Study on Mutual Fund Flows and The Cross-Section of Stock Returns

Description
We use mutual fund flows as a measure of individual investor sentiment for different stocks, and find that high sentiment predicts low future returns. Fund flows are dumb money–by reallocating across different mutual funds, retail investors reduce their wealth in the long run.

Journal of Financial Economics 88 (2008) 299–322
Dumb money: Mutual fund ?ows and the cross-section
of stock returns
$
Andrea Frazzini
a,Ã
, Owen A. Lamont
b
a
University of Chicago, Graduate School of Business, 5807 South Woodlawn Avenue, Chicago, IL 60637, USA
b
Yale School of Management, 135 Prospect Street, New Haven, CT 06520, USA
Received 21 September 2005; received in revised form 22 May 2007; accepted 9 July 2007
Available online 23 February 2008
Abstract
We use mutual fund ?ows as a measure of individual investor sentiment for different stocks, and ?nd that high sentiment
predicts low future returns. Fund ?ows are dumb money–by reallocating across different mutual funds, retail investors
reduce their wealth in the long run. This dumb money effect is related to the value effect: high sentiment stocks tend to be
growth stocks. High sentiment also is associated with high corporate issuance, interpretable as companies increasing the
supply of shares in response to investor demand.
r 2008 Elsevier B.V. All rights reserved.
JEL classi?cation: G14; G23; G32
Keywords: Mutual fund; Individual investors
1. Introduction
Individual retail investors actively reallocate their money across different mutual funds. One can measure
individual sentiment by looking at which funds have in?ows and which have out?ows, and can relate this
sentiment to different stocks by examining the holdings of mutual funds. This paper tests whether sentiment
affects stock prices, and speci?cally whether one can predict future stock returns using a ?ow-based measure
of sentiment. If sentiment pushes stock prices above fundamental value, high sentiment stocks should have
low future returns.
For example, using our data we calculate that in 1999 investors sent $37 billion to Janus funds but only $16
billion to Fidelity funds, despite the fact that Fidelity had three times the assets under management at the
ARTICLE IN PRESS
www.elsevier.com/locate/jfec
0304-405X/$ - see front matter r 2008 Elsevier B.V. All rights reserved.
doi:10.1016/j.j?neco.2007.07.001
$
We thank an anonymous referee, Nicholas Barberis, Judith Chevalier, Christopher Malloy, David Musto, Stefan Nagel, Jeffrey
Pontiff, and seminar participants at Boston University, Barclays Global Investors, Chicago Quantitative Alliance, Goldman Sachs,
Harvard, NBER, University of Chicago, University of Illinois, University of Minnesota, University of Tilburg and Yale for helpful
comments. We thank Breno Schmidt for research assistance. We thank Randy Cohen, Josh Coval and Lubos Pastor, and Antti Petajisto
for sharing data with us.
Ã
Corresponding author.
E-mail address: [email protected] (A. Frazzini).
beginning of the year. Thus, in 1999 retail investors as a group made an active allocation decision to give
greater weight to Janus funds, and in doing so they increased their portfolio weight in tech stocks held by
Janus. By 2001, investors had changed their minds about their allocations, and pulled about $12 billion out of
Janus while adding $31 billion to Fidelity. In this instance, the reallocation caused wealth destruction to
mutual fund investors as Janus and tech stocks performed horribly after 1999.
To systematically test the hypothesis that high sentiment predicts low future returns, we examine ?ows and
stock returns over the period 1980–2003. For each stock, we calculate the mutual fund ownership of the stock
that is due to reallocation decisions re?ected in fund ?ows. For example, in December 1999, 18% of the shares
outstanding of Cisco were owned by the mutual fund sector (using our sample of funds), of which 3% was
attributable to disproportionately high in?ows over the previous three years. That is, under certain
assumptions, if ?ows had occurred proportionately to asset value (instead of disproportionately to funds like
Janus), the level of mutual fund ownership would have been only 15%. This 3% difference is our measure of
investor sentiment. We then test whether this measure predicts differential returns on stocks.
Our main result is that on average, retail investors direct their money to funds which invest in stocks that
have low future returns. To achieve high returns, it is best to do the opposite of these investors. We calculate
that mutual fund investors experience total returns that are signi?cantly lower due to their reallocations.
Therefore, mutual fund investors are ‘‘dumb’’ in the sense that their reallocations reduce their wealth on
average. We call this predictability the ‘‘dumb money’’ effect.
Our results contradict the ‘‘smart money’’ hypothesis of Gruber (1996) and Zheng (1999) that some fund
managers have skill and some individual investors can detect that skill, and send their money to skilled
managers. Gruber (1996) and Zheng (1999) show that the short term performance of funds that experience
in?ows is signi?cantly better than those that experience out?ows, suggesting that mutual fund investors have
selection ability. We ?nd that this smart money effect is con?ned to short horizons of about one quarter, but at
longer horizons the dumb money effect dominates.
We show that the dumb money effect is related to the value effect. This relation re?ects return-chasing
?ows. A series of papers have documented a strong positive relation between mutual fund past performance
and subsequent fund in?ows (see, for example, Ippolito, 1992; Chevalier and Ellison, 1997; Sirri and Tufano,
1998). As a consequence, money ?ows into mutual funds that own growth stocks, and ?ows out of mutual
funds that own value stocks. The value effect explains some, but not all, of the dumb money effect. The fact
that ?ows go into growth stocks poses a challenge to risk-based theories of the value effect, which would need
to explain why one class of investors (individuals) is engaged in a complex dynamic trading strategy of selling
‘‘high risk’’ value stocks and buying ‘‘low risk’’ growth stocks.
In addition to past returns of funds, decisions by individual investors also re?ect their thinking about
economic themes or investment styles, reinforced by marketing efforts by funds (see Jain and Wu, 2000;
Barber, Odean, and Zheng, 2004; Cooper, Huseyin, and Rau, 2005). A paper closely related to ours is Teo and
Woo (2004), who also ?nd evidence for a dumb money effect. Following Barberis and Shleifer (2004), Teo and
Woo (2004) consider categorical thinking by mutual fund investors along the dimensions of large/small or
value/growth. While Teo and Woo (2004) provide valuable evidence, our approach is more general. We do not
have to de?ne speci?c styles or categories. Instead, we impose no categorical structure on the data and just
follow the ?ows.
More generally, one can imagine many different measures of investor sentiment based on prices, returns, or
characteristics of stocks (see, for example, Baker and Wurgler, 2006; Polk and Sapienza, 2008). Our measure is
different because it is based on trading by a speci?c set of investors, and thus allows us to perform an
additional test con?rming that sentiment-prone investors lose money from their trading. If sentiment affects
stocks prices and creates stock return predictability (as prices deviate from fundamentals and eventually
return), as long as trading volume is not zero, it must be that someone somewhere is buying overpriced stocks
and selling underpriced stocks. If some class of investors drives sentiment, it is necessary to prove that these
investors lose money on average from trading (before trading costs).
Our measure of sentiment is based on the actions of one good candidate for sentiment-prone investors,
namely individuals. Using their trades, we infer which stocks are high sentiment and which stocks are low
sentiment. We show that this class of investors does indeed lose money on average from their mutual fund
reallocations, con?rming that they are the dumb money who buy high sentiment stocks. Individual retail
ARTICLE IN PRESS
A. Frazzini, O.A. Lamont / Journal of Financial Economics 88 (2008) 299–322 300
investors are good candidates for sentiment-prone investors because a variety of evidence indicates they make
suboptimal investment decisions. Odean (1999), and Barber and Odean (2000, 2001, 2004) present extensive
evidence that individual investors suffer from biased-self attribution, and tend to be overcon?dent, thus
engaging in (wealth-destroying) excessive trading (see also Grinblatt and Keloharju, 2000; Goetzmann and
Massa, 2002).
If individuals are losing money via their mutual fund trades, who is making money? One candidate is
institutional investors. A large literature explores whether institutions have better average performance than
individuals (see Daniel, Grinblatt, Titman, and Wermers, 1997; Chen, Jegadeesh, and Wermers, 2000).
Unfortunately, since individuals ultimately control fund managers, it can be dif?cult to infer the skills of the
two groups. It is hard for a fund manager to be smarter than his clients. Mutual fund holdings and
performance are driven by both managerial choices in picking stocks and retail investor choices in picking
managers. We provide some estimates of the relative importance of these two effects.
We ?nd that demand by individuals and supply from ?rms are correlated. When individuals indirectly buy
more stock of a speci?c company (via mutual fund in?ows), we also observe that the company increases the
number of shares outstanding (for example, through seasoned equity offerings, stock-?nanced mergers, and
other issuance mechanisms). One interpretation is that individual investors are dumb, and smart ?rms are
opportunistically exploiting their demand for shares.
Although we ?nd that sentiment affects stock prices, we do not attempt to analyze precisely the mechanism
by which sentiment is propagated. Fund ?ows have positive contemporaneous correlations with stock returns
(see, for example, Warther, 1995; Brown, Goetzmann, Hiraki, Shiraishi, and Watanabe, 2002). Although it is
dif?cult to infer causality from correlation, one interpretation is that in?ows drive up stock prices. Wermers
(1999, 2004) presents evidence consistent with ?ow-related additions to existing positions pushing up stock
prices. Coval and Stafford (2007) found evidence of price pressure in securities held in common by distressed
funds when managers are forced to unwind their positions in response to large out?ows, or expand existing
positions in response to large in?ows. We do not test this hypothesis nor draw a causal link between the price
impact of individual funds and future stocks returns. Instead, the hypothesis we wish to test is that stocks
owned by funds with big in?ows are overpriced. We use fund ?ows to construct a measure of investors’
demand for the underlying securities and test the hypothesis that stocks ranking high in popularity have low
future returns. These stocks could be overpriced because in?ows force mutual funds to buy more shares and
thus push stock prices higher, or they could be overpriced because overall demand (not just from mutual fund
in?ows) pushes stock prices higher. In either case, in?ows re?ect the types of stocks with high investor
demand.
This paper is organized as follows. Section 2 discusses the basic measure of sentiment. Section 3 looks at the
relation between ?ows and stock returns. Section 4 looks at a variety of robustness tests. Section 4 puts the
results in economic context, showing the magnitude of wealth destruction caused by ?ows. Section 6 looks at
issuance by ?rms. Section 7 presents conclusions.
2. Constructing the ?ow variable
Previous research has focused on different ownership levels, such as mutual fund ownership as a fraction of
shares outstanding (for example, Chen, Jegadeesh, and Wermers, 2000). We want to devise a measure that is
similar, but is based on ?ows. Speci?cally, we want to take mutual fund ownership and decompose it into the
portion due to ?ows and the portion not due to ?ows. By ‘‘?ows,’’ we mean ?ows from one fund to another
fund (not ?ows in and out of the entire mutual fund sector).
Our central variable is FLOW, the percent of the shares of a given stock owned by mutual funds that is
attributable to fund ?ows. This variable is de?ned as the actual ownership by mutual funds minus the
ownership that would have occurred if every fund had received identical proportional in?ows, every fund
manager chose the same portfolio weights in different stocks as he actually did, and stock prices were the same
as they actually were. We de?ne the precise formula later, but the following example shows the basic idea.
Suppose at quarter 0, the entire mutual fund sector consists of two funds: a technology fund with $20 billion
in assets and a value fund with $80 billion. Suppose at quarter 1, the technology fund has an in?ow of $11
billion and has capital gains of $9 billion (bringing its total assets to $40 billion), while the value fund has an
ARTICLE IN PRESS
A. Frazzini, O.A. Lamont / Journal of Financial Economics 88 (2008) 299–322 301
out?ow of $1 billion and capital gains of $1 billion (so that its assets remain constant). Suppose that in quarter
1 we observe that the technology fund has 10% of its assets in Cisco, while the value fund has no shares of
Cisco. Thus in quarter 1, the mutual fund sector as a whole owns $4 billion in Cisco. If Cisco has $16 billion in
market capitalization in quarter 1, the entire mutual fund sector owns 25% of Cisco.
We now construct a world where investors simply allocate ?ows in proportion to initial fund asset value.
Since in quarter 0 the total mutual fund sector has $100 billion in assets and the total in?ow is $10 billion, the
counterfactual assumption is that all funds get an in?ow equal to 10% of their initial asset value. To simplify,
we assume that the ?ows all occur at the end of the quarter (thus the capital gains earned by the funds are not
affected by these in?ows). Thus, in the counterfactual world, the technology fund would receive
(.20)Ã(10) ¼ $2 billion (giving it total assets of $31 billion), while the value fund would receive
(.80)Ã(10) ¼ $8 billion (giving it total assets of $89 billion). In the counterfactual world the total investment
in Cisco is given by (.1)Ã(31) ¼ $3.1 billion, which is 19.4% of its market capitalization. Hence, the FLOW for
Cisco, the percent ownership of Cisco due to the non-proportional allocation of ?ows to mutual funds, is
25–19.4 ¼ 5.6%.
FLOW is an indicator of what types of stocks are owned by funds experiencing big in?ows. It can be
positive, as in this example, or negative (if the stock is owned by funds experiencing out?ows or lower-than-
average in?ows). It re?ects the active reallocation decisions by investors. What FLOW does not measure is the
amount of stock that is purchased with in?ows; one cannot infer from this example that the technology fund
necessarily used its in?ows to buy Cisco. To the contrary, our assumption in constructing the counterfactual is
that mutual fund managers choose their percent allocation to different stocks in a way that is independent of
in?ows and out?ows. Obviously, there are many frictions (for example, taxes and transaction costs) that
would cause mutual funds to change their stock portfolio weights in different stocks in response to different
in?ows. Thus, we view FLOW as an imperfect measure of demand for stocks due to retail sentiment.
In equilibrium, of course, a world with different ?ows would also be a world with different stock prices, so
one cannot interpret the counterfactual world as an implementable alternative for the aggregate mutual fund
sector. In Section 5, we discuss the effects of ?ows on investor wealth and consider an individual investor (who
is too small to in?uence prices by himself) who behaves like the aggregate investor. We test whether this
individual representative investor bene?ts from the active reallocation decision implicit in fund ?ows. For
individual investors, refraining from active reallocation is an implementable strategy.
2.1. Flows
We calculate mutual fund ?ows using the CRSP Mutual Fund Database. The universe of mutual funds we
study includes all domestic equity funds that exist at any date between 1980 and 2003 for which quarterly total
net assets (TNA) are available and for which we can match CRSP data with the common stock holdings data
from Thomson Financial (described in the next subsection). Since we do not observe ?ows directly, we infer
?ows from fund returns and TNA as reported by CRSP. Let TNA
i
t
be the total net asset of a fund i and let R
i
t
be its return between quarter tÀ1 and quarter t. Following the standard practice in the literature (e.g., Zheng,
1999; Sapp and Tiwari, 2004), we compute ?ows for fund i in quarter t, F
i
t
, as the dollar value of net new issues
and redemptions using
F
i
t
¼ TNA
i
t
À ð1 þ R
i
t
ÞTNA
i
tÀ1
À MGN
i
t
, (1)
where MGN is the increase in total net assets due to mergers during quarter t. Note that (1) assumes that
in?ows and out?ows occur at the end of the quarter, and that existing investors reinvest dividends and other
distributions in the fund.
1
We assume that investors in the merged funds place their money in the surviving
fund. Funds that are born have in?ows equal to their initial TNA, while funds that die have out?ows equal to
their terminal TNA.
Counterfactual ?ows are computed under the assumption that each fund receives a pro rata share of the
total dollar ?ows to the mutual fund sector between date tÀk and date t, with the proportion depending on
TNA as of quarter tÀk. In order to compute the FLOW at date t, we start by looking at the total net asset
ARTICLE IN PRESS
1
We computed our measures under the alternative assumption of middle-of-period ?ows and found no effect on the main results.
A. Frazzini, O.A. Lamont / Journal of Financial Economics 88 (2008) 299–322 302
value of the fund at date tÀk. Then, for every date s we track the evolution of the fund’s counterfactual TNA
using:
^
F
i
s
¼
TNA
i
tÀk
TNA
Agg
tÀk
F
Agg
s
, (2)
d
TNA
i
s
¼ ð1 þ R
i
t
Þ
d
TNA
i
sÀ1
þ
^
F
i
s
, (3)
t À kpspt,
where
^
F
i
and
d
TNA
i
are counterfactual ?ows and TNA. F
Agg
is the actual aggregate ?ows for the entire mutual
fund sector, whileTNA
Agg
tÀk
is the actual aggregate TNA at date t-k. Eqs. (2) and (3) describe the dynamics of
funds that exist both in quarter tÀk and in quarter t. For funds that were newly created in the past k quarters,
d
TNA
i
is automatically zero—all new funds by de?nition represent new ?ows. The resulting counterfactual
total net asset value
d
TNA
i
t
at date t represents the fund size in a world with proportional ?ows in the last k
quarters.
For a detailed numerical example of our counterfactual calculations, see the Appendix, which also discusses
other details on Eqs. (2) and (3). We obtain a quarterly time series of counterfactual total net asset values for
every fund by repeating the counterfactual exercise every quarter t, and storing the resulting
d
TNA
i
t
at the end
of each rolling window.
Consider a representative investor who represents a tiny fraction, call it q, of the mutual fund sector.
Suppose that this investor behaves exactly like the aggregate of mutual fund investors, sending ?ows in and
out of different funds at different times. The counterfactual strategy described above is an alternative strategy
for this investor, and is implementable using the same information and approximately the same amount of
trading by the investor. To implement this strategy, this investor only needs to know lagged fund TNA’s and
aggregate ?ows. For this investor, q
d
TNA
i
t
is his dollar holding in any particular fund.
In designing this strategy, our aim is to create a neutral alternative to active reallocation, which matches the
total ?ows to the mutual fund sector. One could describe this strategy as a more passive, lower turnover,
value-weighting alternative to the active reallocation strategy pursued by the aggregate investor. It is similar in
spirit to the techniques of Daniel, Grinblatt, Titman, and Wermers (1997) and Odean (1999) in that it
compares the alternative of active trading to a more passive strategy based on lagged asset holdings. A feature
of our counterfactual calculations is that they do not mechanically depend on the actual performance of the
funds. A simpler strategy would have been to simply hold funds in proportion to their lagged TNA. The
problem with this strategy is that it tends to sell funds with high returns and buy funds with low returns. Since
we wanted to devise a strategy that re?ected only ?ow decisions by investors (not return patterns in stocks), we
did not use this simpler strategy.
Let x
it
be the total net assets of fund i in month t as a percentage of total assets of the mutual fund sector:
x
it
¼
TNA
i
t
TNA
Agg
t
. (4)
The counterfactual under proportional ?ows is
^ x
it
¼
d
TNA
i
t
d
TNA
Agg
t
. (5)
The difference between x
it
and ^ x
it
re?ects the active decisions of investors to reallocate money from one
manager to another over the past k quarters in a way that is not proportional to the TNA of the funds. This
difference re?ects any deviation from value weighting by the TNA of the fund in making new contributions.
ARTICLE IN PRESS
A. Frazzini, O.A. Lamont / Journal of Financial Economics 88 (2008) 299–322 303
2.2. Holdings
Thomson Financial provides the CDA/Spectrum mutual funds database, which includes all registered
domestic mutual funds ?ling with the SEC. The holdings constitute almost all the equity holdings of the fund
(see the Appendix for a few small exceptions). The holdings data in this study run from January 1980 to
December 2003.
While the SEC requires mutual funds to disclose their holdings on a semi-annual basis, approximately 60%
of funds additionally report quarterly holdings. The last day of the quarter is most commonly the report date.
A typical fund-quarter-stock observation would be as follows: as of March 30, 1998, Fidelity Magellan owned
20,000 shares of IBM. For each fund and each quarter, we calculate w
ij
as the portfolio weight of fund i in
stock j based on the latest available holdings data. Hence the portfolios’ weights w
ij
re?ect ?uctuations of the
market price of the security held.
A particular data challenge is matching the holdings data to the CRSP mutual fund database. This
matching is more dif?cult in the earlier part of the sample period. Further, the holdings data are notably error-
ridden, with obvious typographical errors.
2
Let z be the actual percent of the shares outstanding held by the mutual fund sector,
z
jt
¼
X
i
x
it
w
ijt
TNA
Agg
t
!
=MKTCAP
jt
, (6)
where MKTCAP
jt
is the market capitalization of ?rm j at date t. The ownership that would have occurred with
proportional ?ows into all funds and unchanged fund stock allocation and stock prices would be
^ z
jt
¼
X
i
^ x
it
w
ijt
TNA
Agg
t
!
=MKTCAP
jt
. (7)
For each stock, we calculate our central variable, FLOW, as the percent of the shares outstanding with
mutual fund ownership attributable to ?ows. The ?ow of security j is given by
FLOW
jt
¼ z
jt
À ^ z
jt
¼
X
i
½x
it
À ^ x
it
?w
ijt
TNA
Agg
t
( )
=MKTCAP
jt
. (8)
This ?ow has the following interpretation. If each portfolio manager had made exactly the same decisions in
terms of percent allocation of his total assets to different stocks, and if stock prices were unchanged, but the
dollars had ?owed to each portfolio manager in proportion to their TNA for the last k periods, then mutual
fund ownership in stock j would be lower by FLOW percent. Stocks with high FLOW are stocks that are
owned by mutual funds that have experienced high in?ows.
2.3. Describing the data
Table 1 shows summary statistics for the different types of data in our sample. Our sample starts in 1980. In
Table 1 we describe statistics for FLOW resulting from fund ?ows over the past three years, thus the table
describes data for ?ows starting in 1983.
Panel A shows the coverage of our sample as a fraction of the universe of CRSP equity funds and the
universe of CRSP common stocks. At the start of the sample, in 1983, we cover less than half of all stocks but
93% of the dollar value of the market (re?ecting the fact that mutual funds avoid smaller securities). Our
coverage rises over time as the relative size of the mutual fund sector grows substantially during the period. On
average, over the entire period our sample contains 97% of the total market capitalization and 69% of the
total number of common stocks in CRSP. Our sample of funds includes on average 99% of the total net asset
of US equity funds and 92% of the total number of funds.
ARTICLE IN PRESS
2
The Appendix of the NBER version of this paper, Frazzini and Lamont (2005), describes the matching process, issues of data errors,
and missing reports.
A. Frazzini, O.A. Lamont / Journal of Financial Economics 88 (2008) 299–322 304
Panel C shows summary statistics for three-year FLOW. FLOW is the actual percent ownership by the
mutual fund sector, minus the counterfactual percent ownership. Since the actual percent ownership is
bounded above by 100%, FLOW is bounded above by 100%. In the counterfactual case, there is no
accounting identity enforcing that the dollar value of fund holdings is less than the market capitalization of the
stock. Thus FLOW is unbounded below. Values of FLOW less than À100% are very rare, occurring less than
0.01% of the time for three-year ?ows.
In interpreting FLOW, it is important to remember that FLOW is a relative concept driven only by
differences in ?ows and holdings across different funds holding different stocks. FLOW is not intended to
capture any notion of the absolute popularity of stock. For example, FLOW for Alcoa in December 1999 was
À4.8%. The negative FLOW does not imply that Alcoa was unpopular with mutual funds, nor does it imply
that mutual funds were selling Alcoa. It could be that every mutual fund loved Alcoa, held a lot of it, and
bought more of it in 1999. What the negative ?ow means is that the funds which overweighted Alcoa in 1999
received lower-than-average in?ows (or perhaps out?ows) in 1999.
2.4. Appropriate horizons
Table 1 shows the properties of three-year ?ows. Throughout the paper, we use this three-year horizon as
our baseline speci?cation, because we are interested in understanding the long-term effects of trading on
individual investor wealth. Since we want to understand the net effect of trading, the relevant horizon should
depend on the actual time series behavior of fund ?ows.
Fig. 1 shows evidence on the appropriate chronological unit for fund ?ows. Every quarter, we sort mutual
funds based on ?ows, de?ned as net dollar in?ows divided by TNA at the end of the previous quarter. We
ARTICLE IN PRESS
Table 1
Summary statistics
This table shows summary statistics as of December of each year. Percent coverage of stock universe (EW) is the number of stocks with
a valid three-year FLOW, divided by total number of CRSP stocks. Percent coverage of stock universe (VW) is the total market
capitalization of stocks with a valid three-year FLOW, divided by the total market value of the CRSP stock universe. Percent coverage of
fund universe (EW) is the total number of funds in the sample divided by the total number of equity funds in the CRSP mutual fund
universe. Percent coverage of fund universe (VW) is the total net asset value of funds in the sample divided by the total net asset value of
equity funds in the CRSP mutual fund universe. TNA is the total net asset value of a fund, in millions. x is the fund’s actual percent of
dollar value of the total mutual fund universe in the sample. ^ x is counterfactual percent, using a horizon of three years. z is the percent of
the stock held by mutual funds (the stock’s actual total dollar value of mutual fund holdings divided by the stock’s market capitalization).
^ z is counterfactual z using a three-year horizon, as de?ned in the text.
Min Max Mean Std Level
Full sample, 1983–2003 1983 2003
Panel A: time-series (annual observations, 1983– 2003)
Number of funds in the sample per year 285 9,087 2,159 2,370 285 9,087
Number of stocks in the sample per year 2,710 6,803 4,690 1,516 2,710 4,974
Percent coverage of stock universe (EW) 48.5 92.2 68.7 18.3 48.5 92.2
Percent coverage of stock universe (VW) 92.8 99.4 97.4 2.3 92.8 99.4
Percent coverage of fund universe (EW) 88.0 99.0 92.2 3.0 88.0 99.0
Percent coverage of fund universe (VW) 94.0 99.9 98.9 1.3 94.0 95.0
Panel B: funds (Pooled year-fund observations, 1983– 2003)
TNA, millions of dollars 0.04 109,073 820 3331 245 746
Number of holdings per fund 1 4162 153 257 71 186
x (Percent of fund universe, actual) 0.00 7.86 0.13 0.41 0.49 0.05
^ x (Percent of fund universe, counterfactual) 0.00 11.4 0.17 0.52 0.66 0.06
Panel C: stocks (Pooled stock-fund observations, 1983– 2003)
Number of funds per stock 1 1,202 30 65 5 60
z (Percent owned by funds, actual) 0.00 99.35 9.10 10.13 6.09 10.56
^ z (Percent owned by funds, counterfactual) 0.00 234.32 9.21 4.56 5.02 8.23
FLOW ¼ z À ^ z À188 86.98 0.54 5.61 1.40 1.45
A. Frazzini, O.A. Lamont / Journal of Financial Economics 88 (2008) 299–322 305
assign funds to ?ve quintile portfolios and track the subsequent average ?ows. We plot the subsequent
cumulative difference in ?ows between high ?ow funds and low ?ow funds.
3
Fig. 1 shows that mutual fund
?ows are persistent: funds experiencing high in?ows this quarter tend to experience signi?cant higher ?ows
over the subsequent quarters. The total effect is complete approximately two to three years from portfolio
formation. Thus, fund ?ows tend to cumulate over long horizons. Fig. 1 shows similar results for sorting
stocks based on one quarter FLOW and tracking the subsequent cumulative difference in FLOW between
high ?ow stocks and low ?ow stocks.
Thus, to understand the net effect of fund ?ows on investor wealth, it is not enough to relate short term
?ows to short term performance; one must also take into account how the effects of trading cumulate over
time. If retail investors as a group were purchasing mutual funds in quarter t and redeem their shares in
quarter t+1, then the appropriate measure would be one quarter FLOW. Since Fig. 1 shows that retail
investors as a group are not doing this, longer horizon FLOW is appropriate to study.
3. Flows and stock returns
To test for return predictability, we examine monthly returns in excess of Treasury bills on calendar time
portfolios formed by sorting stocks on FLOW. At the beginning of every calendar month, we rank stocks in
ascending order based on the latest available FLOW and assign them to one of ?ve quintile portfolios. We
compute FLOW over horizons stretching from three months (one quarter, the shortest interval we have for
calculating ?ows) to ?ve years. We rebalance the portfolios every calendar month using value weights.
In Panel A of Table 2, we report time series averages of the sorting variable for each portfolio. The
rightmost column shows the difference between the high ?ow stocks and the low ?ow stocks. The effect of
?ows on mutual fund ownership is fairly sizable. For the top quintile of three-year ?ows, non-proportional
?ows raise the aggregate mutual fund ownership by more than 6% of the stock’s total market capitalization.
For the bottom quintile, ?ows lower ownership by 4% (although one cannot tell this from the table, the
bottom quintile re?ects stocks that are not just experiencing lower-than-average in?ows, they are experiencing
ARTICLE IN PRESS
0
1
2
3
4
5
6
0
10
20
30
40
50
60
70
80
90
100
-1
S
t
o
c
k

f
l
o
w

o
w
n
e
r
s
h
i
p

(
%
)
F
u
n
d

f
l
o
w
s

(
%
)
Quarter t+k
Fund flows Stock flow ownership
1 0 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
Fig. 1. Cumulative ?ows for quarter t+k sorted on ?ows in quarter t. This ?gure shows the average cumulative ?ows in quarter t+k for
mutual funds (stocks) sorted on quarterly ?ows in quarter t. At the beginning of every quarter mutual funds (stocks) are ranked in
ascending order based on their quarterly ?ows. Funds (stocks) are assigned to one of ?ve quintile portfolios. We report the cumulative
average difference in ?ows between the top 20% high ?ow funds (stocks) and the bottom 20% low ?ow funds (stocks). Fund ?ows are
de?ned as dollar in?ows/out?ows divided by the total net assets of the fund at the end of the previous quarter. Stock ?ows are de?ned as
the actual percent of the stock owned by mutual funds minus the counterfactual percent.
3
We compute averages in the spirit of Fama and MacBeth (1973): we calculate averages for each month and report time series means.
This procedure gives equal weight to each monthly observation.
A. Frazzini, O.A. Lamont / Journal of Financial Economics 88 (2008) 299–322 306
out?ows). The difference between the top and bottom quintiles increases with the time horizon, indicating
(consistent with Fig. 1) that ?ows into individual stocks tend to cumulate over time.
Panel B of Table 2 shows the basic results of this paper. We report returns in month t of portfolios formed
by sorting on the last available FLOW as of month tÀ1. The rightmost column shows the returns of a zero
cost portfolio that holds the top 20% high ?ow stocks and sells short the bottom 20% low ?ow stocks. For
every horizon but three months, high ?ow today predicts low subsequent stock returns. The relation is
statistically signi?cant for ?ow computed over horizons stretching from six months to three years. This dumb
money effect is sizable: stocks with high FLOW as a result of the active reallocation across funds over the past
six months to ?ve years underperform low FLOW stocks by between 36 and 85 basis points per month or
approximately between 4% and 10% per year, depending upon the horizon of the past ?ow.
Perhaps surprisingly, Table 2 shows no solid evidence for the smart money effect in stock returns, even at
the shorter horizons where one might expect price momentum to dominate. Gruber (1996) and Zheng (1999)
look at quarterly ?ows and ?nd that high ?ows predict high mutual fund returns: one can see a hint of this in
the three-month ?ow results, although one cannot reject the null hypothesis. We return to this issue in Section
4.5.
Fig. 2 gives an overview of how ?ow predicts returns at various horizons. We show the cumulative average
returns in month t+k on long/short portfolios formed on three-month ?ow in month t. For ko0, the ?gure
shows how lagged returns predict today’s ?ows. The ?gure shows that ?ows into an individual stock are
strongly in?uenced by past returns on that stock. This result is expected given the previous literature
documenting high in?ows to high performing funds. Flows tend to go to funds that have high past returns,
and since funds’ returns are driven by the stocks that they own, ?ows tend to go to stocks that have high past
ARTICLE IN PRESS
Table 2
Calendar time portfolio excess returns and ?ow, 1980–2003
This table shows the average ?ow and excess returns for calendar time portfolios sorted on past ?ow, de?ned as the stock’s actual
percent of the total dollar value of mutual fund holdings divided by the stock’s market capitalization minus the counterfactual percent. At
the beginning of every calendar month stocks are ranked in ascending order based on the last available ?ow. Stocks are assigned to one of
?ve quintile portfolios. L/S is a zero cost portfolio that holds the top 20% stocks and sells short the bottom 20% stocks. Portfolios are
rebalanced monthly to maintain value weights. In Panel A we report averages of the sorting variable for each cell. Flow is in percent. In
Panel B we report average portfolio returns minus Treasury bill returns. Returns are in monthly percent, t-statistics are shown below the
coef?cient estimates.
Panel A: ?ow Q1 (low) Q2 Q3 Q4 Q5 (high) Q5-Q1
3-Month ?ow À0.551 À0.156 À0.025 0.121 0.908 1.459
6-Month ?ow À0.993 À0.266 À0.025 0.248 1.653 2.646
1-Year ?ow À1.768 À0.437 À0.002 0.520 2.856 4.624
3-Year ?ow À4.088 À0.788 0.251 1.652 6.047 10.135
5-Year ?ow À6.319 À1.223 0.438 2.362 8.014 14.333
Panel B: portfolio returns Q1 (low) Q2 Q3 Q4 Q5 (high) L/S
3-Month ?ow 0.628 0.648 0.503 0.546 0.661 0.033
(1.99) (2.28) (1.77) (1.86) (1.82) (0.13)
6-Month ?ow 0.753 0.684 0.689 0.544 0.390 À0.363
(2.52) (2.43) (2.52) (1.87) (1.18) (À2.08)
1-Year ?ow 0.909 0.848 0.760 0.590 0.408 À0.501
(3.02) (3.03) (2.79) (1.97) (1.18) (À2.61)
3-Year ?ow 1.026 0.884 0.695 0.450 0.180 À0.846
(3.19) (3.00) (2.37) (1.34) (0.44) (À3.30)
5-Year ?ow 0.880 0.748 0.671 0.501 0.486 À0.394
(2.67) (2.38) (1.85) (1.36) (1.11) (À1.35)
A. Frazzini, O.A. Lamont / Journal of Financial Economics 88 (2008) 299–322 307
returns. For k40, the ?gure shows the dumb money effect as the downward slope of cumulative returns
becomes pronounced after six or twelve months. High FLOW stocks severely underperform low FLOW
stocks over the course of about two years.
The results in Table 2 and Fig. 2 show that stocks that are overweighted by retail investors due to fund ?ows
tend to have lower subsequent returns. However, in terms of measuring the actual returns experienced by
mutual funds investors, this evidence does not conclusively prove that investors experience returns that are
lower due to their active reallocation, because this evidence does not correspond to the dollar holdings of any
class of investors. One needs to look at all trades and all dollar allocations to different securities over time. In
Section 5, we perform this exercise for the aggregate mutual fund investor, and show that trading does, in fact,
decrease both average returns and the return/risk ratio for an individual who is behaving like the aggregate
mutual fund investor. From this perspective, then, individual investors in aggregate are unambiguously dumb.
4. Robustness Tests
4.1. Controlling for size, momentum, and value
Table 3 shows results for returns controlling for size, value, and price momentum. These variables are
known to predict returns and likely to be correlated with ?ows. Sapp and Tiwari (2004), for example, argue
that the short-horizon smart money effect merely re?ects the price momentum effect of Jegadeesh and Titman
(1993). If an individual follows a strategy of sending money to funds with past high returns in the last year and
withdrawing money from funds with low returns, then he will end up with a portfolio that overweights high
momentum stocks. This strategy might be a smart strategy to follow, as long as he keeps rebalancing the
strategy. However, if the individual fails to rebalance promptly, eventually he will be holding a portfolio with a
strong growth tilt. Thus over long horizons, stocks with high in?ows are likely to be stocks with high past
returns and are therefore likely to be growth stocks. So it is useful to know whether ?ows have incremental
forecasting power for returns or just re?ect known patterns of short horizon momentum and long horizon
value/reversals in stock returns.
The left-hand side of Table 3 shows results where returns have been adjusted to control for value, size, and
momentum. Following Daniel, Grinblatt, Titman, and Wermers (1997) (DGTW), it subtracts from each stock
return the return on a portfolio of ?rms matched on market equity, market-book, and prior one-year return
ARTICLE IN PRESS
0.0%
5.0%
10.0%
15.0%
20.0%
25.0%
-24
R
e
t
u
r
n
Month t+k
-20 -16 -12 -8 -4 4 12 8 0 16 20 24 28 32 36 40
Fig. 2. Average cumulative return in month t+k on a long/short portfolio formed on three-month ?ow in month t. At the beginning of
every calendar month stocks are ranked in ascending order based on the last available ?ow. Stocks are assigned to one of ?ve quintile
portfolios. Portfolios are rebalanced monthly to maintain value weights. The ?gure shows average cumulative returns in event time of a
zero cost portfolio that holds the top 20% stocks and sells short the bottom 20% stocks. The long/short portfolio used here, based on raw
returns, corresponds to ‘‘3-month ?ow, L/S’’ in Table 2.
A. Frazzini, O.A. Lamont / Journal of Financial Economics 88 (2008) 299–322 308
quintiles (a total of 125 matching portfolios).
4
Using DGTW returns, the dumb money effect is substantially
reduced, with the coef?cient falling from À0.85% to À0.42% per month for three-year ?ows, still signi?cant
but approximately half as large. The right-hand side of Table 3 shows alphas and the corresponding factor
loadings from a Fama and French (1993) three factor regression. Here the reduction of the three-year dumb
money effect is not as substantial, as the three-year differential return remains sizeable at À0.74% per month.
The high and negative coef?cient on the HML, the Fama-French value factor, shows that high sentiment
stocks tend to be stocks with high market-book.
In Panel A of Table 4, we take a closer look at the relation between the dumb money effect and the value
effect by independently sorting all stocks into ?ve ?ow categories and ?ve market-book categories, with a
resulting 25 portfolios. We sort on three-year ?ows, and on market-book ratio following the de?nition of
Fama and French (1993). The right-most column shows whether there is a ?ow effect within market-to-book
quintiles. Thus, if the value effect subsumes the dumb money effect, this column should be all zeros. The
bottom row shows whether there is a value effect controlling for ?ows. If the dumb money effect subsumes the
value effect, this row should be all zeros. If the two effects are statistically indistinguishable, then both the row
and the column should be all zeros.
Panel A of Table 4 shows that, generally, neither effect dominates the other. As in Table 3, the dumb money
effect survives the correction for market-book. The dumb money effect is concentrated within growth stocks,
while the value effect is concentrated among high ?ow stocks. High sentiment growth stocks actually
underperform T-bills, while low sentiment growth stocks have very high returns.
Panel B shows double sort portfolios for three-year past stock returns instead of market-book, to explore
the reversal effect of De Bondt and Thaler (1985). In order to make the reversal effect as powerful as possible,
we sort on past returns lagged one year (in other words, we sort on stock returns from month tÀ48 to tÀ12).
The results are similar to Panel A: neither effect subsumes the other. However, the dumb money and value/
reversal effect are clearly quite related, and perhaps re?ect the same underlying phenomenon.
ARTICLE IN PRESS
Table 3
Controlling for value, size, and momentum
This table shows calendar time portfolio abnormal returns. At the beginning of every calendar month stocks are ranked in ascending
order based on the last available ?ow. Stocks are assigned to one of ?ve quintile portfolios. L/S is a zero cost portfolio that holds the top
20% stocks and sells short the bottom 20% stocks. Portfolios are rebalanced monthly to maintain value weights. We report DGTW
average characteristic adjusted returns and Fama and French (1993) alphas. DGTW characteristic adjusted returns are de?ned as raw
monthly returns minus the average return of all CRSP ?rms in the same size, market-book, and one ^ z year momentum quintile. The
quintiles are de?ned with respect to the entire universe in that month and DGTW portfolios are refreshed every calendar month. Fama
French alpha is de?ned as the intercept in a regression of the monthly excess returns on the three factors of Fama and French (1993).
Returns and alphas are in monthly percent, t-statistics are shown below the coef?cient estimates.
DGTW Fama French alpha Loadings on L/S
Q1 Q5 L/S Q1 Q5 L/S MKT SMB HML R2
3-Month ?ow À0.067 À0.016 0.051 À0.197 0.113 0.309 À0.111 0.390 À0.498 0.302
(À1.08) (À0.17) (0.43) (À1.55) (0.85) (1.37) (À1.97) (5.45) (À5.84)
6-Month ?ow À0.024 À0.193 À0.169 À0.030 À0.172 À0.143 À0.056 0.136 À0.426 0.291
(À0.43) (À2.75) (À1.99) (À0.30) (À1.88) (À0.92) (À1.47) (2.78) (À7.30)
1-Year ?ow 0.027 À0.238 À0.265 0.092 À0.238 À0.331 À0.021 0.139 À0.383 0.226
(0.42) (À3.14) (À2.68) (0.93) (À2.13) (À1.86) (À0.48) (2.49) (À5.74)
3-Year ?ow 0.093 À0.329 À0.422 0.260 À0.474 À0.735 0.074 0.151 À0.426 0.229
(1.10) (À3.33) (À2.96) (2.09) (À3.14) (À3.14) (1.27) (2.07) (À4.90)
5-Year ?ow 0.013 À0.168 À0.181 0.083 À0.162 À0.245 0.007 0.526 À0.525 0.541
(0.17) (À1.46) (À1.17) (0.75) (À1.15) (À1.19) (0.14) (8.59) (À6.99)
4
These 125 portfolios are reformed every month based on the market equity, M/B ratio, and prior year return from the previous month.
The portfolios are equally weighted and the quintiles are de?ned with respect to the entire universe in that month.
A. Frazzini, O.A. Lamont / Journal of Financial Economics 88 (2008) 299–322 309
To summarize, the dumb money effect is not completely explained by the value effect. Up to half of the
dumb money effect is explained by value and other characteristics, but a statistically signi?cant portion
remains. Neither the dumb money effect nor the value/reversal effect dominates the other. Thus, investors hurt
themselves by reallocating across mutual funds for two reasons. First, they hurt themselves by overweighting
growth stocks. Second, controlling for market-book, they hurt themselves by overweighting stocks that
underperform their category benchmarks, and in particular, they pick growth stocks that do especially poorly.
4.2. Buy-and-hold long-term returns
The calendar time portfolios reported so far are rebalanced every month. In Fig. 3, we show a slightly
different concept, buy-and-hold returns. For each stock, we calculate the k-month ahead total return
ARTICLE IN PRESS
Table 4
Flows vs. value and reversals
This table shows calendar time portfolio returns. At the beginning of every calendar month stocks are ranked in ascending order based
on the last available ?ow and market-book ratio (M/B). M/B is market-book ratio (market value of equity divided by Compustat book
value of equity). The timing of M/B follows Fama and French (1993) and is as of the previous December year-end. Stocks are assigned to
one of 25 portfolios. L/S is a zero cost portfolio that holds the top 20% stocks and sells short the bottom 20% stocks. Portfolios are
rebalanced monthly to maintain value weights. We report average excess returns. Returns are in monthly percent, t-statistics are shown
below the coef?cient estimates.
Low ?ow High ?ow High ?ow
minus low
?ow
Q1 Q2 Q3 Q4 Q5 L/S
Panel A: 3-year ?ow and value
Q1 (Value) 0.738 0.904 0.968 0.828 0.786 0.048
(2.10) (2.50) (2.66) (2.18) (2.15) (0.17)
Q2 0.812 0.961 0.703 0.704 0.500 À0.312
(2.57) (3.15) (2.13) (2.21) (1.52) (À1.28)
Q3 1.011 0.692 0.573 0.536 0.809 À0.202
(2.91) (2.28) (1.86) (1.63) (2.05) (À0.84)
Q4 0.893 0.670 0.517 0.697 0.472 À0.421
(2.46) (2.01) (1.18) (1.84) (1.07) (À2.51)
Q5 (Growth) 1.322 0.792 0.611 0.480 À0.179 À1.501
(3.23) (2.23) (1.49) (1.13) (À0.33) (À4.33)
Growth minus value 0.583 À0.112 À0.358 À0.347 À0.966
(1.75) (À0.34) (À0.85) (À1.13) (À2.34)
Panel B: 3-year ?ow and reversals
Q1 (Loser) 1.117 1.408 1.171 1.163 1.059 À0.059
(2.25) (2.39) (1.90) (2.13) (1.90) (À0.15)
Q2 1.415 1.044 1.158 0.613 0.712 À0.704
(3.66) (2.60) (2.94) (1.52) (1.61) (À2.76)
Q3 1.162 1.179 0.601 0.712 0.591 À0.570
(3.57) (3.62) (1.84) (2.28) (1.56) (À2.57)
Q4 0.770 0.853 1.094 0.680 0.511 À0.259
(2.47) (2.96) (3.60) (2.41) (1.53) (À1.27)
Q5 (Winner) 0.945 0.839 0.644 0.471 0.109 À0.836
(2.67) (2.43) (1.93) (1.25) (0.25) (À2.98)
Winner minus loser À0.172 À0.568 À0.527 À0.692 À0.950
(À0.45) (À1.11) (À0.99) (À1.80) (À2.39)
A. Frazzini, O.A. Lamont / Journal of Financial Economics 88 (2008) 299–322 310
(for example, the return over the next 36 months).
5
At the beginning of each quarter, we sort stocks based on
past three-month ?ow and calculate the value weighted average of this long-term return for both the high
FLOW and low FLOW stocks. We then take the difference between these two returns, and report the time-
series average of this difference over the entire sample period.
In addition to reporting buy-and-hold (instead of calendar time portfolio) returns, we also use a slightly
different risk adjustment procedure to address concerns that the raw portfolios and the matching portfolios
are constructed using different information.
6
As with the ?ow-based portfolios, we calculate the buy-and-hold
returns for the matching DGTW returns. We construct both the ?ow-based portfolios and the matching
DGTW portfolios at the same frequency. Fund ?ows are available quarterly in our data, so we refresh both
the dumb money portfolio and the matching DGTW portfolios quarterly. This procedure puts the dumb
money and characteristic-matched portfolios on an equal footing.
7
Given the fact that long-run abnormal returns can be very sensitive to the benchmarking technique used, we
also report results for an alternative risk adjustment. Following Barber and Lyon (1997) we measure abnormal
return comparing the return of a stock to the return of a single control stock. Every quarter, we ?rst identify
all ?rms with a market value of equity between 70% and 130% of the market value of equity of the sample
?rm; from this set of ?rms, we then rank potential matches according to book to market and return in the
previous twelve months. We sum ranks across the different characteristics, and select the lowest rank as the
ARTICLE IN PRESS
-10.0%
-8.0%
-6.0%
-4.0%
-2.0%
0.0%
2.0%
1
R
e
t
u
r
n
CAR DTGW Single matched stock
Month t+k from portoflio formation
6 11 16 21 26 31 36
Fig. 3. Buy and hold return in month t+k on a long/short portfolio formed on three month ?ow in month t. This ?gure shows the event-
time average buy and hold return in the 36 months subsequent to the formation date of a long/short portfolio formed on three-month ?ow.
For each horizon k, we calculate for every stock the k-month ahead total return, DGTW-adjusted return, and single matched stock-
adjusted return. DGTW-adjusted return is de?ned as total return minus the total return on an equally weighted portfolio of all CRSP ?rms
in the same size, market-book, and one year momentum quintile. Single matched stock-adjusted return is de?ned as total return minus the
total return of a stock to the return of a single control stock. We rank stocks with a market value of equity between 70% and 130% of the
market value of equity of the sample stock according to book to market and one year momentum. We sum ranks and select the lowest
rank as the matching stock. We maintain the match until the next portfolio rebalancing or the delisting date. If a match is delisted it is
replaced by the second lowest rank stock. At the beginning of every calendar quarter, stocks are ranked in ascending order based on the
last available three-month ?ow. We assign stocks to one of ?ve quintile portfolios and calculate the value-weighted average of long-term
return for both the high ?ow (top 20%) and low ?ow stocks (bottom 20%). This ?gure reports the time-series average of this difference
over the entire sample period. The ?ow portfolio, the DGTW portfolio, and the matched stocks are refreshed quarterly and when
calculating long-term returns; if a ?rm exits the database, we reinvest its weight into the remaining stocks in the portfolio.
5
Both here and everywhere else, we include delisting returns when available in CRSP. If a ?rm is delisted but the delisting return is
missing, we investigate the reason for disappearance. If the delisting is performance-related, we follow Shumway (1997) and assume
aÀ30% delisting return. This assumption does not substantially affect any of the results. When calculating long-term returns, when a ?rm
exits the database, we reinvest its weight into the remaining stocks in the portfolio.
6
As noted in Footnote 4, in our previous results the DGTW portfolios are formed monthly.
7
We also re-ran Table 3 where all portfolios including DGTW matching portfolios were refreshed quarterly. The results were virtually
the same (for example, the three-year differential returns went from À0.422% per month in Table 3 to À0.414% per month).
A. Frazzini, O.A. Lamont / Journal of Financial Economics 88 (2008) 299–322 311
matching stock. We maintain the match until the next portfolio rebalancing or the delisting date. If a match
becomes unavailable at a given point because of delisting then from that point forward it is replaced by the
second lowest rank stock. This procedure ensures that there is no look ahead bias.
Fig. 3 reports this difference in buy-and-hold returns in event-time in the 36 months subsequent to the
formation date, using both raw, DGTW-adjusted and single match-adjusted long-term returns. Looking at
raw returns, the results are similar to Fig. 2. Stocks with high in?ows this quarter underperform stocks with
low in?ows this quarter by about 9% over the subsequent three years.
Looking at DGTW-adjusted buy-and-hold returns, the results are similar to Table 3. Table 3 showed that
DGTW adjustment reduces the total effect by about half. Fig. 3 shows DGTW adjustment reducing the effect
by somewhat more than half, with high ?ow stocks underperforming by an adjusted 3% over the next three
years. As in Table 3, this result re?ects the fact that in?ows tend to go to growth stocks, which have low
average returns. Using a single control stock leads to very similar results.
4.3. Further robustness tests
Table 5 shows the results for different samples of stocks and different methods of calculating returns. First,
it shows results for the sample of stocks which have market cap above and below the CRSP median. The dumb
money effect tends to be larger for large cap securities, and larger for value weighted portfolios than for
equally weighted portfolios. These results may re?ect the fact that we use mutual fund holdings to construct
the FLOW measure. FLOW is probably a better measure of individual sentiment for stocks held mostly by
mutual funds, whose holdings tend to be skewed towards large cap securities.
One concern is that the return predictability in Table 2 may be driven by initial public offerings. To address
this, in Table 5 we de?ne new issues as stocks with less than 24 months of return data on the CRSP tape
at the time of portfolio formation. We split the sample by separating out new issues and computing calendar
time portfolio as before within the two sub-samples. Table 5 shows that excluding new issues only slightly
lowers the dumb money effect. Looking at return predictability within new issues, we ?nd that there is a very
large and signi?cant dumb money effect. Thus, the dumb money effect is much stronger among new issues,
perhaps indicating the sentiment is particularly relevant for this class of stocks. We further consider issuance
in Section 6.
One might ask whether the dumb money effect is an implementable strategy for outside investors using
information available in real time. In constructing calendar time portfolios we use the end of quarter ?le date
ARTICLE IN PRESS
Table 5
Robustness tests
This table shows calendar time returns of a zero cost portfolio that holds the top 20% high ?ow stocks and sells short the bottom 20%
low ?ow stocks. Larger cap stocks are all stocks with market capitalization above the median of the CRSP universe that month, smaller
cap are below median. New issues are de?ned as stocks with less than 24 months of return data on the CRSP tape at the time of portfolio
formation. Returns are in monthly percent, t-statistics are shown below the coef?cient estimates.
Smaller cap Larger cap Equal weight Exclude new
issues
Only new issues Flow lagged 12
months
3-Month ?ow À0.011 0.062 0.071 0.075 0.265 À0.594
(À0.06) (0.21) (0.37) (0.32) (0.64) (À2.70)
6-Month ?ow À0.048 À0.394 À0.204 À0.333 À0.344 À0.678
(À0.34) (À2.02) (À1.95) (À2.04) (À1.24) (À2.99)
1-Year ?ow À0.174 À0.505 À0.304 À0.457 À0.626 À0.674
(À1.10) (À2.44) (À2.21) (À2.49) (À2.06) (À3.00)
3-Year ?ow À0.421 À0.824 À0.502 À0.755 À1.413 À0.023
(À2.09) (À3.18) (À3.26) (À3.11) (À3.99) (À0.12)
5-Year ?ow À0.507 À0.475 À0.173 À0.317 À1.185 À0.031
(À2.49) (À1.58) (À1.38) (À1.17) (À2.88) (À0.14)
A. Frazzini, O.A. Lamont / Journal of Financial Economics 88 (2008) 299–322 312
(FDATE) assigned by Thomson Financial. The mutual funds holdings data re?ect both a ‘‘vintage’’ ?le date
and a report date (RDATE). The report date is the calendar date when a snapshot of the portfolio is recorded.
These holdings eventually become public information and the statutory maximum delay in ?ling after the
report date is 60 days. Thomson Financial assigns ?le dates (FDATE) to the corresponding quarter ends of the
?lings and these dates do not correspond to the actual ?ling date with the SEC. As a result, if the lag between
the report date and Thomson’s ?le date is shorter than 60 days, these holdings are not public information on
Thomson’s ?le date.
8
In our sample, only in 53.15% of fund-quarter observations is the Thomson’s ?le date more than 60 days
beyond the report date. Thus, although our methodology does involve some built-in staleness of ?ows, not all
the variables in Table 2 are in the information set of any investor who has access to all the regulatory ?lings
and reports from mutual funds, as they are ?led with the SEC.
To address this issue, Table 5 shows results with the ?ow variables lagged an additional 12 months. As one
might expect, given Fig. 2, this lagging does not destroy the ability to construct a pro?table trading strategy.
Thus, the dumb money effect is not primarily about short-term information contained in ?ows, it is about
long-term mispricing.
In unreported results, we have also examined the dumb money effect in different categories of funds. First,
we looked at the effect in load funds and no load funds. Second, we looked at the effect across different fund
objective categories (aggressive growth, growth, growth & income, and balanced). In all cases the dumb
money effect was present and about the same size.
In further unreported results, we also examined the extent to which the dumb money effect is an intra-
industry vs. an inter-industry phenomenon. We found that about half of the three-year dumb money effect is
explained by industry performance, with the other half re?ecting industry-adjusted performance. Thus,
investors tend to indirectly select stocks that underperform their industry benchmark, and they also tend to
overweight industries with lower subsequent returns.
4.4. Subsample stability
Table 6 examines the performance of the strategy over time. Since we only have 23 years of returns for
three-year ?ows, inference will naturally be tenuous as we look at subsamples. For each time period, the ?rst
row shows the baseline three-year ?ow results, while the other rows show different versions of the dumb
money effect. First, we split the sample into recessions (as de?ned by the NBER) and non-recessions. While
the dumb money effect appears somewhat higher in recessions, with only 42 recession months, it is dif?cult to
make any strong inference. One clear result is that the dumb money effect is certainly present in non-recession
periods.
The next pair of columns splits the sample in half, pre-1994 and post-1994. Looking at the baseline result,
the dumb money effect is signi?cantly negative in both halves of the sample, although it is much higher in the
second half of the sample. It is not clear how to interpret this difference. Although the dumb money effect is
more than three times as big in the second half of the sample, the difference between the two mean returns is
not signi?cant at conventional levels (we fail to reject the null hypothesis of equality of the two means with a t-
statistic of 1.7) and as discussed previously, in the earlier part of the sample both our coverage of stocks and
the relative size of the mutual fund industry are lower. Thus one might expect weaker results in the early years
of the sample.
The last pair of columns splits the sample pre- and post-1998. The dumb money effect is particularly large in
the 1999–2003 period (although it is statistically signi?cant excluding this period as well). One interpretation
of the time pattern in Table 6 is that the period around 2000 was a time of particularly high irrationality, when
irrational traders earned particularly low returns. Many anomalies grew larger in this period (see Ofek and
Richardson, 2003). Indeed, one might propose that if a return pattern does not grow stronger in this period,
then it is probably not attributable to irrational behavior.
ARTICLE IN PRESS
8
Furthermore, currently, holdings data appear on the SEC Edgar System on the business day following a ?ling, but information lags
were probably longer at the beginning of the sample period.
A. Frazzini, O.A. Lamont / Journal of Financial Economics 88 (2008) 299–322 313
Looking at results for the various speci?cations gives similar results. Every number is negative in every
subsample, although not always signi?cantly different from zero. Controlling for value (in the DGTW and
Fama French rows), the effect is particularly weak in the earlier part of the sample. The effect within new
issues is very large in all subperiods.
To summarize, the dumb money effect is reasonably robust across time periods, although point estimates
are much higher in the second half of the sample. We further examine subsample stability in Section 5, using a
portfolio weighting scheme that is arguably less arbitrary and more economically relevant. There, the results
for stock returns are much more constant across different time periods.
4.5. Comparison to prior results
The prior literature has focused largely on how ?ows predict short-horizon returns. Warther (1995), for
example, looks at aggregate ?ows and the aggregate stock market and found some evidence that high ?ows
today predict high returns over the next four weeks. Similarly, Zheng (1999) and Gruber (1996) largely focus
on how ?ows predict returns over the next few months. Looking at Fig. 3 and the ?rst row of Tables 2 and 3,
one can see a bit of evidence for this smart money effect at the three-month horizon, especially when adjusting
for the value effect.
A previous version of this paper, Frazzini and Lamont (2005), examined mutual fund returns to show how
our results relate to the previous work of Zheng (1999) and Gruber (1996).
9
Using mutual fund returns instead
of stock returns, we found the dumb money effect is still strongly statistically signi?cant at the three-year
horizon. However, in contrast to the results using stock returns, the smart money effect comes in more
strongly at the three-month horizon, and in some speci?cations it is statistically signi?cant.
How should one reconcile these different results at different horizons? Whether behavior is ‘‘smart’’ or
‘‘dumb’’ depends on how it affects ultimate wealth. Despite the fact that individuals may earn positive returns
in the ?rst three months after reallocation, we argue this out-performance is wasted because the individuals as
ARTICLE IN PRESS
Table 6
Subsample stability
This table shows calendar time returns of a zero cost portfolio that holds the top 20% high ?ow stocks and sells short the bottom 20%
low ?ow stocks. DGTW characteristic adjusted returns are de?ned as raw monthly returns minus the average return of all CRSP ?rms in
the same size, market-book, and one year momentum quintile. The quintiles are de?ned with respect to the entire universe in that month
and DGTW portfolios are refreshed every calendar month. Fama French alpha is de?ned as the intercept in a regression of the monthly
excess returns on the three factors of Fama and French (1993). New issues are de?ned as stocks with less than 24 months of return data on
the CRSP tape at the time of portfolio formation. Returns and alphas are in monthly percent, t-statistics are shown below the coef?cient
estimates.
Time period Exclude NBER
recessions
Only NBER
recessions
83–93 94–03 83–98 99–03
] Of months 210 42 132 120 192 60
Stock returns À0.818 À1.183 À0.397 À1.294 À0.501 À1.879
(À3.34) (À0.73) (À2.06) (À2.80) (À2.79) (À1.99)
DGTW À0.353 À0.871 À0.101 À0.731 À0.145 À0.796
(À2.52) (À1.14) (À0.62) (À3.37) (À1.08) (À2.00)
Fama French
alpha
À0.690 À1.074 À0.168 À1.420 À0.224 À1.609
(À2.99) (À1.06) (À0.79) (À3.81) (À1.29) (À2.39)
Exclude new
issues
À0.733 À1.017 À0.363 À1.146 À0.463 À1.628
(À3.16) (À0.65) (À1.98) (À2.63) (À2.64) (À1.63)
Only new issues À1.260 À3.297 À0.865 À1.961 À0.843 À3.124
(À3.56) (À1.85) (À2.41) (À3.23) (À3.00) (À3.12)
9
Although Frazzini and Lamont (2005) had a less complete database than this paper, the basic results using mutual fund returns were
not substantially different.
A. Frazzini, O.A. Lamont / Journal of Financial Economics 88 (2008) 299–322 314
a group are not following a dynamic strategy of buying the best-performing funds, holding them for a quarter,
and then selling them. As revealed in Fig. 1, they are instead in aggregate following a strategy of buying the
best-performing funds, and holding them for a long period of time. So the longer horizon return shows that
investors are not actually bene?ting from their trading. For a more economically relevant measure of how
these two effects balance out, in the next section we look at how the aggregate mutual fund investor is helped
or hurt by his trading.
5. Economic signi?cance to the aggregate investor
5.1. The magnitude of wealth destruction
So far, we have shown that stocks owned by funds with large in?ows have poor subsequent returns. In this
section, we measure the wealth consequences of active reallocation across funds, for the aggregate investor.
We assess the economic signi?cance by measuring the average return earned by a representative investor, and
comparing it to the return he could have earned by simply refraining from engaging in non-proportional ?ows.
We examine both returns on stocks and returns on mutual funds.
De?ne R
ACTUAL
as the return earned by a representative mutual fund investor who owns a tiny fraction of
each existing mutual fund. The returns would re?ect a portfolio of stocks where the portfolio weights re?ect
the portfolio weights of the aggregate mutual fund sector:
R
ACTUAL
t
¼
X
i
x
i;t
X
j
w
ij;t
R
j
t
" #
; (9)
where R
j
is the return on stock j. The return from a strategy of refraining from non-proportional ?ows,
R
NOFLOW
, is
R
NOFLOW
t
¼
X
i
^ x
i;t
X
j
w
ij;t
R
j
t
" #
. (10)
We use three-year ?ows in these calculations. Table 7 shows excess returns on these two portfolios and for
comparison shows the value-weighted market return as well. Since the two mutual fund portfolios use weights
based on dollar holdings, they are, of course, quite similar to each other and to the market portfolio.
ARTICLE IN PRESS
Table 7
Economic signi?cance for the aggregate mutual fund investor
This table shows calendar time portfolio returns. It uses three-year ?ows. R
ACTUAL
is returns on a mimicking portfolio for the entire
mutual fund sector, with portfolio weights the same as the actual weights of the aggregate mutual fund sector. R
NOFLOW
is returns on a
mimicking portfolio for the counterfactual mutual fund sector, with portfolio weights the same as the counterfactual weights of the
aggregate mutual fund sector. R
M
is the CRSP value-weighted market return.
Mean t-Stat SR
Panel A: using stock returns
Actual excess return on mutual fund holdings R
ACTUAL
–R
F
0.657 2.05 0.132
Counterfactual excess return on mutual fund holdings R
NOFLOW
–R
F
0.727 2.27 0.146
Market excess returns R
M
–R
F
0.651 2.26 0.143
Net bene?t of mutual funds R
ACTUAL
–R
M
0.018 0.43 0.028
Dumb money effect R
ACTUAL
–R
NOFLOW
À0.069 À4.10 À0.269
Stock picking R
NOFLOW
–R
M
0.087 1.90 0.123
Panel B: using mutual fund returns
Actual excess return on mutual funds R
ACTUAL
–R
F
0.502 1.75 0.113
Counterfactual excess returns on mutual funds R
NOFLOW
–R
F
0.587 2.08 0.133
Net bene?t of mutual funds R
ACTUAL
–R
M
À0.117 À3.28 À0.210
Dumb money effect R
ACTUAL
–R
NOFLOW
À0.085 À4.09 À0.262
Stock picking R
NOFLOW
–R
M
À0.032 À0.92 À0.059
A. Frazzini, O.A. Lamont / Journal of Financial Economics 88 (2008) 299–322 315
Table 7 shows investor ?ows cause a signi?cant reduction in both average returns and Sharpe ratios (SR)
earned by mutual fund investors. Panel A shows the results using stock returns. A representative investor who
is currently behaving like the aggregate mutual fund sector could increase his Sharpe ratio by 11% (from a
monthly Sharpe ratio of 0.132 to 0.146) by refraining from active reallocation and just directing his ?ows
proportionally.
10
One can assess the signi?cance of this difference in mean returns by looking at the returns on the long-short
portfolio R
ACTUAL
ÀR
NOFLOW
. This return is similar to the long-short portfolio studied in Table 2, except that
here all stocks owned by the mutual fund sector are included, and the weights are proportional to the dollar
value of the holdings. The differential returns are negative and highly signi?cant. Thus investor ?ows cause
wealth destruction. This conclusion is, of course, a partial equilibrium statement. If all investors switched to
proportional ?ows, presumably stock prices would change to re?ect that. But for one individual investor, it
appears that fund ?ows are harmful to wealth.
In Panel B, we repeat the basic analysis, again using three-year ?ows but using funds instead of stocks. We
de?ne R
ACTUAL
and R
NOFLOW
using fund returns instead of stock returns (plugging in actual fund returns for
the term in brackets in Eqs. (9) and (10)). Using mutual fund returns allows us to avoid issues involving
matching funds with holdings. On the other hand, the cost of this speci?cation is that the results now also
re?ect issues such as fund expenses, fund turnover and trading costs, and fund cash holdings. The results in
Panel B are slightly stronger. Using mutual fund returns, the reduction in Sharpe ratio due to ?ows is 17%,
and the magnitude of the dumb money effect (measured by R
ACTUAL
ÀR
NOFLOW
) is somewhat higher. So,
measured using either mutual fund returns or stock returns, investors are lowering their wealth and their
Sharpe ratios by engaging in disproportionate fund ?ows. A simple passive strategy would dominate the
actual strategy of the aggregate mutual fund investors.
Table 7 also helps disentangle the effect of ?ows from the effect of manager stock picking. We start
by considering the average of R
ACTUAL
ÀR
M
, which measures the net return bene?t of owning the aggregate
fund holdings instead of holding the market (ignoring trading costs and expenses). R
M
is the return on the
CRSP value weighted market. The average of this difference consists of two components. The ?rst,
R
ACTUAL
ÀR
NOFLOW
, is the net bene?t of reallocations. We have already seen that this dumb money effect is
negative. The second, R
NOFLOW
ÀR
M
, measures the ability of the mutual fund managers to pick stocks which
outperform the market (using value weights for managers). As shown in the table, using stock returns, this
stock picking effect is 0.087 per month, with a t-statistic of 1.9. Thus, there is some modest evidence that
mutual fund managers do have the ability to pick stocks that outperform the market, once one controls for
their clients’ tendencies of switching money from one fund to another. As shown in the table, this modest skill
is obscured (when looking only at actual holdings) by their clients’ anti-skill at picking funds. Looking at fund
returns, as usual, costs and expenses eat up any stock picking ability managers have, so that the net bene?t of
stock picking in Table 7 is À0.03% per month.
5.2. Economic magnitude
The magnitude of the dumb money effect in Table 7 is on average seven to nine basis points per month
(depending upon whether one uses fund or stock returns). Is this number a large effect? We argue that it is, for
two reasons. First, it results in sizeable reductions in Sharpe ratios of 11–17%. Second, seven to nine basis
points per month is comparable in magnitude to the costs of active fund management. The average expense
ratio for a typical mutual fund is around 1% per year, which translates into eight basis points per month. In
this sense, the dumb money effect costs as much as the entire mutual fund industry.
The results in Panel B give us some context for the economic magnitude of the wealth destruction due to
fund ?ows. The total net bene?t of mutual funds, R
ACTUAL
ÀR
M
, is À0.12% per month, or about 1.4% per
ARTICLE IN PRESS
10
Lamont (2002) ?nds similar results for the policy of refraining from buying new issues.
A. Frazzini, O.A. Lamont / Journal of Financial Economics 88 (2008) 299–322 316
year. Of this, almost 70%, À0.085%, is explained by the dumb money effect.
11
A large literature has
documented that the mutual fund sector does poorly relative to passive benchmarks (see, for example,
Malkiel, 1995). The results here show that fund ?ows appear to account for a large fraction of this poor
performance. Thus, the damage done by actively managed funds comes less from fees and expenses, and more
from the wealth-destroying reallocation across funds.
In Table 8 we explore the robustness of the economic signi?cance in two ways. First, we repeat the basic
analysis for different horizons. It turns out that, at any horizon, individual retail investors are reducing their
wealth by engaging in active reallocation across mutual funds. Even at the three-month horizon, we ?nd no
evidence that trading helps investors earn higher returns.
Second, we report the results for different subperiods. The effect is robust and large across all subperiods,
indicating that the dumb money effect is not only concentrated in the latest part of the sample period. The
results are particularly consistent across time using mutual fund returns.
ARTICLE IN PRESS
Table 8
Robustness tests for economic signi?cance of ?ows
This table shows calendar time portfolio returns for different horizons. R
ACTUAL
is returns on a mimicking portfolio for the entire
mutual fund sector, with portfolio weights the same as the actual weights of the aggregate mutual fund sector. R
NOFLOW
is returns on a
mimicking portfolio for the counterfactual mutual fund sector, with portfolio weights the same as the counterfactual weights of the
aggregate mutual fund sector.
R
ACTUAL
–R
NOFLOW
All sample Exclude
NBER
recessions
Only NBER
recessions
83–93 94–03 83–98 99–03
Panel A: using stock returns
3-Month ?ow À0.015 À0.018 0.024 À0.036 0.007 À0.036 0.048
(À1.23) (À1.46) (0.43) (À2.16) (0.38) (À2.64) (1.94)
6-Month ?ow À0.019 À0.024 0.038 À0.038 À0.000 À0.039 0.041
(À1.54) (À1.89) (0.63) (À2.27) (À0.01) (À2.95) (1.39)
1-Year ?ow À0.039 À0.040 À0.015 À0.050 À0.028 À0.050 À0.003
(À2.69) (À2.80) (À0.21) (À2.92) (À1.19) (À3.75) (À0.08)
3-Year ?ow À0.069 À0.069 À0.069 À0.061 À0.077 À0.064 À0.084
(À4.17) (À4.10) (À0.89) (À2.64) (À3.24) (À3.69) (À2.03)
5-Year ?ow À0.059 À0.058 À0.069 À0.061 À0.058 À0.071 À0.024
(À2.93) (À2.85) (À0.72) (À2.18) (À1.96) (À3.43) (À0.46)
Panel B: using mutual fund returns
3-Month ?ow À0.042 À0.040 À0.068 À0.046 À0.037 À0.042 À0.041
(À2.89) (À2.63) (À1.38) (À2.11) (À1.98) (À2.58) (À1.31)
6-Month ?ow À0.045 À0.042 À0.079 À0.050 À0.039 À0.044 À0.047
(À2.98) (À2.66) (À1.73) (À2.25) (À1.94) (À2.65) (À1.38)
1-Year ?ow À0.055 À0.054 À0.067 À0.056 À0.055 À0.050 À0.071
(À3.23) (À3.00) (À1.54) (À2.35) (À2.21) (À2.82) (À1.63)
3-Year ?ow À0.085 À0.081 À0.147 À0.063 À0.108 À0.057 À0.173
(À4.09) (À3.79) (À1.51) (À2.49) (À3.25) (À3.00) (À2.84)
5-Year ?ow À0.074 À0.068 À0.145 À0.050 À0.094 À0.054 À0.127
(À2.97) (À2.64) (À1.43) (À1.80) (À2.39) (À2.59) (À1.75)
11
Of course, this calculation may be misleading because the return earned by the CRSP value weight portfolio is not a viable free
alternative. We have redone the calculation, substituting the return on the Vanguard 500 Index Fund for R
M
(these returns include fees
and costs). In this case, the total wealth destruction is À0.16% instead of À0.12% (re?ecting the fact the Vanguard fund outperformed the
CRSP value weight portfolio during this period), while the dumb money effect remains of course at À0.085%.
A. Frazzini, O.A. Lamont / Journal of Financial Economics 88 (2008) 299–322 317
6. Issuance
If individual investors (acting through mutual funds) lose money on their trades, who is making money?
Possible candidates include hedge funds, pension funds, other institutions, or individuals trading individual
stocks. Here we focus on another class of traders: ?rms. In contrast to trading by individuals, re?ecting
uninformed and possibly irrational demand, the actions of ?rms represent informed and probably more
rational supply. A substantial body of research studies whether ?rms opportunistically take advantage of
mispricing by issuing equity when it is overpriced and buying it back when it is underpriced (for example,
Loughran and Ritter, 1995). Corporate managers certainly say they are trying to time the market (Graham
and Harvey, 2001).
We measure ?rm behavior using the composite share issuance measure of Daniel and Titman (2006), which
combines a variety of previously documented effects involving repurchases, mergers, and seasoned equity
issues (see also, Pontiff and Woodgate, 2005). Our version of their variable is 1 minus the ?rm’s ratio of the
number of shares outstanding one year ago to the number of shares outstanding today.
12
For example, if the
company has 100 shares and has a seasoned equity issue of an additional 50 shares, the composite issuance
measure is 33%, meaning that 33% of the existing shares today were issued in the last year. The measure can
be negative (re?ecting, for example, repurchases) or positive (re?ecting, for example, executive stock options,
seasoned equity offerings, or stock-?nanced mergers). Issuance and market-book ratios are strongly related:
growth ?rms tend to issue stock, value ?rms tend to repurchase stock. Daniel and Titman (2006) show that
when issuance is high, returns are low over the next year. This pattern suggests that ?rms issue and repurchase
stock in response to mispricing.
Table 9 shows the relation of annual issuance to past three-year ?ows, using the usual format but studying
issuance instead of returns. The table shows issuance between January and December of year t, sorted on
three-year ?ows as of December in year tÀ1. The table uses the standard portfolio logic of forming groups,
taking the average in each group for each of the 20 years available, and reporting the mean and t-statistic for
the resulting 20 time series observations.
The ?rst row shows that ?rms with the lowest three year in?ows issue 1% less stock than ?rms with the
highest in?ows. Thus, in?ows are positively associated with issuance by ?rms. Firms tend to increase shares
outstanding this year when previous year’s ?ows are high. One interpretation of this pattern is that ?rms are
seizing the opportunity to issue stocks when sentiment is high, and repurchase stocks when sentiment is low.
Since average issuance is around 3% (as a fraction of shares outstanding) per year in this sample, 1% is a large
number.
The rest of the table shows robustness tests for this basic result. The next row shows a truncated version of
the issuance variable. Since the issuance variable as de?ned is unbounded below, we de?ne trimmed issuance
as max (À100, issuance). This change has little effect. We also look at the relation in the two different halves of
the sample. As before, the relation is stronger in the second half of the sample, but signi?cant always. Lastly,
because issuance is known to be correlated with valuation, we create characteristic-adjusted issuance in the
same way we create characteristic-adjusted returns in Table 3. The last row of Table 9 shows the average
deviations of issuance from a group of matching ?rms with similar size, valuation, and price momentum as of
December. The results are about the same as with raw issuance, so that once again value does not subsume the
effect of ?ows.
To understand the economic magnitudes shown in Table 9, it is useful to note from Table 2 that the
difference in the sorting variable (three-year ?ow) is about 10% between the top and bottom quintile. That is,
as a result of active reallocation across mutual funds in the past three years, the top quintile has a mutual fund
ownership that is on average 10% more as a percent of shares outstanding than the bottom quintile. This
number is in the same units as the numbers in Table 9 since both ?ows and issuance are expressed as a fraction
of current shares outstanding. Thus, ?rms with ?ows that are 10% higher as a fraction of shares outstanding
tend to increase shares by 1% of shares outstanding. Over three years, the ?rm would issue shares equivalent
to 3% of shares outstanding. Thus, over time, one can loosely say that ?rms respond to $10 billion in ?ows by
issuing $3 billion in stock. Supply accommodates approximately one third of the increase in demand.
ARTICLE IN PRESS
12
We split-adjust the number of shares using CRSP ‘‘factor to adjust shares.’’
A. Frazzini, O.A. Lamont / Journal of Financial Economics 88 (2008) 299–322 318
7. Conclusion
In this paper, we have shown that individual investors have a striking ability to do the wrong thing. They
send their money to mutual funds which own stocks that do poorly over the subsequent few years. Individual
investors are dumb money, and one can use their mutual fund reallocation decisions to predict future stock
returns. The dumb money effect is robust to a variety of different control variables, is not entirely due to one
particular time period, and is implementable using real-time information. By doing the opposite of individuals,
one can construct a portfolio with high returns. Individuals hurt themselves by their decisions, and we
calculate that the aggregate mutual fund investor could raise his Sharpe ratio simply by refraining from
destructive behavior.
Investors achieve low returns by a combination of different channels: they tend to both overweight growth
stocks and select securities that on average underperform their growth benchmarks. Within new issues, they
overweight stocks with especially low subsequent returns. All of the effects above generate poor performance
of the stock portfolio investors indirectly hold via their mutual fund investments.
We have found only weak evidence of a smart money effect of short-term ?ows positively predicting
short-term returns. One interpretation of this effect is that there is some short-term manager skill which is
detected by investors. Another hypothesis, explored by Wermers (2004) and Coval and Stafford (2007), is that
mutual fund in?ows actually push prices higher. Another possibility, explored by Sapp and Tiwari (2004), is
that by chasing past returns, investors are stumbling into a useful momentum strategy. Whatever the
explanation, it is clear that the higher returns earned at the short horizon are not effectively captured by
individual investors. Of course, it could be that some subset of individuals bene?ts from trading, but looking
at the aggregate holdings of mutual funds by all individuals, we show that individuals as a whole are hurt by
their reallocations.
The evidence on issuers and ?ows presents a somewhat nonstandard portrait of capital markets. Past papers
have looked at institutions vs. individuals, and tried to test if institutions take advantage of individuals. Here,
the story is different. Individuals do trade poorly, but these trades are executed through their dynamic
allocation across mutual funds, that is, via ?nancial institutions. As far as we can tell, it is not ?nancial
institutions that exploit the individuals, but rather the non-?nancial institutions that issue stocks and
ARTICLE IN PRESS
Table 9
Issuance
This table shows issuance activity between January and December of year t+1, for portfolios of ?rms sorted on three-year ?ows as of
December in year t. In December stocks are ranked in ascending order based on the last available three-year ?ow. Stocks are assigned to
one of ?ve portfolios. Portfolios are rebalanced every year to maintain value weights. Issuance is de?ned as 1 minus the ?rm’s ratio of the
number of shares outstanding one year ago to the number of shares outstanding today. Issuance is in percent, t-statistics are shown below
the coef?cient estimates. DGTW characteristic adjusted issuance is de?ned as raw issuance minus the average issuance on an equally
weighted portfolio of all CRSP ?rms with non-missing ?ows in the same size, market-book, and one year momentum quintile.
Low ?ow High ?ow High ?ow minus low ?ow
Q1 Q2 Q3 Q4 Q5
Raw issuance 1.828 0.823 0.896 1.607 3.162 1.334
(7.73) (2.74) (2.80) (4.95) (6.35) (2.85)
Trimmed issuance 1.959 1.017 0.974 1.647 3.248 1.289
(8.81) (3.72) (3.27) (5.09) (6.53) (2.69)
Raw issuance 1981–1993 1.394 0.179 0.078 0.922 2.387 0.992
(4.30) (0.49) (0.21) (2.77) (4.68) (2.13)
Raw issuance 1994–2004 2.262 1.466 1.715 2.293 3.937 1.675
(7.56) (3.74) (4.40) (4.77) (4.88) (2.03)
DGTW adjusted issuance À0.654 0.012 0.120 À0.110 0.239 0.893
(À3.63) (0.08) (0.96) (À1.03) (1.68) (3.99)
A. Frazzini, O.A. Lamont / Journal of Financial Economics 88 (2008) 299–322 319
repurchase stocks. Stocks go in and out of favor with individual investors, and ?rms exploit this sentiment by
trading in the opposite direction of individuals, selling stock when individuals want to buy it. We ?nd some
modest evidence that mutual fund managers have stock picking skill, but that any skill is swamped by other
effects including the actions of retail investors in switching their money across funds. In our data, ?nancial
institutions seem more like passive intermediaries who facilitate trade between the dumb money (individuals)
and the smart money (?rms).
Although the dumb money effect is statistically distinct from the value/reversal effect, it is clear these two
effects are highly related. It is remarkable that one is able to recover many features of the value effect without
actually looking at prices or returns for individual stocks. It is clear that any satisfactory theory of the value
effect will need to explain three facts. First, value stocks have higher average returns than growth stocks.
Second, using various issuance mechanisms, the corporate sector tends to sell growth stocks and buy value
stocks. Third, individuals, using mutual funds, tend to buy growth stocks and sell value stocks. One coherent
explanation of these three facts is that individual investor sentiment causes some stocks to be misvalued
relative to other stocks, and that ?rms exploit this mispricing.
Appendix A
A.1. Construction of the counterfactual ?ows
We assign a counterfactual total net asset value of zero to funds that were newly created in the past k
quarters. New funds represent new ?ows, but in the counterfactual exercise they do not receive assets for the
?rst k quarters. The universe of funds we consider when computing the counterfactual ?ows between date tÀk
and date t is funds that were alive at both date tÀk and t.
More speci?cally, consider at generic date t and let F
Agg
s
be the actual aggregate ?ows for all funds alive in
quarter t (including funds that were recently born, but excluding funds that die in month t), for t À kpspt.
Let TNA
Agg
tÀk
be the lagged actual aggregate TNA aggregating only over those funds that exist in both month
tÀk and in month t. We compute the counterfactual ?ows by assigning to each fund a share of the total as
follows:
^
F
i
s
¼
TNA
i
tÀk
TNA
Agg
tÀk
F
Agg
s
(11)
t À kpspt
For funds that die in quarter s+1 (so that their last TNA is quarter s), we set
^
F
i
sþ1
¼ À
d
TNA
i
s
and
d
TNA
i
sþh
¼ 0 for all h40.
Table A1 shows a simpli?ed example where we set k ¼ 1 year. Fund 3 is born in 1981, therefore in 1981 we
register a net in?ow equal to its initial TNA and set the counterfactual TNA to zero. In 1981 two funds are
alive, Fund 1 and Fund 2, and in 1980 they represented two-thirds and one-third of the total fund sector.
Aggregate ?ows in 1981 were equal to $150, hence in the counterfactual exercise we assign a ?ow of $100 to
Fund 1 (as opposed to the actual realized ?ow of $50) and a ?ow of $50 to Fund 2. Given the return of the two
funds between 1980 and 1981, we can compute the counterfactual total net asset value of Fund 1 and 2 in
1981. Proceeding in the same manner whenever a fund is alive at date tÀk and t, we track the evolution of the
fund’s counterfactual TNA using the recursion:
d
TNA
i
t
¼ ð1 þ R
i
t
Þ
d
TNA
i
tÀ1
þ
^
F
i
t
. (12)
Between 1982 and 1993 Fund 2 dies, hence in the counterfactual world we assign an out?ow in 1983 equal to
the TNA in 1982 and set the counterfactual TNA to zero thereafter. Note that Eq. (12) does not guarantee
that counterfactual total net asset values are always non-negative in quarters where we have aggregate
out?ows (F
Agg
t
o0 ). In this case we override Eq. (12), set
d
TNA
i
t
¼ 0 and redistribute the corresponding
counterfactual ?ows to the remaining funds, to keep the total aggregate dollar out?ow the same in both the
ARTICLE IN PRESS
A. Frazzini, O.A. Lamont / Journal of Financial Economics 88 (2008) 299–322 320
counterfactual and actual case. Measuring FLOW over 12 quarters, negative counterfactual TNAs occur for
only 0.08% of the sample.
Finally, we handle mergers as follows: we assume that investors keep earning returns on the existing assets
of the surviving fund. For consistency, when constructing the counterfactual TNA, we also merge the lagged
TNA of the two funds when we compute the ratio TNA
i
tÀk
=TNA
Agg
tÀk
used to determine the pro-rata share of
the total ?ows.
References
Baker, M., Wurgler, J., 2006. Investor sentiment and the cross-section of stock returns. Journal of Finance 61, 1645–1680.
Barber, B., Lyon, J., 1997. Detecting abnormal operating performance: the empirical power and speci?cation of test statistics. Journal of
Financial Economics 41, 341–372.
Barber, B., Odean, T., 2000. Trading is hazardous to your wealth: the common stock investment performance of individual investors.
Journal of Finance 55, 773–806.
Barber, B., Odean, T., 2001. Boys will be boys: gender, overcon?dence, and common stock investment, Quarterly Journal of Economics
116, 261–292.
Barber, B., Odean, T., 2004. All that glitters: the effect of attention and news on the buying behavior of individual and institutional
investors. The Review of Financial Studies 21, 785–818.
Barber, B., Odean, T., Zheng, L., 2004. Out of sight, out of mind: the effects of expenses on mutual fund ?ows. Journal of Business 78,
2095–2120.
Barberis, N., Shleifer, A., 2004. Style investing. Journal of Financial Economics 68, 161–199.
Brown, S., Goetzmann, W., Hiraki, T., Shiraishi, N., Watanabe, M., 2002. Investor sentiment in Japanese and US daily mutual fund
?ows. Yale ICF Working Paper No. 02-09.
ARTICLE IN PRESS
Table A1
Hypothetical example showing counterfactual calculation
Year
1980 1981 1982 1983 1985
Actual data from individual funds
Returns Fund 1 (%) 10 10 5 10 5
Fund 2 (%) À5 10 À10
Fund 3 (%) 10 10 5
TNA Fund 1 100 160 268 395 515
Fund 2 50 105 144 0 0
Fund 3 50 45 100 154
Flows Fund 1 50 100 100 100
Fund 2 50 50 À144 0
Fund 3 50 À10 50 50
Actual data for aggregates
TNA Agg. 150 315 457 494 669
FLOW Agg. 0 150 140 6 150
TNA, last year, of funds existing
this year
Agg. 150 315 313 494
FLOW of non-dying funds Agg. 150 140 150 150
Counterfactual data
TNA Fund 1 100 210 292 449 591
Fund 2 50 105 141 0 0
Fund 3 22 46 79
Flows Fund 1 100 71 128 120
Fund 2 50 47 À141 0
Fund 3 22 22 30
A. Frazzini, O.A. Lamont / Journal of Financial Economics 88 (2008) 299–322 321
Chen, H., Jegadeesh, N., Wermers, R., 2000. The value of active mutual fund management: an examination of the stockholdings and
trades of fund managers. Journal of Financial and Quantitative Analysis 35, 343–368.
Chevalier, J., Ellison, G., 1997. Risk taking by mutual funds as a response to incentives. Journal of Political Economy 105, 1167–1200.
Cooper, M., Huseyin, G., Rau, P., 2005. Changing names with style: mutual fund name changes and their effects on fund ?ows. Journal of
Finance 60, 2825–2858.
Coval, J., Stafford, E., 2007. Asset ?re sales (and purchases) in equity markets. Journal of Financial Economics 86, 479–512.
Daniel, K., Grinblatt, M., Titman, S., Wermers, R., 1997. Measuring mutual fund performance with characteristic-based benchmarks.
Journal of Finance 52, 1035–1058.
Daniel, K., Titman, S., 2006. Market reaction to tangible and intangible information. Journal of Finance 61, 1605–1643.
De Bondt, W., Thaler, R., 1985. Does the stock market overreact? Journal of Finance 40, 793–805.
Fama, E., French, K., 1993. Common risk factors in the returns on stocks and bonds. Journal of Financial Economics 33, 3–56.
Fama, E., MacBeth, J., 1973. Risk, return, and equilibrium: empirical tests. Journal of Political Economy 81, 607–636.
Frazzini , A., Lamont, O., 2005. Dumb money: mutual fund ?ows and the cross-section of stock returns. NBER Working Paper No.
11526.
Goetzmann, W., Massa, M., 2002. Daily momentum and contrarian behavior of index fund investors. Journal of Financial and
Quantitative Analysis 37, 375–390.
Graham, J., Harvey, C., 2001. The theory and practice of corporate ?nance: evidence from the ?eld. Journal of Financial Economics 60,
187–243.
Grinblatt, M., Keloharju, M., 2000. The investment behavior and performance of various investor types: a study of Finland’s unique data
set. Journal of Financial Economics 55, 43–67.
Gruber, M., 1996. Another puzzle: the growth in actively managed mutual funds. Journal of Finance 51, 783–810.
Ippolito, R., 1992. Consumer reaction to measures of poor quality: evidence from the mutual fund industry. Journal of Law and
Economics 35, 45–70.
Jain, P., Wu, J., 2000. Truth in mutual fund advertising: evidence on future performance and fund ?ows. Journal of Finance 55, 937–958.
Jegadeesh, N., Titman, S., 1993. Returns to buying winners and selling losers: implications for stock market ef?ciency. Journal of Finance
48, 65–91.
Lamont, O., 2002. Evaluating value weighting: corporate events and market timing. NBER Working Paper No. 9049.
Loughran, T., Ritter, J., 1995. The new issues puzzle. Journal of Finance 50, 23–51.
Malkiel, B., 1995. Returns from investing in equity mutual funds 1971 to 1991. Journal of Finance 50, 549–572.
Odean, T., 1999. Do investors trade too much? American Economic Review 89, 1279–1298.
Ofek, E., Richardson, M., 2003. DotCom mania: the rise and fall of internet stock prices. Journal of Finance 58, 1113–1137.
Pontiff, J., Woodgate, A., 2005. Shares outstanding and cross-sectional returns. Working Paper, Boston College.
Polk, C., Sapienza, P., 2008. The stock market and corporate investment: a test of catering theory. The Review of Financial Studies,
forthcoming.
Sapp, T., Tiwari, A., 2004. Does stock return momentum explain the smart money effect? Journal of Finance 59, 2605–2622.
Shumway, T., 1997. The delisting bias in CRSP data. Journal of Finance 52, 327–340.
Sirri, E., Tufano, P., 1998. Costly search and mutual fund ?ows. Journal of Finance 53, 1589–1622.
Teo, M., Woo, S., 2004. Style effects in the cross-section of stock returns. Journal of Financial Economics 7, 367–398.
Warther, V., 1995. Aggregate mutual fund ?ows and security returns. Journal of Financial Economics 39, 209–235.
Wermers, R., 1999. Mutual fund trading and the impact on stock prices. Journal of Finance 54, 581–562.
Wermers, R., 2004. Is money really smart? New evidence on the relation between mutual fund ?ows, manager behavior, and performance
Persistence. Working Paper University of Maryland.
Zheng, L., 1999. Is money smart? A study of mutual fund investors’ fund selection ability. Journal of Finance 54, 901–933.
ARTICLE IN PRESS
A. Frazzini, O.A. Lamont / Journal of Financial Economics 88 (2008) 299–322 322

doc_700401706.pdf