Aggregated Market Quality Implications of Dark Trading

GBENGA IBIKUNLE* University of Edinburgh, United Kingdom European Capital Markets Cooperative Research Centre, Pescara, Italy Centre for Responsible Banking & Finance, University of St Andrews, United Kingdom1

MATTEO AQUILINA Financial Conduct Authority

YUXIN SUN University of Edinburgh, United Kingdom

IVAN DIAZ-RAINEY University of Otago, New Zealand

Abstract We test predictions regarding the implications of operating dark pools alongside lit venues for market quality in London ‘City’ venues. We find that dark trading at moderate levels enhances market transparency, and reduces both trading noise and incidences of potential trading manipulation. Results, however, imply that there is a threshold of value when dark trading diminishes transparency, and thus induces deterioration in market quality. We estimate this threshold to be about 15% of the aggregate market trading value. We also observe variations across stocks based on stock liquidity, with 5% and 18% for the highest and lowest liquidity stocks respectively.

JEL classification: G10, G14, G15 Keywords: Dark pools, MiFID, Multilateral Trading Facilities (MTFs), Market Transparency, Adverse Selection, Probability of an Informed Trade (PIN), Trading noise

*Corresponding author: University of Edinburgh Business School, 29 Buccleuch Place, Edinburgh EH8 9JS, United Kingdom; e-mail: [email protected]; phone: +441316515186. We thank seminar participants at the 19th November Seminar at the University of Reading’s ICMA Centre, Alfonso Dufour, Albert J. Menkveld and Marius Zoican for helpful comments and discussions. Ibikunle acknowledges funding from the University of Edinburgh’s First Grant Venture Fund and data support from the Capital Markets Cooperative Research Centre, Sydney. 1. Introduction

Over the course of the last decade regulators have stepped up efforts to make global financial markets more transparent and fairer, especially in the US and the EU. The regulations enacted, the Regulation NMS (RegNMS) in the US and the Markets in Financial

Instruments Directive (MiFID) in the EU, coupled with technological advancements, have in effect led to the proliferation of new classes of trading venues. One of the most prominent new venue type is a class of opaque venues known as ‘Dark Pools’. Trades executed on dark pools have no pre-trade transparency, i.e. market participants, other than the submitter and the pool operator, are not aware of the presence of orders submitted to a dark pool prior to their execution.

Dark pools in Europe are mainly operated as Multilateral Trading Facilities (MTFs); the three largest dark trading venues on the continent are MTFs, which operate both dark and lit order books. These are Chi-X Europe, BATS Europe and Turquoise, all are based in

London, United Kingdom. Considering that a non-negligible proportion of European exchange volumes are now executed in the dark,1 it is vital to investigate how these volumes impact overall market quality in the City. Such an investigation is also of critical importance for investors who already view dark pool activity with some suspicion (see Schacht et al.,

2009). The scepticism appears valid; a lack of pre-trade transparency suggests that dark pools reduce market transparency, thus going against one of the aims of the new market regulations enacted over the last decade. However, recent studies offer differing a view (see Zhu, 2014;

Comerton-Forde and Putniņš, 2015).

This paper contributes to the dark pools debate by presenting the first evidence on the impact of dark trading on aggregated market quality; the study is based on the largest EU equity market, the UK equity market. In a recent dark pools review by the Financial Conduct

1 For example, on an average day in October 2016, approximately €1.3 billion worth of dark transactions were executed on Chi-X’s BATS Europe (BXE) and Chi-X Europe (CXE) dark order books. This figure is based on daily updated data released on BATS Trading website in October 2016.

Authority (FCA), it was noted that “the UK equity market and regulation differs significantly from the US and other markets, including with regard to dark pools”,2 hence a comprehensive investigation of the impact of dark trading within the UK regulatory environment is called for. In carrying out this investigation, this study uses high frequency data for 282 of the largest stocks from the London equity market. The stocks used are constituents of the FTSE 350; thus, the data effectively covers the entire London market over a recent five-year period. In this paper we test the theoretical proposition that the addition of a dark trading venue to a lit trading venue improves overall/aggregate market transparency and quality. Results obtained suggest that it does in the London market once trading in all the major venues is aggregated. Specifically, when the major equity trading venues in London are consolidated into a single order book, the current levels of dark trading in the market over the five-year period between June 2010 and June 2015 enhances market transparency and reduces trading noise in the London equity market.3

Our findings have significant implications for the regulation of European markets given the potential welfare impacts of dark pools in terms of how market quality influences firm production decisions and asset allocation. This potential for impact beyond the financial markets makes it vital that we fully understand the impact of dark pools on market quality characteristics. Increased understanding of the role of dark pools is also critical for adequately crafting policies aimed at ensuring that their benefits are enhanced, while potential pitfalls, if any, are guarded against. However, the study/knowledge base needed for constructing a full picture of the market microstructure impacts of dark pools is still very much in formation at this time. There are several reasons behind the dearth of empirical

2 See the FCA Thematic Review TR16/5: UK equity market dark pools – Role, promotion and oversight in wholesale markets, published in July 2016 and available at: http://www.fca.org.uk/static/documents/thematic- reviews/tr16-05.pdf 3 Often distinctions are made among types of dark trading approaches (see for example, Foley and Putniņš, 2016). In this paper, our data and market structure information available from the examined exchanges indicate that we exclusively deal with ‘midpoint dark trading’ data.

3 academic literature on the dark pool phenomenon. Firstly, modern dark pools are still comparatively new trading venues, hence the relatively small number of studies either published or in development. Secondly, it is difficult to obtain datasets which explicitly identify trades executed via dark pools. The latter of these two issues is being rapidly resolved, as new datasets identifying dark trading activity are now emerging. This study benefits from one of such data sources.

Transparency, which should be enhanced by publicly displayed trading, is a vital ingredient in well-functioning financial markets, and aids efficient price discovery.

Therefore, it implies that venues with no pre-trade transparency, like dark pools, are potentially harmful to financial markets’ trading quality, and they might even have welfare consequences for the economy and society (see also Zhu, 2014). According to the March

2014 edition of the Thomson Reuters Equity Market Share Reporter, dark pool activity for all

European equities accounted for more than €80.23 billion in March 2014, which was about

8.50% of all traded equities on the continent that month. For the 12-month period ending

March 2014, the total value of dark trades stood at more than €898.22 billion, which represented more than 9.55% of all equity trades in Europe during that period. These are economically significant values; therefore it is unsurprising that there has been a lot of apprehension among market participants that dark pools could negatively impact trading quality. A survey conducted among market participants by the CFA Institute finds that 71% of respondents are of the view that dark pools constitute problems for price discovery in the markets (see Schacht et al., 2009).

Despite the concerns evident within the practitioner industry, academic theory suggests that the issue is a much more complicated one, and that dark pools could actually be beneficial to financial markets trading. One of the main theoretical contributors to this emerging debate, Zhu (2014), holds that under natural conditions, uninformed traders

4 gravitate towards dark pools, while informed traders largely trade on the lit markets. This manner of self-selection should result in the reduction of trading noise normally induced by noise/uninformed traders when they submit their orders to lit markets, and therefore improve overall market price discovery. However, Ye (2011), employing a different framework of informed and uninformed traders, disagrees with Zhu’s (2014) predictions. The difference in model predictions could be linked to differences in modelling approaches. While Zhu (2014) assumes that both informed and uninformed traders can freely select their trading venues,

Ye’s (2011) theoretical model gives only informed traders this selection ability.4

Over the past few years, several empirical working papers examining different aspects of dark pool trading activity have been developed. For example, Buti et al. (2011), Preece and

Rosov (2014), Ready (2014) and Brugler (2015) investigate liquidity issues with regards to dark pools. Others, such as Comerton-Forde and Putniņš (2015), Foley and Putniņš (2016) and Aquilina et al. (2016), examine price discovery-related functions of the market. It is important to note that the variations in the regulatory environments, nature of data, and dates covered make these papers distinct. Most of this literature uses US data, and where there is

UK data used it is limited, as are the time periods covered. This paper is the first large sample study of the impact of dark pools on market quality in the UK and under the MiFID regulatory environment. Although there are similarities between the EU (MiFID) regulatory regime and several others around the world, there are also significant market microstructure- impacting differences. For example, the US (RegNMS) regimes differs significantly from the

EU (MiFID) regime in terms of the publishing/reporting of vital market variables such as bid and offer prices. The order protection rule, which mandates that venues re-route orders to where they could achieve better prices execution, a cornerstone of the US regime, is also missing under the MiFID regime.

4 See Kyle (1985) for a discussion about the activities of informed and uninformed/noise traders.

Among the existing working papers on dark trading, only one, Brugler (2015), employs UK data. However, the study could not differentiate between dark and lit trades; it therefore excludes trades from venues operating both lit and dark order books. Unfortunately, however, nearly 100% of the UK’s dark trading activity occurs on three venues operating both lit and dark order books; these are Chi-X, BATS and Turquoise. Furthermore, the sample period for the study is restricted to four months, September to December 2012.

Therefore, the results are unlikely to be representative of dark trading either in the UK or in

Europe as a whole. In order to account for these issues, the results presented in this paper are based on the analysis of a 5-year dataset spanning June 2010 to June 2015. Specifically, using dark and lit trading data for FTSE 350 stocks, we extend the existing literature and make new contributions in three ways. Firstly, as a starting point, we investigate the impact of midpoint dark trading on trading transparency in the overall market. Contrary to the findings of Foley and Putniņš (2016) in the Canadian market, we find that midpoint dark trading improves market quality in the London market. Specifically, we show that moderate levels of midpoint dark trading in the overall London market enhance market transparency. We use two variables to proxy market transparency. The number of incidences of quote stuffing,5 measured when abnormal levels of quote updates are observed, is used as the proxy for market manipulation, which complicates participants’ view of the market. The Easley et al.

(1996; 1997) PIN measure is the proxy for stock-level transparency. Based on these variables, we present empirical evidence showing that moderate levels of mid-point dark pool activity in a market improve overall market transparency and decrease incidences of market manipulation. Secondly, we also investigate the impact of dark trading on the level of trading noise in the market. The results are consistent with the main findings, with trading noise reducing with increases in dark trading values.

5 Quote stuffing is a market manipulation strategy usually employed by high frequency traders. It involves rapidly submitting and cancelling large orders in order to flood the market with quotes requiring processing by competitors, and thus ensures that the competition loses their high frequency trading edge.

2. Dark Pools, Transparency and Price Discovery Efficiency

Examining the impact of dark trading on market quality and manipulation characteristics is fairly complex (see also Comerton-Forde and Putniņš, 2015). This complexity is due to the variety of effects that dark trading could have on market factors such as transparency, trading fragmentation (and segmentation) and market liquidity. Also of significance is the fact that a test for the impact of dark trading could effectively be a test for the impact of market fragmentation, thus creating a joint hypothesis dilemma. While several papers have directly addressed the question of the impact of dark trading on market quality characteristics such as liquidity and price discovery, none has examined the alleged primary impact of dark trading on an aggregate market – market opacity.

Dark transactions are less informative than lit transactions, thus on aggregate they contribute less to price discovery for the entire market. As a consequence, larger proportions of dark trading should lead to less price discovery. This is due to at least two reasons. First, the lack of pre-trade transparency means that information contained in dark orders is not available to the wider market until after the orders are executed. Secondly, information contained in dark orders is assumed to be very low to begin with due to a self-selection pattern that sees uninformed traders lean towards trading on dark pools and informed traders gravitate almost exclusively towards lit venues (see Zhu, 2014). Intuitively, the self-selection process should make the price discovery process on the lit platforms less noisy, since the proportion of uninformed/noise traders at those venues will now have decreased on account of self-selection. This explains why Comerton-Forde and Putniņš’s (2015) findings suggest that self-selection results in the reduction of trading noise in the market. With trading noise in the lit venues now reduced and the price discovery process at those venues enhanced, if the dark pools uses the more efficient price from a lit venue as reference for order execution, it is

7 logical to conclude that dark trading improves overall market quality. Thus, if a network of lit venues exists alongside a system of dark pools, it is logical to expect that aggregate market quality could be improved if traders could self-select the venue they wish to trade in based on their stock of information. For the first time, we put this expectation to the test in this paper by using the London equity market as a case study. The context described above is similar to the case in London where a network of lit and dark pools exists as a virtual market. In this construct the lit exchanges act as the main conduit for transmitting new information to the market and the f price discovery process in them is made more efficient/less noisy/more transparent by the reduction of uninformed/noisy orders migrating to the dark venues.

A ‘fully’ transparent market is deemed to be one where information is evenly distributed, therefore effectively eliminating the presence of traders with significantly superior levels of information. To put this in other words, trades based on superior information are few, or arbitrage opportunities at any frequency are scarce, in such a market.

We contend that this market scenario is one possible consequence of traders self-selecting where they would trade as a result of their information sets. Suppose a number of traders at a lit venue are ‘informed’ about the fundamental value of a tradable instrument, the value of the set of information held by the traders declines in proportion to the sum total of traders holding such information in that venue. If only a small subset of traders hold the valuable information, they stand to profit by trading with a larger subset of (uninformed) traders; however, if the situation is vice versa, the value of the information diminishes rapidly since opportunities for taking advantage of the information set is reduced. The informed traders’ orders are correlated with the underlying value of instruments traded in the market; thus the larger their number, the more prevalent is the incidence of clustering of both buy and sell orders on one side of the market, and also spreads should tighten.

From the foregoing perspective it is difficult to envisage the increase in adverse selection and worsening liquidity predicted by Zhu (2014). Furthermore, the empirical analysis by Foley and Putniņš (2016) fails to show any evidence of deteriorating liquidity in the presence of dark trading.6 However, the rise in the proportion of informed orders in the lit market could still lead to higher non-execution risks if order imbalance is severe in the market. If order imbalance holds at a reasonable level, the differences between the resulting

‘informed’ buy and sell orders could narrow, and thus adverse selection costs all but evaporate and liquidity need not be negatively impacted (see Foley and Putniņš, 2016). Thus, the values of the traded instruments reflect the buy and sell orders as placed by a significant proportion of traders in the lit market. The less informed traders in the lit market would hold even less influence in the direction of trading here, leading to a more transparent market; this is because their proportion of trades in the lit market reduces relative to the proportion of informed trades. Ultimately, adverse selection costs will decline in the overall market if this dynamic holds at a moderate level.

Since the majority of traders in the dark portions of the aggregate market are relatively uninformed, it is practical to obtain pricing information from the more informed lit venues. Dark pools as regulated in the EU for example thus derive their pricing from lit platforms serving as price references, the price generated in the dark pools are in effect also highly correlated with the underlying value of the instruments traded. Hence, we hypothesise that the operation of dark pools alongside lit venues should improve overall/aggregate market transparency, if they are operated as part of a wider market structure. Based on this view, we compute the probability that a trade is informed for each of the stock days in our sample by using the probability of informed trading measure as developed by Easley et al. (1996; 1997).

This measure is well suited to capturing the extent of market transparency and information

6 Results obtained by Comerton-Forde and Putniņš (2015), however, suggest that high levels of dark trading widen the spread.

9 asymmetry, and has already been employed as a proxy for market transparency in a number of existing studies (see as examples, Vega, 2006; Sun and Ibikunle, 2016).

Furthermore, if adding a dark pool to an existing system of lit venues could potentially improve market quality, then we should find evidence of noise reduction in the trading process such that opportunities for market manipulation will decline. The implication here is that dark trading enhances the price discovery process. Therefore, we further hypothesise (and test) that trading noise is negatively linked to dark trading.

A persistent question in the case for our underlying hypothesis is whether an extreme case, where most of the participants in a market are nearly uniformly informed, could exist in practice. We argue that such a case is not as unreal as it might appear, at least for the most active financial instruments. This is due to the high volume of firm-level and economy-wide sets of information readily available to investors and traders across the world. High frequency real time trading decisions are based on these sets of information as they become available.

From all indications, what sets day traders apart in recent times is the speed of exploiting information as it becomes available. In most advanced markets, information by itself no longer provides arbitrage opportunities for traders and investors in stocks with large analyst following – it needs to be exploited at a superior speed in order for a trader to realise the rewards for acquiring it. This market microstructure is not dissimilar to those envisaged by

Kyle (1985) and Glosten and Milgrom (1985). Kyle (1985) and Glosten and Milgrom (1985) imply that the price discovery process is made possible by the ability of a section of the market (informed traders) to acquire information and thus pass it on to the other players in the market. The informed players in the market are compensated for their information gathering activities when they earn arbitrage gains while trading with uninformed traders. Thus when information is costly, and uninformed (or noise traders) are few in the market, the price discovery process may breakdown as traders have no incentive to acquire information.

Our description of a market with a high proportion of informed traders may appear congruent with a market where a breakdown of the price discovery ensues. This is not consistent with the case presented above. The picture we paint is that of a market where the bar for profitable information is raised. Hence, while the value of information declines, faster traders will still be in a position to take advantage of the information they have acquired. By doing so, they are compensated for both their information gathering and ability to convey the information to the market (i.e. faster incorporation of information into prices). This is what has given rise to the technological arms race in financial markets (see for example, Diaz-

Rainey et al., 2015). Thus, the existence of dark trading venues only raises the level of transparency by reducing the noise levels that would have been introduced into the lit market via the presence of uninformed/noise traders (Zhu, 2014). The reduction in the volume of such traders in the lit market then serves to make the overall market more transparent, and increases the quality and volume of information in the market, ultimately making the price discovery process more efficient. Furthermore, increased proportional levels of informed traders in the lit market need not raise adverse selection costs or impair liquidity if order imbalance remains at a reasonable level. And there is no plausible reason to expect that signed order imbalance is jointly determined with market fragmentation. Thus, when uninformed traders increasingly move over to dark venues, we would expect to see as much reduction in buy orders as we observe in sell orders.

3. Policy Background to the study

Due to its competition-enhancing principles, the Markets in Financial Instruments

Directive (MiFID) enforced by the European Union in 2007 led to the proliferation and growth of alternative high-tech trading venues. MTFs are the most prominent of these alternative trading venues and are in direct competition with established national exchanges

(or regulated markets – RMs) such as the London Stock Exchange (LSE). Other alternative venue types include broker crossing networks (BCNs) and systematic internalisers (SIs). The market share held by MTFs especially has grown sharply in recent years; BATS Chi-X

Europe, a ‘paper merger’ of the Chi-X and BATS order books, is now the largest equity trading exchange in Europe by market share.7 By 30th March 2016 there were 151 MTFs listed on the CESR MiFID database managed by the European Securities and Markets

Authority (ESMA).8

Although, under MiFID guidelines, MTFs are required to publish all current bid and ask prices as well as their corresponding bid and ask sizes, the regulations allow exemptions on four grounds. MTFs rely on these exemptions to operate dark pools. The first pre-trade transparency waiver applies to large orders, which could have large market impacts if published prior to their being executed; this is referred to as the Large-in-Scale (LIS) waiver.

Qualification for a LIS requires that trades must be of a minimum size, which is based on the average daily turnover for each instrument. The minimum order size ranges from 50,000 for the least active stocks to 500,000 for the most active ones. The second waiver is the

Reference Price waiver; this is commonly used by MTFs to maintain dark pools of liquidity.

MTFs may avoid abiding with pre-trade transparency requirements if they passively match orders to a widely published reference price obtained from another market. For example,

BATS, Chi-X and Turquoise dark pools passively match orders to LSE’s posted midpoints in the case of FTSE stocks. The third waiver applies to transactions negotiated bilaterally, away from the exchanges, by counterparties. These transactions are usually non-standard and need

7 The order books are still operated separately, although they are ‘integrated’. Also, before 20th May 2013, BATS Chi-X Europe only had a licence to operate MTFs; however, since being granted Recognised Investment Exchange (RIE) status, BATS Chi-X can now operate a listing exchange as well as MTFs. The data employed in this paper is for a period straddling the transition of BATS Chi-X to RIE status. The trading dynamics and rules of the BATS and Chi-X order books employed in this analysis remain essentially the same both before and after the transition. Enquiries made with BATS Chi-X Europe confirm that their current order books are still the same as when BATS Chi-X could only operate MTFs; thus those books are still classic MTFs. As at July 2016, BATS Trading Limited is still listed on the CESR MiFID database as an MTF. However, BATS Europe Regulated Market is also now listed as a regulated market. 8 ESMA builds rule books for financial markets within the EU jurisdiction.

12 to be executed on the basis of prevailing volume weighted bid-ask spread or a reference price if the traded security is not traded continuously. The fourth waiver is for the so-called iceberg orders, and is generally referred to as the Order Management Facility waiver. MTFs can waive pre-trade transparency for orders subject to order management until they are disclosed to the market. Usually only a fraction of a submitted iceberg order is displayed, and once that portion is fulfilled, it is refreshed from the non-displayed portion.

4. Data and Descriptive Statistics

4.1. Data

In this paper, we examine the constituents of the FTSE 350 index of stocks in relation to our research questions; the FTSE 350 includes the 350 largest firms listed on the LSE.

These firms account for about 96.60% of the total market capitalisation of the FTSE All

Share index as at 30th June 2016, the final date in our dataset. All FTSE 350 stocks are traded on several trading venues, and our data consists of trading data from the four main markets where these stocks are traded – the LSE, BATS Europe, Chi-X Europe and Turquoise. The total trading volume from these four trading venues accounts for more than 95% of the FTSE

350 lit trading value. We obtain two sets of intraday tick data from the Thomson Reuters Tick

History (TRTH) database. The first is the time and sales tick data, which includes variables such as the Reuters Identification Code (RIC), date, timestamp, price, volume, bid price, ask price, bid volume, ask volume and 10 levels of bid and ask quotes. We allocate each trade a pair of corresponding prevailing best bid and ask quotes based on the quotes submission information available in the TRTH database. Since we only focus on normal trading hours, we delete the opening auction (7:50hrs – 8:00hrs) and closing auction (16:30hrs – 16:35hrs) periods from the dataset – in any case, these auctions only run on the LSE. The second set of

TRTH data obtained is the 10-level market depth quotes and corresponding sizes for each

13 transaction across the four venues. The TRTH datasets do not identify whether trades are dark or lit, hence we also obtain additional daily dark trading data from the Market Quality

Dashboard (MQD) database managed by the Capital Markets Cooperative Research Centre.9

The final set of sample data employed covers the period from 1st June 2010 to 30th June 2015.

We retain only the stocks that consistently form part of the FTSE 350 index over the sample period and for which we are able to obtain sufficient intraday data for high frequency analysis. Finally, we merge the order book level data for the four trading venues in order to create a single ‘global’ order book for the London market. The final dataset contains 1.152 billion transactions valued at 4.95 trillion British Pounds Sterling executed in 282 stocks over the sample period.

4.2. Descriptive statistics

Panel A of Figure 1 shows a logarithmic time series of trading value for the London market’s total, off-exchange and dark transactions values during the five-year period ending

June 2015. All values are in billions of pounds. The cumulative growth in dark trades appears to be tracking the total trading value for the market throughout the time series, hence the evidence here is that an increasing proportion of trades are now executed in the dark. Overall off-exchange values have also grown at a rate higher than the total market values; on aggregate they are also much higher than dark values. The dynamics are examined more closely in Panel B, which plots the dark and off-exchange values as percentages of the total market trading value. Panel B shows that when compared to dark values, there has been a fall in the proportion of trades executed off-exchange, while dark trade values continue to grow as a proportion of total market values. In June 2010, the average monthly proportion of dark trading value executed (for the stocks in our sample) in the market was 1.31%. This has since

9 In order to ensure comparability, we ascertain that variables occurring in both datasets are sufficient matches. We are fully satisfied that both datasets are very comparable and are sourced from the same trading venues.

14 increased by 350% to attain an average of 4.54% in the first six months of 2015. Over the same period, the proportion of trades executed off-exchange has fallen from 22.08% to

15.62%. Thus, it appears that the lost off-exchange trading value is mainly being shifted to the dark pools. This is important since most of the off-exchange executions are institutional and are viewed as largely liquidity-driven uninformed trades. It thus implies that dark trade is being driven by a search for institutional trading liquidity (see also Ibikunle, 2016). Overall, however, as shown in Panel C, changes in overall lit and dark values appear to be in lockstep throughout the five-year period.

INSERT FIGURE 1 ABOUT HERE

Figure 2 plots the average trade sizes of lit and dark transactions in the London market. The first point to note is that, consistent with the U.S. markets (see Chordia et al.,

2011), there has been a general fall in average trade sizes for both lit and dark trades. Early on in the plot, dark transactions were generally smaller in size than lit transactions; however, they have since become steadily comparable to the lit ones in the last year of the period under review. The trend is perhaps connected with the growth in dark trades as well as the suspected identity of some key actors in the dark pools – institutional traders.

INSERT FIGURE 2 ABOUT HERE

5. Methodology

In order to examine the impact of dark trading on transparency, we first identify two proxies for the concept. Specifically, we select the probability of informed trading (PIN) measure as developed by Easley et al. (1996; 1997) and develop a new proxy for the number

15 of Quote Stuffing Incidences (QSI) in the stocks across all venues as proxies. PIN is a suitable proxy for trading transparency and has been used in previous studies for the same purpose (see Vega, 2006). The number of quote stuffing incidences, which simply captures incidences of abnormally large quotes across the day, could be intuitively interpreted as a measure of the number of times market participants attempt to ‘muddy the trading waters’.

This property of the incidences makes the number of times they are issued a good indication of the relative state of market transparency. Thus trading can be expected to be more transparent on days with lesser numbers of QSIs than on those with larger numbers of QSIs.

As a robustness check, we also compute a measure of the level of noise in each venue relative to the rest of the market and aggregate the estimates for each venue across the entire market, in order to investigate the impact of dark trading on the efficiency of the price discovery process. This measure is called Relative Noise Avoidance (RNA). Our estimation measures of PIN, QSI and RNA are formally defined and discussed below.

5.1. Main Variables

5.1.1. PIN: Inverse proxy for market transparency

Apart from being an inverse proxy for market transparency, PIN could also serve as a proxy for adverse selection cost, especially since it is generally regarded as being strongly correlated with adverse selection risk and information asymmetry (see for example Chung and Li, 2003; Brown et al., 2009). PIN has also been employed as a proxy for priced information risk and information asymmetry in the wider financial economics literature (see for example Vega, 2006; Ellul and Pagano, 2006; Duarte et al., 2008; Chung and Li, 2003).

Recently Lai et al. (2014) compute PIN measures for 30,095 firms across 47 countries over a

15-year period, and conclude that PIN is highly correlated with firm-level private information.

Based on the foregoing, we adopt PIN as an inverse proxy of the daily levels of market transparency. The model is based on the expectation that trading between informed traders, liquidity traders and market makers occurs across the day. Trading commences with the informed traders acquiring a private signal regarding a stock’s value with a probability of

α. Bad news based on private signal will arrive with a probability of δ, and good news based on private signal arrives with a probability of (1 – δ). The market makers generate their bid and ask prices with orders arriving from liquidity traders at the arrival rate ε. Should new private information become available, informed traders will join the trading process, with their orders arriving at the rate μ. It is expected that informed traders will execute a purchase trade if they have acquired a good news signal, and sell should the signal be bad news. We note that the allocation of different arrival rates for uninformed buy and sell orders does not qualitatively change estimations of the probability that an informed trade has been executed

(see Easley et al., 2002).

The PIN model enables us to approximate the unobservable distribution of trades between informed and uninformed traders by modelling observable buy and sell orders.10

Hence, the ‘usual level’ of sales and purchases executed within a stock on a given day over several trading cycles is interpreted as relatively uninformed aggregate trading activity by the model, and this information is employed to estimate ε. Unusual levels of purchase or sale transactions are interpreted as information-based transactions and used to estimate μ. In addition, the frequency of intervals during which ‘abnormal’ levels of purchases and sales are observed is used to estimate the values of α and δ. These estimations are executed in a simultaneous fashion using maximum likelihood estimation method. If we assume that the uninformed and informed trades arrive as a Poisson distribution, the likelihood function for the PIN model for each interval estimated can be expressed as:

10 We infer purchase and sales through the running of Lee and Ready’s (1991) trade classification algorithm.

 B S  B (   )S L((B,S) | )  (1 )eb b es s  eb b e(s ) s B! S! B! S! (1) S (   )B  (1 )es s e(b ) b S! B! where B and S respectively correspond to the total number of purchase and sale transactions for each one hour trading period within each trading day. θ = (α, δ, μ, ε) is the parameter vector for the structural model. Equation (1) represents a system of distributions in which the possible trades are weighted by the probability of a one hour trading period with no news (1 –

α), a one hour trading period with good news (α (1 – δ)), or a one hour trading period with bad news (αδ). Based on the assumption that this process occurs independently across the different trading periods, Easley et al. (1996; 1997) estimate the parameter vector using maximum likelihood estimation procedure. Thus parameters are obtained for each trading day and for each stock in the sample by maximum likelihood estimation. Consistent with Easley et al. (1996; 1997), PIN is computed as:

 PIN    2

5.1.2. QSI: Quote Stuffing Incidences

The quote stuffing variable is simply a measure of the number of times an

abnormal level of quotes updates is observed during a minute interval across the trading

day, at all the venues included in our dataset. An abnormal level of quotes updates is

defined as when a per minute quotes update count is at least one standard deviation above

the mean number of quote updates calculated for that minute using the surrounding period

over days (–30, +30).11

11 For robustness, (–15, +15) and (–60, +60) are also employed with very little variation in the results. The MQD database also has a measure for quote stuffing, called quote stuffing alerts; the time panel yielded by our measure is highly and positively correlated with the MQD measure. Again for robustness, we apply this measure in our regression analysis and the results are not materially altered.

5.1.3. Relative Avoidance of Noise

There are two well-established traditional approaches for measuring the price discovery contributions of different markets/venues; these are Hasbrouck’s (1995) information share (IS) and the Gonzalo and Granger’s (1995) common factor share (CFS).

Essentially, both approaches arise from a vector error correction model (VECM) and aim to decompose price innovations into permanent and transitory components. Yan and Zivot

(2010) and Putniņš (2013), however, contend that correctly allocating a market’s share of price discovery contribution only holds for both measures if the competing venues have similar noise levels. When there are differences in the noise levels, IS and CFS measure to varying degrees a combination of speed of impounding information and relative avoidance of noise, rather than price discovery. According to Putniņš (2013), CFS measures mainly the relative avoidance of temporary shocks in the pricing process and thus allocates price discovery dominance to the venue with lower noise levels. We therefore extend the CFS measure to compute a new variable measuring the relative noisiness of the price discovery process in various venues; the individual stock-venue values for each day are then aggregated across all the venues in order to obtain a global value for the London equity market.

For each stock day we estimate the following VECM using mid-point of quotes12 at one-second intervals, t:

p p v v v v g v g v Pt  0  1 Pt1  Pt1  i Pti  i Pti  t , i1 i1 p p (2) g g g v g g v g Pt  0  1 Pt1  Pt1  i Pti  i Pti  t . i1 i1

12 Our decision to employ quotes here rather than trades is consistent with the literature on decomposing of pricing innovations (see as examples, Huang, 2002; Hupperets and Menkveld, 2002; Theissen, 2002). It is also a valid approach because dark trades are matched against the midpoint of prevailing quotes from lit platforms rather than execution prices.

v g where Pt and Pt correspond to the log of the mid-point of the last bid and ask prices from venue v’s (LSE, Chi-X, BATS or Turquoise) order book and a constituted ‘global’ order book made up of quotes from the three other venues respectively. The CFS values for both price series are then computed following Baillie et al. (2002). For each venue the RNA measure for venue v on day t is given as:

CFS v RNAv  CFS g (3)

where the CFSv and CFSg are the component factor shares for venue v (LSE, Chi-X,

BATS or Turquoise) and the combined other three venues respectively. The higher venue v’s

RNA value is on the day, the lower its noise levels on that day. A venue with a higher RNA has lower noise levels (and thus a higher level of pricing efficiency) than the competing venues.

5.2. Multivariate analysis

The methodological empirical approach employed involves computing a series of stock-day panel estimations relating the market quality variables to dark trading activity and other control variables. Panel estimations are run for all the 282 stocks using their global values as calculated based on the individual exchange values from Chi-X Europe, LSE,

BATS Europe and Turquoise. In order to obtain the global values, following Sun and

Ibikunle (2016), we combine the order book data from the four exchanges and create a single order book using time stamps provided in the TRTH data. For example, in computing the PIN variable, we combine high frequency order book data for the four venues, then compute PIN on the basis of the newly generated order book. The panel estimations are done in two ways:

(i) two-stage least squares (2SLS) instrumental variables (IV) regressions; and (ii) panel generalised method of moments (GMM). PCSE errors are computed in order to obtain

20 heteroscedasticity and autocorrelation robust standard errors. The IV approach models are employed in order to account for the likelihood of the endogeneity of dark and off-exchange trading values.13

The panel regression estimated is of the following form:

6 (4) Qit   DARK DARKit OFFEX OFFEX it HFT HFTit  k Ckit  it k1 where Qit corresponds to PIN, QSI or RNA for stock i on day t, DARKit is the percentage of the stock-day’s total pound trading volume executed in the dark, while OFFEXit is the percentage of the stock-day’s total pound volume of trades executed away from the four

14 exchanges’ downstairs venues for stock i on day t. HFTit serves as a proxy for algorithmic trading and is measured as the ratio of messages to trades for stock i on day t. Ckit is a set of k control variables which includes log of market capitalisation, log of average trade size, log of pound volume of lit trades, and log of effective spread, which is defined as twice the absolute value of the difference between transaction price and prevailing mid-point, the quadratic term

2 DARK it, which is added due to a prediction of a trade-off in the positive and negative effects of fragmentation (see Degryse et al., 2015) and dark trading in the US (see Preece and Rosov,

2014). Since migration to dark pools implies a migration to new trading platforms, accounting for this possible trade-off is prudent. Pother stocks, which is the average of the dependent variable (PIN, QSI or RNA) on the same day for all the other stocks in the same size quintile, is also included to account for commonality across stocks.15

13 Simple Panel Least Squares estimations with stock and time fixed effects are also employed for the IV estimations are more likely to yield statistically significant results; however, the explanatory powers of the models are very comparable. The identical nature of the results hence imply that, just as reported by Comerton- Forde and Putniņš (2015), endogeneity appears not to significantly affect the one-stage least squares estimation results. 14 We also use the log of pound volume of dark and off-exchange trades for stock i on day t as measures of dark and off-exchange trading respectively. The results obtained from this variation are qualitatively similar to the ones presented in the paper. 15 In the GMM IV estimations, a two-stage least squares (2SLS) weighting is used to weigh the various GMM moment conditions to be satisfied by the parameters of interest. The moment conditions are restricted to those that could be written as an orthogonality condition between the residuals of Equation (4) and its right-hand side variables.

With regards to identifying suitable instruments for DARKit and OFFEXit, the usual conditions on good instrument candidates apply: (1) instruments must be highly correlated with the variable to be instrumented, and (2) be largely uncorrelated with eit in Equation (4) above. We employ two sets of IVs. The first set involves using an approach that has become increasingly popular in the recent literature; this is the approach first proposed by Hasbrouck and Saar (2013) and thereafter used by several others such as Buti et al. (2011) and Degryse et al. (2015). Specifically, the variables are instrumented using the average level of dark and off-exchange trading in stocks of similar market capitalisation. In this paper, stocks in the same average daily trading value (in pounds sterling) decile are used for instrumenting a stock’s dark or off-exchange trading value. The second set of IVs are computed based on an approach developed by Ibikunle (2016). In his paper, Ibikunle (2016) aims to maximise the potential for instrument-error term orthogonality by extending the Hasbrouck and Saar (2013) method. Firstly, the initial averages of the trading variables across stocks in the same decile are used in a panel least squares framework, which regresses each of the endogenous variables on their corresponding cross-sectional stock averages and the other control variables. The residuals from this step are then employed as IVs in the GMM estimation. The

IVs obtained satisfy the two conditions stated above to a very high degree. According to

Ibikunle (2016), the reason for the lack of correlation between the IVs and Equation (1)’s residuals is that the common cross-sectional components in the stock averages have been

‘exhausted’ in explaining the changes in the endogenous variables, thus leaving only the stock-dependent factors not explained by the cross-sectional average.

Descriptive statistics for all the employed variables and a correlation matrix for the independent variables are reported in Table 1 and Table 2 respectively. Table 1 shows that large cap stocks generally have the larger trading values across the board, except in the case of average trade sizes. Average trade sizes are shown to be larger in low cap stocks due to

22 their being mainly executed off the main lit order books; Armitage and Ibikunle (2015) show that low volume stocks in the London market are usually executed via upstairs broker-dealer arrangements and then subsequently reported to the exchanges. Further estimates show that about 3,051 lit trades and 140 dark trades are executed in an average stock each day during the sample period. As shown in Figure 1, dark value substantially lags lit and off-exchange values.

Table 2 shows the correlations between the independent variables used in the multivariate analysis. Consistent with Buti et al. (2011) and Comerton-Forde and Putniņš

(2015), the effective spread is seen to be negatively correlated with dark value and off- exchange value. Thus it appears that dark trading serves a similar purpose in the UK, the US and Australia. However, in contrast to previous studies, dark trading is negatively correlated with HFT, the algorithmic trading proxy. This is not surprising since there are fewer advantages to algorithmic trading in the dark (if any) than in the lit markets.

INSERT TABLE 1 AND TABLE 2 ABOUT HERE

6. Results and discussions

The purpose of this paper is to test two central hypotheses: (1) whether dark trading enhances overall market transparency, and (2) whether dark trading contributes to trading noise decline in the London equity market. Proxies employed are discussed in Section 5.

6.1. Dark trading and market transparency

Table 3 presents the results for both the 2SLS SET 1 IV and the GMM SET 2 IV regression results in Panel A and Panel B respectively. For the two sets of results, estimates are presented for the full sample of 282 stocks as well as for each quintile of stocks. In both

23 panels, all of the coefficients for the full sample are statistically significant, and all are at least significant at the 0.001 level, except in the case of the off-exchange value coefficient (0.1 level in Panel A). Most of the coefficient estimates for all of the quintile regressions are also statistically significant, especially when the estimates reflect the expected relationship of the relevant variables with PIN. As predicted, in both sets of results in Table 3 the relationship between PIN and dark value is negative for the full sample (coefficient –0.12 and highly significant with a p-value of –28.163). This implies that increasing levels of dark trading are linked with improvements in market transparency and an elimination of adverse selection risk and information asymmetry (see Chung and Li, 2003; Brown et al., 2009). Thus, based on the estimates in Table 3, we find a reason to support Zhu’s (2014) prediction that the addition of a dark pool to a lit platform improves market quality, due to self-trading venue selection of informed and uninformed traders. The results are also consistent with the suggestion that the self-selection by traders results in declining trading noise in the market (see Comerton-Forde and Putniņš, 2015). As earlier stated, we believe that the inverse relationship between information asymmetry and dark trading is due to the increasing proportion of informed traders in the lit order books, as the uninformed participants migrate to dark order books. This is a valid conclusion in the context of the dark pools we examine in this paper; the dark pools use the lit venues as a price reference point, specifically by matching trades to the lit venues’ mid-points. This relationship implies, and is shown by Ibikunle (2016), that most of the information released into the London market arises from the lit venues, and that dark trades by themselves are not considered as part of the overall market infrastructure impair price discovery (see also Comerton-Forde and Putniņš, 2015).

INSERT TABLE 3 ABOUT HERE

However, the positive Dark2 estimate in Panel A (0.004 and also very significant with a p-value of 36.19) suggests that the negative relationship between PIN and dark trading is quadratic in nature, i.e. there is a level of dark trading in the market which could hurt overall market transparency. This view is not dissimilar to the previously reported between relationship of trading fragmentation and market quality characteristics (see for example,

Degryse et al., 2015; Sun and Ibikunle, 2016). The fragmentation studies are relevant because the migration of trades from lit venues to dark pools in the London market also implies a fragmentation of the trading process. Hence, our main result here is consistent with the existing literature on the effects of market fragmentation on market quality. And this is backed up by the negative coefficient estimates reported for the off exchange value coefficient in both Panel A and Panel B of Table 3. In Figure 3, we plot estimates of the implied effect of dark trading on the inverse proxy for aggregate market transparency, PIN.

Panel A, showing the plot for the full sample, suggests that the turning point of the positive effect of dark trading on market quality is when dark trading is at about 15% of the total trading value in the market. Thus, when dark trading is at about 15% in the market, its impact on aggregate market transparency is zero. However, this single threshold value can be misleading in a market with a large number of stocks with varying levels of liquidity; hence we also plot estimated values for average daily pound volume quintiles as well. As suspected, the threshold estimate varies across quintiles, from 5% for the highest trading stocks to 18% for the lowest trading stocks. The estimates for Quintiles 4, 3 and 2 stocks are 10%, 17% and

17% respectively. Thus, the lowest trading stocks have a higher threshold than the highest trading ones.

Other results are of interest in Table 3. For example, there appears to be asymmetric effects of trade size on market transparency, depending on stock size. While for the overall sample and for the lowest trading stocks (Quintile 1 and Quintile 2) the effect is positive,

25 implying that increasing large trade sizes lead to declining market transparency, the opposite is the case for the highest trading stocks (Quintile 4 and Quintile 5), and all the estimates are statistically significant. This asymmetry can be explained by looking to different streams of the microstructure literature. Early literature contributions (see as examples, Easley and

O'Hara, 1987; Chan and Lakonishok, 1993) suggest that informed traders use large order sizes to take advantage of their informed positions, while more recent studies argue that informed traders prefer to break their orders into smaller pieces to camouflage their trading intentions (see as examples, Admati and Pfleiderer, 1988; Barclay and Warner, 1993;

Chakravarty, 2001). Thus, since large cap stocks are more likely to have order flows large enough to act as camouflage for informed trades, informed traders in large cap stocks are more likely to deploy the latter strategy in those stocks, and for the stocks with lower order flow levels, they may use large orders to convey information to the market.

Also, as expected, the effective spread coefficient estimates are positive and statistically significant for the full sample and for all the quintiles in both Panel A and Panel

B. This is because we expect a rise in adverse selection costs as the spread widens, and thus the probability of informed trading which is positively correlated with adverse selection should also rise with widening spreads. The Pother stocks coefficient is also positive and statistically significant for all but the smallest cap stocks, indicating a high level of commonality across stocks. In relation to the effects of high frequency trading, measured as the ratio of messages to executed trades, the impact is slightly more nuanced. However, consistent with the recent literature, there is sufficient evidence in the results to suggest that the overall impact of high frequency trading is that it improves market transparency. This impact is due to HFTs speeding up the price discovery process (Brogaard et al., 2014;

Chaboud et al., 2014).

For completeness, we also proxy market transparency by computing the number of times abnormal levels of quote updates (QSI) are observed for each stock across the four order books. The log of QSI is then employed as a dependent variable in the two sets of IV regressions as explained above. The results for the full sample and the quintiles are presented in Panel A and Panel B of Table 4. The results for both panels are highly consistent; hence the discussion of Table 4 focuses on one panel. In Panel A the coefficient estimates and the directions of relationships with the quote stuffing variable are as expected. Dark value is found to be negatively related to QSI, and the relationship is statistically significant (–0.49, significant at 0.0001 level). There is no advantage to stuffing quotes in a dark trading environment, since the state of the order book is not observable. Therefore, as the proportion of dark trading in the aggregate market rises, one would expect a reduction in HFT activity and, by extension, a reduction in the incidences of quote stuffing. And consistent with the

PIN analysis, the Dark2 estimate suggests that the relationship is quadratic. However, the results for the largest cap stocks suggest that the opposite relationship holds. This is a reasonable expectation in cases where quote stuffing incidences are fewer relative to the overall size of the order flow. Trading in large cap stocks is characterised by a high rate of order submission, and thus large stocks are less likely to report cases of quote stuffing based on the measure we employ, hence the inconsistency in the statistically significant estimates. It is also important to note that while HFT is seen as enhancing price discovery by speedily eliminating order imbalances and arbitrage opportunities,16 it is found here to be positively related with quote stuffing. This is expected, given that quote stuffing is intuitively more likely in a high frequency trading environment.

INSERT TABLE 4 ABOUT HERE

16 For example, in the literature, the overall impact of HFT is usually found to be positive for large cap stocks (see Chaboud et al., 2014; Hendershott et al., 2011).

Other interesting results include the coefficient estimates for trade size, which is negative and statistically significant for the full sample. The quintiles’ estimates are also negative, but are not statistically significant. This is an expected result because quote stuffing should be easier with smaller trade sizes than with larger ones. Thus as trade sizes decline the propensity for quote stuffing rises, leading to a less transparent market. The effective spread estimates also suggest that quote stuffing thrives when liquidity declines in the market, but not necessarily when trading volume does (positive and significant lit value coefficient estimates for the full sample, medium sized and small sized stocks). As expected, quote stuffing appears to be more prevalent when HFT activity is on the rise – all the HFT coefficient estimates are high (ranging from 0.78 to 1.28) and, except for the largest cap stocks, are also highly significant. Given the negative Pother stocks coefficient estimate, quote stuffing incidences appear to exhibit low commonality across stocks. As shown in Table 1, there is a large variation of quote stuffing incidences across stock sizes.

6.2. Dark trading and trading noise

For robustness, we also test a second hypothesis regarding how dark trading impacts the quality of trading in the London market. The second hypothesis is that, on balance, if dark trading enhances transparency in overall London equity market, then we should also see a reduction in the overall noise levels in the market. Thus, we expect dark trading to be positively related to RNA, a measure of the extent of avoidance of trading noise in one venue relative to its competing venues. A positive relationship implies that in the overall market, increasing dark trade levels lead to declining noise levels, since a rise in RNA values means a fall in market trading noise. Table 5 presents the results for the 2SLS IV regression results testing the relationship of dark trades and other variables with RNA. Consistent with our

28 expectations, and with the PIN and quote stuffing regressions, coefficient estimates for dark value suggest reduction in noise levels in the presence of dark trading. The dark value coefficient is 0.099 and is statistically significant at the 0.0001 level.

INSERT TABLE 5 ABOUT HERE

The result implies that overall trading noise falls by about 10% with every unit rise in dark trading value in the London market. This result is also in line with what we should find if dark trading does really enhance transparency in the overall market as a result of traders self-selecting, trading venue-wise. Similar to this analysis, Comerton-Forde and Putniņš

(2015) investigate the impact of dark trading on informational efficiency in the Australian market. The proxies they use include Autocorrelation and Variance Ratio, which are intuitively similar to our approach. Their results are consistent with ours, as is the insight that the relationship of informational efficiency with dark trading is conditioned on the level of dark trading. As with previous results in Table 3 and Table 4, the Dark2 estimate (–0.004) takes an opposite sign to the dark value coefficient estimate, and the estimate is statically significant. Comparing our results to Zhu (2014) predictions, that operating a dark venue alongside a lit venue will, in equilibrium, enhance price discovery, we find ample evidence that the theoretical prediction holds. However, in line with Comerton-Forde and Putniņš

(2015) findings, our results suggest that at a certain point higher levels of dark trading start to harm market quality.

7. Conclusion

In this study we test the central hypotheses arising from recent theoretical works on the impact of dark trading on market quality. Zhu (2014) argues that the self-selection

29 hypothesis implies that price discovery is enhanced by the addition of a dark pool to a lit exchange. Our results, based on a very large European sample of the 350 largest UK stocks traded across the four main trading venues in London, suggest that the overall market benefits when dark trading occurs at low percentage levels. Using data from mid-point dark order books as well as lit order books, we find that the overall market becomes more transparent as dark trading increases at moderate levels. However, we also observe a non-linear relationship between our proxies of market transparency and dark trading, which implies that at higher levels, dark trading could harm market quality. Also, in line with the positive relationship between dark trading and market transparency, overall trading noise and incidences of market manipulation in the market declines as dark trading rises. These results support the notion that, when given an option to trade in the dark, uninformed traders are more likely to do so than informed traders. Hence, the availability of dark pools provides an opportunity of segmenting trader-types in the market. With the lit venues having a higher percentage of well-informed traders than in a scenario where dark pools do not exist, there arises a higher likelihood of timely information incorporation, and since the trades executed in the dark are based on reference prices as yielded by the lit exchanges, the overall market’s price discovery process is more efficient for each stock traded simultaneously in the dark and lit venues.

The UK and, indeed, European financial markets infrastructures are at a critical point with the imminent implementation of MiFID II/MiFIR, a package of regulations which partly aims to cap the volume of dark trading. This study, suggesting that moderate levels of dark trading is beneficial, is therefore timely and could have significant policy implications.

MiFID II/MiFIR emphasises investor protection and more transparency in financial markets.

We have shown in this study that the latter aim is furthered by the existence of dark pools operating alongside lit exchanges. It is important that policy makers take care not to eliminate the market quality benefits of dark trading by arbitrarily imposing dark trading restrictions.

Careful analysis of the factors determining the levels at which dark trading becomes harmful for each market and instrument type across the EU should be undertaken, in order to craft helpful policies.

References Admati, A. & Pfleiderer, P. (1988) A theory of intraday patterns: volume and price variability. The Review of Financial Studies, 1(1), 3-40.

Aquilina, M., Foley, S., O'Neill, P. & Ruf, T. (2016) Asymmetries in Dark Pool Reference Prices. Financial Conduct Authority, Occasional Paper 21. London.

Armitage, S. & Ibikunle, G. (2015) Price Discovery in London's Upstairs Equity Market. Univeristy of Edinburgh Business School Working Paper. Edinburgh.

Baillie, R. T., Geoffrey Booth, G., Tse, Y. & Zabotina, T. (2002) Price discovery and common factor models. Journal of Financial Markets, 5(3), 309-321.

Barclay, M. J. & Warner, J. B. (1993) Stealth trading and volatility: Which trades move prices? Journal of Financial Economics, 34(3), 281-305.

Brogaard, J., Hendershott, T. & Riordan, R. (2014) High-Frequency Trading and Price Discovery. Review of Financial Studies, 27(8), 2267-2306

Brown, S., Hillegeist, S. A. & Lo, K. (2009) The effect of earnings surprises on information asymmetry. Journal of Accounting and Economics, 47(3), 208-225.

Brugler, J. (2015) Into the Light: Dark Pool Trading and Intraday Market Quality on the Primary Exchange. Working paper.

Buti, S., Rindi, B. & Werner, I. M. (2011) Diving into dark pools. Working paper.

Chaboud, A. P., Chiquoine, B., Hjalmarsson, E. & Vega, C. (2014) Rise of the Machines: Algorithmic Trading in the Foreign Exchange Market. The Journal of Finance, 69(5), 2045-2084.

Chakravarty, S. (2001) Stealth-trading: Which traders' trades move stock prices? Journal of Financial Economics, 61(2), 289-307.

Chan, L. K. C. & Lakonishok, J. (1993) Institutional trades and intraday stock price behavior. Journal of Financial Economics, 33(2), 173-199.

Chordia, T., Roll, R. & Subrahmanyam, A. (2011) Recent trends in trading activity and market quality. Journal of Financial Economics, 101(2), 243-263.

Chung, K. H. & Li, M. (2003) Adverse-Selection Costs and the Probability of Information- Based Trading. Financial Review, 38(2), 257-272.

Comerton-Forde, C. & Putniņš, T. J. (2015) Dark trading and price discovery. Journal of Financial Economics, 118(1), 70-92.

Degryse, H., de Jong, F. & Kervel, V. v. (2015) The Impact of Dark Trading and Visible Fragmentation on Market Quality. Review of Finance, 19(4), 1587-1622.

Diaz-Rainey, I., Ibikunle, G. & Mention, A.-L. (2015) The technological transformation of capital markets. Technological Forecasting and Social Change, 99(277-284.

Duarte, J., Han, X., Harford, J. & Young, L. (2008) Information asymmetry, information dissemination and the effect of regulation FD on the cost of capital. Journal of Financial Economics, 87(1), 24-44.

Easley, D., Hvidkjaer, S. & O'Hara, M. (2002) Is Information Risk a Determinant of Asset Returns? The Journal of Finance, 57(5), 2185-2221.

Easley, D., Kiefer, N. M. & O'Hara, M. (1996) Cream-Skimming or Profit-Sharing? The Curious Role of Purchased Order Flow. The Journal of Finance, 51(3), 811-833.

Easley, D., Kiefer, N. M. & O'Hara, M. (1997) One Day in the Life of a Very Common Stock. The Review of Financial Studies, 10(3), 805-835.

Easley, D. & O'Hara, M. (1987) Price, trade size, and information in securities markets. Journal of Financial Economics, 19(1), 69-90.

Ellul, A. & Pagano, M. (2006) IPO Underpricing and After-Market Liquidity. Review of Financial Studies, 19(2), 381-421.

Foley, S. & Putniņš, T. J. (2016) Should we be afraid of the dark? Dark trading and market quality. Journal of Financial Economics.

Glosten, L. R. & Milgrom, P. R. (1985) Bid, ask and transaction prices in a specialist market with heterogeneously informed traders. Journal of Financial Economics, 14(1), 71- 100.

Gonzalo, J. & Granger, C. W. J. (1995) Estimation of common long-memory components in cointegrated systems. Journal of Business and Economic Statistics, 13(1), 27-35.

Hasbrouck, J. (1995) One Security, Many Markets: Determining the Contributions to Price Discovery. The Journal of Finance, 50(4), 1175-1199.

Hasbrouck, J. & Saar, G. (2013) Low-latency trading. Journal of Financial Markets, 16(4), 646-679.

Hendershott, T., Jones, C. M. & Menkveld, A. J. (2011) Does Algorithmic Trading Improve Liquidity? The Journal of Finance, 66(1), 1-33.

Huang, R. D. (2002) The Quality of ECN and Nasdaq Market Maker Quotes. The Journal of Finance, 57(3), 1285-1319.

Hupperets, E. C. J. & Menkveld, A. J. (2002) Intraday analysis of market integration: Dutch blue chips traded in Amsterdam and New York. Journal of Financial Markets, 5(1), 57-82.

Ibikunle, G. (2016) Trading Places: Price Leadership, Informed Trading and the Competition for Order flow. University of Edinburgh Business School Working paper. Edinburgh.

Kyle, A. S. (1985) Continuous Auctions and Insider Trading. Econometrica, 53(6), 1315- 1335.

Lai, S., Ng, L. & Zhang, B. (2014) Does PIN affect equity prices around the world? Journal of Financial Economics, 114(1), 178-195.

Lee, C. M. C. & Ready, M. J. (1991) Inferring Trade Direction from Intraday Data. The Journal of Finance, 46(2), 733-746.

Preece, R. & Rosov, S. (2014) Dark Trading and Equity Market Quality. Financial Analysts Journal, 70(6), 34 - 48.

Putniņš, T. J. (2013) What do price discovery metrics really measure? Journal of Empirical Finance, 23(0), 68-83.

Ready, M. J. (2014) Determinants of Volume in Dark Pool Crossing Networks. Working paper. Madison, Wisconsin.

Schacht, K., Cronin, C., Allen, C. J. & Preece, R. (2009) Market Microstructure: The Impact of Fragmentation under the Markets in Financial Instruments Directive. Charlottesville, VA: CFA Institute Centre for Financial Market Integrity.

Sun, Y. & Ibikunle, G. (2016) A Decade in the Life of a Market: Visible Market Fragmentation, Quality and Efficiency. University of Edinburgh Working Paper.

Theissen, E. (2002) Price discovery in floor and screen trading systems. Journal of Empirical Finance, 9(4), 455-474.

Vega, C. (2006) Stock price reaction to public and private information. Journal of Financial Economics, 82(1), 103-133.

Yan, B. & Zivot, E. (2010) A structural analysis of price discovery measures. Journal of Financial Markets, 13(1), 1-19.

Ye, M. (2011) A glimpe into the dark: price formation, transaction costs and market share of crossing network. University of Illinois Working paper.

Zhu, H. (2014) Do Dark Pools Harm Price Discovery? Review of Financial Studies, 27(3), 747-789.

Figure 1: Trading values

Panel A plots the total (dark, lit and off-exchange), off-exchange and dark pound trading values for 282 FTSE 350 stocks trading simultaneously on the four main London ‘City’ exchanges/trading venues; these are the London Stock Exchange, BATS, Chi-X and Turquoise between 1st June 2010 to 30th June 2015. Panel B plots the pound values for dark and off-exchange trading as percentages of total market value. Panel C plots aggregate dark and lit trading values for the same period, and for the same stocks and venues. PANEL A

PANEL B

PANEL C

Figure 2: Average trade sizes

The figure plots the average pound sizes per day of lit and dark trades for 282 FTSE 350 stocks trading simultaneously on the four main London ‘City’ exchanges/trading venues; these are the London Stock Exchange, BATS, Chi-X and Turquoise, between 1st June 2010 to 30th June 2015.

Figure 3: Effects of Dark Trading on Market Quality

The figure plots the implied/estimated effects of dark trading on market quality, with the probability of an informed trade (PIN) used as an inverse proxy for market transparency/quality/adverse selection risk. Panels A – F show the full sample and five quintile of stocks, starting with the highest trading quintile to the least trading one. The estimated effects are obtained from stock-day panel regressions as outlined in Table 3. The sample consists of 282 FTSE 350 stocks trading simultaneously on the four main London ‘City’ exchanges/trading venues; these are the London Stock Exchange, BATS, Chi-X and Turquoise, between 1st June 2010 to 30th June 2015. Panel A: Full sample

Panel B: Quintile 5 (stocks with the highest average daily trading value in pounds sterling)

Panel C: Quintile 4 stocks

Panel D: Quintile 3 stocks

Panel E: Quintile 2 stocks

PANEL F: Quintile 1 stocks

Table 1: Descriptive statistics This table reports daily mean values along with medians in parentheses per stock for 282 FTSE 350 stocks trading simultaneously on the four main London ‘City’ exchanges/trading venues; these are the London Stock Exchange, BATS, Chi-X and Turquoise. The sample period covers 1st June 2010 to 30th June 2015. The quintiles are computed on the basis of daily pound volume across the sample period. HFT is a proxy for algorithmic trading and is measured as the ratio of messages to trades. QSI (Quote Stuffing Incidences) is the number of times when per minute quotes update count in a stock is at least one standard deviation above the mean number of quote updates calculated for that minute using the surrounding period over days (–30, +30). Relative Noise Avoidance is based on the component factor share (CFS) of Gonzalo and Granger (1995) and is computed as the ratio of a stock venue’s CFS to that of the other competing venues. The CFS metric is computed using quote mid-point at 1-second intervals. PIN is the Easley et al. (1996, 1997) probability of informed trading measure computed from the parameters yielded by maximising the following likelihood function: (T)B (T)S (T)B ((  )T)S L((B, S) | )  (1 )eT eT  eT e()T B! S! B! S! (T)S ((  )T)B  (1 )eT e()T , S! B! where B and S respectively correspond to the total number of buy and sell orders for the day within each trading interval. θ = (α, δ, μ, ε) is the parameter vector for the model. α corresponds to the probability of an information event, δ is the conditional probability of a low signal of an information event, μ is the arrival rate of informed orders, and ε is the arrival rate of uninformed orders. The probability that a trade is informed for each stock and within each interval is then computed as:  PIN  .  2 Highest cap Lowest cap Variables Full sample 4 3 2 stocks stocks

3050.47 9507.43 3268.75 1487.96 592.82 148.88 Number of lit trades (1156.00) (6644.00) (2670.00) (1221.00) ( 427.00) (83.00 )

139.51 436.44 150.36 63.27 27.98 8.14 Number of dark trades (35.00) (276.00) (100.00) (35.00 ) (12.00) (1.00)

8.07×107 2.82×108 6.74×107 2.39×107 9.06×106 7.97×106 Dark value (£) (1.01×107) ( 1.52×108 ) ( 3.4×107 ) (7.66×106 ) (1.97×106) ( 2.91×105 )

4.52×108 1.53×109 3.05×108 1.44×108 1.14×108 9.21×107 Off-exchange value (£) (9.92×107) ( 7.92×108 ) ( 1.76×108 ) ( 6.80×107 ) ( 5.10×107) ( 2.81×107)

959.23 984.49 761.75 726.67 1228.39 1114.62 Trade size (£) (458.47) (573.30) (496.61) (436.20) (445.34) (327.70)

3.28×1010 1.32×1011 1.36×1010 6.54×109 3.72×109 1.91×109 Market cap (£) ( 6.67×109) ( 5.78×1010) ( 1.22×1010) ( 6.53×109) ( 3.68×109) ( 2.12×109)

1.14 0.35 0.58 1.11 1.59 2.15 Effective spread (bps) ( 0.55 ) ( 0.26) ( 0.40) ( 0.67) ( 1.06) ( 1.58)

1.26×109 4.53×109 9.16×108 2.79×108 1.26×108 1.32×108 Lit value (£) ( 2.04×108) (2.86×108 ) ( 6.02×108 ) ( 1.48×108) ( 6.21×107) ( 2.11×107)

536.55 840.29 727.29 502.07 330.06 201.83 HFT (ratio) ( 194.61 ) ( 221.78 ) ( 216.28) ( 160.96) ( 132.82) ( 132.75)

0.25 0.20 0.20 0.24 0.27 0.32 PIN ( 0.22) ( 0.17) ( 0.17) ( 0.21) (0.24) ( 0.30)

4760.32 8506.13 5423.57 4221.53 3683.04 3821.36 QSI ( 1817 ) ( 3277.5 ) ( 2116.5 ) ( 1667.5) ( 1549.5 ) ( 1388)

Relative Noise 1.47 1.43 1.46 1.49 1.52 1.53 Avoidance (ratio) ( 1.41) ( 1.38) ( 1.41) ( 1.43) ( 1.45) ( 1.44)

Table 2: Correlation matrix The table reports correlations between independent variables calculated using trading data for 282 FTSE 350 stocks trading simultaneously on the four main London ‘City’ exchanges/trading venues; these are the London Stock Exchange, BATS, Chi-X and Turquoise. Dark value is the percentage of the stock-day’s total pound volume executed in dark pools in the London market, and off-exchange value is the percentage of the stock-day’s total pound volume executed away from the downstairs of London’s main four exchanges/trading venues. Lit value is the pound volume of all trades executed on the limit order books of the main four exchanges/trading venues in London. HFT is a proxy for algorithmic trading and is measured as the ratio of messages to trades, and effective spread is computed as twice the absolute value of the difference between transaction price and prevailing mid-point. Market cap, trade size, effective spread (bps) and lit value are in logs. The sample period covers 1st June 2010 to 30th June 2015. The quintiles are computed on the basis of average daily trading value in pounds sterling across the sample period.

Off-exchange Effective Dark value Trade size Market cap Lit value HFT value spread

Dark value 1 0.62 0.63 0.63 –0.67 0.38 –0.49

Off-exchange value 0.62 1 0.76 0.64 –0.52 0.68 –0.23

Trade size 0.63 0.76 1 0.66 –0.27 0.56 0.03

Market cap 0.63 0.64 0.66 1 –0.49 0.69 –0.09

Effective spread –0.67 –0.52 –0.27 –0.49 1 –0.72 0.56

Lit value 0.38 0.68 0.56 0.70 –0.72 1 –0.49

HFT –0.49 –0.23 0.03 –0.09 0.56 –0.49 1

Table 3: The impact of dark trading on market transparency The table reports regression coefficient estimates using a stock-day panel, in which the dependent variable is the Easley et al. (1996, 1997) probability of an informed trade (PIN) measure computed as outlined in Table 1 for 282 FTSE 350 stocks trading simultaneously on the four main London ‘City’ exchanges/trading venues; these are the London Stock Exchange, BATS, Chi-X and Turquoise. The estimated regressions is: 6

PINit   DARK DARKit OFFEX OFFEX it HFT HFTit  k Ckit  it k1 where DARKit is the percentage of the stock-day’s total pound trading volume executed in the dark, while OFFEXit is the percentage of the stock-day’s total pound volume of trades executed away from the four exchanges’ downstairs venues for stock i on day t. HFTit is a proxy for algorithmic trading and is measured as the ratio of messages to trades for stock i on day t. Ckit is a set of k control variables which includes log of market capitalisation, log of average trade size, log of pound volume of lit trades, log of effective spread, which is defined as twice the absolute value of the difference between transaction price and prevailing mid-point, the square of DARKit and Pother stocks. Pother stocks is the average of PIN on the same day for all the other stocks in the same sized quintile. Panel A and Panel B report 2SLS and GMM estimations, respectively, using two different sets of instrumental variables (IVs). DARKit and OFFEXit are instrumented in the 2SLS estimation by using the average level of dark and off-exchange trading in stocks in the same market cap quintile respectively. In the GMM estimation IVs are obtained for DARKit and OFFEXit by first collecting the within-quintile/full sample cross- sectional averages of the trading variables. DARKit and OFFEXit are then each individually regressed on their corresponding cross-sectional stock averages and the other control variables in a panel least squares framework; the residuals yielded by this estimation are each employed as corresponding IVs for DARKit and OFFEXit. The t- statistics are presented in parentheses and derived from panel corrected standard errors (PCSE). *, ** and *** correspond to statistical significance at 0.1, 0.05 and 0.01 levels respectively. The sample period covers 1st June 2010 to 30th June 2015. The quintiles are computed on the basis of average daily trading value in pounds sterling across the sample period. PANEL A

Probability of an informed trade

Highest cap Lowest cap Variables Full sample 4 3 2 stocks stocks –0.121*** –0.073 –0.054 –0.008 0.017 0.036 Dark value (–28.163) (–1.543) (–1.465) (–0.239) (0.442) (0.845)

0.004*** 0.002* 0.001 –0.000 –0.001 –0.001 Dark2 (36.190) (1.890) (1.340) (–0.019) (–0.656) (–0.650)

Off-exchange –0.001* 0.006*** 0.014*** 0.008*** –0.001 –0.010*** value

(–1.881) (2.648) (6.836) (3.432) (–0.593) (–3.217)

0.019*** –0.016*** –0.017*** –0.003 0.018*** 0.019*** Trade size (15.325) (–2.820) (–3.898) (-0.579) (2.907) (2.834)

–0.001*** –0.004*** –0.002*** –0.005*** –0.001 –0.002** Market cap (–3.248) (–2.863) (–2.638) (–3.980) (–1.403) (–2.530)

0.011*** 0.016*** 0.027*** 0.017*** 0.014*** 0.009*** Effective spread (24.766) (5.082) (13.074) (9.205) (8.173) (5.470)

–0.016*** –0.012** 0.006 0.003 –0.009** –0.027*** Lit value (–16.779) (–2.227) (1.623) (0.593) (–1.996) (–6.477)

–0.008*** –0.023** –0.015** 0.010 –0.025*** –0.009 HFT (–5.674) (–2.336) (–2.305) (1.172) (–2.733) (–0.981)

0.004*** 0.020*** 0.007* 0.011*** 0.013*** –0.000 Pother stocks (12.178) (2.712) (1.885) (4.376) (4.527) (–0.118)

1.218*** 1.060** 0.478 0.261 0.238 0.374 Intercept (39.404) (2.248) (1.517) (0.992) (0.883) (1.234)

R2 0.157 0.021 0.010 0.011 0.013 0.019

PANEL B

Probability of an informed trade

Highest cap Lowest cap Variables Full sample 4 3 2 stocks stocks –0.043*** –0.030*** 0.015 –0.004 –0.020** –0.009 Dark value (–3.73) (–2.873) (1.144) (–0.458) (–2.458) (–1.485)

0.001*** 0.001*** –0.000 0.000 0.001*** 0.000 Dark2 (2.17) (2.917) (–0.989) (0.857) (2.920) (1.300)

Off-exchange –0.001*** 0.001 –0.002* –0.001 –0.001 –0.002** value (–4.171) (1.327) (–1.943) (–0.718) (–0.794) (–2.563)

0.037*** –0.001 –0.002 0.001 0.004* 0.019*** Trade size (48.621) (–0.439) (-0.855) (0.534) (1.782) (9.319)

–0.004*** –0.003*** –0.000 –0.005*** –0.002* –0.002* Market cap (–17.032) (–3.406) (–0.204) (–4.035) (–1.929) (–1.884)

0.012*** 0.011*** 0.023*** 0.018*** 0.013*** 0.008*** Effective spread (26.334) (5.963) (12.40) (10.158) (8.646) (6.248)

–0.017*** –0.004*** –0.003** –0.009*** –0.015*** –0.016*** Lit value (–60.622) (–3.284) (–2.500) (–8.472) (–16.301) (–18.985)

–0.005*** 0.001 0.003* –0.003** –0.005*** –0.011*** HFT (–14.260) (0.737) (1.731) (–2.254) (–3.537) (–9.661)

Pother stocks 0.011*** 0.021*** 0.009** 0.012*** 0.011*** 0.001

(28.155) (3.566) (2.474) (5.175) (4.718) (0.509)

–0.029 0.470*** 0.197* 0.521*** 0.658*** 0.469*** Intercept (–1.267) (4.450) (1.663) (6.970) (9.289) (8.412)

R2 0.169 0.010 0.010 0.020 0.028 0.039

Table 4: Dark trading and market manipulation The table reports regression coefficient estimates using a stock-day panel, in which the dependent variable is QSI, defined as the number of times that per minute quotes update count in a stock is at least one standard deviation above the mean number of quote updates calculated for that minute using the surrounding period over days (–30, +30), for 282 FTSE 350 stocks trading simultaneously on the four main London ‘City’ exchanges/trading venues; these are the London Stock Exchange, BATS, Chi-X and Turquoise. The estimated regressions is: 6 log(QSI)it     DARK DARK it  OFFEX OFFEX it   HFT HFTit  k Ckit  it k1 where DARKit is the percentage of the stock-day’s total pound trading volume executed in the dark, while OFFEXit is the percentage of the stock-day’s total pound volume of trades executed away from the four exchanges’ downstairs venues for stock i on day t. HFTit is a proxy for algorithmic trading and is measured as the ratio of messages to trades for stock i on day t. Ckit is a set of k control variables which includes log of market capitalisation, log of average trade size, log of pound volume of lit trades, log of effective spread, which is defined as twice the absolute value of the difference between transaction price and prevailing mid-point, the square of DARKit and Pother stocks. Pother stocks is the average of PIN on the same day for all the other stocks in the same sized quintile. Panel A and Panel B report 2SLS and GMM estimations using two different sets of instrumental variables (IVs). DARKit and OFFEXit are instrumented in the 2SLS estimation by using the average level of dark and off-exchange trading in stocks in the same market cap quintile respectively. In the GMM estimation IVs are obtained for DARKit and OFFEXit by first collecting the within-quintile/full sample cross-sectional averages of the trading variables. DARKit and OFFEXit are then each individually regressed on their corresponding cross-sectional stock averages and the other control variables in a panel least squares framework; the residuals yielded by this estimation are each employed as corresponding IVs for DARKit and OFFEXit. The t-statistics are presented in parentheses and derived from panel corrected standard errors (PCSE). *, ** and *** correspond to statistical significance at 0.1, 0.05 and 0.01 levels respectively. The sample period covers 1st June 2010 to 30th June 2015. The quintiles are computed on the basis of average daily trading value in pounds sterling across the sample period. PANEL A

Quote Stuffing Incidences (QSI)

Highest cap Lowest cap Variables Full sample 4 3 2 stocks stocks –0.492*** 2.988*** 1.918 0.658 –2.419 1.407 Dark value (–4.230) (2.597) (1.306) (0.718) (–0.862) (1.270)

0.007*** –0.074** –0.043 –0.018 0.061 –0.049 Dark2 (2.601) (–2.182) (–1.155) (–0.696) (0.864) (–1.446)

Off-exchange 0.454*** 0.313** 0.011 0.127 0.210 0.177* value (15.094) (2.438) (0.115) (1.468) (1.577) (1.937)

–0.523*** –0.113 –0.187 –0.138 –0.126 –0.082 Trade size (–8.141) (–0.155) (–0.991) (–0.587) (–0.474) (–0.328)

0.064*** –0.008 –0.026 0.142*** 0.033 0.118*** Market cap (6.541) (–0.041) (–0.445) (2.967) (0.856) (3.876)

0.158*** 0.227 0.149 0.214*** 0.244* 0.141** Effective spread (7.413) (0.769) (1.529) (2.842) (1.694) (1.967)

0.518*** 0.194 –0.063 0.151** 0.547 0.389*** Lit value (13.569) (0.965) (–0.372) (1.961) (1.266) (3.414)

0.881*** 1.086 1.280*** 0.777*** 0.861*** 0.820*** HFT (15.452) (1.093) (6.518) (3.935) (4.094) (3.876)

–0.033*** 0.415 –0.047 –0.006 0.093 0.159*** Pother stocks (–4.509) (1.116) (–0.389) (–0.103) (1.320) (2.715)

–3.649*** –35.766** –14.456 –8.716 13.036 –18.869* Intercept (–4.351) (–1.973) (–1.287) (–1.073) (0.767) (–1.897)

R2 0.085 0.477 0.335 0.470 0.230 0.266

PANEL B

Quote Stuffing Incidences (QSI)

Highest cap Lowest cap Variables Full sample 4 3 2 stocks stocks –0.041 0.035 –0.018 0.019 –0.379** –0.238 Dark value (–1.006) (0.108) (–0.058) (0.066) (–2.016) (–1.326)

0.006*** –0.000 0.003 0.003 0.015*** 0.012** Dark2 (4.085) (–0.010) (0.369) (0.367) (2.649) (2.158)

Off-exchange 0.025*** 0.198*** 0.077*** –0.039 0.020 0.019 value (3.682) (5.972) (2.841) (–1.293) (0.837) (0.882)

–0.047*** –0.055 0.002 –0.135** –0.219*** –0.323*** Trade size (–2.682) (–0.649) (0.028) (–2.131) (–4.017) (–6.663)

0.064*** 0.048* 0.070*** 0.092*** 0.040** 0.091*** Market cap (10.272) (1.886) (2.928) (3.392) (2.337) (5.201)

0.097*** 0.275*** 0.277*** 0.186*** 0.133*** 0.109*** Effective spread (10.649) (6.269) (7.993) (5.094) (4.206) (3.930)

0.199*** 0.282*** 0.143*** 0.075*** 0.108*** 0.108*** Lit value (29.786) (6.550) (4.927) (2.701) (4.749) (4.974)

0.671*** 0.820*** 0.819*** 0.802*** 0.883*** 0.825*** HFT (79.001) (16.702) (19.481) (21.936) (27.644) (27.977)

-0.044*** 0.084 0.037 -0.032 0.107** 0.119*** Pother stocks (-7.934) (0.498) (0.376) (-0.533) (2.427) (3.237)

–1.908*** –6.626* –2.820 1.687 4.987*** 3.799** Intercept (–4.295) (–1.839) (–0.894) (0.647) (2.864) (2.195)

R2 0.470 0.553 0.463 0.493 0.521 0.453

Table 5: Dark trading and noise in the price discovery process The table reports 2SLS regression coefficient estimates using a stock-day panel, in which the dependent variable is Relative Noise Avoidance (RNA), a measure based on the component factor share (CFS) of Gonzalo and Granger (1995), and is computed as the ratio of a stock venue’s CFS to that of the other competing venues, for 282 FTSE 350 stocks trading simultaneously on the four main London ‘City’ exchanges/trading venues; these are the London Stock Exchange, BATS, Chi-X and Turquoise. The estimated regressions is: 6

RNAit     DARK DARK it  OFFEX OFFEX it   HFT HFTit  k Ckit  it k1 where DARKit is the percentage of the stock-day’s total pound trading volume executed in the dark, while OFFEXit is the percentage of the stock-day’s total pound volume of trades executed away from the four exchanges’ downstairs venues for stock i on day t. HFTit is a proxy for algorithmic trading and is measured as the ratio of messages to trades for stock i on day t. Ckit is a set of k control variables which includes log of market capitalisation, log of average trade size, log of pound volume of lit trades, log of effective spread, which is defined as twice the absolute value of the difference between transaction price and prevailing mid-point, the square of DARKit and Pother stocks. Pother stocks is the average of PIN on the same day for all the other stocks in the same sized quintile. DARKit and OFFEXit are instrumented by using the average level of dark and off- exchange trading in stocks in the same market cap quintile respectively. The t-statistics are presented in parentheses and derived from panel corrected standard errors (PCSE). *, ** and *** correspond to statistical significance at 0.1, 0.05 and 0.01 levels respectively. The sample period covers 1st June 2010 to 30th June 2015. The quintiles are computed on the basis of average daily trading value in pounds sterling across the sample period.

Relative Noise Avoidance

Highest cap Lowest cap Variables Full sample 4 3 2 stocks stocks 0.099*** 2.807** 1.821** 0.827 0.227 –0.017 Dark value (4.790) (2.330) (2.095) (1.535) (0.509) (–0.042)

–0.004*** –0.072** –0.049** –0.023 –0.008 0.000 Dark2 (–6.119) (–2.291) (–2.105) (–1.535) (–0.663) (0.011)

0.027*** 0.054** 0.056*** 0.036*** 0.061*** 0.036*** Off-exchange value (7.441) (2.279) (5.770) (3.843) (5.203) (3.835)

0.037*** –0.122 –0.031** –0.000 0.039** 0.033 Trade size (7.055) (–1.410) (–1.988) (–0.042) (2.387) (1.507)

0.005*** 0.032** 0.008** 0.003 0.008 0.022*** Market cap (5.882) (2.355) (2.523) (0.304) (0.874) (6.043)

0.064*** 0.156*** 0.084*** 0.077*** 0.060*** 0.048*** Effective spread (17.071) (6.453) (8.091) (8.232) (5.087) (5.033)

0.007** –0.137 –0.007 –0.005 0.028** 0.016** Lit value (2.463) (–1.439) (–0.555) (–0.605) (2.272) (2.240)

0.102*** 0.125** 0.109*** 0.149*** 0.151*** 0.077*** HFT (12.531) (2.489) (5.854) (5.921) (5.822) (3.746)

0.005* –0.142** 0.103*** 0.045*** 0.005 0.050*** Pother stocks (1.690) (–2.523) (2.737) (2.600) (0.220) (5.759)

–1.896*** –30.125** –17.788** –8.360* –4.175 –1.686 Intercept (–7.922) (–2.188) (–2.097) (–1.651) (–1.033) (–0.454)

R2 0.214 0.145 0.172 0.020 0.127 0.133