NBER WORKING PAPER SERIES DUMB MONEY: MUTUAL FUND FLOWS AND THE CROSS-SECTION OF STOCK RETURNS. Andrea Frazzini Owen A. Lamont

Transcription

1 NBER WORKING PAPER SERIES DUMB MONEY: MUTUAL FUND FLOWS AND THE CROSS-SECTION OF STOCK RETURNS Andrea Frazzn Owen A. Lamont Workng Paper NATIONAL BUREAU OF ECONOMIC RESEARCH 1050 Massachusetts Avenue Cambrdge, MA July 2005 We thank Ncholas Barbers, Judth Chevaler, Davd Musto, Stefan Nagel, and semnar partcpants at Goldman Sachs, the NBER and Yale for helpful comments. We thank Breno Schmdt for research assstance. The vews expressed heren are those of the author(s) and do not necessarly reflect the vews of the Natonal Bureau of Economc Research by Andrea Frazzn and Owen A. Lamont. All rghts reserved. Short sectons of text, not to exceed two paragraphs, may be quoted wthout explct permsson provded that full credt, ncludng notce, s gven to the source.

2 Dumb Money: Mutual Fund Flows and the Cross-Secton of Stock Returns Andrea Frazzn and Owen A. Lamont NBER Workng Paper No July 2005 JEL No. G14, G23, G32 ABSTRACT We use mutual fund flows as a measure for ndvdual nvestor sentment for dfferent stocks, and fnd that hgh sentment predcts low future returns at long horzons. Fund flows are dumb money by reallocatng across dfferent mutual funds, retal nvestors reduce ther wealth n the long run. Ths dumb money effect s strongly related to the value effect. Hgh sentment also s assocated hgh corporate ssuance, nterpretable as companes ncreasng the supply of shares n response to nvestor demand. Andrea Frazzn The Unversty of Chcago Graduate School of Busness 5807 South Woodlawn Avenue Chcago, IL andrea.frazzn@chcagogsb.edu Owen A. Lamont Yale School of Management 135 Prospect Street PO Box New Haven, CT and NBER owen.lamont@yale.edu

3 Indvdual retal nvestors actvely reallocate ther money across dfferent mutual funds. Indvduals tend to transfer money from funds wth low recent returns to funds wth hgh recent returns. In addton to lookng at past returns of funds, ndvduals also may consder economc themes or nvestment styles n reallocatng funds. Collectvely, one can measure ndvdual sentment by lookng at whch funds have nflows and whch have outflows, and can relate ths sentment to dfferent stocks by examnng the holdngs of mutual funds. Ths paper tests whether sentment affects stock prces, and specfcally whether one can predct future stock returns usng a flow-based measure of sentment. If sentment pushes stock prces above fundamental value, hgh sentment stocks should have low future returns. For example, usng our data we calculate that n 1999 nvestors sent $37 bllon to Janus funds but only $16 bllon to Fdelty funds, despte the fact that Fdelty had three tmes the assets under management at the begnnng of the year. Thus n 1999 retal nvestors as a group made an actve allocaton decson to gve greater weght to Janus funds, and n dong so they ncreased ther portfolo weght n tech stocks held by Janus. By 2001, nvestors had changed ther mnds about ther allocatons, and pulled about $12 bllon out of Janus whle addng $31 bllon to Fdelty. In ths nstance, the reallocaton caused wealth destructon to mutual fund nvestors as Janus and tech stocks performed horrbly after Accordng to the smart money hypothess of Gruber (1996) and Zheng (1999), some fund managers have skll and some ndvdual nvestors can detect that skll, and send ther money to sklled managers. Thus (n contrast to the Janus example) flows should be postvely correlated wth future returns. Gruber (1996) and Zheng (1999) show that the short term performance of funds that experence nflows (n the last three months) s sgnfcantly better than those that experence outflows, suggestng that mutual fund nvestors have selecton ablty. Dumb money Page 1

4 Our focus s on stocks, not on funds. We are nterested n how nvestor sentment affects stocks prces, and see fund flows as a convenent (and economcally mportant) measure of sentment. To test whether nvestor sentment causes msprcng, one must test whether hgh sentment today predcts low return n the future, and we focus on cross-sectonal stock return predctablty over perods of months and years. We ask the queston of whether, over the longterm, nvestors are earnng hgher returns as a result of ther reallocaton across funds. For each stock, we calculate the mutual fund ownershp of the stock that s due to reallocaton decsons reflected n fund flows. For example, n December 1999, 18% of the shares outstandng of Csco were owned by the mutual fund sector (usng our sample of funds), of whch 3% was attrbutable to dsproportonately hgh nflows over the prevous 3 years. That s, under certan assumptons, f flows had occurred proportonately to asset value (nstead of dsproportonately to funds lke Janus), the level of mutual fund ownershp would have been only 15%. Ths 3% dfference s our measure of nvestor sentment. We then test whether ths measure predcts dfferental returns on stocks. Our man results are as follows. Frst, as suggested the example of Janus and Csco n 1999, on average from 1980 to 2003, retal nvestors drect ther money to funds whch nvest n stocks that have low future returns. To acheve hgh returns, t s best to do the opposte of these nvestors. We calculate that mutual fund nvestors experence total returns that are sgnfcantly lower due to ther reallocatons. Therefore, mutual fund nvestors are dumb n the sense that ther reallocatons reduce ther wealth on average. We call ths predctablty the dumb money effect. Ths dumb money effect poses a challenge to ratonal theores of fund flows. Second, the dumb money effect s related to the value effect. Money flows nto mutual funds that own growth stocks, and flows out of mutual funds that own value stocks. The value Dumb money Page 2

5 effect explans some, but not all, of the dumb money effect. The fact that flows go nto growth stocks poses a challenge to rsk-based theores of the value effect, whch would need to explan why one class of nvestors (ndvduals) s engaged n a complex dynamc tradng strategy of sellng hgh rsk value stocks and buyng low rsk growth stocks. Thrd, demand by ndvduals and supply from frms are correlated. When ndvduals ndrectly buy more stock of a specfc company (va mutual fund nflows), we also observe that company ncreasng the number of shares outstandng (for example, through seasoned equty offerngs, stock-fnanced mergers, and other ssuance mechansms). Ths pattern s consstent wth the nterpretaton that ndvdual nvestors are dumb, and smart frms are opportunstcally explotng ther demand for shares. These results gve a dfferent perspectve on the ssue of ndvduals vs. nsttutons. A large lterature explores whether nsttutons have better average performance than ndvduals. In the case of mutual funds, for example, Danel, Grnblatt, Ttman, and Wermers (1997) show that stocks held by mutual funds have hgher returns, and Chen, Jegadeesh, and Wermers (2000) show that stocks bought by mutual funds outperform stocks sold by mutual funds. Both results suggest that mutual fund managers have stock-pckng skll. Unfortunately, snce ndvduals ultmately control fund managers, t can be dffcult to nfer the vews of fund managers by lookng only at ther holdngs. For example, when the manager of tech fund experences large nflows, hs job s to buy more technology stocks, even f he thnks the tech sector s overvalued. So f we observe the mutual fund sector as a whole holdng technology stocks, that does not mply that mutual managers as a whole beleve tech stocks wll outperform. It s hard for a fund manager to be smarter than hs clents. Mutual fund holdngs are drven by both manageral choces n pckng stocks and retal nvestor choces n Dumb money Page 3

6 pckng managers. We provde some estmates of the relatve mportance of these two effects. Ths paper s organzed as follows. Secton I revews the lterature. Secton II dscusses the basc measure of sentment, descrbes the data, and examnes the relaton of corporate ssuance and flows. Secton III looks at the relaton between flows and stock returns. Secton IV looks at the relaton between flows and mutual fund returns. Secton V uses calendar tme portfolos to put the results n economc context, showng the magntude of wealth destructon caused by flows and provdng evdence on whether mutual fund managers have stock-pckng skll. Secton VI looks at ssuance by frms. Secton VII presents conclusons. I. Background and lterature revew A. Determnants of fund flows A seres of papers have documented a strong postve relaton between mutual fund past performance and subsequent fund nflows (see, for example, Ippolto (1992), Chevaler and Ellson (1997), and Srr and Tufano (1998)). In addton, retal nvestors appear to allocate ther wealth to funds that have caught ther attenton through marketng (see Jan and Wu (2000), and Barber, Odean and Zheng (2004)), or funds wth names that reflect hot nvestment styles (Cooper, Gulen, and Rau (2005)). Benartz and Thaler (2001) report evdence that retal nvestors employ smple rule-of thumbs n allocatng across dfferent types of mutual funds. For ndvdual stocks, the pcture looks dfferent. Odean (1999), and Barber and Odean (2000, 2001, 2004) present extensve evdence that ndvdual nvestors suffer from based-self attrbuton, and tend be overconfdent, thus engagng n (wealth-destroyng) excessve tradng. But n contrast to ther return-chasng behavor n mutual funds, a varety of evdence suggests that ndvdual nvestors act as contrarans when tradng ndvdual stocks (see Grnblatt and Keloharju (2000), Goetzmann and Massa (2002)). Dumb money Page 4

7 Whle ths apparent contradcton between return-chasng and contraransm s nterestng, the hypothess we wsh to test does not depend on resolvng ths ssue. We are nterested n testng whether ndvdual nvestor sentment predcts future returns, so our hypothess s not contngent on measurng whether nvestors are ultmately return-chasng or not. If ndvdual nvestor sentment causes prces to be wrong and prces eventually revert to fundamental value, then sentment should negatvely predct future returns no matter what whether ndvduals over-react or under-react, whether they return-chase or not. As t turns out, n the data we study, mutual fund flows are ndeed return-chasng, and flows tend to go to stocks that have gone up recently. B. Causal effects of flows on prces There s evdence that fund flows have postvely contemporaneous correlatons wth stock returns (see, for example, Brown et al (2002)). Although t s dffcult to nfer causalty from correlaton, one nterpretaton of ths fact s that nflows drve up stock prces. We do not attempt to test ths hypothess wth our data, for three reasons. Frst, we are nterested n whether sentment causes long-term msprcng, not the short term dynamcs of precsely how tradng affects prces. Second, we observe flows and holdngs at farly low frequency (quarterly), so our data s not well suted to studyng short-term prce dynamcs. Thrd, although the fund flows we consder are certanly economcally large, we vew them as an mperfect measure of sentment snce ndvdual nvestor sentment can be manfested n many other ways. Whle ndvduals were sendng mutual fund money to tech funds n 1999, and thus ndrectly purchasng tech stocks, they may have also been buyng tech stocks drectly n ther brokerage accounts, or nvestng n hedge funds that bought tech stocks. In addton, flows can understate the effects of sentment on the mutual fund sector tself. If Janus experences nflows, then other funds Dumb money Page 5

8 experencng outflows mght seek to mtate Janus n order to appeal to whatever s n fashon. Thus flows are a way to measure sentment, but are not the only channel for sentment to work. Thus the hypothess we wsh to test s that stocks owned by funds wth bg nflows are overprced. These stocks could be overprced because nflows force mutual funds to buy more shares and thus push stock prces hgher, or they could be overprced because overall demand (not just from mutual fund nflows) pushes stock prces hgher. In ether case, nflows reflect the types of stocks wth hgh nvestor demand. C. Styles and sentment A paper closely related to ours s Teo and Woo (2004), who also fnd evdence for a dumb money effect. Followng Barbers and Shlefer (2003), Teo and Woo (2004) consder categorcal thnkng by mutual fund nvestors along the dmensons of large/small or value/growth. They show that when a partcular category has large nflows, stocks n that category subsequently underperform. Lke us, they relate mutual fund flows to stock returns, but unlke us they look only at style returns, not ndvdual stock returns. Whle Teo and Woo (2004) provde valuable evdence, our approach s more general. The beneft s that we do not have to defne specfc styles or categores, such as value/growth. Whle categorcal thnkng and style classfcaton are undoubtedly mportant n determnng fund flows, from a practcal pont of vew t s dffcult for the researcher to dentfy all relevant categores used by nvestors over tme. For example, the growth/value category was not wdely used n Instead, we mpose no categorcal structure on the data and just follow the flows. Most strkngly, we are able to document that the fund flow effect s hghly related to the value effect, a fndng that could not have been dscovered usng the method of Teo and Woo (2004). More generally, one could devse many dfferent measures of nvestor sentment based on Dumb money Page 6

9 prces, returns, or characterstcs of stocks (see for example Baker and Wurgler (2005) and Polk and Sapenza (2004)). If sentment effects stocks prces and creates stock return predctablty (as prces devate from fundamentals and eventually return), as long as tradng volume s not zero t must be that someone somewhere s buyng overprced stocks and sellng underprced stocks. To prove that some class of nvestors overweghts hgh sentment stocks, t s necessary to prove that these nvestors lose money on average from tradng (before tradng costs). Our measure of sentment s based on the actons of one good canddate for sentment-prone nvestors, namely ndvduals. Usng ther trades, we nfer whch stocks are hgh sentment and whch stocks are low sentment. We show that ths class of nvestors does ndeed lose money on average from ther mutual fund reallocatons, confrmng that they are the dumb money who buy hgh sentment stocks. II. Constructng the flow varable Prevous research has focused on dfferent ownershp levels, such as mutual fund ownershp as a fracton of shares outstandng (for example, Chen, Jegadeesh, and Wermers, 2000). We want to devse a measure that s smlar, but s based on flows. Specfcally, we want to take mutual fund ownershp and decompose t nto the porton due to flows and the porton not due to flows. By flows, we mean flows from one fund to another fund (not flows n and out of the entre mutual fund sector). Our central varable s FLOW, the percent of the shares of a gven stock owned by mutual funds that are attrbutable to fund flows. Ths varable s defned as the actual ownershp by mutual funds mnus the ownershp that would have occurred f every fund had receved dentcal proportonal nflows (nstead of experencng dfferent nflows and outflows), every fund manager chose the same portfolo weghts n dfferent stocks as he actually dd, and stock Dumb money Page 7

10 prces were the same as they actually were. We defne the precse formula later, but the followng example shows the basc dea. Suppose at quarter 0, the entre mutual fund sector conssts of two funds: a technology fund wth $20 B n assets and a value fund wth $80 B. Suppose at quarter 1, the technology fund has an nflow of $11 B and has captal gans of $9 B (brngng ts total assets to $40 B), whle the value fund has an outflow of $1 B and captal gans of $1 B (so that ts assets reman constant). Suppose that n quarter 1 we observe the technology fund has 10% of ts assets n Csco, whle the value fund has no shares of Csco. Thus n quarter 1, the mutual fund sector as a whole owns $4 B n Csco. If Csco has $16 B n market captalzaton n quarter 1, the entre mutual fund sector owns 25% of Csco. We now construct a world where nvestors smply allocate flows n proporton to ntal fund asset value. Snce n quarter 0 the total mutual fund sector has $100 B n assets and the total nflow s $10 B, the counterfactual assumpton s that all funds get an nflow equal to 10% of ther ntal asset value. To smplfy, we assume that the flows all occur at the end of the quarter (thus the captal gans earned by the funds are not affected by these nflows). Thus n the counterfactual world the technology fund would receve (.20)*(10) = $2 B (gvng t total assets of $31 B), whle the value fund would receve (.80)*(10) = $8 B (gvng t total assets of $89). In the counterfactual world the total nvestment n CISCO s gven by (.1)*(31) = $3.1, whch s 19.4% of ts market captalzaton. Hence, the FLOW for CISCO, the percent ownershp of Csco due to the non-proportonal allocaton of flows to mutual funds, s = 5.6%. FLOW s an ndcator of what types of stocks are owned by funds experencng bg nflows. It s a number that can be postve, as n ths example, or negatve (f the stock s owned by funds experencng outflows or lower-than-average nflows). It reflects the actve reallocaton Dumb money Page 8

11 decsons by nvestors. What FLOW does not measure s the amount of stock that s purchased wth nflows; one cannot nfer from ths example that the technology fund necessarly used ts nflows to buy Csco. To the contrary, our assumpton n constructng the counterfactual s that mutual fund managers choose ther percent allocaton to dfferent stocks n a way that s ndependent of nflows and outflows. Is t reasonable to assume that managers choose ther portfolo weghts across stocks wthout regard to nflows? Obvously, there are many frctons (for example, taxes and transacton costs) that would cause mutual funds to change ther stock portfolo weghts n dfferent stocks n response to dfferent nflows. Thus, we vew FLOW as an mperfect measure of demand for stocks due to retal sentment. In equlbrum, of course, a world wth dfferent flows would also be a world wth dfferent stock prces, so once cannot nterpret the counterfactual world as an mplementable alternatve for the aggregate mutual fund sector. Later, when we dscuss the effects of flows on nvestor wealth, we consder an ndvdual nvestor (who s too small to affect prces by hmself) who behaves lke the aggregate nvestor. We test whether ths ndvdual representatve nvestor benefts from the actve reallocaton decson mplct n fund flows. For ndvdual nvestors, refranng from actve reallocaton s an mplementable strategy. A. Flows We calculate mutual fund flows usng the CRSP US Mutual Fund Database. The unverse of mutual funds we study ncludes all domestc equty funds that exsts at any date between 1980 and 2003 for whch quarterly net asset values (NAV) are avalable and for whch we can match CRSP data wth the common stock holdngs data from Thomson Fnancal (descrbed n the next subsecton). Snce we do not observe flows drectly, we nfer flows from Dumb money Page 9

12 fund return and NAV as reported by CRSP. Let N be the total NAV of a fund and let R be t t ts return between quarter t 1and quarter t. Followng the standard practce n the lterature (e.g. Zheng (1999), Sapp and Twar (2004)), we compute flows for fund n quarter t, dollar value of net new ssues and redemptons usng F t, as the F t = Nt 1+ Rt ) Nt 1 ( MGN (1) t where MGN s the ncrease n total net assets due to mergers durng quarter t. Note that (1) mplctly assumes that nflows and outflows occur at the end of the quarter, and that exstng nvestors renvest dvdends and other dstrbutons n the fund. We assume that nvestors n the merged funds place ther money n the survvng fund. Funds that are born have nflows equal to ther ntal NAV, whle funds that de have outflows equal to ther termnal NAV. Counterfactual flows are computed under the assumpton that each fund receves a pro rata share of the total dollar flows to the mutual fund sector between date t k and date t, wth the proporton dependng on NAV as of quarter t k. More precsely, n order to compute the FLOW at date t, we start by lookng at the net asset value of the fund at date every date s we track the evoluton of the fund s counterfactual NAV usng: t k. Then, for where Fˆ and N = (2) t k Agg Fˆs F Agg s N t k Nˆ = (1 + R )Nˆ + Fˆ (3) s t k s t t s-1 Nˆ are counterfactual flows and NAV s. s Agg F s the actual aggregate flows for the entre mutual fund sector, whle N s the actual aggregate NAV at date t k. Equatons (2) and (3) descrbe the dynamcs of funds that exst both n quarter Agg t-k t k and n quarter t. For funds Dumb money Page 10

13 that were newly created n the past k quarters, Nˆ s automatcally zero all new funds by defnton represent new flows. The resultng counterfactual net asset value represents the fund sze n a world wth proportonal flows n the last k quarters. Nˆ t at date t For a detaled numercal example of our counterfactual calculatons, see the appendx (whch also dscusses other detals on equatons (2) and (3)). We obtan a quarterly tme seres of counterfactual net asset values for every fund by repeatng the counterfactual exercse every quarter t, and storng the resultng Nˆ t at the end of each rollng wndow. Consder a representatve nvestor who represents a tny fracton, call t q, of the mutual fund sector. Suppose that ths nvestor behaves exactly lke the aggregate of mutual nvestors, sendng flows n and out of dfferent funds at dfferent tmes. The counterfactual strategy descrbed above s an alternatve strategy for ths nvestor, and s mplementable usng the same nformaton and approxmately the same amount of tradng by the nvestor. To mplement ths strategy, ths nvestor only needs to know lagged fund NAV s and aggregate flows. For ths nvestor, q Nˆ s hs dollar holdng n any partcular fund. t In desgnng ths strategy, our am s to create a neutral alternatve to actve reallocaton, whch matches the total flows to the mutual fund sector. One could descrbe ths strategy as a more passve, lower turnover, value-weghtng alternatve to the actve reallocaton strategy pursued by the aggregate nvestor. It s smlar n sprt to the technques of Danel, Grnblatt, Ttman, and Wermers (1999) and Odean (1999) n that t compares the alternatve of actve tradng to a more passve strategy based on lagged asset holdngs. A feature of our counterfactual calculatons s that they do not mechancally depend on the actual performance of the funds. A smpler strategy would have been to smply hold funds n proporton to ther lagged NAV. The problem wth ths strategy s that t mechancally tends to sell funds wth hgh returns Dumb money Page 11

14 and buy funds wth low returns. Snce we wanted to devse a strategy that reflected only flow decsons by nvestors (not return patterns n stocks), we dd not used ths smpler strategy. Let x t be the net asset value of fund n month t as a percentage of total asset of the mutual fund sector: N x t = (4) N t Agg t The counterfactual under proportonal flows s: Nˆ xˆ t = (5) N t Agg t The dfference between x t and xˆ t reflects the actve decsons of nvestors to reallocate money from one manager to another over the past k quarters n a way that s not proportonal to the NAV of the funds. Ths dfference reflects any devaton from value weghtng by the NAV of the fund n markng new contrbutons. In theory, ths dfference could reflect rebalancng away from hgh performng funds and nto poorly performng funds, n order to mantan some fxed weghts (nstead of market weghts). In practce, nvestors tend to unbalance (not rebalance), sendng money from poorly performng funds to hgh performng funds. B. Holdngs Thomson Fnancal provdes the CDA/Spectrum mutual funds database, whch ncludes all regstered domestc mutual funds flng wth the SEC. The data show holdngs of ndvdual funds collected va fund prospectuses and SEC N30D flngs. The holdngs consttute almost all the equty holdngs of the fund (see the appendx for a few small exceptons). The holdngs data n ths study run from January 1980 to December Whle the SEC requres mutual funds to dsclose ther holdngs on a sem-annual bass, approxmately 60% of funds addtonally report quarterly holdngs. The last day of the quarter s Dumb money Page 12

15 most commonly the report day. A typcal fund-quarter-stock observaton would be as follows: as of March 30 th, 1998, Fdelty Magellan owned 20,000 shares of IBM. The holdngs data are notably error-rdden, wth obvous typographcal errors. Furthermore, some reports are mssng from the database. 1 We use a seres of flters to elmnate data errors (see appendx). In matchng the holdngs data to the CRSP mutual fund database, we utlzed fund tckers, fund names and total net asset values. For each fund and each quarter, we calculate as the portfolo weght of fund n stock j based on the latest avalable holdngs data. Hence the portfolos weghts w j reflect fluctuatons of the market prce of the securty held. Let z be the actual percent of the shares outstandng held by the mutual fund sector, w j where Agg z j = x wj Nt / MKTCAPj (6) MKTCAP j s the market captalzaton of frm j. The ownershp that would have occurred wth proportonal flows nto all funds and unchanged fund stock allocaton and stock prces would be Agg z ˆ j = xˆ wj Nt / MKTCAPj (7) For each stock, we calculate our central varable, FLOW, as the percent of the shares outstandng wth mutual fund ownershp attrbutable to flows. The flow of securty j s gven by Agg [ x xˆ ] w N MKTCAPj t FLOW j, t = z j, t zˆ j, t =, t, t j t /, (8) Ths flow has the followng nterpretaton. If each portfolo manager had made exactly the same decsons n terms of percent allocaton of hs total assets to dfferent stocks, and f stock prces were unchanged, but the dollars had flown to each portfolo manager n proporton to ther NAV Dumb money Page 13

16 for the last k perods, then mutual fund ownershp n stock j would be lower by FLOW. Stocks wth hgh FLOW are stocks that are owned by mutual funds that have experenced hgh nflows. C. Descrbng the data We frst descrbe the data for funds. Table I shows the top and bottom funds at year end for two years out of our sample, 1988 and 1999, ranked on the dfference between actual fracton of the fund unverse (x) and counterfactual fracton ( xˆ ). In 1999, the Magellan fund has assets that consttuted 3.8% of our sample mutual fund unverse, but had been recevng below average nflows over the past three years. Had Magellan receved flows n proporton to ts sze over the prevous three years, t would have been 5.4% of the unverse nstead of 3.8%. The table shows that n 1999, the funds recevng bg nflows tended to be technology and growth funds. Table II shows some results for ndvdual frms for the years 1999 and The table shows the top and bottom frms ranked on total dollar flows over the past three years (n the analyss, we focus on flows as a percent of market value, but here we rank on dollar flows n order to generate famlar names). The effect of flows on mutual fund ownershp can be farly szeable, wth flows rasng the total ownershp of Sun Mcrosystems n 1999 from 16.6% to 20.5%. In 1999, stocks wth the bggest nflows tend to be technology stocks, whle stocks wth the bggest outflows tend to be fnancal or manufacturng frms, closely correspond to our perceptons of nvestor sentment n the three year perod endng In contrast, n 1988, technology stocks such as DEC and IBM were experencng outflows, whle consumer goods companes lke RJR and Pllsbury were experencng nflows. Thus sentment favors dfferent types of stocks at dfferent tmes. In nterpretng the flow varable, t s mportant to remember that flow s a relatve concept drven only by dfferences n flows and holdngs across dfferent funds holdng dfferent Dumb money Page 14

17 stocks. Flow s not ntended to capture any noton of the absolute popularty of stock. For example, consder Alcoa n The fact the flow varable s large and negatve n Table II does not mean that Alcoa was unpopular wth mutual funds, nor does t mean that mutual funds were sellng Alcoa. It could be that every mutual fund loved Alcoa, held a lot of t, and bought more of t n What the negatve flow means s that the funds whch overweghted Alcoa n 1999 receved lower-than-average nflows (or perhaps outflows) n Indvdual nvestors favored funds whch tlted toward stocks lke Csco more than funds whch tlted towards stocks lke Alcoa. Table III shows summary statstcs for the dfferent types of data n our sample. Our sample starts n In table III we descrbe statstcs for flows over the past three years, thus the table descrbes data for flows startng n One mportant feature of the sample s ts changng nature over tme. We have more nformaton for the latter part of the sample, so our matchng algorthm system works better n the later years. As shown n the table, our coverage of large stocks s much more complete n the early part of the sample: n 1983, we have flow data for 92% of the unverse of stocks on a value-weghted bass, but only 47% on an equal weghted bass. 2 Ths dfference partally reflects the fact that funds tend to own large stocks, but also the fact that we are falng to match some small funds n the early part of the sample perod. One possble concern s survvorshp bas, whch we address n secton IV by usng a method that does not nvolve matchng funds wth stocks. Table III shows summary statstcs for three year flows. One way of descrbng FLOW s that t s the actual percent ownershp by the mutual fund sector, mnus the counterfactual percent ownershp. Snce the actual percent ownershp s bounded above by 100%, FLOW s bounded above by 100%. In the counterfactual case, there s no accountng dentty enforcng that the Dumb money Page 15

18 dollar value of fund holdngs s less than the market captalzaton of the stock. Thus FLOW s unbounded below. Values of FLOW less than -100% are very rare, occurrng less than 0.01% of the tme for three year flows. III. Flows and stock returns A. Excess returns Table IV shows the basc results of ths paper. We form calendar tme portfolos and examne monthly excess returns on portfolos constructed usng our flow measure. We show both equally weghted returns and value weghted returns n month t for fve portfolos formed by sortng on the latest avalable flows as of month t 1. The table shows flows over horzons stretchng from three months (one quarter, the shortest nterval we have for calculatng flows) to fve years. The rghtmost column shows the dfference between the hgh flow stocks and the low flow stocks. Quntle one s the bottom 20 percent of all stocks sorted on flows. It turns out that, for long-horzon flows, the bottom quntle reflects stocks that are not just experencng lower-thanaverage nflows, they are experencng outflows. That s, quntle one contans stocks that ndvdual nvestors are sellng (ndrectly va mutual funds) and quntle fve contans stocks that ndvduals are most heavly buyng. Lookng at the dfference between hgh flow and low flow stocks, t s strkng that for every horzon but three months, hgh flows today predct low future stock returns. Ths relaton s statstcally sgnfcant at the three and fve year horzon. Ths dumb money effect s szeable: lookng at three-year, equal weght results, the dfference between hgh flow and low flow stocks s 61 bass ponts per month or approxmately 8 percent per year. Remarkably, the dumb money effect s slghtly larger usng value weghtng nstead of equal weghtng. Ths result stands n Dumb money Page 16

19 contrast to many other patterns n stock returns, whch tend to be concentrated n small cap stocks. 3 Perhaps surprsngly, n ths table we fnd no sold evdence for the smart money effect n raw returns, even at the horzons of three to twelve months where one mght expect prce momentum to domnate. Gruber (1996) and Zheng (1999) look at quarterly flows and fnd that hgh flows predct hgh returns: one can see a hnt of ths n the three month flow results, although one cannot reject the null hypothess. We return to ths ssue later; t turns out that ths partcular result s senstve to alternatve methods of measurng, and one can fnd specfcatons wth a sgnfcant smart money effect at short horzons. Fgure 1 shows how flows predct returns at varous dfferent horzons. We show the average returns n month t + k on long/short portfolos formed on three month flows n month t. The fgure shows average returns over tme wth accompanyng 95% confdence nterval. For k < 0, the fgure shows how lagged returns predct today's flows. The fgure shows that flows nto an ndvdual stock are very strongly nfluenced by past returns on that stock. Ths result s expected gven the prevous lterature documentng hgh nflows to hgh performng funds. Flows tend to go to funds that have hgh past returns, and snce funds returns are drven by the stocks that they own, flows tend to go to stocks that have hgh past returns. It appears that returns over the past twelve months are especally mportant, wth the effect decreasng as one goes earler than a year. For k > 0, the fgure shows the (nsgnfcant) smart money effect at one month horzon, becomng a sgnfcant dumb money effect after about a year has passed. The predctable negatve returns persst for about a year after that, then fade away. We focus on the three-year results n Table IV. Whch horzon s t approprate to focus on? Why do we focus on the three-year results (where nflows negatvely forecast returns) Dumb money Page 17

20 nstead of the three month results (where nflows postvely forecast returns)? Isn't the horzon arbtrary? Ths s an mportant queston. The answer s that the horzon one should use depends on what one s tryng to measure. Snce our goal s to understand the long-term effects on tradng on ndvdual nvestor wealth, the longer the horzon, the better. The results for longer horzons show that although mutual fund flows do seem to predct short term returns, ths effect s swamped as we look longer horzons (whch cumulate the returns over tme). To understand whether ndvduals are "smart" or "dumb", one needs to measure whether ther tradng s rasng or lowerng ther total wealth (compared to some alternatve nvolvng refranng from tradng) over ther lfetme. To take an example, suppose Joe buys a stock from Sally for $10, and the next day the prce rses to $11. Based on ths evdence alone, one mght conclude that Joe s smart (and Sally s dumb). However, f Joe contnues to hold the stock, and t declnes to $5 (at whch pont he sells t), Joe seems less smart. Joe could ncrease hs wealth by refranng from trade. In ths sense, longer horzons are better horzons for nferrng the net effect of Joe's tradng. In terms of measurng nvestor experence, however, the evdence gven n Table IV cannot resolve the queston of smart or dumb, because ths evdence does not correspond to the dollar holdngs of any class of nvestors. One needs to look at all trades and all dollar allocatons to dfferent securtes over tme. In secton V, we do ths for the aggregate mutual fund nvestor, and show that tradng does n fact decrease both average returns and the return/rsk rato for an ndvdual who s behavng lke the aggregate mutual fund nvestor. In turns out that when one looks at the whole portfolo, the answer no longer depends on the horzon: the dumb money effect exsts at all horzons. From ths perspectve, then, ndvdual nvestors n aggregate are unambguously dumb. Dumb money Page 18

21 B. Robustness tests Fgure 2 shows evdence on the dumb money effect over tme. It shows cumulatve returns on the long/short portfolo formed on three year equal weghted flows (we smply sum each monthly return over tme). Fgure 2 shows that the last few years of the sample are qute nfluental, wth about 100 out of the total of 150 percent comng after But the dumb money effect s not due only to these years. The dfferental return s negatve for 17 of the 21 calendar years avalable. One nterpretaton of the tme pattern s that the perod around was a tme of partcularly hgh rratonalty, when rratonal traders earned partcularly low returns. 4 Many anomales grew larger n ths perod (see Ofek and Rchardson (2003)). Indeed, one mght propose that f a return pattern does not grow stronger n ths perod, then t s probably not attrbutable to rratonal behavor. Table V shows robustness tests. Frst, to address the evdence n Fgure 2, we splt the sample nto pre-1998 and post-1998 perods. Although the magntude of the dumb money effect s much larger n the post-1998 perod, t has the same sgn n both subperods. For equal weghted portfolos, the three year dumb money effect remans statstcally sgnfcant n both subperods, whle for value weghted portfolos (although the magntudes are smlar to equal weghted) the effect s not sgnfcantly dfferent from zero n the frst part of the sample. Table V also shows results for the sample of stocks whch have market cap above and below the CRSP medan market cap. Agan, over the full sample the dumb money effect remans qute strong among large cap stocks. One mght ask whether the dumb money effect s an mplementable strategy for outsde nvestors usng nformaton avalable n real tme. Our methodology nvolves substantal bult-n staleness of flows, largely reflectng the way that Thomson Fnancal has structured the data. 5 So Dumb money Page 19

22 the varables n Table IV are certanly n the nformaton set of any nvestor who has access to all the regulatory flngs and reports from mutual funds, as they are fled. Currently, holdngs data appear on the SEC EDGAR system on the next busness days followng a flng, but nformaton lags were probably longer at the begnnng of the sample perod. To address ths ssue, Table V shows results wth the flow varables generously lagged an addtonal twelve months. Even lagged a full year, the three year flow varable remans a statstcally sgnfcant predctor of equal weght returns and close to sgnfcant for value weght returns (the addtonal laggng decreases the number of avalable observatons, makng nference more dffcult). As one mght expect gven Fgure 1, the laggng produces a substantal and sgnfcant dumb money effect at the short horzons as well. Thus the dumb money effect s not prmarly about short-term nformaton contaned n flows, t s about long-term msprcng. In summary, three year mutual fund flows strongly negatvely predct future stock returns, and there s no horzon at whch flows relable postvely predct excess returns. The dumb money effect s present n both large and small cap stocks, and present n dfferent tme perods. C. Controllng for sze, momentum, and value Table VI shows results for returns controllng for sze, value, and prce momentum. These varables are known to predct returns and lkely to be correlated wth flows. Sapp and Twar (2004), for example, argue that the short-horzon smart money effect merely reflects the prce momentum effect of Jegadeesh and Ttman (1993). If an ndvdual follows a strategy of sendng money to funds wth past hgh returns n the last year and wthdrawng money from funds wth low returns, then he wll end up wth a portfolo that overweghts hgh momentum stocks and underweghts low momentum stocks. Ths strategy mght be a smart (although hgh Dumb money Page 20

23 turnover) strategy to follow, as long as he keeps rebalancng the strategy. However, f the ndvdual fals to rebalance promptly, eventually he wll be holdng a portfolo wth a strong growth tlt. Thus over long horzons, stocks wth hgh nflows are lkely to be stocks wth hgh past returns and are therefore lkely to be growth stocks. So t s useful to know whether flows have ncremental forecastng power for returns or just reflect known patterns of short horzon momentum and long horzon value/reversals n stock returns. The left hand sde of Table VI shows results where returns have been adjusted to control for value, sze, and momentum. Followng Danel, Grnblatt, Ttman, and Wermers (1997), t subtracts from each stock return the return on a portfolo of frms matched on market equty, market-book, and pror one-year return quntles (a total of 125 matchng portfolos). 6 Here the dumb money effect s substantally reduced, wth the coeffcent fallng from to for three year equal weghted flows, stll sgnfcantly negatve but less than half as large. Ths reducton largely reflects the fact that (as we shall see) hgh sentment stocks tend to be stocks wth hgh market-book. The rght-hand sde of Table VI shows alphas from a Fama and French (1993) three factor regresson. Here the reducton of the long-horzon dumb money effect s not as substantal, as the three-year equal weghted dfferental return falls from to Both methods cause the smart money effect to rse for equal weght returns, although t s stll below conventonal sgnfcance levels. In Table VII, we take a closer look at the relaton between the dumb money effect and the value effect by ndependently sortng all stocks nto fve flow categores and fve market-book categores, wth a resultng 25 portfolos. We sort on three year flows, and on market-book rato followng the defnton of Fama and French (1993). The rght-most column n each panel shows Dumb money Page 21

24 whether there s a flow effect wthn market-to-book quntles. Thus f the value effect subsumes the dumb money effect, ths column should be all zeros. The bottom row n each panel shows whether there s a value effect controllng for flows. If the dumb money effect subsumes the value effect, ths row should be all zeros. If the two effects are statstcally ndstngushable, then both the row and the column should be all zeros. Table VII shows that, generally, nether effect domnates the other. Lookng at equal weghted returns, the value effect appears to be much larger than the dumb money effect, wth magntudes of approxmately one percent per month and hgh t-statstcs. As before, the dumb money effect survves the correcton for market-book. However, lookng at value weghted returns, the dumb money effect looks stronger than value effect, wth smlar magntudes and generally hgher t-statstcs. Table VIII shows double sort portfolos for three year past stock returns nstead of market-book, to explore the reversal effect of De Bondt and Thaler (1985). In order to make the reversal effect as powerful as possble, we sort on past returns lagged one year (n other words, we sort on stock returns from month t-48 to t-12). Here, lookng at equal weghted returns, the results are smlar to Table VII, wth the dumb money effect lookng slghtly weaker. Lookng at value weghted returns, the dumb money effect and the reversal effect have smlar magntudes and levels of sgnfcance. To summarze, usng standard adjustment technques, the dumb money effect s not completely explaned by the value effect. Nether the dumb money effect nor the value/reversal effect domnates the other. However, the dumb money and value/reversal effect are clearly qute related, and perhaps reflect the same underlyng phenomenon. Dumb money Page 22

25 IV. Flows and mutual fund returns In ths secton, we set asde our man focus on stock returns, and examne the relaton between mutual fund flows and mutual fund returns. Ths evdence s useful for two purposes. Frst, t shows how our results relate to the prevous work of Zheng (1999) and Gruber (1996). Second, t shows whether our results are drven by problems n matchng the CRSP mutual fund database wth the holdngs database. Table IX shows results usng monthly mutual fund returns (nstead of stock returns) and sortng on flows nto funds nstead of flows nto stocks. The mutual fund returns reflect, n addton to the returns of the stocks held by the fund, the expenses and tradng costs of each fund. The unverse of funds ncludes all domestc equty funds n the CRSP mutual fund database. We show returns for both equally weghted and value weghted portfolos of funds (where the value weghts reflect the NAV of the fund). We frst sort on actual flows mnus counterfactual flows. Table IX shows frst, usng excess returns, that the dumb money effect comes n farly strongly at the 3 year horzon, whle the smart money effect comes n weakly at the 3 month horzon. Turnng next to three-factor alphas, here the smart money effect comes n sgnfcant at the three-month and sx-month horzon, whle the dumb money effect s weaker for equal weghted results, whle stll strong for value weghted results. As a robustness check, we also sort on actual nflows (dollar nflows dvded by assets under management) nstead of actual nflows mnus counterfactual nflows. Ths slghtly dfferent sortng most closely corresponds to the method of Zheng (1999) and Gruber (1996). The results are about the same usng ths sortng varable. How should one nterpret these results? Take for example the equal weghted 3-factor alpha results, where three month nflows predct a postve and sgnfcant dfferental of 19 bass ponts per month, whle three year nflows predct a negatve but nsgnfcant 10 bass ponts. Dumb money Page 23

26 Suppose one beleves that the Fama-French (1993) model s an approprate rsk adjustment. The fact that the dfferental s percent for three year nflows means that the tradng of ndvduals s not helpng them acheve hgher rsk-adjusted average returns. Despte the fact that ndvduals earn sgnfcant and postve 0.19 percent dfferental n the frst three months, ths outperformance s wasted because the ndvduals are not followng a dynamc strategy of buyng the best-performng funds, holdng them for a quarter, and them sellng them. Instead, they are n aggregate followng a strategy of buyng the best-performng funds, and holdng them for a long perod of tme. So the longer horzon return shows that nvestors are not actually beneftng from ther tradng. To summarze, lookng at mutual fund returns, there s a strong dumb money effect among large funds (when value weghtng). Lookng at smaller funds (equal weghtng), the dumb money effect s weaker, especally when correctng for value. Smlar to prevous results, we fnd a smart money effect at the quarterly horzon. However, ths smart money effect s not enough to boost nvestor returns over the long term. For a more economcally relevant measure of how these two effects balance out, n the next secton we look at how the aggregate mutual fund nvestor s helped or hurt by hs tradng. V. Economc sgnfcance to the aggregate nvestor A. The magntude of wealth destructon So far, we have shown that stocks owned by funds wth large nflows have poor subsequent returns. In ths secton, we measure the wealth consequences of actve reallocaton across funds, for the average nvestor. We assess the economc sgnfcance by measurng the average return earned by a representatve nvestor, and comparng t to the return he could have earned by smply refranng from engagng n non-proportonal flows. We examne both returns Dumb money Page 24

27 on stocks and returns on mutual funds. Defne R ACTUAL as the return earned by a representatve mutual nvestor who owns a tny fracton of each exstng mutual fund. The returns would reflect a portfolo of stocks where the portfolo weghts reflect the portfolo weghts of the aggregate mutual fund sector: = j, wj t Rt (9) j x t ACTUAL R t, where R J s the return on stock j. The return from a strategy of refranng from non-proportonal flows, R NOFLOW, s NOFLOW j R t = xˆ, t wj, t Rt (10) j We use three year flows n these calculatons. Table X shows excess returns on these two portfolos, and for comparson shows the value weghted market return as well. Snce the two mutual fund portfolos use weghts based on dollar holdngs, they are of course qute smlar to each other and to the market portfolo. Although very smlar, these portfolos are not dentcal. Table X shows nvestor flows cause a sgnfcant reducton n both average returns and Sharpe ratos earned by mutual fund nvestors. A representatve nvestor who s currently behavng lke the aggregate mutual fund sector could ncrease hs Sharpe rato 9% (from a monthly Sharpe rato of to 0.149) by refranng from actve reallocaton and just drectng hs flows proportonally. 7 One can assess the sgnfcance of ths dfference n mean returns by lookng at the returns on the long-short portfolo R ACTUAL - R NOFLOW. Ths return s smlar to the long-short portfolo studed n Table IV, except that here all stocks owned by the mutual fund sector are ncluded, and the weghts are proportonal to the dollar value of the holdngs. The dfferental returns are negatve and hghly sgnfcant. Thus nvestor flows cause wealth destructon. Ths Dumb money Page 25

28 concluson s, of course, a partal equlbrum statement. If all nvestors swtched to proportonal flows, presumably stock prces would change to reflect that. But for one ndvdual nvestor, t appears that fund flows are harmful to wealth. B. Better dentfcaton of mutual fund manager skll Table X also helps dsentangle the effect of flows from the effect of manager stock pckng. We start by consderng the average of R ACTUAL R M, whch measures the net return beneft of ownng the aggregate fund holdngs nstead of holdng the market (gnorng tradng costs and expenses). R M s the return on the CRSP value weghted market. The average of ths dfference, 0.03, conssts of two components. The frst, R ACTUAL - R NOFLOW, s the net beneft of reallocatons. We already have seen that ths dumb money effect s negatve. The second, R NOFLOW R M, measures the ablty of the mutual fund managers to pck stocks whch outperform the market (usng value weghts for managers). As shown n the table, usng raw returns, ths stock pckng effect s 0.08 per month, wth a t-statstc of 1.8. Thus there s some modest evdence that mutual fund managers do have the ablty to pck stocks that outperform the market, once one controls for ther clents tendences of swtchng money from one fund to another. As shown n the table, ths modest skll s obscured (when lookng only at actual holdngs) by ther clents ant-skll at pckng funds. C. Dfferent measures of economc sgnfcance We explore the robustness of the economc sgnfcance n two ways. Frst, n the bottom part of Table X, we repeat the basc analyss, agan usng three year flows but usng funds nstead of stocks. We defne R ACTUAL and R NOFLOW usng fund returns nstead of stock returns (pluggng n actual fund returns for the term n brackets n equatons (9) and (10)). Agan, as n secton IV, usng mutual fund returns allows us to avod ssues nvolvng matchng funds wth Dumb money Page 26