The impact of latency sensitive trading on high frequency arbitrage opportunities

PACFIN-00874; No of Pages 12 Pacific-Basin Finance Journal xxx (2016) xxx–xxx

Contents lists available at ScienceDirect

Pacific-Basin Finance Journal journal homepage: www.elsevier.com/locate/pacfin

The impact of latency sensitive trading on high frequency arbitrage opportunities Alex Frino a,⁎, Vito Mollica b, Robert I. Webb c, Shunquan Zhang b a b c

University of Wollongong, Australia Macquarie University, Australia University of Virginia, United States of America

a r t i c l e

i n f o

Article history: Received 16 February 2016 Received in revised form 25 July 2016 Accepted 16 August 2016 Available online xxxx JEL classification: G140 G180 D400 Keywords: High frequency trading Statistical arbitrage Co-location ETF Futures

a b s t r a c t This study examines the duration, frequency and profitability of potential high frequency arbitrage strategies between the share price index futures contract and an exchange-traded fund (ETF) written on the S&P/ASX200 constituent securities traded on the Australian Securities Exchange (ASX). We find the frequency and profitability of potential arbitrage opportunities are greater during volatile and high turnover periods—other things equal. We examine the impact of increased competition in high frequency trading (HFT) by identifying the number of ‘co-location connections’ utilized in the ASX's minimum latency liquidity center. We document an increase in the frequency, duration and value (albeit small) of index arbitrage profit opportunities with increased HFT connections. Our results are robust to the inclusion of transaction costs. We conclude that increased HFT activity in markets increases trade execution risk associated with arbitrage (or legging risk) which in turn increases mispricing in markets. © 2016 Elsevier B.V. All rights reserved.

1. Introduction Classical economic theory suggests that excess returns should be competed away as participants enter the market. This is especially true for the profits from riskless arbitrage. Yet, there is conflicting evidence in the financial economic literature over whether high frequency trading (HFT) profits, in general, (Baron et al., 2012; Chordia et al., 2015) and arbitrage profits, in particular (Budish et al., 2015; Chaboud et al., 2014), decline as high frequency or other algorithmic trading increases. There are important public policy implications for market microstructure and the social value of investments by HFT firms in being faster if arbitrage profit opportunities persist (in the absence of limits to arbitrage). Several trading strategies are used by high frequency and latency sensitive traders (O'Hara, 2015). These include: index arbitrage; spread arbitrage/market making; and correlated arbitrage among others. Irrespective of the strategy, traders suffer execution risk if orders in markets are fleeting or stale. Legging risk for arbitragers (Sofianos, 1993) may decline in markets as latency is improved with the adoption and roll-out of enhanced electronic trading platforms, but it may also increase as the number of

⁎ Corresponding author. E-mail address: [email protected] (A. Frino).

http://dx.doi.org/10.1016/j.pacfin.2016.08.004 0927-538X/© 2016 Elsevier B.V. All rights reserved.

Please cite this article as: Frino, A., et al., The impact of latency sensitive trading on high frequency arbitrage opportunities, PacificBasin Finance Journal (2016), http://dx.doi.org/10.1016/j.pacfin.2016.08.004

2

A. Frino et al. / Pacific-Basin Finance Journal xxx (2016) xxx–xxx

market makers and arbitragers compete to try and exploit the same opportunities, and in the process impose negative externalities on markets and impede pricing efficiency (Kozhan and Tham, 2012). In this paper we focus on one HFT strategy—index arbitrage, and examine whether the duration, frequency and profitability of potential arbitrage opportunities between the Australian Securities Exchange (ASX) Share Price Index (SPI) futures contract and the exchange traded fund (ETF), SPDR S&P/ASX 200 Fund (STW), has changed as the number of HFT firms (or intensity of HFT activity) has increased, since the ASX introduced co-location services in February 2012. In addition, we compare the estimated potential arbitrage profits to the cost of being co-located, in order to determine the value of minimum latency. We have two principal reasons for examining ASX data. First, we are able to examine a finer time interval than past studies, with data time-stamped at the microsecond level. Second, we have information on the growth of HFT over our sample period. More specifically, we know the number of co-located cabinets reported by the ASX in its minimum latency co-location center and use this measure as a proxy for HFT competition. Our principal findings are as follows. First, consistent with Budish et al. (2015), the frequency of potential arbitrage opportunities is greater during volatile periods, other things equal. Second, increased HFT in markets changes the speed of convergence between paired instruments, suggesting greater competition among HFTs in their demand and provision of liquidity. Third, contrary to Budish et al. (2015), the average daily profit, frequency and duration of arbitrage opportunity increases, as HFT connections increase in the market. HFTs appear to compete aggressively against each other, resulting in larger and more frequent price discrepancies albeit with economically trivial arbitrage profit opportunities. Our findings are consistent with Kozhan and Tham (2012) who demonstrate that HFTs suffer execution uncertainty from negative externalities inflicted by trading against each other. The remainder of this article is organized as follows. Section 2 reviews the related literature. Section 3 outlines the data, arbitrage strategy, and method utilized to examine the impact of HFT on arbitrage opportunities. Section 4 reports results and robustness tests. Section 5 concludes.

2. Review of the literature Recent surveys highlight the significant debate surrounding the effects of HFTs on markets (see Biais and Woolley (2011), Chordia et al. (2013), Jones (2013), and Goldstein et al. (2014)). O'Hara (2015) further states in light of the fundamentally different way markets are structured today, it is important to understand the strategic behavior the microsecond trading environment affords traders, especially in terms of arising price differentials. Madhavan (2012) for example finds market structure, in particular the increase in market fragmentation and venue competition, is a catalyst for extreme price movements. Equally, Menkveld (2013) finds that the likely success of a new trading venue hinges on the participation of HFTs. Most of the research related to HFT focuses on its impact on market quality. Examining the NYSE and NASDAQ markets, Hendershott et al. (2011a,b) and Hasbrouck and Saar (2013) report that bid-ask spreads and volatility improved during times of increased HFT, while Jarnecic and Snape (2014) document that on the London Stock Exchange (LSE) HFT is associated with shorter order duration and thinner depths that increase the transience of prices. Carrion (2013) and Brogaard et al. (2014a) extend these studies by investigating the relation between HFT and information efficiency/price discovery and document that HFT is associated with improved impounding of information into markets. However, Jain and McInish (2012) find that HFT increases tailrisk in Japan, while Boehmer et al. (2014) in their global study report that HFT increases short-term volatility, leading to further negative externalities in the market, as modelled by Biais et al. (2012). Brogaard et al. (2014b) find following infrastructure upgrades on the LSE, the associated increase in HFT activity does not affect institutional trader costs, while Van Kervel and Menkveld (2016) document that institutional transaction costs increase (decrease) when HFTs trade in the same (opposite) direction as institutional investors who execute a package of trades through order-splitting strategies on the Nasdaq OMX Sweden. Conversely, Toth et al. (2015) find order splitting does not appear to change with the rise of algorithmic trading on the LSE. Extending beyond the issue of the introduction of HFT, Breckenfelder (2013) examines the effect of competition among HFT firms on market quality. Able to identify international entrants into the Swedish market, Breckenfelder (2013) finds HFT competition is associated with an increase in demand for liquidity, volatility and momentum trading. Similarly, examining 13 HFT firms on Canada's Alpha exchange, Brogaard and Garriott (2015), find that HFT firm entrants generally improve liquidity and price efficiency, however their marginal affect differs across stocks and their presence can be disruptive at first. Brogaard and Garriott (2015) also find asymmetric effects in terms of the profits earned by incumbent HFT firms as other HFTs leave and enter the market, leading them to conclude that not all HFT drives competitive pressure on profits. Baron et al. (2012) examine the profitability of HFTs in the (all electronically traded) e-mini S&P 500 stock index futures market during the entire month of August 2010. Using a comprehensive data set identifying the trades of 31 HFT firms, they report all 31 HFT firms were profitable during the month. The HFT firms collectively earned $29 million during the month all while assuming very little risk—the average Sharpe ratio was 9.2. Baron et al. (2012) identify the most profitable HFT firms were not liquidity providers but rather the most aggressive liquidity takers, competing on speed, sophistication and technological innovation. The focus on the continuing investments in speed by HFT firms as a potential explanation for why their profits remain relatively stable extends to Budish et al. (2015). Essentially, Budish et al. (2015) argue that over very short periods of time, correlation between related securities breaks down and creates purely technical arbitrage opportunities, available to whomever is fastest, Please cite this article as: Frino, A., et al., The impact of latency sensitive trading on high frequency arbitrage opportunities, PacificBasin Finance Journal (2016), http://dx.doi.org/10.1016/j.pacfin.2016.08.004

A. Frino et al. / Pacific-Basin Finance Journal xxx (2016) xxx–xxx

3

creating (a socially wasteful) arms race to exploit these arbitrage opportunities. Budish et al. (2015) state the annual technical arbitrage profits one could earn across markets is “substantial.”1 Budish et al. (2015) argue that the duration of one such technical arbitrage opportunity (between the S&P 500 e-mini futures contract and an exchange traded fund for the S&P 500 index), has declined but the profitability of such “arbitrage opportunities is remarkably constant” although the frequency varies directly with market volatility. Simply put, competition among HFTs does not bid away the potential arbitrage opportunities as HFTs compete on investments in being faster vis-a-vis other traders rather than providing better prices. Budish et al. (2015) use these findings to advocate replacing continuous trading with a series of high frequency batch auctions spaced as little as 100 ms or as much as 1 s apart, to reduce the sniping costs in the bid offer spread that results from the race to be fastest. Budish et al. (2015) compare the prices of the e-mini S&P 500 stock index futures contract traded on the Chicago Mercantile Exchange (CME) with prices of the SPDR (“Spider”) S&P 500 stock index ETF traded on multiple venues including the New York Stock Exchange (NYSE) under the ticker symbol SPY. Budish et al. (2015) note their use of direct feeds from the CME and the NYSE in conducting their empirical analyses. The question naturally arises as to whether the data are truly synchronized across trading venues especially since Budish et al. (2015) limit their empirical analysis to only SPDR prices reported on the NYSE. Simply stated, if the two time series are not synchronized “observed” arbitrage profit opportunities may be overstated.2 And, the presumption that there is a large amount of HFT index arbitrage activity may be incorrect. There may also be limits to arbitrage that impede HFTs from exploiting apparent arbitrage opportunities as Kozhan and Tham (2012) point out. In contrast, Chaboud et al. (2014) report evidence of a substantial decline in the frequency (and implicitly the amount) of triangular arbitrage profit opportunities in the foreign exchange market. Indeed, Chaboud et al. (2014) report many instances where no triangular arbitrage profit opportunities exist. Chaboud et al. (2014) attribute the decline in arbitrage profit opportunities to the growth of algorithmic trading. They argue that by reducing the time it takes for price discovery, algorithmic trading increases informational efficiency. To be sure, their use of minute-by-minute data may obscure a decline in the duration of potential arbitrage profit opportunities even though the overall amount of arbitrage profit opportunities is unchanged. That is, declines in the duration of arbitrage opportunities may result in an apparent absence of arbitrage profit opportunities at the minute-by-minute level that are apparent at the millisecond level. 3. Data and methodology The data used in this study are sourced from Thomson Reuters Tick History (TRTH), managed and distributed by SIRCA. We sample tick level data, time stamped to the nearest microsecond, containing price and volume data for each best bid and ask update over the period of March 1, 2012 to January 31, 2014. Two closely related instruments are examined: 1) STW, an ETF tracking the ASX/S&P 200 index, and 2) SPI, an index future contract written over the corresponding equity index. STW is managed by State Street Global Advisors and is available for trading on the Australian equity platforms, ASX and Chi-X.3 Chi-X commenced trading ETFs and other equities in October 2011. In the subsequent analysis, we assume HFTs have direct market access to both venues and thus construct a consolidated tape from the two markets to provide the best representation of the national limit order book.4 On any given day, multiple SPI futures contracts are available in the market with different expiry dates. We sample the most actively traded contract on each trading day. SPI trades solely on the ASX's derivative platform, ASX Trade24, over two sessions, a day session (9:50 am–4:30 pm) and an overnight session (5:10 pm–7:00 am). We examine the day session that overlaps with the corresponding ETF trading hours. We also exclude trades during the opening and closing auctions. This results in a daily sample period extending from 10:10 am to 4:00 pm, during which time both instruments are available for continuous trading. 3.1. Measuring high frequency trading To identify the level of HFT competition in the Australian market, we utilize publicly available information stipulating the usage of the ASX's co-location facility. The ASX introduced its co-location facility, the Australian Liquidity Centre (ALC), for equities

1 The estimate by Budish et al. (2015) of billions of dollars of technical arbitrage profit opportunities conflicts sharply with industry estimates of the total profits of high frequency trading firms in the U.S. equity market. Tabb (2014) estimates that profits among equity HFT firms in the U.S.A. have declined sharply from $7.2 billion to $1.3 billion. If this industry estimate is correct then either HFT firms are leaving substantial profits on the table or Budish et al. (2015) are incorrect in their estimate of the size of such profit opportunities. Given that access to real-time information is a necessary condition to conduct low-latency trading, the question naturally arises as to whether exchanges are pricing their co-location services correctly if Budish et al. (2015) are correct that there are billions of dollars of purely technical arbitrage profit opportunities. See Webb (2003) for a discussion of the history of exchange ownership of real-time price and volume data. 2 Miller et al. (1994) examine index arbitrage in the S&P 500 stock index futures market and point out some of the dangers of inferring arbitrage when data are not synchronized due to infrequent trading. 3 Exchange traded options or future contracts do not exist for STW. 4 Australia does not operate under a RegNMS structure as in the US. Traders are required to provide clients with best execution, however this may not be at the NBBO. Our results are robust to the exclusion of Chi-X data—typically 10% of daily traded volume, and avoiding issues associated with the construction of a consolidated tape or synchronization issues as STW and SPI are traded on the same ASX platform.

Please cite this article as: Frino, A., et al., The impact of latency sensitive trading on high frequency arbitrage opportunities, PacificBasin Finance Journal (2016), http://dx.doi.org/10.1016/j.pacfin.2016.08.004

4

A. Frino et al. / Pacific-Basin Finance Journal xxx (2016) xxx–xxx

products (including ETFs) on February 6, 2012, and futures and options products on February 20, 2012.5 The ALC provides ‘cabinets’ for its customers with minimum latency access to the ASX matching engine and order book—a prerequisite for any HFT operation.6 The number of cabinets rented is reported by the ASX in its yearly and half-yearly reports. There were 76 cabinets occupied at the end of June 2012, the first fiscal year of co-location. The number of cabinets increased to 111 by December 2012, 117 by June 2013 and 133 by December 2013. At the commencement of the ALC co-location facility, a total of 20 occupied cabinets were moved from the ASX's previous data center.7 3.2. HFT arbitrage Breakdowns in correlation between related securities and the potential arbitrage profits they can create has been established by the extant literature across a number of security groups. Epps (1979) for example identifies breakdowns across firms in the same industry, Pontiff (1996) provides evidence of costly arbitrage for closed end-funds, Gagnon and Karolyi (2010) examine arbitrage opportunities in the multi-market trade setting, and Marshall et al. (2013) identify mispricing between the SPDR and iShare ETFs. In theory, SPI and STW should track one another perfectly, since they reference the same underlying securities. Reasons for discrepancies in price levels between the two include the cost of carry, dividends and ETF tracking error in the basket of underlying securities.8 Fig. 1 Panel A, shows the price path for the pair of instruments over the sample period. The SPI futures price is consistently above that of the STW ETF, and the daily price pattern depicts a near perfect parallel shift. This parallel pattern breaks down at finer trading intervals. Fig. 1 Panel B, displays the price sequence from 2 pm to 3 pm on a randomly chosen trading day. Prices still appear to track each other closely, however, the unsynchronized movements of the two instruments at finer snapshots highlight the existence of temporary price discrepancies and potential arbitrage opportunities. Given the high correlation depicted in Fig. 1, any transitory price discrepancies between the two instruments should quickly be resolved by market participants updating their orders. Discrepancies that are not immediately adjusted can be traded by other participants in the market. For instance, if the SPI experiences a sudden increase in price and STW price remains the same, one can short SPI and go long STW, and then liquidate the position following the price reversion. Since the correlation converges rather rapidly, the time it takes to eliminate such a discrepancy is reasonably short. Such statistical arbitrage can be automated by HFTs who invest in minimum latency technology to be the first to take the advantage of such mispricing, and is the focus of our study. We are able to identify these potential arbitrage opportunities and measure their frequency, potential profit and duration. The first step to identify an arbitrage is to quantify the natural price difference between the SPI and STW. A number of strategies can be applied (see Gatev et al. (2006), Hogan et al. (2004), Xiong (2001)) in this paper we follow Budish et al. (2015) and start by defining the immediate spread at any time t, as follows, SPI

STW

St ¼ P t −100P t

STW where PSPI are the mid-quote price levels of the two instruments, multiplying by 100 accounts for the quoted price differt , Pt ential between the SPI and STW. A one-point increment in the ASX 200 index is equivalent to a one index point increase in the SPI futures and $0.01 increase in the quoted price of STW. The immediate spread (St) is not a sufficient basis for an arbitrage strategy as its updating frequency, while timely and relevant, is too volatile and noisy to be used as a reliable signal. By accounting for the average spread over a period of time prior to the current best bid and ask update, one can best identify more consistently realizable arbitrage opportunities. The chosen time period should be at a level that is long enough for the two prices to converge (a stable average spread), but also short enough for the information it contains to be relevant to the current update. In this paper, we determine this time parameter, r, as the time it takes for the returns of two instruments to reach a correlation of 0.90.9 More formally, on a particular trading day, i, the time gap ri is determined by solving the equation Corr(ri) = 0.90, accurate to the millisecond level, using data from the preceding 20-trading days.10

5

See Frino et al. (2014) for institutional details pertaining to the ALC. It can be argued that some market participants may strategically rent a large number of co-location cabinets, without truly occupying the space. However ALC has a capacity of 300 co-location cabinets, almost twice the size of current utilized capacity, rendering such a strategy pointless. 7 ASX established a data centre, in Bondi, Sydney, with a maximum capacity of 20 hosts, that was fully occupied prior to the development of the ALC. Occupants of the old data centre were relocated to the new co-location cabinets once the ALC became operational. See http://www.itnews.com.au/News/289358, new-asx-data-centregoes-live.aspx. 8 There are two major differences between the SPI and STW. First, while the relative minimum tick size (minimum tick divided by price) for SPI and STW are similar, the dollar value of the SPI is substantially larger than the dollar value of the STW. A one tick movement in the SPI, one index point, results in a change of $25(AUD) per contract. An equivalent one tick change in STW, represents only a 1 cent change per unit of STW. This means that one contract in the SPI is equivalent to 2500 STW units. Second, the expiration of the SPI contract at the end of each quarter requires traders to roll over to the next contract if they wish to continue holding their position. STW, on the other hand, pays out accumulated dividends and is required to rebalance its portfolio, to re-weight changes in ASX200 index basket at the end of each quarter. 9 Generally, prices can be sampled as beginning, ending and average values over the interval. Throughout the paper, we apply a time weighted average price approach to ensure the sampled prices are more representative of the individual time interval. 10 There are a few trading days in our sample where convergence in returns is slow (e.g., days with STW trading halts). To ensure ri is not influenced by erroneous data errors, we remove trading days, that require more than 2000 s to reach a correlation of 0.99. This excludes less than 5% of the sample period. Secondly, as shown in Fig. 1, the correlation does not act as a continuously increasing function against the time interval of the return. We use a binary search method to find a solution to Corr(ri) = 0.90. 0.90. Binary search takes approximately 21 iterations of function evaluations to find the required accuracy, while conventional methods would require several hundred thousand evaluations. A caveat with this approach is that it does not guarantee the smallest ri. 6

Please cite this article as: Frino, A., et al., The impact of latency sensitive trading on high frequency arbitrage opportunities, PacificBasin Finance Journal (2016), http://dx.doi.org/10.1016/j.pacfin.2016.08.004

A. Frino et al. / Pacific-Basin Finance Journal xxx (2016) xxx–xxx

5

A: March 2012 January 2014 5300

SPI (ind pts)

55.00

SPI STW

53.00

5100

51.00

4900

49.00

4700

47.00

4500

45.00

4300

43.00

4100

41.00

3900

39.00

STW ($)

5500

37.00

3700 Jan-12

May-12

Sep-12

Jan-13

May-13

Sep-13

Jan-14

B: 2:00 PM 3:00 PM, June 28, 2013 4794

45.15 SPI STW

45.13

4790

45.11

4788

45.09

4786

45.07

4784

45.05

4782 2:00 PM

2:10 PM

2:20 PM

2:30 PM

2:40 PM

2:50 PM

STW ($)

SPI (ind pts)

4792

45.03 3:00 PM

Fig. 1. Price path of the ASX/S&P 200 futures contract and ETF. Panel A depicts the price path of SPI and STW over the sample period March 2012–January 2014. Panel B, depicts the price path of SPI and STW over a randomly chosen one-hour snapshot.

The basis of arbitrage opportunities at each price update is determined as the time weighted spread, St , during the previous r interval as, nr n  o 1X St ¼ St−t i  t i −Max t i−1 ; t nr r i¼1 where {nr | t − tnr −1 b r ≤ t − tnr}. {ti − Max(ti −1,tnr)} represents the time gap between each price update, with the exception of the first observation, since r almost never coincides precisely with the time of quote updates; nr indicates the number of previous spread updates to be included. If a sudden change is observed in either market, and reverts to its starting value within r seconds, the average spread St is unchanged. The basis spread, St , is compared to the best bid and ask updates at time t. We define the bid and ask of the arbitrage strategies as SPI STW STW Sbid (short SPI and long STW), and Sask = askPSPI (long SPI and short STW). By design, the relationship t = bidPt −100askPt t t −100bidPt bid ask St b St b St holds when there is no price divergence. If for instance, there is a sudden decline in the value of SPI (or increase in STW) at ask ask time u, large enough thatSbid u b Su b Su, an arbitrage profit opportunity emerges, in the direction of S . The length of an opportunity ceases when the spread moves back to normal or reverts at some future time v. The difference v−u is referred as the duration per arbitrage. Once the spread reverts and the position is closed out, the potential profit starting at time u is Su −Sask u ; measured in index points per contract. The reverse is true when SPI/STW price suddenly rises/declines.11 The profits from arbitrage are as follows, ( π¼

11

ask

ask

Su −Su ; if Su NSu bid bid Su −Su ; if Su b Su

⇒long SPI and short STW ⇒short SPI and long STW

Note Sask and Sbid t t use price updates at time t, while St contains price information prior to time t.

Please cite this article as: Frino, A., et al., The impact of latency sensitive trading on high frequency arbitrage opportunities, PacificBasin Finance Journal (2016), http://dx.doi.org/10.1016/j.pacfin.2016.08.004

6

A. Frino et al. / Pacific-Basin Finance Journal xxx (2016) xxx–xxx

The change in spread St does not always converge back, due to a permanent shock in the SPI/STW price discrepancy. We follow Budish et al. (2015), if the arbitrage duration, v − u, is larger than r such arbitrages are deemed to be ‘bad-arbs’.12 Expected Dollar profit per arbitrage is calculated by multiplying π by the available quoted Volume defined as, ! 8 bid > ask VolSTW > > > 25  Min VolSPI ; < 2; 500 ! Volume ¼ ask > Vol > bid STW > > 25  Min VolSPI ; : 2; 500

ask

if Su NSu

bid

if Su b Su

where 2500 adjusts the STW share to the same value as the SPI contract and 25 reflects the dollar value per index point. Daily profit is summed and the total frequency of arbitrage is summed at the end of each trading day. Duration per arbitrage is summed on a daily basis to derive daily duration. 3.3. Analysis To investigate the relation between arbitrage opportunities (mispricing) and HFT, we control for a number of factors consistent with the extant literature. Specifically, we control for turnover in the SPI. Consistent with Brailsford and Hodgson (1997), Cummings and Frino (2011) and Baron et al. (2012) we also control for volatility in the futures markets, measured as the log of the ratio between the daily high and low mid-quote in the SPI contract.13 Moreover, Scholtus et al. (2014) find algorithmic trading increases around macroeconomic news releases. To control for news days, we include Bloomberg's news impact measure that ranks macro-economic news from 0 to 100. For days with more than one announcement, we sum the total relevance score to form the variable NewsRelevencet. We also include the previous days S&P/ASX 200 index return to capture the effect of market momentum on potential arbitrage profit opportunities. Specifically, we estimate the following model: Arbt ¼ Intercept þ β1 Colocationt þ β2 logð1=SPI Dollar Volumet Þ þ β3 VolatilitySPI;t þ β4 NewsRelevencet þ β5 LagReturnt þ εt ð1Þ where Arbt is any of the three daily arbitrage variables defined earlier: daily duration, daily arbitrage profit and arbitrage frequency per day; Colocationi is the number of cabinets hosted during each event window and our proxy for HFT competition.14 Our analysis considers five sample periods where we are able to identify the number of ALC cabinets. The periods considered are as follows: Window 1 includes data for the month of March 2012, Window 2 extends May 2012 to July 2012, Window 3 extends November 2012 to January 2013; Window 4 extends May 2013 to July 2013, and Window 5 extends November 2013 to January 2014. These dates are centered on the publication of the ASX's annual and half-yearly financial reports, that identify the number of cabinets in the ALC. Although exact utilization days of individual ALC cabinets are not in the public domain, each event window (except Window 1) represents one-month before and after the reporting date of ASX financial reports, so on average the number of occupied cabinets during each event window is likely to reflect the reported co-location figures in the annual reports.15 4. Results We begin our analysis by reporting return correlations between the two instruments—SPI and STW. Fig. 2 depicts the median daily correlation using mid-price returns sampled at various time intervals. Panel A presents returns sampled from 1 to 1000 ms; Panel B, 1 to 11 s with 10 millisecond intervals; and Panel C, 11 to 111 s measured over 100 millisecond increments. Over the sample period, March 2012 to January 2014, the median daily correlation between the SPI and STW starts at 0.154 at 1 ms and rises to 0.570 by 1 s, 0.765 by 11 s and 0.936 by 111 s. Examining this behavior for each of the five event windows, we observe faster convergence in returns as the number of colocated cabinets increases. Such behavior is particularly evident over finer time intervals (Panel A). However, this does not appear to hold for Window 4. Compared to Budish et al. (2015), who report near zero median correlation for returns measured over 1 ms in the US, we observe a minimum median correlation of 0.154 in Australia. Beyond these finer intervals, the speed of convergence is much slower in Australia vis-à-vis the US. For example, Budish et al. (2015) report a correlation of approximately 0.90 at 11 s, while in Australia the correlation is only 0.765. Table 1 Panel A, reports descriptive statistics for SPI and STW over the sample period. The SPI futures contract is more active and liquid than the STW ETF. The time weighted bid-ask spread in the SPI is 1.13 index points, 13% larger than the minimum tick. The mean bid-ask spread for STW is 2.24 cents, 124% larger than the minimum tick and nearly twice as large as that observed in the SPI. The mean notional depth at the best bid and ask prices in the SPI is $4.67 million, vis-à-vis a notional value of $1.76 12

The mean ‘bad-arb’ time last approximately 88.49 s (median 72.3 s). We also included the ratio of volatility in the SPI and STW, the variable is insignificant and does not alter results. The standard errors are Newey and West (1987) adjusted. Outliers of the three dependent variables, defined as two standard deviations away from the mean, are removed from the analysis. 15 The analysis is also replicated under the assumption of a linear increase in cabinet utilization for all the trading days in the sample period (Mar 2012–Jan 2014), similar results are observed. 13 14

Please cite this article as: Frino, A., et al., The impact of latency sensitive trading on high frequency arbitrage opportunities, PacificBasin Finance Journal (2016), http://dx.doi.org/10.1016/j.pacfin.2016.08.004

A. Frino et al. / Pacific-Basin Finance Journal xxx (2016) xxx–xxx

7

A: 1 to 1000 milliseconds, per millisecond increment Mar-12 May-13 to Jul-13

1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0

100

200

300

May-12 to Jul-12 Nov-13 to Jan-14

400

500 600 Millsecond

Nov-12 to Jan-13 Jan-12 to Jan-14

700

800

900

1000

10

11

B: 1 to 11 seconds, per 10 millisecond increment 1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 1

2

3

4

5

6 Second

7

8

9

C: 11 to 111 seconds, per 100 millisecond increment 1

0.9

0.8

0.7 Mar 12 May 13 Jul 13

0.6 11

21

31

41

May 12 Jul 12 Nov 13 Jan 14 51

61 Second

71

Nov 12 Jan 13 Jan 12 Jan 14 81

91

101

111

Fig. 2. Median correlation convergence. This figure displays the median daily correlation of the mid-price return of the S&P/ASX 200 index future, SPI, and the index ETF, STW, sampled over various time intervals. Panel A presents return correlations sampled from 1 ms to 1000 ms. Panel B presents return correlations sampled from 1 s to 11 s, in 10 millisecond increases. Panel C presents return correlations sampled from 11 s to 111 s, in 100 millisecond increases.

million in the ETF. The SPI futures average daily dollar turnover is 200 times the size of the STW and is 23.21 times more traded. The number of best bid and offer updates on the STW is typically around 12,840 messages per day, while it is 74,500 for the SPI. Daily volatility of the two instruments is relatively similar as expected, given they track the same underlying index. Table 1 Panel B, reports summary statistics in each of the five event windows, and the results are similar to those in Panel A. We do however note heightened bid-ask spreads and quote updates in Window 4. Employing the methodology introduced in Section 3, we quantify the arbitrage opportunities. Summary statistics are reported in Table 2. Our strategy identifies an average of 124 potential arbitrage profit opportunities on a daily basis, corresponding to a potential profit of $529 per trading day. Mean cumulative arbitrage opportunities over a trading day total 606 s.16 Convergence Time (i.e. r) records the time used in each trading day to calculate the basis spread between the SPI and STW. On average, it takes 75 s to reach

16

This does not suggest that arbitrage opportunities were entirely unexploited for 606 s, as such profit opportunities may have been partially exploited.

Please cite this article as: Frino, A., et al., The impact of latency sensitive trading on high frequency arbitrage opportunities, PacificBasin Finance Journal (2016), http://dx.doi.org/10.1016/j.pacfin.2016.08.004

8

A. Frino et al. / Pacific-Basin Finance Journal xxx (2016) xxx–xxx

Table 1 Descriptive Statistics for SPI and STW Trading. This table reports descriptive statistics for the period March 1, 2012 to January 31, 2014 (Panel A) and five event windows (Panel B). Spreads, measured in index points for SPI and cents for STW, and best level Depth are time weighted. Number of trades (# Trades), number of message updates (message traffic) and Dollar Volume are measured as at the end of each trading day. Volatility is measured as the log difference between the daily high and low mid-quote price. Spread (Ind pts./cents) SPI

ETF

Depth ($1000)

# Trades

SPI

SPI

8243 2192 6738 7893 9175

ETF

Panel A: summary statistics March 2012 to March 2014, 480 trading days included Mean 1.130 2.243 4675 1758 Std. Dev. 0.060 0.431 1227 723 Q1 1.085 1.959 3830 1269 Median 1.119 2.215 4464 1662 Q3 1.164 2.461 5317 2196

Dollar Volume ($Million)

Message traffic (000′s)

Volatility (%)

ETF

SPI

ETF

SPI

ETF

SPI

ETF

335 214 196 288 406

1791 543 1430 1697 2059

8.344 6.858 4.621 6.171 9.625

74.50 18.23 60.76 73.58 84.85

12.84 5.96 8.79 12.00 15.91

0.764 0.339 0.515 0.692 0.930

0.762 0.354 0.507 0.686 0.938

4.840 3.255

83.17 10.13

8.93 4.40

0.701 0.222

0.691 0.220

11.658 11.221

78.27 9.38

19.66 3.83

0.892 0.377

0.905 0.411

8.855 6.852

72.34 10.62

19.26 3.17

0.603 0.239

0.590 0.236

9.226 6.266

76.83 19.45

19.78 5.72

0.965 0.429

0.977 0.442

8.121 4.311

66.21 17.05

16.98 6.11

0.806 0.322

0.81 0.322

Panel B: event window summary Window 1 (March 2012, 20 co-located cabinets, 22 trading days included) Mean 1.130 2.345 5179 1882 9206 239 1906 Std. Dev. 0.033 0.342 732 485 1792 87 414 Window 2 (May–July 2012, 76 co-located cabinets, 65 trading days included) Mean 1.089 2.237 4614 2504 9178 276 1769 Std. Dev. 0.030 0.603 626 856 2371 142 522 Window 3 (November 2012–January 2013, 111 co-located cabinets, 60 trading days included) Mean 1.082 1.992 6327 1585 7539 357 1611 Std. Dev. 0.028 0.312 1110 502 1849 199 438 Window 4 (May–July 2013, 117 co-located cabinets, 65 trading days included) Mean 1.208 2.375 3534 1329 8909 363 2049 Std. Dev. 0.052 0.469 591 383 2592 161 679 Window 5 (November 2013–January 2014, 133 co-located cabinets, 59 trading days included) Mean 1.151 2.310 3868 1967 8266 335 1840 Std. Dev. 0.052 0.394 826 811 2102 177 530

Table 2 Summary Statistics of Arbitrage Measures (March 1, 2012–January 31, 2014). This table reports the mean and distribution of arbitrage measures. # of Arbs/Day counts the number of arbitrage opportunities for each trading day. Qty denotes the size of each arbitrage opportunity, measured in number of SPI contracts traded. Daily Profits measures the actual dollar amount of each arbitrage accumulated over the trading day. Daily Duration accumulates the time that each individual arbitrage opportunity lasts over each trading day. Convergence Time records the time used in each trading day to calculate the basis spreads between SPI and STW. % SPI initiated records the percentage of arbitrage opportunities that are initiated by a change in the price of SPI. % Good Arbs presents the percentage of arbitrage opportunities that revert before the daily convergence time. % Buy vs. Sell reveals the proportion of arbitrages that require a bid strategy (long SPI and short STW). Percentile

# of Arbs/Day Qty (SPI Lots) Daily Profit $ Daily Duration Convergence Time (sec) % SPI Initiated % Good Arbs % Buy vs. Sell

Mean

1

5

25

50

75

95

99

124 1.126 529 606 75.36 81.88% 98.15% 50.64%

34 0.001 52 37 16.02

54 0.004 101 94 20.67

81 0.040 244 255 40.90

114 0.258 430 451 65.04

154 1.000 671 762 106.94

236 5.000 1342 1582 150.97

355 9.950 2312 3110 164.85

a 0.90 correlation, ranging from 16 s (1st percentile) to 165 s (99th percentile). Over 80% of the apparent arbitrage profit opportunities arise from quote price movements in the SPI. This is consistent with the greater trading activity in the SPI as compared to the STW. % Good Arbs shows the percentage of arbitrage opportunities that reverse before the daily convergence time, r. More than 98% of the recorded arbitrages are deemed to be “good”, highlighting the close correlation between the two instruments and stability of the adopted arbitrage strategy. Table 3 summarises arbitrage measures (Profit, Duration and # Arbitrage per day) and control variables, during each of the five event windows and over the entire sample period. The variables do not appear to exhibit a clear trend. The average duration is initially 232.9 s (March 2013) and increases to 489.6 s (May–Jul 12) and then 681.5 s (Nov 12–Jan 13). Subsequently, total arbitrage times fall to 370.4 s by mid-2013 and later increase to 749.9 s at the end of January 2014. Additionally, we observe that the arbitrage opportunities are one six-hundredth of the daily profit figure reported by Budish et al. (2015), reflecting the larger trading size and liquidity of the futures and ETF contracts in the US. Table 4 Panel A, reports regression coefficient estimates. Our key variable of interest, Colocation, measures the impact of HFT competition on arbitrage profit opportunities. Firstly, we find an increase in co-located cabinets is associated with an increase in Please cite this article as: Frino, A., et al., The impact of latency sensitive trading on high frequency arbitrage opportunities, PacificBasin Finance Journal (2016), http://dx.doi.org/10.1016/j.pacfin.2016.08.004

A. Frino et al. / Pacific-Basin Finance Journal xxx (2016) xxx–xxx

9

overall daily profit, duration and frequency of potential arbitrage opportunities between SPI and STW. Table 4 Panel A, suggests an increase of 10 co-location cabinets is associated with approximately $9 more profit, six more arbitrage occurrences and an increase in mispricing time of 31 s per day. The increase in duration of high frequency arbitrage profit opportunities contrasts with the findings of Budish et al. (2015) who report a decline in the duration of arbitrage profit opportunities over time, despite total index arbitrage profits remaining constant. Kozhan and Tham (2012) argue that HFT competition increases execution risk as there may be several HFTs submitting the same trade and adversely affecting pricing efficiency. On the other hand, more HFTs can also result in greater competition for liquidity. With closer monitoring of market conditions, HFTs are able to exploit arbitrage profit opportunities that were unlikely to be realized prior to co-location, due to inconsistency in latency that negatively impacts the probability of execution. Collectively, we observe that the over-competition of HFTs may out-weigh the benefits HFTs bring, resulting in a market with larger, more frequent and prolonged price discrepancies between correlated instruments. Among the control variables included in the estimation, the lag index return is not statistically significant in all three model specifications in Table 4 Panel A. This implies that observed arbitrage opportunities are short-lived and not influenced by general equity market behavior.17 SPI trading volume is positively and significantly related to arbitrage profits and occurrences. At the 1 percent significance level, Daily profit and Duration are both negatively related with log(1/SPI Dollar Volume), a 1 percent increase is associated with an increase of $3 in arbitrage profit and almost one extra arbitrage profit opportunity. Price volatility is positively and significantly related to the size and number of arbitrage opportunities. This is consistent with the extant literature that suggests large trading volume and volatility induce greater future index mispricing, or more arbitrage profit opportunities in the market. For robustness, we also use an alternative HFT proxy, log(Message Traffic) and re-estimate Eq. (1).18 Since the variable Message Traffic is a continuous measure and can be observed every day, we utilize the entire sample period (March 1, 2012–January 31, 2014). Results are reported in Table 4 Panel B. The results based on Message Traffic are consistent with aforementioned results. Coefficients for both daily profit and duration are significantly positive, implying that when the two markets are inundated with message traffic, more price discrepancies are observed, resulting in more arbitrage opportunities (in terms of both value and frequency). The Message Traffic variable however is not significant in the estimation of duration of arbitrage opportunities.

4.1. Robustness test: incorporating transaction cost Other than the cost of crossing the bid-ask spread, our analysis so far has not considered exchange fees or the cost of colocation. Such costs can be categorized as fixed or variable. Fixed costs include the cost of setting up servers (e.g. computer hardware and rent), and market data feeds which are not dependent on trading activities.19 For any existing brokers, fixed costs are sunk costs and therefore irrelevant. Variable costs, on the other hand, are crucial to the profitability of the strategy adopted in this paper, since they are directly related to participant's level of trading activity. The ASX charges two types of variable costs, trade and clearing fees. For ETF instruments, trading and clearing fees are 0.15 (subject to total monthly maximum of $75) and 0.90 basis points for each trade value. Futures trading incurs a $1.00 trading fee and $0.90 clearing fee per contract side traded. We incorporate these trading fee structures to measure and identify arbitrage opportunities, after variable costs. Descriptive statistics for arbitrage opportunities net of transaction costs are reported in Table 5. Not surprisingly, the daily frequency of arbitrage opportunities, profit and duration are substantially less than corresponding figures documented earlier. A large proportion of the arbitrage opportunities generate very small profits that are not able to cover the cost of trading. For example, the number of arbitrage opportunities decreases to an average of 24.7 per day, corresponding to a potential daily profit of almost $157. Table 6 reports coefficient estimates of Eq. (1) after incorporating variable transaction costs. The results are consistent with our earlier findings. That is, the number of occupied co-location cabinets remains positively correlated with the three arbitrage variables, suggesting that even after accounting for transaction costs, an increase in HFT activity is associated with an increase in arbitrage opportunities, profitability and duration.

5. Summary and conclusions Does competition among high frequency traders reduce the amount of profits available from exploiting short-lived potential arbitrage opportunities? Limits to arbitrage suggest that all apparent profits are unlikely to be entirely eliminated. This study attempts to answer this question by examining the duration, frequency and amount of arbitrage profit opportunities between the ASX SPI futures contract and its ETF counterpart, STW.

17

For robustness, we also repeat the analysis with the current day index return, results are similar. The presence of both Message Traffic and 1/SPI Dollar Volume forms the order to trade ratio, a widely quoted proxy for algorithmic trading (see Hendershott et al., 2011a,b). 19 It costs $2500/month for a participant to co-locate in the ALC. To execute our strategy, participants also need ASX 24 ITCH order book information ($6000/month) and ASX OUCH to execute trades ($6250/month). 18

Please cite this article as: Frino, A., et al., The impact of latency sensitive trading on high frequency arbitrage opportunities, PacificBasin Finance Journal (2016), http://dx.doi.org/10.1016/j.pacfin.2016.08.004

10

A. Frino et al. / Pacific-Basin Finance Journal xxx (2016) xxx–xxx

Table 3 Summary Statistics of Arbitrage Measures and Factor Variables Across Event Windows. This table shows the mean and standard deviation (in parenthesis) of the arbitrage and control variables used to estimate Eq. (1). Panel A summarises the arbitrage variables from executing the arbitrage strategy between the SPI and STW. Panel B summarises control variables. Colo. Cabinets records the number of cabinets during each event window. Log(Message Traffic) is the log of daily order book message traffic. Index Return records the daily index return. News Release captures total macro-news relevance level, provided by Bloomberg. All

Window 1

Window 2

Window 3

Window 4

Window 5

01/03/12–31/01/14

01/03/12–31/03/12

01/05/12–31/07/12

01/11/12–31/01/13

01/05/13–31/07/13

01/11/13–31/01/14

Panel A: arbitrage variables Daily Duration (sec) 606.1 (551.9) Daily Profit ($) 528.6 (416.3) No. of Arbitrage 124.4 (60.4)

232.9 (117.3) 319.9 (145.4) 77.2 (25.6)

489.6 (375.7) 681.1 (476.9) 104.0 (42.2)

681.5 (372.5) 352.9 (348.3) 100.1 (34.8)

370.4 (278.5) 611.9 (386.9) 156.5 (77)

749.9 (543.4) 639.7 (544.8) 147.3 (56.8)

Panel B: control variables Colo. Cabinets Log(Message Traffic) 11.34 (0.232) Index Return (%) −0.019 (0.506) News Release 100.2 (124.7)

20 11.44 (0.129) 0.038 (0.404) 106.7 (147.4)

76 11.35 (0.245) −0.050 (0.62) 106.5 (131.5)

111 11.29 (0.261) 0.022 (0.377) 107.8 (128.2)

117 11.44 (0.253) −0.173 (0.598) 104.2 (127.7)

133 11.30 (0.266) −0.095 (0.543) 98.9 (124.9)

We find a rise in the number of occupied co-location cabinets is significantly related to an increase in mispricing and index arbitrage opportunities measured in terms of value, frequency and duration. These results suggest that contrary to one received view, increased competition does not improve market efficiency as arbitrage opportunities exist and persist in markets. Our results are also consistent with the notion that increased HFTs create additional risks for arbitragers especially if they are able to easily and quickly modify quotes. While minimum latency should reduce legging risk associated with stale quotes, execution uncertainty increases. The ability of HFTs to modify bid and ask quotes over very short periods creates an opportunity for greater misalignment in prices between two closely related instruments. This results in greater price distortions and price discrepancies in markets.

Table 4 Impact of HFT of Arbitrage Measures. Panel A reports coefficient estimates of the following specification: Arbt ¼ Intercept þ β 1 Colocationt þ β 2 log ð1=SPI Dollar Volumet Þ þ β 3 VolatilitySPI;t þ β 4 NewsRelevencet þ β 5 LagReturnt þ εt ; where Arbt is any one of the three dependent variables: Daily Profit, Daily Duration or # Arbitrage. Colocationi is the number of cabinets hosted during the event window (our measure of HFT).log(1/SPI Dollar Volume) is log of the inverse dollar value of traded SPI contracts on day i; VolatilitySPI,i measures the daily volatility level in SPI, using log of the ratio of the highest and lowest mid-price of SPI contract on day i; NewsRelevencei measures total macro-news relevance level as published by Bloomberg on day i; LagReturn is the return of S&P/ASX200 index in the previous trading day. Panel B reports coefficient estimates where the proxy for HFT is log(Message Traffici), measured as the log of daily number of order book message in both SPI and STW. All other control variables are as previously defined. The model is estimated for the period extending March 1, 2012–January 31, 2014. Newey and West (1987) adjusted t-statistics are reported in parenthesis. Panel A

Panel B

Daily Profit

No of Arb

Daily Duration

Daily Profit

No of Arb

Daily Duration

−6704⁎⁎⁎ (−3.81) 0.928⁎

−1724⁎⁎⁎ (−6.44) 0.652⁎⁎⁎

−949 (−0.43) 3.08⁎⁎⁎

−5430⁎⁎⁎ (−4.09)

−1605⁎⁎⁎ (−6.26)

−2989 (−1.44)

(1.7)

(8.87)

(6.39)

Log(1/SPI dollar volume)

−324⁎⁎⁎

−82.8⁎⁎⁎

Volatility

(−3.84) 30,805⁎⁎⁎

(−6.45) 3188⁎⁎⁎

193⁎⁎ (2.55) −165⁎ (−2.15) 32,430⁎⁎⁎

36.3⁎⁎⁎ (2.65) −60.8⁎⁎⁎ (−4.30) 3713⁎⁎⁎

(4.34) −0.137 (−0.96) 918 (0.22) 0.259

(2.74) −0.045⁎⁎ (−2.34) 195 (0.32) 0.424

(5.89) −0.119 (−1.09) 2221 (0.74) 0.240

(3.59) −0.039⁎⁎ (−2.24) 8.895 (0.02) 0.293

−160 (−1.49) −254⁎⁎ (−2.12) −5967 (−0.78) −0.172 (−1.20) −4851 (−1.28) 0.006

Intercept Co-location Log(Message Traffic)

New Release Lag Return Adjusted R2

−57.1 (−0.54) −3978 (−0.49) −0.298⁎⁎ (−2.04) −1801 (−0.46) 0.074

⁎⁎⁎ Significant at 1%. ⁎⁎ Significant at 5%. ⁎ Significant at 10%.

Please cite this article as: Frino, A., et al., The impact of latency sensitive trading on high frequency arbitrage opportunities, PacificBasin Finance Journal (2016), http://dx.doi.org/10.1016/j.pacfin.2016.08.004

A. Frino et al. / Pacific-Basin Finance Journal xxx (2016) xxx–xxx

11

Table 5 Arbitrage Measures after Accounting for Trading and Clearing Fees (March 1, 2012–January 31, 2014). This table shows the mean and distribution of arbitrage measures adjusted for the cost of trading. # of Arbs/Day records daily counts of arbitrages for each trading day. Qty denotes the size of each arbitrage opportunity, measured in number of SPI contracts traded. Daily Profits measures the actual dollar amount of each arbitrage accumulated over the trading day. Daily Duration accumulates the time that each individual arbitrage opportunity lasts over each trading day. Percentile

# of Arbs/Day Qty (SPI Lots) Daily Profit ($) Daily Duration (sec)

Mean

1

5

25

50

75

95

99

24.7 1.13 156.53 4.678

3 0.00 3.02 0.0042

6 0.00 10.27 0.0146

14 0.04 45.21 0.0742

21 0.26 101.11 0.311

33 1.00 197.72 2.216

57 5.00 497.71 27.252

79 9.95 851.73 61.766

Table 6 Impact of HFT on Arbitrage Measures after Accounting for Trading and Clearing Fees. Panel A reports coefficient estimates of the following specification: Arbt ¼ Intercept þ β 1 Colocationt þ β 2 log ð1=SPI Dollar Volumet Þ þ β 3 VolatilitySPI;t þ β 4 NewsRelevencet þ β 5 LagReturnt þ εt ; where Arbt is any one of the three dependent variables: Daily Profit, Daily Duration or # Arbitrage, after accounting for trading and clearing fees. Colocationi is the number of cabinets hosted during the event window (our measure of HFT).log(1/SPI Dollar Volume) is log of the inverse dollar value of traded SPI contracts on day i; VolatilitySPI,i measures the daily volatility level in SPI, using log of the ratio of the highest and lowest mid-price of SPI contract on day i; NewsRelevencei measures total macro-news relevance level as published by Bloomberg on day i; LagReturn is the return of S&P/ASX200 index in the previous trading day. Panel B reports coefficient estimates where the proxy for HFT is log(Message Traffici), measured as the log of daily number of order book message in both SPI and STW. All other control variables are as previously defined. The model is estimated for the period extending March 1, 2012–January 31, 2014. Newey and West (1987) adjusted t-statistics are reported in parenthesis. Panel A

Intercept Co-location

Panel B

Daily Profit

No of Arb

Daily Duration

Daily Profit

No of Arb

Daily Duration

−955.0 (−1.09) 0.451⁎

−407.8⁎⁎⁎ (−5.69) 0.089⁎⁎⁎

−232.1 (−0.39) 0.522⁎⁎⁎

−888.3 (−1.00)

−326.3⁎⁎⁎ (−4.47)

−222.3 (−0.39)

(1.83)

(4.41)

(3.08) −0.672 (−0.01) −45.58 (−0.83) 11,385⁎⁎⁎

17.381⁎⁎⁎ (3.68) −7.04 (−1.50) 548 (1.44) −0.018⁎⁎⁎ (−3.35) −51.5 (−0.27) 0.275

−18.083 (−0.43) −23.95 (−0.67) 2973 (1.04) −0.103⁎⁎ (−2.37) 143.2 (0.08) 0.014

Log(Message Traffic) Log(1/SPI dollar volume) Volatility New Release Lag Return Adjusted R2

−46.27 (−1.11) 11,226⁎⁎⁎ (3.03) −0.124⁎ (−1.85) 53.68 (0.03) 0.101

−19.78⁎⁎⁎ (−5.73) 443 (1.16) −0.018⁎⁎⁎ (−3.42) −14.88 (−0.08) 0.277

−12.30 (−0.43) 2863 (0.98) −0.101⁎⁎ (−2.39) 350.31 (0.20) 0.037

(3.08) −0.125⁎ (−1.85) −124.4 (−0.06) 0.093

⁎⁎⁎ Significant at 1%. ⁎⁎ Significant at 5%. ⁎ Significant at 10%.

References Baron, M., Brogaard, J., Kirilenko, A., 2012. The trading profits of high frequency traders. Working Paper. Biais, B., Woolley, P., 2011. High Frequency Trading. Toulouse University, IDEI Manuscript. Biais, B., Heider, F., Hoerova, M., 2012. Clearing, counterparty risk, and aggregate risk. IMF Econ. Rev. 60 (2), 193–222. Boehmer, E., Fong, K.Y., Wu, J.J., 2014. International evidence on algorithmic trading. Working Paper. Brailsford, T., Hodgson, A., 1997. Mispricing in stock index futures: a re-examination using the SPI. Aust. J. Manag. 22 (1), 21–45. Breckenfelder, J.H., 2013. Competition between high-frequency traders, and market quality. Working Paper. Brogaard, J., Garriott, C., 2015. High-frequency trading competition. Working Paper. Brogaard, J., Hendershott, T., Riordan, R., 2014a. High-frequency trading and price discovery. Rev. Financ. Stud. 27 (8), 2267–2306. Brogaard, J., Hendershott, T., Hunt, S., Ysusi, C., 2014b. High-frequency trading and the execution costs of institutional investors. Financ. Rev. 49 (2), 345–369. Budish, E., Cramton, P., Shim, J., 2015. The high-frequency trading arms race: frequent batch auctions as a market design response. Q. J. Econ. 130 (4), 1547–1621. Carrion, A., 2013. Very fast money: high-frequency trading on the NASDAQ. J. Financ. Mark. 16 (4), 680–711. Chaboud, A., Chiquoine, B., Hjalmarsson, E., Vega, C., 2014. Rise of the machines: algorithmic trading in the foreign exchange market. J. Financ. 69 (5), 2045–2084. Chordia, T., Goyal, A., Lehmann, B.N., Saar, G., 2013. High-frequency Trading. Johnson School Research Paper Series Vols. 20–2013. Chordia, T., Green, T.C., Kottimukkalur, B., 2015. Do high frequency traders need to be regulated? Evidence from algorithmic trading on macro news. Working Paper. Cummings, J.R., Frino, A., 2011. Index arbitrage and the pricing relationship between Australian stock index futures and their underlying shares. Account. Finance 51 (3), 661–683. Epps, T.W., 1979. Comovements in stock prices in the very short run. J. Am. Stat. Assoc. 74 (366a), 291–298. Frino, A., Mollica, V., Webb, R.I., 2014. The impact of co-location of securities exchanges' and traders' computer servers on market liquidity. J. Futur. Mark. 34, 20–33. Gagnon, L., Karolyi, G.A., 2010. Multi-market trading and arbitrage. J. Financ. Econ. 97 (1), 53–80. Gatev, E., Goetzmann, W.N., Rouwenhorst, K.G., 2006. Pairs trading: performance of a relative-value arbitrage rule. Rev. Financ. Stud. 19 (3), 797–827. Goldstein, M.A., Kumar, P., Graves, F.C., 2014. Computerized and high-frequency trading. Financ. Rev. 49 (2), 177–202.

Please cite this article as: Frino, A., et al., The impact of latency sensitive trading on high frequency arbitrage opportunities, PacificBasin Finance Journal (2016), http://dx.doi.org/10.1016/j.pacfin.2016.08.004

12

A. Frino et al. / Pacific-Basin Finance Journal xxx (2016) xxx–xxx

Hasbrouck, J., Saar, G., 2013. Low-latency trading. J. Financ. Mark. 16 (4), 646–679. Hendershott, T., Charles, M.J., Albert, J.M., 2011a. Does algorithmic trading improve liquidity? J. Financ. 66, 1–33. Hendershott, T., Jones, C.M., Menkveld, A.J., 2011b. Does algorithmic trading improve liquidity? J. Financ. 66 (1), 1–33. Hogan, S., Jarrow, R., Teo, M., Warachka, M., 2004. Testing market efficiency using statistical arbitrage with applications to momentum and value strategies. J. Financ. Econ. 73 (3), 525–565. Jain, P., McInish, T., 2012. Reduced latency and market quality on the Tokyo Stock Exchange. Working Paper. Jarnecic, E., Snape, M., 2014. The provision of liquidity by high-frequency participants. Financ. Rev. 49 (2), 371–394. Jones, C.M., 2013. What Do We Know About High-frequency Trading? Columbia Business School Research Paper Vols. 13–11. Kozhan, R., Tham, W.W., 2012. Execution risk in high-frequency arbitrage. Manag. Sci. 58, 2131–2149. Madhavan, A., 2012. Exchange-traded funds, market structure, and the flash crash. Financ. Anal. J. 68 (4), 20–35. Marshall, B.R., Nguyen, N.H., Visaltanachoti, N., 2013. ETF arbitrage: intraday evidence. J. Bank. Financ. 37 (9), 3486–3498. Menkveld, A.J., 2013. High frequency trading and the new market makers. J. Financ. Mark. 16 (4), 712–740. Miller, M.H., Muthuswamy, J., Whaley, R.E., 1994. Mean reversion of Standard and Poor's 500 Index basis changes: arbitrage-induced or statistical illusion? J. Financ. 49 (2), 479–513. Newey, W.K., West, K.D., 1987. A simple, positive semi-definite, heteroskedasticity and autocorrelation consistent covariance matrix. Econometrica 55, 703–708. O'Hara, M., 2015. High frequency market microstructure. J. Financ. Econ. 116 (2), 257–270. Pontiff, J., 1996. Costly arbitrage: evidence from closed-end funds. Q. J. Econ. 1135–1151. Scholtus, M., van Dijk, D., Frijns, B., 2014. Speed, algorithmic trading, and market quality around macroeconomic news announcements. J. Bank. Financ. 38, 89–105. Sofianos, G., 1993. Index arbitrage profitability. J. Deriv. 1, 6–20. Toth, B., Palit, I., Lillo, F., Farmer, J.D., 2015. Why is equity order flow so persistent? J. Econ. Dyn. Control. 51, 218–239. Van Kervel, V., Menkveld, A.J., 2016. High-frequency trading around large institutional orders. Working Paper. Webb, R.I., 2003. Transitory real-time property rights and exchange intellectual property. J. Futur. Mark. 23, 891–913. Xiong, W., 2001. Convergence trading with wealth effects: an amplification mechanism in financial markets. J. Financ. Econ. 62 (2), 247–292.

Please cite this article as: Frino, A., et al., The impact of latency sensitive trading on high frequency arbitrage opportunities, PacificBasin Finance Journal (2016), http://dx.doi.org/10.1016/j.pacfin.2016.08.004