Co-location and the Comovement of Order Flow: Evidence from Firms that Switch Exchanges

Transcription

1 Co-locaton and the Comovement of Order Flow: Evdence from Frms that Swtch Exchanges Adtya Kaul Unversty of Alberta School of Busness Edmonton, AB, Canada T6G2R6 Tel emal: Vkas Mehrotra a Unversty of Alberta School of Busness Edmonton, AB, Canada T6G2R6 Tel emal: vmehrotr@ualberta.ca Carmen Stefanescu ESSEC Busness School Cergy-Pontose, France Tel emal: stefanescu@essec.edu a Correspondng author.

2 Co-locaton and the Comovement of Order Flow: Evdence from Frms that Swtch Exchanges ABSTRACT Employng a sample of frms that swtch ther exchange lstng from AQ to the, ths paper studes how tradng actvty for a stock s nfluenced by actvty for other stocks sharng ts exchange. Order flow (a measure of sgned tradng actvty) for the swtchng stock co-moves more strongly wth order flow followng the swtch, wth a correspondng weakenng of ts comovement wth AQ order flow. Further tests show that these results are not drven by fundamentals, frm characterstcs, ndexng, or a delayed response to cross-market nformaton. However, stocks wth more small (lkely nose) trades are more susceptble to the local exchange nfluence, whle stocks wth more medum/large (lkely nformed) trades are less so. These fndngs on the nfluence of common locaton on order flow comovement are most consstent wth a vew of comovement proposed by Barbers and Shlefer (2003) and Barbers, Shlefer and Wurgler (2005), where nvestors segregate assets nto ad hoc classes and place correlated trades wthn such classes.

3 1 Introducton In ths paper we examne the nfluence of a common tradng locaton, or co-locaton, on stock-level buy and sell trades (henceforth, order flow). 1 We do ths by studyng the set of stocks that swtch ther exchange lstng from AQ to the, and documentng the change n the comovement of order flow for these stocks wth aggregate order flow from ther old and new exchanges. 2 In dong so, we tackle the fundamental queston of what factors nfluence nvestor buy and sell decsons, the role of local order flow n ths process, and whether the mportance of co-locaton vares wth partcular frm characterstcs and nvestor types. The tradtonal explanaton for why an nvestor buys or sells a securty s that she ether has superor nformaton about value, or faces a lqudty shock (e.g. see Kyle, 1985; Admat and Pflederer, 1988). In ths framework, decsons to buy or sell should be correlated across stocks only to the extent that the nformaton or lqudty needs of nvestors are also correlated. For nstance, faced wth an mpendng recesson, nvestors may decde to convert part of ther stock holdngs nto cash for consumpton, generatng contemporaneous sell orders n many stocks. Smlarly, f nformaton has common components that affect many stocks at once, nvestors recevng ths nformaton wll place correlated trades. Ths nformaton could be publc (e.g. macro-economc) or possbly 1 Order flow captures the dfference between buyng and sellng transactons. Our measure of order flow s calculated as the dfference between buyng and sellng volume, scaled by total tradng volume, where buys and sells are dentfed usng the Lee-Ready (1991) algorthm. 2 Whle t s possble for frms to swtch from the to AQ, very few do so, prmarly because Rule 500(a) made voluntary delstngs dffcult (e.g. by requrng approval of two-thrds of a frm s shareholders). Rule 500 (a) was repealed n August 2003, though the number of frms movng from to AQ remans small. Kalay and Portnaguna (2001) analyze the case of Aeroflex, whch moved voluntarly from the to AQ n 2000.

4 prvate and correlated across stocks (e.g. Albuquerque, De Francsco and Marques, 2008; Albuquerque, Bauer and Schneder, 2009). Conversely, a role for co-locaton n generatng correlated order flow could arse f nvestors mtate others, ndependent of ther lqudty needs or prvate nformaton. Hong, Kubk and Sten (2005) show that fund managers decsons to buy or sell a stock are correlated wth the decsons of other managers n the same cty, consstent wth word-ofmouth effects n nformaton propagaton. Informaton cascade-type herdng models (e.g. Bkhchandan, Hrshlefer and Welch, 1992) suggest that under certan condtons, nvestors wll ratonally abandon mprecse prvate nformaton n favor of a smpler strategy of mtatng other nvestors trades. Barbers and Shlefer (2003) and Barbers, Shlefer and Wurgler (2005; henceforth BSW) submt that nvestors segregate assets nto classes and trade stocks n a class n the same drecton. BSW fnd supportng evdence based on shfts n return comovement for frms that are added to the S&P 500 Index. Green and Hwang (2009) provde addtonal support, usng a sample of stock splts to show that returns for frms n smlar prce ranges co-move strongly. A confoundng ssue wth lnkng order flow comovement to a common locaton s that order flow on a partcular exchange s apt to co-move for a varety of plausbly fundamental reasons. Examnng the level of order flow comovement on an exchange s therefore unlkely to solate the nfluence of co-locaton on nvestor buy and sell decsons. By contrast, snce frm fundamentals are unlkely to change abruptly, whereas the exchange swtch s a dscrete event, changes n comovement surroundng the swtch wll be nformatve about the mportance of co-locaton. That s, a sudden change n comovement Page 2

5 assocated wth the swtch would cast doubt on explanatons based on fundamentals. We examne order flow comovement for 536 frms that relocate from AQ to the between January 1988 and December We start by relatng order flow for each swtchng stock to aggregate AQ and order flows. Ths specfcaton s estmated at the ntraday, daly and weekly frequences, usng one year of tradng before and one year of tradng after the swtch. Our man fndng s that order flow for stocks that swtch to the co-moves more strongly wth aggregate order flow and less so wth AQ followng the swtch, and that these comovement changes are concdent wth the swtch. Ths result s present at every frequency ntraday, daly and weekly and s economcally sgnfcant. For nstance, at the daly level, the mean senstvty of a swtchng frm s order flow to aggregate order flow ncreases three-fold from 0.27 to 0.84, whle ts senstvty to AQ order flow drops from 0.64 to 0.18 n the mmedate aftermath of the exchange venue swtch. The fact that the order flow comovement shft remans large at the lower weekly frequency makes t unlkely that the shft s drven by dfferences n the market structure of the two exchanges or frctons n nformaton transmsson across the two exchanges. Further, the fact that the change n order flow comovement occurs mmedately after the swtch, rather than slowly over tme, supports the concluson that the comovement change s assocated wth the swtch, rather than (gradually changng) frm characterstcs. We nterpret our results as showng that co-locaton, n the form of sharng an exchange, strongly nfluences stock-level order flow. A natural queston s what factors contrbute to ths nfluence. We conduct cross-sectonal tests relatng the change n order Page 3

6 flow comovement to proxes for arbtrage costs, nformaton qualty, lqudty, nose and nformed tradng and nsttutonal ownershp, as well as to frm characterstcs. Somewhat surprsngly, we fnd that stocks that are more lqud and followed by more analysts dsplay larger shfts n comovement, that s, are more nfluenced by order flow on the followng the swtch. These results are nconsstent wth the vew that transactons costs or mpedments to the flow of nformaton are responsble for the nfluence of locaton on trades. Greater nsttutonal ownershp also appears to be assocated wth a larger co-locaton nfluence, consstent wth evdence that nsttutons are prone to herdng (for evdence of nsttutonal herdng, see Sas, 2004; Wermers, 1999; Prnsky and Wang, 2004). Most strkng, we fnd that stocks wth hgher volume n small-szed trades (of shares) are more strongly nfluenced by local order flow, whle stocks wth hgher volume n medum/large-szed trades (500 or more shares) are less so. Ths result s consstent wth behavoral models suggestng that nose traders (who are more lkely to trade n smaller lots) are an mportant source of co-locaton effects on order flow, whle the actons of nformed traders serve to reduce such nfluence. We carry out addtonal analyss to evaluate alternatve explanatons for our fndngs. Frst, gven that frms apply to swtch exchanges, the swtchng frms do not comprse a random sample, and our results are potentally subject to selecton bas. A second concern s that our analyss could be affected by specfc economy-wde trends, e.g. the ncrease n comovement wth order flow could smply be a reflecton of the growng nfluence of the over our sample perod. We deal wth both concerns by repeatng our tests for a Page 4

7 control sample of smlar (.e. ndustry, sze and prce-matched) frms that meet key lstng requrements but reman on AQ. 3 We fnd that the control sample of AQ-resdent frms shows no change n order flow comovement wth AQ or the. The contrastng comovement patterns for the swtchng and control frms ndcate that the results for the swtchng frms are unlkely to be drven by economy-wde trends n comovement. Further, the large dfferences n the magntudes of the comovement changes for the swtchng and control samples suggest that any selecton bas would have to be very large to neutralze our nferences. Second, an alternatve explanaton for our results s that when a stock moves from AQ to the, t swtches membershp between two exchange-level ndexes that are wdely followed by actual or closet ndexers. We conduct several tests to examne such ndexng effects. When we remove the 100 largest stocks by market captalzaton (most lkely to be ncluded n any exchange-level ndexng strategy) from the calculaton of aggregate and AQ order flows, the shft n order flow comovement s materally unchanged. Lkewse, when we drop the last 30 mnutes of daly tradng from our analyss, our results are unaltered. Ths s nconsstent wth ndexers, who have a preference for tradng at the end of the day, drvng our results. Fnally, usng returns (nstead of unavalable order flow) gong back to 1973, we fnd smlarly strong shfts n comovement over perods that pre-date the popularty of ndexng. These results do not support the explanaton that exchange-level ndexng s the man drver for the nfluence of co-locaton on order flow documented n our study. Instead, we beleve that such co-locaton nfluence 3 Keda and Panchaganeshan (2010) show that of the 2,654 AQ frms that meet mportant lstng crtera at the, only 460 actually swtch to the. Page 5

8 on order flow comovement s better descrbed as category preference, as n BSW, rather than a formal ndexng strategy. Thrd, we conduct tests to examne f cross-market nformaton delays can explan our results. When we ncorporate lags of order flow from both markets, the sum of the coeffcents on contemporaneous and lagged exchange order flows s smlar to the coeffcent n our basc test. These results rule out cross-market nformaton frctons as the source of our results. Whle extant work has not looked at the nfluence of co-locaton on order flow, several studes have examned nternatonal tradng venues and return comovement. For example, Chan, Hameed and Lau (2003) document a sgnfcant change n return comovement for the Jardne Group of frms after they delst from Hong Kong and re-lst n Sngapore. Froot and Dabora (1999) document greater return comovement for so-called Samese Twn stocks wth the natonal stock exchange they are lsted on, despte beng perfect substtutes on paper. Usng a sample of nternatonal mergers, Brealey, Cooper and Kaplans (2010) document that the post-merger change n target frm return comovement wth the acqurer frm s home market s large and not explaned by fundamentals. Such excess return comovement across nternatonal venues can arse due to barrers separatng foregn markets. By contrast, our focus on two domestc markets wth hgh lqudty and nformaton flow allows us to solate the effect of co-locaton on order flow Page 6

9 comovement. 4 Our study has several mplcatons. Frst, our results pont to non-ndependence of order flow and returns wthn an exchange. Ths calls nto queston the dversfcaton benefts of nvestng n a large basket of stocks f the basket n queston s tethered to a partcular locaton. Our study also relates to the lterature on fnancal fraglty, whch shows that correlated ownershp and tradng can have real consequences, notably hgher return volatlty. 5 We pont to a specfc channel a common tradng locaton that may engender such fraglty. 6 An unanswered queston n our study, and ndeed n the lterature, s why colocaton matters. In our specfc case, ths could be because locaton confers not just common nformaton cues or convenent asset categorzatons as n BSW, but plausbly common moods as well. For example, Hrshlefer and Shumway (2003) show that daly market returns are affected by mornng sunshne, a result that s consstent wth nvestor moods affectng market-wde order flow. Whether subscrpton to local cues s ratonal or s a behavoral artfact rooted n smple heurstcs s worthy of further examnaton. The rest of the paper s organzed as follows. Secton 2 descrbes the data. Secton 3 studes the change n order flow comovement followng the swtch to the. Secton 4 4 Goyal et al (2008) document dstnct prcng factors for and AQ stocks. However, they cauton aganst concludng that the two markets are therefore segmented. Indeed, the authors argue that the two common factors for these markets are suffcent to brng about ntegraton. 5 For models of fnancal fraglty, see, among others, Allen and Gale (2004) and Greenwood and Thesmar (2009). A common theme n these papers s the presence of correlated shocks that affect market lqudty. 6 In a recent workng paper, Anton and Polk (2010) show that the more connected a stock s (by vrtue of beng owned by the same nsttutons), the more lkely ts return s to co-move wth the returns for other connected stocks. Page 7

10 studes determnants of the changes n order flow comovement. Secton 5 evaluates alternatve explanatons for these results. Secton 6 concludes. 2 Sample and data Usng the Center of Research n Securty Prces (CRSP) fles, we select all frms that swtch ther lstng venue from AQ to the between January 1988 and December We examne ordnary common shares and exclude non U.S. frms, real estate nvestment trusts, and closed-end funds. Frms are requred to have CRSP data and ntraday trade and quote data from ISSM or TAQ for one year before and one year after the swtchng date, as well as book equty data from COMPUSTAT. Ths leaves us wth a sample of 536 frms. We also use a larger sample of 821 AQ frms that swtch to the between January 1973 and December 2000 to study return comovement n a robustness check. Our purpose s to solate the change n order flow comovement assocated wth the move to the, and we attempt to account for exogenous shfts n comovement. For nstance, t s possble for the order flows of all AQ stocks to dsplay greater comovement wth aggregate order flow f the markets have become more ntegrated over our sample perod or f the captures economc prospects more closely than does AQ. Alternatvely, comovement mght ncrease for stocks that resemble stocks, stocks of a certan sze, or stocks from partcular ndustres. To account for these possbltes, we create a control sample of stocks matched to the swtchng frms by sze (market equty), share prce, ndustry (two-dgt SIC code) and date, but that reman on 7 We select our sample of frms by examnng the change n the CRSP exchange code, and dentfy as the event date the frst day on whch the frm appears wth the exchange code. Page 8

11 AQ. 8 The control stocks meet the followng key quanttatve lstng requrements for the : net tangble assets exceedng $40 mllon; market value of equty exceedng $18 mllon before 1995, and $40 mllon after 1995; pre-tax ncome exceedng $2.5 mllon n the precedng fscal year, and $2.0 mllon n the year before that; and shares outstandng exceedng 1.1 mllon. We repeat our tests for the control sample, and test for dfferences between the swtchng and control sample effects. Snce the control frms satsfy key lstng requrements and are smlar to the swtchng frms along mportant dmensons, any dfference n the order flow behavor of the two samples s more lkely to be assocated wth the swtch. Summary statstcs for the swtchng and control samples are presented n Table 1. The fnancal varables are measured as of the fscal year pror to the move to the, whle spreads and share turnover are averages computed over the year pror to the swtch (the month of and the month pror to the swtch are excluded). Insttutonal ownershp s measured as the percentage of outstandng shares held by nsttutons, and together wth the number of analysts, s measured at the end of the quarter pror to the swtch. We obtan prce and volume data from CRSP, balance sheet tems from COMPUSTAT, nsttutonal ownershp from Thomson Fnancal and analyst data from I/B/E/S. Book assets, market equty and nsttutonal ownershp are hgher for the swtchng frms relatve to the control frms, but not apprecably so. For nstance, total assets for the swtchng and control frms average $1.78 bllon and $1.23 bllon, whle average market 8 The matchng procedure follows Huang and Stoll (1996). We match on characterstcs as of the swtchng date and then ensure that the control frm we choose has one year of data before and after the swtchng date. We pck a unque control frm for each swtchng frm. Page 9

12 equty for the swtchng and control frms s $756 mllon and $580 mllon. The correspondng medan values are closer for the two groups. Mean nsttutonal ownershp s 46% for swtchng frms, and 38% for control frms. The mean and medan values of market-to-book are comparable for the swtchng and control frms, suggestng that the samples are smlar n terms of nvestment polcy and style. Daly turnover, a measure of tradng actvty, s also comparable for the two samples (mean and medan turnover are 0.335% and 0.243% for the swtchng sample, and 0.317% and 0.223% for the control sample). 9 Transacton costs are smlar, wth medan relatve spreads of 0.02 for both samples. Analyst coverage s agan close, wth mean and medan values of 7 and 5 for the swtchng frms, and 6 and 5 for the control frms. Overall, Table 1 shows that the test and control samples are alke along most frm and nvestor characterstcs. 3 Order flow comovement We examne the changes n the comovement of order flow for each swtchng and control frm wth aggregate AQ and order flows. The reported results are based on order flow defned n terms of the volume of buys and sells. 10 Trades are classfed as buys and sells usng ntraday data and the Lee-Ready (1991) algorthm, whch uses the quote closest to, but at least fve seconds before, each trade to classfy the trade. A transacton occurrng above (below) the quote mdpont s regarded as a buy (sell). If a 9 The means translate nto annual share turnover of approxmately 80%. 10 Two other measures of order flow are based on the number of trades and the dollar value of trades. We repeat our analyss usng these measures and arrve at dentcal conclusons. Page 10

13 transacton occurs at the quote mdpont, t s sgned usng the last non-zero transacton prce change, as a buy f ths prce change s postve and a sell f t s negatve. Before applyng ths algorthm, we exclude trades wth negatve prces, trades reported out of sequence, trades wth specal settlement condtons, trades and quotes recorded before the open or after the close, and quotes that mply a negatve spread. For each stock, we calculate ntraday order flow as the dfference between the volume of buys and the volume of sells n 15-mnute ntervals through the day,.e. 9:30-9:45 a.m., 9:45-10:00 a.m.,..3:45-4:00 p.m. We also cumulate order flow at the daly and weekly horzons. We standardze order flow for each nterval (15-mnute, daly or weekly) by total volume over that nterval to make t comparable across stocks and through tme. Thus, we study the fractonal order flow at the ntraday, daly and weekly frequency. Aggregate or AQ flow s computed as the equally-weghted (.e. smple) average of the order flows for all ordnary shares tradng n each market. In complng these averages, we exclude the order flow both of the swtchng frm and of ts ndustry. 11 Ths reduces the lkelhood of our fndng a mechancal assocaton between order flow for the swtchng stock and and AQ order flows. For nstance, f a swtchng frm s ndustry has greater representaton on the, comovement wth order flow may be accentuated by common news about ndustry cash flows or rsks. Droppng the swtchng frm s ndustry from the exchange-level order flow calculaton elmnates ths source of comovement, and solates the effect of co-locaton on order flow comovement. 11 The ndustry adjustment s accomplshed by defnng two-dgt ndustres for all stocks, followng Lewellen (2002). Page 11

14 We estmate the followng specfcaton for order flow (OF) for each swtchng stock, and repeat the exercse for the sample of control stocks. OF, t 0 1D OFt OFt OFt D OFt D, t (1) In (1), OF, t s order flow for stock n perod t, OF t and OF t are aggregate order flows for the and AQ, and D s a dummy that s one after the swtch date for stock and zero otherwse. The model s estmated usng order flow data from one year before to one year after the actual date of the swtch. We exclude the tradng months mmedately before and after the swtchng date to remove any effects related to the actual swtch. In ths specfcaton, and measure the base, pre-swtch levels of comovement wth and AQ order flows. Smlarly, and, the coeffcents on the nteracton terms, measure the change n comovement wth and AQ order flows followng the swtch. Table 2, Panel A through Panel C, presents the results for order flow measured over ntraday (15-mnute), daly and weekly wndows. In each panel, we nclude the mean coeffcents for the swtchng and control samples from (1), the t-statstc (n talcs) from a test of the null hypothess that the mean coeffcent s zero, and the p-value from a t-test of the null hypothess that the swtchng and control sample mean coeffcents are equal. Lookng frst at the level of comovement pror to the swtch, we fnd that the mean value of, the coeffcent on aggregate AQ order flow, s 0.82 at the 15-mnute frequency, Page 12

15 0.64 at the daly frequency, and 0.47 at the weekly frequency. 12 The pre-swtch coeffcent on aggregate order flow,, s smallest at the ntraday frequency (0.10), and ncreases at the daly and weekly frequences (0.27 and 0.40). The patterns n the slopes as the measurement nterval lengthens (declnng for AQ order flow and ncreasng for order flow) are consstent wth the noton that cross-exchange nformaton s more dffcult to access at hgher frequences. Ths could create a tendency for stock order flow to co-move more strongly wth local aggregate order flow at hgher frequences. The patterns are also consstent wth behavoral models of comovement (e.g. Barbers and Shlefer, 2003) where correlated trades wthn certan baskets arse due to nose trader sentment, and countervalng trades only arrve wth a lag. Our man nterest les n the changes n comovement, reflected n and. At each frequency, we fnd that the mean value of s sgnfcantly above zero and that of sgnfcantly below zero. The post-swtch comovement wth and AQ order flows s gven by and respectvely. Over 15-mnute measurement ntervals, the mean slope on order flow ncreases from 0.10 before the swtch to 0.53 after the swtch ( =0.43), whle the slope on AQ order flow drops from 0.82 to 0.44 ( = 0.38). Wth daly order flow, the mean slope ncreases from 0.27 to 0.84 whle the mean AQ slope declnes from 0.64 to The mean slope n the weekly order flow regresson rses from 0.40 to 0.80 and the AQ slope drops from 12 All of these coeffcents are sgnfcantly dfferent from zero. To save space, we wll not always dscuss statstcal sgnfcance. Unless we explctly menton t, the coeffcents are sgnfcant. Page 13

16 0.47 to Thus, there s a strong ncrease n the comovement of swtchng stock order flow wth aggregate order flow, and an equally strong decouplng wth AQ order flow. Moreover, the change n comovement s evdent at every frequency we examne, wth no tendency for the comovement n order flow to weaken over longer measurement wndows. Table 2 also shows that the mean values of and for the control stocks that reman lsted on AQ are not generally sgnfcant, and ther magntudes are usually no more than one-tenth those for the swtchng stocks. The one excepton s at the weekly horzon; even here, the mean coeffcent s less than half as large as the swtchng stock mean. The small values of and for non-swtchng but otherwse smlar control stocks suggest that the changes n order flow comovement for the swtchng stocks are drven not by economy-wde trends but by the shft to the. Note that the mean values of and are generally smlar for the control and swtchng stocks, ndcatng that the samples are well-matched n terms of pre-swtch comovement. The explanaton that co-locaton nfluences tradng becomes more compellng f the changes n comovement occur mmedately after the swtch. Fgure 1 plots the mean levels of ntraday order flow comovement for the swtchng and control stocks by month, startng one year before and endng one year after the swtchng date. We defne monthly ntervals n event tme, wth 22 tradng days consttutng each event month. As before, we exclude the event months mmedately before and after the swtch date, so month +1 s the month 13 We also examne the medans nstead of the means n all our tests. These yeld dentcal conclusons. Page 14

17 startng 22 days after the swtch date, and month -1 s the month endng 22 days before the swtch date. Wth 15-mnute order flow, we have 26 observatons per day or approxmately 570 observatons per month, and ths allows us to estmate the monthly slope coeffcents wth precson. For the swtchng stocks, Fgure 1 shows that the mean coeffcent on ntraday order flow jumps from 0.11 to 0.42 between month -1 and month +1. Smultaneously, the mean coeffcent on AQ order flow drops from 0.79 to Outsde of ths wndow, the slope coeffcents are relatvely flat both before and after the swtch. Unreported t-tests show that the mean slopes n month +1 are sgnfcantly dfferent from those n month -1. However, the slopes n month +2 through month +12 are not sgnfcantly dfferent from those n month +1, nor are the slopes n months -12 through -2 dfferent from those n month -1. Thus, the change n order flow comovement appears to occur mmedately after the swtch, rather than gradually over the months leadng up to or followng the swtch. Ths supports the concluson that the order flow comovement changes n Table 2 are assocated wth the swtch to the, rather than reflectng slow movng trends, e.g. changng asset characterstcs of the swtchng frms. Also shown n Fgure 1, the slope coeffcents are flat and never sgnfcant for the control stocks. In summary, ths secton shows that order flow for a stock lsted on AQ comoves strongly wth aggregate AQ order flow and relatvely weakly wth aggregate order flow. After the stock swtches to the, the nfluence of aggregate AQ order flow wanes whle that of aggregate order flow ncreases, and these changes are economcally and statstcally sgnfcant. The results show that order flow for a stock s Page 15

18 strongly nfluenced by local trades, but do not tell us why ths occurs, or what factors affect the mportance of local order flow. We turn to these questons n the next secton. 4 The determnants of order flow comovement The nfluence of exchange-level order flow could arse for a varety of reasons, ncludng costly or mperfect nformaton, nvestor nattenton, nsttutonal ownershp, nose tradng and arbtrage costs. We start by summarzng these canddate explanatons and descrbng the varables that we use to evaluate them. 1. Imperfect/costly nformaton. Ratonal herdng models (e.g. the nformaton cascade models of Bkchandan et al., 1992) suggest that comovement can arse due to mprecse prvate nformaton. For nstance, an nvestor observng everyone around her sellng stocks may ratonally abandon her prvate nformaton and sell shares. Under ths explanaton, aggregate exchange-level order flow could be more mportant when dosyncratc nformaton s scarce or less precse. Other models of ratonal excess comovement such as the costly nformaton model of Veldkamp (2006) make smlar predctons. We use analyst coverage, NumAnalysts (measured one quarter before the swtch) as a proxy for the qualty of frm-level nformaton. 14 Under the mperfect/costly nformaton explanaton, we expect that greater analyst coverage wll be assocated wth weaker local nfluence. Thus, frms wth more analysts should see smaller shfts n order flow comovement followng the swtch to the. Frm sze (Sze) may also be a proxy for 14 Lang and Lundholm (1996) and Healy, Hutton, and Palepu (1999) fnd that greater analyst followng s assocated wth ncreased dsclosure though both studes cauton aganst drawng causal nferences. Page 16

19 nformaton qualty, wth publc nformaton beng more scarce for smaller frms. Consstent wth ths argument, Lakonshok, Shlefer and Vshny (1992) and Wermers (1999) fnd that small stocks see more herdng on the part of money managers. Hence, we expect order flow for small stocks to show sharper changes n comovement after the swtch. 2. Investor nattenton. It s possble that nvestors do not become aware of lower vsblty stocks untl the stocks move to ther preferred exchange. For nstance, Kadlec and McConnell (1994) report an ncrease n the number of regstered shareholders for stocks after they lst on the. Under the nvestor nattenton explanaton, traders are more lkely to be aware of AQ stocks that are vsble, and more vsble AQ stocks wll see smaller ncreases n comovement wth order flow. Barber and Odean (2008) suggest that two proxes for nvestor awareness are extreme past returns and hgh volume, and we use these varables and adapt ther classfcaton scheme. The former s defned va a dummy varable, BgRet, that s one f a frm s cumulatve return over the year pror to the swtch s above the 90 th or below the 10 th percentle n the cross-sectonal dstrbuton of swtchng frm returns, and zero otherwse. The latter s defned usng a dummy varable, BgVol, that equals one f average daly market-adjusted turnover over the year pror to the swtch s above the 90 th percentle n the cross-sectonal dstrbuton of swtchng frm turnover, and zero otherwse. Under the nvestor nattenton explanaton, we expect smaller changes n order flow comovement for stocks wth extreme returns and turnover snce these stocks are lkely to be more vsble pror to the swtch. 3. Insttutonal ownershp. Nagel (2005) shows that several return anomales are less Page 17

20 marked for hgh nsttutonal ownershp frms. To the extent that nsttutonal ownershp eases market frctons, the nfluence of exchange order flow should be smaller for stocks wth larger nsttutonal ownershp. A contrary effect s suggested by the fact that nsttutons may themselves contrbute to herdng (e.g. see Wermers, 1999; Prnsky and Wang, 2004). If nsttutonal ownershp s correlated wth herdng, t should also correlate wth changes n order flow comovement. We measure nsttutonal ownershp, Inst, as of the quarter pror to the swtch. The sgn of the coeffcent on Inst s an emprcal ssue. 4. Nose tradng. Barbers and Shlefer (2003) suggest that correlated nose tradng gves rse to excess comovement n returns. Ther tradng based explanaton suggests that order flow for stocks wth more nose tradng should dsplay a stronger local nfluence, whle order flow for stocks wth more nformed tradng should show a weaker local nfluence. We dentfy nose and nformed traders on the bass of trade sze. Followng Barclay and Warner (1993), we compute the mean values of daly share turnover comng from small trades ( shares), Turn S, and from medum/large trades (500 or more shares), Turn B, as proxes for the presence of nose and nformed traders. We use turnover to make the measures comparable across stocks wth dfferent levels of tradng actvty. If nose tradng contrbutes to exchange-level comovement, stocks wth hgher values of Turn S wll see larger changes n order flow comovement followng the swtch. Smlarly, f nformed tradng reduces local nfluence, we should see smaller changes n order flow comovement for stocks wth hgher values of Turn B. Turn S and Turn B are measured over the year pror to the swtch. 5. Transacton costs. As transacton costs ncrease, arbtrage trades wll tend to declne, and Page 18

21 we expect the nfluence of local order flow to ncrease. We use the relatve bd-ask spread (RelSpr) as a proxy for tradng costs, and calculate ts mean over the year pror to the swtch. Sze (market captalzaton) s lkely to be nversely related to transacton costs. We expect to see larger changes n comovement for stocks wth hgher transacton costs, so the coeffcent on RelSpr should be postve, and that on Sze negatve. 6. Other controls. We nclude a dvdend dummy (Dv) and market-to-book (MB) as style proxes. Dv s one f the frm pays a dvdend n the year pror to the swtch and zero otherwse, whle MB s calculated usng the book and market equty values from the last fscal year-end before the swtch. As a control for the level of pror comovement, we nclude the stock s pre-swtch comovement coeffcent for the market concerned. Fnally, we nclude year dummes. We evaluate these explanatons by estmatng the followng cross-sectonal regressons (2a) and (2b), where summarzed n Table 2) are the dependent varables: and (estmates from equaton (1), dstrbutons b b BgRet 6 0 b1 b BgVol 7 b Turn 2 S b Inst 8 b Turn 3 B b Sze 9 b RelSpr 4 b 10 MB b NumAnalysts b 5 11 Dv YearDum e (2a) c 6 c BgRet 0 c1 c 7 BgVol c Turn 2 S c Inst 8 c Turn 3 B c Sze 9 c 4 c RelSpr 10 MB c c 5 11 NumAnalysts Dv YearDum e (2b) The results are reported n Panel A of Table 3. The frst column has the coeffcent estmates from (2a), whch models the determnants of. The coeffcent on β s negatve and sgnfcant; thus, stocks whose order flows co-move more strongly wth Page 19

22 order flow before the swtch experence smaller ncreases n comovement wth order flow. The coeffcent on RelSpr s also negatve and sgnfcant, mplyng that order flow for more lqud (lower spread) stocks shows larger changes n comovement wth order flow followng the swtch. Under the transacton cost explanaton, order flow for stocks wth hgher relatve spreads whch are more expensve to trade ought to be more strongly nfluenced by local order flow. Thus, ths result s nconsstent wth the tradng cost explanaton for exchange effects on order flow. The coeffcent on analyst coverage s postve and sgnfcant. That s, order flow for frms wth more analysts tends to be more strongly nfluenced by aggregate order flow. The mperfect /costly nformaton explanaton suggests that the post-swtch change n the nfluence of order flow should be smaller for frms wth more analysts. Hence, ths result s at odds wth the ratonal herdng explanaton based on mperfect nformaton. The coeffcents on the two proxes for nvestor attenton, the BgRet and BgVol dummy varables, are nsgnfcantly dfferent from zero. Thus, there s no evdence of smaller changes n order flow comovement for more vsble stocks. Ths result s nconsstent wth the hypothess that nvestor nattenton nduces local effects n order flow comovement. The coeffcent on Inst s postve and sgnfcant, meanng that order flow for frms wth greater nsttutonal ownershp s more subject to local nfluence. Ths result s consstent wth evdence that nsttutonal nvestors tend to herd (e.g. Wermers, 1999) and contrbute to comovement n lqudty and returns (see Kamara et al., 2008, for lqudty; Prnsky and Wang, 2004, for returns). Page 20

23 The coeffcent on Sze s nsgnfcantly dfferent from zero. Snce large frms have hgh ndex weghts, the lack of sgnfcance of frm sze s nconsstent wth an ndexng explanaton for order flow comovement. We explore ndexng effects n more detal n the next secton. Ths result s also nconsstent wth the mperfect/costly nformaton explanaton, snce large frms should have hgher qualty nformaton. Our result on frm sze s dfferent from that n Wermers (1999), who fnds greater herdng n small stocks. The dfference n our results mght stem from the fact that Wermers (1999) studes the level of comovement n portfolo allocatons whle we examne the change n order flow comovement assocated wth a specfc event. The coeffcents on small share turnover and medum/large share turnover address the nose trader explanaton for comovement. We fnd that the coeffcent on Turn S s sgnfcantly postve whle that on Turn B s sgnfcantly negatve. That s, stocks wth hgher small share turnover show larger changes n comovement wth order flow, whle stocks wth hgher medum/large share turnover show smaller changes n comovement followng the swtch. Ths fndng s consstent wth the nose trader explanaton, where the presence of small, lkely unnformed, nvestors leads to greater comovement wth exchange order flow, and the presence of nformed nvestors, who trade n larger lots, reduces the nfluence of locaton. The coeffcents on MB and Dv are sgnfcant, negatve for MB and postve for Dv. The negatve coeffcent on MB mples that order flow for value (lower MB) frms co-moves more strongly wth order flow after the swtch. Smlarly, the postve coeffcent on Dv means that the post-swtch order flow for frms that pay dvdends co-moves more Page 21

24 strongly wth order flow. To the extent that the conssts of lower MB stocks and dvdend-payng stocks relatve to AQ, these results show that style effects strengthen once a stock shares a tradng locaton wth ts style cohort, wth the exchange swtch beng a catalyst that promotes such comovement. The second column n Panel A of Table 3 reports coeffcents from (2b), whch models the determnants of the change n comovement wth AQ order flow,. We see that the coeffcent on the relatve spread s not sgnfcant, nor are the coeffcents on frm sze or the dummes for hgh turnover and large returns. The sgns of the coeffcents on the remanng varables are the opposte of those n the model for : negatve for analyst followng, nsttutonal ownershp, Turn S, and Dv; and postve for Turn B and MB. These coeffcents ndcate that stocks that decouple to a greater extent from AQ after the swtch: () are value rather than growth (have lower MB and are dvdend-payng); () have a more transparent nformaton envronment (greater analyst followng); () have greater nsttutonal ownershp; and (v) are characterzed by more nose trades and fewer nformed trades. In Panel B of Table 3, we explore the possblty that nvestors are more lkely to be nfluenced by local tradng when market condtons devate sgnfcantly from fundamentals. Ths possblty s suggested by Chrste and Huang (1995), who argue that stock returns are apt to be less dspersed durng perods of extreme market movements. The latter part of our sample perod ( ) concdes wth what we now know was the tech bubble. If nvestors are more lkely to dsregard ther nformaton and trade wth local order flow durng bubbles, we expect the changes n comovement wth aggregate AQ and Page 22

25 order flows to be larger durng ths perod. The specfcaton reported n Panel B dffers from that n Panel A n that we replace the year dummes wth a sngle dummy varable, Bubble, that s one for swtches occurrng between 1996 and 2000 and zero otherwse. The coeffcent on Bubble s postve and sgnfcant n the specfcaton for and negatve and sgnfcant n the specfcaton for. Ths s consstent wth the noton that durng bubbles, locaton has a more pronounced nfluence on stock order flow. The coeffcents on the remanng varables n Panel B are smlar to the values n Panel A n terms of sze and sgnfcance. We examne whether the sgnfcance of Bubble has other orgns. For nstance, snce the tech bubble occurs at the end of our sample perod, the sgnfcance of Bubble could reflect a secular ncrease n the nfluence of local order flow. We cannot nclude both year dummes for and Bubble due to collnearty, but we can re-ntroduce the year dummes for When we do ths, we fnd that the coeffcent on Bubble remans hghly sgnfcant n the specfcaton (t= 2.6), and s margnally sgnfcant n the specfcaton (t= -1.80). We regard ths analyss as exploratory, but t supports the dea that locaton has a stronger nfluence on stock order flow durng extreme market condtons. The cross-sectonal analyss casts doubt on the order flow comovement shfts beng assocated wth transacton costs, nvestor nattenton, or mperfect nformaton. Rather, the results n Table 3 provde support for a strong role played by nose traders, as well as by nsttutonal nvestors, n generatng local comovement n order flow. Page 23

26 5 Alternatve explanatons and robustness checks Our results pont to a strong nfluence of locaton on stock-level order flow that s dffcult to reconcle wth transacton costs or nformaton frctons, but appears consstent wth behavoral explanatons. In ths secton, we conduct robustness checks to rule out alternatve explanatons. Frst, the nfluence of exchange order flow could be an artfact of de facto and AQ ndexng strateges followed by nsttutons. A second possblty s that the nfluence of local order flow s due to cross-market nformaton affectng trades wth a lag. Fnally, t s possble that systematc nfluences such as market returns drve both stock and exchange-level order flows. If these explanatons are not supported, the nference that locaton engenders ncremental order flow comovement s strengthened. 5.1 Indexng effects It s possble that when a stock moves from AQ to the, t swtches membershp between two exchange-level ndexes that are wdely followed by actual or closet ndexers. If these ndexers experence correlated nflows and outflows, a stock s order flow wll decouple wth AQ order flow and co-move wth order flow after t moves to the. Ths argument s no dfferent from that artculated n BSW s study of ndex addtons. We address ths possblty by carryng out several tests. Our frst test examnes comovement over a perod that predates the popularty of ndexng. The sample conssts of 285 frms that swtch to the over the perod, whch commences wth the formaton of AQ and ends the year before our Page 24

27 ntraday sample perod begns. Snce order flow and volume data are not avalable for both exchanges over ths perod, we use returns to gan nferences about order flow comovement. Ths approach s justfed by mcrostructure work (e.g. Hasbrouck, 1991; Chorda and Subrahmanyam, 2004) that shows a postve correspondence between prce changes and order flow. Thus, we estmate the followng return comovement regresson: R, t 0 1D Rt Rt Rt D Rt D, t (3). Here, R, t s the return n perod t for stock, Rt and R t are the perod t equallyweghted and AQ returns, and D s, as before, a dummy varable that s one after the swtch date for stock and zero otherwse. We estmate (3) usng return data from one year before to one year after the swtch, agan excludng the tradng months adjacent to the swtch. The and AQ market returns are calculated usng all ordnary common shares from the two markets. As wth order flow, we exclude the return for the swtchng (or control) frm and all frms belongng to ts ndustry from aggregate and AQ returns. Table 4 presents the mean coeffcent estmates, t-tests of the null hypothess that the mean s zero, and t-tests comparng the swtchng and control sample means. We start by documentng that, over the perod, where we have data for both returns and order flow, the patterns n return comovement resemble those n order flow comovement. Accordngly, Panel A provdes ntraday (15-mnute) betas, Panel B daly betas, and Panel C weekly betas for Here, we use TAQ and ISSM md-quote returns. Correspondng to the change n order flow comovement documented earler, there s Page 25

28 a strkng and statstcally sgnfcant shft n return comovement for the swtchng frms, evdent at every frequency. Startng wth 15-mnute returns, we see that the mean slope on returns ncreases from 0.38 before the swtch to 0.57 after the swtch (mean =0.19), whle the slope on AQ returns drops from 0.75 to 0.31 (mean = 0.44). Wth daly returns, the mean slope ncreases from 0.60 to 0.75 whle the mean AQ slope declnes from 0.60 to The mean slope n the weekly return regresson rses from 0.66 to 0.81 and the AQ slope drops from 0.55 to Overall, the smlarty between the return comovement and order flow comovement results suggests that we can draw nferences about order flow comovement over the perod by studyng return comovement over ths perod. 15 Panel D contans the coeffcents from (3), estmated usng daly returns for the perod (we do not report the weekly results to save space). The mean beta for swtchng stocks ncreases by 0.27 (t= 5.9), whle the mean AQ beta declnes by 0.45 (t = -7.6). These comovement shfts are smlar to those over the perod, reported n Panel B. Thus, sharp changes n return comovement are also evdent durng a perod where ndexng (exchange-level or otherwse) s unlkely to have been popular. In our second ndexng test we re-estmate (1) for stock order flow after droppng order flow for the 100 largest and AQ stocks (by market value) from aggregate order flow n each perod (ntraday, daly and weekly; n the weekly specfcaton, we use the begnnng-of-week market values to rank frms). Large cap stocks are lkely to be a 15 We do not explore comovement n monthly returns because wth a one-year pre- and post-swtch measurement wndow, monthly returns would gve us only 12 observatons per frm. Page 26

29 central part of any exchange-based ndexng strategy, and ther order flows should reflect ndexer tradng. Thus, f ndexng s the man cause of exchange nfluences on order flow, comovement wth aggregate order flow based on the remanng and AQ stocks ought to be much weaker. These results are presented n Panel A of Table When we drop order flow for the 100 largest frms from exchange order flow, the changes n ntraday comovement wth and AQ order flows are 0.41 and -0.38, compared to 0.43 and wth the full set of frms, reported n Table 2, Panel A. (Not tabulated, the correspondng daly changes are 0.52 and and the weekly changes are 0.38 and -0.33, also nearly dentcal to the values n Table 2.) Overall, t does not appear that the comovement patterns are drven by a desre to mmc large cap stocks on an exchange. Our thrd test re-estmates (1) after droppng the last 30 mnutes n the tradng day. Index funds are lkely to trade near the end of the day n order to mnmze trackng error: Harford and Kaul (2005) show that order flow comovement for S&P consttuent stocks ncreases materally at the end of the day. Panel B of Table 5 has the results. We see that the mean values of and are 0.42 (t=22.70) and (t=-13.17). These are smlar to the values for the full day n Table 2, panel A. The fact that comovement s strong for the entre tradng day, and not just at the end of the day, suggests that ndexng s not the chef drver of our results. Fnally, an mportant dfference between ndex addtons and exchange swtches s 16 In the nterests of brevty, Table 5 presents the results for the swtchng sample alone. We have estmated every model for the control sample and our nferences from earler tables reman unaltered. Page 27

30 that when a stock s added to an ndex, ndexers have lttle choce but to buy and sell t n conjuncton wth other stocks n the ndex. There are no such compulsons for stocks that swtch exchanges, gven the lkely nformal bounds of such ndexng. Indeed, we beleve that nvestor preference for stocks on a partcular exchange s better descrbed as category or habtat preference, as n BSW, than as a formal ndexng strategy. 5.2 Informaton lags Informaton lags can explan the shfts n comovement f nvestors trade more rapdly on local nformaton than on cross-market nformaton. Suppose that nformaton relevant to the economy orgnatng on the (and therefore reflected n aggregate order flow) reaches AQ wth a delay. Before a frm moves to the, ts order flow wll react to ths nformaton wth a lag, so we wll see a small coeffcent on contemporaneous order flow and postve coeffcents on lagged order flow. After the swtch, the coeffcents on lagged order flow should declne and the coeffcent on contemporaneous order flow should ncrease. The reverse happens wth AQ order flow snce the reacton to AQ news slows after the swtch. However, n ths case, the sum of the contemporaneous and lagged coeffcents for each market should reman roughly constant after the swtch. The fact that the order flow comovement shfts reman strong at the weekly frequency suggests that nformaton lags are unlkely to be the man explanaton. Nevertheless, we formally evaluate ths explanaton by regressng stock order flow on lagged and contemporaneous values of aggregate and AQ order flows. We use fve lags of and AQ order flows n the ntraday analyss, and nteract these terms Page 28

31 wth the post-swtch dummy. The results are presented n panel C of table 5, and can be summarzed as follows. The sums of the changes n the coeffcents on lagged and AQ order flows have small and nsgnfcant means (the mean changes n the ndvdual coeffcents, not reported, are lkewse nsgnfcant). Thus, we fnd no evdence of a drop n the cumulatve senstvty to lagged order flow to offset the ncreased senstvty to contemporaneous order flow followng the swtch (or the opposte for AQ order flow). Delays n the transmsson of nformaton across exchanges do not explan the nfluence of co-locaton on stock-level order flow. 5.3 Other controls Specfcaton (1) models stock order flow as a functon of exchange level order flow alone. Mssng from these specfcatons are other systematc factors, such as market returns. For example, f the market return n the pror perod s postve, perhaps nvestors wll sell part of ther postons n many stocks (such contraran behavor s found n Chorda, Roll and Subrahmanyam, 2002). To address ths possblty, we re-estmate the order flow regressons ncludng lagged frm and market returns, as well as lagged frm and market order flows. Table 5, Panel D reports the results (only the coeffcents on order flow are tabulated to save space). The changes n the coeffcents on contemporaneous exchange order flows declne slghtly (the mean values of and are now 0.32 and -0.27), but reman hghly sgnfcant. We fnd a sgnfcantly postve coeffcent on lagged own order flow, a sgnfcantly negatve Page 29