth accounting for 9.8% of total depth above 5VQ in large cap stocks and 32.5% of depth above 5VQ in small cap stocks. HFTs and NHFTs tend to use hidden liquidity differently, with HFTs showing a larger variation across the market cap groups. NHFTs use a greater percentage of hidden depth in their orders than HFTs in large cap stocks, but a smaller percentage then HFTs in small cap stocks. In large cap stocks, HFTs and NHFTs tend to use more hidden depth closer to the BVQ, with 7.6% of HFT depth and 18.4% of NHFT depth at the BVQ being hidden, while only 2.3% of HFT depth and 9.7% of NHFT depth above the 5VQ is hidden. The opposite pattern occurs in small cap stocks, with only 7.3% of HFT depth and 15.5% of NHFT depth being hidden at the BVQ, while 50.6% of HFT depth and 22% of NHFT depth is hidden at prices above the 5VQ. 15. Data from the offer side of the limit order book is quantitatively similar to the bid side. 13

16 3.3 Limit Order Book Imbalances The primary method used to evaluate the differences in information content between displayed and hidden limit orders is by sorting stocks into quintiles based on imbalances in shares between the bid and offer side of different parts of the limit order book. Order book imbalances are calculated for each part by subtracting the shares in the part on the offer side of the limit book from shares on the bid side of the order book. 16 For some parts of the study percentage imbalances are used to more easily compare imbalances across stocks. Percentage imbalances are constructed for each part by taking the imbalance for that part of the limit order book and dividing it by the total shares in the limit order book, the sum of all hidden and displayed depth for all quoted prices on the bid and offer side of the limit order book. Table 3 provides descriptive statistics of the daily standard deviation of the constructed order imbalance measures in thousands of dollars for the different market cap groups. For each firm the time-series mean of the daily standard deviation of imbalances in the one-minute limit order book snapshots is constructed. Data in the table is reported from the cross-sectional distribution constructed from the time series means for each firm in the sample. Comparing median firms across the market cap groups, total dollar imbalances are greatest in the large cap stocks and dollar imbalances are greater deeper in the book than at the BVQ. Hidden imbalances tend to be smaller than displayed imbalances in large cap stocks, but greater in medium and small cap stocks, with this pattern holding across the different book locations for each of the market cap groups. NHFT total book imbalances tend to be greater than HFT imbalances, the standard deviation of NHFT imbalances is 6.1 times greater for the median firm then HFT imbalances, despite NHFTs accounting for only 2.7 times as much total book depth as HFTs, as shown in Table 2. The difference between NHFT and HFT imbalances is smaller in medium and small cap stocks, with NHFT imbalances being only 2.5 and 2.7 times greater in medium and small cap stocks. Similar to total hidden imbalances, both NHFT and HFT hidden imbalances are smaller than displayed imbalances in large cap stocks, but greater in medium and small cap stocks. In order to ease comparison across stocks, order book imbalances are expressed as a percentage of total order book depth when grouping stocks into portfolios based on limit order book imbalances. Table 4 reports the distribution of percentage order book imbalances from the pooled time-series of all stock-minute observations. As expected, the median order book imbalances are 0. For over 80% of the sample, the hidden and displayed imbalances are less than 25% of total book depth.. Hidden Imbalances are less than displayed imbalances for the 10th through the 90th percentiles, with the 10th and 90th hidden imbalance percentiles representing 23.1% and 19.7% of total book depth, and the 10th and 90th displayed imbalances representing 23.6% and 24.3% of total book depth. Hidden imbalances tend to be 16. A positive imbalance would indicate there are more shares on the bid side of the order book, while a negative imbalance would indicate more shares on the offer side. 14

17 more extreme than displayed imbalances in the tails of the distribution, with 1st and 99th percentile hidden imbalances of 66.0% and 62.0% and 1st and 99th percentile displayed imbalances of 56.6% and 53.9%. NHFT imbalances tend to be greater than HFT imbalances, with NHFT 10th and 90th percentile total book imbalances of 33.2% and 32.3% and HFT 10th and 90th percentile total book imbalances of 9.2% and 8.0%. The relationships between HFT hidden and displayed imbalances and NHFT hidden and displayed imbalances both appear to follow similar trends, with hidden imbalances being smaller than displayed imbalances for the middle 80% of the sample and hidden imbalances being greater than displayed imbalances in the tails of the distribution. 3.4 Additional Measures Table 5 contains descriptive statistics for additional measures used in this study. Each measure is constructed based on the one minute average or standard deviation for each stock day. The stock day means and standard deviations are then averaged over each stock and the table reports the cross-sectional distribution of stock averages for each market cap group. The table reports the standard deviation of the one minute NBBO midpoint returns in basis points, which is constructed from the NBBO displayed quotes across all eligible exchanges in the US listed on TAQ. Small cap stocks are the most volatile, with a mean NBBO midpoint standard deviation of 16.2 bp. The NASDAQ transaction data is used to construct intraday measures of trading volume and order flow for different time intervals. The net marketable buying for each time interval is constructed by taking the total shares traded in the time interval that were seller initiated (i.e. initiated by a marketable sell order) and subtracting them from the total shares that were buyer initiated. 17 Table 5 reports the standard deviation of one minute Net marketable buying, in thousands of dollars, calculated separately for HFTs, NHFTs, and all combined traders. As expected, Net Marketable Buying is greatest for large cap stocks, with a mean standard deviation of $247,000 per minute. The standard deviation of NHFTs net marketable buying is greater than HFTs across all market cap groups, with mean NHFT and HFT net marketable buying standard deviations of $185,000 and $125,000 in large cap stocks. Data on displayed orders added and canceled from the limit order book is calculated from NASDAQ TotalView ITCH. Table 5 shows the average limit order dollar volume, in thousands of dollars, added and canceled from the bid and offer side of the NASDAQ limit order book each minute. The average one minute dollar volume added and canceled from each side of the limit order book is 12.9 and 12.4 times the average trading volume for large cap stocks, 24.5 and 23.9 times for medium cap stocks, and 40.6 and 39.5 times for small cap stocks. 17. A positive value means that the number of shares in marketable buy orders was greater than the number of shares in marketable sell orders during the time period. 15