Managers, Investors, and Crises: Mutual Fund Strategies in Emerging Markets

Transcription

1 Publc Dsclosure Auhorzed Publc Dsclosure Auhorzed Publc Dsclosure Auhorzed Publc Dsclosure Auhorzed JEL: F3 G1 G2. Managers Invesors and Crses: Muual Fund Sraeges n Emergng Markes Gracela Kamnsky Rchard Lyons Sergo Schmukler Keywords: muual funds; managers; nvesors; radng sraeges; emergng markes; momenum; feedback radng; crss; conagon. Respecve afflaons are George Washngon Unversy UC Berkeley and NBER and he World Bank. Correspondence o Sergo Schmukler The World Bank 1818 H Sree NW Washngon D.C Tel: sschmukler@worldbank.org. We hank he followng for valuable commens: Jeff Frankel Mke Gavn George Hogue Andrew Karoly Federco Surzenegger and parcpans a he World Bank/Unversdad Torcuao D Tella conference on Inegraon and Conagon (June 1999) he Cancun Meeng of he Economerc Socey he Lan Amercan Economc Assocaon and he IDS-U. of Sussex. For help wh daa we hank he World Bank (Eas Asa and Pacfc Regon) Erk Srr from he SEC Konsannos Tsasarons from he BIS and Ian Wlson from Emergng Marke Funds Research. For excellen research asssance we hank Jon Tong Sergo Kurla Ccla Harun Jose Pneda and Allen Cheung. (The effors of Sergo Kurla and especally Jon Tong were prodgous.) For fnancal suppor we hank he NSF and he World Bank (Lan Amercan Regonal Sudes Program and Research Suppor Budge).

2 Managers Invesors and Crses: Muual Fund Sraeges n Emergng Markes I. Inroducon Fnancal crss n 1997 engulfed no only Asa spread o counres as dsan as Souh Afrca he Czech Republc and Brazl. To undersand why a leraure has developed ha examnes why he spreadng of crss mgh be due o fnancal lnks. There s evdence ha banks for example were mporan n spreadng he 1997 crss. The ransmsson channel was lendng: counres were exposed o he same banks (Kamnsky and Renhar 1999). Porfolo nvesors have also been scrunzed parcularly nsuons such as hedge funds penson funds and muual funds (Brown e al Echengreen and Maheson 1998 Km and We 1999 Frankel and Schmukler 1998 among many ohers). A common concluson s ha nsuons somemes panc dsregardng fundamenals and spreadng crss even o counres wh srong fundamenals. The leraure noes ha ndvduals oo can conrbue o hs panc by fleeng from funds parcularly muual funds forcng fund managers o sell when fundamenals do no warran sellng. Ths paper conrbues o hs leraure on fnancal lnks by examnng he radng sraeges of an mporan class of nvesor: U.S. muual funds. Surprsngly sysemac analyss of muual funds nernaonal sraeges does no ye exs. 1 Consequenly our resuls are of more general neres han our crss movaon mgh sugges. A he same me he lack of sysemac analyss of funds behavor durng crses warrans specal aenon. Though here s some evdence ha funds help crss o spread ha evdence s ndrec and hghly aggregave. Frankel and Schmukler (1998) for example use closedend muual funds o show ha he Mexcan crss n 1994 was no ransmed o Asa drecly bu ndrecly va New York where he funds are raded. The oppose vew ha funds do no spread crss also has some suppor n aggregae daa. For example ne 1 Funds domesc (U.S.) sraeges have been analyzed exensvely however. See Grnbla e al. (1995) Warher (1995) and Wermers (1999) among ohers. 1

3 redempon by muual-fund nvesors durng crss perods s no large and ouflows ha occur end o be small and shor-lved (a leas durng Mexco s crss see Marcs e al and Rea 1996). Froo e al. (1998) presen a smlar pcure based on aggregaed flows ha mx muual funds wh oher ypes of nernaonal nvesor. They fnd ha ne nflows durng he Mexcan and Asan crses decreased bu here s lle evdence of ne ouflows. 2 Our paper depars from he more aggregaed analyss above by effecng analyss a he porfolo level. We develop a novel daa se ha ncludes ndvdual porfolos whch allows us o examne radng sraeges a much hgher resoluon. The daa nclude he quarerly holdngs of 13 muual funds from Aprl 1993 o January All 13 funds are dedcaed Lan Amerca funds. (A year-end 1998 here were 25 Lan Amerca funds; he 13 we rack accoun for 88% of he value of hese 25 funds.) We use hese daa o address wo ses of quesons. The frs se relaes o wheher funds engage n momenum radng sysemacally buyng wnnng socks and sellng losng socks (Jegadeesh and Tman 1993 Grnbla e al. 1995). The second se of quesons relaes o wheher funds engage n conagon radng by whch we mean sysemacally sellng socks from one counry when sock prces are fallng n anoher. In addressng hs second se of quesons we esablsh a frs drec emprcal lnk beween conagon and radng sraeges. The mehodologcal conrbuon of he paper s our approach o arbung acons o fund managers versus underlyng nvesors. Despe a vas leraure on he behavor of domesc (.e. U.S.) funds o our knowledge we are he frs o dsenangle he wo. In effec he rades of muual funds reflec boh nsuonal and ndvdual decsons. To undersand hose rades parcularly n he nernaonal conex ensurng ha he decsons are no commngled s an mporan sep. Our resuls show ha emergng-marke funds do ndeed engage n momenum radng. Ther sraeges exhb posve momenum hey sysemacally buy wnners and sell losers. Ths s due o momenum radng a boh he fund-manager level and he nvesor level (hrough redempons/nflows). We furher dsngush beween 2 Though a lovely daa se he Froo el al. (1998) daa do no nclude ransacons seled n foregn currences e.g. ADR rades n New York and Brady bonds. These rades can be especally mporan n mes of crss when local-marke lqudy s a a mnmum. Our fund-porfolo daa nclude hese rades. 2

4 conemporaneous momenum radng (buyng curren wnners and sellng curren losers) and lagged momenum radng (buyng pas wnners and sellng pas losers). Conemporaneous momenum radng s sronger durng crses and sronger for fund nvesors han for fund managers. Lagged momenum radng on he oher hand s sronger durng non-crss perods and sronger for managers. We also fnd ha funds engage n conagon radng by whch we mean ha hey sysemacally sell asses from one counry when asse prces fall n anoher. Ths conagon radng s due prmarly o underlyng nvesors no managers. The paper s organzed as follows. The nex secon oulnes our approach o measurng momenum radng and conagon radng. Secon III descrbes our daa. Secon IV presens our momenum and conagon resuls. Secon V addresses wheher reurn auocorrelaon whn Lan Amerca can raonalze our secon-iv resuls. Secon VI concludes. The appendx provdes some relaed regresson-based analyss. II. Sraeges: Momenum Tradng and Conagon Tradng Ths secon presens our approach o esng wheher funds employ momenum and conagon radng sraeges. Momenum radng also called posve feedback radng s he sysemac purchase of socks ha have performed well and sale of socks ha have performed poorly ( wnners and losers ). Conagon radng s he sellng of asses from one counry when asse prces are fallng n anoher. Conagon radng s hus a crosscounry phenomenon n conras o momenum radng whch s a whn-counry phenomenon. (Ths ype of cross-counry analyss s no possble usng recen snglecounry daa ses such as hose of Km and We 1999 and Choe Kho and Sulz 1999.) Frs we revew he exsng fnance leraure on momenum radng. Second we presen our approach o esng for momenum radng an approach ha draws from hs earler leraure. Then we urn o conagon radng presenng frs a bref revew of he conagon leraure followed by our approach o esng for conagon radng. The approach we adop n esng for conagon radng s n he same spr as our es for momenum radng. 3

5 II.1. Inroducon o Momenum Tradng The leraure on momenum radng ncludes wo lnes of work one based n asse prcng and he oher based n nernaonal fnance. The asse-prcng lne begns wh he fndng ha a sraegy of buyng pas wnners and sellng pas losers generaes sgnfcan posve reurns over 3- o 12-monh holdng perods (Jegadeesh and Tman 1993 Asness e al Rouwenhors 1998). 3 Once esablshed hs resul nspred work on wheher nvesors acually follow momenum radng sraeges. Grnbla e al. (1995) for example examne he domesc sraeges of U.S. muual funds and fnd ha hey do sysemacally buy pas wnners. They do no sysemacally sell pas losers however. They also fnd ha funds usng momenum radng sraeges realze sgnfcanly beer performance. Evaluaon of performance s a cenral heme for all he papers n hs asse-prcng lne of he leraure. The second lne of work on momenum radng s based n nernaonal fnance. Is organzng heme s he lnk beween reurns and nernaonal capal flows. A he cener of hs leraure s he posve conemporaneous correlaon beween capal nflows and reurns. Early work esablshes hs correlaon usng daa aggregaed over boh me and ypes of marke parcpan (Tesar and Werner 1994 Bohn and Tesar 1996). Laer work relaxes he aggregaon over me o address wheher he conemporaneous correlaon n quarerly daa s ruly conemporaneous (Froo e al Choe e al Km and We 1999). Hgher frequency daa can dsngush hree possbles. Reurns may precede flows ndcang posve feedback radng (whch s no necessarly rraonal per he asse-prcng leraure noed above). Reurns and flows may be ruly conemporaneous ndcang ha order flow self may be drvng prces. 4 And reurns may lag flows 3 The reurn connuaons ha are mpled by hs resul are no nconssen wh he reurn reversals documened elsewhere n he leraure. Horzon lengh s he key o undersandng hs: connuaons appear a md-range horzons 3 o 12 monhs. Reurn reversals n conras appear a shor horzons (up o 1 monh see Jegadeesh 1990 and Lehmann 1990) and a long horzons (3 o 5 years see De Bond and Thaler 1985). Reversals call for conraran (or negave feedback) radng sraeges. Parenhecally all hese me-seres anomales are dsnc from he cross-seconal anomales ha have receved much aenon n he asse-prcng leraure recenly (e.g. sze and book-o-marke effecs). 4 Mcrosrucure fnance provdes hree channels for ruly conemporaneous prce mpac. The frs s nformaon f he buyer has superor nformaon abou a secury s payoffs hen he purchase sgnals ha nformaon shfng expecaons and hereby ncreasng prce. The second s ncomplee rsksharng a he markemaker level he buyer s purchase emporarly dsurbs he markemaker s poson 4

6 ndcang flows ably o predc reurns. Usng hgh-frequency daa aggregaed across ypes of marke parcpan Froo e al. (1998) fnd evdence of all hree wh he frs posve feedback radng beng he mos mporan for explanng quarerly correlaon. Choe e al. (1999) and Km and We (1999) use hgh-frequency daa from Korea o examne posve feedback radng around he 1997 currency crss. Choe e al. fnd ha foregn nvesors as a group engage n posve feedback radng before he crss bu durng he crss feedback radng mosly dsappears. Km and We examne foregn nsuonal nvesors separaely and fnd ha hey engage n posve feedback radng a all mes before durng and afer he crss. Our analyss s relaed o and borrows from boh he nernaonal-fnance and asse-prcng lnes of he leraure. Lke he work n nernaonal fnance we are more concerned abou nernaonal flows and crss ransmsson han porfolo performance. Lke work n asse prcng however we manan a drec lnk o nvesmen sraegy and s measuremen. In parcular we focus on a specfc class of nernaonal nvesor muual funds. A benef of focusng on a specfc nvesor class s ha we can characerze he evoluon of acual porfolos and how ha evoluon relaes o reurns n varous counres. Anoher benef s ha our daa allow us o analyze only he behavor of fund managers and her underlyng nvesors. On he cos sde focusng on funds as a specfc nvesor class means ha we lose resoluon n erms of daa frequency: our daa are quarerly. whch requres he buyer o pay compensaon n he form of a hgher prce (so-called nvenory effecs ). The hrd s mperfec subsuably he buyer s purchase may be a large enough porfolo shf relave o he marke as a whole ha permanenly hgher prce s requred o clear he marke (even f s common knowledge ha he buyer does no have superor nformaon abou he secury s payoffs). 5

7 II.2. Measurng Momenum Tradng Our momenum-radng measure s akn o ha used o analyze funds domesc sraeges (e.g. Grnbla e al. 1995). The measure capures he relaon beween secury ransacons and reurns. I s based on he mean of ndvdual observaons of he varable: 5 1 M = R k (1) where s he holdng by fund of sock (n shares) a me s ( + -1 )/2 and R -k s he reurn on sock from -k-1 o -k. When k=0 hs measure capures he conemporaneous relaon beween rades and reurns referred o as lag-zero momenum radng (L0M). When k=1 he measure capures he lagged response of rades o reurns and s referred o as lag-one momenum radng (L1M). Parenhecally noce he mplcaon of he subscrp: he mean of M measures he nensy of momenum radng a he level of ndvdual socks. Tesng he null of no momenum radng s a es of wheher he mean of M over all and s zero. Ths measure of momenum radng has wo mporan advanages. Frs s no conamnaed by passve prce momenum. Passve prce momenum arses n momenum radng measures lke hose of Grnbla e al. where he erm n brackes s a change n porfolo wegh raher han a percenage quany adusmen. When usng a porfolo wegh a prce ncrease n one sock (relave o prces of oher holdngs) produces a posve relaon beween weghs and reurns ha has nohng o do wh radng sraegy. (A smlar posve relaon arses for losng socks.) The second advanage of our measure over one based on porfolo weghs s ha our measure s no conamnaed by anoher passve effec passve quany momenum. When usng porfolo weghs a large rade n one sock can have subsanal effecs on he weghs of holdngs ha nvolve no ransacons. Our man concern here as n he res of he nernaonal-fnance-based 5 Ths mean esmae does no value-wegh he ndvdual sock posons. Ths could make a dfference f he nensy of momenum radng dffers dependng on poson value. Afer calculang boh ways we dd no fnd any qualave dfference n he resuls. 6

8 leraure on momenum radng s he relaon beween reurns and ransacon flows. 6 Accordngly we wan our realzaons of M o reflec acual ransacons he buyng and sellng of wnners and losers. Separang Manager and Invesor Momenum Tradng An mporan ssue n he conex of muual-fund sraeges s he effec of ne redempons. Many funds experence subsanal redempons durng crss perods. If on average funds sell shares o mee redempons when R -k s negave hen our momenum radng measures wll be posve. Ths resul s no spurous. Bu does reflec sraeges of underlyng nvesors raher han sraeges of he fund manager. We conrol for hs redempon effec by measurng he quany ransaced n each sock relave o a fund-specfc benchmark. Ths benchmark reflecs he quany ha would be ransaced f a fund's ne flows from nvesors produced proporonal adusmen n all socks. Specfcally o solae he manager's conrbuon o momenum radng we calculae ndvdual observaons of: M ( ) 1 1 = R k (2) P P where P s he prce of secury a me and P s (P +P -1 )/2. The second erm n brackes s a erm ha s fund-specfc.e. for a gven fund and me s nvaran across socks. I capures he percen ncrease n porfolo sze due o ne nflows. Here we use he noaon o denoe all hose socks held by fund. The overall momenum radng measure n equaon (2) herefore reflecs he degree o whch he manager of fund buys wnners and sells losers beyond any average quany adusmen due o fund nflows/ouflows. To undersand why noe ha he numeraor of he second erm n 6 Ths conrass wh he asse-prcng-based leraure on momenum whose man concern s porfolo performance n whch case s necessary o consder he reurn mpac of all porfolo posons. Noe oo ha emergng-marke funds are subec o large and rapd redempons whch dependng on lqudy n specfc markes can produce sgnfcan passve quany momenum. 7

9 brackes s he change n porfolo value due o nflows/ouflows usng he P erm facors ou capal gans/losses and he denomnaor s he average porfolo value. (As wh our frs momenum radng measure M when k=0 M capures he conemporaneous relaon beween rades and reurns L0M and when k=1 M capures he lagged response of rades o reurns L1M). Under he null hypohess of no momenum radng a he manager level he mean of he observaons M s zero. We can also examne he nvesor-level erm n solaon. Tha s we can calculae ndvdual observaons of M ( ) 1 P = R k. (3) P Henceforh we refer o momenum radng sascs calculaed from equaons (1)-(3) as whole-fund manager-only and nvesor-only momenum respecvely. A Second Invesor-Level Measure Before movng on s mporan o recognze wha our nvesor-only measure s capurng and wha s no capurng. Wha our nvesor-only measure does capure s nvesor effecs on our whole-fund measure; ha s he sum of he nvesor-only and manager-only measures equals he whole-fund measure. (Ths s no que rue n our repored resuls because we om some ouler observaons for robusness as descrbed below n secon IV.) Though hs nvesor-only measure s ceranly an obec of neres does no recognze ha nvesors decsons are made a he level of he fund no a he level of ndvdual socks. (Manager decsons n conras are made a he level of ndvdual socks). To capure hs we also esmae an nvesor-only measure a he fund level. Specfcally we esmae he mean of he sasc: M ( 1 ) P R k = P. (4) 8

10 Clearly hs reduces he number of observaons we lose he sock dmenson bu beer corresponds o he decson ha nvesors acually face. Condonal Momenum Tradng In addon o he momenum measures L0M and L1M we are also neresed n condonal momenum radng. Specfcally we spl our sample no sub-perods: crss and non-crss. The crss poron of our full sample (Aprl 1993 o January 1999) ncludes four sub-perods: December 1994 o June 1995 (Mexco) July 1997 o March 1998 (Asa) Augus 1998 o Ocober 1998 (Russa) and January 1999 (Brazl). 7 Sascal Inference Several nference ssues deserve furher aenon. Frs he percenage quany changes he erm n brackes n equaons (1) hrough (4) may have fund-specfc volales. Two facors could accoun for dfferng volales a he fund level. Facor one s he consderable cross-seconal dfference n fund sze sze can affec radng sraeges. Facor wo s fund dfferences ha are dsnc from sze such as urnover raos redempon penales and oher facors. Below we es for heeroskedascy across funds and afer fndng we correc for. 8 Whle he frs nference ssue peraned o heerogeney across funds a second nference ssue perans o dependence across observaons whn funds. Specfcally ndvdual observaons of our varous momenum radng sascs M are unlkely o be ndependen across socks whn a gven fund. Our mean esmae should accoun for hs 7 We also examned a second condonal momenum measure by splng our sample no buys and sells (as n Grnbla e al. 1995). Buyng pas wnners and sellng pas losers need no be symmerc. We found however ha our resuls were exremely sensve o he specfcaon of expeced reurns an adusmen ha s necessary when splng buys from sells (see Grnbla e al. page 1091). We do no repor hose resuls due o her fragly. 8 Because our heeroskedascy correcon affecs only sandard errors each observaon of M ges equal wegh n he calculaon of a momenum measure s mean. Our correcon for heeroskedascy herefore does no aler he fac ha funds wh more observaons have more effecve wegh. Regreably we have lle sascal power o explore wheher funds dffer apprecably n he nensy of her momenum radng. As for heeroskedascy n he me-seres dmenson our sample paron no crss and non-crss perods accouns for he mos obvous correcon. 9

11 cross-sock whn-fund correlaon. Our esmaes of he mean cluser observaons whn funds and allow he weghs assgned o ndvdual observaons o vary wh he covarance srucure. A hrd nference ssue ha warrans aenon s he possbly ha our momenum radng measures mgh be based due o hgh reurn volaly whch s clearly a feaure of our crss-rdden sample (see Forbes and Rgobon 1998). In fac we are no exposed o hs bas under our null because under our null he sascs we repor n Tables 1-5 are equal o zero. In hs case he bas s no problemac. 9 II.3. Inroducon o Conagon The fnancal crses of he 1990s n Europe Mexco Asa Russa and Brazl spread rapdly across counres ncludng counres wh dverse marke fundamenals. 10 These evens spawned a leraure o make sense of he seemng conagon. The erm conagon s used que dfferenly by dfferen auhors however so le us be more specfc. From he ouse however was clear ha auhors use ha erm que dfferenly. Presenly he leraure on conagon denfes hree ypes: fundamenal-spllover conagon common-cause conagon and non-fundamenal conagon. Fundamenalspllover conagon occurs when an nsde dsurbance s rapdly ransmed o mulple economcally nerdependen counres. Common-cause conagon occurs when an ousde dsurbance s rapdly ransmed o mulple counres (e.g. a fall n commody prces or learnng abou common fundamenal facors). Fundamenal dsurbances underle boh of hese frs wo ypes. The hrd ype non-fundamenal conagon can sem from any knd 9 Under he alernave hypohess of non-zero measures however precse sascal comparsons across crss and non-crss sub-samples would requre adusmens for he volaly-specfc naure of he sample spl. Ths ype of comparson s no cenral o our paper. Neverheless we dd re-esmae our man comparave resuls wh a Forbes-Rgobon correcon (n hs case a correcon o esmaed covarance raher han correlaon) and found no qualave change n he resuls. 10 Wness Indonesa n Nobody can dsagree ha here were sgns of weakness n he Indonesan economy a he ouse of he Asan crss: he bankng secor was fragle he economy was no growng and here was a curren accoun defc. Sll hese problems were no nsurmounable. Kamnsky (1998) for example esmaes ha he probables of crss n Indonesa by June 1997 amouned o only 20 percen. Ths probably sands n sharp conras o he lkelhood of a currency crss n Thaland whch skyrockeed o 100 percen a he begnnng of Sll he Indonesan rupah collapsed only weeks afer he floang of he Tha bah. 10

12 of dsurbance; he defnng characersc s ha he rapd ransmsson o mulple counres s beyond wha s warraned by fundamenals (.e. conrollng for fundamenals canno accoun for ). Ths hrd ype s somemes referred o as pure or rue conagon. Many auhors focus on he frs wo ypes of conagon hose drven by fundamenals. For example Echengreen Rose and Wyplosz (1996) examne wheher conagon s more prevalen among counres wh eher mporan rade lnks or smlar marke fundamenals. In he frs case devaluaon n one counry reduces compeveness n parner-counres prompng devaluaons o resore compeveness (fundamenalspllover conagon). In he second case devaluaon acs lke a wake-up call: nvesors seeng one counry collapsng learn abou he fragly of smlar counres and speculae agans hose counres' currences (common-cause conagon). The Echengreen e al. evdence pons n he drecon of rade lnks raher han smlar fundamenals. Corse e al. (1998) also clam ha rade lnks drve he srong spllovers durng he Asan crss. Kamnsky and Renhar (1999) focus nsead on fnancal-secor lnks. In parcular hey examne he role of common bank lenders and he effec of cross-marke hedgng (a ype of common-cause conagon). They fnd ha common lenders were cenral o he spreadng of he Asan crss (as hey were o he spreadng of he Deb Crss of he 1980s). The non-fundamenal caegory of conagon has araced more aenon han he wo fundamenals-drven caegores. Theorecal work on non-fundamenal conagon focuses on raonal herdng. For example n he model of Calvo and Mendoza (1998) he coss of gaherng counry-specfc nformaon nduce raonal nvesors o follow he herd. In he model of Calvo (1999) unnformed nvesors replcae sellng by lqudy-squeezed nformed nvesors because he unnformed msakenly (bu raonally) beleve hese sales are sgnalng worsenng fundamenals. Kodres and Prsker (1999) focus on nvesors who engage n cross-marke hedgng of macroeconomc rsks. In ha paper nernaonal marke comovemen can occur n he absence of any relevan nformaon and even n he absence of drec common facors across counres. For example a negave shock o one counry can lead nformed nvesors o sell ha counry s asses and buy asses of anoher counry ncreasng her exposure o he dosyncrac facor of he second counry. Invesors hen hedge hs new poson by sellng he asses of a hrd counry compleng he chan of conagon from he frs counry o he hrd. 11

13 The leraure on non-fundamenal conagon also has an emprcal branch. Kamnsky and Schmukler (1999) fnd ha spllover effecs unrelaed o marke fundamenals are que common and spread quckly across counres whn a regon. Valdes (1998) examnes he degree o whch comovemen of Brady-bond prces s unexplaned by fundamenals. Ineresngly conagon n hs paper s symmerc applyng boh on he downsde durng crses and on he upsde durng perods of rapd capal nflow. A dfferen lne of emprcal work on non-fundamenal conagon examnes wheher crses are spread by parcular nvesor groups. For example Choe Kho and Sulz (1998) use ransacon daa n he Korean equy marke o examne wheher foregn nvesors desablze prces. They fnd evdence of herdng by foregn nvesors before Korea s economc crss n lae 1997 bu hese effecs dsappear durng he peak of he crss and here s no evdence of desablzaon. Snce her daa nclude only ransacons on he Korean Sock Exchange hese auhors canno examne he ransmsson of crss across counres. II.4. Measurng Conagon Tradng Our approach o esng for conagon s dfferen from he leraure revewed above. Daa on ndvdual porfolos allow us o address conagon n a new way from he radng-sraegy perspecve. We wll use he erm conagon radng o mean he sysemac sellng (buyng) of socks n one counry when he sock marke falls (rses) n anoher. 11 To do hs we nroduce a new measure a conagon radng measure. Our conagon radng measure s based on he mehodology oulned above for measurng momenum radng. Lke he momenum measures we presen conagon radng measures a hree dfferen levels: whole-fund conagon radng (C) manager-only conagon radng 11 Noce ha hs defnon does no ake accoun of he fundamenal-versus-non-fundamenal dsncon nroduced above. The appendx nroduces a regresson-based approach ha allows us o es for conagon wh conrols for varous fundamenal facors. 12

14 13 (C') and nvesor-only conagon radng (C''). These hree measures are he sample averages of he varables: f R C 1 = (5) ( ) f R P P C 1 1 = (6) ( ) f R P P C 1 = (7) Insead of esng for a relaon beween quany changes and own-sock reurns our conagon radng measure ess for a relaon beween quany changes and foregn-counry equy reurns. In effec we are esng for wha mgh be called "cross-counry momenum radng." Here R f s he reurn on he foregn-counry ndex f from -1 o. For each of he hree measures above (C C' and C'') we consder fve dfferen conagon radng measures each one consruced from a dfferen foregn equy ndex. Those foregn equy ndexes nclude Brazl Mexco Asa Russa and he U.S. Naurally when calculang he conagon radng measure when f=brazl we do no nclude observaons where sock s from Brazl (smlarly for Mexco). Under he null hypohess of no conagon radng he mean of he observaons C s zero. Our conagon radng measure n equaons (5)-(7) allows us o address many of he ssues we address wh our momenum radng measure. For example we examne crss versus non-crss sub-samples and we paron he crss sub-sample furher o solae he effecs of parcular crses. We do no offer a conagon-radng analogue o

15 equaon (4) nvesor-only a he fund level only because he resuls we shall fnd for ha measure are n he end smlar o he nvesor-only resuls from equaon (3) III. Daa Our daa on muual-fund holdngs come from wo sources. The frs source s he U.S. Secures and Exchange Commsson (SEC). Muual funds are requred o repor holdngs o he SEC wce a year. The second source s Mornngsar. Mornngsar conducs surveys of muual fund holdngs a a hgher frequency: quarerly surveys are he norm for mos funds. For our purposes quarerly daa are avalable from Mornngsar for abou 50% of he funds we examne. In hose nsances where our measure of M s based on porfolo holdngs ha are no measured hree monhs apar hese observaons of are mulpled by 3/x where x s he number of monhs beween and -1. Our sample ncludes he holdngs of 13 Lan Amerca equy funds (open-end) from Aprl 1993 o January 1999 (24 quarers). Those funds are (1) Fdely Lan Amerca (2) Morgan Sanley Dean Wer Insuonal Lan Amerca (3) Van Kampen Lan Amerca (formerly Morgan Sanley) (4) BT Invesmen Lan Amerca Equy (5) TCW Galleo Lan Amerca Equy (6) TCW/Dean Wer Lan Amerca Growh (7) Excelsor Lan Amerca (8) Gove Lan Amerca (9) Ivy Souh Amerca (10) Scudder Lan Amerca (11) T. Rowe Prce Lan Amerca (12) Merrll Lynch Lan Amerca and (13) Templeon Lan Amerca. No all of hese funds exsed from he begnnng of our sample; on average we have abou 10 quarers of daa (ou of a possble 24) per fund. Our hrd source of daa s Bloomberg and he Inernaonal Fnance Corporaon (IFC). Bloomberg provdes monhly prce seres for all eques held by he 13 funds ncludng ADRs. (The need for monhly prce daa arses n our analyss of lag-one momenum radng.) These prce seres are correced for spls and dvdends. The IFC provdes nformaon on sock marke ndexes whch we need for our conagon radng analyss. Our conagon radng analyss uses he IFC Lan Amerca Sock Marke ndex he IFC Asa Sock Marke ndex and several IFC counry sock marke ndexes. The U.S. equy reurn s he S&P 500 reurn. All reurn daa are expressed n percen. 14

16 IV. Resuls: Momenum and Conagon Tradng We presen our resuls n four pars. Frs we presen aggregae evdence on he rades of muual funds n mes of crss. Then we presen resuls on whn-counry momenum radng (equaons 1-4). We follow hese wh cross-counry conagon radng resuls (equaons 5-7). In he appendx we also presen some regresson-based resuls relang momenum and conagon radng wh oher deermnans of radng sraegy. IV.1. Aggregae Evdence on our Sample of Funds Durng Crss Though our daa se does nclude ndvdual porfolos le us frs consder evdence based on he aggregaon of hose porfolos. We focus hs aggregae evdence on funds experence wh nvesor nflows and ouflows. Durng he fourh quarer of 1997 he peak of he Asan crss Lan Amercan funds suffered large ouflows (Fgure 1). 12 The reversal from nflows o ouflows durng he Asan and Russan crses s more severe han ha durng he Mexcan crss n December In he Mexcan crss funds ended o pull ou of Mexco Argenna and Brazl all of whch are relavely lqud; funds ended no o pull ou from more llqud markes such as Colomba. Moreover he Mexco-nduced pullou was emporary by he hrd quarer of 1995 fund nflows o Lan Amerca had resumed (conssen wh he fndngs of Marcs e al and Rea 1996). Relave o he Mexcan crss he Asan and Russan crses of 1997 and 1998 were more broad-based and perssen. In hose crses he rerea from Lan Amerca was more ndscrmnae wh heavy sales reachng even he mos llqud markes. On average ne sales n 1998 were abou 32 percen. Ths resul dffers from ha of Froo e al. (1998) who fnd lle evdence of ne ouflows durng he Asan crss. A possble explanaon s ha he aggregaed daa used by Froo e al. nclude nsuon ypes ha counerac he clear ne sellng by muual funds (hedge funds?). Anoher possble explanaon s ha he Froo el al. daa do no nclude ransacons seled n dollars euros or yen e.g. ADR 12 Ne sellng n Fgure 1 s calculaed as he change n number of shares as a percenage of average shares held durng he quarer valued a he begnnng-of-quarer prce. The average shares held durng he quarer s he mean of he begnnng- and end-of-quarer holdngs. 15

17 rades n New York and dollar denomnaed bonds. Ths s very mporan n Lan Amerca. Our daa se ncludes all hese rades. One echnque avalable o managers s usng cash (e.g. lqud money-marke nsrumens such as U.S. Treasury blls) o buffer her porfolos from redempons. Holdng cash allows managers o mee redempons whou he need o sell less-lqud asses. In prncple hs can mue he effec of nvesor ouflows on he underlyng socks. However managers can also renforce nvesors acons f hey ncrease her lqud posons n mes of nvesor rerenchmen. For our whole sample funds kep an average of 4.4 percen of her ne asse value n cash. We hen spl our sample no wo sub-samples one where on average hese funds receved nflows and one where on average hese funds suffered ouflows. In he nflows sub-sample we fnd an average cash poson of 4.6 percen whereas n he ouflow sample we fnd an average cash poson of 4.3 percen. Average cash posons are remarkably sable. Managers choce of cash poson does no appear o eher mue or renforce nvesor acons. 13 IV.2. Momenum Tradng Resuls In our full sample we fnd srong evdence of lag-zero momenum radng a all hree levels: whole-fund manager-only and nvesor-only (Table 1 column 1). Ineresngly conemporaneous momenum radng s especally srong durng crses. In erms of arbuon s nvesors ha accoun for he lon s share of he conemporaneous momenum radng a he whole-fund level. Sgnfcan lag-one momenum radng s presen only n he non-crss poron of our sample and s concenraed a he manager level. 14 For robusness we esmae each cell based only on observaons of M whn hree sandard devaons of s mean. Ths s he reason why whn any column of Table 1 he 13 A naural queson s wheher hese cash posons are sable because managers face some knd of consran. The realy s ha funds are far less consraned han our cash-holdng resuls mgh ndcae n any de ure sense. De faco however managers are sensve abou deparng oo much from her benchmarks. The classc example s he hapless manager a Fdely s Magellan Fund n he lae 90s who fel ha he sock marke was over-valued swched heavly no cash wached he marke rse furher and was fred for he decson. 14 In our esmaon L1M always relaes he ransaced quanes beween -1 and wh he reurn over he monh precedng -1. Increasng he lengh of he perod over whch lagged reurns are measured dmnshes explanaory power n general. 16

18 manager-only and nvesor-only esmaes do no sum o he whole-fund esmae exacly. 15 To nerpre he sze of he coeffcens consder he whole-fund L0M esmae of Gven he uns of our daa an L0M esmae of 2.5 mples ha on average he produc of ( / ) and R over a quarer s 2.5 percen (a represenave example would be a reurn of 10% and a poson reducon of 0.25 or 2.5%). 16 Table 2 presens esmaes of our nvesor-only measure a he fund level raher han a he sock level as n Table 1. Recall ha hs fund-level varan of he nvesor-only measure recognzes ha nvesors decsons are made a he level of he fund no a he level of ndvdual socks. Despe fewer observaons from losng he sock dmenson our resuls are sharpened n erms of sascal sgnfcance hough he overall paern remans he same. The only noable change n he paern s he sgnfcance of L1M a he nvesor leve: s now sgnfcan a he 1 percen level whereas was nsgnfcan n Table 1. Table 3 presens momenum radng measures for hree crss-perod sub-samples: he Mexcan Crss (December 1994 o June 1995) he Asan Crss (July 1997 o March 1998) and he Russan Crss (Augus 1998 o Ocober 1998). The neresng queson here s wheher momenum radng s equally srong across dfferen crses. The answer s no. Whn our Lan Amercan sample we fnd ha posve momenum radng was sronges durng he 1994 Mexcan Crss. IV.3. Conagon Tradng Resuls Tables 4 and 5 presen our conagon radng resuls. Table 4 presens he allsample resuls as well as he crss versus non-crss sub-samples. Table 5 spls he crss sub-sample furher no he Mexcan Asan and Russan crses. In Table 4 we fnd more sgnfcance a he nvesor level han a he manager level. Thus nvesors clearly engage n conagon radng bu managers are less ap. Of he fve dfferen reurn benchmarks (Brazl Mexco Asa Russa and he U.S.) Russa clearly has he sronges effecs 15 Usng all observaons ends o ncrease boh pon esmaes and -sascs. 16 Reurns are measured n percen. The quany-adusmen erm n momenum s unransformed (e.g. he 0.25 n he example). Noe ha he quany-adusmen erm uses he average quany n he denomnaor so ha he poson reducon n our parenhecal example s only approxmae. Noe oo ha our L1M measures below are based on monhly reurns no quarerly reurns as n our L0M measures so her sze s correspondngly smaller. 17

19 funds are sysemacally buyng Lan Amercan eques when Russa s reurns are hgh and vce versa. Ths s especally rue durng he Russan Crss whch squares wh nformal accouns of he exraordnarly nense conagon a ha me. Even durng he Russan Crss however fund managers remaned cool-headed: here s no evdence hey engaged n conagon radng. The conemporaneous relaon wh U.S. equy reurns s he only one of he fve reurn benchmarks ha s concenraed a he manager level. I s also he only sgnfcan effec ha s negave. Ths negave LOC sasc for he U.S. reurn mples ha fund managers sysemacally buy Lan Amercan eques when U.S. reurns are low (conrollng for fund nflows/redempons). Though pas work has shown clear lnks beween emergng-marke reurns and U.S. neres raes hs s he frs evdence of whch we are aware ha lnks acual porfolo shfs o U.S. equy reurns. Table 5 focuses on conagon radng durng hree specfc crses: he Mexcan he Asan and he Russan. The reacon of nvesors o Russan equy reurns durng he Russan crss was parcularly srong: nvesors sysemacally sold Lan Amercan eques when Russan equy reurns were low. Noe hough ha hs lnk o Russa s no operave a he manager level. In he case of he Mexcan crss he effec s smaller bu sll sgnfcan and here s some evdence ha managers were nvolved n ha case. In he case of he Asan crss here s no dscernable lnk o he radng of Lan Amercan eques. The las hree columns show he lnk o U.S. marke reurns durng each of hese hree crses. Gven he proxmy o Mexco and he mporance of economc lnks beween he wo counres s no surprsng ha he lnk beween Lan-Amercan porfolos and U.S. reurns s sronges durng he Mexcan crss. Ineresngly he conagon radng sasc s negave and s sgnfcan a boh he manager and nvesor levels. Ths suggess ha durng he Mexcan crss managers and nvesors ended o sell Lan Amercan eques when U.S. reurns were hgh and vce versa. One nerpreaon s ha srong U.S. reurns n he face of Mexco s crss bodes well for Mexcan eques whch nduces a porfolo shf away from he res of Lan Amerca. In closng hs secon on conagon radng s worhwhle re-emphaszng he qualave dfference beween he resuls above and he exsng conagon leraure. The dfference s ha we measure quanes as well as prces and address her on behavor whereas much of he leraure focuses on correlaon n prces only. 18

20 V. Raonalzng Momenum Tradng: Reurn Auocorrelaon? In an envronmen wh posvely auocorrelaed reurns momenum radng s a naural response. The prevous secon presened evdence of posve L0M and a leas durng non-crss perods posve L1M. Ths rases he queson of wheher reurns whn Lan Amerca exhb posve auocorrelaon. One common way o es for reurn auocorrelaon s usng varance raos. If reurns follow a random walk hen reurn varance s a lnear funcon of horzon lengh. Tha s he varance of reurns over k perods s k mes he varance of reurns over one-perod. If nsead reurns are posvely auocorrelaed he varance of k-perod reurns s larger han he sum of one-perod reurns varances grow faser han lnearly. Thus varance raos larger han one are conssen wh raonal posve momenum radng. Alernavely when reurns are negavely auocorrelaed he varance of k-perod reurns s smaller han k mes he varance of one-perod reurns. Varance raos smaller han one would call for negave momenum (or conraran) radng. Table 6 repors he values of he varance-rao es sasc a dfferen horzons ogeher wh p-values for seven Lan Amercan counres. For comparson we also provde resuls for he U.S. sock marke. 17 Ineresngly sock reurns n several Lan Amercan markes are hghly perssen (varance-rao sascs larger han one) even a hree and four-year horzons. In conras U.S. reurns show no perssence a any horzon. Though ceranly no proof ha he posve momenum radng we fnd n Lan Amerca s raonal afer all hs perssence n reurns s a he ndex level hese resuls do pon o he possbly of raonalzng our momenum resuls a leas for some counres (e.g. Mexco Chle Colomba and Venezuela). I s mporan o noe however ha whle posve auocorrelaon s necessary for raonalzng posve L1M s ceranly no necessary for raonalzng posve L0M. As noed n Secon II.1 reurns and rades may be ruly conemporaneous f order flow self s drvng prces. Ths s possble where fund ransacons are large relave o lqudy 17 See Campbell Lo and MacKnlay (1997) for he asympoc dsrbuon of he varance-rao es. 19

21 n he marke (he mperfec subsuably channel noed n foonoe 4) or when fund managers rades are perceved as conanng superor nformaon. VI. Concluson Dscrmnang among he varous ways ha fnancal markes can spread crss requres a sharper pcure of acual behavor. Who s dong he radng? Wha are her radng sraeges? In hs paper we examne porfolos of an mporan class of nernaonal nvesor US muual funds. We address wo ses of quesons. The frs relaes o wheher and when hese funds engage n momenum radng sysemacally buyng wnnng socks and sellng losng socks. We fnd ha nernaonal funds do engage n momenum radng. Ther radng exhbs posve momenum due o momenum a wo levels: he fund manager level and he nvesor level (hrough redempons/nflows). Funds also engage n momenum radng n boh crss and non-crss perods. Conemporaneous momenum radng s sronger durng crses and sronger for fund nvesors han for fund managers. Lagged momenum radng on he oher hand s sronger durng non-crss perods and sronger for managers. The second se of quesons we address relaes o funds use of conagon radng sraeges sellng asses from one counry when asse prces fall n anoher. We fnd ha funds do engage n conagon radng. Per he appendx hs resul s robus o conrollng for own-sock reurns he local-marke facor and he US-marke facor. Srcly speakng whle hese conrols have a sound heorecal bass hey are no suffcen o conclude ha hs conagon radng s non-fundamenal (or pure) conagon radng. In any even we have uncovered several sylzed facs ha are useful for evaluang hypoheses abou he emergng-marke crses and her ransmsson. Beyond hese sylzed facs hs paper ncludes several mehodologcal nnovaons. For example he dsncon beween momenum radng a he manager and nvesor levels s new o he leraure as s our mehod for dsngushng he wo. Our mehod of measurng conagon radng va ransacon quanes s also new. Fnally our regresson-based approach o conrollng for sysemac reurn facors n measurng momenum and conagon radng provdes a valuable check on he blaeral measures robusness. 20

22 An mporan queson we have no addressed s Who akes he oher sde of hese momenum and conagon rades? Someone ceranly mus. Ths queson s unforunaely beyond he feasble scope of our analyss. We can offer some parng houghs however. Consder for example he followng queson: If he model n our managers and nvesors heads s one of undershoong prces followed by posvely auocorrelaed reurns hen mus be ha her couner-pares beleve he oppose model? No hs s no necessary. The leraure n mcrosrucure fnance whch we ouch on n secon II.1 provdes many models of lqudy provders who do no have oppose models or vews hey smply requre compensaon for provdng lqudy n he form of ransacon coss (revenues from her perspecve). I s also approprae o keep n mnd ha ogeher he muual funds we examne own only abou 10 percen of he marke capalzaon of he counres we consder. If hey were a more subsanal fracon hen fndng couner-pares for her rades would be much more dffcul. Indeed he premse ha funds respond o conemporaneous reurns raher han causng hem would be become raher enuous. 21

23 Appendx: A Regresson-Based Approach The bvarae relaons examned va equaons (1)-(7) draw from and herefore allow drec comparson wh pas emprcal work on momenum radng. Bu hese bvarae relaons provde no means of esng on sgnfcance. Is lag-one momenum radng sll sgnfcan afer conrollng for lag-zero momenum radng (.e. afer conrollng for conemporaneous prce effecs)? Is cross-counry conagon radng sll sgnfcan afer conrollng for own-prce effecs va lag-zero and lag-one momenum radng? Are hese relaons robus o ncludng local-marke ndex reurns and he USmarke ndex reurn? A regresson-based approach provdes a naural framework for addressng hese quesons. A he whole-fund level he quesons of he prevous paragraph can be addressed by esmang: 1 = α + β R + β R + β R + β R + β R + ε LA 4 LM 5 US. (A1) Here R and R -1 are own-sock reurns as before. These varables capure lag-zero and lag-one momenum radng respecvely. The varable R LA s he conemporaneous reurn on a Lan Amercan equy ndex. 18 Ths varable capures cross-counry conagon radng. The fourh varable R LM s he local-marke ndex reurn. Ths varable does no ener he analyss nroduced n he prevous secons and s nended here as a conrol for counrylevel sysemac facors. The las varable R US s he US-marke ndex reurn. Ths varable also does no ener n he prevous secons and s nended here as a conrol for sysemac U.S. facors whch have well esablshed effecs on emergng equy markes. A he manager-only and nvesor-only levels he dependen varable n equaon (A1) s replaced wh: 18 We do no aemp o remove he own-counry poron of he broader Lan Amercan ndex. Noe hough ha he own-counry ndex s also n he regresson and our resuls are able o dsngush que sharply beween hem. In fac he own-counry ndex s never sgnfcan so s hghly unlkely he effecs are confounded. 22

24 23 Manager-only: ( ) P P (A2) Invesor-only: ( ) P P 1 1. (A3) Ths follows he separaon of he manager-only and nvesor-only levels n our analyss of bvarae momenum and conagon radng. Resuls Tables A1-A3 presen OLS esmaes of he models n equaons (A1) (A2) and (A3). A he whole-fund level (Table A1) he full sample generaes sgnfcan posve coeffcens on all of he frs hree varables. Thus momenum and conagon radng are robus o movng from bvarae measures o mulvarae measures and ncludng conrols for he overall local and U.S. markes. Ineresngly he local-marke conrol s never sgnfcan. The U.S.-marke conrol n conras s que sgnfcan and negave. Ths squares wh pas emprcal work showng ha U.S. nvesors end o chase emergng markes when reurns a home are low. When he sample s spl no crss and non-crss sub-perods we fnd ha conagon radng s largely a crss-perod phenomenon. Tables A2 and A3 presen resuls for he manager-only and nvesor-only regressons respecvely. A he manager level we fnd sgnfcan posve momenum radng (boh lag zero and lag one) and sgnfcan conagon radng wh respec o he U.S. marke bu no evdence of conagon radng wh respec o oher Lan Amercan markes (β 3 ) excep n mes of crss. Our nvesor-level resuls ell a dsncly dfferen sory. Once we conrol for he local ndex reurn we fnd ha nvesors do no engage n sock-specfc momenum radng. Ths s no surprsng: one would no expec nvesors o respond o ndvdual socks bu o he marke as a whole. They do respond srongly however o he conemporaneous local-ndex reurn. And hey also respond srongly whn

25 he quarer o oher Lan Amercan markes per he sgnfcan posve coeffcen β 3. Noe hough ha hese laer wo effecs are concenraed n he non-crss perods. 24

26 Fgure 1: Muual Funds' Ne Buyng/Sellng of Socks n Lan Amercan Counres Argenna Mexco Jan-93 Jun-94 Oc-95 Mar-97 Jul-98 Dec Jan-93 Jun-94 Oc-95 Brazl Mar-97 Jul-98 Dec-99 Chle Jan-93 Jun-94 Oc-95 Mar-97 Jul-98 Dec-99 Colomba Jan-93 Jun-94 Oc-95 Mar-97 Jul-98 Dec-99 Jan-93 Jun-94 Oc-95 Mar-97 Jul-98 Dec Jan-93 Jun-94 Oc-95 Peru Mar-97 Jul-98 Dec-99 Venezuela Jan-93 Jun-94 Oc-95 Mar-97 Jul-98 Dec-99 Lan Amerca Jan-93 Jun-94 Oc-95 Mar-97 Jul-98 Dec-99 Source: Our daa se. Ne Buyng/Sellng s equal o he value-weghed percenage change n quarerly holdngs of all funds n each counry where he value weghng uses he begnnng-of-perod share prce. All fgures are n percen. However snce quarerly change n he number of shares s dvded by he mean number of shares (a he begnnng and end-of-perod) changes can be greaer han 100 percen. 25

27 Table 1 Lag-0 and Lag-1 Momenum Tradng All Sample Non-Crss Crss Whole-Fund Momenum L0M T-sasc Observaons 2.36*** *** *** L1M T-sasc Observaons ** Manager-Only Momenum L0M T-sasc Observaons 0.86*** *** L1M T-sasc Observaons ** Invesor-Only Momenum L0M T-sasc Observaons 1.70*** *** *** L1M T-sasc Observaons L0M s he pon esmae for he mean lag-0 momenum radng measure. L1M s he pon esmae for he mean lag-1 momenum radng measure (measured from reurn over he prevous monh). Whole-Fund momenum ess wheher he mean of ( / )R -k s zero per equaon (1). Manager-Only momenum conrols for nvesor redempon effecs as n equaon (2). Invesor-Only momenum reflecs only nvesor redempon effecs as n equaon (3). All -sascs are correced for heeroskedascy across funds. Full sample: quarerly daa from Aprl 1993 o January The crss poron of he sample s December 1994-June 1995 July 1997-March 1998 Augus 1998-Ocober 1998 and January The non-crss poron s he res of he sample. The oal of roughly 4400 observaons s 13 funds mes an average of abou 35 socks per fund mes an average of abou 10 quarers of avalable daa per fund. For robusness resuls n each cell are based only on observaons whn hree sandard devaons of he mean. * Sascally Sgnfcan a he 10-percen level ** Sascally Sgnfcan a he 5-percen level *** Sascally Sgnfcan a he 1-percen level 26

28 Table 2 Invesor-Only Momenum a he Fund Level All Sample Non-Crss Crss Invesor-Only: Fund Level L0M T-sasc Observaons 1.99*** *** *** L1M T-sasc Observaons 0.54*** * ** L0M s he pon esmae for he mean lag-0 momenum radng measure. L1M s he pon esmae for he mean lag-1 momenum radng measure (measured from reurn over he prevous monh). Invesor-Only: Fund Level reflecs only nvesor redempon effecs a he fund level as n equaon (4). All -sascs are correced for heeroskedascy across funds. Full sample: quarerly daa from Aprl 1993 o January The crss poron of he sample s December 1994-June 1995 July 1997-March 1998 Augus 1998-Ocober 1998 and January The non-crss poron s he res of he sample. The oal of roughly 127 observaons s 13 funds mes an average of abou 10 quarers of avalable daa per fund. For robusness resuls n each cell are based only on observaons whn hree sandard devaons of he mean. * Sascally Sgnfcan a he 10-percen level ** Sascally Sgnfcan a he 5-percen level *** Sascally Sgnfcan a he 1-percen level 27

29 Table 3 Momenum Tradng Resuls by Crss Mexcan Crss Asan Crss Russan Crss Whole-Fund Momenum L0M T-sasc Observaons 12.11*** *** *** L1M T-sasc Observaons 1.00* Manager-Only Momenum L0M T-sasc Observaons 6.56** ** L1M T-sasc Observaons 1.00*** Invesor-Only Momenum L0M T-sasc Observaons 7.56** ** *** L1M T-sasc Observaons L0M s he pon esmae for he mean lag-0 momenum radng measure. L1M s he pon esmae for he mean lag-1 momenum radng measure (measured from reurn over he prevous monh). Whole-Fund momenum ess wheher he mean of ( / zero per equaon (1). Manager-Only momenum conrols for nvesor redempon effecs as n equaon (2). Invesor-Only momenum reflecs only nvesor redempon effecs as n equaon (3). All -sascs are correced for heeroskedascy across funds. The Mexcan Crss poron of he sample s December 1994-June The Asan Crss poron of he sample s July 1997-March The Russan Crss poron of he sample s Augus 1998-Ocober For robusness resuls n each cell are based only on observaons whn hree sandard devaons of he mean. * Sascally Sgnfcan a he 10-percen level ** Sascally Sgnfcan a he 5-percen level *** Sascally Sgnfcan a he 1-percen level )R -k s 28