Investor Behavior over the Rise and Fall of Nasdaq

Transcription

1 nvestor Behavor over the Rse and Fall of Nasdaq JOHN M. GRFFN, JEFFREY H. HARRS, AND SELM TOPALOGLU * September 4, 2003 * Grffn s vstng at Yale Unversty and on faculty at the Unversty of Texas at Austn, Harrs s at Unversty of Delaware and s a former Vstng Academc Fellow at Nasdaq, and Topaloglu s at Queen s Unversty. Grffn can be reached at john.grffn@yale.edu. We are grateful for helpful comments and dscusson from Erc Falkensten, Ken French, Wll Goetzmann, Davd Hrshlefer, Roger bbotson, Steve Jordan, Jason Karcesk, Patrck Kelly, Stephen F. LeRoy, Federco Nardar, Spencer Martn, Bob Parrno, Avr Ravd, Jay Rtter, Geert Rouwenhorst, Laura Starks, René Stulz, Fabo Trojan, Kent Womack, and semnar partcpants at Arzona State Unversty, Dartmouth College, The Oho State Unversty, Unversty of Maryland, Unversty of New South Wales, Unversty of Texas, 2003 European Fnance Assocaton, and 2003 Western Fnance Assocaton.

2 nvestor Behavor over the Rse and Fall of Nasdaq Abstract The large theoretcal lterature about bubbles ncludes models where naïve ndvduals cause excessve prce movements and smart money trades aganst (and potentally elmnates) a bubble or where sophstcated nvestors follow market prces and help drve a bubble. We examne these competng vews by focusng on nvestor actvty over the spectacular rse and fall of Nasdaq from September 999 through 200. We fnd that both nsttutonal ownershp levels and volume on Nasdaq were hgh. nsttutons bought shares from ndvduals the day after market up-moves and nsttutons sold on net followng market dps. These patterns are pervasve throughout the market run-up and subsequent crash perod. Ths short-term nsttutonal trend-chasng behavor does not appear to be mechancally nduced by flows nto and out of mutual funds. Our evdence supports the vew that nsttutons contrbuted more than ndvduals to the Nasdaq rse and fall.

3 Optmal plannng for ndvduals and organzatons hnges crtcally on the level and mpled future return of the aggregate stock market. Rapd swngs n market prces can cause costly msallocatons of resources. The market run-up n the 990s was the greatest n U.S. hstory and the subsequent prce correcton has been the largest snce the Great Depresson. Durng ths unque perod, we examne the nteracton of nvestor groups on the Nasdaq market. Bubbles n asset prces have been the subject of extensve debate. Whle numerous theoretcal models examne the forces and assumptons necessary to generate prce bubbles, relatvely lttle emprcal research has dstngushed between these theores. Much of the theoretcal bubble lterature posts dfferent roles for nvestor groups, ncludng the nteracton of sophstcated traders wth less nformed agents. The goal of ths paper s to examne whether nsttutonal and ndvdual tradng actvty surroundng the rse and fall of the Nasdaq ndex s consstent wth any of the varous proposed vews of stock prce bubbles, abstractng away from whether the prce levels were justfable. Some models of stock prce bubbles post that nose traders smply chase past prces (postve feedback tradng) and lead prces away from fundamental values (DeLong et al. (990a)). n support of ths vew, Ofek and Rchardson (2003) demonstrate that nternet stocks n 2000 had low levels of nsttutonal ownershp and fewer block trades than non-nternet securtes. The tradtonal ratonal markets vew (e.g., Fredman (93), Fama (96)) recognzes that agents may trade rratonally but contends that such tradng does not substantally affect prces snce sophstcated traders (arbtrageurs) quckly trade aganst these agents to elmnate devatons from true economc values. n ths settng, rratonal bubbles do not occur. However, f arbtrageurs face lmted captal and are evaluated under a fnte tme horzon (as descrbed by Shlefer and Vshny (997)) then they may not trade aganst what they vew as a clear devaton from the true economc value. n ths vew, arbtrageurs may not elmnate a bubble but they would also not actvely partcpate n t. Shller (2000, p. 6-9) documents that n real terms the S&P prce run-up s even larger than those n the 920s and also shows that the prce-to-earnngs ratos n 2000 are much hgher than n the entre hstory of U.S. stock prces.

4 Alternatvely, DeLong et al. (990b) and Abreu and Brunnermeer (2002, 2003) propose a world where smart money, or arbtrageurs, actually see the drecton that unnformed nvestors are tradng and move ahead of these nvestors, further drvng stock prce movements. n support of these theores, Brunnermeer and Nagel (2003) fnd compellng evdence that 3 hedge fund managers generally took larger postons n hgh prce-to-sales stocks durng the Nasdaq run-up. 2 Shller (2000) provdes another reason why nsttutonal captal may move wth the market, notng that (p. 8), Professonal nvestors are not mmune from the effects of the popular nvestng culture that we observe n ndvdual nvestors. Our man tool of analyss s a unque dataset that allows us to classfy nvestors by type n Nasdaq 00 securtes over the perod from September, 999 through December 3, 200. We classfy order flow as ndvdual (retal) or nsttutonal based on the brokerage house that handles the order. Although the ndvdual/nsttutonal nvestor brokerage house classfcaton may not exactly ft the nose trader/arbtrageur dstncton, t does ft the popular noton that ndvduals are less nformed and that nsttutons are smart money. 3 We frst examne nsttutonal tradng and holdngs of Nasdaq 00 stocks and fnd that durng the dramatc market run-up from March 999 to March 2000 nsttutons generally ncreased ther postons n Nasdaq securtes. n March 2000 nsttutons held over 0 percent of Nasdaq s market captalzaton, smlar to levels of ownershp n other securtes. n addton, we fnd that nsttutons were responsble for more tradng volume than ndvduals n the sx-month perod pror to the Nasdaq peak n March We also examne the relaton between short-term daly stock prce movements and changes n ownershp. f a group of nvestors had a propensty to sell when the market rose and buy when the market fell then t could be nferred that they tend to have a dampenng effect on market movement. On a 2 Nevertheless, they document that hedge funds own less than one-thrd of one percent of the hgh prce-to-sales Nasdaq securtes, suggestng that hedge fund actvty seems unlkely to be the predomnant drver of the market. 3 Ths dea s supported by recent emprcal evdence from Chen, Jegadeesh and Wermers (2000) who show that stocks that mutual funds buy outperform those they sell, potentally at the expense of poor tmng of trades by ndvduals (Odean (999)). Coval, Hrshlefer, and Shumway (2002) fnd that there s a subset of ndvduals who do earn abnormal returns. 2

5 contemporaneous daly and lagged bass we fnd that ndvduals (and not nsttutons) play such a role. Large negatve (postve) return days are accompaned by nsttutonal sellng (buyng) and ndvdual buyng (sellng). There s an even stronger relaton on the next tradng day. For example, on the day followng a large negatve Nasdaq return, nsttutons sell approxmately $08 mllon worth of shares on average and buy $88 mllon followng large postve return days. We nvestgate these relatons usng daly VAR regressons and fnd that a one standard devaton ncrease n today s Nasdaq market return s assocated wth a 0.48 standard devaton ncrease n tomorrow s net nsttutonal buyng actvty. We also examne the tmng of nsttutonal/ndvdual tradng n relaton to the level of Nasdaq prces. t could be that, whle nsttutons are generally trend chasers, they move aganst prces durng the market run-up. We fnd ths s not the case. nsttutons buy stocks on the day followng Nasdaq market ncreases and sell securtes (to ndvduals) on the day followng market decreases durng the market runup as well as durng the crash, suggestng that sophstcated (nsttutonal) nvestors actvely contrbuted to the Nasdaq run-up and run-down. One explanaton for our fndngs s that the nsttutonal nvestors are merely reactng to redemptons and purchases of mutual funds. Goetzmann and Massa (999) fnd a strong contemporaneous relaton between S&P ndex fund flows and the contemporaneous daly ndex returns and Edelen and Warner (200) fnd evdence of mutual fund flows followng returns. We examne the jont relaton between daly nsttutonal order mbalances, mutual fund flows, and returns n the VAR framework. We fnd that, whle nsttutonal mbalances are postvely related to mutual fund flows, mutual fund flows cannot explan the strong contemporaneous or lagged relaton between nsttutonal mbalances and returns or the persstence throughout the rse and fall of the Nasdaq ndex. To gan a clearer pcture of the tmng of daly tradng we also examne ntradaly tradng behavor n fve-mnute wndows. ndvdual (nsttutonal) tradng s strongly postvely (negatvely) assocated wth contemporaneous fve-mnute market moves. However, ths actvty quckly reverses wth nsttutons followng short-term market returns (or the nformaton whch nduced the returns) and ndvduals tradng to the contrary. Overall, we fnd that nsttutonal traders (or smart money) do not 3

6 attempt to trade aganst market movements, but rather actvely partcpate n both the run-up and rundown of Nasdaq 00 prces. Our evdence s most consstent wth models where smart money follows past stock prce movements leadng to larger stock prce bubbles than would exst n ther absence. The outlne of our paper s as follows. Secton dscusses the role that the nteracton among traders plays n bubble theores. Secton descrbes the data. Secton examnes the level of nsttutonal ownershp and the fracton of tradng volume that can be assgned to nsttutons and ndvduals. Secton V detals the overall daly tradng behavor of nsttutonal and ndvdual nvestors n relaton to daly prce movements over the entre perod and tests whether these relatons have changed pre- and post-march We examne how captal nflows affect nsttutonal behavor n Secton V and the ntradaly tradng behavor of ndvdual and nsttutonal nvestors n Secton V. We address robustness ssues n Secton V and our conclusons follow.. Bubble Theores and Testable mplcatons A. Prce Bubbles and the nteracton between nvestor Groups What s a bubble? Kndleberger s (978) book on bubbles merely defnes t as an upward prce movement over an extended range that then mplodes. When referrng to the rapd ncrease and fall of the Nasdaq ndex, we use the term n ths smple manner. Yet people also often assocate bubbles wth a de-couplng of market prces from the underlyng fundamental values. 4 Ofek and Rchardson (2002) show compellng evdence that nternet stock prce levels were too hgh to be justfed by even exceptonal levels of expected earnngs growth. The purpose of our study s not to examne the relaton between Nasdaq prces and fundamentals but to examne the connecton between models that can generate bubbles (and those that argue they wll not exst) and the tradng behavor of nvestor groups. 4 Hndsght benefts from knowng the prce realzaton that actually occurred whereas n real tme, nvestors face only uncertan future probabltes. For example, the famous Dutch Tulp mana s often cted as a psychologcally drven bubble, yet Garber (989) argues that tulp bulb prce patterns may be consstent wth fundamentals. Whle our analyss centers on the lterature on bubbles, many of our predctons are related to the lterature on nsttutonal herdng. nsttutonal nvestors may trade together and wth past prce movements as a result of slowly 4

7 We focus on predctons about bubbles that call for the nteracton between nvestor groups. 6 The most basc conjecture about prce bubbles s that the overzealous tradng actvty of rratonal agents causes prces to dverge from fundamental value. The effcent markets response s that arbtrageurs quckly elmnate any prcng dscrepances. As descrbed by Fama (96), f there are many sophstcated traders n the market, however, they may cause these bubbles to burst before they have a chance to really get under way. However, DeLong et al. (990a) show that nose traders who follow past prce movements create nose-trader rsk the rsk that they may drve prces further away from fundamentals. Furthermore, Shlefer and Vshny (997) show that, wth delegated portfolo management and captal constrants, arbtrage s lmted. Funds may lose money and experence captal outflows precsely when the msprcngs on the securtes they are arbtragng are the largest. Thus, the above lterature argues that prce bubbles: a) wll not exst, or b) exst and are drven by nose traders wth sophstcated traders stayng relatvely neutral or takng bets aganst these agents. A more unconventonal vew of bubbles s that sophstcated traders actually help drve them. DeLong et al. (990b) show that ratonal speculators may actually start prce movements knowng that postve feedback traders wll follow and purchase these securtes at hgher prces tomorrow. Lkewse, n Abreu and Brunnermeer (2002, 2003) arbtrageurs know that the market s overvalued but maxmze profts by rdng the bubble. Because arbtrageurs face lmted captal and short-sale constrants, they can only burst a bubble when a coordnated sellng effort among arbtrageurs occurs. These models predct that sophstcated nvestors generally move n the same drecton as the prce bubbles. Smlarly, agency dffusng prvate nformaton (Froot, Scharfsten, and Sten (992), Hrshlefer, Subrahmanyam, and Ttman (994)) or to nfer nformaton from other traders (Bkhchandan, Hrshlefer, and Welch (992)). 6 Most ratonal bubble models n smultaneous or sequental markets do not rest on nvestor heterogenety. Santos and Woodford (997) argue that condtons used to generate ratonal bubbles are specal crcumstances. However, LeRoy (2003) revews these models and argues that the theoretcal condtons whch rule out ratonal bubble models are mplausble. He argues that evdence of nformed nvestors falng to move aganst prces s support of ratonal bubble models (see LeRoy, p. 6). Brunnermeer (200) provdes an evaluaton of bubble models, focusng on asymmetrc nformaton. Summares of related behavoral prcng ssues are also provded n Shlefer (2000), Hrshlefer (200) and De Bondt (2003). Emprcal connectons wth bubbles are dscussed n Shller (2003).

8 problems (as descrbed n Allen and Gale (2003)) can cause ratonal portfolo managers to take addtonal rsk and move asset prces to excessve levels. 7 B. Emprcal mplcatons n our emprcal analyss we relate the behavor of nsttutonal nvestors (more sophstcated) relatve to ndvdual (perhaps nose traders) to the above lterature along several dmensons. Frst, we examne ther overall level of actvty as proxed for by holdngs and tradng volume. Second, we examne whether nsttutonal or ndvdual nvestors react dfferently n response to past market movements. 8 We nterpret changes n ownershp n the same drecton as short-term past prce moves (trend chasng) as a behavor that lkely leads to larger prce swngs than those that mght occur n the absence of that nvestor group. Thrd, we also examne whether tradng n response to contemporaneous and past market returns changes durng the run-up and fall of Nasdaq. We are not the frst to examne the relaton between nsttutonal tradng and past prce movements. Grnblatt, Ttman, and Wermers (99) and Wermers (999) fnd evdence that mutual funds chase past ndvdual securty stock returns at the quarterly horzon. At shorter frequences, Odean (998), Choe, Kho, and Stulz (999), Barber and Odean (2002), and Grffn, Harrs, and Topaloglu (2003) fnd that ndvdual nvestors are often contrarans and more lkely to sell (purchase) a stock f ts prce recently decreased (ncreased). 9 These papers focus on the relaton between cross-sectonal frm performance and changes n ownershp and do not examne market-wde relatons. 0 f nsttutons or ndvduals keep ther net amount of nvestment fxed but reallocate nvestments between stocks, such as descrbed n the model of Barbers and Shlefer (2003), one would see lttle relaton between changes n ownershp and aggregate stock market return performance. 7 Smlarly, Allen, Morrs and Postlewate (993) and Allen and Gale (993) present models where bubbles arse due to asymmetrc nformaton and moral hazard. 8 We loosely refer to the Nasdaq 00 ndex as the market. Separatng effects on Nasdaq versus a broader ndex would lkely not be frutful as the correlaton between daly Nasdaq 00 and S&P 00 returns s over our sample perod of September, 999 to December 3, Sas, Starks, and Ttman (2000) and Ca, Kaul, and Zheng (200) use quarterly data but also nvestgate short-term trend chasng and prce mpact of nsttutonal nvestors. 0 For example, n Grffn, Harrs, and Topaloglu (2003) all changes n ownershp and returns are measured relatve to Nasdaq averages. Ownershp for a partcular frm s measured relatve to the average change n ownershp for all Nasdaq frms and abnormal returns are n excess of the Nasdaq ndex. 6

9 Recent evdence by Shapra and Veneza (2003) fnds that ndvdual nvestors n srael are market trend chasers and buy more than they sell followng market up-moves and professonal money managers are contraran wth respect to the srael market between However, n survey evdence Shller (999) documents that 6 (9) percent of U.S. ndvdual nvestors thought that a three percent one-day drop would be followed by an ncrease (decrease) n prces n 999 a large ncrease from 3 (34) percent n 989. Gven the common percepton by ndvdual nvestors that dps were buyng opportuntes n 999, one mght see net ndvdual buyng after downward prce moves. We study the relaton between leadng, contemporaneous, and lagged market returns and nsttutonal and ndvdual tradng actvty over our September 999 to December 200 sample of Nasdaq 00 ndex stocks.. Data The prmary data set for ths paper conssts of aggregate trades and quotes n Nasdaq 00 stocks from September, 999 to December 3, 200. We use only the Nasdaq 00 stocks to avod lqudty problems n less actve securtes and to make drect comparsons to the rse and fall of the ndex. The propretary data we use s drectly from Nasdaq clearng records and ncludes the date, tme, tcker symbol, trade sze and prce of each transacton for each Nasdaq 00 securty. The data also nclude addtonal dentfyng felds (related to the settlement process) about the partes nvolved n each trade. These addtonal felds nclude three features that allow us to assgn tradng volume to nsttutons and ndvduals. Frst, each trade s lnked to the partes (market maker or Electronc Communcaton Network (ECN)) on both sdes of the trade. Second, each sde of each trade s classfed as to whether the partes are tradng for ther own account (as a market maker) or are smply handlng a trade for a retal or nsttutonal clent (agency tradng). Thrd, each trade s marked as to whch party s buyng and sellng. Denns and Strckland (2002) examne the relaton between nsttutonal ownershp levels and the cross-sectonal volatlty from 988 to 996. They fnd that stocks wth hgher levels of nsttutonal ownershp move more than other stocks n the 30 days when the market ndex moves up or down more than two percent. However, nsttutons prefer lqud stocks (Falkensten (996)) that are more lkely to move more when the market moves and nsttutonal holdngs do not necessarly mply nsttutonal actvty. Boyer and Zheng (2002) document that mutual funds, foregners, and penson funds move wth the market on a quarterly bass from Cohen (999) fnds that nsttutons buy shares from ndvduals durng economc downturns also usng quarterly ownershp data. 7

10 Ths desgnaton helps us avod problematc trade classfcatons that commonly result from tck-test rules. Wth these three peces of propretary nformaton, tradng volume s assgned to brokerage houses that prmarly deal wth ndvdual nvestors, to brokerage houses prmarly handlng nsttutonal order flow, or to market makers. Although for ease of dscusson we refer to the data as nsttutonal and ndvdual, a more accurate but cumbersome classfcaton would be to denote the trades as those executed through brokerage houses prmarly dealng wth nsttutonal or ndvdual clents. We frst examne the relaton among ndvdual/nsttuton classfcaton and trade-sze groups as reported n Table A. The average ndvdual-to-market maker trade s for 483 shares and nsttuton-tomarket maker trades average,972 shares. A note of cauton s n order as the percent of volume by all partes s understated due to 9.43 percent of volume (8 percent of trades) that we are unable to classfy because of dscrepances n the trade routng records. We also use trade-sze groups. Small trades are those for fewer than 00 shares; medum-sze trades are from 00 to 0,000 shares; and trades for greater than 0,000 shares are classfed as large trades (Barclay and Warner (993)). Small, medum, and large trades consttute 66.6, 33.0, and 0.83 percent of all trades but 7.0, 4.44, and 28. percent of volume, respectvely. Whle some papers use the large/small trade dstncton as a proxy for nsttutonal and ndvdual tradng, t s mportant to note that ths procedure gnores a large proporton of medum-sze tradng, whch Barclay and Warner show s most lkely to contan nformaton. Fgure reports the percentage of the total volume n the small and large trade groups. ndvdual-to-ndvdual and ndvdual-to-market maker trades together account for percent of volume n trades for less than 00 shares, whereas nsttutons tradng wth other nsttutons or market makers accounts for 0.23 percent of all small-sze trade volume. Conversely, for large trades, ndvduals tradng wth market makers or other ndvduals account for 4.73 percent of tradng volume. Total nsttutonal tradng accounts for percent of tradng volume. Whle the assgnment of tradng volume s sure to capture some ndvduals tradng at nsttutonal brokerage houses and vce versa, f one 8

11 accepts that large blocks are more lkely to be orgnated by nsttutons and small trades by ndvduals, these fndngs strongly support our classfcaton as a useful mechansm for assgnng tradng volume. 2 To examne long-run trends n ownershp levels and to compare trends across markets, we compute aggregate ownershp from the 3F flngs compled on the Spectrum database. Spectrum classfes nsttutons nto fve groups: () banks, (2) nsurance companes, (3) mutual funds, (4) nvestment advsors, and () other (ncludng penson and endowment funds). 3 Note that 3F flngs are not requred for state penson funds, hedge funds, nsttutons wth less than $00 mllon under management, or for ndvdual securty postons below 0,000 shares or $200,000. Gven these lmtatons, holdngs from Spectrum should form a close but mperfect proxy for the level of quarterly holdngs. Grffn, Harrs, and Topaloglu (2003) fnd a strong relaton between quarterly changes n ownershp from Spectrum and the Nasdaq ndvdual/nsttutonal data. Furthermore, they fnd that Nasdaq nsttutonal ownershp changes are most strongly assocated wth mutual funds and the nvestment advsor categores.. Ownershp levels and aggregate actvty To determne the role of nsttutonal behavor over the long run, we examne three man questons: ) Do nsttutons ncrease or decrease ther holdngs n Nasdaq securtes over ts rse and fall? 2) s there a wde dfference between nsttutonal holdngs on Nasdaq and the overall market? 3) Do nsttutons or ndvduals domnate tradng actvty n the Nasdaq 00? We frst examne the overall level of ownershp on Nasdaq usng quarterly Spectrum holdngs from the begnnng of the Nasdaq ndex n 98 through 200. We then compare equal and valueweghted ownershp levels and ownershp levels by nsttutonal type to ownershp levels n all CRSP 2 Nevertheless, we also nvestgate our man fndngs usng the addtonal flter that nsttutonal trades must also be block trades and obtan smlar results. 3 The other category also ncludes foundatons, trusts, fnancal nsttutons, government, mscellaneous, and nonfnancal companes. 9

12 stocks and S&P 00 stocks. We also examne the fracton of volume classfed as nsttutonal or ndvdual over the September 999 to December 200 perod. A. Ownershp levels over tme Fgure 2, Panel A plots value-weghted ownershp levels along wth the correspondng ndex return for the Nasdaq 00, the CRSP value-weghted, and the S&P 00 ndex. f a partcular group of traders s responsble for the large prce ncreases on Nasdaq relatve to the S&P 00 then we mght expect dstnctly dfferent ownershp levels between the two ndces. Aggregate nsttutonal ownershp reported on Spectrum ncreases throughout the perod from late 98 through the md 990s. t s mportant to note that some of ths ncrease could be due to nsttutons and frms reachng thresholds where they are requred to report ther postons. 4 Gompers and Metrck (200) and Bennett, Sas, and Starks (2003), document these aggregate ncreases across all securtes from 980 to 997. By the end of 99 nsttutons held approxmately 8 percent of the Nasdaq 00 market captalzaton. Durng the ntal stages of the Nasdaq 00 run-up (from 996 to March 999), nsttutons generally decreased ther postons n Nasdaq ndex stocks, although the shft n ownershp was not smooth. However, when prces more than doubled durng the fnal stages of the Nasdaq 00 run-up from March 999 to March 2000, nsttutonal ownershp ncreased from 46 percent to 0. percent of the Nasdaq 00. Ofek and Rchardson (2003) fnd that the mean (medan) nsttutonal holdngs n nternet stocks s 3.33 (2.92) percent n March Our fndngs show that nsttutons held a much larger fracton of Nasdaq 00 securtes. Furthermore, durng the Nasdaq collapse nsttutons actually slghtly ncreased ther aggregate holdngs, consstent wth ncreased hedge fund ownershp n hgh growth stocks documented n Brunnermeer and Nagel (2003). We also compare the ownershp patterns to those from all CRSP securtes and the S&P 00. At the ncepton of the Nasdaq 00 nsttutons held approxmately the same percentage n Nasdaq stocks as 4 Frms wth less than $00 mllon under management are not requred to report nor are frms requred to report securty postons below 0,000 shares or $200,000. Obvously, more postons are reported followng upward stock prce moves and stock splts. They argue that the combnaton of low nsttutonal ownershp and the dffculty of shortng shares can lead to msprcng. 0

13 they dd n all CRSP stocks. By 996, Nasdaq ownershp ncreased to levels hgher than the average CRSP and S&P 00 stock. However, nsttutons reduced ther postons n Nasdaq 00 securtes n both 996 and 997 so that by March 3, 998 nsttutonal ownershp n the Nasdaq 00 was smlar to that of all other frms covered by Spectrum. From March of 998 to December 200, Nasdaq ownershp levels closely track all CRSP and S&P 00 securtes. We present equal-weghted levels of ownershp across securtes for the three ndces n Fgure 2, Panel B. Small stocks have substantally less nsttutonal ownershp. Panel B shows that the nferences made for Nasdaq 00 and S&P 00 securtes are not drven by a few large cap stocks n the ndces. On March 3, 2000 nsttutons held 8.7 of Nasdaq 00 shares on a equal-weghted bass, a number not far below the 6.0 percent held n the average S&P 00 securty. Snce we do not see large ownershp changes n 2000 t suggests that a combnaton of weak nsttutonal and ndvdual demand are lkely culprts for the burst. Fgure 3 detals the ownershp patterns of the fve dfferent nsttutonal groups for Nasdaq 00 securtes n Panel A and then compares those to S&P ownershp levels n Panel B. As Fgure 3 shows, mutual fund ownershp has generally ncreased from below fve percent n 98 to over percent n 200. Ths trend s present for both S&P and Nasdaq 00 stocks. Ownershp by nvestment advsors ncreases from 98 through 993 but falls dramatcally n Nasdaq stocks untl March 999. These advsors slghtly ncreased ther postons n Nasdaq stocks pror to and durng the Nasdaq crash. Other nsttutonal types saw less varaton from 98 through 200. From March 998 to December 200 most nsttuton types slghtly ncreased ther ownershp n Nasdaq 00 equtes except for a small decrease by banks. Partcularly from 998 to 200 ownershp trends n Nasdaq 00 and S&P 00 stocks were qute smlar, ndcatng that nsttutons dd not actvely underweght postons n rsng Nasdaq stocks. B. Tradng Volume nsttutons can own a large fracton of the Nasdaq market but f they trade nfrequently they wll lkely have lttle mpact on short-run prcng. We break down the proporton of ndvdual tradng volume

14 relatve to nsttutonal volume over the perod from September 999 through December 200 n Fgure 4. 6 nterestngly, up untl md-march of 2000, nsttutonal tradng volume was generally larger than the proporton of tradng volume due to ndvduals. Ofek and Rchardson (2003) argue that nternet stock volume, composed of few block trades wth a large proporton of ndvdual ownershp, was due to ndvdual nvestors. Pror to the market fall, ths was not the case for Nasdaq 00 stocks. After March 2000, nsttutons and ndvduals trade roughly equal share volumes, but n early 200 the fracton of nsttutonal tradng volume gradually decreases. Over the entre perod the shft s dramatc. Durng the frst two months of the sample, nsttutonal tradng represents 3.3 percent of nonmarket maker volume. Yet by the last two months of the sample perod nsttutons only account for 40.4 percent of volume. We also examne value-weghted proportons and fnd a smlar (but not as dramatc) ncrease n ndvdual volume n That nsttutons account for a majorty of the tradng volume and a large, growng fracton of holdngs pror to the Nasdaq bubble s peak suggests that ndvduals were not solely responsble for the Nasdaq run-up. V. Daly tradng behavor What s the nteracton between contemporaneous market returns and nsttutonal and ndvdual buyng? Are nsttutons or ndvduals more lkely to buy after hgh prevous-day stock returns? We examne the relaton between returns and ownershp type over the whole perod as well as whether the nature of the relatons changed across the rse and fall of the market. We also brefly examne whether the relaton s dfferent for market up and down moves. A. Whole Perod Results We calculate the tradng mbalance for each stock as the dfference between the buy and sell volumes each day scaled by the number of shares outstandng. We then examne equal- and valueweghted averages across stocks to obtan a relatve measure of the magntude of the nsttutonal or 6 As shown n Table A, 2.86 percent of volume s due to market makers and 9.43 percent s unassgned. 7 n unreported results we also examne both equal- and value-weghted turnover for the Nasdaq 00 and fnd dramatc turnover ncreases n late 2000 and 200 but 200 value-weghted turnover s smlar to earler years. 2

15 ndvdual tradng mbalance. Unless otherwse noted, we refer to ths as the mbalance. Snce for every buyer there must be a seller, n the absence of a market maker, net nsttutonal buyng actvty would perfectly offset ndvdual sellng actvty. However, gven that we cannot assgn 9.43 percent of trades and market makers do partcpate, the nsttutonal and ndvdual mbalances are close substtutes but not perfectly negatvely correlated. Whle we focus on the daly nsttutonal buy-sell mbalance measure, one could also nterpret fndngs from the ndvdual sell-buy mbalance perspectve. To examne the relaton between nsttutonal/ndvdual actvty and Nasdaq 00 returns, we dvde all tradng days nto sx categores, those wth returns: less than negatve two percent; n four ncreasng one percent ntervals; and those wth greater than two percent. Over our sample perod, Nasdaq 00 returns are extremely volatle wth 2. percent of days havng returns greater than two percent and 27.2 percent of days havng returns less than negatve two percent. Table reports the average returns, equal-weghted nsttutonal mbalance, dollar value of the total mbalances, as well as the buy-sell volume as a fracton of total volume. We examne all fgures on the day of the return move as well as the two days before and after. nsttutonal tradng actvty exhbts a strong contemporaneous relaton wth market returns days wth the lowest returns have the largest nsttutonal sellng. However, there s an even stronger relaton between returns and next-day nsttutonal tradng. For returns moves of less than negatve two percent, the mbalance of 9.98 ndcates that percent of total outstandng shares are sold by nsttutons, whereas for return moves of greater than two percent, of the typcal Nasdaq 00 securty s bought by nsttutons. However, nsttutons sell percent of shares on the next day for large negatve market moves, and buy 0.03 percent of shares on the day followng large market upmoves. Dfferences n the net nsttutonal buyng for wnners and losers are hghly statstcally and economcally sgnfcant. Usng begnnng of the day market values, we also estmate the dollar values of the net actvty. The next panel shows that on days followng market down-moves, nsttutons sell (and ndvduals approxmately buy) slghtly more than half-a-bllon dollars of Nasdaq shares ( mllon) and 3

16 followng market up-moves nsttutons buy 88. mllon dollars worth of securtes. Two days followng the return move there s lttle dfference n the behavor of nsttutons and ndvduals. We also examne nsttutonal buy-sell volume scaled by total volume. These patterns show a weaker contemporaneous daly relaton between nsttutonal tradng and stock prces but an even stronger relaton on the followng day. On the day followng a large negatve return move, -0.8 percent of the daly volume s nsttutonal sellng (to ndvduals). For large market up-moves,.9 percent of the followng day s volume conssts of net nsttutonal buyng. Snce followng past prces n a postve fashon can lead to larger prce moves than n the absence of trend chasng, the evdence here does not support the noton that ndvduals were entrely responsble for excess prce movements durng the Nasdaq bubble. Aggregate nsttutonal tradng mbalances are correlated wth both pror day returns and mbalances. Therefore, t s not clear f mbalances are assocated wth prevous day returns or whether ths s merely an artfact of nsttutonal tradng beng correlated across days. To nvestgate these nferences jontly and more rgorously, we estmate a VAR model for both equal- and value-weghted aggregate buy-sell mbalances and Nasdaq 00 ndex returns. We standardze both varables by ther tme-seres mean and standard devaton for ease of nterpretaton and estmate the followng system of structural VAR equatons: R t α + β Rt + λ t + δ t, R = () t α + β 0Rt + β Rt + λ t + δ t, = (2) where R t s the return on the Nasdaq 00 ndex on date t and t s the nsttutonal buy-sell mbalance at tme t. We estmate the system wth fve lags. 8 8 The Akake nformaton crteron tests shows that fve lags s the approprate lag-length for the equal-weghted VARs and fnds four lags for the value-weghted regressons. For consstency, we use fve lags n all specfcatons. Rather than usng a VAR, some prevous lterature such as Warther (99) uses expected and unexpected flows from a frst-pass estmaton. However, the VAR approach has the advantage of smultaneously modelng the persstence of mbalances wthn the VAR (rather 4

17 Panel A of Table dsplays results for equal-weghted market returns and equal-weghted aggregate nsttutonal tradng mbalances for a standard VAR wthout contemporaneous effects. n the mbalance equaton, a one standard devaton ncrease n nsttutonal mbalances s postvely assocated wth the prevous two days nsttutonal mbalances. nsttutons systematcally move n and out of Nasdaq 00 stocks n opposte drecton over wndows averagng three days. The return equaton shows that nsttutonal mbalances have some ablty to predct next-day returns. However, the predctablty n returns s small compared to that n mbalances. The adjusted R 2 s for the return equaton ndcate that roughly only 2.2 percent of the varance of returns s explaned by past returns and mbalances. n Panel B, we use a structural VAR model where nsttutonal mbalances depend on past mbalances and both contemporaneous and lagged returns. The contemporaneous return s ncluded n the structural VAR to jontly estmate the contemporaneous relaton between mbalances and returns. We are not assumng causalty. A one standard devaton ncrease n returns s assocated wth a contemporaneous 0.36 standard devaton ncrease n daly nsttutonal buyng and a shock to the prevous-day return leads to a 0.48 standard devaton ncrease n net nsttutonal buyng. After controllng for past mbalances, mbalances are even slghtly more strongly assocated wth the prevous day s return than the contemporaneous return. One standard devaton shocks to lagged two- and three-day returns are assocated wth a 0.09 and 0.3 standard devaton decrease n nsttutonal net actvty. We also examne value-weghted Nasdaq returns and value-weghted nsttutonal mbalances n Panel C of Table. The value-weghted Nasdaq 00 s nterestng to examne snce t s a wdely traded asset. Agan, past returns are an mportant determnant of mbalances, wth a coeffcent of 0.4 on the contemporaneous return and a coeffcent of 0.47 on the lagged return. nsttutons (ndvduals) move nto (out of) the market on the day of and followng a market ncrease. than from a frst-pass regresson) and avodng the error-n-varables problem that arse from usng estmated regressors as dependent varables n a second-stage regresson.

18 B. Tradng Behavor n Dfferent Market Envronments The trend chasng behavor observed thus far suggests that nsttutons could be at least partally responsble for the Nasdaq bubble. Yet, the tmng of ther tradng actvty s mportant. n partcular, f nsttutons prmarly follow negatve prce moves durng the run-up and correcton phase of the bubble, then ther tradng behavor could take prces towards fundamental values qucker than n the absence of ther tradng. Conversely, f they engage n trend chasng durng the run-up, then ther behavor may lead to hgher prces than would have occurred n ther absence. We examne these ssues by studyng the nsttutonal and ndvdual tradng behavor durng the Nasdaq run-up, crash and subsequent perods separately. We defne the run-up as the perod from September, 999 (the begnnng of our sample) to the peak of the Nasdaq 00 on March 27, 2000 and the crash perod from March 28, 2000 to Aprl 4, 200. The more-stable perod starts on Aprl, 200 and runs untl the end of our sample on December 3, 200. The more-stable perod s pcked from Aprl, 200 onwards because the market rebounded for fve months before gong below levels set on Aprl 4, 200. Table examnes the relaton between returns and contemporaneous daly mbalances as well as the mbalances on the two days before and two days after a return move durng the three sub-perods. Table shows that there s a relatvely weak relaton between returns and contemporaneous nsttutonal mbalances n the run-up perod. However, there appears to be a strong relaton between returns and nextday nsttutonal mbalances. For the largest negatve return moves, nsttutons sell percent of the shares to ndvduals whereas for the most postve return moves nsttutons buy percent of the shares. n the crash and more-stable perods, the relaton between returns and the next day s nsttutonal buyng s even stronger. For return moves less than mnus two percent, n the crash (more-stable) perod nsttutons sell (0.0462) percent of the shares to ndvduals whereas for return moves greater than two percent nsttutons buy (0.0339) percent of the shares. n all three perods the nsttutonal 6

19 tradng n relaton to returns s greater on the day after the return move than on the same day. No statstcally sgnfcant patterns are observed after two days. n unreported results we also examne estmates of the dollar values of tradng relatve to return moves. The three perods are not drectly comparable because the market value of Nasdaq falls rapdly over the second sub-perod but the three perods can be used as a gauge of the overall economc mportance. Followng large negatve market moves, roughly 472 mllon dollars are sold n the run-up perod, 82 n the crash perod, and 390 n the more-stable perod. The dollar value of purchases followng market up-moves s even greater n each perod. Overall, the sortng evdence suggests that nsttutons are strong trend chasers throughout the sample perod. We formally test whether the dynamcs of nsttutonal/ndvdual tradng has changed throughout the perod usng a VAR. Panel A of Table V reports VAR results by sub-perod. Consstent wth our sortng results, the contemporaneous relaton between nsttutonal tradng and returns s weakest n the market run-up perod (wth a 0.7 coeffcent). However, the relaton between returns and the next day s nsttutonal mbalance s strong n ths perod (wth a coeffcent of 0.47). n the crash (more-stable) perod the contemporaneous return relaton s 0.34 (0.44). Smlarly, a one standard devaton shock to returns leads to a 0.4 ncrease n the followng day s nsttutonal buyng durng the crash perod and an ncrease of 0.2 n the more-stable perod. nsttutonal mbalances are negatvely related to the pror two- and three-day returns n all three perods but the relaton s only sgnfcant n the fnal sub-perod. n the crash perod, there s some weak evdence that nsttutonal mbalances foreshadow next-day market return movements. To judge the dfference among perods, we estmate a sngle regresson over the entre sample and nteract dummy varables wth returns and nsttutonal mbalances for both the run-up and crash perod. We nterpret all varables relatve to the more-stable perod. n the mbalance equaton, Panel B shows that nsttutons moved less wth the market durng the run-up perod than n the more-stable perod (coeffcent of -0.2). Ths coeffcent s also negatve although nsgnfcant durng the crash perod. We fnd no sgnfcant dfference n trend chasng actvty relatve to the more-stable perod. Overall, 7

20 nsttutonal trend chasng s qute strong throughout the perod. Rather than provdng a stablzng force for the market, nsttutons strongly followed past prces even durng the market run-up perod a practce that enhances any potental prce dstortons. One potental ssue s whether there are asymmetres n the relaton between both postve and negatve return moves and nsttutonal tradng. n partcular, t may be the case that nsttutonal tradng follows postve but not negatve market moves or vce versa. n Table V we examne ths wth daly VAR regressons by sub-perod where returns are nteracted wth a dummy varable (equal to one when returns are negatve). 9 Although the dfferences are not statstcally sgnfcant, n the run-up perod nsttutons dd not move wth market down moves as much as they dd wth up moves. n contrast, there s weak evdence that nsttutonal sellng was more rampant on market down days durng the crash perod. Such s not the case n the more stable perod where nsttutons sold sgnfcantly less on market down days than they bought on market up days. Durng both the run-up and crash perods, nsttutons sold slghtly more on the day followng up-moves than they bought followng down-moves. Overall, the evdence that nsttutons moved wth the market on the day of and the day followng market up-moves n the pre-crash perod suggests that nsttutons helped drve Nasdaq s rse. V. nsttutonal Actvty and Mutual Fund Flows One queston rased by our analyss s whether the nsttutonal actvty seen n Nasdaq mbalances merely reflects nsttutonal managers reactng to aggregate nflows and outflows from ndvduals n professonally managed funds. f ndvdual nvestors sell stocks n ther ndvdual accounts followng large up-moves and at the same tme buy securtes va mutual funds then the net market exposure of ndvdual nvestors could be unchanged. ndeed, from February 998 through June 999, Edelen and Warner (200) document a postve contemporaneous relaton between daly mutual fund net flows and NYSE ndex returns and an even stronger postve relaton between net flows and the 9 The coeffcents on returns nteracted wth the dummy varables for negatve return days can be nterpreted as the ncremental mpact of returns on negatve return days as compared to those day zero returns that are postve. 8

21 prevous day s market return. Goetzmann and Massa (999) examne flows n and out of several large S&P ndex funds and fnd a contemporaneous relaton, but lttle lagged relaton, between daly fund flows and returns. They argue that these nflows may be responsble for the large S&P ndex prce ncrease. We obtan aggregate net flows nto U.S. mutual funds from TrmTabs. Both Edelen and Warner (200) and Goetzmann, Rouwenhorst, and vkovch (200) fnd that the accuracy and tmng of the TrmTabs reportng s hgh, but TrmTabs covers just to 20 percent of mutual fund assets. Because these flows only represent mutual funds and our mbalance data conssts of all nsttutonal brokerage houses, our examnaton of the market s somewhat ncomplete. nvestors are thought to be more actve n redeemng and purchasng mutual funds compared to other types of nsttutonal captal (.e., managed accounts, penson or hedge funds). ndeed Grffn, Harrs, and Topaloglu (2003) show that cross-sectonal changes n mutual fund holdngs have the strongest relaton of all nvestor types to quarterly mbalances computed from smlar Nasdaq data for ndvdual securtes. Emprcally, we estmate a VAR wth returns, nsttutonal mbalances and mutual fund flows together. The model conssts of the followng three equatons: R t α + β Rt + λ t + γ Ft + δ t, R = (3) t α + β 0Rt + β Rt + λ t + γ 0Ft + γ Ft + δ t, = (4) F t α + β 0Rt + β Rt + λ t + γ Ft + δ t, = () where F t s the mutual fund flows scaled by the prevous month s aggregate S&P market captalzaton. Returns and mbalances are defned as above. We nclude contemporaneous flows n the mbalance equaton to test whether there stll exsts a postve contemporaneous (and lagged) relaton between mbalances and past returns after controllng for contemporaneous flows. Because flows are value-weghted, Table V presents the results of the value-weghted VAR model. To see whether our nsttutonal mbalance measure captures ndvdual nvestors tradng through 9

22 nsttutonal channels, we examne the relaton between contemporaneous or lagged fund flows and our mbalance measure. The relaton between returns and mbalances reman relatvely unchanged but several other results emerge. Frst, fund flows are margnally contemporaneously correlated wth our nsttutonal mbalance measure (p-value of 0.). Lagged flows, however, have lttle sgnfcant relaton wth nsttutonal mbalances. Second, the magntude and sgnfcance of the return and mbalance coeffcents reman largely unchanged even after ncludng contemporaneous and lagged mutual fund flows n the mbalance equaton. A one standard devaton ncrease n returns s stll assocated wth a contemporaneous 0.4 standard devaton ncrease n daly nsttutonal buyng, dentcal to the coeffcent found wthout controllng for fund flows n Panel C of Table. Smlar results hold for prevous-day returns (coeffcent of 0.42 vs n Panel C of Table ). Past returns reman mportant determnants of nsttutonal mbalances even after controllng for mutual fund flows. Return lags 2 to are negatve and half sgnfcantly so, presentng some evdence that nsttutons are somewhat contraran at longer lags. Thrd, the mportance of the prevous day s returns for explanng flows s strkngly smlar to that found n Edelen and Warner (200). Mutual fund flows are negatvely related to past one- and twoday flows but sgnfcantly postvely correlated at three- to fve-day horzons. n addton, there s lttle evdence that mutual fund flows have any relaton to future Nasdaq 00 returns. n unreported results, we demean the flow varables by day-of-the-week ndcators but found lttle change n the VAR coeffcents. 20 We also examne whether these relatons change across the three sub-perods. Overall our man fndngs are smlar controllng for mutual fund flows does not substantally reduce the strong relaton between mbalances and contemporaneous or prevous-day returns n any of the three subperods. Although the ndvdual VAR coeffcents are nformatve, they only measure the statc lead-lag relatonshps among nsttutonal mbalances, mutual fund flows and returns. The jont mpact of past 20 Mutual fund flows are much greater on Frdays and somewhat larger on Mondays and Tuesdays but exhbt no clear day-of-the-month patterns. 20

23 shocks s more mportant for determnng the dynamc behavor of the system. Through the use of mpulse response functons we trace out the cumulatve mpact of an nnovaton n nsttutonal mbalances, mutual fund flows, and returns on subsequent mbalances, flows, and returns over tme. Usng equatons (3) through () wthout the contemporaneous varables, we examne the cumulatve mpact for ten days wth two standard devaton confdence ntervals. Fgure shows that returns have a strong mpact on nsttutonal mbalances at lag one but the effect reverses and goes to zero n the subsequent perods. Ths reversal, consstent wth the negatve relaton between mbalances and returns at lags 2 through shown n Table V, ndcates that the shortterm ncrease n nsttutonal buyng n response to the prevous day s stock return s counteracted by nsttutonal sellng. n contrast, shocks to returns lead to an ncrease n mutual fund nflows, yet only part of the flow ncrease s transtory. Even at lag ten, a one standard devaton ncrease n returns leads to a 0.42 percent ncrease n mutual fund flows. There are few short-term dynamc effects of mutual fund flows on nsttutonal mbalances and nsgnfcantly postve effects at longer horzons. As expected, there s no relaton between nsttutonal actvty and future mutual fund flows. The VAR system shows no mpact of mutual fund flows on returns but nsttutonal mbalances seem to have a small but sgnfcant mpact on returns. Overall, these results renforce the concluson that mutual fund flows do not have a large effect on mbalances even though both mbalances and mutual fund flows are strongly related to past Nasdaq 00 returns. nterestngly, whle a shock to returns has a lastng effect on mutual fund flows, the postve relaton between returns and nsttutonal mbalances s transtory. V. ntradaly nvestor Tradng Our prevous analyss shows a strong contemporaneous postve relaton between daly returns and nsttutonal mbalances. However, t s unclear what s occurrng wthn the day. f ndvduals act as nose traders usng extraneous nformaton or past prces then one mght expect to see that nsttutons ncorporate news nto prces wth ndvduals followng these prce movements. n contrast, theoretcal 2

24 models such as Abreu and Brunnermeer (2002) call for the smart money to watch the drecton of tradng by the naïve nvestors and then follow sut. To examne the ntraday relaton, we dvde each tradng day nto 78 fve-mnute ntervals from 9:30 a.m. to 4:00 p.m. We use the prevalng nsde bd and ask quotes to calculate the bd-ask mdponts and construct mdpont returns at fve-mnute ntervals. We lag the bd-ask mdponts by two seconds before computng the returns snce nternal Nasdaq analyss ndcates that trades are on average reported two seconds later than quotes. We defne buy-sell mbalances as the dfference between the buy and sell volumes for each fve-mnute nterval scaled by the total number of outstandng shares. Equal-weghted aggregate nsttutonal mbalances and returns are constructed by averagng across all 00 Nasdaq stocks for each of the fve-mnute ntervals. Because market makers take postons throughout the day, ntradaly nsttutonal and ndvdual mbalances are not essentally offsettng so we examne both nsttutonal and ndvdual mbalances. We estmate standardzed ntradaly VARs wth sx lags (30-mnutes) for returns and nsttutonal and ndvdual buy-sell mbalances durng the Nasdaq run-up, crash, and more-stable perods. To avod crossng day boundares for lagged varables, we exclude the frst half hour of each tradng day. Table V shows a strong negatve relaton between contemporaneous nsttutonal buy-sell mbalances and returns and a strong and economcally large relaton between ndvdual mbalances and contemporaneous market returns. A one standard devaton ncrease n return s contemporaneously assocated wth a 0.4 standard devaton decrease n net nsttutonal buyng and a 0.74 standard devaton ncrease n ndvdual buyng. Ths evdence s nconsstent wth models such as DeLong et al. (990b) where nsttutons move prces and ndvduals follow. However, the relaton reverses quckly wth nsttutonal mbalances postvely related to past return moves and ndvdual tradng negatvely related to the prevous 20-mnute returns. nsttutonal net buyng s negatvely related to the past fve-mnute nsttutonal mbalance but postvely related to all other nsttutonal mbalances. ndvdual mbalances have an economcally small negatve relaton to nsttutonal tradng over the prevous 30-mnute perod and an economcally larger postve relaton to 22

25 past ndvdual tradng over the precedng 20-mnute wndow. n general, even after controllng for past returns, there s persstence n the trade executons of both nsttutonal and ndvdual brokerage houses. f one group trades on news faster than another group then one mght expect the tradng of that partcular group to lead prces. Examnng the return equaton n the run-up perod shows that there s lttle evdence of ether nsttutonal or ndvdual mbalances leadng the market. We also examne the crash and more-stable perod. n both there s an economcally large contemporaneous negatve relaton between returns and nsttutonal net buyng and a postve relaton between returns and ndvdual buyng. Both perods also show nsttutons buyng and ndvduals sellng followng postve return moves. n general, ndvduals trade ether wth news or for other reasons and move the market. nsttutons seem to follow the market, or alternatvely, are not able to effectvely mask ther tradng ntent and suffer prce mpact accordngly. Ths behavor of nsttutons chasng past prces durng all stages of the bubble potentally exacerbates stock prce movements. V. Robustness ssues A. Tmng of Trades Our results all show strong evdence of nsttutonal trades beng executed n the same drecton as the prevous day s market moves. Such actvty could be due to nsttutons makng decsons to trade after the market moves or due to nsttutonal orders movng the market and then orders beng executed. Market wde prce pressure by nsttutons would predct that the contemporaneous effect should be much stronger than the lagged effect we observe the opposte. Nevertheless, we examne ths relaton by usng sortng results smlar to those n Table except that we also requre nsttutons to move n the opposte drecton of the market on day zero. nterestngly, for days where the market drops by more than two percent and nsttutons are net buyers, nsttutons are stll strong net sellers on the followng day. Smlarly, nsttutons are strong net buyers followng days where the market goes up accompaned by nsttutonal net sellng. 23

26 Ths evdence agan suggests nventory rebalancng s an ncomplete explanaton and that nsttutons are makng actve decsons to trade n response to market returns or the nformaton assocated wth those market moves. One possble explanaton why nsttutons mght follow market moves (deemed probable by some leadng practtoners) s that peer benchmarkng leads managers to shft from cash nto equtes when the market rses, hedgng the rsk of relatve underperformance. 2 B. Trade Classfcaton Snce our method of dentfyng trades may capture some ndvduals who trade at nsttutonal brokerage houses, we replcate our key results usng the more strngent classfcaton that trades executed through an nsttutonal brokerage house must also be for amounts greater than 0,000 shares. The fndngs n Table V are comparable to those n Panel B of Table. They show a slghtly weaker contemporaneous relaton but a nearly dentcal relaton between mbalances and past returns. Results n Tables V and V are also re-examned and show that our man fndngs are robust to the method of trade classfcaton. V. Concluson Ths paper examnes the tradng behavor of nsttutonal and ndvdual nvestors surroundng the rse and fall of Nasdaq. We fnd that nsttutons held a large fracton of the Nasdaq market captalzaton and were also responsble for more tradng volume than ndvduals durng the market run-up. On the day of and the day followng a large postve return, nsttutons buy. Conversely, ndvduals sell when the market rses and buy the day after the market dps. Overall, a one standard devaton ncrease n 2 Mechancal constrants mght also mpose restrctons on managers. However, the most probable constrants seem to actually lead to an opposte effect from the behavor we observe. Suppose a manager thnks the market s overvalued and would lke to mantan a 0 percent weghtng. Due to restrctons mposed n the fund s prospectus must mantan at least 80 percent of captal n equtes and hence s currently at the lower 80 percent allocaton lmt. A market ncrease would ncrease the manager s equty allocaton above 80 percent and hence a bearsh manager could now sell equtes. n contrast, f a manager s judgng themselves relatve to other managers who are more heavly weghted n equtes than the market up-move mght cause the manager who s underweghted n the market to move more money nto equtes out of fear of beng left behnd the pack. Nevertheless, t s nterestng to note that ths decson s an actve one on the part of the manager and not nduced by a (legal) constrant on holdngs. 24

27 yesterday s market return s followed by an economcally large 0.48 standard devaton ncrease n today s net nsttutonal buyng. Ths pattern of strong nsttutonal daly trend chasng behavor s present throughout the Nasdaq run-up and collapse. n addton these patterns are symmetrc wth respect to both postve and negatve returns. We examne whether the patterns reflect the behavor of nsttutonal nvestors or f nsttutons (lke mutual funds) merely react to ndvdual nvestor demand va ther funds. Although we fnd that nsttutonal actvty s related to daly mutual fund flows, these flows explan only a small proporton of short-term nsttutonal actvty and cannot explan the strong daly contemporaneous and lagged relaton between nsttutonal mbalances and returns. Furthermore, we fnd an extremely strong postve relaton between ntradaly ndvdual nvestor tradng and stock returns n the same fve-mnute nterval. However, nsttutonal trades soon follow recent prce changes and subsequent ndvdual trades move aganst the past prces. Ths actvty s pervasve durng the Nasdaq run-up, crash, and more-stable perods. Overall our results do not support models where ndvduals move market-wde prces and nsttutons (smart money) ether passvely stand by or actvely move aganst nose traders. Our evdence s most consstent wth models where nsttutons, at least over short one-day ntervals, follow recent market moves a practce whch may exacerbate prce trends. Future research should examne the underlyng motvatons behnd the complex nteractons between ndvdual and nsttutonal traders. 2

28 Appendx Based on nternal conversatons wthn Nasdaq, nsttutonal brokers, wrehouses (e.g., Salomon Smth Barney and Morgan Stanley) and nstnet are classfed as prmarly handlng nsttutonal order flow. Other ECNs (excludng nstnet), regonal frms and wholesalers (e.g., Schwab and Natonal Fnancal Servces Corporaton) are classfed as prmarly handlng ndvdual order flow. Small frms and the two regonal exchanges consttute about 8 percent of the total trades over the perod and contan a mx of ndvdual and nsttutonal tradng volume. These brokerage houses are assgned as prmarly nsttutonal or ndvdual tradng houses based on ther overall dstrbuton of tradng volume over the perod. Small frm and regonal exchanges are classfed as an nsttutonal dealer f the thrd quartle of trade-sze dstrbuton s,000 shares or greater. The,000-share cutoff s consstent wth the thrd quartle of trade-sze dstrbuton for nsttutonal brokers, wrehouses, and nstnet whereas handlers of ndvdual order flow (other ECNs, regonal frms and wholesalers) have trade szes smaller than 600 shares at the thrd quartle of the trade sze dstrbuton. Further analyss of the data s avalable n Grffn, Harrs, and Topaloglu (2003). We note that 9.43 percent of volume cannot be assgned to ether an nsttuton or ndvdual brokerage because of nconsstences n the trade routng process. Because a dsproportonate number of these unclassfed trades are sell trades, we typcally observe slghtly more buy than sell trades n the ndvdual (and possbly n the nsttutonal) categores. Snce the proporton of unclassfed trades s small relatve to the total daly mbalances t seems unlkely to bear much mpact on our daly regresson results. However, because the buy trade bas affects cumulatve ownershp levels we do not cumulate ownershp by nvestment type. 26

29 References Abreu, Dlp, and Marcus K. Brunnermeer, 2002, Synchronzaton rsk and delayed arbtrage, Journal of Fnancal Economcs 66, Abreu, Dlp, and Marcus K. Brunnermeer, 2003, Bubbles and crashes, Econometrca 7, Allen, Frankln and Douglas Gale, 993, Churnng bubbles, Revew of Economc Studes 60, Allen, Frankln and Douglas Gale, 2003, Asset prce bubbles and stock market nterlnkages, n Asset Prce Bubbles, edted by W.C. Hunter, G.G. Kaufman and M. Pomerleano, MT Press, Allen, Frankln, Stephen Morrs, and Andrew Postlewate, 993, Fnte bubbles wth short sale constrants and asymmetrc nformaton, Journal of Economc Theory 6, Barber, Brad M., and Terrance Odean, 2002, All that gltters: The effect of attenton and news on the buyng behavor of ndvdual and nsttutonal nvestors, Workng paper, Unversty of Calforna, Davs. Barbers, Nck, and Andre Shlefer, 2003, Style nvestng, forthcomng n Journal of Fnancal Economcs. Barclay, Mchael J., and Jerold B. Warner, 993, Stealth tradng and volatlty: Whch trades move prces? Journal of Fnancal Economcs 34, Bennett, James, Rchard Sas, and Laura Starks, 2003, Greener pastures and the mpact of dynamc nsttutonal preferences, forthcomng n the Revew of Fnancal Studes. Bkhchandan, Sushl, Davd Hrshlefer, and vo Welch, 992, A theory of fads, fashons, custom and cultural change as nformatonal cascades, Journal of Poltcal Economy 00, Boyer, Bran and Lu Zheng, 2002, Who moves the market? A study of stock prces and nvestment cashflows, Workng Paper, Unversty of Mchgan. Brunnermeer, Marcus K., 200, Asset Prcng under Asymmetrc nformaton: Bubbles, crashes, techncal analyss and herdng, Oxford Unversty Press, Oxford. Brunnermeer, Marcus K. and Stefan Nagel, 2003, Arbtrage and ts lmts: hedge funds and the technology bubble, Workng paper, Prnceton Unversty. Ca, Fang, Gautam Kaul, and Lu Zheng, 200, nsttutonal tradng and stock returns, Workng paper, Unversty of Mchgan. Chen, Hsu-Lang, Narasmhan Jegadeesh, and Russ Wermers, 2000, The value of actve mutual fund management: An examnaton of the stockholdngs and trades of fund managers, Journal of Fnance and Quanttatve Analyss 3, Choe, Hyuk, Bong-Chan Kho, and Rene M. Stulz, 999, Do foregn nvestors destablze stock markets? The Korean experence n 997, Journal of Fnancal Economcs 4,

30 Cohen, Randolph, 999, Asset allocaton decsons of ndvduals and nsttutons, Workng paper, Harvard Busness School. Coval, Joshua D., Davd Hrshlefer, and Tyler Shumway, 2002, Can ndvdual nvestors beat the market? Workng paper, Unversty of Mchgan. De Bondt, Werner, 2003, Bubble psychology, n Asset Prce Bubbles, edted by W.C. Hunter, G.G. Kaufman and M. Pomerleano, MT Press, DeLong, J. Bradford, Andre Shlefer, Lawrence H. Summers, and Robert Waldmann, 990a, Nose trader rsk n fnancal markets, Journal of Poltcal Economy 98, DeLong, J. Bradford, Andre Shlefer, Lawrence H. Summers, and Robert Waldmann, 990b, Postve feedback nvestment strateges and destablzng ratonal speculaton, Journal of Fnance 4, Denns, P., and D. Strckland, 2002, Who blnks n volatle markets, ndvduals or nsttutons? Journal of Fnance 7, Edelen, Roger M., and Jerold B. Warner, 200, Aggregate prce effects of nsttutonal tradng: A study of mutual fund flow and market returns, Journal of Fnancal Economcs 9, Falkensten, E. G., 996, Preferences for stock characterstcs as revealed by mutual fund portfolo holdngs, Journal of Fnance, -3. Fama, Eugene F., 96, The behavor of stock-market prces, Journal of Busness, 38, Fredman, Mlton, 93, The case for flexble exchange rates, n: Essays n postve economcs, Unversty of Chcago Press, Chcago. Froot, Kenneth A., Davd S. Scharfsten, and Jeremy C. Sten, 992, Herd on the street: nformatonal neffcences n a market wth short-term speculaton, Journal of Fnance 47, Garber, Peter M., 989, Tulpmana, Journal of Poltcal Economy 97, Goetzmann, Wllam, and Massmo Massa, 999, ndex funds and stock market growth, Workng paper, Yale Unversty. Goetzmann, Wllam, Geert Rouwenhorst, Zoran vkovch, 200, Day tradng nternatonal mutual funds: Evdence and polcy solutons, Journal of Fnancal and Quanttatve Analyss 36, Gompers, Paul A., and Andrew Metrck, 200, nsttutonal nvestors and equty prces, Quarterly Journal of Economcs 6, Grffn, John M., Jeffrey H. Harrs, and Selm Topaloglu, 2003, The dynamcs of nsttutonal and ndvdual tradng, forthcomng n Journal of Fnance. Grnblatt, Mark, Sherdan Ttman, and Russ Wermers, 99, Momentum nvestment strateges, portfolo performance, and herdng: A study of mutual fund behavor, Amercan Economc Revew 8,

31 Hrshlefer, Davd, 200, nvestor psychology and asset prces, Journal of Fnance 6, Hrshlefer, Davd, Avandhar Subrahmanyam and Sherdan Ttman, 994, Securty analyss and tradng patterns when some nvestors receve prvate nformaton before others, Journal of Fnance 48, Kndleberger Charles P., 978, Manas, pancs, and crashes: A hstory of fnancal crses, Macmllan, London, 3rd edton, 996. LeRoy, Stephen, 2003, Ratonal exuberance, Workng paper, Unversty of Calforna at Santa Barbara. Odean, Terrance, 998, Are nvestors reluctant to realze ther losses? Journal of Fnance 3, Odean, Terrance, 999, Do nvestors trade too much? Amercan Economc Revew 89, Ofek, El and Matthew Rchardson, 2003, DotCom mana: The rse and fall of nternet stock prces,, Journal of Fnance 8, Ofek, El and Matthew Rchardson, 2002, The valuaton and market ratonalty of nternet stock prces, Oxford Revew of Economc Polcy 8, Santos, Manuel, and Mchael Woodford, 997, Ratonal asset prcng bubbles, Econometrca 6, 9-7. Shapra, Zur, and tzhak Veneza, 2003, On weekend effects, tmng, and herdng: Do professonal nvestors behave dfferently from amateurs?, Workng paper, Yale Unversty. Shller, Robert J., 999, Measurng Bubble Expectatons and nvestor Confdence, Workng paper, Yale Unversty. Shller, Robert J., 2000, rratonal exuberance, Prnceton Unversty Press, Prnceton, New Jersey. Shller, Robert J., 2003, From effcent markets theory to behavoral fnance, Journal of Economc Perspectves Shlefer, Andre, 2000, neffcent markets - An ntroducton to behavoral fnance, Oxford Unversty Press, Oxford. Shlefer, Andre, and R. W. Vshny, 997, The lmts of arbtrage, Journal of Fnance 2, 3 -. Sas, Rchard W., Laura T. Starks, and Sherdan Ttman, 2000, The prce mpact of nsttutonal tradng, Workng paper, Unversty of Texas. Warther, Vncent A., 99, Aggregate mutual fund flows and stock returns, Journal of Fnancal Economcs 39, Wermers, Russ, 999, Mutual fund herdng and the mpact on stock prces, Journal of Fnance 4,

32 Table Varous Characterstcs for Return Groups Each day from September 3, 999 to December 27, 200 s assgned to one of sx groups based on the equal-weghted average return on the stocks comprsng the Nasdaq 00 ndex. Ths table reports the equal-weghted average return, equal-weghted nsttutonal buy-sell mbalance, nsttutonal net buyng and equal-weghted nsttutonal net volume for the fve tradng days centered around the rankng day. For each stock, nsttutonal buy-sell mbalance (expressed n /000 of a percent) s the dfference between the nsttutonal buy and sell volumes for that day scaled by the total number of outstandng shares. nsttutonal net volume s the dfference between the nsttutonal buy and sell volumes scaled by the total number of shares traded on a gven day. Returns and nsttutonal net volume are expressed n percent per day. nsttutonal net buyng, expressed n mllon dollars, s the dfference between the total value bought and sold by nsttutons. a and b ndcate sgnfcance at the one and fve percent level. EW Return N R< a <=R< a <=R< a <=R< a <=R< a R>= b a b Dff a b EW nsttutonal mbalance N R< a a <=R< b.93 -<=R< <=R< <=R< a R>= a 3. a 2.66 Dff a a 6.8 nsttutonal Net Buyng N R< a a <=R< b <=R< b <=R< <=R< a a R>= a 88. a 9. Dff a a EW nsttutonal Net Volume N R< a 0.0-2<=R< <=R< b 0.94 a b 0.3 b 0<=R< a 0.4 b 0.0 b <=R< b 0.8 a 0.86 a 0.09 R>= a.9 a 0.8 a Dff a 2.44 a 0.48 b 30

33 Table Daly VAR Estmates of Nasdaq 00 Returns and nsttutonal Buy-Sell mbalances Ths table presents the coeffcent estmates, p-values and adjusted R 2 s from the daly vector autoregresson (VAR) wth lags and the followng structural VAR: R t α + β Rt + λ t + δ t, R = (A) t α + β 0 Rt + β Rt + λ t + δ t, = (B) where R t s the average daly return and t s the daly average nsttutonal buy-sell mbalance for the stocks comprsng the Nasdaq 00 ndex. For each stock, nsttutonal buy-sell mbalance s the dfference between the nsttutonal buy and sell volumes for that day scaled by the total number of outstandng shares. Panels A and B report the results for equal-weghted averages and Panel C presents the results for value-weghted averages. To facltate nterpretaton, both varables are standardzed pror to estmaton of the VAR. p-values are n parentheses. Panel A Return nst. mbal. Adj. R 2 Dep. Var. α β 0 β β 2 β 3 β 4 β λ λ 2 λ 3 λ 4 λ Return (0.99) (0.24) (0.00) (0.26) (0.) (0.22) (0.0) (0.88) (0.23) (0.9) (0.42) nst. mbal (.00) (0.00) (0.00) (0.0) (0.42) (0.03) (0.00) (0.0) (0.6) (0.07) (0.00) Panel B Return nst. mbal. Adj. R 2 Dep. Var. α β 0 β β 2 β 3 β 4 β λ λ 2 λ 3 λ 4 λ Return (0.99) (0.24) (0.00) (0.26) (0.) (0.22) (0.0) (0.88) (0.23) (0.9) (0.42) nst. mbal (.00) (0.00) (0.00) (0.02) (0.00) (0.) (0.07) (0.00) (0.00) (0.96) (0.08) (0.00) Panel C Return nst. mbal. Adj. R 2 Dep. Var. α β 0 β β 2 β 3 β 4 β λ λ 2 λ 3 λ 4 λ Return (0.99) (0.04) (0.00) (0.66) (0.27) (0.23) (0.06) (0.) (0.04) (0.8) (0.20) nst. mbal (.00) (0.00) (0.00) (0.03) (0.00) (0.00) (0.8) (0.00) (0.00) (0.88) (0.06) (0.43) 3

34 Table nsttutonal mbalances for Return Groups for Subperods Each day from September 3, 999 to December 27, 200 s assgned to one of sx groups based on the equal-weghted average return on the stocks comprsng the Nasdaq 00 ndex. Ths table reports the equal-weghted average nsttutonal buy-sell mbalance and nsttutonal net buyng for the fve tradng days centered around the rankng day for each subperod. The perod from September, 999 to March 27, 2000 comprses the run-up perod. March 28, 2000 to Aprl 4, 200 s the crash perod. The morestable perod starts on Aprl, 200 and ends on December 3, 200. For each stock, nsttutonal buysell mbalance (expressed n /000 of a percent) s the dfference between the nsttutonal buy and sell volumes for that day scaled by the total number of outstandng shares. Run-up N R< b a 9.0-2<=R< a <=R< b.88 b 0<=R< b 3.2 b <=R< b 4.22 b 9.76 a 20.6 a -. R>= b 26.8 a 3.7 b Dff a 4.6 Crash N R< a a <=R< <=R< b <=R< b 8..0 <=R< b b R>= b 3.33 a.47 Dff a a 4.37 More-stable N R< a a <=R< b b <=R< b -24. b 0<=R< <=R< a a b R>= a 33.9 a Dff a a 9.2 a ndcates sgnfcance at percent. b ndcates sgnfcance at percent. 32

35 Table V Daly VAR Estmates of Nasdaq 00 Returns and nsttutonal mbalances for Subperods Panel A presents the coeffcent estmates, p-values and adjusted R 2 s from the followng daly structural vector autoregresson (VAR) wth lags for each subperod: R t α + β Rt + λ t + δ t, R = (A) t α + β 0 Rt + β Rt + λ t + δ t, = (B) where R t s the daly equal-weghted average return and t s the daly equal-weghted average nsttutonal buy-sell mbalance for the stocks comprsng the Nasdaq 00 ndex. For each stock, nsttutonal buy-sell mbalance s the dfference between the nsttutonal buy and sell volumes for that day scaled by the total number of outstandng shares. Panel B presents the results for a VAR where returns and nsttutonal mbalances are nteracted wth ndcator varables for the run-up and crash perods and ncluded as exogenous varables. The perod from September, 999 to March 27, 2000 comprses the run-up perod. March 28, 2000 to Aprl 4, 200 s the crash perod. The more-stable perod starts on Aprl, 200 and ends on December 3, 200. To facltate nterpretaton, all varables are standardzed pror to estmaton of the VAR. p-values are n parentheses. Panel A Run-up Return nst. mbal. Adj. R 2 Dep. Var. α β 0 β β 2 β 3 β 4 β λ λ 2 λ 3 λ 4 λ Return (0.0) (0.34) (0.0) (0.9) (0.72) (0.24) (0.9) (0.64) (0.66) (0.09) (0.63) nst. mbal (0.6) (0.04) (0.00) (0.90) (0.4) (0.44) (0.73) (0.03) (0.20) (0.78) (0.7) (0.29) Crash Return nst. mbal. Adj. R 2 Dep. Var. α β 0 β β 2 β 3 β 4 β λ λ 2 λ 3 λ 4 λ Return (0.03) (0.04) (0.00) (0.9) (0.3) (0.0) (0.04) (0.28) (0.4) (0.0) (0.8) nst. mbal (0.30) (0.00) (0.00) (0.2) (0.09) (0.2) (0.60) (0.0) (0.4) (0.83) (0.6) (0.09) More-stable Return nst. mbal. Adj. R 2 Dep. Var. α β 0 β β 2 β 3 β 4 β λ λ 2 λ 3 λ 4 λ Return (0.70) (0.) (0.8) (0.0) (0.6) (0.94) (0.27) (0.67) (0.40) (0.76) (0.92) nst. mbal (0.07) (0.00) (0.00) (0.09) (0.02) (0.6) (0.0) (0.0) (0.02) (0.62) (0.74) (0.0) 33

36 Panel B Run-up Dummy Return Crash Dummy Return Run-up Dummy nst. mbal. Crash Dummy nst. mbal. Return nst. mbal. Adj. R 2 nt Return (0.6) (0.48) (0.7) (0.43) (0.03) (0.9) (0.0) (0.82) (0.84) (0.3) (0.38) (0.46) (0.73) (0.38) (0.2) (0.07) (0.38) (0.4) (0.9) nst. mbal (0.72) (0.0) (0.92) (0.09) (0.44) (0.09) (0.4) (0.27) (0.4) (0.84) (0.40) (0.9) (0.77) (0.) (0.7) 0.00 (0.00) (0.0) (0.00) (0.00) (0.00) (0.7) 34

37 Table V Asymmetres Ths table presents the coeffcent estmates, p-values and adjusted R 2 s from the daly structural vector autoregresson (VAR) wth 3 lags for each subperod. The endogenous varables are the daly equalweghted average return and the daly equal-weghted average nsttutonal buy-sell mbalance for the stocks comprsng the Nasdaq 00 ndex. For each stock, nsttutonal buy-sell mbalance s the dfference between the nsttutonal buy and sell volumes for that day scaled by the total number of outstandng shares. Returns are nteracted wth ndcator varables for negatve returns and ncluded as exogenous varables for the nsttutonal buy-sell mbalance. The perod from September, 999 to March 27, 2000 comprse the run-up perod. March 28, 2000 to Aprl 4, 200 s the crash perod. The more-stable perod starts on Aprl, 200 and ends on December 3, 200. p-values are n parentheses. Run-up Neg. Ret. Dummy Return Return nst. mbal. Adj. R 2 nt Return (0.0) (0.32) (0.0) (0.66) (0.8) (0.6) (0.44) nst. mbal (0.70) (0.20) (0.09) (0.39) (0.79) (0.03) (0.04) (0.6) (0.20) (0.02) (0.4) (0.86) Crash Neg. Ret. Dummy Return Return nst. mbal. Adj. R 2 nt Return (0.6) (0.03) (0.00) (0.24) (0.04) (0.06) (0.40) nst. mbal (0.9) (0.08) (0.20) (0.76) (0.87) (0.0) (0.00) (0.34) (0.28) (0.00) (0.) (0.6) More-stable Neg. Ret. Dummy Return Return nst. mbal. Adj. R 2 nt Return (0.7) (0.27) (0.9) (0.9) (0.6) (0.90) (0.86) nst. mbal (0.0) (0.02) (0.89) (0.4) (0.2) (0.00) (0.00) (0.06) (0.6) (0.00) (0.0) (0.99) 3

38 Table V Daly VAR Estmates of Nasdaq 00 Returns, nsttutonal mbalances and Mutual Fund Flows Ths table presents the coeffcent estmates, p-values and adjusted R 2 s from the followng daly structural vector autoregresson (VAR) wth lags: R t α + β Rt + λ t + γ Ft + δ t, R = (A) t α + β 0Rt + β Rt + λ t + γ 0Ft + γ Ft + δ t, = (B) F t α + β 0Rt + β Rt + λ t + γ Ft + δ t, = (C) where R t s the daly value-weghted Nasdaq 00 return, t s the daly value-weghted average nsttutonal buy-sell mbalance for the stocks comprsng the Nasdaq 00 ndex and F t s the mutual fund flows scaled by the aggregate S&P market captalzaton. For each stock, nsttutonal buy-sell mbalance s the dfference between the nsttutonal buy and sell volumes for that day scaled by the total number of outstandng shares. To facltate nterpretaton, all varables are standardzed pror to estmaton of the VAR. p-values are n parentheses. Return nst. mbal. Mut. flow Adj. R 2 α β 0 β β 2 β 3 β 4 β λ λ 2 λ 3 λ 4 λ γ 0 γ γ 2 γ 3 γ 4 γ Return (0.99) (0.0) (0.02) (0.40) (0.43) (0.8) (0.04) (0.4) (0.0) (0.38) (0.) (0.4) (0.04) (0.04) (0.3) (0.4) nst. mbal (.00) (0.00) (0.00) (0.09) (0.07) (0.00) (0.0) (0.00) (0.00) (0.89) (0.09) (0.23) (0.) (0.6) (0.69) (0.8) (0.0) (0.44) Mut. flow (0.98) (0.26) 0.00 (0.3) (0.) (0.09) (0.4) (0.0) (0.33) (0.77) (0.0) (0.47) (0.00) (0.0) (0.03) (0.00) (0.00) 36

39 Table V ntradaly VAR Estmates of Nasdaq 00 Returns and nsttutonal and ndvdual Buy-Sell mbalances Ths table presents the coeffcent estmates, p-values and adjusted R 2 s from the followng ntradaly vector autoregresson (VAR) wth 6 lags for each subperod: R t α + βrt + λ t + γ Jt + δ t, R = (A) t α + β 0 Rt + β Rt + λ t + γ J t + δ t, = (B) J t α + β 0 Rt + β Rt + λ t + γ J t + δ t, J = (C) where R t s the fve-mnute equal-weghted Nasdaq 00 return and t (J t ) s the fve-mnute equal-weghted average nsttutonal (ndvdual) buysell mbalance. Ths measure s the mean of the nsttutonal (ndvdual) buy-sell mbalances for the stocks comprsng the Nasdaq 00 ndex. For each stock, buy-sell mbalance s the dfference between the nsttutonal (ndvdual) buy and sell volumes for that fve-mnute nterval scaled by the total number of outstandng shares. To avod crossng day boundares for lagged returns and buy-sell mbalances, the frst half hour of each tradng day s excluded from the analyss. The perod from September, 999 to March 27, 2000 comprses the run-up perod. March 28, 2000 to Aprl 4, 200 s the crash perod. More-stable perod starts on Aprl, 200 and ends on December 3, 200. To facltate nterpretaton, all three varables are standardzed pror to estmaton of the VAR. Run-up Return nst. mbal. nd. mbal. Adj. R 2 α β 0 Β β 2 β 3 β 4 β β 6 λ λ 2 λ 3 λ 4 λ λ 6 γ γ 2 γ 3 γ 4 γ γ 6 Return (0.4) (0.00) (0.04) (0.0) (0.08) (0.89) (0.88) (0.26) (0.67) (0.02) (0.36) (0.40) (0.83) (0.68) (0.00) (0.) (0.09) (0.) (0.82) nst. mbal (0.00) (0.00) (0.02) (0.00) (0.00) (0.4) (0.0) (0.0) (0.00) (0.00) (0.00) (0.00) (0.00) (0.00) (0.00) (0.00) (0.) (0.72) (0.24) (0.08) nd. mbal (0.00) (0.00) (0.00) (0.00) (0.00) (0.00) (0.84) (0.46) (0.00) (0.00) (0.00) (0.06) (0.02) (0.00) (0.00) (0.00) (0.20) (0.00) (0.7) (0.00) 37

40 Crash Return nst. mbal. nd. mbal. Adj. R 2 α β 0 β β 2 β 3 β 4 β β 6 λ λ 2 λ 3 λ 4 λ λ 6 γ γ 2 γ 3 γ 4 γ γ 6 Return (0.28) (0.00) (0.00) (0.49) (0.49) (0.77) (0.93) (0.24) (0.0) (0.62) (0.72) (0.0) (0.33) (0.00) (0.0) (0.93) (0.) (0.90) (0.07) nst. mbal (0.48) (0.00) (0.00) (0.00) (0.00) (0.0) (0.00) (0.6) (0.00) (0.00) (0.00) (0.00) (0.00) (0.00) (0.00) (0.2) (0.4) (0.40) (0.2) (0.2) nd. mbal (0.00) (0.00) (0.00) (0.00) (0.00) (0.00) (0.00) (0.00) (0.00) (0.03) (0.) (0.3) (0.22) (0.09) (0.00) (0.00) (0.00) (0.00) (0.00) (0.00) More-stable Return nst. mbal. nd. mbal. Adj. R 2 α β 0 β β 2 β 3 β 4 β β 6 λ λ 2 λ 3 λ 4 λ λ 6 γ γ 2 γ 3 γ 4 γ γ 6 Return (0.24) (0.00) (0.00) (0.0) (0.82) (0.) (0.28) (0.9) (.00) (0.3) (0.00) (0.88) (0.2) (0.4) (0.8) (0.2) (0.47) (0.3) (0.60) nst. mbal (0.0) (0.00) (0.00) (0.00) (0.00) (0.40) (0.) (0.2) (0.00) (0.00) (0.00) (0.00) (0.00) (0.00) (0.00) (0.00) (0.23) (0.8) (0.87) (0.00) nd. mbal (0.00) (0.00) (0.00) (0.00) (0.00) (0.0) (0.32) (0.26) (0.00) (0.07) (0.07) (0.36) (0.79) (0.88) (0.00) (0.00) (0.00) (0.00) (0.0) (0.00) 38

41 Table V nsttutonal Block Trades Ths table presents the coeffcent estmates, p-values and adjusted R 2 s from the followng daly structural vector autoregresson (VAR) wth lags: R t α + β Rt + λ t + δ t, R = (A) t α + β 0 Rt + β Rt + λ t + δ t, = (B) where R t s the daly equal-weghted average return and t s the daly equal-weghted average nsttutonal buy-sell mbalance for the stocks comprsng the Nasdaq 00 ndex. For each stock, nsttutonal buy-sell mbalance s the dfference between the nsttutonal buy and sell volumes for that day scaled by the total number of outstandng shares, where the buy and sell volumes only nclude trades for greater than 0,000 shares. To facltate nterpretaton, both varables are standardzed pror to estmaton of the VAR. p- values are n parentheses. Return nst. mbal. Adj. R 2 Dep. Var. α β 0 β β 2 β 3 β 4 β λ λ 2 λ 3 λ 4 λ Return (0.99) (0.42) (0.00) (0.) (0.0) (0.9) (0.03) (0.62) (0.) (0.9) (0.29) nst. mbal (0.99) (0.00) (0.00) (0.02) (0.06) (0.32) (0.20) (0.00) (0.23) (0.63) (0.07) (0.6) 39

42 Fgure Dstrbuton of Volume by Trade Sze Ths fgure plots the percentage of volume that can be explaned by each trade assgnment over the September, 999 to December 3, 200 perod for small and large trades. The market maker (dealer) on each sde of each trade s tradng for hs/her own account (MM) or s smply actng as an agent and handlng a trade for a customer. All agency trades are classfed nto nsttutonal (nst.) or ndvdual (nd.) based on whether the market maker prmarly deals wth nsttutons or ndvduals. Both sdes of the trades are classfed as to whether they trade wth another nsttuton, an ndvdual or a market maker. The trades wth nconsstences n assgnng whether a market maker acted as a prncpal or an agent for each leg of the trade form the non-classfed group. Trades for less than 00 shares are labeled as small trades, and trades for greater than 0,000 shares are classfed as large trades. 40