Synchronicity, Instant Messaging and Performance among Financial Traders
aa r X i v : . [ phy s i c s . s o c - ph ] O c t Synchronicity, Instant Messaging and Performanceamong Financial Traders
Serguei Saavedra ∗ † , Kathleen Hagerty † , and Brian Uzzi ∗ † ∗ Northwestern Institute on Complex Systems, Northwestern University, Evanston, Illinois, USA, 60208, and † Kellogg School of Management, Northwestern University, Evanston,Illinois, USA, 60208Submitted to Proceedings of the National Academy of Sciences of the United States of America
Successful animal systems often manage risk through synchronousbehavior that spontaneously arises without leadership. In criticalhuman systems facing risk, such as financial markets or militaryoperations, our understanding of the benefits associated to syn-chronicity is nascent but promising. Building on previous work il-luminating commonalities between ecological and human systems,we compare the activity patterns of individual financial traders withthe simultaneous activity of other traders—an individual and spon-taneous characteristic we call synchronous trading. Additionally, weexamine the association of synchronous trading with individual per-formance and communication patterns. Analyzing empirical dataon day traders’ second-to-second trading and instant messaging, wefind that the higher the traders’ synchronous trading, the less likelythey lose money at the end of the day. We also find that the dailyinstant messaging patterns of traders are closely associated withtheir level of synchronous trading. This suggests that synchronicityand vanguard technology may help cope with risky decisions in com-plex systems and furnish new prospects for achieving collective andindividual goals. collective behavior | synchronicity | communication | data mining | complexsystems S ynchronous behavior has been found to enhance individ-ual and group performance across a variety of domainseven though the individuals might make no conscious effortto coordinate their behavior [1–6]. Similarly, in systems of col-laboration and competition, synchronous behavior can eludesimple associations with individual benefits [5,7–10]. Cicadasthat chirp simultaneously with others find the best balancebetween risk and reward [10]. Cicadas that chirp in advanceor in delay of the full chorus relish the best chances of findinga mate but may suffer the greatest risk of being spotted by apredator [8,9]. Congruently timed humans’ actions have beensurmised to provide potential benefits revealed by the mean-ingful coincidence of synchronous behavior [1, 11–15]. For in-stance, it has been found that simultaneous discoveries, orthe times when multiple individuals arrive at a similar con-clusion simultaneously, is collective evidence that a solutionis valid [11, 12, 14].In this paper, we studied the association between individ-ual performance and the simultaneous activity patterns fol-lowed by independent decision makers under risk. These con-ditions exist in many high-frequency decision contexts but areuniquely well documented in financial systems [16], where con-tinuous change in information creates recurring uncertaintyabout when to trade and the second-to-second actions of fi-nancial traders are recorded [17,18]. Reducing the risk of los-ing money is the essence of trading [17,19–21]. Over time therisks of trading can decrease as information is disambiguated.However, as this happens, the increasing certainty of informa-tion is incorporated into low-return prices. Thus, racing to bethe first to discover the right time to trade is the critical prob-lem to be solved [20,21]. By analogy, this optimal timing mayrepresent the mating sweet spot observed for cicadas. Chirpin the sweet-spot and the chance of mating/returns is rela-tively high and predation/losses is relatively low. This sug- gests that as separate traders disambiguate their local viewof news, they can spontaneously and simultaneously react asa group, without intention to coordinate, producing a syn-chronous behavior that might reveal the right time to tradein the market. Here, we tested whether traders’ performanceis relatively better when trading simultaneously with othertraders—an individual and spontaneous characteristic we callsynchronous trading.Additionally, traders need to assess whether informationis positive or negative for a stock, the potential magnitude ofthe information’s impact, and the degree to which the infor-mation is already reflected in the price [20, 22]. Social cor-roboration is key to making these assessments [19, 23, 24]. Itreduces cognitive overload and ambiguity when diverse viewsconverge on a common interpretation [12, 14] and typicallytakes place among persons tied through social network rela-tions [25]. We tested whether daily instant messaging patternsof traders are associated with the rise of synchronous trading.We believe our results can have broad implications for un-derstanding fast collective action solutions to decision makingunder uncertainty. Empirical Setting
We observed all the second-to-second trades and instant mes-sages of all 66 stock day traders in a typical trading firm from9/07-02/09 (see Material and Methods). These day traderstraded only stocks and made > .
8% oftheir trades are live, non-computerized trades (computerizedtrades were omitted from the analysis). Day traders typicallydo not hold stocks for more than a day. They typically enterand fully exit all their positions daily, which creates a stan-dardized measure of performance: whether the trader made orlost money at the end of each day. On average, these tradersmake money on just 55% of their trades.Despite sitting in the same firm, these traders generallytrade independently rather than in teams. This is becausethey typically trade different stocks. One trader might tradehigh tech, one trader health care, another trader autos, andso on. Trading different stocks helps them diversify the firm’sholdings, exploit their specialized sector knowledge, and avoidtrading against each other. This means that traders have littleincentive to simply mimic each other’s trades or trading be-
Reserved for Publication Footnotes
Issue Date
Volume Issue Number 1 – avior. Nevertheless, despite trading different stocks, tradersdo process common market information. Common informa-tion includes Federal Reserve announcements, new job figures,housing market change, speculations about bankruptcies, orother global socio-economic data that traders attempt to dis-ambiguate by exchanging information with others as they en-deavor to discover the right time to trade [17,19]. A key formof information exchange here, and increasingly in other humancomplex systems, is instant messaging [26]. Instant messagesamong traders and their network are based on elective re-lationships. Each trader has the autonomy to communicatewith persons of their choice. Hence, they are hierarchical tiesin which orders are dictated from managers to subordinates.Within the communications, information is both professionaland personal. Typical information includes interpretationsof market news, expectations of where the market might bemoving, rumors, and forms of personal information commonlyexchanged among business friends [27].Our extensive field research and interviews with tradersat the firm confirmed that the content of these traders’ mes-sages included information consistent with the content foundin other research [27]. All the traders in the firm exchangedinstant messages throughout the day with their network. In-stant messages were sent and received from their terminalsor mobile devices. By federal law all instant messages tiedto trading go through the firm’s capture system. The im-portance of instant messages to these traders is instantiatedby the intensity its use. We analyzed the full population of > Results
Synchronous trading.
To measure the synchronous trading ofeach trader with other traders, we defined a measure thatquantifies the extent to which an individual’s specific selectionof time to trade is the same as the selection of other traders.To compare the synchronous trading among all traders acrossour observation period, we quantified the degree to which thenumber of traders T ij trading within the same time windowsas trader i in day j compares to the same value when ran-domizing just the trades of trader i (Fig. 1). Specifically, thisrandomization ensures normalizing individual activity, whilekeeping the trading structure and information heterogeneitiesof each specific trading day constant (e.g., number of traders,total number and timing of interactions, and number of inter-actions per trader). Mathematically, we defined synchronoustrading as s ij = ( T ij − h T ∗ ij i ) /σ T ∗ ij , where T ij is the observednumber of simultaneous traders and h T ∗ ij i and σ T ∗ ij are the av-erage and standard deviation of simultaneous traders acrossan ensemble of random replicates within which the trades oftrader i were randomly shuffled. The greater the degree of atrader i ’s daily s ij , the greater her synchronous trading canbe, and vice versa. Additionally, we defined the advancedtrading s − ij and delayed trading s +1 ij in a similar fashion asthe synchronous trading at zero-time lag s ij , but quantifiedthe number of traders trading one window late and one win-dow early, respectively.We examined multiple time windows and reported the fullanalysis for 1-second windows. This window size was chosenfor the main results for several reasons. First, the 1-secondlevel of resolution comports well with the frenetic informationenvironment and fast reaction time dynamics of modern mar-kets [18]. The time scale in which information heterogeneitiesexist has increasingly shortened with the growth of computer-ized trading, which now accounts for between 30% to 60% ofthe trading volume on financial exchanges [18,20]. In comput- erized trading, preset algorithms trade very large volumes ofshares in hundreds of a second, which means that traders mustreact to market opportunities that appear and disappear ona second-by-second scale. Also consistent with the view thatinformation moves at high speeds in modern markets and thattraders react at that level, we found that the traders in thisfirm do display a propensity to trade on 1-second time scales.The average empirical interval of consecutive 1-second tradesis 1 .
01 with a standard deviation of 0 .
14. The maximum in-terval was 9 and it occurred only twice in the data. Similarly,reaction time research has found that human reaction occursin less than 1-second time frames [28, 29]. Second, we chose1-second intervals for synchronous trading because it is thefinest, most conservative time resolution in our data. Largerthan 1-second intervals require a priori knowledge to find theappropriate balance between a window large enough to encap-sulate changes in slow, non-computerized information hetero-geneities and yet not so large that unrelated activities appearsynchronized because they occur in a large window. Workingempirically to estimate this balance, we tested larger than 1-second intervals. We found that our results exist for intervalsup to 15 seconds. This window size seems to be a realistic limitfor slower moving types of information and suggests that syn-chronicity is associated with individual thresholds that rangeacross different information heterogeneities in this complexsystem [29, 30].Our examination of the existence of synchronous tradingrevealed three interesting findings. First, Figure 2A pools allof our data and shows the probability density of synchronoustrading at zero-time lag s ij , advanced trading s − ij and de-layed trading s +1 ij . We observed that synchronous trading issignificantly different ( p < − using Kolmogorov-Smirnovtest) to advanced and delayed trading. Values greater than s ij >
10 comprise 0 .
03% of the entire data and are due totwo traders, who could have a better access to informationor better reaction times [29, 30]. We conservatively omittedthese outliers from our statistical analyses and found that ourresults did not change, confirming that synchronous tradingis a special characteristic of collective behavior [2, 4, 8]. Thisalso suggests that timing is a key factor driving the decisionof traders, and reminds us about the high-frequency changesin the market [18].Second, Figure 2B shows that the average synchronoustrading h s ij i increases with the market’s daily uncertainty( p < − using Markov randomizations), as given by thestandard market volatility index, the VIX [17]. This find-ing supports the idea that collective behavior is associatedwith uncertainty as in the case of biological systems [6, 31].The greater the level of uncertainty faced by individuals, themore likely is a collective behavior such as schooling or flock-ing to arise. These findings suggest a parallel in human sys-tems. As the level of uncertainty in the market increases,the more likely is synchronous behavior to occur. Underhigh-uncertainty days, the average synchronous trading canincrease to h s ij i ≈
2, i.e. the average synchronous trading ofall traders is almost two standard deviations higher than theexpected by chance.Third, we found evidence that synchronous trading doesnot appear to be due to coordination. Unlike coordinated be-havior, where pairs or sets of actors consistently align theirbehavior, synchronous activity commonly displays the oppo-site pattern [4]. We found that 98% of all pairwise correla-tions between activity patterns of two different traders arenon-significant p > .
15. This is probably because no two in-dividuals consistently follow the same strategic behavior andthe same two actors are not always correct in their assess-ments about the market. Similarly, if synchronous trading as driven by coordination, we should observe simultaneoustrades of predominantly the same stock [20, 21]. However,we find that 96% of our simultaneous trades are of differentstocks. Moreover, the trades are of different types: 60% of thesimultaneous trades involve both buying and selling activities. Individual performance.
Individual daily performance p ij canbe assessed by whether the trader i loses or makes money atthe end of the day j . Moreover, because the amount of moneymade by a trader at day’s end depends on various factors,such as market volatility, number of stocks traded, and size oftrades, a simple binary outcome variable appropriately stan-dardizes their performance by considering whether the traderlost ($ <
0) or made money ($ > p ij = 0and p ij = 1, respectively.We quantified the relationship between a trader’s syn-chronous trading s ij and performance p ij with a logistic re-gression of the form logit( p ij ) = β + β s ij (see Materials andMethods). We found that synchronous trading s ij was signif-icantly ( p < − ) and positively associated with a trader’sperformance (Fig. 3). Using the same logistic analysis, wecompared advanced and delayed trading with end-of-day per-formance p ij . The results indicate that both advanced anddelayed trading are statistically unrelated ( p > .
15) to end-of-day performance. This reveals that synchronous trading,though arising without apparent coordination, indicates auniquely beneficial time to trade that neither advanced nordelayed trading can reveal.
Instant messaging patterns.
An important proposed contribu-tion of our work is to identify those factors associated withthe level of collective human behavior. In biological systems,local communication channels have been identified as a cor-relate of the rise of synchronous behavior [1, 3, 6, 7, 32]. Fol-lowing this line of reasoning, we found distinctive associationsbetween traders’ instant messaging patterns and synchronoustrading. First, we found that instant messaging volume isassociated with trading volume throughout the day, suggest-ing a close connection between the two. Pooling all traders’IM and trade activity over our observation period, Figure 4Ashows that IMs have a significant correlation ( p < − ) withtrades over the day. On average, IM and trade activity riserapidly after the 9:30am opening bell, peak at 10am, declineat lunch time, uptick again from 1-3 pm, and finally declineprecipitously at the 4 pm closing bell.Second, research has shown that collective synchronous ac-tivity can arise when a coupling mechanism delays or pausesthe timing of individual activities [1–3, 8, 33, 34]. For exam-ple, cicadas have been found to have an internal clock thatstimulates chirping. This clock would induce a cicada to reg-ularly chirp whether or not it was exposed to the chirps ofother cicadas. Synchrnous chirping arises because the inter-nal and individual chirp activity is delayed by the chirp ofanother cicada, coupling the timing of the internal chirp withthe collective chirps of other cicadas [9]. To determine if in-stant messages can play a coupling role, we observed whetherthey were associated with the rise of synchronous trading. Asnoted above, instant messages play the important function oftransferring information that helps traders disambiguate mar-ket information, and, because a trader cannot trade and IMat the same time, instant messaging necessarily can delay hisor her trades. This would suggest that if we were to comparethe observed intensity of synchronous trading of a trader rel-ative to the synchronous trading expected by chance, as wedid above, but this time we randomized the trader’s tradesacross all 1-second time windows of the day except those 1- second time windows where there was an instant message,we would expect the intensity of synchronous trading to in-crease in the presence of non-random instant messages. Toquantify the IM-trade coupling, we compared the degree towhich synchronous trading changes when randomizing a sin-gle trader’s trades in any second in which an IM was not sentor received. Methodologically, we quantified the IM-trade cou-pling by θ ij = | s ij − ˆ s ij | , where ˆ s ij is the synchronous tradingcalculated by randomly shuffling over the seconds with no IMactivity.Consistent with our conjectures, we found a positive andsignificant association ( p < − using Markov randomiza-tions) between the IM-trade coupling θ ij and the synchronoustrading s ij of each trader (Fig. 4B, Materials and Methods).This reveals that the instant messaging patterns of traders isassociated to their trades such that the observed level of syn-chronous trading increases as the communication pattern isincreasingly different from what would be expected by chance.The more non-random the instant messaging pattern, thegreater the synchronous trading. This suggests that the localcommunication patterns of individuals have an important as-sociation with the rise of simultaneous activity, which in turnis associated to their performance. If one assumes that IMsare used to corroborate the meaning of the market through-out the day, then our findings suggest that the increasinglystructured communication strategically aims to help each in-dividual trader make a decision about when to trade. Discussion
Synchronicity is a pervasive and mysterious drive in nature [1].In animal, biological, and physical systems synchronicity re-duces uncertainty, as when school of anchovies evade preda-tion, neurons co-fire to process complex information, or per-turbations reduce noise in physical systems. Synchronicityalso apparently arises from local interactions without the aidof centralized leadership. This suggests that while more re-search needs to be done on synchronicity’s functional role incomplex human systems, it may furnish a functional alterna-tive to leadership in rich informational environments.We examined the association between synchronicity andperformance in a complex system where performance increaseswith uncertainty reduction. Examining a typical proprietarytrading firm wherein the traders individually race to be thefirst to disambiguate a constant stream of uncertain marketinformation in an effort to make profitable trades, we foundthat when a stock trader in that firm trades at the same timeas other traders in the firm, his or her financial performance issignificantly increased. We also found a coupling mechanism;we found that traders’ instant message communication pat-terns are positively associated with the rise of sync. Buildingon synchronicity principles found in other complex systems,we speculate that the mechanisms underlying these empir-ical patterns involve rapid information aggregation throughinstant messaging networks. Because separate traders in thefirm have different instant message contacts in the market,each trader samples, separate local inferences about the even-tual meaning of market information. When these diversepoints of view converge, the traders trade in synchronicitysuch that the synchronous timing of trades reflects a pointof crowd wisdom despite no conscious intention to do so onthe part of any individual trader. These mechanisms suggestthat synchronicity in human systems reflect some of the sameprinciples found in animal systems, namely that synchronic-ity appears to arise with attention to local information ratherthan centralized leadership. In the human system we exam-ine, and in human systems where quick response behavior is
Footline Author PNAS
Issue Date
Volume Issue Number ikely a mix of being reactive and thoughtful about the infor-mation presented, we also purport that the rapid aggregationof local information from diverse points of view plays a rolein the performance benefits of the synchronicity we observe.If one assumes that each of the traders have their own exper-tise, training, and assumptions that go into deriving inferencesfrom the market information they process, it suggests that ac-tions consistent with the corroboration of diverse viewpointsare likely to be a better approximation of the true meaning ofinformation than singular or myopic points of view.The powerful information processing capacities of humansin complex systems may furnish unique opportunities to applythe ideas developed here about human synchronicity to othercontexts. We would speculate that in many increasingly richinformation environments there are benefits to understandingsynchronicity. For example, currently in the domain of in-telligence and national security, many security officers face afrenetic pace of information not unlike the traders we stud-ied. They too receive constant feeds of information—videos,text, voice, blogs, RSS newsfeeds, and tweets—and are inconstant communication with their own instant message net-work throughout the day. Moreover, like traders they arealso racing to disambiguate news. Disambiguating informa-tion quickly means a potential pre-empt of an attack whereasadvanced or delayed disambiguation can mean “jumping thegun” or waiting too long respectively [30]. Disease controlagencies around the world all monitor large amounts of timesensitive data in attempts to identify possible outbreaks. Inboth situations, and more generally, in situations where infor-mation overloads might overwhelm individual decision mak-ing ability and information is time sensitive, creating systemsthat can capture moments of synchronicity may help identifywhether an action is functional or not.Using observational data to describe this phenomenon pro-vided a rich mix of real data but we are unable to completelytest these mechanisms. Future work might devise experi-ments, perhaps in one of the mock trading labs now in ex-istence in universities, by manipulating the content and rateof change of market news, providing access vs. no access toinstant message communication networks, and changing theinformation sampled from the instant message networks frommyopic to diverse. Future research might also begin to ex-amine the potential dysfunctions of synchronicity in humancomplex systems. Under what conditions does collective ge-nius turn into mob madness? Another direction for future research is to explore the differences between synchronicityand other collective behavior mechanisms. Materials and Methods
Data.
We observed all the 66 day traders at an anonymous trading firm from9/26/2007-02/20/2009. Day traders keep short-term positions and do not hold in-ventories of stocks; they enter and exit positions each day, normally between 9.30am-4:00pm. Our traders are “point-and-clickers”– they make trades in real-time in 98%of the time (the 1.2% of the trades done algorithmically were omitted and did notaffect the results). 40-70% of the trading on NYSE is point-and-click. We observedthese traders trading approximately 4500 different stocks over various exchanges,which suggests that they sample a large part of the market. As in most trading firms,traders do not trade everyday of every week for various reasons. Similarly, in thisfirm, no more than 22 traders were at their desks on any one day We analyzed all the > million intraday stock trades of these day traders and their > million instantmessages exchanged across their networks. The performance data was calculated bythe firm using standard industry metrics. Traders cannot trade via IMs. Logistic regression.
To check the robustness of the association between syn-chronous trading and individual performance to other potential influences such asnumber of trades and daily uncertainty, we performed the same analysis with termsfor number of trades, an interaction term for number of trades and s ij , and wealso controlled for market volatility (i.e. VIX). The extended model has the form logit( p ij ) = α + βs ij + γk ij + δs ij k ij + ǫv j . Additionally, to controlfor unobservable factors of each particular trader, we used fixed effects (dummy vari-ables for each trader) in the logistic regression. Under all circumstances, synchronoustrading was significantly and positively associated with individual performance. Association between synchronous trading and IM-trade coupling.
To check therobustness of the association between IM-trade coupling and synchronous trading toother potential influences such as number of IMs and daily uncertainty, we performed alinear regression with terms for number of IMs, an interaction term for number of IMsand θ ij , and we also controlled for market volatility (i.e. VIX). The extended modelhas the form s ij = α + βs ij + γθ ij + δs ij θ ij + ǫv j . Additionally, to controlfor unobservable factors of each particular trader, we used fixed effects (dummy vari-ables for each trader) in the regression. Under all circumstances, IM-trade couplingwas significantly and positively associated with synchronous trading. ACKNOWLEDGMENTS.
We would like to thank Alex Arenas, Jordi Bascompte,Jordi Duch, William Kath, Tae-Hyun Kim, Eduardo L´opez, Hani Mahmassani, DeanMalmgren, Alejandro Morales Gallardo, Mason Porter, Mark Rivera, Daniel Stouf-fer, Felix Reed-Tsochas, Marta Sales-Pardo, Uri Wilensky, and the members of theNICO weekly seminar series for useful discussions that led to the improvement of thiswork. We also thank the Kellogg School of Management, Northwestern University, theNorthwestern Institute on Complex Systems (NICO) for financial support. Researchwas also sponsored by the Army Research Laboratory and was accomplished underCooperative Agreement Number W911NF-09-2-0053. The views and conclusions con-tained in this document are those of the authors and should not be interpreted asrepresenting the official policies, either expressed or implied, of the Army ResearchLaboratory or the U.S. Government. The U.S. Government is authorized to repro-duce and distribute reprints for Government purposes notwithstanding any copyrightnotation here on.
1. Strogatz S-H (2003) SYNC: The Emerging Science of Spontaneous Order, (Hyperion).2. Vicente R, Gollo L-L, Mirasso C-R, Fischer I, Pipa G (2008) Dynamical relaying canyield zero time lag neuronal synchrony despite long conduction delays. Proc. Natl.Acad. Sci. USA 105: 17157-17162.3. Arenas A, D´ıaz-Guilera A, Kurths J, Moreno Y, Zhou C (2008) Synchronization incomplex networks. Physics Reports 469: 93-153.4. Schneidaman E, Berry M-J, Segev R, Bialek W (2006) Weak pairwise correlations im-ply strongly correlated network states in a neural population Nature 440: 1007-1012.5. Sumpter D-J-T (2006) The principles of collective animal behaviour. Philos. Trans. R.Soc. Lon. B Biol. Sci. 361: 5-22.6. Couzin I-D (2007) Collective minds. Nature 445: 715.7. Cole B-J (1991) Short-term activity cycles in ants: generation of periodicity throughworker interaction. Am. Nat. 137: 244-259.8. Greenfield M-D, Roizen I (1993) Katydid synchronous chorusing is an evolutionarystable outcome of female choice. Nature 364: 618-620.9. Ermentrout B (1991) An adaptive model for synchrony in the firefly Pteroptyx malac-cae. Journal of Mathematical Biology 29: 571-585.10. Cooley J-R, Marshall D-C (2001) Sexual signaling in periodical cicadas, Magicicadaspp. Behaviour 138: 827-855.11. Merton R-K (1961) Singletons and multiples in scientific discovery: A chapter in thesociology of science Proc. of the American Phil. Soc. 105: 470-486.12. Galton F (1907) Vox populi Nature 75: 450.13. N´eda Z, Ravasz E, Brechet Y, Vicsek T. Barab´asi A-L (2000) Self-organizing processes:The sound of many hands clapping. Nature 403: 849-850. 14. Hong L, Page S-E (2004) Groups of diverse problem solvers can outperform groups ofhigh-ability problem solvers. Proc. Natl. Acad. Sci. 101: 16385-16389.15. Eckmann J-P, Moses E, Sergi D (2004) Entropy of dialogues creates coherent struc-tures in e-mail trafiic. Proc. Natl. Acad. Sci. 101: 14333-14337.16. May R-M, Levin S-A, Sugihara G (2009) Complex Systems: Ecology for bankers Na-ture 451: 893-895.17. Whaley R-E (2000) The investor fear gauge J. Portfolio Management 26: 12-17.18. Moro E, Vicente J, Mayano L-G, Gerig A, Farmer J-D, Vaglica G, Lillo F, MantegnaR-N (2009) Market Impact and Trading Profile of Hidden Orders in Stock MarketsPhys. Rev. E 80: e066102.19. Tetlock P (2007) Giving content to investor sentiment: The role of media in the stockmarket The J. of Finance 62: 1139-1168.20. Bouchaud J-P, Farmer J-D, Lillo F (2009) How Markets Slowly Digest Changes inSupply and Demand, in: Handbook of Financial Markets: Dynamics and Evolution(North-Holland, Elsevier).21. Brock W, Lakonishok J, Lebaron B (1992) Simple technical trading rules and thestochastic properties of stock returns J. Finance 47: 17311764.22. Back K, Pedersen H (1998) Long-lived information and intraday patterns J. of Finan-cial Markets 1: 385-402.23. Antweiler W, Frank M-Z (2004) Is all that talk just noise? The information contentof internet stock message boards. The J. of Finance 59: 1259-1294.24. Hong H, Stein J-C (2003) Differences of opinion, short-sales constraints, and marketcrashes Rev. Financ. Stud. 16: 487-525.
5. Zhao Z, et al. (2010) Effect of social group dynamics on contagion Phys. Rev. E 81:056107.26. Leskovec J, Horvitz E (2008) Planetary-Scale Views on a Large Instant-MessagingNetwork. Proceedings of the International WWW 2008 Conference.27. Cetina K-K, Bruegger U (2002) Global Microstructures: The Virtual Societies of Fi-nancial Markets. American J. of Sociology 107: 905-950.28. Steven R, Galloway T, Berka C, Behneman A (2010) A neurophysiologic approach forstudying team cognition. Proceedings of the Interservice/industry training, Simulation,and Education Conference (I/ITSEC): 1-8.29. Lo C-C, Wang X-J (2006). Cortico-basal ganglia circuit mechanism for a decisionthreshold in reaction time tasks. Nature Neuroscience 9: 956-963. 30. Kiani R, Hanks T-D, Shadlen M-N (2006). When is enough enough? Nature Neuro-science 9: 861-863.31. Torney C, Neufeld Z, Couzin I-D (2009). Context-dependent interaction leads to emer-gent search behavior in social aggregates. Proc. Natl. Acad. Sci. 106: 22055-22060.32. Strogatz S-H, Abrams D-M, McRobie A, Eckhardt B, Ott E (2005) Theoretical me-chanics: Crowd synchrony on the Millennium Bridge. Nature 438: 43-44.33. Hunt D, Korniss G, Szymanski B-K (2010) Network Synchronization in a Noisy Envi-ronment with Time Delays: Fundamental Limits and Trade-Offs Phys. Rev. Lett. 105:068701.34. Hunt D, Korniss G, Szymanski B-K (2011) The impact of competing time delays incoupled stochastic systems Phys. Lett. A 375: 880-885.
Footline Author PNAS
Issue Date
Volume Issue Number Seconds of the day T r ade r Number of traders trading simultaneously with trader i
Fig. 1.
Calculating synchronous trading. The synchronous trading s ij of a trader i in day j (e.g., the trader whose trades are highlighted in blue) is defined as the degreeto which the observed number of other traders trading within the same seconds as trader i (top values) compares to the same value when randomizing just the trades of thatparticular trader. For advanced and delayed trading, we calculated the number of other traders trading one second late and one second early, respectively. −5 0 5 10 1500.10.20.30.40.50.60.7 Synchronous trading P r obab ili t y den s i t y
10 20 30 40 50 60 70 80 90−0.500.511.522.5
Daily uncertainty, VIX A v e r age sy n c h r onou s t r ad i ng A B
Fig. 2.
Synchronous trading and uncertainty. A shows the probability density for synchronous trading s ij (bottom blue bars), advanced trading s − ij (middle green bars)and delayed trading s +1 ij (top orange bars) for all traders across the observation period. The bar size is the sum of the 3 probability values, and colors correspond to therelative contribution each distribution makes to the total sum. We found that synchronous trading is significantly different to advanced and delayed trading ( p < − usingKolmogorov-Smirnov test). B shows the positive association ( p < − using Markov randomizations) between the average synchronous trading h s ij i and level of dailyuncertainty in the market, as given by the market’s standard volatility index (VIX) [17]. The dashed line depicts the relationship estimated via a linear regression. P r obab ili t y den s i t y −6 −4 −2 0 2 4 6 8 100.40.60.8 Synchronous trading P r obab ili t y o f m a k i ng m one y AB Fig. 3.
Individual performance. A depicts the probability density of synchronous trading for traders that make money (left green bars), and for those that do not (rightyellow bars). The two distributions are significantly different considering all values ( p = 0 . using Kolmogorov-Smirnov test), within -2 and 2 exclusively ( p = 0 . usingKolmogorov-Smirnov test) and outside -2 and 2 ( p = 0 . using Kolmogorov-Smirnov test). B shows the relationship between synchronous trading s ij and the probabilityof making money p ij . The curve depicts the probability of performance (making money) estimated via a logistic regression (Methods). For any trader under consideration,the probability of making money increases as the synchronous trading increases. The gray region corresponds to the confidence interval. Hour P r obab ili t y den s i t y IM−trade coupling S y n c h r onou s t r ad i ng A B
Fig. 4.
Association between IMs and trading. Panel A presents the probability density of observing any trade (black line) and IM (green dashed line) in each hour onaverage across the observation period. Approximately 95% of trades and IMs are done between 9.30am and 4pm, which correspond to the main operation hours of NYSE.Panel B shows the empirical relationship ( p < − using Markov randomizations) between the IM-trade coupling θ ij and synchronous trading s ij for all traders acrossthe observation period. The dashed line depicts the association estimated via a linear regression (Methods).Footline Author PNAS Issue Date