Author Draft for Review Only

O nl y

The Effects of Algorithmic Trading on Security Market Quality

I

A

ut

ho

r

D

ra

ft

fo

r

is professor of economics and finance at Wake Forest University’s School of Business in WinstonSalem, NC. [email protected]

mplementation of the Markets in Financial Instruments Directive 1 (MiFID1) in 2007 transformed the European equity markets. Although not designed to promote algorithmic trading (AT), the new trading rules regarding best execution, market transparency, and market organisation resulted in precisely that outcome. At the exchange level, infrastructure development to reduce latency supported increased AT. For example, in 2011 the London Stock Exchange (LSE) introduced MillenniumIT, which reduced latency below 120 microseconds, making LSE one of the fastest trading venues in the world. Later the same year, the LSE Group fully acquired Turquoise, whose platform achieves latencies of 90 microseconds. The already substantial volume of AT in London soon exploded. This article investigates the marketquality effects of AT. Assessments of security market quality should include a focus on market fairness or integrity as well as on market efficiency. Aitken and Harris [2012] propose a market-quality framework in which five elements of market design (technology, information, participants, regulation, and instruments traded) are related to market manipulation, information leakage, and broker-agency conf lict as well as to trading costs and price discovery efficiency. This study develops metrics of trade-based manipulation and information leakage ahead of price-

SPRING 2015

JOT-HARRIS.indd 1

sensitive announcements in London and Paris equities and then provides a systems estimation for assessing the simultaneous effect on manipulation, leakage, and effective spreads four years before and after MiFID1. Our empirical research on security market quality has policy implications for algorithmic trading. High-frequency trading (HFT) is a subset of AT; not all computerdetermined trading operates through HFT servers. For example, institutional trading desk instructions often initiate computerized execution programs that scan multiple platforms to slice and dice large portfolio rebalancing orders, to execute pairs trades, or to conduct hedging strategies without involving HFT servers. Just as important, manipulation and insider trading surveillance alert data can often be shown to be artifacts of complex algorithmic trading strategies not demonstrative of illegal intent. After a sixfold increase in AT following the implementation of MiFID1, we find that the average trade size, trade value, and the bid-ask spread declined significantly in both the London and Paris national exchanges. Employing a 3SLS estimation, we find a reduced incidence of both information leakage and market manipulation with greater AT. The results on spreads are more mixed: greater AT lowered effective spreads in Paris but only in the top-quintile stocks. In London, effective spreads in all

ev

DE

R

FREDERICK H. B. H ARRIS

ie w

FREDERICK H. DE B. HARRIS

THE JOURNAL OF T RADING

1

3/9/15 10:21:16 AM

R

ev

ie w

O nl y

instead, HFT realize profits from small multivenue price deviations in hundreds and sometimes thousands of intraday transactions accomplished in a few milliseconds. But in addition to HFT, AT is routinely employed in agency trading of large institutional orders, where it is designed to stealthily capture liquidity, to reduce order instruction footprints, or simply to minimize implementation shortfall for a given price impact. AT often employs fast order cancellation and replacement to arbitrage between competing execution channels based on real-time forecasting models of the stock-specific state of the market. AT is therefore distinguished from pure low-latency trading, which seeks to profit on a f leeting informational advantage in acquiring and sending faster message traffic to exchange servers without optimization or real-time forecasting involved. In a market environment where computer algorithms are competing against one another without human intervention, market integrity remains crucial to achieving security market quality. Trading algorithms have agency duties to clients and ethical duties to society as a controlable extension of their designers and operators. This perspective ref lects the explicitly dualistic nature of regulatory mandates in practically every capital market worldwide to assure fairness as well as enhance efficiency. Grossman and Stiglitz’s [1980] noisy rational expectations equilibrium for efficient markets established that competitive security market prices do not simply aggregate continuously all the available information relevant to the value of a security. Instead, efficient security market prices reveal a noisy signal. These real-time observed prices differ from the security’s full-information value by a zero mean random disturbance, the variance of which depends on the precision of the informed traders’ information, on the fraction of informed traders in the marketplace, and on the degree of risk aversion. As a consequence, observed security prices executable in algorithmic trading strategies must be informationally efficient not just in the static sense of a mean price that “contains all relevant information” but also and more crucially in the dynamic sense of a real-time price discovery process. The dynamics of price discovery are not an add-on attachment to other more central concepts, they are the sine qua non concept of informationally efficient markets.

D

ra

ft

fo

r

stocks declined with greater AT. In addition, in June of 2009 one of the principal MTFs facilitating AT, Chi-X, sharply reduced the explicit transaction costs of settlement in London, thereby further enhancing transactioncost efficiency. Our research highlights the simultaneous interrelationships between market integrity and market efficiency and seeks to estimate the magnitude of tradeoffs in market quality that arise from particular market designs. An increase in market manipulation raises price volatility, which reduces order aggressiveness, leading to higher effective spreads. Ceteris paribus, higher spreads reduce the incidence of manipulation, especially in more thinly traded stocks, because of the higher costs of trading. Finally, both integrity violations and trading costs will inf luence the execution channel decision to employ AT. And given AT’s capabilities in assessing the real-time state of the market, an increase in AT may lead to a reduced potential to manipulate the close, greater liquidity, and lower spreads. By estimating market integrity, market efficiency, and algorithmic trading as a system, we are able to isolate and control for changes in market design across exchanges and over time that would otherwise confound the estimated effects of AT. In this article, we highlight two such market design differences over the 20032011 sample period: (1) MiFID1 and (2) the LSE hybrid market, with its worked principal agreements by London broker-dealers for large institutional orders intertwined with an electronic limit order book versus the NYSE Euronext’s fully electronic limit order book supplemented for less liquid stocks by liquidite provideurs.

ho

r

ALGORITHMIC TRADING STANDARDS IN NOISY RATIONAL EXPECTATIONS EFFICIENT MARKETS

A

ut

AT trading decisions are predesigned by tradeoptimization search engines, and submissions are automated and executed without human intervention. AT is employed by proprietary traders to accomplish rapid order submissions and cancellations, often accentuated by co-locating servers beside an exchange. A broad range of AT trading strategies is employed from pseudo market-making to pinging for pockets of undisplayed liquidity to statistical arbitrage by quantitative hedge funds. High-frequency traders lose money on positions taken for more than five seconds (Menkveld [2013]); 2

THE EFFECTS OF A LGORITHMIC T RADING ON SECURITY M ARKET QUALITY

JOT-HARRIS.indd 2

SPRING 2015

3/9/15 10:21:16 AM

R

ev

ie w

O nl y

message traffic is exceedingly costly, and regulators and exchange operators (for example, on the LSE) are thoroughly warranted in passing along pro rata fees to recover the direct fixed costs of a trading desk’s weekly or monthly message traffic. But low fill-to-order ratios triggered by high incidence cancellations are optimizing and nondeceptive behavior. Liu [2009] has shown this to be axiomatic when spreads for market-making are small and/or information monitoring costs are low. Low fillto-order ratios accompanying the HFT’s extraordinary rates of cancellation are just what one would expect in an informationally efficient market. As to the mandates of regulators, perceived fairness is almost universally recognized as crucial to the effective functioning of equity markets. In contrast, in bond markets financial professionals encounter one another in repeated games of negotiated price with extremely asymmetric information. Under those circumstances, reputation effects assure credible commitments as to incomplete and unaudited but largely true information revelation. In bond market equilibrium, haircuts off the asking price for bonds vary with the reputation of the known counterparty. But not so in traditional equity markets, where traders enter into largely anonymous transactions at fixed prices in one-shot games. The information in equity markets is certainly not more asymmetric, but these radically different conditions of exchange necessitate disclosure regulation of audited financial statements. Without such preconditions for assuring fairness, quoted price equity markets with trader anonymity cannot function effectively to price and transfer risk at minimum transaction cost. To the extent that competition between competing equity algorithms parallels bond market exchange characteristics, a stronger argument can be made for somewhat different standards for assuring fairness in equity transactions. Long before trade-based manipulation or front running of client orders becomes rampant, the implicit transaction costs to do an equity market round trip increase. Aitken et al. [2014] show that spreads are 6% –11% higher across the world’s equity markets when the incidence of trade-based manipulation at the close doubles. In addition, when deceptive practices such as trade-based manipulation and spoofing are allowed, artificially induced price volatility also rises. And as we argued earlier, with manipulation displacing the market equilibrium price or spoofing falsely signaling an artificial volume, price discovery efficiency is directly

A

ut

ho

r

D

ra

ft

fo

r

Nobel laureate Clive Granger’s (Gonzalo and Granger [1995]) concept of the common factor share has defined price-discovery efficiency in light of permanent versus transitory decompositions in vector error correction models (VECM). An execution channel or trading algorithm is said to be more price-discovery efficient if its prices chase fewer transitory shocks and instead impound more permanent stochastic trends than a competing channel or algorithm. So recognizing noisy rational expectations (RE) equilibrium in competitive markets has expanded the definition of efficient markets to include metrics of not only bid-ask spreads representing transaction costs and variance ratios representing information precision but also price discovery efficiency (PDE) representing informational efficiency itself (Yan and Zivot [2011]). In a VECM of co-integrated price series, PDE is the relative impact multiplier of prices in a competing execution channel to transitory shocks versus the impact multiplier in the focal channel to transitory shocks (Harris et al. [2014]). Thus, PDE provides evidence of what senior traders call “bad trades” that inhibit price discovery because they chase liquidity-based transitory shocks that contain no new information about permanent stochastic trends. The significance of noisy RE modeling and the new concept of PDE is that trading algorithms designed to deceive competing algorithms about the permanence of a transitory shock inhibit discovery of the true information-based price. Lower price discovery efficiency available in that channel or marketplace necessarily emerges from such deception. And lower PDE is a basis for prohibiting those trading practices that give effect to this intended deception. The prohibited trading behaviors of ramp-and-dump manipulation and spoofing are two practices that fit easily into this category of intentionally deceptive trading practices that inhibit price discovery. Whether the computer algorithms can be said to intend to deceive or not, the designers and operators have created a trading practice that inhibits price discovery by deceiving other traders into misperceiving a false trend in prices or volumes. The pivotal role of the price discovery process in the functioning of a noisy RE equilibrium in efficient competitive markets warrants a security market policy that prohibits these deceptive trading practices. On the other hand, the cancellation of executable limit orders does not fit into this category of intentionally deceptive trading practices that inhibit price discovery. Server infrastructure to support explosively growing

SPRING 2015

JOT-HARRIS.indd 3


3

3/9/15 10:21:16 AM

A

ut

ho

r

D

ra

ft

ie w

fo

r

R

ev

RELATED LITERATURE ON HFT AND AT

In a theoretical model of HFT executed through an electronic limit order book, Cvitanic and Kirilenko [2010] show that the introduction of HFT lowered the duration between trades, reduced the average trade value, and moved the distribution of prices closer to the mean, resulting in reduced volatility. A key assumption was that high-frequency traders act as uninformed market-makers. In contrast, Jovanovic and Menkveld [2010] construct a theoretical model in which ATs are more informed than their counterparts. Such ATs would reduce spreads since they are better able to detect stale orders and avoid adverse selection; so on the one hand, ATs can improve informational efficiency. However, the enhanced speed of execution of AT reduces the willingness of human participants to enter the market, reducing liquidity. Jovanovic and Menkveld [2010] theorize that such a situation could widen spreads especially in the top liquidity deciles and thereby reduce market efficiency. Hasbrouck and Saar [2013] conclude that increased AT improves other traditional market-quality measures such as short-term volatility, spreads, and displayed depth. Both Brogaard [2010] and Jarnecic and Snape [2010] find that higher HFT results in a reduction in the bid-ask spread. Hendershott et al. [2011] find that higher AT lowers both the effective and the realized spread. In contrast, Hendershott and Moulton [2011] show that lower-latency trading systems on the New York Stock Exchange (NYSE) increase spreads but reduce noise, making prices more efficient. Hendershott et al. [2011] address how AT impacts information asymmetry. Using the volume of message traffic (normalized by the number of trades) as a proxy,

they find AT significantly reduces both quoted and effective spreads as a result of a decline in price impact. In short, better and faster assessments of the “state of the market” information reduce the adverse-selection component of the spread. Consistent with the results of Henderschott et al. [2011, 2013], McInish and Upson [2013] find that AT increases the profits to liquidity providers (the realized spread) relative to slow traders not employing algorithmic trading. Zhang [2010] finds greater HFT has worsened price discovery by focusing price adjustment dynamics too often on transitory shocks ref lecting order imbalance as opposed to permanent price moves ref lecting changes in fundamental value. Predatory market stuffing by HFTs to congest message traffic and prevent cancellations of large institutional orders (that have been sliced and diced but are leaving footprints) would illustrate this view. But using Gonzalo-Granger’s time-series decomposition techniques (Yan and Zivot [2011]), Brogaard et al. [2014] show that HFT trades are more price-discovery efficient in trading in the direction of permanent price moves rather than in the direction of transitory shocks. Given the dominance of algorithmic and HFT traders especially prevalent on Chi-X, tighter quoted spreads on Chi-X complete the picture of a more informationally efficient marketplace. Our research asks the further question whether increased AT has enhanced market integrity as well.

O nl y

inhibited. These are the bedrock dimensions of market efficiency—transaction costs, price volatility, and price discovery. Perceived unfairness in equity trading is detrimental to all three. So why would society wish to prohibit algorithmic trading practices that seek not just to mask a trader’s footprints but rather to actively mislead a competing trader’s algorithm? The answer is simple—to not do so means that spreads will rise, induced volatility will increase, and price discovery will be inhibited. In short, the equity markets will necessarily become less efficient than they could have been if these deceptive practices were prohibited, surveilled, reported, and prosecuted.

4

DATA

Using a proprietary NASDAQ dataset that directly identifies 26 HFTs over several periods during 2008, 2009, and 2010, Brogaard [2010] investigates whether HFTs f lee the market during times of heightened volatility. He finds that when prices f luctuate more than normal, HFTs supply more liquidity and demand less. Spending more time at the NBBO, but providing less depth, HFTs are better able to avoid trading at stale prices, and their presence improves pricing efficiency. Using trade and quote data in the millisecond environment to identify strategic runs of trades, Hasbrouck and Saar [2013] observe interactions between traders separated by as little as three milliseconds. These strategic runs are used to create a cancellation-based proxy for the level of HFT. Unlike the very short-term studies by Jarnecic and Snape [2010], Hasbrouck and Saar analyze a month-long period of high market stress in 2007


JOT-HARRIS.indd 4

SPRING 2015

3/9/15 10:21:16 AM

and another in 2008. Dividing the day into 10-minute intervals and applying a two-stage simultaneous equation system, they allow for an endogenous relationship between a cancellation-to-trade (CTR) ratio and volatility, depth, and spreads. Their CTR proxy for AT is found to lower volatility, increase quoted depth, and reduce quoted spreads, even in times of market stress. We adopt this method of identifying AT using as a proxy the cancel-to-trade (CTR) ratio. For all listed securities on the LSE and Euronext, we retrieve the bid and ask quotes (time stamped to the nearest millisecond) for up to 10 levels of the limit order book from 2003 through 2011. The data come from Thompson Reuters Tick History (TRTH) and to construct our trading-ahead measure of market integrity, we also use Reuters News Announcement (RNA) feed. We access intraday information related to market prices, volume, number of level 1 orders/trades, as well as all changes to the level 10 order book.

R

ev

ie w

O nl y

In defining our proxy measure of AT, we focus on cancellation orders rather than total message traffic. Specifically, the cancellations that reduce only depth are compared to the number of trades to form a cancelto-trade ratio (CTR). To determine the number of order cancellations, we match trade executions against the limit order book immediately prior to trades. When an incoming marketable order executes against more than one standing limit order, multiple messages are generated for each standing limit order. Following the methodology employed in Hasbrouck and Saar [2013], we combine into a single order all marketable order arrivals within the same millisecond and in the same direction that are unbroken by any nonexecution message. Exhibit 1 shows the sixfold increase in CTR from the onset of MiFID1 in November 2007 to mid-2010. By fragmenting the market and introducing a passport rule, MIFID1 was expected to increase HFT/AT and therefore CTR. The pre-MiFID mean CTR rose from 3.07 to 17.91 post-MiFID in London (see Exhibit 2,

EXHIBIT 1

A

ut

ho

r

D

ra

ft

fo

r

Cancel-to-Trade (CTR) Ratio, a Proxy for Algorithmic Trading

SPRING 2015

JOT-HARRIS.indd 5


5

3/9/15 10:21:16 AM

MARKET-QUALITY METRICS OF INTEGRITY AND EFFICIENCY

Panel A) and from 1.42 to 8.01 in Paris (see Exhibit 2, Panel B), both significant at 1%. Cutting the data into end-of-quarter (EOQ) expiry days in Panels C and D (when some integrity violations are more numerous), CTR is again much higher post-MiFID1 introduction (3.82 to 17.57 in London and 1.33 to 7.58 in Paris). Subsequently, Exhibit 1 shows that CTR returned to its pre-MiFID1 levels once the LSE planned and then introduced a sliding scale cost-recovery fee (in November 2010) and even earlier in Paris where the French planned for several years and then introduced a transaction tax (from August 2011-2012).

Market Integrity: Information Leakage

O nl y

Measurement of LEAKAGE begins by identifying unusual price and volume behavior prior to price-sensitive announcements. First, we calculate: Abnormal Returnit = Returnit − βi (

mt

)

(1)

ie w

where Returni is the daily return on stock i, βi is the stock-specific correlation with the market return con-

EXHIBIT 2 Descriptive Statistics

ev

We present descriptive statistics for each of our variables. MTC (marking the close) are instances of end-of-day prices that are dislocated by trade-based manipulation. The price change during the last 15 minutes is compared with its historical distribution over the previous 30 trading days to identify cases that are outliers of more than three standard deviations. Then from this sample a subset that has a price reversion of 50% or more in the first 15 minutes the next trading day are selected.

fo

r

R

LEAKAGE is the ratio of the number of information leakages to the number of clean event windows ×100. An event is the occurrence of an information announcement, an event window comprises the days t − 6 through t + 2 around the announcement, and a clean event window is an event window with no other information announcements. Abnormal returns are calculated for each announcement as the difference between the individual stock return for an event window and the return on the market index. We retain a sample of price-sensitive announcements, which are announcements for which the abnormal return is more than three standard deviations from the mean abnormal return for the 250-day base period ending at t − 10.

ra

ft

SPREAD, relative effective spread, is 200 times the trade price times the trade direction indicator (1 for buys and −1 for sells) times the difference between the trade price and the quote midpoint relative to the quote midpoint in basis points. Our proxy for high-frequency trading, AT, is the ratio of the number of cancellations to the number of trades (CTR). After eliminating reductions in depth due to trade executions, we classify the remaining quote updates that only reduce depth as cancellations.

A

ut

ho

r

D

OTR is the L10 order book’s order-to-trade ratio, a metric of order instruction message traffic. For each variable for each panel, we test the null hypothesis of equality of means, pre- and post-MiFID1, using a Wilcoxon rank sum difference in means test. ∗∗ and † signify significant difference at 95% and 99%.

6


JOT-HARRIS.indd 6

SPRING 2015

3/9/15 10:21:17 AM

(Continued)

ev

ie w

O nl y

EXHIBIT 2

where day t occurs between event day −6 and −1,

and where there are no days in which

A

ut

ho

r

D

ra

ft

fo

r

structed over the benchmarking period and Returnm is the daily market return. Our metric is based on pricesensitive company announcements, where an announcement is regarded as price sensitive when the return of the underlying security between days t − 6 and t + 2 is more than three standard deviations (3 σi ) away from the average seven-day abnormal return ( dy l) constructed using a bootstrap procedure during a 250trading-day benchmarking period ending at t − 10. In addition to a market model of returns, our metric uses a clean-event window (similar to the FSA’s occasional paper 25 “Measuring Market Cleanliness”). For an announcement to be included in the leakage sample, it must also occur within an event window that lacks any other stock-specific announcements over a six-day preevent period.1 Information leakage is defined as an event where there is an abnormal price movement on one or more days in the t − 6 to t − 1 pre-announcement period in the same direction as the overall announcement return. For a positive announcement,

R

† and ∗∗ indicate significance at 1% and 5%, respectively.

Information Leakage Event Definition Any A Abnorrmal Return r it where ARl =

SPRING 2015

JOT-HARRIS.indd 7

l

3 * σi

1 −260 r l Return r it , ∑ Abnorma 250 t =−10

Abnorma r l Return r it

l

− 3 * σi

for any day in the preevent window −6 to − 1. Our metric of information LEAKAGE is the ratio of information leakage events to clean event windows: LEAKAG E GE =

Number of inform f ation leakage a es Number of clean event windows w

(2)

Using a monthly metric over the period 2003–2011, Exhibit 3, Panel A presents the information leakages for the LSE and Euronext, respectively. Market Integrity: Market Manipulation

Market manipulation involves creating a false or misleading representation with the intent to dislocate the market price. We estimate suspected instances of ramping manipulation at the close using surveillance industry alert procedures. The surveillance industry devotes much time and effort and has developed considerable expertise in distinguishing purely abnormal closing prices from ramping manipulations at the close.


7

3/9/15 10:21:17 AM

EXHIBIT 3 Panel A: Information Leakage Events

fo

r

R

ev

ie w

O nl y

From the set of all clean company announcements, we plot the subset for which the return during an event window from day t − 6 to day t + 2 is more than three standard deviations away from a 250-day benchmark period ending at t − 10. The information leakage variable we analyze is this number of surveillance alert events divided by the number of clean announcements.

Panel B: Manipulate-the-Close (MTC) Events

A

ut

ho

r

D

ra

ft

We show the number of MTC events for the LSE and Euronext. We calculate the return for the last 15 minutes of each trading day (including the closing auction) and retain the returns that are plus or minus three standard deviations from the mean over a base period for the previous 30 trading days and that also have a price correction of at least 50% to the next day’s opening price, plus fifteen minutes.

† and ** indicate significance at 1% and 5%, respectively.

8


JOT-HARRIS.indd 8

SPRING 2015

3/9/15 10:21:18 AM

We have relied on that expertise in developing the following MTC metric. Manipulative behavior at the close is suspected when an end-of-day (EOD) percentage price change (in the last fifteen minutes of trading plus the closing auction) Rit exceeds 3σi above or below the mean of the distribution of 30 days prior observations and then mean reverts at least 50% in the first fifteen minutes of trading the next day:

bid-ask spread immediately prior to trades, multiplied by 200 to ref lect both entry and exit trades: Relative Effective Spread Definition 200 × Dt * ((T ((Trade d Price P i et − Midt )/Mid )/Midt )

Ramping Manipulation Event Definition f

it

ΔEODi + 3 * σ i

or each stock i and trading day t, =

ΔEO

i

=

Peod −155m ,i 1 t =−1 ∑ ΔEODit 30 t =−31

For both the LSE and Euronext samples, Exhibit 2 presents descriptive statistics before and after the introduction of MiFID1. The implementation of MiFID1 in November 2007, accompanied by direct market access and co-location, allowed algorithmic traders using an MTF such as Chi-X to avoid all structural barriers to reaching the national exchanges. Posting and canceling and revising quotes with dizzying frequency, the Level 10 limit order message traffic peaked at 200 orders for every trade in August 2010. Soon thereafter, Exhibit 1 shows that our empirical proxy for AT (the canel-totrade ratio) reached a high point in November 2010 in London and in January 2011 in Paris. Remembering that these univariate effects of introducing MiFID1 may well not hold up as ceteris paribus, marginal effects when we estimate the concurrent effect of AT in the equation system (i)–(iii), Exhibit 3, Panel A shows MTC declined significantly in London following the introduction of MiFID1 (compare mean events of 31.85 before to 7.55 after, significant at 99%). On the other hand, the univariate impact of MiFID1 on LEAKAGE is in the opposite direction. Controlling for many confounding variables, we find in our multivariate analysis (reported in Aitken et al. [2015] and summarized in Exhibit 4) that on net, MiFID1 increased EOD Manipulation and LEAKAGE in London and Paris, respectively. Value-weighted SPR exhibits significant reductions at 99% confidence in univariate analysis of the post-MiFID period for both exchanges in both samples. Exhibit 5 on relative effective spreads in these European markets shows spreads eroding from 20-35 basis points to 10-15 basis points earlier in the decade. Then when MiFID1 was introduced, value-weighted spreads

ev

it

d −15 15m ,it

ut

ho

r

D

ra

ft

fo

r

where ΔEOD is the percentage return between the closing price and the price 15 minutes prior to the close, ΔEOD is the average return over a rolling window of thirty trading days prior to the day being analysed, and σi is the standard deviation of ΔEODi over the same period. These statistically abnormal events are then screened against Reuters database for textual evidence of false rumors or company announcements. The remainder are defined as successful ramping manipulation attempts to mark the close (MTC). Exhibit 3, Panel B shows the incidence of MTC for the LSE and NYSE Euronext-Paris. MTC exhibits considerable variation over time as well as much elevated levels in London during 2006 and 2007 and then an abrupt reduction after the introduction of MiFID1 in November 2007. In Paris, MTC is sometimes quite high, especially in January and August 2008 and then again in April 2010 and July 2011.

R

Δ O

Peod,, t − P

O nl y

where Dt is the direction of the trade (1 for a buyerinitiated trade and −1 for a seller-initiated trade), and Midt is the midpoint price of the ask and the bid quote at time t. Daily and monthly spreads are equally weighted averages of all the securities traded that day or month.

A

Market Efficiency: Effective Spreads

The execution cost to trade various sizes is measured by the volume-weighted relative effective spread on a per trade basis and then averaged across the day. We calculate relative effective spreads in the standard fashion as the trade price minus the midpoint of the

SPRING 2015

JOT-HARRIS.indd 9


9

3/9/15 10:21:18 AM

initially spiked and thereafter fell throughout 20082010 consistently, by 40% on the LSE and by 60% on Euronext-Paris. This is precisely when an increasing market share of the trading migrated to the Chi-X multilateral trading system, a venue that proved especially attractive to algorithmic traders. Why do spreads narrow with greater HFT/AT? For one thing, technological innovation substantially reduced order-processing costs that are normally recovered through the spread. In addition, Hendershott et al. [2011] suggest that the reduction in spreads has occurred because high-frequency and algorithmic traders are better able to manage information asymmetries. Given

O nl y

AT’s capabilities in fast assessment and execution of alternative trading strategies triggered by changes in the state of the market, adverse selection may well decline at these times of market stress. As summarized in Exhibit 4, we find ultimately that AT reduced and MiFID1 raised spreads in London among both the most liquid and all listed stocks. SUMMARY AND CONCLUSION

ie w

Trade-to-trade data for 2003–2011 from the London Stock Exchange and NYSE Euronext-Paris exhibit a massive increase in algorithmic trading fol-

EXHIBIT 4

Simultaneous Determinants of Alerts, Effective Spread, and Algorithmic Trading

ev

Let alert incidence AIt be either MTC = Number of Successful f MTC or LEAKAGE = % Information Leakage; SPRt = Effective Spread; ATt = CTR (Cancel-to-Trade Ratio). Using 3SLS we estimate the following three-equation system to account for the cross-equation correlation t′ , ∈ t″ , ∈ t′″ (Aitken, Harris, Aspris, and Foley [2015]): of the ∈ (i)

SPR P t = α ′ + β 9 + β10

t″ + Year + ∈

(ii)

t′″ +∈

(iii)

+ β188

+ β12 t

t

+ β13 + β14

r

t

t

+ β19

fo

AT Tt = α ′′ + β166 t + β17

+ β11

R

′ AI t = α + β1Vlumt + β 2Stdt + β 3 A T + β 4 MiFID1 + β 5 SPR P + β 6 Ret t + β 7 EOQ + β 8 AI t −1,2 1,2 + ∈t

+ β 20 t + β 21

+ β15 t

+ β 22

1, 1,2

t 1,2 ,

A

ut

ho

r

D

ra

ft

where Stdt = mean standard deviation of the natural log of daily returns; Rett = mean of the natural log of the daily return; Vlumt = mean of the natural log of turnover; MiFIDt = a dummy variable for the introduction of MiFID1 (Equals 0 prior to November, 2007, and 1 subsequent); EOQ = a dummy variable to account for the end of each financial quarter (=1 for March, June, September, and December, and 0 for t′ , ∈ t″ , ∈ t′″ = observational error terms. all other months; Prt = mean of the natural log of price; Year = yearly fixed effects are applied; and ∈ We estimated the equations separately for the LSE and Euronext, having tested and found the pooled estimations invalid.

10


JOT-HARRIS.indd 10

SPRING 2015

3/9/15 10:21:19 AM

lowing the 2007 introduction of the MiFID1. We examine the impact of AT on the market quality of these two exchanges. Unlike previous studies that have focused exclusively on market efficiency and liquidity as measured by variables such as effective spreads (SPR), price impact and volatility, we also examine the impact of AT on market integrity. We construct two integrity proxies—one for information leakage (LEAKAGE) and the other for end-of-day manipulation—that is, marking the close (MTC). We then develop and estimate a threestage simultaneous equations system that accounts for the endogenous relationships that exist between AT, SPREAD, and LEAKAGE or MTC, while controlling for the changing regulatory environment and for the likely cross-equation correlation of disturbances. We document a significant change in both market efficiency and market integrity on the LSE and Euronext-Paris, especially apparent after the introduction of MiFID1. The policy implications of increased AT and the MiFID1 market-design change are summarized in

R

ev

ie w

O nl y

Exhibit 4. Although the average trade size and value declined significantly over 2003–2011, our multivariate results indicate that the increasing level of AT unambiguously increased market efficiency by reducing relative effective spreads in London and in top-quintile stocks in both markets. This reduction in SPREAD is consistent with theoretical predictions that suggest that because algorithmic traders are better able to update their information sets, they incur lower adverse-selection and inventoryholding costs relative to other market participants. Distinctively, we provide the first estimates on the relationship between AT and two metrics of market integrity (Aitken et al. [2015]). We show that although AT increases substantially toward the end of the day, the closing price is less manipulated the higher the incidence of AT in both London and Paris. By providing additional liquidity and reducing price impact, increased AT appears to raise the cost of implementing ramping manipulations. Moreover, we conjecture that like some hedge fund prop desks, ATs can reverse positions to

fo

A

ut

ho

r

D

ra

ft

Effective Spreads by Month 2001–2011

r

EXHIBIT 5

SPRING 2015

JOT-HARRIS.indd 11


11

3/9/15 10:21:20 AM

profit from ramping events faster than the manipulators themselves, thereby removing the manipulators’ profits. Our results regarding information leakage are more prevalent in the most liquid stocks. Increased AT lowers the incidence of leaking nonpublic information prior to price-sensitive announcements in the top two liquidity deciles on both Euronext-Paris and on the LSE. We conjecture that the speed and intraday forecasting accuracy of ATs in separating permanent and transitory components makes more abrupt the price reaction to trading based on information leakage. This would markedly increase the probability of detection, prosecution, and conviction for insider trading, thereby discouraging information leakage and trading ahead of price- sensitive announcements. The explosion in AT has saved the London and Paris markets from integrity violations attendant to the fragmented market design brought on by MiFID1.

Aitken, M.J., F.H.de B. Harris, T.H. McInish, A. Aspiris, and S. Foley. “High Frequency Trading—Assessing the Impact on Market Efficiency and Integrity.” Foresight Project, The Future of Computer Trading in Financial Markets, U.K. Government Office for Science, Driver Review DR28, 2012.

ENDNOTES

Grossman, S.J., and J.E. Stiglitz. “On the Impossibility of Informationally Efficient Markets.” American Economic Review, Vol. 70, No. 3 (1980), pp. 393-409.

O nl y

Brogaard, J.A. “High-Frequency Trading and Its Impact on Market Quality.” Working paper, Northwestern University, 2010.

ie w

Brogaard, J.A., T. Henderschott, and R. Riordan. “HighFrequency Trading and Price Discovery.” Review of Financial Studies, Vol. 22, No. 8 (2014), pp. 2267-2306.

R

ev

Gonzalo, J., and C. Granger. “Estimation of Common LongMemory Components in Cointegrated Systems.” Journal of Business and Economic Statistics, 13 (1995), pp. 27-35.

ho

REFERENCES

r

D

ra

ft

fo

r

The U.K. Treasury sponsored this research as part of its Foresight Project, The Future of Computer Trading in Financial Markets: An International Perspective, U.K. Government Office for Science (Aitken et al. [2012]). We wish to thank seminar participants at the U.S. Securities and Exchange Commission, the Australian Securities and Investments Commission, the Italian CONSOB, and the Six Swiss Exchange, as well as the Capital Markets CRC development team and SIRCA for providing the raw data. 1 If a company has more than one announcement within the six-day window, only the first announcement is considered.

Aitken, M.J., F.H.de B. Harris, A. Aspiris, and S. Foley. “Market Fairness: The Poor Country Cousin of Market Effcieincy.” Working paper, Wake Forest University, 2015.

ut

Aitken, M.J., and F.H. de B. Harris. “Evidence-Based Policy Making for Financial Markets: A Fairness and Efficiency Framework for Assessing Market Quality.” The Journal of Trading, 6 (2011) pp. 22-31.

A

Aitken, M.J., F.H.de B. Harris, and S. Ji. “A Worldwide Examination of Exchange Market Quality: Greater Integrity Increases Market Efficiency.” Journal of Business Ethics, forthcoming, online, 10 (August 10, 2014).

12

Harris, F.H. de B., T. McInish, F. Sensenbrenner, and R.A. Wood. “Fragmentation and Price Discovery: A Comparison of RegNMS and MiFID1.” The Journal of Trading, Vol. 9, No. 4 (2014), pp. 6-36. Hasbrouck, J., and G. Saar. “Low-Latency Trading.” Journal of Financial Markets, 16 (2013), pp. 646-679. Hendershott, T., C. Jones, and A. Menkveld. “Does Algorithmic Trading Improve Liquidity?” Journal of Finance, 66 (2011), pp. 1-33. Hendershott, T., and P. Moulton. “Automation, Speed, and Stock Market Quality: The NYSE’s Hybrid.” Journal of Financial Markets, 14 (2011), pp. 568-604. Hendershott, T., and R. Riordan. “Algorithmic Trading and the Market for Liquidity.” Journal of Financial and Quantitative Analysis, 48 (2013), pp. 1001-1024. Jarnecic, E., and M. Snape. “An Analysis of Trades by HighFrequency Participants on the London Stock Exchange.” Working paper, 2010.


JOT-HARRIS.indd 12

SPRING 2015

3/9/15 10:21:20 AM

Kumar, P., M. Goldstein, F. Graves, and L. Borucki. “Trading at the Speed of Light: The Impact of High-Frequency Trading on Market Performance, Regulatory Oversight, and Securities Litigation.” In Finance-Current Topics in Corporate Finance and Securities Litigation, pp. 1-12. The Brattle Group, Issue 02/2011 (2011).

O nl y

Liu, W. “Monitoring and Limit Order Submission Risks.” Journal of Financial Markets, Vol. 12, No. 1 (2009), pp. 107-141. McInish, T., and J. Upson. “The Quote Exception Rule: Giving High-Frequency Traders an Unintended Advantage.” Financial Management, 42 (2013), pp. 481-501.

ie w

Menkveld, A.J. “High-Frequency Trading and the New Market Makers.” Journal of Financial Markets, 16 (2013), pp. 712-740.

R

r

Zhang, F.X. “High Frequency Trading, Stock Volatility, and Price Discovery.” Working paper, Yale University, 2010.

ev

Yan, B., and Zivot, E. “A Structural Analysis of Price Discovery Measures.” Journal of Financial Markets, 13 (2011), pp. 1-19.

A

ut

ho

r

D

ra

ft

fo

To order reprints of this article, please contact Dewey Palmieri at dpalmieri@ iijournals.com or 212-224-3675.

SPRING 2015

JOT-HARRIS.indd 13


13

3/9/15 10:21:20 AM