Asset Returns and the Listing Choice of Firms

Transcription

1 Asset Returns and the Lstng Choce of Frms Shmuel Baruch Gdeon Saar 1 Frst Draft: July 2003 Ths Verson: December Baruch s from the Davd Eccles School of Busness, Unversty of Utah, Salt Lake Cty, UT Tel , E-mal: fnsb@busness.utah.edu. Saar s from the Johnson Graduate School of Management, 455 Sage Hall, Cornell Unversty, Ithaca, NY Tel , Fax , E-mal: gs25@cornell.edu. We wsh to thank Nkolay Halov and Yang Lu for research assstance, and Hendrk Bessembnder, Robert Bloomfeld, Ekkehart Boehmer, Danel Del, Therry Foucault, Robert Jarrow, Ron Mchaely, Maureen O Hara, and semnar partcpants at Cornell Unversty and the NBER Market Mcrostructure Group for helpful comments.

2 Abstract We propose a mechansm that relates asset returns to the frm s optmal lstng choce. The crucal element n our framework s not a dfference n the structure or rules of the alternatve markets, but a dfference n the return patterns of the securtes that are traded on these markets. We use a smple tradng model wth asymmetrc nformaton to show that a stock wll be more lqud when t s lsted on a market where smlar securtes, or securtes wth whch ts value nnovatons are more correlated, are traded. We emprcally examne the mplcatons of our model usng NYSE and Nasdaq securtes, and document that the return patterns of securtes lsted on the NYSE ndeed look dfferent from the return patterns of Nasdaq securtes. Stocks that are elgble to lst on the other market but do not swtch have return patterns that are smlar to those of other securtes on ther own market but dfferent from the return patterns of securtes lsted on the other market. We show that the return patterns of stocks that swtch markets change n the two years pror to the move to be more smlar to the return patterns of securtes lsted on the new market. Furthermore, the greatest mprovement n lqudty s experenced by the swtchng stocks whose return patterns resemble most the return patterns of securtes lsted on the new market. Our results suggest that managers choose the market on whch to lst optmally to enhance the lqudty of ther stocks.

3 Asset Returns and the Lstng Choce of Frms Many frms have a choce to make about a prmary market on whch to lst ther stocks, e.g., they could qualfy for lstng on ether the NYSE or Nasdaq. Does t really matter where frms lst? Why do frms n the same ndustry tend to cluster on the same market? Are observed lstng choces consstent wth ratonal decson makng? These are the questons that we study n ths paper. There s emprcal evdence suggestng that the lstng decson can affect the lqudty of the stock or the frm s vsblty. 1 As a result, pror lterature has suggested several characterstcs of markets that can brng about an optmal lstng choce, such as market structure, lstng requrements or fees, and regulatory oversght (ncludng corporate governance rules). 2 In ths paper we propose a new determnant of the lstng decson. We abstract from the partcular features of the alternatve markets and nstead put forward a mechansm that relates the frm s asset returns to the optmal lstng choce. The crucal element n our framework s not a dfference n the structure or rules of the markets, but rather a dfference n the return patterns of the securtes that are traded on these markets. We begn by developng a smple tradng model where multple securtes are traded n one of two markets and there s nformaton asymmetry among nvestors. We show that a stock s more lqud when t s lsted on a market where smlar securtes, or securtes wth whch ts value nnovatons are more correlated, are traded. The drvng force behnd the result s that market makers can extract nformaton about the value of the stock from the order flows of other securtes n the market. Naturally, ths nformaton s more relevant when assets are smlar, and therefore trades have smaller prce mpacts (or lower adverse selecton costs), reflectng greater confdence market makers have n the prces they set. If 1 See Grammatkos and Papaoannou (1986), Sanger and McConnell (1986), Edelman and Baker (1990), Sanger and Peterson (1990), Cowan, Carter, Dark, and Sngh (1992), Chrste and Huang (1994), Kadlec and McConnell (1994), Aggarwal and Angel (1997), Baker, Powell, and Weaver (1999a, 1999b), Elyasan, Hauser, and Lauterbach (2000), Kalay and Portnaguna (2001), and Bessembnder and Rath (2002). 2 See, for example, Cowan, Carter, Dark, and Sngh (1992), Wan (2001), Hedle and Huang (2002), Lpson (2002), and Foucault and Parlour (2004). 1

4 managers of frms care about the lqudty of ther stocks, our model suggests that they should lst ther stocks on a market wth smlar securtes. 3 Ths nsght can explan why managers of frms often cte ther desre to be on a market wth smlar frms as a motve for choosng to lst on a partcular market. For example, the Charman and CEO of Alled Captal commented on ther move from Nasdaq to the NYSE sayng, There aren t a lot of dvdend-payng stocks on Nasdaq, and, as a dvdend payer, we really thnk our stock s better suted to the NYSE. The CFO of CARBO Ceramcs noted on the move of hs frm from Nasdaq to the NYSE, Although the Nasdaq Natonal Market has been extremely helpful to us snce our ntal publc offerng n 1996, we beleve that our move to the New York Stock Exchange s consstent wth the NYSE lstng of the majorty of publcly-traded companes n the olfeld servces ndustry. 4 These quotes suggest that whle managers of frms may not be thnkng explctly n terms of the comovement of ther stocks returns wth those of other frms, they probably have an ntutve noton of what consttutes a smlar frm. Therefore, our model provdes the ratonale for why frms n the same ndustry, presumably havng more commonalty n ther return patterns, tend to lst on the same market. Stll, the quote about Alled Captal s move suggests that ndustry membershp may not always be the determnng crteron. A mature frm n a certan ndustry may have more n common wth a mature frm n another ndustry than wth a start-up n ts own ndustry. When a frm matures and ts return-generatng process changes to reflect more of the general condtons n the economy, t may be better off swtchng to a market wth other mature frms. Ths seems to correspond to the path many frms take: lstng for the frst tme on Nasdaq and then movng to the NYSE. Snce not all Nasdaq frms that are elgble to lst on the NYSE swtch markets, one could look 3 Managers mght care about the lqudty of ther stocks for several reasons. For example, adverse selecton and other mpedments to lqudty could be prced (e.g., Amhud and Mendelson (1986) and Easley and O Hara (2004)). Lqudty could also be a determnant of the cost of rasng external captal (Butler, Grullon, and Weston (2005)). 4 The quotes are taken from: () Alled Captal Corp. Moves to Bd Board by Robyn Kurdek n the Venture Captal Journal, June 1, 2001, and () CARBO Ceramcs Announces Completon of Publc Offerng and Move to the New York Stock Exchange on the PR Newswre, May 19,

5 at the return pattern of a stock n relaton to other securtes lsted on ts market to see f stayng put s ndeed the optmal thng to do. We therefore proceed to emprcally examne the mplcatons of our model usng NYSE and Nasdaq securtes over a three-year sample perod ( ). There s more than one way n whch we can defne smlarty n the return patterns of securtes, and each defnton suggests a dfferent emprcal methodology. Our frst defnton draws on an nterpretaton of the value nnovatons n the model as prvate nformaton about common factors n returns (see also Subrahmanyam (1991) and Caballe and Krshnan (1994)). We therefore examne smlarty n returns by lookng at the loadngs of securtes on estmates of common factors from a prncpal component procedure. Alternatvely, f prvate nformaton s only relevant to a subset of stocks, our model suggests that better estmates of the senstvtes to the value nnovatons (whch determne the optmal lstng choce) can be obtaned by elmnatng the market component n returns. We therefore examne smlarty measures constructed from correlatons of market model resduals. Our frst test provdes a check on the basc assumpton of the model that the two markets dffer n terms of the return patterns of the securtes that are lsted on them. We conduct a prncpal component analyss and fnd that Nasdaq securtes load more heavly on the frst prncpal component and that NYSE securtes load more heavly on the second prncpal component, consstent wth the assumpton of our model. When testng the model s predctons, we look both at frms that swtch markets and at frms that are elgble to swtch but stay put. Whle managers of the swtchng frms make an actve decson to move, managers of frms that are elgble to move (but choose not to do so) make passve decsons that mpact the lqudty of ther stocks. We start by examnng the managers passve decsons. We create four groups: () Nasdaq Natonal Market common, domestc stocks that are elgble to lst on the NYSE, () all other Nasdaq Natonal Market securtes, () NYSE common, domestc stocks that are elgble to lst on Nasdaq, and (v) all other NYSE securtes. If managers of frms seek to 3

6 mprove lqudty by ther choce of a lstng venue, our model suggests that stocks that are elgble to swtch markets would have return patterns more smlar to those of other securtes n ther own market and less smlar to those of securtes n the other market. We fnd that ths s ndeed the case, supportng the noton that the passve decsons managers make by remanng lsted on ther markets are optmal wth respect to lqudty. We then examne the actve choces made by managers of frms that move from Nasdaq to the NYSE durng our sample perod. 5 We fnd evdence of a change n ther return patterns two years pror to the move n the drecton of beng more smlar to the return patterns of NYSE securtes. These results support the concluson that managers make optmal lstng decsons. Lookng at the return patterns of swtchng frms before they swtch also demonstrates the robustness of our conclusons to the alternatve explanaton that correlated lqudty tradng s drvng the smlarty n return patterns that we document for securtes that are lsted on the same market. Furthermore, we show usng several prce mpact measures that the lqudty of the swtchng stocks mproves upon movng to the NYSE. Whle consstent wth an optmal lstng decson, ths result s hardly surprsng because such lqudty gans were documented n pror studes. However, we also fnd that the degree of lqudty mprovement s greater for stocks wth more smlar return patterns to those of NYSE securtes, evdence consstent wth the nsghts of our model where smlarty n return patterns among stocks lsted on the same market affects ther lqudty. Much of the pror lterature analyzes the determnants of managers optmal choce n the context of the two domnant markets n the U.S., the New York Stock Exchange (NYSE) and the Nasdaq Natonal Market. 6 Because there are dfferences n the market structure, rules, 5 There was one frm, Aeroflex Incorporated, that voluntarly swtched from NYSE to Nasdaq durng our three-year sample perod (see Kalay and Portnaguna (2001)). We repeated all the tests wth Aeroflex n the sample and our conclusons were unchanged. We chose to present the results wthout Aeroflex for two reasons. Frst, t smplfes the exposton. Second, t does not obscure the fact the results are essentally drven by the 86 stocks that moved from Nasdaq to the NYSE rather than the sngle stock that moved n the opposte drecton. 6 The optmal lstng decson we nvestgate concerns the frm s prmary market. We do not study the decson of a manager to cross-lst hs frm s stock on addtonal markets abroad (see, for example, Pagano, 4

7 and lstng requrements of these two markets, ths research effort has focused on dentfyng the structural element or the rule that ether makes one market superor to the other or explans the reasonng behnd the decsons of some frms to lst on Nasdaq and others on the NYSE. For example, Hedle and Huang (2002) and Lpson (2002) observe that the NYSE s a centralzed floor-aucton market whle the Nasdaq s a fragmented screen-based market wth multple dealers and alternatve tradng systems. If some nvestors have prvate nformaton and they can better hde n a fragmented screen-based market, movng from Nasdaq to the NYSE would beneft nvestors by reducng the extent of adverse selecton. Foucault and Parlour (2004) note that dfferent markets may charge dfferent lstng fees, a stuaton that characterzes the NYSE and Nasdaq. They provde a model where frms self-select to the most approprate market as a tradeoff exsts between lstng fees and transacton costs. 7 Wan (2001) argues that the dfferent market structures gve rse to dfferent volume fgures on the NYSE and Nasdaq, and that SEC rules restrctng the tradng of nsders are therefore less bndng on Nasdaq. Hs argument mples that ncentves of nsders that are not shared by outsde shareholders may affect the lstng choce. 8 Cowan, Carter, Dark, and Sngh (1992) suggest that NYSE rules dscouragng unequal votng rghts of multple share classes may also nfluence the decson of managers on where to lst. 9 One contrbuton of our work s that t proposes a new determnant of the lstng choce that does not rely on dfferences n the structure or rules of the markets. We also demonstrate how the lstng venue can affect the lqudty of stocks through the learnng process of market partcpants about prvate nformaton n prces. We do not suggest that ths s the only consderaton that managers have when makng a choce about lstng on a market, and our Röell, and Zechner (2002) and references theren on the queston of why do frms lst abroad). 7 See also Macey and O Hara (2002) on the economcs of stock exchange lstng fees and lstng requrements. 8 See also the models of Huddart, Hughes, and Brunnermeer (1999) and Chemmanur and Fulgher (2003) that demonstrate how nsders lstng decsons can be affected by publc dsclosure requrements. 9 See also Aggarwal and Angel (1997), Bessembnder (2000), Corwn and Harrs (2001), and Jan and Km (2004). 5

8 approach does not state that ths should be the only determnant. Stll, our model provdes a way to emprcally evaluate whether managers care about lqudty and make lstng decsons consstent wth mprovng the lqudty of ther stocks. We test the model (both the assumpton and the mplcatons) usng varous methodologes and fnd that managers seem to behave optmally n the sense of our model as f they want to maxmze the lqudty of ther stocks. When frms behave ths way, the mechansm we propose as a determnant of the lstng choce would perpetuate tself. In other words, when frms make actve decsons to lst on markets already populated by smlar frms, the assumpton of our model that the two markets dffer n terms of the return patterns of the frms that are lsted on them wll contnue to hold. As long as the assumpton holds, the optmal lstng choce wll be to jon the venue where smlar frms are lsted. Hence, ths determnant of the lstng choce seems robust to changes n the structure or rules of markets. Another contrbuton of our approach s to propose a relaton that goes from asset return patterns to the decsons of managers through an nformaton-asymmetry-drven market mcrostructure tradng model. Dow and Gorton (1997) and Subrahmanyam and Ttman (1999, 2001) recently presented models where a manager can learn useful nformaton from hs frm s stock prce. Our analyss suggests that certan manageral decsons would beneft from examnng the frm s return pattern alongsde the return patterns of other securtes n the market. And whle we fnd support for the optmal lstng choces of NYSE and Nasdaq frms, the nature of our approach provdes an ntuton that s more general than the specfcs of these two markets. The rest of the paper proceeds as follows. We present the theoretcal model and derve the mplcatons for the relaton between the lstng choce and lqudty n Secton I. Secton II s devoted to the emprcal work, and Secton III states our conclusons. 6

9 I Theory The purpose of ths secton s to develop a smple model that relates the lstng choce to lqudty. We frst descrbe the market pror to the lstng of a new asset. We consder an economy wth one rsk-free asset and two rsky assets (asset 1 and asset 2). Wthout loss of generalty, we set the return on the rsk-free asset to zero. Each rsky asset s traded n a separate market organzed as n Kyle (1985), where prces are set by compettve and rskneutral market makers. 10 After each round of tradng, there s a publc release of nformaton and the compettve market makers agree that the values of rsky assets 1 and 2 have changed by the value nnovatons s 1 + θ 1 and s 2 + θ 2, respectvely. We further assume that s 1 and s 2 are standard normal random varables, ndependent of each other and ndependent of θ 1 and θ The random varables θ 1 and θ 2 have zero means and can be correlated wth each other. We vew each asset n the model as representng a group of smlar assets traded on a sngle market. To smplfy the exposton, we assume that assets lsted on one market have a value nnovaton that does not exst n the values of assets lsted on the other market. Smlar results can be obtaned f value nnovatons for assets n both markets are weghted averages of both s 1 and s 2, but the weght on s 1 s greater n one market and the weght on s 2 s greater n the other market. The prmtve of our approach therefore s that smlar assets (those wth a common value nnovaton) are lsted on the same market. We then nvestgate the mplcaton of ths assumpton to the choce of a frm that consders where to lst or whether to move from one market to another when the dstrbuton of ts value nnovatons changes. The economy s populated by lqudty traders, nformed traders, and two groups of 10 See also Baruch, Karoly, and Lemmon (2005) who nvestgate nternatonal cross-lstngs n a multmarket model n the sprt of Kyle (1985). 11 An mplct assumpton that we are makng here s that the varances of s 1 and s 2 are not too small relatve to the varances of θ 1 and θ 2. In other words, we beleve that market makers can actually learn useful nformaton from the order flow. Also, note that we assume equal varances of the value nnovatons s 1 and s 2. When we carry out the emprcal work to test the mplcatons of the model, we standardze the varables used to mplement these value nnovatons so that they ndeed have unt varance. 7

10 market makers. We assume that the aggregate demand of the lqudty traders for each asset s a standard normal random varable that s ndependent of all other nnovatons n the market. There are two rsk-neutral nformed traders. The frst one observes the realzaton of s 1, whch s an unbased sgnal of the value nnovaton of asset 1, and the second nformed trader observes the realzaton of s 2. As n Kyle (1985), anonymty of traders mples that market makers observe only aggregate net orders. Snce we post two markets that are dentcal wth respect to ther structures and rules, we need to ntroduce some sort of segmentaton n order to have a meanngful dstncton between them. Therefore, we assume that market makers observe only the aggregate order flow that arrves n ther own market before settng clearng prces. Ths s not the frst theoretcal paper to use such a frcton. Chowdhry and Nanda (1991), for example, model a sngle asset that s traded on dfferent exchanges and assume that market makers n a gven exchange observe only the order flow that arrves to ther exchange (ths s the feature that dstngushes one market from another n ther model). 12 Emprcal work by Benvenste, Marcus, and Wlhelm (1992) and Coval and Shumway (2001) suggests that traders n one market ndeed have access to valuable nformaton that s not shared by traders n another market. 13 It s mportant to note, though, that the segmentaton we consder exsts only at the tme the order flow arrves n the market. After a trade has taken place, market makers n one market may observe prces set n the other market. Snce we assume publc release of nformaton after each round of tradng, the model allows for economy-wde reportng of last-trade prces. We now ntroduce another rsky asset, asset 3, that could potentally be lsted on ether 12 Ths assumpton s also used n Foucault and Gehrg (2004) and Baruch, Karoly, and Lemmon (2005). 13 Our emprcal work n Secton II uses NYSE and Nasdaq as the two markets. If, as n Benvenste, Marcus, and Wlhelm (1992), human nteracton on the NYSE floor conveys nformaton, Nasdaq market makers have no access to nformaton that NYSE specalsts observe. Lnkages such as ITS (the Intermarket Tradng System) do not allevate ths nformatonal frcton because only the orders a market wshes to pass on to another market travel through ITS, as opposed to the entre order flow. Also, conversatons wth practtoners suggest that many tradng desks on Wall Street are organzed such that traders n lsted stocks st together and traders n over-the-counter stocks st together, facltatng better nformaton sharng on stocks that are traded on the same market. 8

11 market. To have the most general case, the nnovaton of asset 3 s gven by a lnear combnaton of s 1 and s 2 : a s 1 + b s 2 + θ 3. The scalars a and b are the senstvtes of asset 3 s value to the nnovatons s 1 and s 2. The magntudes of a and b determne whether asset 3 s more smlar to asset 1 or asset 2. The random varable θ 3 has zero mean and s ndependent of s 1 and s 2 but possbly correlated wth θ 1 and θ 2. The lqudty demand for ths asset s a standard normal random varable, ndependent of all other random varables. 14 Say asset 3 s lsted on market 1, where asset 1 s traded. Then, market makers prce asset 3 based not only on ts own aggregate demand but also on the demand they observe for the other asset lsted on the same market. Smlarly, f asset 3 s lsted on market 2, market makers can observe the aggregate demands for both assets 2 and 3 when settng the prce of asset 3. Segmentaton of markets and rsk-neutralty of the nformed traders mply that the tradng strateges and prce rules n one market are unaffected by the tradng actvty takng place n the other market. We can therefore study the outcome of each market separately, and we focus on the market where asset 3 s lsted. Consder frst the case n whch asset 3 s lsted on market 1. We can wrte the value of the assets traded n market 1 usng matrx notaton as Ṽ = µ + F S + Θ, where Ṽ =(ṽ 1, ṽ 3 ), µ R 2 s the value of the assets pror to the nnovaton, F s a matrx of scalars gven by ( ) 1 0 F =, (1) a b S =( s 1, s 2 ) s the vector of value-relevant sgnals of the nformed traders, and Θ =( θ 1, θ 3 ). Let Z R 2 be the orders submtted by the lqudty traders to market 1. Let X 1 R 2 and X 2 R 2 be the orders submtted by the frst and second nformed traders, respectvely. Note that each nformed trader can submt orders for both assets 1 and 3. Let P R 2 be 14 In Secton II.3 we dscuss an extenson of the model that relaxes the ndependence assumpton by allowng lqudty tradng n one asset to be correlated wth lqudty tradng n another asset that s lsted on the same market. We show that ths does not change the man mplcaton of the model. 9

12 the clearng prces of the two assets, X = X 1 + X 2 be the aggregate demand of the nformed traders, and Y = X + Z be the net order flow submtted to the market. An equlbrum s a prce rule P : R 2 R 2 and strateges X 1,X 2 R 2 such that: () gven the strateges, the prce rule satsfes the condton P = E[Ṽ Ỹ ], and () gven the prce rule, the -th nformed trader ( = {1, 2}) maxmzes the expected profts E[(Ṽ P )X s ]. Theorem 1. There exsts a lnear equlbrum n whch () the prce rule s gven by P (Ỹ )= µ +ΛỸ, where Λ s a 2 2 matrx of scalars, and () aggregate nformed tradng can be wrtten as X = βs, where β s a 2 2 matrx of scalars. The matrces Λ and β are the solutons to the system of equatons (2) below satsfyng the second order condton that (Λ T +Λ)s a negatve semdefnte matrx: 15 β = (Λ+Λ T ) 1 F (2) Λ = Fβ T (I + ββ T ) 1 Proof of the theorem can be found n Appendx A. In equlbrum, the dagonal entres of the matrx Λ are the prce mpacts of market orders n asset 1 and asset 3. Let λ 3 (1) be the prce mpact of market orders n asset 3 when t s lsted on market 1 (.e., the second row, second column entry of the matrx Λ). Smlar to λ n the sngle-asset Kyle (1985) model, λ 3 (1) measures the lqudty of asset 3. Indeed, an unnformed trader who demands z of asset 3 expects to pay z 2 λ 3 (1) for mmedacy. The off-dagonal entres represent prce changes n one asset nduced by observng order flow n the other asset. These affect the nformatonal effcency of the asset s prce, but lke other cases where the prce of an asset changes due to publc nformaton, these prce movements are not manfestatons of the llqudty of a stock. Lqudty s measured by how much the order flow n an asset moves ts own prce (the dagonal entres). It follows from the proof provded n Appendx A (see equaton (6)) that λ 3 (1) = a2 + b (1 + b ). 2 a 2 +(1+ b ) 2 15 Superscrpt T denotes the transpose operaton and I denotes the dentty matrx. 10

13 In order to compare lqudty when an asset s lsted on one market versus the other, we need to fnd the prce mpact of market orders when asset 3 s lsted on market 2. It s straghtforward to show that a smlar lnear equlbrum exsts n ths case as well, where the prce mpact of market orders n asset 3 s gven by 16. λ 3 (2) = b2 + a (1 + a ) 2 b 2 +(1+ a ) 2 To determne whether lqudty s better when asset 3 s lsted on market 1 or on market 2, we calculate the dfference λ 3 (1) 2 λ 3 (2) 2 = b4 a 4 +2( b 3 a 3 )+b 2 a 2 4(a 2 +(1+ b ) 2 )(b 2 +(1+ a ) 2 ) (3) The followng proposton follows mmedately: Proposton 1. If a > b then lqudty s better when asset 3 s lsted on market 1, and f b > a then lqudty s better when asset 3 s lsted on market 2. Proposton 1 states that f the magntude of the senstvty of asset 3 to s 1 (the valuerelevant prvate nformaton of the frst nformed trader) s greater than the magntude of ts senstvty to s 2, then lqudty wll be better f the asset s lsted on market 1. Conversely, f the magntude of the senstvty of asset 3 to s 1 s smaller than ts senstvty to s 2, then lqudty wll be better f the asset s lsted on market 2. What s the ntuton behnd ths result? The key can be found n the off-dagonal terms of the matrx Λ that are specfed n the proof of Theorem 1. They represent the market makers nference from order flow of one asset that s relevant to the prce of the other asset. In the model, f we assume that θ 1 and θ 3 are uncorrelated, a s the comovement of the value nnovatons of asset 1 and asset 3. Heurstcally, the greater the magntude of a, the more that market makers can learn about the value nnovaton n asset 3 by observng the order 16 The proof s analogous to the one of Theorem 1 and s therefore omtted for brevty. 11

14 flow n asset As such, they do not need to change ther belefs (and hence the prce) to the same extent n response to the order flow n asset 3, and therefore the prce mpact of market orders n asset 3, λ 3 (1), s smaller. In other words, the lqudty of asset 3 when lsted on market 1 wll be better than ts lqudty when lsted on market 2 f asset 3 comoves more wth asset 1 than wth asset 2. In a smlar fashon, when asset 3 comoves more wth asset 2 than wth asset 1, or when b > a, lqudty wll be better f asset 3 s lsted on market It s nterestng to note that the lqudty of the exstng asset n the market on whch asset 3 lsts also mproves. For example, f asset 3 lsts on market 1, the ablty of market makers who trade asset 1 to learn about the prvate sgnal s 1 ncreases because they can observe the order flow n asset 3. Therefore, the new lstng provdes a postve externalty to the market. We can further show that when a > b, the mprovement to the lqudty of asset 1 when asset 3 lsts on market 1 s greater than the mprovement to the lqudty of asset 2 when asset 3 lsts on market 2 (and vce versa f b > a ). 19 Ths result s rather ntutve because the effcency wth whch market makers learn about prvate nformaton depends on the strength of the commonalty n the value nnovatons of the exstng and the 17 See also Strobl (2001) who nvestgates the allocaton of multple stocks to specalsts on the NYSE n a nosy ratonal expectatons framework. Bhattacharya, Reny, and Spegel (1995) look at market breakdowns drven by adverse selecton when multple securtes are traded on a sngle market. They demonstrate the effect of destructve nterference that mght occur when the value nnovatons of the traded securtes are too hghly correlated. 18 One could clam that ths logc should also apply to allocaton of stocks to specalsts on the floor of the NYSE. In other words, that lqudty would be enhanced f stocks wth correlated value nnovatons are traded by the same specalst. Unfortunately, the lmtatons of any sngle person wth respect to nformaton processng or nteractons wth brokers and computer screens prevent specalsts from tradng stocks that are very smlar (lke two large technology stocks). The reason s that when there s news about a frm or an ndustry, the number of orders that arrve for one actve stock makes the specalst (and the specalst s clerk) unable to handle tradng n any other securty. Ths s because most trades on the NYSE requre the specalst (or the clerk) to manually approve the executon. Havng another actvely traded securty that s nfluenced by the same news event at the same tme would prevent the specalst from mantanng an orderly market n ether securty. 19 To prove ths clam we verfed that whenever a > b, (1 + b ) 2 a 2 +(1+ b ) 2 (1 + a ) 2 b 2 +(1+ a ) 2 < 0 where the frst term s the prce mpact of asset 1 when asset 3 lsts on market 1 and the second term s the prce mpact of asset 2 when asset 3 lsts on market 2. 12

15 new assets. II Emprcal Evdence If managers care about lqudty, Proposton 1 suggests that frms would lst on markets where smlar securtes are lsted. We use ths mplcaton of our model to study the lstng choces of frms on the two man U.S. markets: NYSE and Nasdaq. The sample and data sources are dscussed n Secton II.1. The tests n Secton II.2 have two goals. Frst, we examne whether the basc assumpton of the model that the two markets dffer n terms of the return patterns of the securtes lsted on them holds. Second, we study the passve decsons of frms that are elgble to move to the other market but do not. Our model mples that elgble stocks should not move f ther returns are more correlated wth securtes lsted on ther own market than wth the securtes on the other market. In Secton II.3 we present tests of the actve decsons of managers usng a sample of frms that swtch markets. Our model suggests that a frm should swtch f ts return pattern has changed to more closely resemble the return patterns of stocks on the other market. We also dscuss n ths secton the robustness of our conclusons to return behavor nduced by correlated lqudty tradng among stocks that are lsted on the same market. Fnally, we examne whether dfferences n lqudty mprovement upon swtchng are related to the extent of smlarty between the return patterns of the swtchng frms and those of securtes lsted on the new market. II.1 Sample, Data, and Defntons of Return Smlarty Our sample perod s (three years). In order to nvestgate the passve lstng choces of managers n Secton II.2 we need to dentfy the frms that actually have a choce: those that are lsted on one market and satsfy the lstng requrements of the other market. Therefore, we want to dentfy Nasdaq common, domestc stocks that were elgble to lst on 13

16 the NYSE on the frst day of the sample perod, as well as NYSE common, domestc stocks that were elgble to lst on Nasdaq on that date. We use nformaton n the CRSP and COMPUSTAT databases to evaluate each common, domestc stock and see f t satsfes the ntal lstng requrements of the other market. Most crtera specfed by the NYSE and Nasdaq can be mapped rather well to the nformaton n these two databases. Appendx B contans the varables we use from CRSP and COMPUSTAT for dfferent lstng requrements. Some slppage, however, s nevtable as the NYSE and Nasdaq evaluate nformaton provded by the frms themselves that s not necessarly dentcal to what we observe n the databases. Therefore, despte our best efforts, the procedure we use to determne elgblty may ntroduce some nose nto the analyss. Our procedure dentfes 1,155 NYSE common, domestc stocks that were elgble to lst on Nasdaq on January 1, We wll refer to them throughout the analyss as the NYSE1 group. Smlarly, our procedure yelds a lst of 408 Nasdaq Natonal Market (NNM) common, domestc stocks that were elgble to lst on the NYSE at the begnnng of our sample perod, henceforth NNM1. Panel A of Table 1 presents summary statstcs for the securtes n both groups usng nformaton from the CRSP database. For the tests of the actve lstng decson of managers n Secton II.3 our sample conssts of all common, domestc stocks that voluntarly swtched from the Nasdaq Natonal Market to the NYSE durng our sample perod. There were 86 such moves n the years Whle the NYSE revsed Rule 500 n 1999 to make t easer for frms to voluntarly delst, only one frm (Aeroflex) chose to move from the NYSE to Nasdaq durng our sample perod. 20 When we nclude Aeroflex n our tests, the results are unchanged. However, to smplfy the exposton and not to obscure the fact that the results are drven by swtches n one drecton, we present n the paper the results of the tests on the sample of frms that moved from Nasdaq to the NYSE. 21 Panel B of Table 1 provdes summary statstcs on the swtchng sample. 20 See Kalay and Portnaguna (2001) for a dscusson of Aeroflex s move from the NYSE to Nasdaq. 21 The results of the tests that nclude Aeroflex are avalable from the authors upon request. 14

17 Movng from the theoretcal model to the emprcal tests requres us to carefully defne what we mean by smlarty of frms. In the model, smlarty was formalzed by a common source of varaton n the value nnovatons of the assets. In the emprcal work, we would lke to capture ths smlarty by lookng at the daly return patterns of securtes. Stll, there s more than one way n whch the model can be used to motvate emprcal defntons of such smlarty, and we are usng two separate defntons n our tests. For the frst defnton of return smlarty we nterpret the prvate nformaton n the model as nformaton about common factors n returns. Pure Nasdaq stocks may be more senstve to a certan common factor, say s 1, whle pure NYSE stocks may be more senstve to another common factor, s 2. Accordng to ths nterpretaton, the value nnovaton components modeled by θ 1, θ 2, and θ 3 represent unsystematc rsk and are therefore uncorrelated wth each other. Our model predcts that Nasdaq stocks that are elgble to lst on the NYSE would stay on Nasdaq f they are more senstve to s 1 than to s 2,.e., f a > b (see Proposton 1). Smlarly, NYSE stocks that are elgble to lst on Nasdaq reman on the NYSE because ther returns are more senstve to s 2, lke the other NYSE securtes, than to s 1 ( b > a ). If managers of swtchng frms behave optmally, we should see pror to ther move that they become more senstve to s 2 and less to s 1. The emprcal methodology we use to examne ths defnton of return smlarty s prncpal component analyss, where we look at the loadngs of NYSE and Nasdaq securtes on the prncpal components that consttute estmates of common sources of return varaton. Our second defnton of return smlarty nterprets a prvate sgnal n the model as nformaton that can be relevant to certan frms but not necessarly to the entre market. Snce θ n the model could be correlated across securtes, t can represent the stock-specfc senstvty to the market portfolo multpled by the excess return on the market. Therefore, θ assumes the role of the common element n returns, whle the ndependent sgnals, s 1 and s 2, nduce some comovement n the returns of certan stocks even f they are not common to 15

18 the entre economy. 22 Under ths nterpretaton, we would obtan a cleaner pcture of return varaton due to s 1 and s 2 by removng the nfluence of the market portfolo and constructng emprcal estmates of a and b from the absolute values of correlatons of return resduals. These estmates could then be used to look at the mplcatons of Proposton 1. II.2 Smlarty n Return Patterns: NYSE and Nasdaq Stocks In ths secton we look at the return patterns of NYSE and Nasdaq securtes. Frst, we would lke to see f the assumpton of our model that the two markets dffer n terms of the return patterns of the frms that are lsted on them ndeed holds. We also nvestgate n ths secton the passve decsons of frms to reman lsted on a gven market. If ndeed these are optmal n the sense of our model (.e., they maxmze lqudty), then we should observe that stocks that are elgble to move to the other market but do not swtch are more smlar to securtes lsted on ther market than to securtes lsted on the other market. To carry out the tests, we create two addtonal groups of securtes. NYSE2 conssts of all securtes that were contnuously lsted on the NYSE durng the sample perod and are not ncluded n NYSE1. Smlarly, NNM2 conssts of all securtes that were contnuously lsted on the Nasdaq Natonal Market durng our sample perod and are not part of NNM1. 23 We examne the frst defnton of return smlarty usng a prncpal component analyss. Because there are 752 days n the sample perod and several thousand securtes, we form 15 portfolos from the securtes n each group (for a total of 60 portfolos). 24 Snce each group contans a dfferent number of securtes, N, we randomly assgn approxmately N/15 securtes to each portfolo. For example, the 1,155 stocks n the NYSE1 group are dvded 22 Ths does not change f one adds uncorrelated sources of nose, ɛ 1 and ɛ 2, to the values of securtes 1 and 2, respectvely. 23 NYSE2 and NNM2 nclude non-common stocks (ADRs, REITs, etc.). We nclude them because market makers n our model can potentally learn useful nformaton from all other securtes that are lsted on the same market. We see no reason to restrct the market makers nformaton set n the emprcal work. We focus on common, domestc stocks for NYSE1 and NNM1 n order to make the determnaton of elgblty to lst on the other market less complex (and therefore to reduce potental msclassfcatons). 24 We aggregate securtes nto portfolos because the prncpal component analyss cannot dentfy loadngs when the number of observatons (days) s smaller than the number of varables (securtes). 16

19 nto 15 portfolos contanng 77 stocks each. 25 Portfolo returns are computed as averages of the daly returns on the stocks n the portfolo. We then perform a prncpal component analyss of daly portfolo returns n the three-year sample perod, retan the frst two prncpal components, and use an orthogonal rotaton. 26 The procedure provdes estmates of the loadngs on the two prncpal components. These loadngs can be nterpreted as the bvarate correlatons between the portfolos returns and the prncpal components. Panel A of Table 2 shows the means and standard devatons of the rotated factor loadngs for the two markets and the four groups: NNM1, NNM2, NYSE1, and NYSE2. NYSE securtes seem to load more heavly on the frst prncpal component (0.797) than do Nasdaq securtes (0.474). In contrast, Nasdaq securtes load more heavly on the second prncpal component than do NYSE securtes. A t-test demonstrates that the mean loadng of NYSE securtes s dfferent from that of Nasdaq securtes (p-value < ). Furthermore, the mean loadng on the frst component of the NNM1 group (Nasdaq common, domestc stocks that are elgble to move to the NYSE) s closer to the mean of all other Nasdaq securtes (NNM2) than to that of ether of the NYSE groups. The same can be sad of the common, domestc stocks n NYSE1: ther mean loadng on the frst prncpal component (0.868) looks more lke that of other NYSE securtes (0.726) than those of NNM1 (0.558) or NNM2 (0.389). A smlar pattern s observed n the loadngs on the second prncpal component (e.g., NYSE1 s closer n magntude to NYSE2 than to NNM1 or NNM2). The results of the prncpal component analyss are mportant n two respects. Frst, f we take the prncpal components to represent estmates of common factors, Nasdaq stocks are 25 We repeated the analyss wth an equal number of securtes n each of the 60 portfolos by randomly drawng wthout replacement 25 securtes from each group to form each portfolo. The number of stocks was chosen such that even the smallest group (NNM1) would enable us to form 15 dfferent portfolos wth 25 stocks each. The results of ths analyss were smlar to the results presented below and are therefore not reported. 26 We apply the commonly used VARIMAX rotaton (see, for example, Kaser (1958) and Hatcher (1994)). To assst us n makng the decson on how many prncpal components to retan, we used the tests proposed by Conway and Renganum (1988) and Connor and Korajczyk (1993). The tests suggested retanng two prncpal components, whch explan over 69% of the varance, both n the analyss here and n the analyss performed n Secton II.3. We also looked at what happens when we retan three or four prncpal components. In general, we observed smlar results to what we report below n that there was one prncpal component on whch NYSE securtes had hgh loadngs and another one on whch Nasdaq securtes had hgh loadngs. 17

20 more senstve to one factor ( s 1 ) and NYSE stocks are more senstve to another factor ( s 2 ). Ths fndng s consstent wth the assumpton that we use as the prmtve of our approach: that there are two groups of frms wth dfferent return patterns and each group s lsted on a dfferent market. Second, snce our test looks separately at common, domestc stocks that are elgble to move to the other market, we can say somethng about the optmalty of managers passve decson to reman lsted on a market. We fnd that common, domestc stocks that have the opton to swtch but opt not to do so have loadngs that are more smlar to those of other securtes n the same market than to loadngs of securtes lsted on the other market. The evdence s consstent wth managers makng optmal (passve) lstng decsons by not movng to the other market. We also carry out a test of the passve decsons usng the second defnton of smlarty n return patterns that was dscussed n Secton II.1. To remove the effect of common varablty n returns and focus on s 1 and s 2, we run a market model for each securty usng daly returns over the sample perod and the value-weghted portfolo of NYSE, AMEX and Nasdaq stocks (from CRSP) as a proxy for the market portfolo. We then take the resduals from the market model (to elmnate the common varaton n returns represented by θ ) and normalze them to have unt varance. Note that n the model, the comovement of the value nnovatons of asset 1 and asset 3 n market 1 after elmnatng θ 1 and θ 3 s smply a. Smlarly, the comovement of asset 2 and asset 3 n market 2 after elmnatng θ 2 and θ 3 s equal to b. Proposton 1 states that the dfference between the absolute values of a and b determnes the optmal choce of a manager between the two markets. Our market model procedure s meant to elmnate θ from the returns of all securtes. Therefore, to estmate a, we compute for each stock n NYSE1 and NNM1 the correlaton between ts normalzed resdual and the normalzed resduals of all securtes n the NNM2 group. We denote the average of the absolute values of these correlatons as ā. To estmate b, we compute for each stock n NYSE1 and NNM1 the correlaton between ts normalzed resdual and the normalzed resduals of all securtes n 18

21 the NYSE2 group. We denote the average of the absolute values of these correlatons as b. If Nasdaq common, domestc stocks that are elgble to move to the NYSE optmally stay on Nasdaq, t should be that the average absolute value of ther correlatons wth other Nasdaq securtes (NNM2) s greater than the average absolute value of ther correlatons wth the other NYSE securtes (NYSE2). On the other hand, for elgble NYSE common, domestc stocks we would predct that ther estmates of b (the magntude of comovement wth NYSE2 securtes) wll be greater than ther estmates of ā (the magntude of comovement wth NNM2 securtes). Therefore, we test whether ā b > 0 for NNM1 stocks and b ā > 0 for NYSE1 stocks. Panel A of Table 3 provdes the means and medans of ā, b, and ā b for the 408 NNM1 stocks. The t-test ndcates that the mean of ā b s postve and hghly statstcally sgnfcant (p-value < ). Smlarly, a Wlcoxon sgned-rank test shows that the medan s postve and statstcally sgnfcant. Panel B of Table 3 provdes the results for the 1,155 NYSE1 common, domestc stocks. The mean and medan of the dfferences b ā are postve and hghly statstcally dfferent from zero. Hence, all our fndngs n ths secton pont to the concluson that managers of frms choose to lst ther companes on the market where smlar frms are lsted, consstent wth lqudty maxmzaton. We qualfy ths concluson, though, because so far we have only tested the passve choces of managers where we nfer the choce from the fact that no acton was taken to move the stock to a dfferent market. Next we examne the actve decsons of managers to swtch markets. II.3 Tests Usng Swtchng Frms In ths secton, we test the mplcatons of our model usng frms that swtched from Nasdaq to the NYSE durng the three-year sample perod ( ). Our model suggests that a stock s return pattern affects ts lqudty va the nteracton wth other securtes that are lsted on the same market. If the return patterns of the stocks that swtch changed to be less 19

22 smlar to those of other securtes n the old market and more smlar to the return patterns of securtes n the new market, the decson to swtch markets could be motvated by the desre to mprove lqudty. The use of swtchng frms has another advantage n that t helps to emprcally dsentangle the mplcatons of our model from an alternatve explanaton: that correlated lqudty tradng s the drvng force behnd the return patterns we documented n Secton II.2. More specfcally, say each market has ts own class of lqudty (or nose) traders who trade the assets lsted on ther market but not assets lsted on the other market. An example of such traders s someone buyng and sellng ndex funds. As some ndexes are market specfc (e.g., the Nasdaq 100 Index), t s concevable that cash flows nto and out of such ndex funds, or creaton and redempton of Exchange Traded Funds that follow market-specfc ndexes (e.g., the QQQ), may cause the prces of assets lsted on the same market to move together. We frst look at what happens f correlated lqudty tradng among assets that are lsted on the same market s ntroduced nto the model of Secton I. Let ρ denote that correlaton. Fgure 1 shows how the dfference λ 3 (1) 2 λ 3 (2) 2 from Proposton 1 changes wth ρ. 27 To mantan a two dmensonal fgure, we fx a = 1 and draw 5 lnes: two for a > b, twofor a < b, and one for a = b. We see that the two curves for whch a > b are n the negatve doman, mplyng better lqudty on market 1, and the dfference λ 3 (1) 2 λ 3 (2) 2 becomes more negatve as ρ ncreases. The opposte happens when a < b : the two curves are n the postve doman, mplyng that lqudty s better on market 2, and the dfference s monotoncally ncreasng wth ρ. Therefore, the man mplcaton of our model (Proposton 1) stll holds when correlated lqudty tradng s consdered. From an emprcal pont of vew, there are clear dstnctons between our model and the correlated-lqudty-tradng explanaton. The model uses return patterns as a prmtve and suggests that these determne optmal lstng choces that renforce the stuaton where frms lsted on the same market have more smlar return patterns. On the other hand, the 27 The soluton to the model wth correlated lqudty tradng s avalable from the authors upon request. 20

23 alternatve explanaton takes the act of lstng as a prmtve and clams that lstng on a market brngs about correlated lqudty tradng and thus nduces smlarty n the return patterns of the newly-lsted frm and other frms traded on the same market. 28 To examne the robustness of our conclusons to ths alternatve explanaton, we use the swtchng frms to see whether changes n return patterns occur before a frm s lsted on a market, and whether the lstng decson s consstent wth movng to a market populated by more smlar frms. There are a couple of ssues that should be mentoned up front wth respect to ths exercse. Frst, there may be other reasons besdes lqudty to swtch markets (e.g., a preferred regulator). Swtches that are motvated by other consderatons would ntroduce nose nto the tests and make t more dffcult for us to fnd an effect. Second, there are really no gudelnes for how long pror to a swtch we should analyze the data to detect the changes n return patterns. It s reasonable to look at the year pror to the move, as the decson to move was probably made at that pont, but the decson could have been made followng a perod n whch the return pattern changed to be more lke that of frms lsted on the other market before managers decded to swtch. 29 Therefore, the choce of a perod to analyze pror to a swtch s arbtrary n nature. Wth these reservatons n mnd, we proceed to use the methodologes from Secton II.2 to examne the return patterns of the stocks n the swtchng sample. Our frst test s based on the prncpal components methodology. The 86 frms that swtched from Nasdaq to the NYSE dd so at dfferent ponts n tme durng the sample perod. Ideally we should look at the tme perods (say, one or two years) defned ndvdually for each frm gong back from ts swtch date. We are able to do that usng our second defnton of return smlarty, whch nvolves measures constructed from correlatons of market 28 See also Chan, Hameed, and Lau (2003). 29 One could also argue that managers may be able to foresee that ther frms would become more smlar to NYSE frms before the return pattern changes, and so the move would come n antcpaton of ths process. Of course, the mplcaton of ths nterpretaton s smlar to that of the alternatve explanaton, so our tests would not be able to separate the two. 21

24 model resduals. However, for the frst defnton of return smlarty, whch nvolves prncpal component analyss, we need to have the same tme nterval for all stocks. Therefore, we use calendar tme to defne the perods we nvestgate. In other words, for the frms that swtch n the year 2000 we look at the smlarty of return patterns n calendar years 1999 (year t 1) and 1998 (year t 2). Smlarly, for the frms that swtch n the year 2001 we look at return patterns n calendar years 2000 and 1999, and for the frms that swtch n 2002 we look at return patterns n 2001 and The frst step n our prncpal component methodology s to generate from the CRSP database a unverse of all securtes that were traded contnuously each year (1998, 1999, 2000, and 2001) on ether the NYSE or the Nasdaq Natonal Market (excludng the 86 stocks n our swtchng sample). To explan our procedure we wll focus on the 21 stocks that swtched n For these stocks, year t 1 s As n Secton II.2, we form 30 portfolos from the securtes n each market n 1999 (for a total of 60 portfolos) by randomly assgnng approxmately N/30 securtes to each portfolo. We also form a swtchng portfolo by computng the equal-weghted returns n 1999 of the stocks that moved from Nasdaq to the NYSE n Usng these 61 portfolos we then perform a prncpal component analyss of daly portfolo returns n 1999 and retan the loadngs on the frst two prncpal components. We want to look at the dstance of the swtchng portfolo s loadng from the loadngs of the NYSE and NNM portfolos, and how ths dstance changes over tme. We therefore compute 30 dstances of the swtchng portfolo s loadng from NYSE stocks by takng the absolute value of the dfference between the loadng (say, on the frst prncpal component) of the swtchng portfolo and the loadngs on the same prncpal component of the 30 NYSE portfolos. Smlarly, we compute 30 dstances of the swtchng portfolo s loadng from the loadngs of the NNM portfolos. We repeat ths process of randomly assgnng securtes nto portfolos and performng a prncpal component analyss 100 tmes. Ths procedure s repeated for the same swtchng stocks usng daly returns n 1998 to get ther loadngs n year t 2. We then look at the 27 common, domestc stocks that swtched n 2001 and 22