Trade-Based Manipulation and Market Efficiency: A Cross-Market Comparison Michael J. Aitken Chair of Capital Markets Technologies University of New South Wales Australian School of Business [email protected] Frederick H.deB. Harris John B. McKinnon Professor of Economics and Finance Wake Forest University Schools of Business and Australian School of Business University of New South Wales [email protected] Shan Ji University of New South Wales and Smarts Group International, Inc. [email protected]

Abstract We develop a testable hypothesis that trade-based manipulation as proxied by the daily incidence of ramping alerts raises execution costs for completing larger trades on 34 security markets worldwide 2000-2005. The alternative hypothesis is that ramping alerts represent information arrivals that are delayed, unmasked as rumors, or proven false. Using observational error components to represent the presence of a manipulator or the arrival of information in a random effects model, we show that effective spreads are positively related to the incidence of ramping alerts across 8 of 10 liquidity deciles. The magnitude is economically significant; cutting ramping manipulation by half reduces the effective spread 31 to 59 basis points in the middle liquidity deciles worldwide. In addition, we identify the determinants of ramping manipulation by estimating a simultaneous equations model of alert incidence, spreads, and the probability of deploying real-time surveillance (RTS) across all listed securities in 2005. Closing call auctions, direct market access, specific regulations, RTS procedures and enforcement assure better market integrity enhancing market efficiency. Keywords: Market manipulation, market integrity, spreads, surveillance JEL Classification: G28 (Financial Institutions & Services, Government Policy & Regulation) This Draft: 18 November 2009 Acknowledgements: We wish to thank the Security Industry Research Centre of Asia-Pacific (SIRCA) and the Capital Markets Cooperative Research Centre (CMCRC) II in Sydney for data and financial support, respectively. Doug Cumming, Tom Smith, and Kumar Venkataraman provided invaluable advice. The opinions reported herein are personal and do not reflect the policies, procedures, or opinions of any of our employers.

Trade-Based Manipulation and Market Efficiency: A Cross-Market Comparison 1. Introduction Not all market volatility is natural, deriving from fundamentals. Instead, some volatility is induced by failing to assure market integrity or by adopting poor market design. Along with fraudulent disclosure and insider trading, trade-based manipulation is one of the most significant threats to market integrity. With the CBOE VIX and other indices of volatility reaching record levels worldwide in 2008-2009, major security exchanges have begun to declare their commitment to surveillance procedures and market designs that best serve the twin goals of market integrity and efficiency. For example, NASDAQ’s website states NASDAQ is among the world’s most regulated stock markets, employing sophisticated surveillance systems…to protect investors and provide a fair and competitive trading environment... fostering innovative technologies…[ that] continue to build the most efficient trading environment…to the benefit of all market participants and investors. Although many exchanges have invested in significant surveillance resources to detect trade-based manipulation and improve market efficiency, little is known about the direct relationship between the two. Primarily this reflects the extreme difficulty of collecting large samples of detailed data about security market manipulations. In this paper we employ a random effects model to analyze the possible presence of a manipulator as well as the possible arrival of information each security day, both of which are observed with error in the market place. In this error components framework, we then investigate whether a reduction in trade-based manipulation actually achieves tighter spreads. Our analysis covers all the listed securities on 34 exchanges worldwide over the period 2000-2005. 1.1. Theory of Trade-Based Manipulation: An Overview Allen and Gale (1992) define trade-based manipulation as a trader attempting to manipulate a stock price simply by buying and then selling (or vice versa), without releasing any false information or taking any other publicly observable action designed to alter the security’s value.1 Traditional full-information financial theory asserts that such speculation stabilizes prices because manipulators like all rational speculators buy when the prices are low and sell when the prices are high. In contrast, with incomplete and asymmetric information, Hart and Kreps (1986) show that speculation can destabilize prices and increase volatility because uninformed traders can not distinguish between the rational speculators and strategic insiders with private information. As a result of this pooling equilibrium, profitable manipulation can exist under quite general conditions of expected utility maximization and rational expectations by manipulators, strategic insiders, 1

Trade-based manipulation is thereby distinguishable from other failures of market integrity like insider trading (Bhattacharya and Daouk 2002) or the spreading of false rumours (Van Bommel 2008).

1

market makers, and noise traders pursuing a momentum strategy (Allen and Gale 1992, Allen et. al. 2006). Even without momentum traders, Aggarwal and Wu (2006) show that if information seekers can not distinguish between a manipulator and an informed trader acting strategically, trade-based manipulation can be profitable. A strategic insider can make unprofitable initial trades against the direction of his information, set in motion a price trend among partially informed followers, and then profitably unwind his position against still less informed market makers and other liquidity providers (Chakraborty and Yilmaz 2004, 2007). With a manipulator present in these nested information environments, the efficient market no longer serves as an aggregator of equilibrium price information alone. Instead, security price trends may represent induced volatility rather than the arrival of new information. When markets begin to trend, liquidity suppliers who were content to earn the spread in mean-reverting markets choose instead to go flat. This reduction in liquidity may not change the quotes for trivial size, but it does raise the effective spread. Moreover, with reduced non-execution risk because of the higher volatility, Foucault’s (1999) theory of order placement implies that liquidity suppliers will then submit orders less aggressively. Aitken, Almeida, Harris and McInish (2007) confirm empirically that liquidity suppliers in electronic markets will then layer orders further from the BBO thereby raising the effective bid-ask spread for completing larger trades. 1.2. Prior Empirical Findings on Manipulation Given this theoretical link between securities market manipulation, induced volatility, and effective spreads, what does prior empirical research show about the effect of laws, regulations and surveillance designed to prevent manipulation? When a securities market is laden with manipulators, investors choose to invest elsewhere. Cumming and Johan (2008) establish that trading activity increases if exchanges adopt surveillance procedures and regulations that assure market integrity. A number of other empirical studies of particular events in particular exchanges confirm that manipulation also increases volatility (e.g., Stoll and Whaley (1987, 1991), Chamberlain, Chueng and Kuan (1989), and Chiou, et al (2007). However, there has been no direct test of the relationship between manipulation and spreads across securities and exchanges. Herein, we estimate a random effects model of the error components relationship between trade-based manipulation and quoted and effective spreads. A doubling of manipulation alert incidence (AI) is associated with a 31 to 59 basis point increase in effective spreads across 7 of 10 liquidity deciles, which represents approximately a 10% increase in the moderate liquidity securities across 34 exchanges worldwide 2000-2005. We then investigate the determinants of ramping alert incidence with a cross-sectional simultaneous equations model of AI, spreads, and the probability of adopting real time-time surveillance (RTS). The data is collected daily across all listed stocks in 2005 and then aggregated to the exchange-specific liquidity decile as a unit of analysis to mirror market surveillance in practice. Factors that affect market quality (trading

2

regulations, market technology, market design infrastructure, market participants) are measured across exchanges and then related to alert incidence, spreads, and Prob(RTS). 1.3. Our contribution Using random effects modeling, we show that less ramping manipulation reduces effective spreads by approximately 10% across the middle liquidity deciles where manipulation is predicted to be most prevalent. Moreover, we find that a particular design choice (call auctions at the close), a particular technology (direct market access lines), and a regulation specifically prohibiting ramping manipulation reduce manipulation alert incidence. Market integrity regulations more generally reduce effective spreads. And real-time surveillance of trade-based manipulation is more likely to be adopted the lower the volatility, the larger the trading volume, the higher the foreign direct investment, the faster the execution speeds, the fewer the competing regulatory objectives, and the greater the risk of market manipulation as proxied by ramping alert incidence. The paper proceeds as follows: the next section documents the myriad forms and evidence of tradebased manipulation. Section 2 explains the random effects methodology, develops our testable hypothesis, distinguishes a competing alternative, and specifies the empirical models we estimate. Section 3 describes the data, our measurement of ramping alert incidence, and the distributional properties of AI and spreads worldwide, concluding with several data-driven limitations of our research design. Section 4 presents and discusses the empirical results of the error components model. Section 5 addresses the drivers of trade-based manipulation including an analysis of the dimensions of market quality. Section 6 develops a crosssectional empirical model of alert incidence, spreads, and the probability of deployment of real-time surveillance systems for all listed securities on 24 exchanges in 2005. Section 7 presents a systems estimation of the simultaneous structural equations for AI, QSpr, and Prob(RTS). A summary and conclusion are provided.

1.3.1. The Myriad Forms of Market Manipulation Security market manipulation exists in a wide variety of forms – ramping, wash trades, layering the order book, churning, cornering, squeezing, front running, bait and switch and other intentionally misleading orders and trades.2 No matter the form, successful market manipulation temporarily distorts a security’s price. The overall aim is to drive the price in the direction beneficial to the manipulator who then liquidates his holdings or covers his short position at a price better than the implicit efficient price in a full-information equilibrium.

1.3.2. Ramping Manipulation Marking the Close Ramping manipulation is the focus of the present research because this is the only form of manipulation for which the requisite data are publicly available. Statistically rare price movements (e.g., less than α = .005) that revert the following day may indicate market manipulation to artificially inflate or deflate the price of a 2

An Appendix describing each of these in more detail is available from the authors.

3

security. A Ramping manipulation normally involves two surveillance alerts: Marking the Close and then Reversal at the Start of the next Trading Day. Marking the Close refers to the practice of executing purchase or sale orders at or near the close of the trading session in order to raise or lower the closing price, the bid or the offer artificially. In one context Ramping is also referred to as Painting the Tape when a fund manager manipulates a security’s closing price at the end of the evaluation period. The manipulator’s purpose is to reduce margin or net capital requirements for enhancing profit and loss, or to influence the mark-to-market calculations mandated by regulatory authorities for credit authorization or reporting purposes if holding a large position in derivative contracts. Figure 1, Panel A illustrates a Marking the Close alert incident screen developed by Smarts Group International for assisting regulators and brokers in detecting trade-based manipulation. Trades at the ask are shown in red, and trades at the bid are shown in green. The size of the trade is indicated by the diameter of the circle. Blue circles represent off-market trades that can be negotiated at prices away from the continuous auction price beyond a minimum size. Trading volume is shown in the bar diagrams colored yellow. This particular 21% appreciation observed in the last 15 minutes of trading occurs in less than ½ of 1% of the trading days for this stock. Such a significant price increase at the end of the day normally reflects anticipated positive information arrivals. If, however, these price trends are reversed the next morning, a ramping incident alert is triggered, as illustrated in Figure 1, panel B. This paper will use such Ramping Alerts (which consists of a Marking the Close Alert and a Reversal the Next Morning Alert) as a proxy for securities market manipulation. Of course, such price reversals may also be explained by information announcements that are quickly reversed, unmasked as rumors, or proven false. Our empirical testing with random effects/error components modeling is designed to distinguish between these two competing hypotheses. An Australian Securities Exchange (ASX) case involving the tracking stock on the Standard & Poor’s ASX 200 Index illustrates the behavior underlying Ramping Alerts. On Friday, 29 June 2001 between 4 and 4:15 p.m., this stock increased 45.5 points or approximately 30% following the closing auction. The last trading day of the financial year almost always pushes share prices somewhat higher on the ASX, but on 29 June 2001 the All Ordinaries Index rose by only 2%, and the ASX became concerned that market manipulation may have been involved in the tracking stock. By market open on the following Monday, this unusual increase was reversed. Specifically, on 2 July, the index tracking stock fell by 54 points, as the ramping buyers (believed to be fund managers and derivative players) dumped the index tracking stock and withdrew. Figure 1 Panel B presents a surveillance screen that displayed the Marking the Close and Reversal the Next Morning alerts for those two days.

1.3.3. Other evidence of closing price manipulations Flexison and Pelli (1998) and Hillion and Souminen (2004) find that brokers manipulate the closing price of a stock preceding large agency trades in order to improve their customers’ impression of execution

4

quality. Carhart, Kaniel, Kusto and Reed (2002), and Bernhardt and Davies (2005) show that mutual funds paint the tape by manipulating closing prices at the end of evaluation periods to improve fund performance against a closing VWAP benchmark. Stoll and Whaley (1987), and Chamberlain, Chueng and Kuan (1989) find empirical evidence in the North American markets that on the expiration day of index futures/option contracts, the price meanreversals are significantly higher than month-ends or quarter-ends without index futures/options expiration. Stoll and Whaley (1991) suggests that the change of settlement procedure to use next day’s opening price in the New York Futures Exchange and New York Stock Exchange would shift expiration day timing but not affect the motivation to manipulate closing prices.

2. Hypothesis Development For the purpose of this research, we focus not on the relationship between manipulation and informational efficiency but rather on manipulation and execution costs. Across the 34 exchanges we study worldwide, we can measure two widely used and accepted relative spreads: (1) the cost of a round trip transaction at the best bid and offer relative to the quote midpoint (the quoted spread), and (2) the average cost beyond the quote midpoint to complete all trades relative to the quote midpoint (the effective spread). By maintained hypothesis, market manipulation increases price volatility. Foucault (1999) develops a theory of order placement relating volatility to effective spreads. Order placement strategy consists of two components, order type and order aggressiveness. Traders choose from market orders and limit orders. When non-execution risk is high, traders employ market orders to gain immediate execution. When pickingoff risk is high, limit orders are preferable. Order aggressiveness refers to how close the limit order price is to the prevailing best bid or offer (BBO) when the order is entered or amended. Foucault’s theory predicts that when the volatility increases, traders will tend to place limit orders rather than market orders to reduce their picking-off risk at the cost of higher non-execution risk. When nonexecution risk is also particularly high, liquidity demanders are under pressure to trade immediately upon arrival and are therefore willing to place market orders at less favorable prices. This induces liquidity suppliers to post less aggressive limit orders (farther away from the BBO) in order to take advantage of the impatient traders. The implication is that effective bid-ask spreads (volume-weighting the trade prices that walk up and down the book) will rise as volatility increases. If the volatility of security markets is indeed higher in the face of market manipulation, this effective spreads measure of market efficiency will be adversely affected. Aitken, Almeida, Harris and McInish (2007) provide empirical evidence supporting the implications of Foucault’s theory. They hypothesize that hedge funds and proprietary trading desks tend to have short-

5

lived information about valuation and/or the state of the market. As a result, these classes of traders face higher costs of non-execution and lower picking-off risk than insurance companies and mutual funds. They hypothesize that the order placement strategy of proprietary trading desks and hedge funds will therefore be more aggressive than insurance companies and mutual funds. Using a large sample of trading desk instructions, they show that insurance companies and mutual funds do tend to be less aggressive order placers, ceteris paribus. Our testable hypothesis is that trade-based manipulation also reduces order aggressiveness. Specifically, the null hypothesis of our research is : A higher incidence of trade-based manipulation (as proxied by ramping alert incidence AI) is associated with wider effective spreads, ceteris paribus where AI is the number of alerts per security day triggered by an algorithm implemented by the world’s leading surveillance technology and consulting firm, Smarts Group International.3 If mean reversion of prices the next morning following an extraordinary price increase (or decrease) marking the close represents trade-based manipulation, then volatility will have increased, ceteris paribus. Foucault’s theory predicts liquidity suppliers will therefore reduce their order aggressiveness. Quoted spreads for trivial depth on electronic markets may remain unchanged as technical transaction costs are unchanged. Nevertheless, limit orders for larger volume would be spread farther away from the BBO to avoid being triggered by a manipulator’s walking up/down the order book and then quickly liquidating/covering his position before the liquidity traders could react. This rational response of liquidity suppliers to the anticipated presence of a possible manipulator, ceteris paribus, would show up as a higher volume-weighted effective spread. If a security or exchange exhibits repeated manipulations, we would expect higher effective spreads to persist, a disreputation effect of failing to assure market integrity.

2.1. Alternative hypothesis The alternative hypothesis is that ramping alerts represent not market manipulations but rather information arrivals that are delayed, unmasked as rumors, or proven false. Such information arrivals all of which are quickly reversed would not lead to wider effective spreads when averaged over longer periods of months, quarters or years. Rather, under the alternative hypothesis, spreads would quickly mean revert. In contrast, confirmed information arrivals (both positive and negative) trigger herding and other information-based trading that cause markets to trend (rather than mean revert). Markets that trend do exhibit increased spreads as market makers and liquidity traders go flat and protect themselves against picking off risk. But not so with information arrivals that are quickly reversed. Consequently, under the alternative hypothesis, information arrivals that are quickly reversed lead to mean reversion of the spread,

3

At present, the SMARTS Market Integrity Platform is deployed in 40+ national exchanges and regulators and 150+ brokers across 35 countries. See www.smartsgroup.com.

6

and ramping alert incidence (AI) if mistakenly capturing information arrivals that are quickly reversed will be unrelated to long-term average spreads.

2.2. A random effects/error components methodology for trade-based manipulation research Both information arrivals and market manipulations are inherently unobservable variables, subject a priori to observational errors. Moreover, since spreads in round trip transactions, and the spread to complete larger trades, are substantive costs of engaging in market manipulations, we also expect alert incidence to be determined by relative spreads. Hence, our prior is that relative spreads and AI are endogenously determined. For both of these reasons, this research utilizes a Random Effects Model to test the correlation between Ramping Alert Incidence and bid/ask spreads. In light of the lognormality of the cross-sectional spread, we hypothesize that the theoretical relation between relative effective spread and ramping alert incidence is (1) and transform for estimation to the regression relation ( 1’ ) where = the relative spread of market i at time t = a constant = the percent change of the relative spread with respect to a percent change of Ramping Alert incidence (i.e., an elasticity) = the number of Ramping Alerts per security day in market i in period t = an observational error on information arrivals at time t = an observational error on Ramping Alert detection in market i = a residual error term When a Hausmann specification test fails to reject Cov(

) = 0 and Cov(

) = 0, the

Ramping Alert incidence AI regressor is orthogonal to observational errors across exchanges and over time. In this circumstance, we employ Da Silva’s autoregressive model: (2’) which assumes a mixed variance-component moving average for the error structure of order k determined by minimizing the RMSE. The Hausmann’s (1978) m test statistic we employ to distinguish these two cases is distributed with 1 d.o.f.

7

On the other hand, when the Hausman specification test rejects (

: Cov(

) = 0 and/or

), then we conclude that autoregressive models would be a misspecification because Ramping Alert incidence is measured with observational error (i.e., is itself stochastic). In that event, fixed effect dummy variables are used to control for the systematic effects of observational errors in measuring the arrival of information over time Dt and the presence of a manipulator across securities Di: (3’) It is essential to understand that the Random Effects Model specified above will be used to test the correlation rather than the causality between relative spread and Ramping Alert incidence across securities markets over time. The Random Effects Model is based on priors about the error components. That is, we expect spreads to reflect both

– the observational error on information arrivals at time t, as well as

the observational error on manipualtor detection in market i. If Cov(

) = 0 and Cov(

) = 0), then

and

is independent of

and

–

(i.e., if

are not jointly dependent on random

observational errors in detecting informational arrivals over time or market manipulations across securities. In that event, the correlation between relative spreads and manipulation alert incidence can be estimated directly without bias. However, when the Hausman specification test reveals that Cov( Cov(

)

)

0 and/or

0, then we proceed to fixed effects modeling to address the endogeneity of SPR and AI

both being dependent on idiosyncratic factors across exchanges and over time. The exchange-specific dummy variables capture idiosyncratic surveillance, detection, prosecution, or enforcement reasons why manipulation may be observed only with error. The time-series dummy variables control for idiosyncratic reasons in each time period why information may arrive but be observed only with error. Despite all these fixed effects control variables, the null hypothesis remains a positive association between

2.2.1. Cov(

and

.

) Illustrated:

The following examples illustrate various possible findings for the information arrival observational errors relevant to testing

(i.e., that the covariance between Ramping Alert incidence and information

arrival observational errors over time is zero): (1) A Ramping Alert is triggered by a Reversal the Next Trading Day Alert following a Marking the Close Alert. The price reversal detected the next morning could result from an information arrival plus the associated liquidity appearing thereafter in resilient limit order books. Or alternatively, it could result from ramp-and-dump trading behaviour on the day following a Marking the Close Alert triggered by a true closing price manipulation. Because in both cases the Ramping Alert would be triggered, the null hypothesis of zero systematic relationship (i.e., zero covariance) between information arrivals and Ramping Alert incidence over time would in this example be accepted;

8

(2) In opposition to (1), if reinforcing positive (or negative) news is announced overnight when the Marking the Close Alert is triggered, there will be no price reversal the next morning. As the market continues to trend in response to fundamentals, this stops the Ramping Alert. In such an event, the null hypothesis of zero covariance between information arrivals and Ramping Alert incidence over time would be rejected; (3) What we have been labelling an information event could also have been a non-event. For example, a pure white noise announcement by a small listed company could be out of traders’ sight and therefore doesn’t trigger a Ramping Alert. In such a scenario, the null hypothesis of zero covariance between information arrivals and Ramping Alert incidence over time would be accepted.

2.2.2. Cov(

) Illustrated:

The following examples illustrate various possible findings for the presence of a manipulator observational errors relevant to testing

(i.e., that the covariance between Ramping Alert incidence and manipulator

detection error across securities or markets is zero). (1) When closing price manipulation is truly present for a security from a market and a Ramping Alert is triggered for that security, the null hypothesis of zero covariance between Ramping Alert incidence and manipulation detection error across securities markets would be accepted; (2) When closing price manipulation is truly present for a security, but due to insufficient monitoring by momentum traders or arbitrage traders, no one emulates the manipulators’ trading activities, the manipulation strategy therefore fails and no Ramping alert is triggered. In such a scenario, the null hypothesis of zero covariance between Ramping Alert incidence and manipulation detection error across securities markets would be rejected; (3) The alert detection surveillance process could also generate pure white noise. For example, an arbitrary time period before closing (e.g., 15 minutes) is used in all markets to detect closing price manipulation (AI). But the manipulation could occur 30 to 15 minutes before closing or at other randomly chosen time distances from closing. In such a case, the null hypothesis of zero covariance between manipulation detection error and Ramping Alert incidence over time would also be accepted.

3. Data and Measurement The data for this research is obtained from the Reuters database maintained by the Securities Industry Research Centre of Asia-Pacific (SIRCA). This database contains intra-day trade and quote data since 2000 for more than 200 exchanges worldwide. We analyze the entire sample of listed securities in all ten liquidity deciles from each of 34 securities markets for which we could obtain surveillance procedure information. Table 1 lists the exchanges studied. Because ramping manipulation alerts are so rare at the individual security-day level, we aggregated security-day observations to the liquidity-decile-day level. Liquidity

9

deciles are determined by dividing the total number of securities in each market into 10 groups, based on their monthly trading turnover. The period of analysis for our study extends for six years from January 2000 to December 2005.

3.1. Ramping Alert Algorithm The Ramping Alert algorithm applied in this study is from the SMARTS’ Real-time Securities Market Surveillance Platform. The algorithm of the Ramping Alert is described below. Benchmark Period and Threshold For date T, a historical price change distribution for the past month (the benchmarking period) is created for each security. The observations in this distribution are sampled wherever on-market trades occur throughout the benchmarking period. Fifteen minutes after the market opens, we calculate the percentage change between the trade price and the true price 15 minutes earlier. True price is defined as (1) the previous trade price; or (2) the best bid (offer) price at time t-15 minutes if the previous trade price is below (above) the best (offer) price at t-15 minutes. Finally, we take the absolute value of the calculated percentage price change and add it to the historical distribution. At the end of the benchmarking period, we check the number of observations from each security’s historical price change distribution. If there are more than 50 observations, then we set the ramping price change threshold for that security as the 99.5% histogram distribution cut-off. If there are 50 or less observations from the distribution, then we determine the ramping price change threshold for that security as follows: Price $0.001 to 0.10—20%, $0.11 to 0.25—15%, $0.26 to 0.50—12%, $0.51 to 1.00—10%, $1.01 to 5.00—8%, $5.01 to 10.00—5%, and above $10.01—3%. The purpose of the benchmark process is to identify the top ½ of 1 percent of least frequent price changes for a security during the benchmark period. Assuming that there are approximately 20 trading days in a month and 100 trades in each trading day (assuming 6 trading hours per day), there are approximately 2000 price change observations each month. If these observations are sorted, the value of the 10th largest price change is where the threshold for ramping for that security is set. For example, if the 10th highest price change for BHP Billiton is 0.5% during September, then the security is deemed to have been subject to ramping if the return in the last 15 minutes of 10 October, 20 October, or 31 October was greater than 0.5%. Conditions for Marking the Close Alert After market i closes on date t, for each security, we trigger a Marking the Close Alert if the absolute percentage difference between the closing price and the true price 15 minutes prior is greater than the ramping price change threshold for that security. Conditions for Reversal the Next Morning Alert On date T+1, for each security, we trigger a ramping alert if the following conditions are satisfied: (1) if there was a Marking the Close Alert triggered for that security on date T; and

10

(2) if during the first 15 minutes of trading on date T+1, if the Marking the close Alert triggered on date T is for driving up (down) closing price by X%, at least one trade occurs at price P that is below (above) the closing price on date T by more X% or more (a.k.a., Reversal the Next Morning). The above algorithms are run daily across the 34 securities exchanges over 6 years (2000-2005) to derive the alert incidence (AI) of ramping manipulation per security day. The AI of daily ramping manipulation per security day is calculated as the number of ramping alerts triggered for all securities from a liquidity decile in market i in year Y normalized by the average number of listed securities.

3.2. Time-weighted quoted spread and volume-weighted effective spread We employ standard time-weighted calculations for the Quoted Spread for each security: The time weight was calculated by taking the proportion of trading time that each spread existed during a trading day. The Effective Spread for each security was calculated as:

where Di,t = the trade direction (D = 1 for buyer initiated trades with trade price above the midpoint price and D = -1 for seller initiated trades with trade price below the midpoint price). The volume weight was calculated by taking the volume of each trade as a proportion of the total daily traded volume for each security. In both cases, negative average spreads per security day and instances where one side of the spread was absent were removed from the sample.

3.3. Descriptive statistics 3.3.1. Spreads The descriptive statistics for the average spreads per security year across the 34 exchanges over the 6 year period (2000-2005) are presented in Table 2. It can be seen that the 204 average Quoted Spreads and Effective Spreads are demonstrably non-normal. For example, the effective spread (“ES”) has mean (

) of 6.64% and standard deviation (

) of 0.1383 with skewness of 5.1899 and

kurtosis of 28.7423. After a natural log transform, we observe the distribution of approximately normal (

= -3.4645 and

to be

= 1.2172) with skewness of -0.1480 and kurtosis of 0.9826.

The same is also observed for quoted spread. Figure 2 presents the histogram for the two spread measures before and after the natural log transform. Using the properties of the lognormal distribution and assuming exact log-normality for our observations, an estimator of = 6.56%. This mean estimate differs from

would be

=

of 6.64% because our sample differs slightly from an exact

lognormal distribution. The Quoted Spread varies enormously across the ten deciles (see Table 3, Panel A) from 0.73% for the most liquid decile to 20.98% for the least liquid decile. Effective Spread varies from 2.19% to 13.56%. In general, the three thinnest-trading deciles exhibit spreads that are an order of magnitude wider than the

11

three most liquid deciles. These descriptive statistics suggest Deciles 4, 5, 6 and 7 should be grouped as a separate class of transactions, separate from the thickest and thinnest-trading deciles, and we perform Chow tests on the estimated models to confirm whether or not to do so. The pooling of the data is rejected. Consequently, we report below and discuss the regressions for thickly-traded (1,2,3), moderately liquid (4,5,6,7), and thinly-traded (8,9,10) subsets of the ten liquidity deciles. 3.3.2. Ramping Alert Incidence (AI) Table 1 shows the mean annual alert incidence of daily manipulation in each of the 34 exchanges. The grand mean of means is 0.21, meaning 1 in approximately 500 trading days or one every five days per 100 securities. Mean AI ranges from 0.02, 0.03, 0.03, 0.06 and 0.07 in Korea, Istanbul, Shanghai, Shenzhen, and Hong Kong to 0.45 and 0.44 in the Taiwan and Bombay stock exchanges. The alert incidence in Singapore and the Euronext markets (0.30 to 0.37) are quite high in the distribution. NASDAQ has a 0.08 AI. Table 3, Panel B presents the mean annual alert incidence of ramping manipulation by security day aggregated to the decile level across all 34 exchanges. It can be seen that across all securities from 34 markets, the alerts incidence is monotonically decreasing from 0.31 in decile 1 to 0.13 in decile 10. In the least liquid securities, detection of manipulation strategies by surveillance authorities is too likely. On the other hand, in the most liquid securities, the capital required to ramp a security is too extensive. Therefore, we would expect manipulations to be highest in the moderate liquidity deciles 4, 5, 6, and 7. 3.4. Limitations of the error components research design The higher alert incidence at the top of Table 3, Panel B could be caused by alert detection errors especially in thickly-traded deciles 1, 2 and 3. In fact, much surveillance workflow is designed to validate alerts that can be triggered by a variety of legitimate reasons not involving manipulation. It is quite common that a large proportion of the alerts triggered each day prove to be false positives. And it has become common surveillance practice to adjust various attributes (e.g., alerting conditions, thresholds, etc.) to have alerts issued more or less often based on the client exchanges’ capabilities in screening for false positives. So the problem of false positives in ramping alerts is well known and is likely to be highly correlated with trading volume. By definition, the surveillance alert count data adjusts for the larger number of securities in higher liquidity-deciles but not for the fact that these securities are traded more frequently than those from lower liquidity-deciles. Given the design of surveillance alerts, the higher incidence of false positives in more liquid securities probably explains the rising monotonic mean AI statistics from decile 10 to decile 1. We should like to point out however that this false positives data collection issue on AI in fact our findings against (not in favor of) our testable hypothesis of a positive spread-AI relationship. When mean reversion occurs as an equilibrating response to random information shocks in resilient limit order books, liquidity providers tend to tighten the spreads. Decreased picking-off risk in mean-reverting (flat as opposed to trending) markets results in lower spreads, the opposite of our hypothesized positive relationship between

12

SPR and AI. Hence false positives in our AI data collection would increase the chances of falsifying our hypothesis, not the reverse. A second limitation of the error components model structure is that we make no attempt to estimate the structural equations. Spreads and AI are simultaneously determined or at least determined by common shocks as seemingly unrelated regressions. Higher AI raises volatility, and volatility reduces order aggressiveness raising spreads. Spreads are an execution cost of market manipulation; higher quoted spreads reduce the incidence of manipulation, ceteris paribus. Controling for exogenous volatility in the structural equation for spreads, our manipulation hypothesis would be that higher AI results in still less order aggressiveness as induced volatility reduces non-execution risk, again raising spreads. So, either greater exogenous volatility or greater volatility attributable to manipulation raises spreads. Here, we are content to trace the simple correlation between spreads and surveillance alerts. Note that the two can not be spuriously correlated through exogenous volatility, since the spreads-exogenous volatility relationship is a priori positive, while the basic AI-exogenous volatility relationship is a priori negative. Specifically, increased exogenous volatility substantially raises the transaction costs of ramping a market, thereby reducing equilibrium AI. Structural equation estimation is required to sort out these SPRvolatility and AI-volatility relationships, and we undertake that research in sections 5 and 6 below. 4. Empirical Results of Error Components Model Table 4 Panel A shows that across all liquidity deciles, effective spreads averaging 664 basis points across our 34 exchanges are increased by 11.92% (i.e., e0.1126 – 1) or about 79 basis points when ramping alert incidence doubles. These elasticity parameter estimates from the Full Random Effects Model are unbiased by covariance between AI and the observational errors (since the Hausman specification test is insignificant) but they explain only about 5% of the variation in spreads. With full fixed effects introduced for n-1 exchanges and years, the model can explain 81.7% of the variation in spreads with F=18.52, and the SPR-AI relationship remains positive and significant at 0.01. Again a doubling of ramping alerts is estimated to increase spreads by 11.14% or 74 basis points. Allowing for a moving average of error components, the AI parameter estimate declines to a 7.44% increase in spreads or about 49 basis points when AI doubles. Table 4 Panel B shows that even the relationship between quoted spreads for doing trivial size at the BBO and AI is positive, though weaker. A doubling of alert incidence raises quoted spreads averaging 699 basis points across our 34 exchanges by 4.00% or 28 basis points. Here, the Hausman test indicates the Full Random Effects estimate of the SPR-AI relationship is biased by joint interdependence on fixed effects across exchanges and over time. Controling for exchange-specific dummy variables (31 of 33 of which are significant) and controlling for annual dummy variables (4 of 4 of which are significant), the model explains 94.3% of the variation in spreads with F=71.3, significant at 0.01, and the AI elasticity parameter (i.e., e0.0392 – 1) remains positive and significant.

13

Chow tests indicate however that these pooled estimations across all ten liquidity deciles cannot be validly pooled ( F = 10.09 with p-value less than 0.01) -- i.e., that the All Deciles results are masking enormous heterogeneity in the SPR-AI relationship across thickly versus thinly-traded stocks. In disaggregated results available from the authors, Quoted SPR-AI elasticities of less than ½ of 1% are observed in thickly-traded deciles 1, 2 and no effect in decile 3. Similarly, deciles 4 and 5 shows no effect. In the moderately liquid deciles 6 and 7, doubling ramping alerts raises quoted spreads by 3.2% or 18 basis points and by 2.0% or 15 basis points, respectively. Similarly, in decile 8, the estimated SPR-AI elasticity parameter is a minuscule 1% or 11 basis points. Indeed, it is only in the most thinly-traded stocks that we observe any economically significant effect of surveillance alerts on quoted spreads. A doubling of ramping alerts in decile 10 securities raises quoted spreads by 3.2% or 67 basis points. The BBO measure of market efficiency seems to be affected by failures to assure market integrity only in the thinnest stocks. Quite the opposite is true of effective spread-AI relationships however. Again, in disaggregated results available from the authors, in 8 of 10 liquidity deciles the effect of doubling ramping alert incidence is associated with 31 to 59 basis point of increased effective spread. Among the thickly-traded deciles, decile 2 exhibits a significant SPR-AI parameter. Doubling alerts increases effective spreads by 4.6% or 11 basis points in the Random Effects Model (and by 3.7% in the Full Fixed Effects Model). In almost all the moderately liquid and thin-trading deciles, the estimated elasticities of effective spreads with respect to alert incidence are very substantial: 3.9% x 790 b.p. of mean spread in decile 4 = 31 b.p.,4 7.5% x 523 b.p. of mean spread in decile 6 = 39 b.p., 8.2% x 721 b.p. in decile 7 = 59 b.p., 5.1% x 770 in decile 8 = 40 b.p., 5.4% x 1091 b.p. in decile 9 = 59 b.p., and 3.4% x 1356 in decile 10 = 46 b.p. Across 8 of 10 liquidity deciles, assuring that ramping manipulation is halved provides a very substantial gain of market efficiency in completing larger trades.

4.1. Discussion of Results What liquidity demanders must pay to get substantial size done (i.e., the effective spread), in the face of a doubled incidence of ramping alerts increases substantially even though the quotes for trivial size are largely unchanged. We find in examining the surveillance data across all listed securities on 34 exchanges 2000-2005 that the market efficiency-integrity relationship is much more sensitive when we account for actual trading prices rather than the quoted spreads on offer for doing trivial size. In 8 of 10 deciles, doubling alert incidence increases effective spreads from 3.7% to 8.2% whereas in price quotes, only one decile (i.e., decile 10) shows an effect on quoted spreads of this magnitude. So, it is institutional clients seeking larger trades and the pseudo market makers who provide liquidity to them who would be expected to pressure exchanges to detect potential manipulators and exclude them from the marketplace.

4

Decile 5 exhibits a 2.7% estimated increased in effective spread on a mean spread of 542 = a 15 basis point increase from doubling AI.

14

To understand further the roles of various market participants potentially caught up in a manipulator’s ramping transactions, we describe below the traders involved in capturing intraday profit. Closing price manipulations must be distinguished from legging patterns that arise in the normal functioning of quote-driven or order-driven markets. “Legging” refers to one side of the book changing while the other side does not (usually as large orders “walk the book”). In mean-reverting, stationary price sequences, most traders desire to get flat when a legging pattern develops. In non-stationary price sequences, however, trading profits are available in legging patterns by shorting one side of the market and inventorying the other. For that reason, intraday traders with sufficient access to order flow data to detect when a market begins to trend, often desire to participate in the trend. Manipulators know this and attempt to mimic the other natural players involved in legging patterns. Legging patterns arise from the buy-side trading of momentum traders, basket traders, and valuetraders operating with no discretion as to timing. And on the sell side, legging patterns arise from the trading activities of specialists, arbitrageurs, day traders, and fair-weather market makers. We now briefly describe the behavior of each in turn. On the buy side, momentum traders typically are impatient, especially if they are buying/selling into a rising/falling market. Such traders have a significant likelihood of walking the book with market orders of substantial size, thereby triggering a legging pattern. Basket traders rebalance institutional portfolios, submit arbitrage trades, and implement portfolio insurance. Baskets may trade as limit, marketable limit or market orders and hence will periodically walk the book, again triggering legging patterns. Also on the buy side, value traders may or may not be given discretion by their portfolio managers regarding their execution strategy. On the one hand, if the portfolio manager requires the value trader to acquire a particular stock quickly, the result is likely to be a significant price impact resulting from walking the book. Value traders who have discretion tend to be patient traders, willing to supply liquidity to earn the spread, or issuing marketable limit orders carefully so as to minimize price impact. However, some value traders with discretion will use the Value Weighted Average Price (VWAP) trading strategy to attempt to attain or better the day’s VWAP for traded stocks. When “mousetrapped” some distance from the VWAP at the end of the day, such trading may be difficult to distinguish from a manipulator’s ramping the close. On the sell side, specialists and designated liquidity providers have an affirmative obligation to refresh the book by selling into rising markets and buying into declining markets. If the ask side of the limit order book is being repeatedly hit, perhaps signaling movement toward a higher equilibrium price, these traders will temporarily ignore the bid side of the book--a dynamic adjustment pattern that is consistent with legging as well as ramping. Also on the sell side, those placing limit orders with no affirmative obligation are likely to fall into three categories: patient traders with money funds, arbitrageurs/day traders, and fairweather market makers.

15

Patient traders with money funds are attempting to minimize trading costs and therefore are not likely to create substantial price impact. Arbitrageurs and day traders may employ algorithms that monitor the market in real time examining various metrics to provide very short term estimates of the state of the market. Limit orders can be profitable in mean-reverting markets, but are picked off (if not cancelled) or left in the dust in trending markets. Thus arbitragers and day traders will cancel limit orders if their algorithm indicates a shift from a mean-reverting to a trending state. Such traders, depending upon the strength of their signal and their aggressiveness, may grab available liquidity in the direction of the trend and feed it back into the markets after the market settles into a mean-reverting state with limit orders. Such aggressiveness exacerbates legging patterns and may be mistaken for closing price manipulations. Fair-weather pseudo market makers will post buy and sell limit orders in mean reverting markets wherein supplying liquidity is profitable. Their trading algorithms will scan news feeds in real time and if significantly good/bad news is indicated the algorithm will cancel all limits, grab available limit orders and perhaps submit market orders to profit on the anticipated move in appropriate investment instruments. As markets again become mean reverting, the acquired instruments are fed back into the market with limits, and fair weather market-making resumed. Such pseudo market makers are highly dependent upon rapid execution and their aggressive behavior can exacerbate and make more feasible ramping manipulations.

4.1.2. Parameter magnitudes As to the magnitudes, a doubling of mean ramping alert incidence in decile 4 of 0.23 to 0.46 security days, increases the 790 basis points of mean effective spread across all listed securities in our 34 exchanges by approximately 3.9%. That means twice the average daily ramping incidence would be associated with 31 basis points of additional execution costs. In decile 1 (and also decile 3), no significant correlation is found with alert incidence which is consistent with the expectation that manipulations are more costly and difficult to implement in highly liquid securities. The elasticity of spreads with respect to alert incidence in the moderately-liquid deciles is much larger, as expected. Decile 6, for example, exhibits an elasticity of effective spread with respect to AI of 7.47, and decile 7 of 8.18%. The basis point impact on effective spreads is 39 b.p and 59 b.p., respectively, as AI doubles. In decile 10, a doubling of ramping alert incidence from its mean of 0.12 to 0.24, increases a much larger effective spread of 1356 basis points by only 3.4 percent in the Full Fixed Effects Model or 46 basis points. Other illiquid deciles exhibit similar results as well. For example, in decile 9, 5.4% elasticity in the Da Silva Autoregressive Moving Average Model raises the effective spread of 1091 by 59 basis points. In decile 8, 5.13% elasticity in the Da Silva MA model raises effective spreads by 40 basis points. The lower elasticities in deciles 8, 9 and 10 can be explained by the fact that detection of manipulation activities is too likely in the least liquid deciles, which drives manipulators away.

16

5. Determinants of Trade-Based Manipulation Although these correlations between AI and effective spreads using observational error components are compelling, and the magnitudes are economically significant, the determinants of cross-sectional variation in ramping incidence and the causal relations remain unknown. To identify the potential drivers of market manipulation, Aitken (2009) hypothesizes a role for each of the following: Regulation, Trading and Surveillance Technology, Security Market Infrastructure, and Market Participants. We discuss each of these hypotheses below. 5.1. Dimensions of Market Quality 5.1.1.

Regulation

Bhattacharya and Daouk (2002) confirm a negative relationship between the cost of equity capital and the enforcement of insider trading laws across 108 countries. Cumming and Sofia (2008) investigate the number of trading regulations with corresponding surveillance technology to monitor alerts and the existence of a trading regulation specifically against ramping across 25 exchanges. They find that comprehensive rules prohibiting trade-based manipulation generate higher turnover and larger market cap. We hypothesize that security exchanges with regulations specifically against ramping (Variable name: RampReg) are expected to have fewer ramping manipulations. A larger number of trading regulations (Variable name: Regs) that have corresponding surveillance alerts (e.g., against front running or other broker-client conflicts of interest) signify an atmosphere of more aggressive surveillance and should lower technical transaction costs of trading but may deflect surveillance resources away from the effort to stop ramping manipulation. Hence, we hypothesize a negative relationship between AI and RampReg, but a positive relationship between AI and Regs. 5.1.2.

Trading and Surveillance Technology

The recent revolution in internet-based order filing has provided investors access to a real-time and centralized order book with an expedited channel for order submission.

The latest development on some

exchanges (e.g., the London Stock Exchange, Direct Edge etc) is the effort to launch an Enhanced Liquidity Provider Program (ELP), which provides subscribed traders an integrated view of both displayed and dark pool order books. Historically, traders had to seek executions in either the displayed market or a single "dark pool". ELP now offers a comprehensive solution for traders looking to aggregate liquidity of all types before implementing particular execution strategies.5 To keep up with more sophisticated trading, real-time surveillance (RTS) technology has also been gradually replacing traditional T+N market surveillance or transaction log books. For example, SMARTS, the leading real-time market surveillance platform from Smarts Group International Ltd., has been deployed 5

According to Reuters (2008), Direct Edge's ELP Program broke the 100 million shares traded/per day mark on 15th July 2008 while the overall trading volume was 1.23 billion shares on that same day.

17

by more than 50 national securities exchanges and regulators around the world. But there has been no prior research studying the relationship between RTS technology and market quality. We will use the deployment of SMARTS as a proxy for eexperience with Real Time Surveillance (Variable name: RTS) technology. In cross-section we expect exchanges more vulnerable to trade-based manipulations to adopt RTS. Over time RTS should help reduce the incidence of market manipulation. Another technology advance is Direct Market Access (DMA) defined as electronic facilities which allow brokers to offer clients direct access to the exchange trading system through the broker’s infrastructure without manual intervention by the broker. DMA facilitates algorithmic trading and makes market manipulation more difficult. To be successful, market manipulators must avoid “signature footprints” and exit faster than counterparty hedge funds or proprietary trading desks who often adopt algorithmic trading with computer “bots.” Hal Varian refers to such datarati as “firm[s] whose business hinges on making smart, daring choices…gleaned from algorithmic spelunking and executed with the confidence that comes from really doing the math.” Such businesses are difficult to mousetrap on the wrong side of VWAP, to mislead into chasing false trends, or to manipulate at the close. DMA we predict will be inversely related to ramping alert incidence. 5.1.3.

Security Market Infrastructure

The following dimensions of infrastructure are expected to impact market integrity: (1) the presence of a closing call auction, (2) volatility defined as the standard deviation of daily returns, (3) market liquidity defined as the market turnover, and (4) the technical transaction costs measured by quoted spreads. Many securities exchanges have introduced closing call auctions (Variable name: CallAucDum) to improve market quality but have achieved mixed results (Pagano and Schwartz 2003, Comerton-Forde and Rydges 2006, etc). In our context, by allowing traders to unwind their intraday positions and go flat overnight if so desired, closing call auctions should reduce the exposure to manipulation resulting in tighter spreads as AI declines. Higher volatility (Variable name: Vol) leads to less aggressive order placement as non-execution risk declines for any given picking off risk (Foucault 1999). Thus, volatility results in wider spreads, which would increase the cost of manipulation resulting in a smaller number of alerts. But manipulation carried out on securities with higher volatilities should have a lower probability of detection and enforcement since in volatile price environments less ramping alerts are triggered and legal safe harbors widen. Securities that are highly liquid normally have big market capitalization and are therefore difficult to manipulate due to the higher total costs involved to entice momentum traders to chase a false trend. Hence, we expect higher market liquidity (Variable name: Liq) to be associated with a lower number of alerts. However, a large proportion of alerts triggered each day are false positives. The problem of false positive ramping alerts is well known in surveillance research to be correlated with liquidity.

18

Quoted spreads (Variable name: QSpr), as a measure of the percentage cost of a round trip transaction at the BBO, is directly proportional to the technical transactional costs involved in manipulations. It is expected that higher quoted spreads will cause lower market manipulation and alert incidence, ceteris paribus. In contrast, we showed in section 4 above that effective spreads are positively correlated with alert incidence. The two are not inconsistent. With so little depth at the BBO in electronic markets, quoted and effective spreads do diverge. Moreover, in theory, we expect this divergence in relationship to AI. When non-execution risk is high (for any given picking off risk), liquidity traders prefer market orders or limit order close to the BBO to gain immediate execution. When non-execution risk declines or picking-off risk rises, limit orders away from the BBO are preferred. Foucault’s (1999) theory predicts that when price volatility increases (perhaps because of manipulation) and non-execution risk consequently declines (for any distance from the BBO), liquidity traders will tend to place less aggressive limit orders in order to reduce picking off risk, which results in wider effective spreads. 5.1.4. Market Participants

Several emerging markets have encouraged foreign capital investment in their equity markets hoping that overseas hedge funds and proprietary trading desks would boost liquidity and tend to stabilize the financial market. But research on several financial crises in emerging markets has tied those foreign investors to excessive volatilities or bubbles. Chiou, et al (2007) has two related findings. The first is that extremely low transaction costs and fast adjustment of order placement with cancellations, pinging, etc. create greater chances for the informed to manipulate the market and take advantage of the uninformed. This reasoning implies quoted spreads would be a negative determinant of alert incidence. Chiou also finds that informed foreign direct investors have a greater incentive to manipulate the market because they are beyond the reach of national security market regulators. Using the United Nation’s data on foreign direct investment (FDI) as a percentage of GDP, we predict a positive relationship with AI. 5.2. Research Design To study the cross-sectional determinants of trade-based manipulation we decided to focus not on heterogeneous enforcement actions in highly divergent regulatory regimes across exchanges but rather on the comparable data that is publicly-available worldwide, ramping alert incidence data. As a maintained hypothesis we assume that trade-based manipulation more generally can be well represented by this ramping alert proxy. Again, regulators and exchange officials investigate nine other forms of securities market manipulation using proprietary databases (Ji 2009). The empirical model structure is a simultaneous set of three structural equations describing ramping alert incidence (AI), the quoted spread (QSpr), and the deployment of real-time surveillance (RTS) systems: AI = f (QSpr, RTS, Control variables, Fixed effects)

(I)

19

QSpr = g (AI, RTS, Control variables, Fixed effects)

( II )

RTS = h (AI, QSpr, Control variables, Fixed effects)

( III )

The empirical specifications prove to be highly non-linear with lognormal transformations and a probit equation. Nevertheless, because each of the endogenous variables could in principle affect the others, we assure the order condition for identification by excluding from each equation two control variables (3 endogenous-1) present elsewhere in the system. In each equation, the excluded variables are control variables found to be insignificant in preliminary single-equation estimations of the focal equation but highly significant in other equations. In addition, we test for and thereafter incorporate into the model, exchangespecific fixed effects to address the idiosyncratic institutional features prevalent throughout the sample of 24 security markets.6 We hypothesize the following regression relation between the level of ramping manipulation or conversely the level of integrity of an exchange and the measures for Regulation, Technology, Security Market Infrastructure, and Market Participants discussed above: 7 +

(I’)

where = Mean number of daily ramping alerts per security in market i, = Dummy variable for the deployment of a Real Time Surveillance system in market i, = Mean Standard deviation of daily returns of securities in market i (a potentially endogenous variable) = Market turnover per security in market i = Mean quoted spread measuring the round-trip transaction costs at the BBO for securities in market i at time t (another potentially endogenous variable) = Dummy variable for the existence of a Closing Call Auction in market i = Dummy variable for the existence of Direct Market Access in market i = Dummy variable for the existence of a Ramping Regulation in market i = the residual error term (perhaps a negotiated fee for block execution)

6

These 24 are the only members of our 34 exchange sample for which detailed regulatory data are available.

7

The absence from this list of information generation and disclosure (Pagano and Roell 1996, Lang and Lundholm 1996, and Oved 2002) is intentional because trade-based manipulation rather than insider trading or false rumor dissemination is the focus of our research.

20

In deciding where (in which securities and markets) and when to execute, ramping manipulators consider the round-trip technical transaction costs (the quoted bid-ask spread), any requisite fees and commissions, the trading volume required to elicit a desired price impact, and the frequency and severity of civil and criminal penalties. A security’s baseline price volatility (i.e., unramped) often determines the likelihood of detection by surveillance officials as well as the availability of legal “safe harbors” that reduce the probability of indictment and conviction. Real-time surveillance (RTS) tends to be deployed in exchanges that perceive a greater vulnerability to manipulation and at least initially therefore often experience the higher volatility that accompanies more manipulation alerts. 5.2.1 Simultaneity Given the potential endogeneity of RTS and volatility(VOL), Hausman-Wu specification tests are conducted to determine whether simultaneity issues significantly bias the estimates from the above OLS model of alert incidence. We estimate RTS and VOL as well as QSPR as instrumental variables with a twostage least squares (2SLS) procedure and then test for parameter equivalence between the IV and OLS estimates. For example, we predict Volatility by using the following IV regression relation:

where

= the number of trading regulations that have corresponding surveillance alerts in market i,

and other variables are as defined previously. Although Stoll and Whaley (1987, 1991), Chamberlain, Chueng and Kuan (1989), and Chiou, et al (2007) all find evidence that price volatility is higher during the time period of manipulation within a security market, we find little evidence that higher AI increases volatility across securities or exchangespecific liquidity deciles. Specifically, aggregating individual securities into 10 liquidity deciles (by turnover) for each of 24 exchanges in 2005, volatility and quoted spread are both exogenous in the crosssectional AI equation we study (Hausman-Wu χ2 test 2.09 with fixed effects for 1 d.o.f. yields α = 0.143). In contrast, in a time-series cross-sectional model structure, volatility would be endogenous a priori. Similarly, we predict the decision to deploy real-time surveillance using the following specification:

where

= Foreign direct investment as % of Gross Fixed Capital to the country of market i,

21

and other variables are as defined previously. Prob(RTS) proves to be endogenous in the structural equations model I, II, III, as expected.8 Volatility and the rest of the independent variables are specified in raw data form as control variables, after confirming their exogeneity. 5.2.2. Other Econometric Issues We find in the Probit analysis of the decision to deploy real-time surveillance that the determinants of alert incidence and of quoted spreads influence the decision by an exchange as to whether to adopt RTS. That is, whether to have an RTS capability in Hong Kong may depend upon the presence of order filling with Direct Market Access (DMA) or the ability to unwind positions in a closing auction (CloseAucDum). The desire to assure market integrity by reducing the vulnerability to manipulators would lead to more adoptions of RTS, ceteris paribus. And the combination of RTS and DMA or RTS and CloseAucDum would then have a rather different impact on spreads and manipulation alert incidence than DMA or CloseAucDum taken alone. Consequently, the parametric effects of market design changes in those exchanges that adopt RTS may differ from those that do not. In the estimation of the three structural equations, we therefore explored the possible impact of full interaction terms between the deployment of RTS and all the r.h.s. variables using ML estimation of a Heckman-style selectivity bias model. The results are qualitatively almost identical to and beyond the scope of the present study. Beyond potential simultaneity and selectivity bias, there are several other econometric issues to resolve: 1) the pooling of trading and surveillance data across all liquidity deciles, 2) the relevance of exchange-specific fixed effects, and 3) the likely cross-equation correlation of the error terms and the consequent need for systems estimation of equations I, II, III. We address each of these issues below. 5.3 Regulatory Data We again employ the Reuters database maintained by the Securities Industry Research Centre of Asia-Pacific (SIRCA). This database contains intra-day trade and quote data for more than 200 world markets 19992005. The Trading Regulations Database from Cumming and Sofia (2008) covers 25 security exchanges for the years 2005 to 2008. Consequently, our analysis of the determinants of trade-based manipulation addresses the single year 2005 across all ten liquidity deciles on the overlapping 24 exchanges. Table 1 lists the securities exchanges studied—14 in the Asia-Pacific region, 6 in Europe, 2 in the U.S., and 2 in Africa. The analysis is conducted on the entire universe of trades and quotes for all listed securities. We aggregate the data to obtain a unit of analysis that is an exchange-specific liquidity decile based on their monthly trading turnover. We term this unit of analysis an “exchange decile” for short.9

8

The pooled OLS and IV (2SLS) estimations across all 10 deciles and 24 markets that underlie these Hausman-Wu tests are available from the authors. The disaggregated results for thickly-traded deciles 1,2,3 and moderately-liquid deciles 4, 5, 6 and 7, and for thinly-traded deciles 8, 9, and 10 are reported below in Tables 9, 10, and 11.

22

5.3.1. Descriptive Statistics on Determinants of AI The descriptive statistics for the following variables are presented in Table 7 below. •

Annual Alert incidence of daily ramping manipulation per security per decile across the 24 securities exchanges in year 2005 (AI);

•

Annual Average Quoted Spread per security per decile across the 24 securities exchanges in year 2005 (SPR);

•

Annual Average Standard Deviation of Logarithmic Daily Return per security per decile across the 24 securities exchanges in year 2005 (Vol);

•

Annual Average Turnover per decile across the 24 securities exchanges in year 2005 (Liquidity);

•

Number of Trading Regulations that are surveillance monitored across the 24 securities exchanges in year 2005 (Regs).

It can be seen that except for the number of trading regulations that are surveillance monitored (Regs), all the other 4 variables are demonstrably non-normal. For example, the quoted spread has a mean of 5.36% and a standard deviation of 0.0956 with skewness of 3.4166 and kurtosis of 13.1504. After a natural log transform, we observe the distribution of

to be approximately normal (

XXX = -3.9911 and

= 1.4776) with skewness of 0.2425 and kurtosis of -0.6729. The same is also observed for AI, Vol and Liquidity. Using the properties of the lognormal distribution and assuming exact log-normality for our observations, an estimator of

would be

. The Reuters data yields such an estimate of

= 5.51%. This figure closely approximates but differs from the observed mean of 5.36% because our sample differs slightly from a pure lognormal distribution. A quoted spread of 536 basis points at the mean conveys that this sample of 24 securities exchanges is very different from the lowest execution costs worldwide where the DJIA stocks trade in New York for 11 basis points. However, the more appropriate comparison is to the universe of all ten liquidity deciles where Aitken, Cook, Harris, and McInish (2009) report 61 b.p. for NYSE, 257 b.p. for ASX, 283 b.p. for TSE, 303 b.p. for NASDAQ, 371 for Euronext, and 381 for Xetra in matched samples. 5.3.2. Final Empirical Specification In the error components model of sections 2, 3 and 4, we found that exchange-specific fixed effects contribute significant explanatory power to the relationship between relative spreads and ramping alert 9

To check for any aggregation bias, we introduced fixed effects for the liquidity deciles, omitting decile seven. Not surprisingly, because a continuous measure of liquidity itself is a right-hand-side variable in all of our models, the results were qualitatively identical.

23

incidence in many deciles. Therefore, we also investigated the effects of these idiosyncratic institutional factors here in the structural equations. Three parameters in the AI equation change sign and significance when the fixed effects are included (Spr, DMA, and RTS), and the R-squared rose from 0.062 to 0.443. All the parameters in the spreads equation (except Regs) were stable, but R-squared again rose from 0.416 to 0.499 and from 0.597 to 0.906 in the RTS equation. Based on those findings and the four variables that are near log-normally distributed, we transform equation I for estimation to the regression relation:10

+

where

(I’)

= a dummy variable for each of 23 securities exchanges listed in Table 1; the Egyptian

Stock Exchange (CAI) is omitted as the modal observation. We retain in each final specification all exchange dummy variables found to be significant at α ≤ 0.05. In estimating I’, the pooling of thickly-traded stocks in liquidity deciles 1,2, and 3, the moderatelyliquid stocks in deciles 4-7, and the thinly-traded stocks in deciles 8,9, and 10 is rejected by a Wald test (F yielding α < 0.01). Estimation of the quoted spread and real-time surveillance equations II’ and III’) yields the same result. Possible heteroskedastic error variances across the liquidity decile groupings necessitate Wald tests, and these too reject the pooling of the liquidity decile data. Consequently, we perform separate estimations for these 3 subsets of the ten liquidity deciles throughout our subsequent analysis.

6. A Simultaneous Equations Model 6.1 Endogeneity 6.1.1. OLS, 2SLS Results for Moderately-traded Deciles In the OLS estimation for moderately-liquid securities (shown in Table 8 Panel A), closing call auctions and direct market access (DMA) reduce alert incidence. These design features may well allow counterparties to unwind their intraday exposures before potential manipulators can execute ramping manipulation strategies. Closing Call Auctions (Call) also lower quoted spreads while increasing volatility. Real time surveillance 10

As a robustness check, we relax the log linear functional form specification of the AI and Spread models in equation (1) by employing maximum likelihood estimation.

24

(RTS) is deployed in asymmetric information environments more vulnerable to manipulation and therefore is associated in the cross section with higher alert incidence and higher spreads. Higher technical transaction cost measured by the quoted spread are themselves associated with higher volatility, but controlling for this effect, the deployment of RTS lowers volatility. Similarly, the presence of regulations specifically prohibiting ramping (RampReg) suggest a perceived vulnerability to manipulators and proves to be associated in cross section with higher spreads but like RTS also results, ceteris paribus, in lower volatility. The 2SLS results for the determinants of AI shown in Table 8 Panel B are qualitatively identical to the OLS estimates in Panel A. The instruments for volatility (Volhat) and Spreads (Sprhat) are introduced one at a time and tested separately for exogeneity. Hausman-Wu specification tests fail to reject the parametric equivalence between the OLS and IV (2SLS) estimates (e.g., Hausman-Wu χ2 test 2.92 for 1 d.o.f. yields α = 0.09). We conclude that simultaneity bias in the thickly-traded deciles is not material for the AIVolatility equation pair or for the AI-Spread equation pair. Subsequent system estimation will address, however, the remaining possibility of cross-equation correlation of the error terms, using seemingly unrelated regressions and maximum likelihood estimation. As to the IV equations for volatility and spreads also in Panel B, larger numbers of security market regulations about market integrity accompanied by compliance monitoring (Regs) result in lower spreads but increased volatility. For example, surveillance effort devoted to insider trading violations or front running may deflect attention from the ramping regulations that most directly influence volatility. Greater turnover (Liq), as expected, also lowers quoted spreads but raises volatility. Five of 23 exchange dummy variables were statistically significant in the AI equation, 13 in the volatility equation, and 20 of 23 in the spreads equation. To uncover so many significant determinants despite the presence of all these fixed effects echoes the validity of the modeling framework being proposed. Overall, the F-stats are 6.86***, 10.09***, and 161.97***, and R-squared rises to 0.71, 0.70, and 0.98 for the AI, Vol, and QSpr models, respectively. 6.1.2. OLS, 2SLS Results for Thickly-traded Deciles In the OLS results in Table 9 Panel A, several differences emerge between moderately-liquid and thicklytraded stocks. As hypothesized, the extraordinarily high transaction costs of manipulation in the thicklytraded securities causes higher quoted spreads to reduce volatility whereas higher spreads increase volatility in the moderately-liquid deciles. Positive and significant AI-spread relationships in OLS estimation disappear in 2SLS estimation. Alert incidence declines with DMA infrastructure but is otherwise difficult to detect and unrelated to the hypothesized determinants. The Hausman-Wu test for the exogeneity of volatility rejects the equivalence of the OLS and IV estimates (Hausman-Wu χ2 test 11.53 for 1 d.o.f. yields α < 0.01). Volatility among thickly-traded securities does appear to be a function of alert incidence and its determinants. Specifically, referring to the IV estimation in column 2, volatility increases with closing call auctions and declines with spreads, RampReg, and other monitored integrity Regs. In addition, the

25

deployment of real-time surveillance systems decreases return volatility by an even larger magnitude in these thickly-traded than in the moderately-liquid stocks. As to the AI-Spread (market integrity–market efficiency) equation pair, we again conclude based on the Hausman-Wu test that simultaneity bias is minimal (Hausman-Wu χ2 test 0.40 for 1 d.o.f. yields α = 0.53). In the IV estimation for spreads, shown in the final column, quoted spreads decrease with closing call auctions and integrity regulations that are monitored. In cross section, higher spreads are associated with the presence of real-time surveillance and a regulation specifically prohibiting ramping. These results are consistent across both moderately-liquid and thickly-traded stocks. Again, subsequent investigation will address SURL and ML estimation because of possible cross-equation correlation of the error terms. Three of 23 exchange dummy variables were statistically significant in the AI equation, 12 in the volatility equation, and 18 of 23 in the spreads equation. Overall, the F-stats are 6.86***, 10.09***, and 161.97***, and R-squared for the thickly-traded securities increases with fixed effects to 0.71, 0.93, and 0.97 for the AI, Vol, and QSpr models, respectively. 6.1.3. OLS, 2SLS Results for Thinly-traded Deciles It can be seen from Table 10 that in the thinly-traded securities deciles, higher spreads are associated with greater volatility in cross-sectional OLS and 2SLS estimates. Call auctions at the close to unwind intraday exposures lowers spreads. Direct Market Access in these thinly-traded securities appear to advantage the manipulators and raise the incidence of alerts, perhaps because with little liquidity available counterparties at the end of a ramping scenario cannot find buyers even though execution speed is very fast. Unlike in thickly-traded and moderately-traded deciles, real time surveillance and ramping regulations lower both alert incidence and volatility in OLS estimates while continuing to be associated with higher spreads. Volatility, as usual, raise spreads suggesting more asymmetric information. Market quality assurance in the form of regulations promoting fair and orderly markets lowers the spread. Liquidity (such as it is) in these thinly-traded stocks is also sufficient to lower the spread despite higher volatility. Again, Hausman-Wu tests reject the hypothesis that IV estimations for Vol and QSpr are needed; OLS single equation results are unbiased. Six of 23 exchange dummy variables were statistically significant in the AI equation, 5 in the volatility equation, and 20 of 23 in the spreads equation. Overall, the F-stats are 3.48***, 19.90***, and 32.47***, and R-squared for the thickly-traded securities increases with fixed effects to 0.65, 0.92, and 0.95 for the AI, Vol, and QSpr models, respectively. 6.2. Probability of Real-time Surveillance Surveillance of financial markets has a long history. The reasons why are both obvious and subtle. The assurance of market integrity typically requires an aggressive surveillance regime in tandem with regulatory enforcement against those who conduct prohibited practices. In addition, however, as self-regulatory organizations (SROs), many exchanges have more extensive obligations to monitor trading, detect

26

manipulative behavior, and punish violators than might exist in an industry like insurance that operates under detailed and continuous regulatory review and approval. Real-time surveillance has grown more sophisticated in the last decade concurrent with the growth of electronic (and especially algorithmic electronic) trading. Today, lower latency and an explosion of trade executions barely imaginable a few years ago, today necessitate real-time mechanisms for capturing and processing surveillance data. RTS systems have become a more prevalent response to heightened SRO obligations and are now deployed in 8 of the 24 exchanges we study. 6.2.1. Probit Model Specification (III’)

where

= The probability of the deployment of Real Time Surveillance System (i.e., SMARTS) in market i = an instrument for the mean security days with Ramping Alerts in market i = the average standard deviation of daily returns of market i = The average market turnover of market i = The dummy variable for the existence of direct market access of market i = Foreign direct investment as % of gross fixed capital to the country of market i = Residual error term

A priori, we expect the deployment of real-time surveillance systems to increase with an instrumental variable for greater alert incidence (AIhat), greater vulnerability to manipulation by foreign investors (FDI) especially those using DMA, and with higher turnover (Liq). Increased turnover magnifies the problem of false positives in surveillance monitoring, and RTS can help distinguish true from false positives. In addition, we expect RTS to decline when regulatory attention is deflected to client-agent issues, insider trading, or other integrity regulations (Regs) and when higher return volatility (Vol) makes prosecuting and convicting manipulators more difficult. 6.2.2. Empirical Results for Prob(RTS) Table 11 reports our PROBIT analysis of real-time surveillance system deployment. Panel A pools the results from all liquidity deciles; subsequent panels B, C and D, report the sub-groupings of liquidity

27

deciles. In the All Deciles results in Panel A as well as in the moderately-liquid securities in Panel B, alert incidence is positively related to the adoption of RTS, indicating a perceived vulnerability to manipulation that RTS can help mitigate. Secondly, direct market access (DMA) facilitates quick responses by both manipulators and counterparties, requiring an expanded capability by the market surveillance officials to monitor the situations as they evolve. More extensive integrity regulations (REG) serve as something of a substitute for RTS, deflecting attention from and reducing the likelihood of real-time surveillance. Higher liquidity (LIQ) increases RTS in thickly-traded and moderately liquid stocks, perhaps because higher turnover accentuates the problem of false positives in scrutinizing potentially manipulative trades. And RTS assists in separating the true and false positives in a surveillance program. Finally, a higher percentage of foreign direct investment (FDI) raises the vulnerability of an exchange to ramp-and-dump manipulative schemes, so RTS increases to combat it. Overall 16 out of 23 exchanges have significant idiosyncratic effects, and the LR rises from 94.22 to 289.75 when these exchange-specific fixed effects are incorporated into the specification. In Panel B (deciles 4, 5, 6 and7), alert incidence is at its peak because costs of trade-based manipulation are moderate (unlike in highly liquid securities), yet detection is still quite difficult (unlike in thinly-traded stocks). For these moderately liquid stocks, higher alert incidence, DMA, higher liquidity, and foreign direct investment increase RTS. In addition, more numerous integrity regulatory burdens separable from ramping regulations reduce the adoption of RTS. Overall 11 out of 23 exchanges have significant idiosyncratic effects, and the LR rises from 59.52 to 115.9 when the exchange-specific fixed effects are incorporated into the specification. In Panel C for the most liquid deciles, many of the aforementioned results are quite comparable: DMA, higher liquidity, and foreign direct investment increase RTS, whilst more numerous regulations unrelated to ramping reduce the adoption of RTS. On the other hand, higher volatility reduces the deployment of RTS. In these most liquid stocks not only is trade-based manipulation more difficult to detect whatever the means of surveillance, but in addition, higher volatility offers a safe harbor legal defense to alleged manipulators. This makes successful prosecution and enforcement actions much more difficult. In addition, alert incidence itself is negatively-related to RTS suggesting that with the safe harbors attributable to high volatility and the surveillance false positives attributable to high turnover, these most liquid stocks are not the intended target for the deployment of sophisticated RTS systems. Rather many exchanges and regulators are more focused on the moderately-liquid stocks where detection of manipulators remains difficult without sophisticated RTS technology. Overall 3 out of 23 exchanges have significant idiosyncratic effects, and the LR rises from 43.91 to 86.92 when the exchange fixed effects are incorporated into the specification. Finally, in Panel D for the least liquid stocks, only DMA and FDI consistently increase the deployment of RTS. Again, a larger number of regulations unrelated to ramping reduce the deployment of

28

RTS. Variation in volatility probably does not matter because manipulation in these thinnest trading stocks is easy to detect. Liquidity probably does not matter because so little successful manipulation is attempted in these 8,9,10 decile stocks that false positives seldom arise. Overall 16 out of 23 exchanges have significant idiosyncratic effects in these thinnest stocks, and the LR rises from 24.73 to 86.92 when the exchangespecific fixed effects are incorporated into the specification.

7. Systems Estimation of the Structural Equations Having established that the probability of deployment of RTS is itself related to alert incidence, volatility, and the determinants of spreads, we estimate a simultaneous system of equations I’’, II’’, and III’’ characterizing integrity (AI), efficiency (QSpr), and the likelihood of RTS (ProbRTS): lnAIi = f (lnQSpri, ProbRTSi, Controls{Calli, DMAi,, RampReg i, lnVoli}, 23 Fixed effectsi) + ui ( I’’ ) lnQSpri = g (ProbRTSi, Controls{Calli, RampRegi, lnVoli, lnRegsi}, 23 Fixed effectsi) + vi

( II’’)

Prob(RTS)i = h (lnAIi, Controls{DMAi, lnRegsi, lnVoli, lnLiqi, lnFDIi}, 23 Fixed effectsi) + wi

( III’’)

The estimation is cross-sectional for the entire year of daily trading data in 2005, encompassing all the listed securities across 24 exchanges aggregated into 10 exchange-specific liquidity deciles. As throughout our previous work, all ten exchange-deciles can not be validly pooled and are instead grouped for estimation into thickly-traded deciles 1, 2, 3, moderately-liquid deciles 4, 5, 6, 7, and thinly-traded deciles 8, 9, 10. A priori, as we hypothesized in explaining our research design in section 5.2, each endogenous variable could affect all the others. Our previously-discussed empirical findings on simultaneity and model structure indicate two exceptions. Alert incidence does not affect quoted as opposed to effective spreads, consistent with our study of error components in section 2, although QSpr is a determinant of AI. And second, quoted spreads do not affect Prob(RTS). The order condition for identification is satisfied by excluding Regs and FDI from the AI equation, FDI and DMA from the QSpr equation, and RampReg and Call from the Prob(RTS) equation. In each case, these identifying variables were insignificant at α = 0.05 in the equation from which they were omitted. Because of possible cross-equation correlation of the error terms ui, vi, and wi, we first estimated seemingly unrelated regressions (SURL) and found the results qualitatively identical to the OLS regressions discussed above, suggesting Cov(ui,vi) = Cov(ui,wi) = Cov(vi,wi) = 0.11 The 2SLS estimates reported in Table 12 provide therefore a valid model structure for estimating the determinants of AI and the effects of the dichotomous decision to deploy RTS. However, RTS is a limited dependent variable. The appropriate

11

Hausman-Wu tests also indicate parameter equivalence between OLS and SURL.

29

estimation method for the (AI, QSpr, ProbRTS) system of equations I’’, II’’, III’’ involving a probit equation is maximum likelihood; we employ the SAS QLim procedure. 7.1. Real-time Surveillance Equation Prob(RTS) Table 12 Panel A reports on the determinants of the decision to deploy real-time surveillance systems. Some of the results do not differ by liquidity decile grouping. Higher alert incidence raises Prob(RTS) in all deciles. The presence of DMA lines as well as higher foreign participation in the security industry (FDI) increase the likelihood of RTS systems deployment in all deciles. Sophisticated algorithmic execution using DMA and the foreign investment banks and proprietary trading desks that often trade algorithmically necessitate the more sophisticated surveillance that RTS systems offer. Higher FDI is widely thought to increase threats to market integrity from hedge funds and proprietary trading desks offshore. Higher turnover (Liq) in both the most liquid deciles 1, 2 and 3 and the moderately-liquid deciles 4,5,6,7 also increase Prob(RTS) primarily because of the accentuation of the false positives problem as trading volume rises. Again, RTS systems can help distinguish true from false positive alerts. In the thick and thin-trading deciles, more numerous integrity regulations (REGS) serve as a partial substitute for or deflect attention from ramping surveillance and decrease the likelihood of deploying an RTS system. For example, client-broker conflicts compete for surveillance monitoring attention and regulatory enforcement. Also in the thick and thin-trading deciles, increased volatility reduces the likelihood of deploying RTS systems. Not so in the moderately liquid deciles. Apparently, the legal safe harbor created for ramping defendants by more volatile prices and returns does not dissuade the regulators and exchanges from deploying sophisticated RTS-based surveillance systems where the chances of trade-based manipulation are greatest. In the moderate liquidity deciles 4, 5, 6, 7, the Probit model exhibits a Likelihood Ratio of 115.9 and an Aldrich-Nelson statistic of 0.55. In the thick-trading and thin-trading deciles, the Likelihood Ratio declines to 86.9, but the Probit model still exhibits an Aldrich-Nelson statistic of 0.55. Overall, the All Deciles estimation exhibits a Log Likelihood of 289.7 with again a 0.55 Aldrich-Nelson statistic, suggesting the determinants of Prob(RTS) are more comparable across all ten liquidity deciles than are the determinants of AI and QSpr. We wish to reemphasize two points of difference however. Two variables in deciles 1, 2, and 3 and in deciles 8, 9 and 10 reduce the likelihood of RTS system development: Volatility and Regs. First, recall that more volatility offers legal safe harbors to alleged manipulators seeking to avoid prosecution, so RTS is less valuable as a tool for assuring market integrity in those cases because of the difficulty of detecting violators and securing convictions. Second, an emphasis on insider trading or other integrity regulations unrelated to ramping appears to decrease the willingness to invest in RTS in exchanges characterized by thin trading and in the exchanges with highly liquid securities. 7.2. Integrity Equation (AI)

30

In the moderate liquidity deciles 4,5,6,7 in Panel B of Table 12, we find that the presence of a closing call auction (Call) reduces the incidence of ramping alerts. Trade-based manipulation proves more difficult when a manipulator’s counterparties can use closing auctions to unwind their intraday exposures. High speed execution on direct access (DMA) lines also results in fewer ramping alerts and increased market integrity. Our interpretation is that counterparties are able to employ DMA to set preventative algorithms in place that render ramp-and-dump manipulation unprofitable. The probability of deployment of RTS is significantly positively related to alert incidence. In the absence of any multi-year panel data on the dynamic effects of RTS, what we are observing in cross section is the perceived vulnerability of certain exchanges to manipulation and their consequent deployment of RTS plus the regulatory regimes required to assure better market integrity. Two exchanges (Shanghai and Shenzhen) have significant dummy variable effects at α = 0.05 (both negative, suggestive of greater market integrity), and the alert incidence model for moderate liquidity deciles exhibits an OLS adjusted R-squared of 0.327 with a 99% F test (5.20,11). For the most liquid deciles 1, 2 and 3, also in Panel B, direct access execution on DMA lines again results in fewer ramping alerts and increased market integrity. In addition, we find that higher quoted spreads (which are associated with more asymmetric information in these most liquid securities) increase the incidence of ramping alerts. Manipulation always necessitates substantial trading volume but especially so in these highly liquid securities, and the more asymmetric the information, the more likely momentum traders will become involved in reaching that trading volume required for manipulators to execute a rampand-dump strategy. None of the numerous other determinants from the moderately-liquid subsample are significant, suggesting that manipulation is often prohibitively costly to execute and true positive surveillance alerts are less effectively distinguished from false positives in these most liquid stocks. Three exchanges (Shenzhen, Korea and New Zealand) have significant dummy variable effects at α = 0.05 (all negative, suggestive of greater market integrity). The alert incidence model exhibits an OLS adjusted Rsquare of 0.665 and a 99% F test (16.69, 9). For this moderately-liquid set of securities, the Log Likelihood statistic for the ML estimation is -209.3. In the thin liquidity deciles 8,9,10, displayed in the third row of Panel A, we find that the presence of a regulation specifically prohibiting ramp-and-dump manipulation (RampReg) reduces the incidence of ramping alerts. Potential manipulators appear convinced that the exchange will pursue violators who are relatively easily detected in these thinly-traded securities, and consequently fewer ramping alerts are observed. Two exchanges (Xetra and New Zealand) have significant dummy variable effects at α = 0.05 (the first negative, indicative of more market integrity in thinly-traded securities, but the second positive, indicative of lower market integrity). None of the other determinants from the more liquid deciles are significant, suggesting again that the “action” in trade-based manipulation is focused not here but on the moderately-liquid stocks where surveillance-based detection is much more difficult. The alert incidence model for these thinly-traded deciles

31

exhibits an OLS adjusted R-square of 0.401, a 99% F test (6.29, 9), and the Log Likelihood statistic for the ML estimation is -175.79. 7.3. Efficiency Equation (QSpr) Although our primary focus remains on alert incidence and real time surveillance, Table 12 also displays results for the determinants of time-weighted quoted spreads (Panel C). In all deciles, closing call auctions and more regulations in pursuit of market integrity (Regs) lower quoted spreads. The probability of deployment of RTS and a regulation specifically prohibiting ramping (RampReg) indicate in cross-section the perceived likelihood of more ramping which as we have seen in the previous OLS and 2SLS estimations is associated with higher quoted spreads. Twenty of 23 exchanges have significant fixed effects at α = 0.05 (both negative and positive) in the moderately liquid and thinly-traded deciles. Adjusted R-squares range from 0.862 to 0.904 with 99% F tests. For the most liquid deciles, again closing call auctions and more regulations in pursuit of market integrity decreases the quoted spread.

8. Summary and Conclusion Our research examines trade-based market manipulation at the close and the relationship between this market integrity concept and the relative spread, a standard measure of market efficiency. Using market surveillance techniques we calculate ramping alert incidence (AI) across securities on a daily basis. AI is a proxy for the inherently unobservable presence of a manipulator. The event of information arrival is also inherently unobservable. Consequently, we first model the error components of AI and SPR across stocks and over time as observational errors in a random effects model. Later we provide a factor analysis of the determinants of AI. The unit of analysis is the exchange-specific liquidity decile across 34 securities markets worldwide in 2000-2005. We develop a hypothesis that trade-based manipulation as proxied by the incidence of ramping alerts raises execution costs for completing larger trades. If mean reversion of prices the next morning following a statistically extraordinary price marking the close represents ramping manipulation, then we would expect limit order placers to reduce their order aggressiveness. Quoted spreads for trivial depth might remain unchanged but limit orders for larger volume would be spread farther away from the BBO to avoid being triggered by a manipulator’s walking the order book and then liquidating/covering his position before other traders could react. This rational response to the anticipated presence of a possible manipulator would show up as higher effective spreads in both order-driven and quote-driven markets averaged over long periods. The alternative hypothesis is that ramping alerts represent information arrivals that are delayed, unmasked as rumors, or proven false. Such information arrivals that are quickly reversed would not lead to wider effective spreads when averaged over long periods. Our findings on the relationship between market integrity and efficiency are two-fold. First, variation in trade-based manipulation across security exchanges and over time does raise the execution costs of larger

32

trades. We find ramping alert incidence positively related to effective spreads in 8 of 10 deciles from the most liquid to the thinnest-trading securities. In contrast, as expected, quoted spreads are largely unaffected by alert incidence, after controlling for full time-series and cross-sectional fixed effects. Second, the magnitude of the decrease in effective spreads when ramping manipulation incidence halves is economically significant -- i.e., 31 to 59 basis points worldwide. This compares with an average effective spread of 27 basis points for all deciles of securities on the NYSE (Aitken, Cook, Harris, McInish 2008), and 219 basis points in the most liquid decile 1 stocks worldwide. As we have shown, ramping manipulation is most likely in the moderate liquidity deciles 4, 5, 6, and 7 where surveillance detection is less likely than in illiquid stocks and yet capital requirements for a successful manipulation are somewhat reduced relative to the most liquid stocks. In these moderate liquidity deciles, effective spreads are much higher averaging 523 to 790 basis points across the 34 exchanges. Therefore, the effect of doubling or halving AI is approximately a 10% change in execution costs for completing larger trades. In sum, we find market integrity does substantially affect market efficiency worldwide. We conjecture that ramping manipulation has little or no effect on the quoted spread, yet decreases market efficiency farther up the book because so little depth is posted at the BBO on electronic limit order books (typically only 50 to 100 shares). Time-weighted quoted spreads only reflect technical transaction costs like inventory, clearing and settlement costs, not active management of trading interests. Volumeweighted effective spreads reflect the cost of completing larger trades that are material to trading desk and portfolio management performance. Layering of order placements up and down the limit order book is a rational way for active traders to participate in price moves yet capture intraday trading profit (Aitken, Almeida, Harris, and McInish 2007). As ramping manipulation activity increases, rational layering requires less aggressive order placement than would be optimal in the absence of ramping manipulation. As to the determinants of ramping incidence from our simultaneous equations cross-sectional model, we find direct market access (DMA) reduces ramping manipulation. Our interpretation is that DMA lines and the algorithmic countertrading strategies they facilitate can circumvent the pump and dump tactics of a ramping manipulator. Similarly, call auctions at the close provide an escape mechanism for counterparties to market manipulator trades; ramping manipulation tactics are better executed in pure limit order books. Consistent with empirical analysis of legal institutions by Cumming and Johan (2008) and Eleswarapu and Venkataraman (2006), we find regulations specifically prohibiting ramping manipulation also reduce the incidence of ramping alerts. Finally, the probability of adopting real time surveillance (RTS) systems is found to be positively related to ramping alert incidence. In the absence of any panel data on the dynamic effects of adopting RTS, what we are observing in cross section is the perceived vulnerability of certain exchanges to manipulation and their consequent adoption of RTS plus the regulatory regimes required to assure better market integrity.

33

As to the probability of adopting real time surveillance, we find Prob(RTS) across markets rises with increased ramping alert incidence, increased trading volume, and increased foreign direct investment (FDI), as expected. Mid-liquidity deciles remain the focus of manipulator activity but across markets, as trading volume increases, exchanges and exchange regulators are more likely to implement RTS to keep track of the complexity introduced by layering, slicing and dicing, and order cancellations in thickly-traded securities. Increased FDI also causes stock exchanges to adopt RTS in order to match the more sophisticated trading strategies of global private equity and hedge fund investors they encounter. Predictably, the adoption of RTS is also more likely when DMA lines make the speed of trading difficult for human surveillance analysts to monitor. Conversely, RTS is less likely to be adopted when volatility rises. Price volatility in especially noisy markets like Thailand raises the incidence of false positives, so less sensitive monitoring tools than RTS prove attractive. Finally, RTS is also less likely to be adopted when surveillance and enforcement activity and resources are dispersed across a larger number of market integrity regulations (e.g., against insider trading, fraudulent disclosure, broker-client conflicts, etc.) Our findings have important policy implications for many securities exchanges in terms of market design and market surveillance. First, the exhibited relationship between alert incidence and effective spreads indicates trade-based manipulation has a significant impact on execution costs. Therefore, the prevention of securities market manipulation not only serves the indirect purpose of improving an exchange’s reputation for market integrity but also contributes directly to achieving a more efficient marketplace. Second, our results indicate that some market design changes can enhance the regulatory efforts to prevent securities market manipulations. For example, to prevent manipulators from marking the closing price, some exchanges could choose to adopt a closing auction or a random closing time, which would make manipulation more costly. Nevertheless, no securities exchange can be designed perfectly. Consequently, exchange and broker-level surveillance backed by effective regulatory enforcement is a necessary and pivotal complement to good design choices.

REFERENCES Aggarwal, R., Wu, G. (2006). "Stock Market Manipulation." Journal of Business 79(4): 1915-1953. Aitken, M.J. (2009). “Dimensions of Market Quality.” Working paper, University of New South Wales.

34

Aitken, M.J., Almeida, N., Harris, F.H.deB., McInish, T.H. (2007). "Liquidity Supply in Electronic Markets." Journal of Financial Markets 10: 144-168. Aitken, M.J., Cook, R., Harris, F.H.deB., McInish, T.H. (2008). "Market Design and Execution Cost for Matched Securities Worldwide." A Guide to Liquidity 2(Winter): 1-39. Allen, F., Gale, D. (1992). "Stock-Price Manipulation.” The Review of Financial Studies 5: 503-529. Allen, F., Gorton, G. (1992). "Stock price manipulation, market microstructure and asymmetric information " European Economic Review 36: 624-630. Allen, F., Titov, L., Mei, J.P. (2006). "Large Investors, Price Manipulation, and Limits to Arbitrage: An Anatomy of Market Corners." Review of Finance 10: 645-693. Bernhardt, D., Davies, R.J. (2005). "Painting the tape: Aggregate evidence." Economic Letters 89: 306-311. Bhattacharya, U., Daouk, H. (2002). "The World Price of Insider Trading." The Journal of Finance 7(1): 75108. Bhattacharya, U., Spiegel, M. (1991). "Insiders, Outsiders and Market Breakdowns." The Review of Financial Studies 4(2): 255-282. Bommel, V. (2003). "Rumors." Journal of Finance 58: 1499-1521. Carhart, M. M., Kaniel, R., Kusto, D.K., Reed, A.V. (2002). "Leaning for the Tape: Evidence of Gaming Behavior in Equity Mutual Funds." Journal of Finance 57(2): 661-693. Chakraborty, A., Yilmaz, B. (2004). "Manipulation in a Market Order Market." Journal of Financial Markets 7, 187-206. Chakraborty, A., Yilmaz, B. (2008). "Microstructure Bluffing with Nested Information." American Economic Review 98(2): 280-284. Chakravarty, S., Harris, F.H.deB., and Wood, R.A. (2009). "The Changing State of the NYSE Market: New Common Factors, Players, and Impulse Responses." Journal of Trading 4 (1): 68-94. Chamberlain, T. W., Chueng, C.S., Kuan, C.C.Y. (1989). "Expiration-Day Effects on Index Futures and Options: Some Canadian Evidence." Financial Analysts Journal 45: 67-71. Chiou, J. S., Wu, P.S., Wang, A.W., Huang, B.Y. (2007). "The Asymmetric Information and Price Manipulation in Stock Markets." Applied Economics 39: 883-891. Comerton-Forde, C., Rydge, J. (2006). "The Current State of Asia-Pacific Stock Exchanges: A Critical Review of Market Design." Pacific-Basin Finance Journal 14: 1-32. Comerton-Forde, C. and Talis Putnins (2008). “Measuring Closing Price Manipulation,” Working paper, University of Sydney. Cumming, D., Johan, S. (2008). "Global Market Surveillance." American Law and Economics Review, forthcoming.

35

Eleswarapu, V., Venkataraman, K. (2006), “The Impact of Political and Legal Institutions on Equity Trading Costs.” Review of Financial Studies, 19: 1081- 1111. Felixson, K., Pelli, A. (1998). "Day End Returns - Stock Price Manipulation." Journal of Multinational Financial Management 9: 95-127. Foucault, T. (1999). "Order flow composition and trading costs in a dynamic limit order market." Journal of Financial Markets 2: 99-134. Harris, F. H. deB. (2008). "Execution costs and market design worldwide: A panel discussion." Journal of Trading 3(1): 9-25. Harris, L. (1989). "A Day-End Transaction Price Anomaly." Journal of Financial Quantitaive Analysis 24(1): 29-45. Harris, L., Sofianos G., Shapiro, J.E. (1994). "Program Trading and Intra-day Volatility." The Review of Financial Studies 7(4): 653-686. Hart, O. D. (1977). "On the Profitability of Speculation." Quarterly Journal of Economics 91(4): 579-597. Hart, O. D., Kreps, D.M. (1986). "Price Destablizing Speculation." The Journal of Political Economy 94(6): 927-952. Hillion, P., Souminen, M. (2004). "The Manipulation of Closing Prices." The Journal of Financial Markets 7: 351-375. Lang, M. H., Lundholm, R.J. (1996). "Corporate Disclosure Policy and Analyst Behavior." The Accounting Review 71(4): 467-493. Merrick, J. J., Naik, N.Y., Yadav, R.K. (2005). "Strategic Trading Behavior and Price Distortion in a Manipulated Market: Anatomy of a Squeeze." Journal of Financial Economics 77: 171-218. Oved, Y. (2002). "Information Disclosure Costs and the Choice of Financing Source." Journal of Finance 4(1): 3-21. Pagano, M. S., Schwartz, R.A. (2003). "A Closing Call's Impact on Market Quality at Euronext Paris." Journal of Financial Economics 68: 439-484. Pagano, Marco, Roell, A. (1996). "Transparency and Liquidity: A Comparison of Auction and Dealers Markets with Informed Trading." Journal of Finance 51(2): 579-611. Stoll, H. R., Whaley, R.E. (1987). "Program trading and expiration-day effects." Financial Analysts Journal 43(2): 16-28. Stoll, H. R., Whaley, R.E. (1991). "Expiration Day Effects: What has changed?" Financial Analysts Journal: 58-72.

Table 1 List of Securities Exchanges Covered and Mean Alert Incidence (All Stocks) 2000-2005

36

Index Securities Exchange 1 American Stock Exchange 2 Australian Securities Exchange 3 Bombay Stock Exchange 4 Cairo Stock Exchange 5 Colombo Stock Exchange 6 Copenhagen Stock Exchange 7 Deutsche Bourse-Extra 8 Euronext Amsterdam 9 Euronext Brussels 10 Euronext Paris 11 Euronext Portugal 12 Hong Kong Stock Exchange 13 Istanbul Stock Exchange 14 Jakarta Stock Exchange 15 Johannesburg Stock Exchange 16 Korea Stock Exchange 17 Kuala Lumpur Stock Exchange 18 London Stock Exchange 19 Madrid Stock Exchange 20 Milan Stock Exchange 21 NASDAQ 22 National Stock Exchange 23 New York Stock Exchange 24 New Zealand Stock Exchange 25 Oslo Bors 26 Shanghai Stock Exchange 27 Shenzhen Stock Exchange 28 Singapore Stock Exchange 29 Stock Exchange of Thailand 30 Stockholm Borsen 31 Swiss Stock Exchange 32 Taiwan Stock Exchange 33 Tokyo Stock Exchange 34 Toronto Stock Exchange Mean Std. Dev. Obs. (34 exchanges x 6 years)

Home Country U.S. Australia India Egypt Sri Lanka Denmark Germany Netherlands Belgium France Portugal China Turkey Indonesia South Africa Korea Malaysia U.K. Spain Italy U.S. India U.S. New Zealand Norway China China Singapore Thailand Sweden Switzerland Taiwan Japan Canada

Alert Incidence 0.19 0.25 0.44 0.14 0.16 0.17 0.10 0.32 0.35 0.30 0.37 0.07 0.03 0.21 0.07 0.02 0.40 0.32 0.15 0.23 0.08 0.37 0.26 0.28 0.17 0.03 0.06 0.36 0.13 0.23 0.15 0.45 0.16 0.22 0.21 0.18 204

Table 2 Descriptive Statistics Quoted and Effective Spreads, 34 Markets, 2000-2005 Panel A: Moments for Raw Spreads

37

Quoted Spread

Effective Spread

Mean

6.99%

6.64%

Std. Dev

0.4994

0.1383

Skewness

1.2689

5.1899

Kurtosis

1.7947

28.7423

204

204

No. of Observations

Panel B: Moments for Natural Log of Spreads Quoted Spread Effective Spread Mean

-3.1555

-3.4645

Std. Dev

1.1513

1.2172

Skewness

-0.6831

-0.1480

Kurtosis

-0.3900

0.9826

204

204

No. of Observations

38

Table 3, Panel A Mean Spreads by Deciles Quoted Spread

Effective Spread

Decile 1 (Most Liquid)

0.73%

2.19%

Decile 2

1.28%

2.88%

Decile 3

1.85%

3.26%

Decile 4

2.56%

7.90%

Decile 5

3.72%

5.42%

Decile 6

5.36%

5.23%

Decile 7

7.52%

7.21%

Decile 8

10.89%

7.70%

Decile 9

15.02%

10.91%

Decile 10 (Least Liquid)

20.98%

13.56%

Grand Mean

6.99%

6.64%

Panel B Mean Alert Incidence by Decile Decile 1 (Most Liquid)

0.31

Decile 2

0.27

Decile 3

0.22

Decile 4

0.23

Decile 5

0.23

Decile 6

0.22

Decile 7

0.19

Decile 8

0.16

Decile 9

0.12

Decile 10 (Least Liquid)

0.13

39

Table 4 Error Components Model of Efficiency-Integrity Relationship for All Deciles Panel A Effective Spreads All Deciles

Full Random Effects (N: 204) MA Error Components ( N:204, MA Order=3) Full Fixed Effects ( N:204)

Mean Spread: 6.64%

Intercept: -3.23*** (t= -15.62)

Mean Spread: 6.64%

Intercept: -3.31*** (t=-16.47)

Mean Spread: 6.64% Exchange Dum: 30 out of 33

Intercept: -4.75*** (t=-17.95) Year Dum: 4 out of 4 (F=3.47***)

AI Parameter: 0.1126*** (t=3.26) R-square: 0.0500 AI Parameter: 0.0718***(t= 4.00) R-square: 0.0246 AI Parameter: 0.1081*** (t=3.02) R-square: 0.8117

Hausman M Test: 0.23 (>0.64) F Test: 10.63*** F Test: 5.09*** F Test: 18.51***

Panel B Quoted Spreads All Deciles

Full Random Effects (N: 204) MA Error Components ( N:204, MA Order=4) Full Fixed Effects ( N:204)

Mean Spread: 6.99%

Intercept: -3.07*** (t= -15.23)

Mean Spread: 6.99%

Intercept: -3.10*** (t= -15.43)

Mean Spread: 6.99% Exchange Dum: 31 out of 33

Intercept: -4.50*** (t= -32.44) Year Dum: 4 out of 4 (F=5.5***)

AI Parameter: 0.0432** (t=2.32) R-square: 0.0260 AI Parameter: 0.0264 R-square: 0.0111 AI Parameter: .0392** (t= 2.09) R-square: 0.9426

Hausman M Test: 4.08 (>0.04) F Test: 5.39*** F Test: 2.26 F Test: 71.30***

40

Table 5 Descriptive Statistics for 24 Markets in year 2005 Panel A: Moments for Raw Observations AI SPR Vol

Liquidity

Regs

Mean

0.1775

5.36%

0.0409

97120486

6.7500

Std. Dev

0.1876

0.0956

0.0446

365031297

3.9053

Skewness

1.7931

3.4166

3.9966

7.3444

0.6062

Kurtosis

3.2889

13.1504

17.1601

63.4832

-0.2879

240

240

240

240

240

No. of Observations

Panel B: Moments for Natural Log of Observations AI SPR Vol Liquidity

Regs

Mean

-2.5002

-3.9911

-3.4882

15.1705

1.3483

Std. Dev

0.8547

1.4776

0.7877

3.3841

2.2723

Skewness

0.1977

0.2425

-1.5412

-0.7449

-4.1936

Kurtosis

3.4063

-0.6729

3.3485

0.2987

16.9851

240

240

240

240

240

No. of Observations

41

Table 6 OLS and 2SLS AI Equations and IV equations for potentially endogenous variables Vol and QSpr for moderately-traded deciles in 2005 Panel A: OLS With Fixed Effects

AI Model

Volatility Model

Spreads Model

F-test

6.86***

10.09***

161.97***

R-square

0.7102

0.7023

0.9830

βspr

-0.3522(t= -1.21)

0.1995*** (t=3.55)

N/A

βcall

-1.3882***(t=2.66)

2.1948***(t=9.23)

-3.9521***(t=-17.99)

βDMA

-0.8916 (t=-0.76)

N/A

N/A

βRTS

1.1498***(t=3.18)

-2.6347*** (t=-10.07)

10.5837***(t=20.16)

βRampReg

0.6178(t=1.45)

-2.7780** (t=-8.16)

22.1870***(t=19.87)

βvol

0.0803(t= 0.82)

N/A

0.4125***(t=4.68)

βregs

N/A

0.0473**(t=2.13)

-1.4952***(t=-17.38)

βliq

N/A

0.2327(t=4.74)

-0.3485***(t=-15.03)

βexch

5 out of 23

13 out of 23

20 out of 23

α

-1.4953( t=-1.55)

-2.7532***(t =-12.51)

5.5170***(t=13.43)

Panel B: 2SLS With Fixed Effects

Instrument: Volatility

Instrument: Spread

Hausman-Wu Stats

2.92

0.03

AI Model

Volatility Model

AI Model

Spread Model

F-test

5.16***

4.61***

5.12***

16.07***

R-square

0.4037

0.5352

0.4015

0.8446

βspr

-0.1594(t=-1.33)

0.4970**(t=2.51)

0.2482( t=1.19)

N/A

βcall

-1.3477***(t=2.67)

1.9304***(t=5.36)

-1.3473**(t=-2.61)

-3.7108***(t=-5.35)

βDMA

-0.4566(t=-1.20)

N/A

-0.3400(t=-0.78)

N/A

βRTS

1.1392***(t=3.13)

-2.3076***(t=-5.90)

0.9515**(t=2.53)

11.8457***(t=7.17)

βRampReg

0.5642*(t=1.75)

-2.3800***(t=-4.69)

0.5707*(t=1.69)

26.4773***( t=8.00)

42

βvol

0.0979(t=0.38)

N/A

0.1409(t=0.27)

0.9962***(t=4.09)

βregs

N/A

0.2137(t=0.67)

N/A

-1.9181***(t=-7.88)

βliq

N/A

0.1671*** (t=3.47)

N/A

-0.2325***(t=-4.27)

βexch

13 out of 23

8 out of 23

4 out of 23

20 out of 23

α

-1.3900( t=-1.32)

-3.9844***(t =-8.82)

1.6026(t=-0.71)

2.3468***(t=2.77)

43

Table 7 OLS and 2SLS AI Equations and IV equations for potentially endogenous variables Vol and QSpr in thickly liquid deciles of 24 Securities Exchanges in 2005 Panel A: OLS With Fixed Effects

AI Model

Volatility Model

Spread Model

F-test

16.69***

26.39***

55.97***

R-square

0.7079

0.9309

0.9682

βspr

0.2550*(t= 1.99)

-0.1596*** (t=-2.97)

N/A

βcall

0.3995(t=1.06)

1.7444***(t=9.76)

-2.3986***(t=-7.87)

βDMA

-0.7370** (t=-2.00)

N/A

N/A

βRTS

-0.1391(t=-0.38)

-2.0008*** (t=-7.97)

7.8736***(t=9.10)

βRampReg

0.0043(t= 0.01)

-1.2003*** (t=-3.50)

18.5677***(t=8.92)

βvol

-0.3466(t= -1.39)

N/A

-0.1900(t=-1.01)

βregs

N/A

-0.1138***(t=-3.79)

-1.4223***(t=-8.51)

βliq

N/A

0.0194(t=0.46)

βexch

3 out of 23

-0.3184***(t=-11.25) 18 out of 23

With Fixed α Effects

-1.7210*** ( t= -2.89) -3.8249***(t =-17.38) Instrument: Volatility

12 out 23 Panel B:of2SLS

Hausman-Wu Stats

3.2526***(t=4.36) Instrument: Spread

11.53***

0.40

AI Model

Volatility Model

AI Model

Spread Model

F-test

8.53***

0.61

15.58***

29.30***

R-square

0.5531

0.2370

0.6934

0.9374

βspr/sprhat

0.2940(t=1.61)

-0.0752(t=-0.11)

0.1488( t=0.68)

N/A

βcall

0.9707(t=1.12)

0.7978(t=0.91)

0.4894(t=1.20)

-2.4694***(t=-4.80)

βDMA

-2.2920(t=-1.15)

N/A

-0.9907**(t=-2.08)

N/A

βRTS

1.9729(t=0.72)

2.5008 (t=0.46)

0.2303(t=0.60)

7.6212***(t=10.50)

βRampReg

0.6088(t=0.68)

2.6388(t=0.78)

0.0541(t=0.14)

17.9909***( t=5.27)

44

βvolhat/vol

-1.3295(t=-1.10)

N/A

-0.5944(t=-1.27)

-0.4863(t=-0.22)

βregs

N/A

-0.3716**(t=-1.96)

N/A

-1.3883***(t=-3.58)

βliq

N/A

0.0934*** (t=2.64)

βexch

3 out of 23

2 out of 23

N/A 3 out of 23

-0.2845(t=-1.49) 17 out of 23

α

-5.3393( t= -1.24) -3.3627(t =-1.34)

-3.2541**(t=-1.97)

1.6922(t=0.18)

45

Table 8 OLS and 2SLS AI Equations and IV equations for potentially endogenous variables Vol and QSPR in thinly-traded deciles of 24 Securities Exchanges in 2005 Panel A: OLS With Fixed Effects

AI Model

Volatility Model

Spread Model

F-test

3.48***

19.90***

32.47***

R-square

0.6543

0.9154

0.9464

βspr

-0.0009(t= -0.02)

0.4132***(t=3.02)

N/A

βcall

2.1942(t=1.08)

0.8322(t=1.09)

-3.3923***(t=-5.88)

βDMA

3.6283*(t=1.82)

N/A

N/A

βRTS

-6.6846**(t=-2.07)

-1.5364(t=-0.80)

9.8075***(t=7.84)

βRampReg

-5.7615**(t=-2.09)

-2.5914(t=-0.67)

19.9007***(t=8.02)

βvol

-0.8529(t= -0.63)

N/A

0.3999***(t=5.39)

βregs

N/A

0.0434(t=0.16)

-1.3158***(t=-7.20)

βliq

N/A

0.3760***(t=6.86)

βexch

6 out of 23

5 out of 23

-0.3351***(t=-5.71) 20 out of 23

α

-5.6848**( t=-2.46)

-6.3500***(t =-9.63)

4.7542***(t=5.39)

Panel B: 2SLS With Fixed Effects

Instrument: Volatility

Instrument: Spread

HausmanWu Stats

5.45

0.47

AI Model

Volatility Model

AI Model

Spread Model

F-test

4.76***

12.73***

5.96***

0.02

R-square

0.4087

0.6488

0.4640

0.0015

βspr

-0.5534(t=-1.02)

0.4971**(t=2.49)

0.1943( t=0.62)

N/A

βcall

0.2628(t=0.36)

0.4125(t=1.53)

0.7056(t=1.17)

-4.2293(t=-0.05)

βDMA

0.7475(t=0.87)

N/A

-0.0439(t=-0.07)

N/A

46

βRTS

-1.1742(t=-1.62)

-0.7679***(t=-2.99)

-1.2001*(t=-1.81)

61.7170(t=0.03)

βRampReg

-1.3938*(t=-1.98)

-1.1679***(t=-3.28)

-1.3280**(t=-2.14)

117.3303( t=0.03)

βvol

-0.7525(t=-1.27)

N/A

-0.0096(t=-0.03)

-42.2121***(t=-0.03)

βregs

N/A

-0.0418(t=-1.19)

N/A

-11.1745***(t=-0.03)

βliq

N/A

0.4442*** (t=8.10)

βexch

2 out of 23

3 out of 23

N/A 2 out of 23

9.1848(t=0.03) 0 out of 23

α

7.0739*( t=-1.92)

-7.0548***(t =-17.03) -2.0224(t=-0.94)

-248.988(t=-0.03)

47

Table 9 PROBIT Analysis of Real-time Surveillance System Deployment Panel A: All Deciles Discrete Response Frequency 0 (65%) Without Fixed Effects

1 (35%) With Fixed Effects

N

240

240

Log Likelihood

-97.76

-3.4718E-8

Likelihood Ratio

94.22

289.75

Aldrich-Nelson

0.28

0.55

βai

0. 1327** (t=2.13)

1.3573( t=0.25)

βDMA

1.5636***( t =6.48)

142.39***(t=59.88)

βregs

-0.0988***( t= -3.29)

-3.2514(t=-0.09)

βvol

-0.3184**( t=-1.97)

7.0394 (t=0.70)

βliq

0.0968*** ( t= 2.67)

-2.4852(t=-0.08)

βFDI

1.7875*** (t= 3.46)

α

--3.3926*** ( t= -3.36)

83.6882***(t=87.90) -37.7427***(t=-15.87)

βexch

N/A

16 out of 23

Panel B: Decile 4, 5, 6, 7 Without Fixed Effects

With Fixed Effects

N

96

96

Log Likelihood

-28.19

-6.4724E-8

Likelihood Ratio

59.52

115.9

Aldrich-Nelson

0.38

0.55

βai

1.1149*** (t=3.19)

1118.7557**( t=2.17)

βDMA

2.8118***( t =3.91)

2315.6456***(t=7.68)

βregs

-0.2222***( t=-2.94)

-94.0829(t=-0.62)

βvol

-0.1999( t=-0.45)

207.2831 (t=0.98)

48

βliq

0.3120** ( t= 2.53)

102.5849(t=0.71)

βFDI

0.6523 (t= 0.77)

α

-4.3470* ( t= -1.65)

704.7527***(t=39.28) -378.0217***( t =9.92)

βexch

N/A

11 out of 23

Panel C: Decile 1, 2, 3 Without Fixed Effects

With Fixed Effects

N

72

72

Log Likelihood

-21.50

-3.2837E-8

Likelihood Ratio

43.91

86.92

Aldrich-Nelson

0.38

0.55

βai

0. 4221* (t=1.65)

-382.0626***( t=-7.26)

βDMA

1.5285**( t =2.49)

4224.9291***(t=1283.78)

βregs

-0.1133*( t=-1.72)

-9.5370(t=-0.02)

βvol

-2.0903**( t=-2.52)

-1143.8055*** (t=-105.47)

βliq

0.4128*** ( t= 2.97)

49.9484(t=0.57)

βFDI

2.5194* (t= 1.69)

α

-14.8848*** ( t= -3.20)

3155.9683***(t=1191.19) -10306***(t=-3130.69)

βexch

N/A

3 out of 23

49

Panel D: Decile 8, 9, 10 Without Fixed Effects

With Fixed Effects

N

72

72

Log Likelihood

-31.09

-5.944E-8

Likelihood Ratio

24.73

86.92

Aldrich-Nelson

0.26

0.55

βai

0.0372 (t=0.51)

2.1517( t=0.19)

βDMA

1.5268***( t =3.56)

56.5862***(t=15.74)

βregs

-0.0910*( t=-1.64)

-3.3795(t=-0.13)

βvol

-0.1113( t=-0.47)

-1.1287 (t=-0.08)

βliq

0.0606( t= 0.78)

-1.6398(t=-0.04)

βFDI

1.8022* (t=1.96)

α

-2.2170 ( t=-1.31)

37.6652***(t=141.53) -9.3265***( t=-2.59)

βexch

N/A

16 out of 23

50

Table 10 Simultaneous Equation System Estimates Panel A: Prob(RTS) Equation βAI

βDMA

βregs

βvol

βliq

βFDI

βexch Α

0.9138*** (t=6.27)

1.5227*** (t=43.79)

-0.0600 (t=-0.27)

0.0969 (t=0.42)

0.1094*** (t=8.30)

1.0354*** (t=39.19)

7/23 -0.7022*** (t=-20.19)

N: 72 0.6544*** LLH: -89.3 (t=55.64)

1.5227*** (t=406.53)

-0.0718*** -0.1033*** 0.1109*** (t=-5.46) (t=-7.48) (t=3.18)

0.6186*** (t=2463.8)

1/23 -1.7417*** (t=-11.44)

N: 72 0.5512*** LLH: -175. (t=8.77)

2.3054*** (t=12.15)

-0.0883*** -0.1247** (t=-8.07) (t=-1.94)

0.9240*** (t=184.19)

6/23 -0.3590*** (t=8.31)

Deciles Group 4,5,6,7 ML

1, 2, 3 ML 8, 9,10 ML

N: 96 LLH: -209

0.0058 (t=0.26)

Panel B: Integrity (AI) Equation Deciles Group

βspr

4,5,6,7 N: 96 LLH: -209.3 ML 1, 2, 3 ML

N: 72 LLH: -89.33

8, 9,10 N: 72 LLH: -175.8 ML

βcall

βRampReg

βvol

-0.0269 -0.2287*** -1.3116*** 1.6521*** (t=-0.15) (t=-3.60) (t=-3.66) (t=4.73)

0.0147 (t=0.04)

-0.4314 2/23 -3.3295*** (t=-1.42) (t=-3.44)

0.2297* (t=1.74)

0.3458 (t=0.98)

-0.1803 0.3619 (t=-1.11) (t=0.64)

βDMA

βProbRTS

βexch α

-0.8107** (t=-2.35)

0.1364 (t=0.39)

-0.0005 -0.3078 3/23 -1.8397* (t=-0.01) (t=-1.34) (t=-1.91)

0.0074 (t=0.01)

-0.6112 (t=-0.93)

-1.319** -0.0826 2/23 -3.3267*** (t=-2.18) (t=-0.43) (t=-3.49)

51

Panel C: Efficiency (Quoted Spread) Equation Deciles Group

βcall

βProbRTS

βRampReg

βvol

βregs

βexch

α

2.6807*** (t=2.63)

7.7208*** (t=2.77)

0.1463 (t=0.52)

-0.5519** (t=-2.24)

20/23

2.4904*** (t=2.80)

18.5797*** -0.2018 -1.9436*** 13/23 (t=11.11) (t=-1.34) (t=-6.31)

3.2416*** (t=5.41)

4,5,6,7 ML

N: 96 LLH: -209.3

-0.9427* (t=-1.61)

1, 2, 3 ML

N: 72 LLH: -89.3

-2.4053*** 7.9109*** (t=-9.83) (t=11.36)

8, 9, 10 ML

N: 72 LLH: -175.8

-5.1504*** 11.5069*** 21.8500*** -0.0660 -1.4930*** 20/23 (t=-9.97) (t=8.83) (t=8.31) (t=-0.61) (t=-7.76)

0.6328 (t=1.19)

52

Figure 1, Panel A A Marking the Close Alert

Figure 1, Panel B A Ramping Alert

5.

53

Figure 2 Histograms for Spread Measure before and after log transform

54