Computer-based trading in the cross-section

Transcription

1 Computer-based trading in the cross-section Torben Latza, Ian Marsh and Richard Payne June 15, 212 Abstract We investigate low-latency, computer-based trading in almost 3 stocks on the London Stock Exchange. We demonstrate that the proportion of trades that we identify as computer-based is increasing in stock liquidity and decreasing in the ratio of tick size to price. Signed computer-based trading activity leads other signed trades. Computer-based trading results in lower execution costs for the least active stocks and also for stocks with high tick to price ratios. Computerbased trades enhance trading efficiency in this regard. The information content of computer-based trades is high relative to non-computer-based trades but this advantage decreases with stock liquidity and with tick to price ratios. Computer-based trades contain relatively high levels of information in hard to trade stocks, but it is not clear whether this is value-related information or, given their ability to predict subsequent flows, whether it reflects an ability to predict the evolution of the order book. Keywords: Market microstructure; computer-based trading; high-frequency trading; transactions costs; asymmetric information; London Stock Exchange. Thanks to the London Stock Exchange for providing the data used in this study and to Cass and Warwick Business Schools for supporting this work with research pump-priming grants. All errors are our own. Faculty of Finance, Cass Business School, City University, London. Correspondence: richard.payne@city.ac.uk.

2 1 Introduction In recent times, the effects of automated trading and quoting by computers on price dynamics and execution costs in equity markets has been put firmly under the spotlight. The US flash crash of May 6th 21 where, in a matter of minutes, US stock indices and single stock prices dropped precipitously before quickly recovering has been blamed by many on computer-generated trading and quoting activity. Regulators around the world are looking closely at how computer-based trading (CBT) affects the quality of markets and trying to decide whether and, if so how, market microstructures should be altered so as to limit the scope of CBT. 1 Much of the discussion of CBT seems to treat it as synonymous with high-frequency trading (HFT) activities of pure trading firms, proprietary trading desks in investment banks and hedge funds. Broadly speaking, high-frequency traders look to profit from correcting small, transient mis-pricings in markets using technology that allows them to identify trading opportunities at millisecond timescales. 2 However, an enormous quantity of computer-based trades emanates from a completely different source. Investment banks use algorithmic trading engines to fill the orders of their clients. Here the motivation is to execute efficiently, usually dynamically across time, an order for stock in the sense of minimising transactions costs relative to some benchmark. 3 The original client orders are unlikely to be related to high-frequency mis-pricings in markets, although they may be designed to exploit longer-run forecasting rules. Given the preceding definitions, one would expect high-frequency trading and algorithmic execution to have very different effects on markets. HFT activity is likely to carry more information relevant at very short timescales and thus contribute more 1 In the UK, which is the focus of this study, the Department for Business, Innovation and Skills has set up a Foresight programme to investigate how CBT is impacting upon UK markets. See 2 Note that this does not only entail liquidity demand. If the bid-ask spread is greater than the level justified by microstructure fundamentals then a high-frequency trader can tighten the spread and thus profit through liquidity supply. 3 One such benchmark might be the volume weighted average price of the stock over the interval that the order was traded. Another might be the price of the stock at the time the oerder was submitted to the broker. 1

3 towards price discovery as measured on a millisecond timescale. Algorithmic trades will likely carry less high-frequency information, but will likely be timed to take advantage of temporarily favourable trading conditions. Previous work in this area has verified that high-frequency trading carries information and contributes towards market quality. Brogaard (21) identifies quoting and trading by high-frequency trading firms for 12 Nasdaq stocks observed in February 21. He shows that high-frequency trades contribute more to efficient prices than do other trades and argues that, overall, high-frequency trading firms are beneficial to market quality. Hendershott and Riordan (211) show that computer-based traders in DAX stocks contribute to price discovery, both through liquidity supply and liquidity demand. Our aim is to understand whether the positive and negative effects of CBT are uniform in the cross-section of stocks. If algorithmic trading brings execution cost benefits to clients, is it able to do so for both the most heavily and least heavily traded stocks? Our prior is that CBT is most likely to be useful in less liquid markets where being the first to spot fleeting favourable trading conditions has value. Conversely, in more liquid stocks, with low tick sizes, trading costs are likely to be uniformly low such that even a high-latency trader can execute reasonably well. Similarly, we proceed to ask whether HFT engines identify profitable mis-pricings in even the most liquid of stocks, despite the enormous quantity of attention focussed on such stocks? Alternatively, are the gains from HFT trading less liquid stocks large enough to outweigh the increased costs of aggressive trading? Is the information content of HFT activity uniform in the cross-section? These are empirical questions and we let the data speak. In order to investigate how the effects of CBT change in the cross-section, we are left with the problem of identifying computer-based trades. Unfortunately our quote and trade data do not contain any identifiers for the traders involved and thus we cannot separate activity based on the identities of the counter-parties. Thus we use an alternative mechanism. We examine each marketable order in the data and classify it as being generated by a computer if the order that it executes against is less than 2

4 1 milliseconds old. Essentially we argue that the small reaction time of computers relative to humans means that incredibly quick executions of new limit orders must come from a computer. Of course, this is not guaranteed. Just by chance, two human orders may be entered within milliseconds of one another and match. At the other end of the spectrum, a computer algorithm might decide to hit an order much older than 1ms (if other market conditions have changed). Thus our classification scheme is clearly not perfect. However, we feel that it is very likely to separate out a set of computer-based trades from all others. Our data-based approach to the identification of computer-based trading resembles the analysis in Hasbrouck and Saar (211). We apply our CBT identification mechanism to data on all stocks electronically traded on the London Stock Exchange over the period from July to December 28. In the cross-section we have just under 3 stocks. For each stock, we reconstruct the order book tick-by-tick for each trading day in the sample and then apply our scheme for identifying CBT to all market orders and all marketable limit orders. Based on this classification of trades, we show that; The proportion of trades that we identify as CBT is increasing in stock liquidity and strongly decreasing in the ratio of tick size to price. The information content of computer-based trades relative to non-computer based trades decreases with stock liquidity and decreases with tick size. CBT leads to reduced trading costs in low activity stocks and stocks with large tick sizes. Thus, the effects of computer-based trading alter in very clear fashion in the crosssection: CBT does carry more information and can also lead to more efficient execution than other trades, but only for less liquid stocks. Attempts to control or limit computer-based trading should take note of this, in that in making it harder for high-frequency trading firms to do business, will likely impede information flow into less liquid securities and make algorithmic trading of such securities less efficient. 3

5 This could have the effect of increasing implementation shortfall for mutual funds, for example, and thus reducing returns to ultimate investors. A small point to note from our analysis is that if, however, one desires a mechanism to control the scale of CBT then tick size is an obvious candidate. Increased tick sizes would see CBT reduced. However, increasing tick sizes leads to the price impact of CBT rising relative to that of other trades. Thus while increasing tick size might discourage CBT, it seems that the computer-based trades that remain carry more information. The rest of the paper is set out as follows. The next section provides a brief review of relevant literature and our data is described in Section 3. Section 4 describes our empirical analysis and our results. Section 5 concludes. 2 Literature Review Prompted by market developments such as the May 211 Flash Crash in the US, regulatory interest and an emerging empirical literature, the last few years has seen the growth of a body of work that theoretically analyses the impact of computerbased trading on market outcomes (i.e. prices, volumes and liquidity). These papers differ in their characterisation of computer-based traders such that their predictions also differ quite dramatically. Some papers offer a largely positive view of computer-based trading. Gerig and Michayluk (21) model an automated market-maker that, unlike human marketmakers, can trade multiple securities and process information from related markets and embed such a trader in a standard Glosten and Milgrom (1985) sequential trade model. The existence of the automated market-maker leads to greater informational efficiency, larger volumes and lower trading costs for liquidity traders. Martinez and Rosu (211) also conclude that computer-based traders may be beneficial to markets. They model them as aggressive traders who exploit a speed advantage and in doing 4

6 so ensure that information is more quickly reflected in prices. Another subset of papers suggests that computer-based trading can be harmful for markets. Cartea and Penalva (211), for example, introduce a class of parasitic computer-based traders who are quick enough to trade in between retail traders and professional dealers. In a Grossman-Miller style world this leads to retail traders paying higher costs as the computer-based traders extract some surplus from trading activity. Jarrow and Protter (211) argue that, in aggregate, the activity of CBTs, who possess advantages in trading speed and trade independently but on correlated signals, may create mis-pricings in securities that increase volatility and which damage the welfare of non-automated traders. 4 Hoffmann (211) incorporates speed differentials into the Foucault (1999) framework where the key advantage of speed is the fact that, after revelation of public information, fast traders orders are less likely to be picked off than those of slow traders. Hoffman s model predicts that, in certain scenarios, the advantage enjoyed by HFTs leads to non-hfts suffering higher execution costs and also that total social welfare may be reduced. Similar results are obtained by Jovanovic and Menkveld (211). In contrast to the variation in the results of theory work on the effects of CBT on market quality, empirical work on the issue by and large suggests a positive relationship between the two. We have mentioned above the contributions of Brogaard (21) and Hendershott and Riordan (211), which demonstrate, using Nasdaq and Xetra data respectively, that high-frequency trading enhances price discovery and informational efficiency. Hendershott, Jones, and Menkveld (211) measure the extent of CBT using the level of message traffic on the NYSE and argue that CBT leads to more informative price quotes, increased liquidity and smaller asymmetric information problems. Menkveld (212) studies the entry of a single new high-frequency trading player to the market for Dutch stocks in 27-28, and shows that this new player behaves as a market-maker, with the majority of its trades passive and thus 4 Cohen and Szpruch (211) build on the Jarrow and Protter (211) paper to show how HFTs may profitably front-run slower traders and how this may reduce market efficiency. They go on to analyse whether a Tobin Tax might limit the effects of HFTs on market quality. 5

7 contributing to lower bid-ask spreads. Finally, Kirilenko, Kyle, Samadi, and Tuzun (211) analyse the role played by high-frequency traders during the US flash crash of May 21. Looking at detailed trading data for the CME s S&P E-mini contract they argue that HFTs did not generate the flash crash, attributing the decline in prices to the actions of fundamental sellers. They do, however suggest, that after the start of the decline in prices, HFTs became net sellers of futures also, exacerbating the overall size of the crash. One thing to note about both the theoretical and empirical work above is that there is no focus on how the effects of CBT might differ in the cross-section of stocks. Theoretically, there are no clear predictions as to how to expect CBTs to behave in less versus more liquid stocks and empirically there has, thus far been no work that seeks to compare the interactions of CBT and market quality in the cross-section. 3 Data 3.1 Data sources and coverage We base our analysis on order level data from the London Stock Exchange for the final six months of 28. The data provided by the exchange contains every message arriving at the SETS order book. Our starting set of securities contains 313 stocks that were SETS traded during this interval. These range from the most liquid Blue Chip stocks (e.g. BP PLC or Rio Tinto PLC) which see thousands of trades on the average trading day, to much smaller issues that trade relatively infrequently. As the data contains every order level message we can build the order book precisely, millisecond by millisecond, and we can measure the lifetimes of orders precisely (to the millisecond). Moreover, we can compute the age of any limit order at the time of its (partial or complete) execution. We exclude the first five minutes of trading each day to remove the impact of the 6

8 opening auction. Further, since financial stocks in the UK were subject to restrictions on short selling in the second half of our sample we also exclude all financial stocks from the majority of our analysis. We consider the impact of the short sales ban separately. Finally, some outlier stocks are removed where they have been affected by mergers and other unusual corporate activities. Finally, we are left with just less than 3 stocks in our analysis. 3.2 Identifying computer-generated orders We use a very simple scheme to classify executions as being computer generated. Any execution that results from a new order filling a standing order that is less than 1 milliseconds old is assumed to be computer generated. 5 Essentially, we argue that 1 milliseconds is too short a time interval for a human to identify and act upon an order book event and thus that the preponderence of the events we isolate must come from automated trading systems. 6 Of course, our classification scheme is not foolproof. A market order emanating from a human may, by chance, execute against a limit order that was entered only a few microseconds previously. However, the chances of this happening are slight relative to the probability of such a trade resulting from computer-based interaction. At the other end of the spectrum, a computer may decide to hit an order that is multiple seconds old, perhaps due to changes in markets conditions in a related stock. This is much more likely and probably happens quite frequently in our data. On balance, however, we feel that our trade classification scheme is likely to do a reasonable job of separating out the form of low-latency computer-generated trades that regulators seem to be focussed on at present and thus is both empirically and economically interesting. 5 Note that trades are aggregated so that executions at the same price, with the same direction and at the same millisecond are assumed to be one execution. 6 A time interval of.1 seconds is regarded by many to be the lower bound of human reaction time. For example, in the 1 metres sprint, athletes reacting to the starting gun within 1ms of it firing are deemed to have false started. 7

9 Some graphical evidence on the time between order entries and execution times are given in Figures 1 and 2. These figures show that around 15% of executions meet our definition of a CBT. This percentage is slightly larger for more liquid stocks and somewhat smaller for stocks with larger tick sizes relative to price. The figures also show that there is a spike of probabililty mass for orders less than 1 ms old. 3.3 Basic statistics for computer-based and other trades We briefly consider the simple cross-section variation in CBT-activity and some basic relationships between CBT activity and market conditions. We observe the following: Figure 3 and Table 1: CBT activity monotonically increases with stock liquidity Figure 4 and Table 1: CBT declines as the ratio of tick size to price increases Computers appear to find high tick size stocks, which tend to have larger depths and more sticky prices, less easy to trade. Table 2 suggests the following realtionships between CBT activity and spreads, volume and volatility when we condition on both the normal activity level (characterised by ADV) and expensiveness to tarde (tick to price ratios): there is no clear correlation between CBT and average bid-ask spreads at either a 5 min or 3 min sampling frequency and regardless of whether one focusses on particularly liquid or illiquid stocks or high or low tick size stocks. there is a tendency, most clear for the most liquid and lower tick size stocks, for the proportion of CBT activity to increase in high volume times. there is a similar result for the relationship between the share of CBT in overall trading and volatility. 8

10 finally, the relationship between CBT and volatility is stronger for low tick size stocks: perhaps low tick sizes make it easier for CBT to pick off stale limit orders in volatile times and thus their share of activity is increased. These last three realtionships are interesting because, as we will show below, the advantages of CBT over non-cbt is minimised (and may even be negative) in periods of high trading activity and high volatility. Thus, at the time when CBT activity levels are highest the advantages they bring are minimised. 4 Empirical analysis 4.1 Order flows: auto and cross correlations For each stock we compute fast and slow order flow measures defined as the sum of signed traded quantities within a one minute interval resulting from CBT and other trades. Note that the LSE data allows us to sign trades precisely. We first present auto- and cross-correlations in these flow measures. Both fast and slow order flows are positively autocorrelated (Figure 5). Averaging across all stocks, the first order autocorrelation coefficient is.16 for slow flow and.6 for fast flow. Small though significantly positive autocorrelations in both types of flow remain even when considering lags of up to ten minutes. On average, slow flows are always more positively correlated than fast flows. However, for stocks with high ADV the difference between autocorrelations in fast and slow stocks is very small after the first lag. Based on these results, it is tempting to regard our low-latency, CBT order flow as being more opportunistic than slow flows. The latter are easier to classify as resulting from traders splitting large orders and feeding them into the market over time. To facilitate cross-stock analysis we proceed to standardise flows by dividing fast (slow) flow by fast (slow) volume to give an imbalance measure. We then compute 9

11 cross-correlations between fast and slow flows for a stock at various displacements. Averaging across stocks we find that strongly positive cross-correlations are found between slow flows and lagged fast flows (Figure 6). The peak cross-correlation of +.37 is found between minute t slow flows and t 1 fast flows, but correlations are greater than +.1 even when fast flows are lagged by five minutes. The one minute-lag cross-correlation is even higher than the contemporaneous correlation of.2. Conversely, correlations between fast flows and lagged slow flows are much lower, typically less than.5, though still positive on average. Splitting stocks according to ADV reveals some clear patterns. All cross-correlations are lower than average for high volume stocks, and higher than average for low volume stocks. For high volume stocks, correlations between fast flows and lagged slow flows are negative, though magnitudes are very small. Nevertheless, the asymmetry in cross-correlations remains and fast flows appear to lead slow flows for all three activity subsamples. Peak cross-correlations are always at the -1 displacement, and are very large for the less active stocks. It is important to emphasise the directional element to these results. The results do not just say that fast trades precede slow trades, they indicate that fast buys (sells) tend to precede slow buys (sells). These results suggest that CBT anticipate non-cbt flows. This might be at the core of what CBT does, in the sense that they make money by executing in front of slower traders in a Brunnermeier and Pedersen (25) predatory trading sense. Alternatively, it may simply be explained by CBT being quicker to react to public revelation of stock or market level news. In order to better understand the nature of fast and slow order flows we run a regression of stock i s standardised fast order flow on the contemporaneous market fast order flow (standardised fast order flow in all stocks excluding stock i). Results suggest that fast flows are positively correlated across stocks - fast buy (sell) orders in stock i take place at the same time as fast buy (sell) orders for other stocks. We find the same for slow orders but the regression coefficients and goodness of fit measures are much higher. Thus, the commonality of order flow found elsewhere in the litera- 1

12 ture (Hasbrouck and Seppi, 21) may be driven by commonality in slow flows but is also true for fast flows. There is significant cross-section variation in these findings as the least liquid stocks have more strongly correlated fast flows than the most liquid stocks, while slow flow cross-stock correlations are highest among the more liquid. 4.2 Execution Costs For each stock we assign every trade an execution cost. This cost is equal to the distance, in basis points, between the trade price and the midquote just prior to the execution of the trade. For buy orders this measure (in an electronic market) is guaranteed to be positive and for sell orders it is always negative. Thus, we take the negative of the measure for sell trades. Denote that measure for trade t in stock i by z i,t. We then regress our measure of execution costs on a constant, a dummy to indicate whether the trade was computer-generated and controls for the size of the trade, overall stock volume in the 1 minutes preceding the current trade and realised stock volatility over the preceding 1 minutes; z i,t = β,i + β 1,i CBT i,t + β 2,i T radesize i,t + β 3,i V olume i,t + β 4,i σ i,t + u i,t (1) Given that SETS is a pure limit order book, these execution cost regressions are simply asking whether, controlling for market conditions, there are systematic differences in bid-ask spreads at times of low latency CBT versus the times of other trades. Alternatively, do computer-based executions tend to capture smaller spreads than do human executions? Table 4 presents the results. Consider first the control variables. As expected, execution costs are higher in more volatile periods and for larger transactions. Conversely, but still as expected, execution costs are generally lower in high volume periods. The magnitudes of these effects all decrease as we consider more liquid stocks as proxied by average daily volume (ADV) measures. The constant terms indicate that execution costs for stocks 11

13 with high tick to price ratios (T/P) are higher than for low T/P stocks within ADV categories, however there are no systematic patterns when we compare estimated coefficients for control variables. Moving to the key variable in the analysis, the results show that CBT execute at significantly lower costs than non-cbt in all regressions. The gain is between onequarter and two-thirds of a basis point for high ADV stocks and over 3bp for low ADV stocks. The gains fall as a proportion of the non-cbt execution costs (i.e. the constant) for more liquid stocks. These results suggest that algorithms trading more liquid stocks do not offer that many advantages in execution cost terms over human trades. Of course we don t have dynamics in the analysis and so we cannot distinguish worked CBT orders from worked human orders. In less liquid stocks, though, it appears that the speed advantage offered by CBT translates directly into them picking off better prices. When trading and quoting is less continuous (due to low natural trading interest, or high tick sizes), CBTs can jump in and take good prices as they appear and before humans can react. CBTs still execute on better terms than do non-cbts in the most liquid stocks, but their advantage is much smaller. 4.3 Information contents We estimate price impacts for CBT and other trades with the following simple regression which is run in transaction time: r t = α + P α i r t i + i=1 P j= β j x CBT t j + P k= γ k x Other t k + ɛ t (2) where now x CBT t x Other t is a signed, transaction indicator for computer-based trades and is a similar signed indicator for non-cbt activity. The price impacts for CBT and other trading respectively are given by P j= β j and P k= γ k. The regressions 12

14 use K = 1 for all stocks. 7 Table 5 presents our results. The first two panels show that, as expected, price impact (PI) is higher for less liquid stocks, irrespective of whether it is a CBT or non-cbt. Price impacts are also higher for high T/P stocks than for low T/P stocks. Panel C demonstrates three key findings. First, the ratios of CBT price impact to that of other trades are all greater than unity, suggesting that on average CBT contain more information than do other trades. Second, this price impact ratio decreases with liquidity. CBT in the least liquid stocks contains more information (relative to other trades) than it does in more liquid stocks. Third, the price impact ratio is lower for low T/P stocks than for high T/P stocks within a given ADV liquidity category. That CBT brings information to the market is worth emphasising. CBT is often characterised as predatory and of no social value. However, our results suggest that it brings, on average, more information to the market than do other trades. From this perspective, our results accord very clearly with those of Hendershott and Riordan (211) who examine one month of trading in German DAX stocks. Our key contribution is to examine how this informational advantage of CBT varies across stocks. The results suggest that price impacts of CBT or non-cbt in the most liquid stocks are essentially the same, probably since these are highly researched and very actively traded assets. Conversely, and arguably as expected, CBT has a clear edge over non-cbt in less liquid stocks with higher degrees of informational asymmetry. Per trade, therefore, CBT contains less information for more liquid stocks than it does for less liquid stocks. And compared to an average non-cbt trade, a CBT trade contains relatively more information in the less liquid stocks than it does in more liquid stocks. At least two explanations for this suggest themselves. First, it is possible that better information about asset payoffs in less liquid stocks give CBT 7 Results are robust to running a full VAR as in Hasbrouck (1991) and estimating price impacts using impulse response functions. Note that this entails causally ordering the trade variables, but we have tried this both ways. 13

15 this pattern of advantage. Second, we might hypothesise that CBT benefits from better information about the state of the order book for a stock, and the shallower the order book the better this advantage becomes. 8 This is supported by our earlier findings that fast trades lead slow trades on the same side of the order book. We explore this further below. The other aspect to the findings in Table 5 is the role of T/P. The absolute level of information content of CBT (and non-cbt) increases with T/P and in relative terms the advantage of CBT over non-cbt is higher for high T/P firms. Again this is suggestive of CBT carrying more information for more difficult to trade stocks. Table 6 presents similar statistics where the sample period is split in two. The first half of our sample was relatively stable, while the second half, corresponding to the last quarter of 28, was both very volatile and saw a higher than usual trade volume as the financial crisis developed rapidly and spilled outside the financial sector. 9 We observe that, as expected, price impacts of CBT and non-cbt are both higher in the volatile period. More surprisingly, the relative price impact of CBT to non-cbt is lower during the volatile second half of the sample. Indeed, for the most easily traded stocks with high ADV and low T/P, the price impact of CBT is lower than that of non-cbt. This might suggest that the CBT advantage is not simply news flow related since the last three months of 28 was characterised by a high level of important economywide and value-relevant news. Our results would be consistent with the hypothesis that in non-turbulent times, CBT brings information to market by quickly exploiting high frequency predictabilities in markets. However, in very turbulent times the patterns that CBT trade on become clouded by more important informational issues, such that the statistically regularities they trade on may actually break down. This 8 It is hard to imagine that CBT has a material advantage over non-cbt in understanding the evolution of the order book for the most liquid stocks since the bid and ask sides resemble queues with exceptionally large depth available at the inside spread. An advantage is more likely in more complex books where liquidity drain and replenish cycles, for example, might be exploitable by a fast-acting algorithm. 9 Remember that we have excluded financial sector stocks from our analysis. 14

16 leads to their information advantage falling. We explore this hypothesis by examining conditional information contents 4.4 Conditional information contents The cross-sectional and time-series variation of information contents of CBT versus other trades highlighted in the previous section demands further analysis. In this section we compute conditional price impacts, first allocating stocks into one of six categories based on three ADV and two T/P buckets, and then conditioning on (i) time of day, (ii) trading volume in the preceding 1 minutes and (iii) midquote return volatility in the preceding 1 minutes. We present these results graphically [regression-based analysis to follow]. Conditioning by time-of-day reveals some interesting patterns (Figure 7). First, it is clear that price impacts for CBT and non-cbt are highest at the opening. The opening is generally considered to be a period of higher than usual information asymmetry (driven by out of trading hours news flow, for example, generating information differences between traders). On many exchanges it is also a period of higher than usual trading activity but the LSE is an exception to this rule. Outside of the opening call auction (which is excluded from our sample), trading volume and quoting activity are both relatively low in the first hour of trading, especially for less liquid stocks. Second, CBT have much higher price impacts for the least actively traded stocks, irrespective of T/P (top row). Third, as we consider more actively traded stocks the information advantage of CBT declines. For the most easily traded stocks with high ADV and low T/P (bottom left graph) there is essentially no difference between the information content of CBT and non-cbt. Figure 8 plots price impacts conditioning on recent trading volume. In the lowest volume quartiles (quartile 1) CBT contain more information than non-cbt. The advantage is largest for less liquid stocks and for stocks with higher T/P ratios, but is present for all categories. In periods of high trading volume (quartile 4) there is 15

17 essentially no difference between the information content of CBT and non-cbt. Finally, Figure 9 shows that conditioning on quote activity also indicates that CBT have relatively high information content in low activity periods (quartile 1) but, if anything, non-cbt contain more information in periods of high activity levels (quartile 4). It is important to note that these findings hold across the spectrum of usual trading activity as captured by ADV. When stocks trade a lot relative to normal volume, or trade in a volatile fashion, the informational advantage of CBT seems to disappear. In summary, for the most liquid, low tick size stocks it is not clear that CBT have any information advantage at all. Therefore, in such stocks they bring very little to the table in terms of either execution quality or information. It is in the less liquid stocks that CBT matters. CBT can reduce trading costs relative to non-cbt, and fast trades appear to have more information (or can react quicker to information) in these stocks. However, even this advantage can be eroded by high levels of trading activity and/or high levels of price volatility. 5 Conclusion We study how the effects of computer-based trading vary in the cross-section of stocks. Using a data-based definition of computer-based trading activity we show that; Based on this classification of trades, we show that; The proportion of CBT that we identify is increasing in stock liquidity and decreasing in the ratio of tick size to price. Signed CBT leads signed other activity. CBTs are not only quicker than non- CBTs, but their buys (sells) precede the buys (sells) of other traders. Are they predatory, as in Brunnermeier-Pedersen, or parasitic as in Cartea- Penalva? 16

18 CBT leads to greater reductions trading costs for low activity stocks and for stocks with large tick sizes than for more active stocks with small tick sizes. This is evidence of the trading efficiency that CBT can deliver in hard to trade stocks and, indirectly, suggests that CBT might be generating positive effects for those clients who use algorithmic execution systems. The information content of computer-based trades relative to non-computer based trades decreases with stock liquidity and with tick size. The basic result echoes earlier work, but the fact that the advantage is larger in least liquid stocks deserves thinking about. Surely these are the stocks with the largest mis-pricings, and so it is conceivable that CBT brings value-related information to the market place. However, these are also stocks where knowledge of likely developments in the order book are of most value. In liquid cheap to trade stocks, order books are generally deep and stable. Knowledge of likely evolutions in this type of order book are unlikely to be of value for trading. Conversely, in less liquid, and more expensive to trade stocks, order books are thin and potentially volatile. Computer-based traders able to predict changes in the state of such order books have the potential to gain. The relatively high information content of CBT may therefore reflect both value-related information and/or order flow information. Policy: tick size is a possible lever. It reduces CBT by our definition and doesn t seem to damage the ability of CBT to save on execution costs all that much. It systematically increases the information advantage of CBT trading relative to other trades, though. While the former might be something that regulators quite like, the latter is arguably a nasty side effect. 17

19 References Brogaard, J., 21, High Frequency Trading and Market Quality, SSRN elibrary. Brunnermeier, M., and L. Pedersen, 25, Predatory Trading, Journal of Finance, 6. Cartea, Ã., and J. Penalva, 211, Where is the Value in High Frequency Trading?, SSRN elibrary. Cohen, S. N., and L. Szpruch, 211, A Limit Order Book Model for Latency Arbitrage, SSRN elibrary. Foucault, T., 1999, Order flow Composition and Trading Costs in a Dynamic Limit Order Market, Journal of Financial Markets, 2, Gerig, A., and D. Michayluk, 21, Automated Liquidity Provision and the Demise of Traditional Market Making, SSRN elibrary. Glosten, L. R., and P. Milgrom, 1985, Bid, ask, and transaction prices in a Specialist arket with heterogeneously informed agents, Journal of Financial Economics, 14, Hasbrouck, J., 1991, Measuring the Information Content of Stock Trades, Journal of Finance, 46(1), Hasbrouck, J., and G. Saar, 211, Low-Latency Trading, SSRN elibrary. Hasbrouck, J., and D. Seppi, 21, Common Factors in Prices, Order Flows and Liquidity, Journal of Financial Economics, 59. Hendershott, T., C. M. Jones, and A. J. Menkveld, 211, Does Algorithmic Trading Improve Liquidity?, Journal of Finance, 66, Hendershott, T. J., and R. Riordan, 211, Algorithmic Trading and Information, SSRN elibrary. Hoffmann, P., 211, A Dynamic Limit Order Market with and Traders, SSRN elibrary. Jarrow, R. A., and P. Protter, 211, A Dysfunctional Role of High Frequency Trading in Electronic Markets, SSRN elibrary. Jovanovic, B., and A. J. Menkveld, 211, Middlemen in Limit-Order Markets, SSRN elibrary. Kirilenko, A. A., A. P. S. Kyle, M. Samadi, and T. Tuzun, 211, The Flash Crash: The Impact of High Frequency Trading on an Electronic Market, SSRN elibrary. 18

20 Martinez, V. H., and I. Rosu, 211, High Frequency Traders, News and Volatility, SSRN elibrary. Menkveld, A. J., 212, High Frequency Trading and the New-Market Makers, SSRN elibrary. 19

21 Table 1: Proportion of computer-based trading: sorted by stock liquidity and tick/price Low T/P High T/P Low ADV.131 (55).115 (48) Med ADV.169 (42).143 (62) High ADV.184 (61).147 (45) Notes: for each of our stocks, we estimate the proportion of trades that are computer-generated. Stocks are split into 3 ADV based groups (at the 33rd and 66th percentiles of the ADV distribution) and 2 tick size groups (stocks above and below the median value of tick to price). The median proportion of computer-generated trades in each of the 6 bins is presented above. In parentheses, the number of stocks in each group is presented 2

22 Table 2: Correlations between computer-based trading activity and other variables: 5 min sampling : sorted by stock liquidity and tick/price Panel A: bid-ask spreads Low T/P High T/P Low ADV Med ADV High ADV Panel B: volume Low T/P High T/P Low ADV Med ADV High ADV Panel C: volatility Low T/P High T/P Low ADV Med ADV High ADV Notes: for each of our stocks, we compute the correlation between the proprtion of trades generated by a computer in a 5 minute interval and spreads, volumes and volatilities respectively at the same sampling frequency. Stocks are split into 3 ADV based groups (at the 33rd and 66th percentiles of the ADV distribution) and 2 tick size groups (stocks above and below the median value of tick to price). The median correlation in each of the 6 bins is presented above. 21

23 Table 3: Correlations between computer-based trading activity and other variables: 3 min sampling : sorted by stock liquidity and tick/price Panel A: bid-ask spreads Low T/P High T/P Low ADV Med ADV High ADV.2.21 Panel B: volume Low T/P High T/P Low ADV Med ADV High ADV Panel C: volatility Low T/P High T/P Low ADV Med ADV High ADV Notes: for each of our stocks, we compute the correlation between the proprtion of trades generated by a computer in a 3 minute interval and spreads, volumes and volatilities respectively at the same sampling frequency. Stocks are split into 3 ADV based groups (at the 33rd and 66th percentiles of the ADV distribution) and 2 tick size groups (stocks above and below the median value of tick to price). The median correlation in each of the 6 bins is presented above. 22

24 Table 4: Trade by trade execution cost regressions: CBT versus non-cbt Constant CBT dummy Trade size Volume Volatility Panel A: Low ADV Low T/P (33.22) (-4.57).378 (4.435) (-1.59).325 (7.561) High T/P (28.349) (-3.734).156 (3.975).95 (.333).316 (5.527) Panel B: Med ADV Low T/P 9.96 (56.82) (-7.965).42 (5.515) -.1 (-4.264).142 (8.798) High T/P (52.61) (-4.7).37 (5.442) -.4 (-1.757).176 (8.678) Panel C: High ADV Low T/P (79.56) (-7.25).15 (7.19) -.4 ( ).148 (2.221) High T/P 1.34 (7.95) (-2.49).12 (6.415) -.3 (-7.339).156 (12.71) Notes: results from regressions of executions costs for individual trades on a constant, a dummy to indicate whether a trade was computer generated or not, trade size (in 1s of shares), total market volume and stock return volatility. Regressions are run separately for each stock and then split into 6 groups based on a bivariate sort on ADV and the ratio of tick size to price. Stocks are split into 3 ADV based groups (at the 33rd and 66th percentiles of the ADV distribution) and 2 tick size groups (stocks above and below the median value of tick to price). The results above are the mean coefficients and t-values in each group. [RGP comment: we needs regression R-squared for this table. Also, we need to remove MECOM from these results.] 23

25 Table 5: Price impacts of computer-based versus other trades: sorted by stock liquidity and tick/price Panel A: Price impact of CBT Low T/P High T/P Low ADV Med ADV High ADV Panel B: Price impact of other trades Low T/P High T/P Low ADV Med ADV High ADV Panel B: Price impact ratio Low T/P High T/P Low ADV Med ADV High ADV Notes: for each of our stocks, we estimate the price impacts of computer-based trades and other trades using the regression model in equation(2). Price impacts should be interpreted as the basis point movement in price associated with a trade of median size (in shares). Stocks are split into 3 ADV based groups (at the 33rd and 66th percentiles of the ADV distribution) and 2 tick size groups (stocks above and below the median value of tick to price). The mean price impacts of the two trade groups and the ratio of the price impacts is presented above. 24

26 Table 6: Price impacts of computer-based versus other trades across sub-samples: sorted by stock liquidity and tick/price First Half Second Half Panel A: Price impact of CBT Low T/P High T/P Low T/P High T/P Low ADV Med ADV High ADV Panel B: Price impact of other trades Low T/P High T/P Low T/P High T/P Low ADV Med ADV High ADV Panel B: Price impact ratio Low T/P High T/P Low T/P High T/P Low ADV Med ADV High ADV Notes: for each of our stocks, we estimate the price impacts of computer-based trades and other trades using the regression model in equation(2). The sample is split into two halves. Price impacts should be interpreted as the basis point movement in price associated with a trade of median size (in shares). Stocks are split into 3 ADV based groups (at the 33rd and 66th percentiles of the ADV distribution) and 2 tick size groups (stocks above and below the median value of tick to price). The mean price impacts of the two trade groups and the ratio of the price impacts is presented above. 25

27 Figure 1: Time from order entry to execution: sorted by ADV Low ADV Med ADV High ADV.14 Propn of trades Milliseconds Notes: ADV is average daily volume. Figure 2: Time from order entry to execution: sorted by ratio of tick size to price Low T/P Med T/P High T/P.14 Propn of trades Milliseconds 26

28 Figure 3: The proportion of CBT activity: sorted by ADV Propn of trades Stocks: ADV rank Notes: ADV is average daily volume. Figure 4: The proportion of CBT activity: sorted by ratio of tick size to price Propn of trades Stocks: T/P rank 27

29 Figure 5: Autocorrelations of flows: sorted by ADV Autocorrelations: fast and slow flows: all stocks Autocorrelations: large ADV Displacement Displacement Autocorrelations: medium ADV Autocorrelations: small ADV Displacement Displacement Notes: ADV is average daily volume. 28

30 Figure 6: Cross autocorrelations of flows: sorted by ADV.4 Cross correlations: fast and slow flows (negative displacment = fast lead): all stocks.3 Cross correlations: large ADV Displacement Displacement.4 Cross correlations: medium ADV).7 Cross correlations: small ADV Displacement Displacement Notes: ADV is average daily volume. 29

31 Figure 7: The price impact of a trade, sorted by ADV and T/P, conditional on time of day Price impacts by time of day: ADV SS 1, Tick SS Price impacts by time of day: ADV SS 1, Tick SS Hour of day Hour of day Price impacts by time of day: ADV SS 2, Tick SS Price impacts by time of day: ADV SS 2, Tick SS Hour of day Hour of day Price impacts by time of day: ADV SS 3, Tick SS Price impacts by time of day: ADV SS 3, Tick SS Hour of day Hour of day Notes: ADV is average daily volume. 3

32 Figure 8: The price impact of a trade, sorted by ADV and T/P, conditional on trading volume Price impacts by aggvolume subsample: ADV SS 1, Tick SS Price impacts by aggvolume subsample: ADV SS 1, Tick SS aggvolume quartile aggvolume quartile Price impacts by aggvolume subsample: ADV SS 2, Tick SS Price impacts by aggvolume subsample: ADV SS 2, Tick SS aggvolume quartile aggvolume quartile Price impacts by aggvolume subsample: ADV SS 3, Tick SS Price impacts by aggvolume subsample: ADV SS 3, Tick SS aggvolume quartile aggvolume quartile Notes: ADV is average daily volume. 31

33 Figure 9: The price impact of a trade, sorted by ADV and T/P, conditional on quote price volatility Price impacts by quotevol subsample: ADV SS 1, Tick SS Price impacts by quotevol subsample: ADV SS 1, Tick SS quotevol quartile quotevol quartile Price impacts by quotevol subsample: ADV SS 2, Tick SS Price impacts by quotevol subsample: ADV SS 2, Tick SS quotevol quartile quotevol quartile Price impacts by quotevol subsample: ADV SS 3, Tick SS Price impacts by quotevol subsample: ADV SS 3, Tick SS quotevol quartile quotevol quartile Notes: ADV is average daily volume. 32