WSJ Category Kings

Journal of Financial Economics 123 (2017) 337–356
Contents lists available at ScienceDirect
Journal of Financial Economics

journal homepage: www.elsevier.com/locate/jfec
WSJ Category Kings – The impact of media attention on

consumer and mutual fund investment decisionsR
Ron Kaniel a, Robert Parham b,∗
a
University of Rochester - United States, CEPR, and IDC
b
University of Rochester, United States
a r t i c l e i n f o a b s t r a c t
Article history: We exploit a novel natural experiment to establish a causal relation between media at-
Available online 14 November 2016 tention and consumer investment behavior, independent of the conveyed information. Our
findings indicate a 31% local average increase in quarterly capital flows into mutual funds
JEL Classification:
mentioned in a prominent Wall Street Journal “Category Kings” ranking list, compared to
D14
those funds which just missed making the list. This flow increase is about seven times
D83
G11 larger than extra flows due to the well-documented performance-flow relation. Other
G14 funds in the same fund complex receive substantial extra flows as well, especially in
G23 smaller complexes. There is no increase in flows when the Wall Street Journal publishes
similar lists absent the prominence of the Category Kings labeling. We show mutual fund
Keywords: managers react to the incentive created by the media effect in a strategic way predicted
Mutual funds
by theory, and present evidence for the existence of propagation mechanisms including
Investor attention
increased fund complex advertising subsequent to having a Category King and increased
Incentives
Flows efficacy of subsequent fund media mentions.
Regression discontinuity © 2016 Elsevier B.V. All rights reserved.
1. Introduction in the media impacts financial decision making, indepen-

dently of the information conveyed. To provide evidence
It is widely accepted that information disseminated by of a relation between media attention and financial deci-
the media informs consumer decision making in financial sions while mitigating the confounding effects of informa-
markets.1 Our goal, however, is to show that appearance tion simultaneously released in the media announcement,
we exploit a clean natural experiment in which the Wall
R
The research leading to these results has received funding from the
Street Journal (WSJ) has prominently published the top ten
European Research Council under the European Union’s Seventh Frame- mutual funds, ranked within various commonly used in-
work Programme (FP7/2007-2013) / ERC Grant Agreement no. [312842]. vestment style categories, every quarter since 1994. Rank-
We wish to thank Susan Christoffersen (the referee), René M. Stulz (the ings are simply based on previous 12-month returns, en-
editor), Jonathan Reuter, Laura Starks, and seminar participants at the
suring both minimal editorial impact and quasi-random as-
University of Rochester, University of Connecticut, the American Economic
Association, the American Finance Association and the European Financial signment around the publication cutoff of rank = 10. The
Management Association annual meetings for helpful comments and sug- top-10 ranking lists are part of an independent section, “In-
gestions. Shelly Massachi, Robert (Jay) Kahn, Arash Amoozegar, Yihua Nie vesting in funds - A quarterly analysis,” and have an eye-
and Diego Leal Gonzalez provided valuable research assistance. All errors catching heading, “Category Kings.”
are our own.
∗ We show a clear discontinuity in capital flows fol-
Corresponding author.
E-mail addresses: [email protected] (R. Kaniel), robert. lowing publication between funds which appeared in the
[email protected] (R. Parham). ranking and those which did not. Using a regression dis-
1
See, e.g., Peress (2014). continuity design (RDD), we find a significant local average
http://dx.doi.org/10.1016/j.jfineco.2016.11.003
0304-405X/© 2016 Elsevier B.V. All rights reserved.
338 R. Kaniel, R. Parham / Journal of Financial Economics 123 (2017) 337–356
treatment effect, between funds ranked 10 (published) role in propagating the effect. By analyzing fund complex
and 11 (unpublished), of 2.2 percentage points increase advertising behavior and media coverage, we are able to
in flow of capital into the published funds during the show mutual fund complexes increase advertising activi-
post-publication quarter.2 This represents a hefty 31% in- ties (expenditure, average ad size, and number of ads) in
crease in capital flows during the post-publication quarter, response to appearance in the WSJ rankings, and enjoy
indicating consumers strongly react to media attention increased efficacy for mentioning their ranking in ads or
directed at these funds. The publication effect on flows is for being mentioned in news and business articles. We
roughly seven times larger in magnitude than the effect of thus establish several possible propagation mechanisms
the well-documented performance-flow relation.3 of the media effect. These protracted propagation mech-
We establish that the prominence of the publication anisms are also consistent with our finding that capital
and its visibility are key to driving the media effect. Simi- flow increases are gradual throughout the quarter, imply-
lar WSJ ranking tables, based on year-to-date return, which ing consumers do not rush to change investment alloca-
were published monthly on a regular basis, yet less promi- tions following the WSJ publication, but rather are influ-
nently, caused no significant increase in flows following enced by it when making allocation and rebalancing de-
publication. The lack of increase in flows holds even when cisions throughout the quarter. We also show that small,
restricting to December ranking lists, which rank based on young funds from small complexes, which are ex ante less
11-month returns. Thus, the impact of the Category Kings visible, enjoy a higher “bang for the buck” from being pub-
lists on investor decisions is mostly due to their visibility lished, again consistent with the importance of visibility.
and prominence, and not to their “information role” (Del In sharp contrast to the sizable effects we identify for
Guercio and Tkac, 2008). Furthermore, even after the WSJ the highly visible Category King lists, and as additional
made all rankings readily available on its website, starting supportive evidence of the special role these lists play,
in 2007, thus making the ranking information readily avail- we observe no significant discontinuities in capital flows
able for all funds and not only the top ten in each cate- for falsification tests in which: we only examine cate-
gory, there was no significant decrease in the discontinu- gories which were not published in the WSJ; the analysis
ous flows garnered by the funds prominently published in is shifted in time to not coincide with the Category Kings
the WSJ Category Kings lists. publication; the ranking is based on the most recent 11-
We further show that, subsequent to publication of the rather than 12-month return. We also show that the dis-
Category Kings lists, consumers not only “chase” published continuity at rank = 10 for the Category King lists is unique
funds. They also change their attitude towards the entire and does not exist for other plausible cutoffs.
brand/complex: there is a sizable spillover effect of 1.8 Section 2 discusses the related literature and puts our
percentage points increase in capital flows into the other study in context. Section 3 describes the data used and
funds of the complex in the subsequent quarter. This find- provides summary statistics. Section 4 presents our full
ing is consistent with an impact of media attention and empirical strategy and results. Section 5 concludes. The
visibility on brand-name recognition at the complex level, Appendix presents further evidence for the validity of the
and is less consistent with an information channel. RDD and the robustness of our results to empirical design
The existence of a media effect on consumer financial choices.
decision making implies that fund managers’ payoffs re-
semble a call option due to the implicit asymmetric incen-
2. Related literature
tives induced by the extra flows.4 Consistent with theoreti-
cal predictions by Basak et al. (2007) and Cuoco and Kaniel
The existence of a pure media effect is a natural theo-
(2011), we show that funds ranked near the rank = 10 cut-
retical result of costly information gathering by consumers
off at the beginning of the last ranking month, and only
in the spirit of Grossman and Stiglitz (1980). When search
these funds, “diverge from the herd” by increasing tracking
is costly, the mere appearance of a financial instrument in
error volatility relative to their category in an attempt to
the media leads consumers to add the instrument to their
make the list. A closer analysis reveals that funds are well
limited “consideration set,” as proposed by Merton (1987).5
aware of the trade-offs induced by this risk-shifting: within
Mutual funds are specifically useful in exploring the im-
funds ranked near the cutoff, only those that are unlikely
pact of media attention on investment choices as fund cap-
to be ranked as top performing funds next quarter increase
ital flows are readily available at a fairly high frequency,
tracking error volatility.
in contrast to other investor allocation decisions which are
We further establish that both subsequent fund adver-
much harder to obtain. Moreover, mutual funds represent
tising and efficacy of subsequent media mentions play a
a significant component of many U.S. households’ finan-
cial holdings. In 2013, 69% of U.S. households with income
above $50,0 0 0 owned mutual funds.6 Finally, for many fi-
2
As described below, capital flows are measured as percent increase nancial instruments, demand forces resulting from media
in fund assets. To avoid confusion, we refer to flow values in percentage
appearance change the price of the instrument in the short
point terms, while reserving the percent sign, %, for denoting proportions
of other values, as usual.
3
E.g., Gruber (1996), Carhart (1997), Chevalier and Ellison (1997).
4 5
And the fact that management fees are determined as a percent of See Corwin and Coughenour (2008) for a discussion on the impact of
fund size. See Brown et al. (1996) and Chevalier and Ellison (1997) for effort allocation due to limited attention in financial markets.
6
tests of fund manager risk-shifting in the presence of call-option-like pay- Source: Investment Company Institute 2014 Fact Book available at
offs. www.icifactbook.org.
R. Kaniel, R. Parham / Journal of Financial Economics 123 (2017) 337–356 339
term, whereas mutual funds’ prices are related to the per- sever the link between local content publication and local
formance of the underlying portfolio. This decoupling of trading.
demand and price greatly simplifies the analysis of media A shortcoming of using population splits is the need
impact.7 to control for determinants of a media outlet’s decision to
Several authors examine and establish the correlation publish specific content and for characteristics of the sub-
between media attention and consumer investment be- populations exposed to the content, which may compli-
havior. Sirri and Tufano (1998) consider media atten- cate the identification. Engelberg and Parsons (2011), for
tion as one of three proxies for the magnitude of search example, utilize controls for earnings, investor, and news-
costs associated with purchasing a mutual fund. They use paper characteristics, in addition to controls aimed at cap-
Lexis/Nexis mentions of mutual funds in the media and turing home bias on the part of investors and local me-
correlate them with capital flows while controlling for dia. We are able to eliminate concerns regarding selection
fund characteristics, with mixed results. Similarly, Barber bias, by focusing on regularly appearing style categories in
and Odean (2008) construct a measure based on mentions the WSJ top-10 ranking tables, and taking advantage of the
of companies in the Dow Jones News Service daily feed, fact that the WSJ uses a pre-specified fixed explicit algo-
as one of three proxies for media attention. They find that rithm to rank the funds. Thus, our setting eliminates the
investors are more likely to be net buyers of stocks men- bias arising from the endogeneity of the decision to pub-
tioned in the news than of those not mentioned.8 Kaniel lish a specific media article regarding a specific investment
et al. (2007) correlate the existence and frequency of me- vehicle.
dia coverage of mutual funds to subsequent capital flows, One channel through which the media can affect con-
and Solomon et al. (2014) further correlate media men- sumer investment behavior is the “information digestion”
tions of fund holdings to subsequent flows into the fund. channel proposed by Del Guercio and Tkac (2008), who
Tetlock (2007) uses textual analysis of a WSJ opinion col- use Morningstar ratings to explore the effect of ranking
umn to create a proxy for media sentiment towards the on fund flows. Del Guercio and Tkac (2008) conduct event
stock market and finds that it is associated with past and studies of over 10,0 0 0 Morningstar rating changes and
future returns of the Dow Jones Industrial Average and show that these discrete rating changes lead to changes
with future trading volumes on the New York Stock Ex- in mutual fund flows, above and beyond those predicted
change. Finally, Fang et al. (2014) consider the impact of by a time-series benchmark regression of fund fundamen-
media coverage of stocks on mutual fund trades. tals. They conclude that repackaging of fund-quality infor-
A limitation of these inquiries is that they are restricted mation into simple discrete ratings like Morningstar assists
in their ability to make causal claims regarding the impact investors facing search costs to digest information easily.
of media visibility, due to the endogeneity of media report- It is important to note, however, that as all star rankings
ing – an item is in the news if there is news to report. are published simultaneously on Morningstar’s website, it
Our identification strategy is tailored to alleviate such en- is not possible to distinguish the effect of pure media vis-
dogeneity concerns. ibility from information content under their setting. We
Prior attempts to alleviate endogeneity concerns re- aim to fill this gap, and are able to attribute a significant
garding the impact of media coverage have generally component of quarterly flows to a single-day appearance
employed population splits in which different groups of in a WSJ category ranking table, while showing that simi-
agents are exposed to different media outlets. For example, lar ranking tables published less prominently do not garner
in the literature concerning the effects of media on voter similar investor response.
political leaning and behavior, using a population splits An important distinction between the WSJ rankings and
approach, previous researchers have shown that both the Morningstar ratings used by Del Guercio and Tkac
television (DellaVigna and Kaplan, 2007; Enikolopov et al., (2008) is that the WSJ rankings are simply based on previ-
2011) and newspapers (Gerber et al., 2009) have an effect ous 12-month returns, and this fact is made explicitly clear
on political attitudes and voting patterns [see DellaVigna in the publications. Morningstar ratings, on the other hand,
and Gentzkow (2010) for a survey]. In a financial context, are calculated using multiple return horizons (three, five,
Engelberg and Parsons (2011) use micro-level trading ten years) with opaque “proprietary” weights given to the
data to show that sub-populations exposed to different different horizons (based on style drift), and the returns
local newspapers differ in investment behavior following are also risk-adjusted.9 As such, Morningstar ratings give
the publication of articles discussing earnings releases of a perception that there is an elaborate evaluation mecha-
Standard & Poors (S&P) 500 Index firms. Engelberg and nism, beyond just information summary, behind them, and
Parsons (2011) further demonstrate how extreme weather are certified by the Morningstar brand. This is not the case
events which may disrupt the delivery of local newspapers with the WSJ rankings.
Our research also contributes to the literature that
analyzes how implicit and explicit incentives impact fund
managers’ investment decisions. Brown et al. (1996) find
7
Admittedly, fund flows may potentially impact a fund’s ability to gen- that the ratio of fund volatility in the second part of
erate subsequent returns, but for the purpose of our study this is a the year to the first is higher for interim losers than
second-order concern.
8
Da et al. (2011) use a direct revealed investor attention measure, de-
rived from Google search frequency of Russell 30 0 0 stock tickers, to pro-
9
vide support to the hypothesis of Barber and Odean (2008) that investors See Reuter and Zitzewitz (2014) for a detailed discussion of the Morn-
are net buyers of attention-grabbing stocks. ingstar rating construction.
for winners, and argue this is consistent with an annual are sectors (e.g., gold, Japan), changing every quarter based
tournament structure. Chevalier and Ellison (1997) show on editor’s choice.11 In our analysis, we concentrate on
that risk-taking behavior in the last quarter, measured by the 12 major categories to eliminate the effect of editorial
tracking error volatility relative to the market, is consistent bias, though including the sector categories in the analy-
with flow-induced incentives implied by the performance sis does not materially change the magnitude or signifi-
in the first three quarters. Carhart et al. (2002) show cance of our results. We note that during within-quarter
winning funds trade to temporarily inflate their fund net months, in which the special issue “Investing in funds”
asset value (NAV) on the very last day of the year, and was not published, the WSJ nevertheless published similar
argue this is to shift performance between years. Del ranking tables. These within-quarter tables were published
Guercio and Tkac (2008) suggest that contrary to the much less prominently in the back pages of the newspa-
implicit assumptions of Brown et al. (1996) and Chevalier per, along with the fund quotes, and the ranking was based
and Ellison (1997), this tournament behavior is much more on year-to-date returns rather than previous 12-month re-
of an ongoing and high frequency tournament, which we turns. We use data on within-quarter months to compare
are able to confirm in our tests. high vs. low visibility publications and establish the im-
Our evidence suggests fund managers are well aware of portance of prominence. Data on published funds and cat-
the impact on fund flows of making the top-10 lists. Fur- egories, as well as precise publication date for each issue,
thermore, we show they understand that for funds close were collected by directly searching for the published ta-
to the publication cutoff, the appropriate strategy to in- bles in microfiche archives of the WSJ.
crease the likelihood of making it onto the list is to in- The regression discontinuity analysis critically depends
crease tracking error volatility relative to other funds in the on the WSJ ranking procedure. This procedure starts with
same ranking category, rather than just increasing volatil- the assignment of each fund to a category. During our
ity. Even more striking, we show that managers under- sample period, category definitions were supplied to the
stand and react to the trade-off involved with this risk- WSJ by an external data vendor, Lipper Analytical Ser-
shifting behavior: among funds near the cutoff, a month vices. At the end of every quarter, funds in each category
before the ranking, only those unlikely to be in a simi- are ranked based on previous 12-month returns. However,
lar position next quarter engage in such diversionary risk- many of these funds have several share classes with dif-
shifting. Finally, we further contribute to the literature by ferent fee structures, and consequently slightly different
showing that in addition to the ex ante impact on invest- net 12-month returns. To ensure each fund is only ranked
ment decisions, ex post family advertising is impacted by once, the WSJ retains the largest share class of each fund,
having a fund in the Category King lists, and this impact is based on total net assets (TNA).
stronger for funds unlikely to be in a similar position next Our method requires the complete list of ranked fund
quarter. classes rather than just the top-10 published. We therefore
replicate the WSJ ranking procedure using data on mutual
3. Data and summary statistics fund returns and characteristics from CRSP. We use cate-
gory definitions and monthly return data to construct pre-
We use data from the following sources: Wall Street vious 12-month returns for all funds in the categories ex-
Journal Category Kings tables – lists of top-10 mutual funds amined, and TNA data to choose the largest share class.
by category appearing in a special quarterly section of Several of our tests require daily return data, which were
the Journal, as well as monthly tables appearing during also obtained from CRSP. As the replication process uses a
within-quarter months at the back pages of the Journal; different data set than the one used for publication in the
Center for Research in Security Prices (CRSP) – returns, WSJ, we do not achieve full replication accuracy. Our rank-
monthly flows, and other fund characteristics; Trimtabs – ing successfully matches that of the WSJ 89% of the time.12
daily flows; Kantar Media – fund complex advertising; In the analysis to follow we replace the CRSP-based top-10
Factiva – print media coverage; Morningstar – star ratings. with the actual published top-10, though our results are
Below we describe how these are used. nearly identical when using the CRSP-based list without a
We consider 52 quarters, from 20 0 0Q1 to 2012Q4, in top-10 correction. This robustness helps alleviate concerns
which the “Investing in funds – A quarterly analysis” sec- regarding replication accuracy and its effects on reported
tion was published in the WSJ.10 The publication typi- results.
cally contains lists of top-10 mutual funds in 22 invest- One mutual fund characteristic not available on CRSP
ment style categories, each ranked based on previous 12- but required for some of our analysis is the Morningstar
month returns. The 12 major categories are {small cap, mid star rating of each fund, assigned for funds with at least
cap, large cap, multi cap}× {growth, core, value}, and are three years of return history. We obtain the historical star
included in the publication every quarter. The remaining ratings directly from Morningstar.
Table 1 reports summary statistics of mutual fund char-
10
This period is chosen due to data availability on mutual fund cate-
acteristics for a set of 111,780 fund observations on 5,334
gorization used by the WSJ, which is crucial to correctly replicating the unique funds over the sample period. Results for the full
WSJ rankings. While the rankings were published in the WSJ starting in
1994, CRSP does not report Lipper categories before 20 0 0. Furthermore,
11
Lipper changed their categorization scheme starting late 1999, after being Over 200 Lipper sectors exist; 39 sectors appear at least once in the
acquired by Thomson Reuters in 1998. Using stale categories is therefore rankings.
12
impossible. Lipper is unable to provide a data set of the categorization of Such that a typical CRSP-based top-10 list will contain approximately
funds prior to 20 0 0. nine of the ten funds mentioned in the WSJ.
Table 1
Summary statistics by rank cross-sections.
This table reports summary statistics for a set of 111,780 fund observations on 5,334 unique funds over the period 20 0 0Q1–2012Q4 across several cross-
sections of the data by rank, which is the ranking of the fund based on the previous 12-month return at the end of each quarter. Also reported are the
p-values on the difference in mean between funds with rank = 10 and rank = 11. The characteristics reported are: TNA (total net assets held by the fund,
in millions), Fund age (years since fund inception), Yearly return (12-month return on which the ranking is based), Expense ratio (the percentage of fund
assets claimed as expenses every year), Stars (the Morningstar star ranking of the fund), Beta (the fund’s beta vs. the S&P500 portfolio), Next Q return (the
return the fund generated during the quarter following publication), and Next Q flow (the percentage net capital flow into the fund, during the quarter
following publication).
Full Rank = 1 Rank = 10 Rank = 11 Rank = 20 Rank = 50 p-val(11–10)
TNA Mean ($M) 910.256 550.926 761.336 919.244 1100.958 1273.192 (0.330)
SD 3678.845 1443.850 2251.147 3361.632 3934.336 5099.416
Fund age Mean (Y) 11.446 9.572 10.311 11.257 11.724 11.526 (0.133)
SD 12.115 12.073 10.747 11.473 12.163 11.609
Yearly return Mean (percent) 6.056 41.667 19.682 19.097 15.683 10.012 (0.707)
SD 24.641 50.886 27.637 27.407 25.434 22.860
Expense ratio Mean (percent) 1.227 1.560 1.244 1.229 1.209 1.183 (0.594)
SD 1.592 1.259 0.518 0.460 0.503 0.501
Stars Mean 3.047 3.717 3.636 3.635 3.544 3.168 (0.985)
SD 1.064 1.325 1.113 1.087 1.043 0.955
Beta Mean 1.015 1.035 0.997 0.992 0.996 1.017 (0.711)
SD 0.247 0.545 0.282 0.255 0.233 0.206
Next Q return Mean (percenta ) 1.188 2.170 1.273 1.221 1.407 1.279 (0.933)
SD 10.793 12.795 10.760 10.677 10.599 10.882
Next Q flow Mean (percenta ) 1.373 17.166 10.325 7.203 5.478 2.217 (0.027)∗∗
SD 17.343 33.937 28.809 20.487 20.688 16.506
a
Next quarter return and flow are in quarterly percentage terms.
∗∗
Significant at the 5% level.
panel and five rank cross-sections are presented, as well 20 during our sample period. As we are interested in test-
as the p-value on equality of mean for each characteris- ing the effect of media mentions subsequent to the WSJ
tic between rank = 10 and rank = 11. The only characteris- publication, we limit the search to start from the day fol-
tic for which there is a significant difference in mean be- lowing publication in the WSJ.
tween rank = 10 and rank = 11 is capital flows during the For a more in-depth analysis of capital flows during the
post-publication quarter. post-publication quarter, we use data on daily capital flows
One possible propagation mechanism we consider is di- purchased from TrimTabs. The TrimTabs data set relies on
rect mutual fund advertising. To that end, we obtain from voluntary disclosure by mutual funds, however, and there-
Kantar Media, an advertising consulting firm, a data set on fore has limited coverage of the funds in our CRSP/WSJ
mutual fund complex advertising activity, including the ad- sample. We observe daily flows for a subset ranging from
vertising expenditure of the complex in print media, and 5% of fund share classes at the beginning of the inspected
a PDF copy of each magazine advertisement.13 We man- period to approximately 20% towards the end of the pe-
ually tag approximately 6,800 ad images to extract useful riod. We use the TrimTabs data to analyze the duration
ad characteristics such as the exact funds mentioned in the and impact of the media effect during the post-publication
ad, the ad size, and whether a fund’s WSJ rank was men- quarter, and to test for a possible “announcement day” ef-
tioned in the advertisements. After removing ads which fect. While the exact publication date in the WSJ varies
were found to not be related to mutual funds, and ac- within the first week of the publication quarter, the flow
counting for the fact that each ad may be published multi- variables constructed from CRSP consider the entire quar-
ple times, we end up with 9,446 observations for 127 mu- ter. The TrimTabs daily flow data also help us verify our
tual fund complexes over the sample period. results are not driven by the fact that these few pre-
To facilitate an investigation of whether subsequent publication days are included in the CRSP flow calculations.
media coverage plays a propagation mechanism role, we
construct a data set containing the number of times each
4. Empirical strategy and results
mutual fund was mentioned in 89 major US news and
business publications.14 We use the Dow Jones Factiva
4.1. Media exposure causes increased investment
news collection and conduct an automated search for me-
dia mentions of almost 75,0 0 0 fund-months, searching by
The foundation of our empirical strategy is a com-
either fund ticker or name. We find over 13,0 0 0 articles
parison between capital flows into published and unpub-
mentioning 2,722 mutual funds which made it to the top-
lished mutual funds using a regression discontinuity de-
sign (RDD).15 A significant discontinuity in capital flows
13
A similar data set, albeit using a shorter time period, is used by
Phillips et al. (2016); they provide a useful Appendix describing the data
further. 15
Reuter and Zitzewitz (2014) use a methodology somewhat similar to
14
List of publications as defined by the Factiva category of the same ours to test, and mostly reject, the decreasing returns to scale hypothesis
name. Full list available from the authors upon request. of Berk and Green (2004). They exploit differences in mutual fund capital
during the post-publication quarter will indicate that me- in the estimation and observations with ranks 2 and 19
dia exposure has a causal effect on consumers’ investment have about 6% of the weight of observations with ranks 10
behavior. and 11.
Capital flows into fund i during quarter q are defined as Formal discontinuity tests are reported in Table 2. We
percent increase in the fund’s assets beyond asset appreci- perform five types of discontinuity tests. Our main test
ation: uses Eq. (2) to compute the discontinuity using a local lin-
T NAiq+1 − T NAiq (1 + Riq,q+1 ) ear kernel regression. The second test repeats this analysis
F lowiq,q+1 = , (1) but adds controls for fund size, age, expense ratio, Morn-
T NAiq ingstar rating, and fund return during the last quarter be-
in which T NAiq is the total net assets of fund i at the before publication, to control for any abnormal flows stem-
ming from recent high returns.19 The third test is a Z-test
ginning of quarter q, and Riq,q+1 is the return on the fund’s
for difference in mean capital flow between funds ranked
assets between the beginning of quarter q and the begin-
10 and funds ranked 11. The fourth test compares the mean
ning of quarter q + 1.16 All flows are winsorized at the 1%
capital flow on one side of the discontinuity, at rank = 10,
level to decrease the effect of outliers.
with an extrapolation of the data trend from the other side
The independent variable for the RDD is a fund’s rank
of the discontinuity (the data with rank > 10). The ex-
within its style category at the end of the 12-month rank-
trapolated value at rank = 10 using the data with rank >
ing period. Due to the discrete nature of the rank variable,
10 is obtained by computing an LLR regression using Eq.
the exact cutoff in the [10, 11] segment is an empirical de-
(2), in which cutoff = 10. In such a regression, α 0 will cap-
sign choice. The choice which minimizes extrapolation er-
ture the predicted value from the right (i.e., using the data
ror is to use cutoff = 10.5. To find the predicted value of
with rank > 10) at rank = 10. The fifth test is similar, but
capital flows at 10.5, we employ a local linear kernel re-
compares mean capital flow at rank = 11 with α0 + α1 , the
gression (LLR) around the cutoff, as advocated by the ex-
predicted value from the left (i.e., using data with rank
tant RDD literature.17
< 11) at rank = 11 of an LLR regression using Eq. (2), in
The basic LLR equation is:
which cutoff = 11.
FlowQrank,cat,q = α0 + α1 × D + β0 × (rank − cutoff ) + All five tests yield statistically significant discontinuity
β1 × D × (cutoff − rank ) + rank,cat,q , (2) in capital flows around the cutoff.20 Our main test indi-
cates a local average increase in fund capital flows of 2.2
in which D = 1 if rank < 10.5 and D = 0 otherwise, and percentage points, representing a 31% average increase in
FlowQrank, cat, q is percent capital flow during quarter q (the capital flows during the post-publication quarter, relative
quarter subsequent to publication) into the fund ranked to a predicted value of 7.1 percentage points at rank =
rank within category cat. The discontinuity is thus the 10. We find no indication of discontinuity in pre-ranking
value of α 1 . It is important to note that LLR are not es- returns, on which the ranking is based, or any signifi-
timated by using ordinary least squares (OLS) but by using cant discontinuity in mutual fund returns during the post-
weighted least squares (WLS), with the weights defined by publication quarter that may indicate the existence of scale
the kernel and bandwidth used. We determine the optimal diseconomies.21
bandwidth using the estimator proposed by Imbens and A graphical view of the main test is presented in
Kalyanaraman (2012), who derive a closed-form, fully data- Fig. 1(a). The plot in this figure is constructed using two
driven estimator for an optimal RDD bandwidth. We use a nonparametric kernel regressions, from the right and left,
triangular kernel, K (t ) = max{0, 1 − |t |}, shown by Cheng with the same kernel and bandwidth as in the tabulated
et al. (1997) to have optimality properties for boundary es- result. This procedure guarantees that the tabulated result
timation, as in the RDD case.18 Using WLS ensures obser- obtained using Eq. (2) and the discontinuity value derived
vations closer to the discontinuity (e.g., ranks 9–12) have from the figure at rank = 10.5 are precisely the same. It
a stronger effect on the resulting estimation than obser- is evident that capital flows into mutual funds during the
vations further away (e.g., ranks 4–5 and 16–17). Specifi- post-publication quarter exhibit a clear discontinuous in-
cally, for next-quarter flow estimation in our base setting, crease.22
the optimal bandwidth is estimated to be 9, which means These results comprise evidence of a causal link from
the observations with ranks 1 and 20 are not included media attention to consumer investment behavior. Unlike
previous work trying to ascertain such a causal link, these
flows between funds with different Morningstar ratings, which are close
to the discrete ratings cutoff points, as a source of exogenous variation in
19
fund size. The precise method of using controls in LLR is described by
16 Eq. (4) below.
Our results are nearly identical when using the alternative measure,
F lowiq,q+1 = (T NAiq+1 − T NAiq (1 + Riq,q+1 ))/(T NAiq (1 + Riq,q+1 )). The alterna- 20
We verify they remain significant when using heteroskedasticity ro-
tive measure assumes new capital flows take place at the beginning of bust standard errors (s.e.), jackknifed s.e., and s.e. clustered by quarter.
the quarter whereas the main definition assumes new capital flows take 21
This is in line with the findings of Reuter and Zitzewitz (2014), who
place at the end of the quarter. focus on testing the existence of scale diseconomies but find little evi-
17
E.g., Hahn et al. (2001), Imbens and Lemieux (2008). We also consider dence that fund size erodes returns.
local quadratic regressions and high order global polynomials from the 22
While for a convex function an RDD methodology could result in dif-
left and right. Our conclusions hold. ferent slopes on the two sides of the discontinuity, due to the fact that
18
In Appendix Fig. A.3, we show our discontinuity result is robust to observations on one side of the cutoff are essentially ignored when com-
the choice of bandwidth. In unreported results, we also verify robustness puting the slope on the other side, the two slopes in the figure are sta-
to different choices of kernel. tistically indistinguishable.
Table 2
Tests for discontinuity in flows and returns around the rank = 10 cutoff.
This table reports the results of five discontinuity tests performed with dependent variables: Next Q flow (percentage net capital flow into published
fund-class during post-publication quarter), Prev Y return (previous 12-month return on which ranking is based), Next Q return (return fund generated
during quarter after publication), Next Q flow entire fund (flow into all share classes of published fund), and Next Q flow complex spillover (flow into all
funds of a fund complex except published one). The first test computes a locally weighted linear regression of the dependent variable on fund rank with an
indicator for rank < 10.5 and reports the discontinuity (coefficient on the indicator). The second test repeats this analysis but adds controls for fund size,
age, expense ratio, last quarter return, and Morningstar rank. The third test is a Z-test for difference in means between funds ranked 10 and funds ranked
11. The last two tests compare actual mean at rank = 10 with regression predicted value from the right and actual mean at rank = 11 with regression
predicted value from the left, using Z-tests for difference in means. Regression predicted values at 10 and 11 are calculated using locally weighted linear
regressions of the dependent variable on fund rank, with the data limited to rank > 10.5 and rank < 10.5, respectively. Values in (brackets) are p-values on
the coefficient being zero (no discontinuity). For complex spillovers, if a complex appears more than once in the rankings in the same quarter, only one of
the occurrences is kept, at random, to avoid biasing the standard errors. The last column reports N, the number of fund-quarter observations participating
in the test.
Next Q Prev Y Next Q Next Q flow Next Q flow N

flow return return (entire fund) (complex spillover)
Discontinuity at 10.5 2.203∗∗ −0.600 0.108 1.862∗ 1.820∗∗ 11,232

(0.025) (0.677) (0.414) (0.053) (0.017)
Discontinuity w/ controls 2.377∗∗ 0.850 0.090 1.549∗ 1.623∗∗ 7,880
(0.020) (0.259) (0.438) (0.092) (0.049)
Actual 10 vs. actual 11 3.122∗∗ 0.585 0.051 3.127∗∗ 2.694∗∗ 1,248
(0.027) (0.707) (0.933) (0.022) (0.032)
Fitted vs. actual at 10 3.223∗∗ 0.473 0.033 2.792∗∗ 2.706∗∗ 6,240
(0.011) (0.371) (0.476) (0.027) (0.021)
Fitted vs. actual at 11 2.103∗ −0.487 0.126 2.198∗∗ 1.809∗∗ 6,240
(0.066) (0.613) (0.424) (0.046) (0.042)
∗∗
∗
results do not depend on population splits in which differ- ing in funds – A quarterly analysis” was not published.
ent subgroups are exposed to different media. The fact that These ranking tables were published less prominently, in
the WSJ ranking tables are published every quarter and the back pages of the regular section of the newspaper.
their content is “algorithmically” (but quasi-randomly) de- This feature allows us to test whether the prominence of
termined means we require no additional controls for con- the publication and the media visibility it garners are the
sumer characteristics or controls capturing the decision of key drivers of the increased flows, by testing whether de-
the media outlet to publish specific articles. livering a similar, but low prominence, ranking table still
causes changes to investment behavior.
4.2. Information vs. prominence channels Panel A of Table 3 presents the results of this analy-
sis. We replicate the published within-quarter ranking ta-
The results of Section 4.1 show that appearing in the bles and analyze flows into funds during the three months
media causes increased flows into those mutual funds post-publication. In the first row of Panel A, we find
which were mentioned, but these results do not delineate no significant increase in flows following the publication
the role of different possible channels in causing changes of the within-quarter months’ rankings. A distinction be-
in investor behavior. We turn next to evidence that high- tween the Category Kings and the within-quarter rankings
lights the central role of visibility and prominence in giving is that the within-quarter rankings are based on year-to-
rise to this effect, distinct from an informational channel. date rather than 12-month returns. The rankings published
We begin by noting that all the information provided during December are the most closely correlated to the
in the publication was publicly available prior to publica- rankings published during the subsequent January in the
tion. As mutual fund returns and their classifications into WSJ special issue. We also separately analyze the rank-
categories are widely available, any interested party could ing tables published every December and based on Jan-
create the ranking tables in advance of publication. Hence, uary to November returns, in the second row of the panel.
no new information is being provided by the publication. Here too, we find no evidence of discontinuity in capi-
The information channel proposed by Del Guercio and Tkac tal flows into mentioned mutual funds during the three
(2008) remains, however, a possible candidate channel. As months post-publication. A graphical view of the within-
search is costly, the role of the ranking tables in summa- quarter-based results is presented in Fig. 1(b) and 1(c), and
rizing information into discrete ratings and decreasing the it is evident no discontinuity exists for both these settings,
information acquisition cost for investors is likely to be an in stark contrast to the discontinuity in Fig. 1(a).
important channel. However, we show that the pure media A second feature of the empirical setup useful in test-
visibility channel we identify is distinct from the informa- ing the prominence hypothesis is the fact that in addition
tion channel. to publishing the “Category Kings” tables, the WSJ made
A useful feature of our empirical setting is the fact that all rankings (and not just the top-10) readily available on
similar ranking tables were also published by the WSJ for its website starting in 2007. Importantly, while one could
within-quarter months (months that do not follow the end get to these rankings with a few clicks of the mouse,
of a calendar quarter), in which the special issue “Invest- they were not in the highly visible independent section
Fig. 1. RDD analysis of post-publication fund flows by rank. Each dot represents the mean net percentage capital flow into the funds at the given rank
during the quarter after publication (Next Q flow). The figures include local linear kernel regression lines and 95% confidence intervals for the segments
[1,10] and [11,50]. For Panel (a), funds are ranked based on the most recent 12-month return within their investment category at the beginning of every
quarter from 20 0 0Q1 to 2012Q4 using data from CRSP, and each dot is the mean of 624 data points per rank, corresponding to 12 categories over 52
quarters. For Panel (b), funds are ranked based on year-to-date return within their investment category at the beginning of every within-quarter month
from 20 0 0M2 to 2012M12, and each dot is the mean of 1,248 data points (two within-quarter months for each of 12 categories over 52 quarters). Panel
(c) concentrates only on rankings based on January to November returns, which were published during December in the WSJ, and each dot is the mean of
156 data points. The top-10 funds are replaced with the actual funds published in the Wall Street Journal.
“Investing in funds - A quarterly analysis,” under the “Category Kings” rankings, as the dependent variable. We
heading “Category Kings.” Hence, this change affects in- find a significant discontinuity in aggregate capital flows
formation availability (e.g., WSJ rankings now have similar into all share classes of the published funds. This is to be
availability to that of Morningstar rankings), though not expected, as different fund share classes typically only dif-
the information visibility (funds ranked 1–10 are still fer in their fee structure, but acts as further evidence for
much more visible due to the publication of the special the robustness of the impact of media visibility.
issue). We cannot reject the hypothesis that the disconti- The fifth column of Table 2 uses flows into all other
nuities pre- and post-2007 are the same (p=0.762), with funds within the same complex as the fund-class which
the post-2007 discontinuity being less than 10% (0.19 per- appeared in the WSJ rankings, excluding the published fund.
centage points) smaller than the pre-2007 one. This result Strikingly, we find a significant discontinuity of 1.8 per-
shows the effect was not weakened by better availability centage points in the aggregate capital flows into all other
of information regarding mutual fund rankings over time, funds of the same fund complex during the quarter after
similar to the results of Phillips et al. (2016). publication of the prominent rankings tables.23 To put our
The analysis so far was conducted at the fund-class
level, as this is the unit of observation used in the WSJ
23
ranking tables. The fourth column of Table 2 reports Aggregate capital flows to the rest of the complex are computed by
a weighted average of the percent flows into each fund class, with the
the results of the five discontinuity tests described in
total net assets of each class used as weight. When more than one fund
Section 4.1 but using capital flows into all share classes of of the same complex is ranked in the top-20, only one of the occurrences
the fund-class which appeared in the prominent quarterly is kept, at random, to avoid attenuating the standard errors.
Table 3
Tests for discontinuity in flows and returns in no-attention settings.
Panels A and B repeat the main discontinuity test of Table 2 over the same five dependent variables for settings which lack media attention: “All
YTD” analyzes flows and returns for rankings used in the monthly tables published in the WSJ during within-quarter months based on year-to-date
returns; “December YTD” uses January to November returns to replicate the rankings used in the monthly tables published in the WSJ during December;
“Unpublished categories” uses category-quarters not published on the WSJ; “Within-quarter” uses rankings based on 12-month returns ending in months
which do not follow the end of a calendar quarter (and so are not published); “11-Month ranking” constructs the fund rankings at the end of a quarter
based on the most recent 11- (rather than 12-) month return. Rankings in Panel A were actually published, though with low visibility, while rankings in
Panel B are counterfactual and were not published. Complex spillover correction, p-values and N are as in Table 2. Panel C reports the results of conducting
the main discontinuity test for Next Q flow in the original setting but at various possible cutoffs. For each cutoff X, we report the intercept discontinuity
based on a locally weighted linear regression at rank = (X + 0.5 ), along with its p-value.
Panel A: Published YTD settings
Next Q Prev Y Next Q Next Q flow Next Q flow N

Flow return return (entire fund) (complex spillover)
All YTD 0.235 −0.251 −0.041 −0.464 0.139 22,464

(0.388) (0.717) (0.547) (0.722) (0.614)
December YTD 0.639 −0.422 −0.079 −0.630 0.192 2,808
(0.661) (0.840) (0.698) (0.844) (0.828)
Panel B: Counterfactual settings
Unpublished categories 0.043 −0.246 −0.182 0.162 −0.019 24,612

(0.902) (0.676) (0.617) (0.385) (0.659)
11-Month ranking 0.154 −0.539 0.014 −0.752 0.011 11,232
(0.446) (0.676) (0.488) (0.661) (0.441)
Within-quarter 0.517 −0.639 −0.025 −0.540 0.217 22,464
(0.252) (0.769) (0.529) (0.670) (0.619)
Panel C: Different cutoffs
6 7 8 9 10 11 12 13 14 15 20 50
Discontinuity 1.360 0.039 0.362 0.175 2.203∗∗ 0.791 −1.435 −0.789 −1.169 0.359 −0.332 0.275
p-value 0.148 0.487 0.380 0.440 0.025 0.236 0.141 0.360 0.184 0.466 0.712 0.497
∗∗
finding in perspective, we note that focusing on monthly nuity reported in Section 4.1 is a unique feature of the
flows, Nanda et al. (2004) find a 4.4% (on an annual basis) data, driven by the publication and ensuing media atten-
increase in flows into complexes of “star” funds. We find a tion. The first setting simply repeats the analysis in the
7.3% increase (on an annual basis) in flows to the rest of first row of Table 2, but reports the discontinuity in capi-
the complex. This significant spillover effect is consistent tal flows when considering fund categories which were not
with an impact of media attention on brand-name recogni- published in that quarter’s WSJ issue. The second setting
tion at the complex level, and is less consistent with an in- uses the published categories, but constructs their rankings
formation channel, as we explicitly exclude the fund which based on the most recent 11- (rather than 12-) month re-
was published from the flow calculation.24 turns, ending at the end of a calendar quarter, and like our
The analysis above indicates that media exposure (vis- baseline test, considers flows during the subsequent quar-
ibility) indeed has a role distinct from that of informa- ter. The third setting combines the publication schedule
tion dissemination in creating increased capital flows. We of the within-quarter tables described in Section 4.2 with
note that this does not mean well-summarized information the ranking method of the quarterly “Category Kings,” de-
is unimportant to generating flows. The WSJ rankings do scribed in Section 4.1. Funds are ranked based on previ-
help investors by summarizing how funds’ 12-month re- ous 12-month returns, ending at the within-quarter month
turns stack up relative to the category into a simple dis- (e.g., February 2001 to January 2002 or March 2005 to
crete rank. But it is the prominence of the “Category Kings” February 2006), and we analyze flows in the subsequent
discrete ranking lists, within the quarterly “Investing in three months (e.g., February to April 2002 or March to
funds” WSJ special issue, that is key to driving increased July 2006). We find no significant discontinuity in any of
capital flows from investors to the published mutual funds these counterfactual settings. A graphical view is available
and their complexes. in Fig. 2, and it is evident capital flows change smoothly
around the cutoff, in contrast with Fig. 1(a). Pre-ranking
4.3. Discontinuity robustness 12-month returns, the variable driving the ranking, is also
smooth as expected and exhibits no discontinuity around
Panel B of Table 3 reports the results of three coun- the cutoff (Appendix Fig. A.2).
terfactual settings, which aim to verify that the disconti- Panel C of Table 3 reports the results of a falsifi-
cation test for discontinuity in capital flows using the
original setting of Table 2, but at cutoffs other than rank =
24
For a discussion of information spillover between related products,
10.5. We expect to find no discontinuity at other cut-
see Hendricks and Sorensen (2009). For a discussion of spillovers in the
context of mutual fund management companies, see Sialm and Tham offs, and the results in Panel C confirm that the only
(2016). statistically significant discontinuity is at rank = 10.5. In
Fig. 2. RDD analysis of counterfactual settings. Each dot represents the mean net percentage capital flow into the funds at the given rank during the
quarter after ranking (Next Q flow). The figures include local linear kernel regression lines and 95% confidence intervals for the segments [1,10] and [11,50].
For Panel (a), funds are ranked based on previous 12-month returns within their investment category at the beginning of every quarter from 20 0 0Q1 to
2012Q4 using data from CRSP, but only categories which were not published in the WSJ that quarter are included. The number of data points per rank in
Panel (a) varies from 1,882 observations for rank = 1 and down to 502 observations for rank = 50. For Panel (b), funds are ranked based on the most recent
11- (rather than 12-) month return within their investment category, and each dot is the mean of 624 data points per rank. For Panel (c), funds are ranked
based on the most recent 12-month returns within their investment category at the beginning of every within-quarter month from 20 0 0M2 to 2012M12
(in which the rankings actually published were based on year-to-date returns), and each dot is the mean of 1,248 data points.
Appendix Section A.1 we also verify that rankings are suffi- flows. Rising from rank = 11 to rank = 10 is, however, cor-
ciently fluid around the cutoff to satisfy the quasi-random related with a 2.49 percentage points increase in flows.25
assumption, that pre-ranking fund characteristics exhibit This incentive scheme is related to, but distinct from, the
no discontinuity, and that the discontinuity is robust to dif- one created by the flow-return relationship, discussed by
ferent choices of kernel and bandwidth. Brown et al. (1996), Chevalier and Ellison (1997), and Sirri
and Tufano (1998). Finding such a managerial response to
4.4. Funds’ response pre-publication the media effect can also further increase our confidence
in the validity of the discontinuity results reported in the
Next, we hypothesize that the media effect leading to previous section.
increased capital flows into published mutual funds should Theory predicts that managers of mutual funds around
affect optimal risk-shifting behavior of mutual fund man- the publication cutoff will respond to this incentive
agers pre-publication. Discontinuity in capital flows im- by increasing the fund’s tracking error volatility rela-
plies that, for funds around the rank = 10 cutoff, there is tive to its respective category; see, for example, Basak
a greater upside to increased rank than a downside to de- et al. (2007) and Cuoco and Kaniel (2011). Tracking error
creased rank. For example, a fund ranked 11 a month be-
fore the end of a ranking period which drops from rank =
11 to rank = 20 by the end of the ranking period will, on 25
Of which 2.2 percentage points are due to publication, and 0.29 per-
average, see a 0.8 percentage points decrease in capital centage points are from the return-flow relationship.
volatility captures how much a fund’s investment policy Table 4

Changes in fund investment behavior during last month of a ranking
deviates from a baseline portfolio, by analyzing the volatil-
period.
ity of the difference in their daily returns. As measures Panel A reports changes, between month 11 and month 12 of a ranking
based on daily returns are inherently noisy, our first step is period, in: average tracking error volatility with respect to (w.r.t.) the cat-
to aggregate funds into portfolio groups consisting of four egory portfolio (TE); average tracking error volatility w.r.t. the S&P500
consecutive ranks within each category-quarter. We con- portfolio (TESP); average volatility of fund returns (VOL); average beta
w.r.t. the S&P500 portfolio (BETA). Funds are ranked based on their re-
centrate our analysis on a four-fund portfolio’s tracking er-
turn during the first 11 months of a ranking period, and then grouped
ror volatility relative to an equal-weighted portfolio of all into portfolios of four funds by rank (e.g., funds ranked 9 to 12 within
mutual funds in the same category, though we also con- a given category are formed into a portfolio). The tracking errors of the
sider S&P500 as an alternative baseline portfolio. portfolio during the pre-formation month and the post-formation month
(months 11 and 12, respectively) are calculated as the standard deviations
We denote the return on day t of an equal-weighted
[i,i+3] of the difference between the daily return of the four-fund portfolio and
portfolio of funds ranked i, i + 1, i + 2, and i + 3 by rt , that of the respective baseline portfolio (fund category or S&P500). The
[9,12]
for example, rt . The tracking error volatility (TE) for a mean difference is reported, along with the respective p-value on differ-
four-fund portfolio during month m relative to the baseline ence from zero. Reported tracking errors are in daily percentage terms,
and each value is calculated using 99,340 fund-day observations. Panel B
category portfolio cat is defined as:
reports the results of repeating the tracking error volatility analysis sep-
arately for the top half and bottom half of funds based on their return
T Ei,cat,m = StDev rt[i,i+3] − rtcat | t ∈ m , (3) in the first quarter of a ranking period (months 1 − 3). Panel C repeats
the analysis of changes in average tracking error volatility w.r.t. the cate-
in which StDev represents the standard deviation operator. gory portfolio (TE), but forms portfolios based on 10-month returns and
A significant increase in the TE measure will indicate that compares tracking errors during the two months prior to portfolio forma-
tion (months 9,10) to tracking errors during the post-formation months
managers are attempting to increase ranking volatility, and (months 11,12).
the expected payoff of the call-like option they hold, by
“deviating from the herd.” Panel A: Full sample
To uncover such a possible increase, we first rank [1,4] [5,8] [9,12] [13,16] [17,20]
funds within each category based on the first 11 months TE 0.035 −0.716 1.777∗∗ 0.194 0.028
of a ranking period (e.g., returns from January 20 0 0 to (0.985) (0.560) (0.031) (0.773) (0.965)
November 20 0 0), and calculate the four-fund TE measures TESP −1.109 −2.064 0.237 0.137 −0.910
during month 11 (pre-formation) and during month 12 (0.572) (0.140) (0.822) (0.877) (0.272)
VOL −1.182 −1.455 −1.251 −0.512 −0.714
(post-formation). We demean the TE measure relative
(0.657) (0.418) (0.360) (0.653) (0.504)
to the average TE of all four-fund portfolios within that BETA 0.073 −0.024 −0.108 −0.065 0.087
category-month, to remove the effects of market- or (0.898) (0.933) (0.756) (0.806) (0.753)
portfolio-wide volatility within that month: Panel B: Top vs. bottom Q1 performers
TE i,cat,m = T Ei,cat,m − Mean j (T E j,cat,m ), TE - Top 0.049 −0.087 3.016 ∗∗∗ 2.571∗∗ 0.645
(0.979) (0.953) (0.005) (0.025) (0.536)
with Meanj denoting the average across all four-fund TE - Bottom 3.956 −0.283 0.208 −0.149 −0.724
groups. We finally calculate the average difference per (0.269) (0.891) (0.895) (0.904) (0.546)
four-fund group [i, i + 3], between the T E during month Panel C: Two-month change
12 and during month 11, across all categories and ranking TE 0.131 0.381 1.219∗ 0.716 −0.682
periods. (0.852) (0.701) (0.089) (0.221) (0.691)
The first row of Panel A in Table 4 lists these aver- ∗∗∗
age differences per four-fund group. As predicted, there is ∗∗
a large and statistically significant increase in tracking er- ∗
ror relative to the category portfolio for funds close to the
cutoff. We also report changes in: tracking error volatil-
ity relative to the S&P500 Index (TESP); the volatility of to weigh the benefit of potentially rising in the rank and
funds’ daily returns (VOL); and the funds’ beta relative to making the list this quarter versus the distortions induced
the S&P500 Index (BETA), though theory makes no predic- on fund allocations. Funds close to the ranking list cut-
tions regarding these metrics. We observe no statistically off that have their best quarter about to “expire” have
significant increase in either, as is evident from the second, the highest incentive to engage in such risk-shifting, as
third, and fourth rows of Panel A, respectively. they are unlikely to be in a position to make the list next
An interesting feature of our empirical setting is the re- quarter.26
peated quarterly publication based on previous 12-month To test this hypothesis, Panel B splits the sample into
returns. This feature implies that a fund which enjoys ex- funds above and below the median quarterly return for
ceptionally high returns in a given quarter benefits from the first quarter of every 12-month ranking period, and
these exceptional returns, in terms of the likelihood of be- repeats the analysis of Panel A independently for each
ing published in the WSJ, during the subsequent four top- sample split. Strikingly, only funds in the top first quarter
10 ranking publications, before that quarter is no longer
accounted for. As discussed by Phillips et al. (2016), in- 26
E.g., a fund with an exceptional Q1 in 2001 may get published in
vestors do not differentiate between new and stale in- the WSJ in 4/20 01, 7/20 01, 10/20 01, 1/20 02. The publication occurring on
formation components of fund performance. A fund man- 1/2002 is the last one for which the fund returns during Q1 2001 are still
ager considering engaging in risk-shifting behavior needs relevant, and we say the quarter expires after that publication.
return group exhibit TE increase, at rank groups [9, 12] and funds post-2001. Given that our sample starts in 20 0 0, we
[13, 16].27 The stark differences between the top and bot- lack sufficient data to distinguish any changes in increased
tom groups in Panel B strongly support our hypothesis re- leaning for the tape behavior for funds close to the cutoff.
garding strategic risk-shifting behavior.
To verify these results are not driven by a systematic 4.5. Media effect propagation post-publication
difference between the funds in terms of Morningstar rat-
ings, we also attempt to split the sample into high-rated In Section 4.4 we provided evidence that the presence
funds (5 or 4 stars, 5,116 observations) and low-rated funds of the media effect impacts funds’ trading strategies prior
(3 or less stars, 3,754 observation). In unreported results, to publication. We now consider the way in which the me-
we find no significant difference in risk-taking behavior be- dia effect propagates post-publication and causes the in-
tween these groups. This also holds when considering only creased flows. This is done both to provide more texture to
funds with 5 Morningstar stars to be high-rated, or when the results so far, and to ascertain that observed propaga-
considering only the subsample of high Q1 performers and tion patterns are consistent with the visibility and promi-
splitting it based on Morningstar rankings. nence findings.
While it may seem that changing tracking error volatil-
ity over one month will do little to help managers in- 4.5.1. Timing of flow increase
crease their returns as measured over 12 months and as- We start by examining the duration of the media ef-
sist them in entering the rankings, it is important to note fect within the post-publication quarter to shed light on
that rankings around the cutoff are extremely fluid. Even consumer behavior in response to the media stimuli, and
minute changes in returns can affect the probability of a whether consumers have an immediate or protracted re-
fund being published, as can be seen in Appendix Fig. A.1. sponse to the media stimuli. Fig. 3(a) begins this analy-
For example, a fund which started the last day of a 12- sis by presenting differences between mean percent cumu-
month ranking period with rank = 10 (rank = 9) has more lative capital flows into funds ranked 10 and 11, for the
than 25% (10%) chance of ending that day (and the entire year following publication in the WSJ, in monthly intervals.
12-month ranking period) with rank > 10, thus not being We can see that, ex ante, the media effect is expected to
published. A fund which started that day with rank = 11 last up to six months post-publication. Fig. 3(b) presents
(rank = 12) has about 20% (10%) chance of ending that the difference in flows after removing funds which were
day within the top-10 and being published. Our empirical published in the following quarter’s Category Kings lists.
strategy benefits greatly from this fact, as it ensures strong The effect diminishes and is no longer significant after
local randomization around the cutoff–a necessary feature the third month, suggesting the effect in months 4–6 may
for a strongly valid RDD. This fluidity of ranking is also the be driven by funds being republished in the WSJ in sub-
reason why tracking error changes during the last month sequent quarters. Fig. 3(c) repeats the analysis of differ-
help the managers increase the likeliness of moving their ence in flows, in daily intervals for the first 60 trading
ranking measured over 12 months. For robustness, Panel days post-publication, using the TrimTabs daily flow data.
C of Table 4 presents tests of TE changes during the last The TrimTabs data are missing flow data for five trading
two months of a ranking period, in which funds are ranked days each month, on average, making it impossible to pre-
based on the first ten months of a ranking period, and we cisely calculate cumulative flows throughout the quarter.
calculate four-fund TE measures during months 9–10 (pre- We overcome the missing observations limitation by calcu-
formation) and during months 11–12 (post-formation). We lating median daily flows for post-ranking days, and then
again find significant changes only for the [9, 12] group. Fi- cumulating these daily medians. Standard errors on the cu-
nally, we find that funds about to lose a strong first quar- mulated medians are obtained using the bootstrap with
ter, which are those shown to engage in risk-shifting the 1,0 0 0 repetitions.
most, are also 2.3% more likely (p=0.091) to be published The evidence in Fig. 3 indicates a fairly smooth in-
conditional on being within the [9, 12] rank a month prior crease in capital flows throughout the post-publication
to publication, further corroborating the hypothesis that quarter. We find no evidence of an immediate flow re-
risk-shifting when close to the publication cutoff is useful sponse following publication (no “announcement day ef-
in getting a fund published. fect”), or concentration of increased flows at the early part
A second possible managerial response for funds close of the quarter. These results indicate protracted propaga-
to the cutoff is to lean for the tape, artificially increas- tion of the media effect, in which the WSJ publication ei-
ing fund value by inflating the value of stocks they al- ther changes investor perceptions and attitudes towards
ready hold using strategically placed buy orders on the the published funds and their complexes, later leading to
very last day of the quarter. However, Duong and Meschke investment, or is a first step in a chain, followed by other
(2015) show that following circulation of early versions investor stimuli leading to purchase. We discuss evidence
of Carhart et al. (2002), that uncovered the practice, the supporting both these channels below.
Securities and Exchange Commission enacted enforcement
actions against mutual funds participating in this prac- 4.5.2. Less visible funds
tice, and consequently, it has disappeared among mutual Next, we examine how heterogeneity in mutual funds’
sizes (in terms of assets under management) and ages
(since fund inception), as well as the size of the fund
27
The TE increase for each of these top-rank groups is also statistically complex to which the funds belong affect increased flows
different from the corresponding bottom groups. into published vs. unpublished funds. We hypothesize that
1
0.8
0.8
p-value on diff
p-value on diff
0.6
0.4 0.6
0.4
0.2
0.2
0
0
(a) Full sample (b) Excluding repeatedly published
1
1
Diff in cumulative median flow 10 to 11 (%)
0.8
0.8
p -value on diff
0.6
0.6
0.4
0.4
0.2
0.2
0
0
(c) Daily
Fig. 3. Time trend of difference in capital flows into mutual funds ranked 10 and 11. Using the CRSP data on monthly flows, Panel (a) presents the
difference between mean percent cumulative capital flows into funds ranked 10 and 11, during the 12 months following publication in the WSJ, along
with its p-value. Panel (b) repeats this analysis, but omits funds which were published in the WSJ ranking lists in the next quarterly publication (i.e., in
month 4). Panel (c) concentrates on the first 60 trading days following exact publication day using daily flow data obtained from TrimTabs. The coverage
of the TrimTabs data ranges from 5% of the funds in the early years of the sample to approximately 20% towards the end of the inspected period. Panel (c)
therefore reports the difference, between funds ranked 10 and 11, in cumulative median daily flows. We obtain p-values on the differences in cumulated
medians using the bootstrap with 1,0 0 0 repetitions.
younger and smaller funds, from smaller fund complexes, between the effect of a characteristic (e.g., fund size) on
will have a higher “bang for the buck” from media visibil- capital flows into unpublished vs. published funds.
ity, as they are less visible ex ante. We consider each characteristic independently, such
To quantify the impact of heterogeneous fund character-
that Controls rank,cat,q is a scalar, and then include all char-
istics on the increase in capital flowing into the fund, we
acteristics simultaneously, such that Controls rank,cat,qis a
add controls to the local linear kernel regression described
vector. For ease of interpretation, all characteristics are
by Eq. (2). The controlled LLR is described by:
standardized, such that the 1 coefficients capture the ef-
FlowQrank,cat,q = α0 + α1 × D + β0 × (rank − cutoff ) + fect of a one standard deviation change in a characteris-
tic. Panel A of Table 5 reports the results of the indepen-
β1 × D × (cutoff − rank ) +
dent tests for each characteristic, and indicates that a one

0 × Controlsrank,cat,q + standard deviation decrease in the size of a published fund
yields a 1.32 percentage points higher increase in capital
1 × D × Controlsrank,cat,q + rank,cat,q , (4)
flows into the published fund than a similar decrease in
with cutoff = 10.5, D = 1 if rank < 10.5 and D = 0 other- the size of an unpublished fund. A similar pattern holds
a vector of demeaned fund characteris-
wise, and Controls for younger funds as well. Panel B reports the results of
tics. The mean is calculated as a weighted average and the the test in which all three characteristics are included si-
weights are given by the bandwidth and kernel used in the multaneously, along with the interaction of fund size and
weighted regression, as described for Eq. (2). This demean- complex size, and while the coefficient for fund size is no
ing procedure is required to prevent the controls from hav- longer significant, the results in this panel indicate that
ing any intercept in the regression and skewing the results smaller fund complexes also enjoy a higher “bang for the
of α 0 and α 1 . The differential impact of the jth control buck” from being published. We conclude that media at-
variable is measured by 1, j , and captures the difference tention amplifies the inverse relation between fund size
Table 5 bility channel in which the WSJ publication increases con-

Determinants of increased capital flows.
sumer brand awareness to the published funds and their
Panel A of this table reports the results of three independent locally
weighted linear regressions of Next Q flow (percentage net capital flow complexes.
into published fund class during post-publication quarter) on fund rank
with an indicator for rank < 10.5, while controlling for: Fund size (log to- 4.5.3. Propagation via advertising and subsequent news
tal net assets of fund class), Fund age (years since fund inception), and mentions
Complex size (log total net assets of all funds within the complex), along
Finally, we investigate the impact of fund advertising
with the corresponding (p-value) and N, the number of fund-quarter ob-
servations participating in the test. α 1 is the flow discontinuity (coeffi- post-publication, as a potential propagation mechanism of
cient on the indicator) and 1 is the differential impact of the control the media effect from publication to altering a financial
variable around the discontinuity (coefficient on the interaction of the in- consumer’s behavior.29 We then proceed by considering
dicator and control variable). All control variables are standardized. Panel
several other possible propagation mechanisms. First, in-
B repeats this test but controls for all control variables simultaneously
within a single regression, and adds the interaction of Size and Complex
creased effectiveness of fund advertising efforts, for exam-
size to the set of controls. Panel C reports the results of four indepen- ple, by being able to state the ranking on the WSJ list
dent locally weighted linear regressions of Next Q flow complex spillover in ads.30 Second, increased news coverage subsequent to
(percentage net capital flow into all funds of a published fund complex appearance in a Category King list. And third, increased
except the published one during post-publication quarter) on fund rank
efficacy of subsequent news coverage in generating fund
with an indicator for rank < 10.5, while controlling for Fund size and age,
Complex size, and the number of funds in the complex. If a complex ap- flows. We find evidence supporting most of these.
pears more than once in the rankings in the same quarter, only one of To investigate fund advertising, we utilize a data set
the occurrences is kept, at random, to avoid biasing the standard errors. provided by Kantar Media that examines advertising be-
Panel A: Next Q flow–independently controlled havior by mutual fund complexes using data on more than
6,600 published mutual fund ads. The data set includes
Fund size Fund age Complex size
ad size and expenditure, among other features. We ex-
Discontinuity at 10.5 (α 1 ) 2.132 ∗∗
1.969 ∗∗
2.270∗∗ tended the data by manually extracting features such as
(0.027) (0.038) (0.022) the names and tickers of mutual funds mentioned in the
1 −1.324∗∗ −1.134∗∗ −0.418 ad, or the fact that an ad mentions the ranking of a fund
(0.036) (0.022) (0.282)
N 11,016 11,016 10,999
based on the WSJ ranking scheme, from the ad images.
To analyze the frequency of post-publication news cover-
Panel B: Next Q flow–simultaneously controlled age (and its efficacy), we compiled a data set by execut-
ing more than 75,0 0 0 Factiva searches, each counting the
α1 1 [Size] 1 [Age] 1 [CplxSize] 1 [CplxSize∗ Size] N
number of times a fund or its ticker were mentioned in
2.610∗∗ −0.973 −0.810∗ −1.491∗ 0.509∗ 10,999 articles published in 89 major US news and business pub-
(0.014) (0.133) (0.061) (0.060) (0.052) lications during the quarters before and after publication.
Panel A of Table 6 reports results using our main dis-
Panel C: Next Q flow complex spillover–independently controlled
continuity test, for a specification similar to the one de-
Fund size Fund age Complex size Funds in complex scribed by Eq. (2), but with different dependent variables.
The first three columns of Panel A describe discontinuity
α1 1.731 ∗∗
1.683∗∗
1.725 ∗∗
1.745∗∗
(0.035) (0.039) (0.035) (0.033) results of indicators for increase in: Ad size–the average
1 −0.546 −0.374 −1.530∗∗∗ −1.835∗∗∗ size in square inches of all ads published by the mutual
(0.221) (0.347) (0.001) (0.002) fund complex; Amount spent–the dollar amount the fund
N 5,550 5,550 5,550 5,550 complex spent on advertising during the given quarter;
∗∗∗
Significant at the 1% level. Ads published–the number of ads published by the fund
∗∗
Significant at the 5% level. complex. We observe a significant discontinuity for each
∗
of these indicators, showing that appearing on the top-10
publication is associated with a 27% increase in the proba-
bility that a fund complex will increase the average size of
/ age / complex size and capital flows, as predicted by a
ads it publishes following publication, compared with pre-
visibility channel and the lower ex ante visibility of such
publication size. Similar increases in Amount spent and
funds.28
Ads published strongly indicate an increase in advertising
As discussed in Section 4.2, media attention also causes
activity by a fund complex, following one of its funds ap-
an increase in complex spillover flows (capital flows into
pearing in a “Category Kings” ranking.
all other funds in the same complex of the published fund,
Panel B of Table 6 reports results of affecting charac-
excluding the one published). In Panel C of Table 5 we con-
teristic tests, using a specification similar to the one de-
duct an analysis similar to that of Panel A, but using the
scribed by Eq. (4) in the previous section, but in which
complex spillover flows as the dependent variable. Here,
the dependent variable is capital flows into published mu-
too, we find that ex ante less visible complexes–smaller
tual funds, while controlling for the fund advertising activ-
complexes in terms of assets under management or num-
ity indicators described above. For the first three columns,
ber of funds managed–enjoy a higher “bang for the buck”
from being published. These facts are in line with a visi-
29
For a discussion of mutual fund advertising behavior and effects on
subsequent flows, see Jain and Wu (20 0 0).
28 30
As expected, the magnitudes of the discontinuity in both panels, mea- It is possible that existing, non-advertising, marketing efforts are
sured by the α 1 coefficients, are similar to the one reported in Table 2. more intense or more effective as well.
Table 6
Subsequent advertising and media publications.
Using data on actual fund advertising activity and the number of times each fund is mentioned in major U.S. news and business media outlets, Panel
A reports the results of five discontinuity tests in which the dependent variables are indicator variables testing whether mutual fund complexes increase
advertising activity or have their funds being mentioned more in the media, when comparing the pre-publication to the post-publication quarter (indicators
on increase denoted I+ [ ]), using locally weighted linear regressions of the dependent variable on fund rank with an indicator for rank < 10.5. Also reported
are the corresponding (p-value) and N, the number of fund-quarter observations participating in the test. The advertising activities tested are: the average
ad size published by the complex; the dollar amount spent on advertising by the complex; the number of ads published by the complex; the number
of times a fund’s rank is mentioned in ads; the number of times the fund is mentioned in (non-ad) news media. The tests in Panel B use Next Q flow
(percentage net capital flow into published fund class during post-publication quarter) as the dependent variable, while independently controlling for each
of the advertising activity indicators. The tests in Panel C repeat those of Panel B but replace Next Q flow with Next Q flow complex spillover (percentage
net capital flow into all funds of a published fund complex except the published one during post-publication quarter). If a complex appears more than once
in the rankings in the same quarter, only one of the occurrences is kept, at random, to avoid biasing the standard errors. α 1 is the discontinuity (coefficient
on the indicator) and 1 is the differential impact of the control variable around the discontinuity (coefficient on the interaction of the indicator and
control variable).
Panel A: Discontinuity tests
I+ [Ad size] I+ [Amount spent] I+ [# Ads published] I+ [# Rank mentions] I+ [# Media mentions]
Discontinuity at 10.5 (α 1 ) 0.265 ∗∗

0.202 ∗
0.195∗
0.042 0.026
(0.014) (0.061) (0.095) (0.822) (0.206)
N 1,225 567 1,165 471 10,488
Panel B: Effect of ads and media on increased fund-class flows
α1 2.852 1.637 3.241 2.627 2.079∗∗

(0.373) (0.664) (0.337) (0.616) (0.043)
1 -0.001 -0.008 -0.001 0.093∗ 0.032∗∗
(0.994) (0.825) (0.941) (0.071) (0.026)
N 1,225 567 1,165 471 10,488
Panel C: Effect of ads and media on increased complex flows
α1 2.158∗ 2.094 3.861∗∗ 1.533 1.936∗∗

(0.074) (0.134) (0.020) (0.292) (0.021)
1 −0.006 −0.028 −0.014 0.050∗ 0.026∗∗
(0.687) (0.166) (0.342) (0.089) (0.021)
N 384 186 357 114 6,429
∗∗
∗
testing the effect of increased advertising activity on flows, top-10 lists, within the top-10 there is a disproportionate
we observe no significant differential effect around the dis- amount of mentions to funds ranked 6–10 relative to those
continuity. This indicates that the efficacy of increasing the ranked 1–5: within the top 20, funds ranked 6–10 com-
number of ads published by a fund which made the list prise 41% of mentions, while funds ranked 1–5 account
is similar to that of a fund which just missed it, as is the for only 22% of mentions. Funds ranked 11–15 and 16–20
efficacy of increasing advertising budget or the actual ad garner 22% and 15% of mentions, respectively. The fourth
size. But while the efficacy per-ad does not change, the in- column uses an indicator variable indicating whether the
crease in ads published observed in Panel A will still result complex has increased the number of times it mentions
in increased flows. Additionally, it is possible that funds do the fund’s rank in its ads, following publication. We do
gain an increased efficacy for ads following publication, but not observe an increase in the number of times fund rank
this direct effect is offset by decreasing returns to scale on is mentioned pre- and post-publication (Panel A), but we
the efficacy of advertising, driven by the increase in ad- do observe some increased efficacy of mentioning funds’
vertising activity. Results are similarly insignificant when WSJ ranks more, if these funds were mentioned in the WSJ
considering the efficacy of advertising activity on spillover rankings, as indicated by the significant coefficients on the
flows into all other funds of the fund complex, as can be control ( 1 ) in Panels B and C.
seen in the first three columns of Panel C. In the last column of Table 6 we test whether funds
We further investigate whether funds that made it to are mentioned more in subsequent news (non-ad) articles
the top-10 are mentioned more, relative to those which in the media, based on the Factiva searches, and whether
did not, by extracting fund names mentioned in complex such subsequent news mentions have higher efficacy in
ads and matching them with our CRSP-based ranking. Out driving flows for funds which were mentioned relative
of 910 mutual fund names mentioned in the ads, we match to those which were not mentioned in the WSJ ranking.
693 (76%) to the CRSP rankings. Naturally, better perform- While we do not observe more news mentions for men-
ing funds are mentioned more: top-20 ranking funds ac- tioned funds (last column of Panel A), indicating that be-
count for only 5% of the sample but 22% of mentions, ing ranked in the WSJ does not cause more news articles to
within the regularly appearing 12 categories. More impor- be published regarding the mentioned fund, we do observe
tantly, and consistent with the hypothesis that fund com- increased efficacy of news mentions for ranked funds, both
plexes try to utilize the appearance of their funds in the in driving fund-class flows and spillover flows into the
entire complex (Panels B and C). These subsequent news portance of prominence and visibility by considering two
mentions for published funds may act as a reminder or re- natural experiments, one analyzing flow response to sim-
inforcement for interested consumers, whereas subsequent ilar WSJ rankings lists published regularly yet less visi-
news mentions for unpublished funds lack such a role. For bly, and one leveraging a change in the information en-
a discussion of the role of repetition in consumer persua- vironment after 2007, in which all rankings, not only the
sion see, e.g., Cacioppo and Petty (1979) and Campbell and top-10, were made readily available on the WSJ website.
Keller (2003). Both experiments show that prominence of the Category
We also attempt to test whether being mentioned in Kings rankings, and not the information they convey, is
the WSJ tables increases observed investor attention by the key driver of increased capital flows from investors to
considering the Google Search Volume Index (SVI), fol- published mutual funds: appearance on a similar yet less
lowing Da et al. (2011). Unfortunately, Google censors re- prominent top-10 list does not garner additional flows; the
sults with too few searches (with an undisclosed censoring impact of being a Category King on flows is similar pre-
threshold), such that only 4.2% of the fund-month obser- and post-2007.
vations in our data set have nonzero SVI. We are therefore We show that the presence of these extra flows, which
unable to test for an increase in SVI following publication stem from an appearance in the Category Kings lists, im-
in the WSJ. pacts fund decisions both prior to and after publication of
The results in this section support a channel in which the rankings. As predicted by theory, we find that the in-
the WSJ media mention is a first step in a chain, fol- centive created by the extra flows due to appearance in
lowed by other investor stimuli, such as exposure to mu- the rankings leads funds close to the cutoff, but only those
tual fund ads or exposure to news mentions, which then unlikely to be top-ranked in the subsequent quarter, to
leads to purchase. This is again consistent with a brand- increase tracking error volatility relative to the respective
name recognition channel, but less so with an information category. This is done in an attempt to “make the list.” The
channel. fact that only funds unlikely to be top-ranked next quarter
increase tracking error volatility highlights that managers
5. Conclusion are well cognizant of trade-offs associated with this risk-
shifting behavior.
We exploit a novel natural experiment to establish a In investigating potential propagation mechanisms of
causal link between media attention and consumer invest- the effect, we provide evidence showing that subsequent
ment behavior, independent of any conveyed information. to a fund’s appearance in the rankings, the fund family in-
Our identification strategy precisely controls for the pub- creases advertising activity, and fund families with Cate-
lication’s underlying information content, overcoming the gory King funds that are unlikely to be top-ranked next
typical challenge in the literature of decoupling effects of quarter increase advertising expenditures more. Further-
media attention from those of potential information reve- more, mentions of the fund rank in ads, as well as men-
lation due in part to the endogeneity of media coverage. tions of the fund name or ticker in media articles, have
We show that a single mention of a fund in a promi- an increased efficacy in generating flows for funds which
nent Wall Street Journal “Category Kings” ranking table, made the list relative to those which did not. These results
that appears once a quarter, leads to a 31% local aver- are in line with our finding that the increase in flows is
age increase in subsequent quarterly capital flows, along not limited to a short period of time close to publication
with a significant spillover effect to other funds of the day, but rather builds throughout the quarter and abates
same complex. A back-of-the-envelope calculation suggests once a new ranking is published, which is to be expected
that the mere presence on a top-10 list in a single rank- if advertising and subsequent media exposure play impor-
ing period allows a fund to collect almost $1.5 million in tant roles in the effect’s propagation.
increased fees, on average.31 This is in addition to, and Even in the 21st century, being the proclaimed King,
much higher than, increased fees stemming from the well- even if only for a day, helps.
documented return-flow relation, which amounts to an es-
timated $20 0,0 0 0.32 When also considering spillover flows Appendix A
to other funds in the complex, the increased fees from ap-
pearance on the list amount to a sizable $36 million.33 A1. Validity of the RDD
Our results reveal that media attention affects flows
even when it does not contribute new information, as long The literature discussing RDD best-practices (e.g., Hahn
as it appears sufficiently prominently. We highlight the im- et al., 2001; Imbens and Lemieux, 2008; Lee and Lemieux,
2010) suggests several tests to verify the validity of an
31
RDD. In our setting, a valid RDD requires quasi-random
The mean TNA and expense ratio for funds ranked 10 are $771M and
1.24%, respectively. Sirri and Tufano (1998) find that the typical holding
ranking around the cutoff. This requirement will be sat-
period for mutual fund investors is seven years. The 2.2 percentage points isfied if funds’ rankings are highly volatile. Additionally, a
local average increase in flows from being ranked 10 rather than 11 trans- discontinuity in any of the mutual funds’ observable char-
lates to $1,472,302 in extra fees over the seven-year holding period. acteristics pre-ranking may question the validity of the
32
The LLR predicted value from the left (right) of capital flows at rank =
RDD.
10 (rank = 11) is 9.56 percentage points (7.07 percentage points). Hence,
being ranked 10 rather than 11 increases flows by 2.49 percentage points, Quasi-random ranking is necessary to guarantee that
2.2 of which are due to the discontinuity. differences between mutual funds just above and just
33
The mean TNA for the complex of a fund ranked 10 is $24,115M. below the publication threshold are caused by media
Fig. A1. Frequency of entering/exiting top-10. For rank ∈ [1, 10], the graph depicts the empirical probability of not appearing in the top-10 by publication
date, conditional on holding that rank (a) three months, (b) two months, (c) one month, and (d) one day before the end of a ranking period. For rank ∈ [11,
20], the graph depicts the probability of appearing in the top-10 by publication date, conditional on holding that rank (a) three months, (b) two months,
(c) one month, and (d) one day before the end of a ranking period. See Fig. 1 for details of rank calculation.
attention rather than reflecting a spurious correlation. the cutoff. To verify this predication and the validity of the
Fig. A.1 provides evidence of high fund ranking volatility quasi-random assignment assumption, Table A.1 reports re-
in our data. We consider the empirical ex post probability sults of tests for discontinuity in several fund pre-ranking
of being in the top-10 list by publication date conditional characteristics. We find no statistically significant evidence
on the rank held by the mutual fund three months, two of a discontinuity in any of the tested characteristics, or in
months, one month, and one day before the end of a rank- other unreported characteristics such as Morningstar rank-
ing period. More than 50% of the time, a fund ranked ten ing, pre-publication beta, and 12b1 fees. The findings re-
a month before publication will not remain in the top-10 ported in Fig. A.1 and Table A.1 ensure that the “Local Ran-
by the time of publication, and almost 40% of the time, a domization” assumption of Lee and Lemieux (2010) holds.
fund ranked 11 will be part of the top-10 come publica- Finally, as is common in regression discontinuity de-
tion. Even when considering daily ranking volatility, a sim- sign studies, we verify the results are not driven by the
ilar pattern holds. Approximately 25% of the time, a fund choice of bandwidth. Fig. A.3 presents the magnitude and
ranked ten at the beginning of the last ranking day will not significance of the discontinuity in capital flows, based
be in the top-10 by the end of that day, and a fund ranked on a range of possible bandwidths. The discontinuity is
11 at the beginning of the day will cross the publication significant at the 5% level for all bandwidths between 4
cutoff and get published. Furthermore, Fig. A.2 shows that and 10, and the magnitude of the discontinuity in capi-
previous 12-month return, the driving variable, is remark- tal flows ranges between 2 p.p. and 3.5 p.p. the quarter
ably smooth, as expected. after publication. Bandwidth selection does not seem to
This quasi-random assignment of mutual funds around drive our results. As further robustness tests, we verify our
the rank = 10 cutoff implies there should be no disconti- discontinuity estimates by using the robust bias-correction
nuity in observable fund pre-ranking characteristics around RDD standard errors calculation method described by
Fig. A2. RDD analysis of pre-ranking returns by rank. Mutual funds are ranked based on previous 12-month returns within their investment category at
the beginning of every quarter from 20 0 0Q1 to 2012Q4 using data from CRSP. The figure depicts the two one-sided local linear kernel regressions of the
previous 12-month returns on the funds’ WSJ ranks.
Fig. A3. Effects of bandwidth on discontinuity estimation. This figure presents the magnitude and significance of discontinuity in capital flows as a function
of the bandwidth used in local linear kernel regressions of flows on ranks around the rank = 10.5 cutoff. The vertical dotted line is at the actual bandwidth
used. The actual bandwidth was chosen based on the optimal bandwidth estimator of Imbens and Kalyanaraman (2012). See Fig. 1 for details of rank
calculation, and Table 2 for details of discontinuity calculation.
Table A1
Discontinuity test for fund characteristics.
We repeat the discontinuity tests of Table 2 for several observable features of mutual funds in our sample: The fund’s total net assets (TNA—$M), age
(years), expense ratio (percent), management fee (percent), and front load fee (percent). All characteristics are measured at the end of the corresponding
12-month ranking period, before publication.
TNA Fund age Exp. ratio Mgmt. fee Front load
Discontinuity at 10.5 −26.691 −0.278 −0.008 −0.003 −0.178

(0.838) (0.488) (0.625) (0.237) (0.211)
Discontinuity w/ controls 22.018 −0.891 −0.002 −0.002 −0.162
(0.901) (0.159) (0.922) (0.515) (0.231)
Actual 10 vs actual 11 −157.909 −0.946 0.015 −0.001 −0.203
(0.276) (0.149) (0.580) (0.115) (0.376)
Fitted vs actual at 10 −100.751 −0.748 0.0 0 0 −0.003 −0.220
(0.419) (0.137) (0.989) (0.241) (0.321)
Fitted vs actual at 11 −83.849 −0.476 0.006 −0.001 −0.161
(0.592) (0.420) (0.796) (0.160) (0.399)
Table A2
Discontinuity test for other possible effects.
This table tests for several other possible effects of the publication in the WSJ by repeating the discontinuity tests of Table 2, but for: the expense ratio
of the fund a quarter after publication (Exp. ratio+Q); the management fee the fund charges a quarter after publication (Mgmt. fee+Q); the expense ratio
of the fund a year after publication (Exp. ratio+Y); the management fee the fund charges a year after publication (Mgmt. fee+Y).
Exp. ratio+Q Mgmt. fee+Q Exp. ratio+Y Mgmt. fee+Y
Discontinuity at 10.5 −0.004 −0.060 0.001 −0.005

(0.837) (0.453) (0.949) (0.830)
Discontinuity w/ controls 0.004 −0.027 −0.004 −0.004
(0.853) (0.433) (0.883) (0.887)
Actual 10 vs actual 11 0.008 −0.112 0.016 −0.046
(0.762) (0.121) (0.627) (0.309)
Fitted vs actual at 10 −0.005 −0.055 −0.005 −0.027
(0.797) (0.401) (0.890) (0.488)
Fitted vs actual at 11 0.008 −0.117 0.022 −0.025
(0.744) (0.110) (0.488) (0.563)
Calonico et al. (2014), as well as by using different ker- Cuoco, D., Kaniel, R., 2011. Equilibrium prices in the presence of dele-
nels. Our results (unreported) are unaffected. We also note gated portfolio management. Journal of Financial Economics 101 (2),
264–296.
that using rank as the forcing variable guarantees a similar Da, Z., Engelberg, J., Gao, P., 2011. In search of attention. Journal of Finance
number of observations on both sides of the cutoff and so 66 (5), 1461–1499.
the density test of McCrary (2008) does not apply. Del Guercio, D., Tkac, P.A., 2008. The effect of Morningstar ratings on mu-
tual fund flows. Journal of Financial and Quantitative Analysis 43 (4),
907–936.
A2. Supplementary discontinuity tests DellaVigna, S., Gentzkow, M., 2010. Persuasion: empirical evidence. An-
nual Review of Economics 2, 643–669.
DellaVigna, S., Kaplan, E., 2007. The Fox News effect: media bias and vot-
Though we focus our attention on increased capital
ing. Quarterly Journal of Economics 122 (3), 807–860.
flows, we examine other possible consequences of media Duong, T.X., Meschke, F., 2015. The rise and fall of portfolio pumping
exposure. We test and find no significant effect on fund among US mutual funds. Iowa State University and University of
Kansas. Unpublished working paper.
management fees and expense ratios a quarter, and a year,
Engelberg, J.E., Parsons, C.A., 2011. The causal impact of media in financial
after publication. Results of discontinuity tests for these markets. Journal of Finance 66 (1), 67–97.
possible outcomes are reported in Table A.2. In further un- Enikolopov, R., Petrova, M., Zhuravskaya, E., 2011. Media and political per-
reported results, we test for but fail to find a significant suasion: evidence from Russia. American Economic Review 101 (7),
3253–3285.
difference in capital flows into retail- and institutional- Fang, L.H., Peress, J., Zheng, L., 2014. Does media coverage of stocks affect
targeted fund classes. The magnitude of flows into all other mutual funds’ trading and performance? Review of Financial Studies
retail classes of the same fund is similar to that of flows 27, 3441–3466.
Gerber, A.S., Karlan, D., Bergan, D., 2009. Does the media matter? A field
into all other institutional classes of the same fund as well. experiment measuring the effect of newspapers on voting behav-
ior and political opinions. American Economic Journal: Applied Eco-
References nomics 1 (2), 35–52.
Grossman, S.J., Stiglitz, J.E., 1980. On the impossibility of informationally
Barber, B.M., Odean, T., 2008. All that glitters: the effect of attention and efficient markets. American Economic Review 70 (3), 393–408.
news on the buying behavior of individual and institutional investors. Gruber, M.J., 1996. Another puzzle: the growth in actively managed mu-
Review of Financial Studies 21 (2), 785–818. tual funds. Journal of Finance 51 (3), 783–810.
Basak, S., Pavlova, A., Shapiro, A., 2007. Optimal asset allocation and risk Hahn, J., Todd, P., Van-der Klaauw, W., 2001. Identification and estimation
shifting in money management. Review of Financial Studies 20 (5), of treatment effects with a regression-discontinuity design. Econo-
1583–1621. metrica 69 (1), 201–209.
Berk, J.B., Green, R.C., 2004. Mutual fund flows and performance in ratio- Hendricks, K., Sorensen, A., 2009. Information and the skewness of music
nal markets. Journal of Political Economy 112 (6), 1269–1295. sales. Journal of Political Economy 117 (2), 324–369.
Brown, K.C., Harlow, W.V., Starks, L.T., 1996. Of tournaments and tempta- Imbens, G.W., Kalyanaraman, K., 2012. Optimal bandwidth choice for the
tions: an analysis of managerial incentives in the mutual fund indus- regression discontinuity estimator. Review of Economic Studies 79 (3),
try. Journal of Finance 51 (1), 85–110. 933–959. doi:10.1093/restud/rdr043.
Cacioppo, J.T., Petty, R.E., 1979. Effects of message repetition and position Imbens, G.W., Lemieux, T., 2008. Regression discontinuity designs: a
on cognitive response, recall, and persuasion. Journal of Personality guide to practice. Journal of Econometrics 142, 615–635. doi:10.1016/
and Social Psychology 37 (1), 97–109. doi:10.1037//0022-3514.37.1.97. j.jeconom.20 07.05.0 01.
Calonico, S., Cattaneo, M.D., Titiunik, R., 2014. Robust nonparametric confi- Jain, P.C., Wu, J.S., 20 0 0. Truth in mutual fund advertising: evidence on fu-
dence intervals for regression-discontinuity designs. Econometrica 82, ture performance and fund flows. Journal of Finance 55 (2), 937–958.
2295–2326. Kaniel, R., Starks, L., Vasudevan, V., 2007. Headlines and bottom lines: at-
Campbell, M.C., Keller, K.L., 2003. Brand familiarity and advertising repe- tention and learning effects from media coverage of mutual funds.
tition effects. Journal of Consumer Research 30 (2), 292–304. University of Rochester and University of Texas at Austin. Unpublished
Carhart, M.M., 1997. On persistence in mutual fund performance. Journal working paper.
of Finance 52 (1), 57–82. Lee, D.S., Lemieux, T., 2010. Regression discontinuity designs in eco-
Carhart, M.M., Kaniel, R., Musto, D.K., Reed, A.V., 2002. Leaning for the nomics. Journal of Economic Literature 48, 281–355.
tape: evidence of gaming behavior in equity mutual funds. Journal of McCrary, J., 2008. Manipulation of the running variable in the regression
Finance 57 (2), 661–693. discontinuity design: a density test. Journal of Econometrics 142 (2),
Cheng, M.-Y., Fan, J., Marron, J.S., 1997. On automatic boundary correc- 698–714. doi:10.1016/j.jeconom.20 07.05.0 05.
tions. Annals of Statistics 25 (4), 1691–1708. Merton, R.C., 1987. A simple model of capital market equilibrium with in-
Chevalier, J.A., Ellison, G.D., 1997. Risk taking by mutual funds as a re- complete information. Journal of Finance 42 (3), 483–510.
sponse to incentives. Journal of Political Economy 105 (6), 1167–1200. Nanda, V., Wang, Z.J., Zheng, L., 2004. Family values and the star phe-
Corwin, S.A., Coughenour, J.F., 2008. Limited attention and the alloca- nomenon: strategies of mutual fund families. Review of Financial
tion of effort in securities trading. The Journal of Finance 63 (6), Studies 17 (3), 667–698. doi:10.1093/rpd/hhg054.
3031–3068.
Peress, J., 2014. The media and the diffusion of information in financial Sialm, C., Tham, T.M., 2016. Spillover effects in mutual fund companies.
markets: evidence from newspaper strikes. Journal of Finance 69 (5), Management Science 62, 1472–1486.
2007–2043. doi:10.1111/jofi.12179. Sirri, E.R., Tufano, P., 1998. Costly search and mutual fund flows. Journal
Phillips, B., Pukthuanthong, K., Rau, P.R., 2016. Past performance may be of Finance 53 (5), 1589–1622.
an illusion: performance, flows and fees in mutual funds. Critical Fi- Solomon, D.H., Soltes, E.F., Sosyura, D., 2014. Winners in the spotlight:
nance Review. forthcoming. media coverage of fund holdings as a driver of flows. Journal of Fi-
Reuter, J., Zitzewitz, E., 2014. How much does size erode mutual fund per- nancial Economics 113, 53–72.
formance? A regression discontinuity approach. Boston College and Tetlock, P.C., 2007. Giving content to investor sentiment: the role of media
Dartmouth College. Unpublished working paper. in the stock market. The Journal of Finance 62 (3), 1139–1168.

WSJ Category Kings

Uploaded by

Copyright:

Available Formats

WSJ Category Kings

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

WSJ Category Kings

Uploaded by

Copyright:

Available Formats

Journal of Financial Economics 123 (2017) 337–356

Contents lists available at ScienceDirect

Journal of Financial Economics

WSJ Category Kings – The impact of media attention on

1. Introduction in the media impacts ﬁnancial decision making, indepen-

Full Rank = 1 Rank = 10 Rank = 11 Rank = 20 Rank = 50 p-val(11–10)

Next Q Prev Y Next Q Next Q ﬂow Next Q ﬂow N

Discontinuity at 10.5 2.203∗∗ −0.600 0.108 1.862∗ 1.820∗∗ 11,232

Panel A: Published YTD settings

Next Q Prev Y Next Q Next Q ﬂow Next Q ﬂow N

All YTD 0.235 −0.251 −0.041 −0.464 0.139 22,464

Unpublished categories 0.043 −0.246 −0.182 0.162 −0.019 24,612

volatility captures how much a fund’s investment policy Table 4

Table 5 bility channel in which the WSJ publication increases con-

Panel A: Discontinuity tests

Discontinuity at 10.5 (α 1 ) 0.265 ∗∗

α1 2.852 1.637 3.241 2.627 2.079∗∗

α1 2.158∗ 2.094 3.861∗∗ 1.533 1.936∗∗

TNA Fund age Exp. ratio Mgmt. fee Front load

Discontinuity at 10.5 −26.691 −0.278 −0.008 −0.003 −0.178

Exp. ratio+Q Mgmt. fee+Q Exp. ratio+Y Mgmt. fee+Y

Discontinuity at 10.5 −0.004 −0.060 0.001 −0.005

You might also like