* Trading on Algos Johannes A. Skjeltorp(cid:132) Norges Bank Elvira Sojli Erasmus University Rotterdam Wing Wah Tham Erasmus University Rotterdam and Tinbergen Institute Preliminary Draft Comments welcome Abstract This paper studies the impact of algorithmic trading (AT) on asset prices. We find thattheheterogeneityofalgorithmictradersacrossstocksgeneratespredictablepatterns in stock returns. A trading strategy that exploits the AT return predictability generates a monthly risk-adjusted performance between 50-130 basis points for the period 1999 to 2012. We find that stocks with lower AT have higher returns, after controlling for standard market-, size-, book-to-market-, momentum, and liquidity risk factors. This effect survives the inclusion of many cross-sectional return predictors and is statistically and economically significant. Return predictability is stronger among stocks with higher impedimentstotradeandhigherpredatory/opportunisticalgorithmictraders. Ourpaper is the first to study and establish a strong link between algorithmic trading and asset prices. Keywords: Asset pricing, Algorithmic trading, Market quality, Liquidity. JEL Classification: G10; G20; G14. *The authors thank Schmuel Baruch, Tarun Chordia, Thierry Foucault, Amit Goyal, Terry Hendershott, AlbertMenkveld,RyanRiordan,NormanSchurhoff,Bernt-ArneØdegaard,andparticipantsattheNorthern Finance Association Meeting 2013, ESSFM Gerzensee 2013, and the International Workshop on Market Mi- crostructureandNonlinearDynamicsforusefulcommentsandsuggestions. Theviewsexpressedarethoseof the authors and should not be interpreted as reflecting those of Norges Bank (Central Bank of Norway). (cid:132)Corresponding author. Address: Central Bank of Norway, P.O.Box 1179, Sentrum, NO-0107 Oslo, Nor- way. Email: [email protected], Phone: (+47) 22316740. Other authors’ email addresses: [email protected] (Sojli), [email protected] (Tham). 1 Introduction Algorithmic trading (AT) has increased enormously in recent years and it is estimated to account for 53% of U.S. daily equity trading volume. The episodes of the Flash crash in 2010 and the runaway trading code of Knight Capital in August 2012, which cost $440 million for its shareholders, highlight the economic importance of understanding the impact of algorithmic trading on asset prices. Despite the growing importance of AT in financial markets, there is no work studying its impact on asset returns. This paper fills the gap by examining the cross-sectional relation between AT and stock returns. A major challenge in studying the impact of AT on market quality and on asset prices is the availability of long time series of public data. We overcome this issue by constructing an AT measure based on the order-to-trade ratio (OTR) using (publicly available) Trade and Quote (TAQ) data and examine its evolution for the period 1999 to 2012.1 Using detailed high-frequency trading data (hereafter HFT data) obtained from NASDAQ OMX, wecompareourmeasurewiththefractionoftradesandquotesbyHFTperstockasprovided by NASDAQ.2 This data set is also used by Brogaard, Hendershott, and Riordan (2013) to study the impact of high-frequency trading on price discovery. We find that our measure is highly correlated, with a correlation of 65%-75%, to several measures of AT from the HFT data. Using a data sample of NYSE, AMEX, and NASDAQ-listed firms from January 1999 to October 2012, we find a raw return differential between the low AT and the high AT portfolio of 9.4% per year. The AT effect is robust to adjustments for risk factors as well as firm characteristics. A portfolio of stocks with low AT outperforms a portfolio of stocks with high AT by 50 to 130 basis points per month after adjusting for the market, size, book-to- market, momentum, and liquidity factors. The negative relation between AT and returns is significant even after controlling for other widely documented return predictors. The return 1Hendershott, Jones, and Menkveld (2011) propose the use of message-to-trade ratios as an AT measure for studying the impact of AT on liquidity in NYSE. SEC Chairman Mary Schapiro supports the use of OTRasameasureofATforpoliciestocurbexcessivemessaging. TheEuropeanParliament’sEconomicand MonetaryAffairsCommittee(ECON)suggeststolevyafeefortradingmemberswhoexceedanOTRof250:1, see http://www.thetradenews.com/news/Regions/Europe/MEPs_squeeze_HFT_in_MAD_proposals.aspx. 2Hagstro¨mer and Nord´en (2013) show that 98.2% of HFT messages and 96.7% of their trades are gener- ated using algorithms, while only 13.5% of non-HFT messages and 63.6% of messages from hybrid firms are generated by algorithms in the OMXS30 index in 2012. 1 difference is larger among smaller stocks, more illiquid stocks, and stocks with higher delay in information diffusion. The results are robust to different holding periods and various asset pricing tests. WeproposetwopotentialexplanationsfortheexistenceoftheATeffect. First, thehigher returnsofstockswithlowATmayreflectadelayininformationdiffusionamongthesestocks. Biais, Hombert, and Weill (2010) suggest that AT reduces the cognitive inability of human traderstoexecutetheirtasksefficientlyandquickly. ATcanimprovethespeedofinformation diffusion through trading algorithms that parse information from news wires and electronic sources almost instantly. More AT represents not only quicker response to news arrivals and the ability to identify short-lived mis-pricings, but also more information coverage and aggregation. Thus, more AT decreases delays in information diffusion and reduces trading frictions, see Chaboud, Chiquoine, Hjalmarsson, and Vega (2009), Hendershott et al. (2011), and Hendershott and Riordan (2012). Using the Hou and Moskowitz (2005) measure of how quickly stock prices respond to information, we find little support for the cognitive inability of human traders hypothesis. Thesecondexplanationisbasedontheheterogeneityofalgorithmictradersacrossstocks. The diversity of AT strategies implies a considerable heterogeneity in algorithmic trader types. In general, AT strategies are classified as market making and predatory algos, and these different algo types participate in stocks with different characteristics. Hagstr¨omer and Nord´en(2013)showthatmarketmakingalgoshavehigherquote-to-traderatiosandaremore prevalent in smaller stocks with wider tick size, since the profit of market making per trade increases with the tick size of a stock. However, predatory algos use their speed advantage overslowertraderstorespondtonews, toanticipatelargeordersofbuy-sideinstitutions, and to exploit cross-market arbitrage activities. Thus, we expect to find more predatory algos in stocks with slower traders and with more buy-side institutions. Given the pick-off risk, slower traders and buy-side institutions will require higher returns. We test the AT heterogeneity hypothesis using the detailed NASDAQ HFT data with information on liquidity demanders and suppliers as well as trades between faster algorithms and slower human traders. We find that risk adjusted returns are higher in stocks with more active HFT trading against slower passive non-HFT traders and lower for stocks with more 2 market making HFTs. Furthermore, the AT effect is more prevalent in stocks with more institutional investors. These results suggest that the higher risk-adjusted returns in lower AT stocks are associated with higher pick-off risk created by opportunistic algos that prefer to trade with humans and in stocks with more institutional investors. We also study the role of market frictions/impediments-to-trade on the persistence of the AT effect. A profitable trading opportunity can persist only if there are market frictions that discourage arbitrageurs from exploiting it. We find that the AT effect is stronger among smaller stocks and stocks with higher transaction costs, supporting the impediments-to-trade hypothesis. Overall, these findings suggest that the AT effect arises from the heterogene- ity of algorithmic traders across different stocks and impediments-to-trade perpetuate the phenomenon. This paper is closely related to the literature on liquidity and asymmetric information and asset prices. O’Hara (2003) argues that financial markets can play an important role for asset prices through liquidity and price discovery. Amihud and Mendelson (1986), Brennan and Subrahmanyam (1996), Brennan, Chordia, and Subrahmanyam (1998), Amihud (2002), Chordia, Subrahmanyam, and Anshuman (2002b), Jones (2002), and Brennan, Chordia, Subrahmanyam, and Tong (2012), among many others, provide evidence that liquidity is an important determinant of expected returns. Brennan and Subrahmanyam (1996), Easley, Hvidkjaer, and O’Hara (2002) and Easley and O’Hara (2004), among others, argue that stocks with more information asymmetry should have higher expected returns. We provide supportfor thesetwostreams ofliteraturebyshowing howAT affectsasset pricesvia adverse selection and stock liquidity levels. The paper is also related to the growing literature on understanding the impact of al- gorithmic and high frequency trading in financial markets. Biais and Woolley (2011) and Jones (2013) provide a survey of the literature on HFT. In theory, algorithmic trading can be beneficial for financial markets as it may mitigate traders cognition limits (Biais et al., 2010), but algorithmic traders can increase predatory behavior and adverse selection (Biais, Foucault, and Moinas, 2011; Foucault, Hombert, and Ro¸su, 2012) and increase imperfect competition (Biais et al., 2011). The empirical literature addresses the effects of AT on trade execution, liquidity, and market efficiency. Algorithms appear to reduce execution costs and 3 risks (Domowitz and Yegerman, 2005; Engle, Ferstenberg, and Russell, 2012) and improve arbitrageurs’abilitytoeliminateassetmispricings(Chaboudetal.,2009). Severalpapersfind that AT increases competition among trading venues and liquidity providers, provides liq- uidity when it is scarce, and improves price efficiency (see Hendershott et al., 2011; Chaboud et al., 2009; Hendershott and Riordan, 2012; Brogaard et al., 2013). However, Kirilenko, Kyle, Samadi, and Tuzun (2012) provide empirical evidence of adverse selection by showing that HFT are able to predict price changes at the expense of slower traders. Chaboud et al. (2009) find more correlated trades among algorithmic traders, which can potentially increase systemic risk in the spirit of Khandan and Lo (2011). Overall, the evidence on whether mar- ket quality is higher or lower with AT is mixed. Differently from the existing empirical work in the market microstructure literature, we investigate the impact of AT on financial markets through an asset pricing perspective by studying the relation between AT and asset returns across portfolios with varying algorithmic activity. Our results highlight the importance of accounting for heterogeneity in algos when studying their impact on market quality and asset prices. This paper is among the first to construct a proxy for AT using publicly available data, which can be used to inform the intense public and academic debate about the impact of AT on market quality. Previous studies have mainly focused on market microstructure issues due to the lack of long time series of publicly available data. So far, measures of AT come from propriety databases and the time series of these measures is short. Chaboud et al. (2009) study the impact of AT on market efficiency in the foreign exchange market using proprietary ElectronicBrokingSystem(EBS)datafrom2003to2007. Hendershottetal.(2011)construct measuresofATusingelectronicmessagetrafficandtradesinNYSE’sSuperDOTsystemfrom theproprietary NYSE’sSystemOrderDatadatabasefrom2001to2005. HasbrouckandSaar (2013)usetwomonthsofNASDAQ-ITCHdatain2008toconstructameasureofproprietary algorithmsparticipationinstocks. HendershottandRiordan(2012)useproprietary Deutsche Boerse data which identify whether or not the order was generated with an algorithm for 30 DAX stocks in January 2008. Brogaard et al. (2013) study the role of HFT, a subset of AT, using NASDAQ data that identifies HFT trading activity on a stratified sample of stocks (120 stocks) in 2008 and 2009. Boehmer, Fong, and Wu (2012) construct a proxy for AT 4 in 39 individual exchanges around the world from 2001 and 2009 to study the impact of AT on market quality using the Thomson Reuters Tick History database. Using TAQ data, we construct an AT measure consistent with Hendershott et al. (2011) and Boehmer et al. (2012) for the whole U.S. equity market. We validate the reliability of OTR from TAQ as an appropriate AT measure by comparing it with the shorter but more detailed NASDAQ HFT data. The validation facilitates future research and analysis on the role and impact of AT in broader U.S. based asset pricing and corporate finance studies. 2 Data We employ the trades and quotes reported in TAQ for the period January 1999 to October 2012, to construct a long time series of stock level AT. AT has been taking place since 1999, aftertheU.S.SecuritiesandExchangeCommissionapprovedtheoptionforelectronicsystems to register as full-fledged exchanges in 1998.3 Therefore our sample starts in 1999. We retain stocks listed on the NYSE, AMEX, and NASDAQ for which information is available in TAQ, Center for Research in Security Prices (CRSP), and Compustat. Following the literature, we use only common stocks (Common Stock Indicator Type=0), common shares (Share Code 10 and 11), and stocks not trading on a “when issued” basis. Stocks that change primary exchange, ticker symbol, or CUSIP are removed from the sample (Hasbrouck, 2009; Goyenko, Holden,andTrzcinka,2009;Chordia,Roll,andSubrahmanyam,2000). Wealsoremovestocks that have a price lower than $5 and higher than $1,000 at the end of a month. To avoid look ahead biases, all filters are applied on a monthly basis and not on the whole sample. There are5,978individualstocksinthefinalsample. Throughoutthepaper, weusereturnsof-30% for the delisting month (delisting codes 500 and 520-584), as in Shumway (1997). All returns are calculated using bid-ask midpoint prices, adjusted for splits and cash distributions, to reduce market microstructure noise effects on observed returns (Asparouhova, Bessembinder, and Kalcheva, 2010, 2013). Table A.1 reports the definitions and the construction details for all variables and Table A.2 in the Appendix provides the summary statistics. 3Informal talks with HFT trading firms confirm this. Oxford Advisors Ltd launched Oxford High Fre- quency Trading Equity Fund (OHFT) in January 2001, the Sun Herald discusses benefits and costs of high frequencytradinginFebruary2001. InstitutionalInvestorcoversalgorithmictradinginJanuary2002,claiming that it has been around since the late 80’s. 5 2.1 Proxy for algorithmic trading To investigate the link between AT and asset prices, we need a proxy for stock level AT. Our proxy is the monthly number of quote updates in TAQ relative to the number of trade executions (order-to-trade ratio) across all exchanges. We define a quote update as a change inthebestbidorofferpriceorinthequantityatthebestbid/offerpriceinanyexchange. The order-to-trade ratio has recently been applied by several exchanges to reduce the explosion of message traffic resulting from the increase in AT.4 As many trading algorithms quote and cancel orders very rapidly to detect how the market is moving, to trigger responses by other traders, to identify hidden liquidity, etc., the strategies used by algorithmic traders, and high frequency traders in particular, have contributed to a huge increase in the amount of message traffic relative to trade executions. Thus, the order-to-trade ratio should capture changes in AT both in the cross-section and over time. We calculate the AT proxy at the monthly frequency for each stock by summing the daily number of quote updates and the number of trades, across all exchanges. AT for stock i in month t is measured by the monthly number of messages across all exchanges divided by the number of trades: N(quotes) i,t AT = , (1) i,t N(trades) i,t whereN(quotes) andN(trades) denotethemonthlynumberofquoteupdatesandtrades, i,t i,t respectively, for stock i in month t.5 2.2 Consistency of the AT measure The AT measure proposed is based on the consideration that the more AT there is, the more quote updates there will be. Nonetheless, we need to ensure that the measure we have constructed is reliable and captures AT. We use a set of detailed HFT data obtained 4The London Stock Exchange was the first to introduce an “order management surcharge” in 2005 based on the number of trades per orders submitted. They revised this charge in 2010. Euronext, which comprises theParis,Amsterdam,Brussels,andLisbonstockexchanges,hasoperatedonesince2007. In2012DirectEdge introduced the “Message Efficiency Incentive Program”, where the exchange pays full rebates only to traders thathaveanaveragemonthlymessage-to-traderatiolessthan100to1. InMay2012theOsloStockExchange introducedanorder-to-executefee,wheretradersthatexceedaratioof70foramonthincurachargeofNOK 0.05 (USD 0.0008) per order. Deutsche Bo¨rse and Borsa Italiana have announced similar measures in 2012. 5Another way to measure algorithmic activity is to construct a ratio of trading volume to quote updates, seeHendershottetal.(2011). Ourresultsremainqualitativelythesamewhenusingthismeasure. Theresults are available from the authors upon request. 6 from NASDAQ OMX to compare AT with the fraction of quotes by HFT per stock. The data consists of trade data for a stratified sample of 120 randomly selected stocks listed on NASDAQ and NYSE for 2008 and 2009 and trade and quote data for February 22-26, 2010. Trades identify the liquidity demander and supplier as HFT or non-HFT. Quotes identify the best bid and offer by HFTs or non-HFTs and any updates to the best prices and quantities. More details on this dataset are provided in Brogaard et al. (2013), who use it to study the impact of high-frequency trading on price discovery. To evaluate our AT measure, we use the HFT trade and quote data from February 2010, because it allows us to generate an HFT quote measure that we cannot do with the earlier HFT data from 2008 and 2009. Foreachstockandeachdate, wecalculatethenumberoftotalquoteupdates(HFT+non- HFT) in the NASDAQ HFT database N(quotes)NASD and the number of quote updates reportedbyNASDAQinTAQ,N(quotes)NASD|TAQ. Thecross-sectionalcorrelationbetween these two measures in Table 1 is 94%. We then calculate the daily number of quote updates by HFT, N(quotes)HFT, and the daily number of trades by HFT, N(trades)HFT. The correlationbetweenN(quotes)HFT andN(quotes)NASD|TAQ is87%,whichmeansthatalarge proportion of NASDAQ quote updates reported in TAQ are HFT related. We then calculate ATHFT = N(quotes)HFT which is comparable with our market-wide AT measure based on the N(trades)HFT TAQ data. The correlation between AT and ATHFT is 57% and highly significant with a t-value of 15.44. We also compare the AT proxy for the NASDAQ TAQ data, where we only use NASDAQ trades and quotes in TAQ to calculate the AT proxy, ATNASD|TAQ = N(quotes)NASD|TAQ/N(trades)NASD|TAQ with ATHFT. The correlation between ATNASD|TAQ and ATHFT is 75% and highly signifi- cant. Overall, the results from comparing the AT measure with actual HFT data show that our AT measure is a valid measure of HFT and AT. 7 2.3 Algorithmic trading characteristics Panel A of Figure 1 shows the equally weighted monthly AT over the sample period. AT has increased substantially over time. Panel B shows that the driver behind the increase in AT is the explosion in quote updates relative to executed trades. Panel C shows the time series of AT for market capitalization (MCAP) quintiles constructed at the end of each month. The MCAP1 portfolio includes the smallest stocks and MCAP5 the largest stocks. AT is higher for lower market capitalization stocks, similar to what Hendershott et al. (2011) document using NYSE only data to measure AT. Table 2 examines the determinants of AT in a regression setting. The dependent variable is the yearly AT measure. We run a two-way fixed effects panel regression with standard errors clustered at the stock level. Table 2 shows that AT is higher in stocks with fewer analyst following and lower institutional ownership, stocks with higher prices and larger spread, and smaller stock with lower trading. To summarize, the descriptive statistics how that there is significant variation in AT across stocks and a dramatic increase in AT over time. Moreover, cross-sectionally, our proxy for AT decreases with market capitalization, consistent with evidence in Hendershott et al. (2011). 3 Algorithmic trading and returns 3.1 Raw returns We examine the raw return characteristics of AT portfolios across stock characteristics, to analyze the relationship between AT and returns. Table 3 shows the average monthly returns in excess of the risk free rate for portfolios cross-sorted on various characteristics and AT. We usemonthlyconditionalsorts,wherefirstthesampleisdividedintothreeportfoliosbyvarious firm characteristics, such as size, every month t. We then sort characteristic-based terciles intofiveATportfolios. Thereareapproximately145stocksineachportfolioeachmonth. We show the equally weighted average excess return of each portfolio and the return difference between the low and high AT portfolios in month t+1. The double-sorts control for firms characteristicsoneatatimeandshowthelowtohighATportfoliodifferenceisnotassociated with a particular characteristic, but is pervasive across all cross sorts: Size (MCAP), Book- 8 to-Market(BM),relativespread(SPREAD),USDtradingvolume(USDVOL),pastmonth return (R1), and past 12 month return (R212). 3.2 Risk adjusted returns Next, we test whether the return differential between the low and high AT stocks can be explained by the market, size, value, momentum, and liquidity factors. Each month, all stocks are divided into portfolios based on AT at time t. Portfolio returns are the equally weighted average realized returns of the constituent stocks in each portfolio in month t+1. We estimate individual portfolio loadings with a 24 months rolling window regression: J (cid:88) r = α + β X +ε , (2) p,t+1 p p,j j,t p,t+1 j=1 wherer isthereturninexcessoftheriskfreerateformontht+1ofportfoliopconstructed p,t+1 at month t AT level, and X is the set of J risk factors: excess market return (r ), value j,t m HML (r ), size SMB (r ), Pastor and Stambaugh (2003) liquidity (r ), and momentum hml smb liq UMD (r ).6 umd Table 4 reports alphas for 5, 10, 25, and 50 AT portfolios. There are 433, 217, 87, and 43 stocksineachportfoliorespectively. Thelow-ATportfolio(AT1)hasastatisticallysignificant monthlyalpha(α )thatrangesbetween0.9and1.3%acrossvariousportfoliosplitsandasset 1 pricing models. The high-AT portfolio alphas range from -0.2% to 0.3%, but are statistically not different from zero in most specifications and portfolio splits. This suggests that the high AT portfolios are priced well by the factor models. However, the risk-adjusted returns between the low and high AT portfolios are statistically significant and vary between 0.5% to 1.3% per month across different AT portfolios. The profitability of the long-short strategy derivesmainlyfromthelongpositionortheperformanceoflow-ATportfolio(AT1)insteadof theshortpositionofAT10,whichlimitsconcernsabouttheimpactofshort-sellingconstraints on this strategy. So far we have only considered one month holding (portfolio re-balancing) periods, but 6Since we are using portfolios conditional on AT, we only have portfolio returns from February 1999. We usea24monthestimationwindowtoincreasethesampleperiod. Fortheindividualstockregressions,weuse a 48 month rolling window to estimate factor loadings. 9