Published on: **Mar 4, 2016**

Published in:
Economy & Finance Business

- 1. THE JOURNAL OF FINANCE • VOL. LXIV, NO. 5 • OCTOBER 2009 The Relation between Price and Performance in the Mutual Fund Industry JAVIER GIL-BAZO and PABLO RUIZ-VERDU∗ ´ ABSTRACT Gruber (1996) drew attention to the puzzle that investors buy actively managed equity mutual funds, even though on average such funds underperform index funds. We uncover another puzzling fact about the market for equity mutual funds: Funds with worse before-fee performance charge higher fees. This negative relation between fees and performance is robust and can be explained as the outcome of strategic fee-setting by mutual funds in the presence of investors with different degrees of sensitivity to performance. We also find some evidence that better fund governance may bring fees more in line with performance.MANY STUDIES HAVE attempted to determine whether equity mutual funds areable to consistently earn positive risk-adjusted returns.1 Although these studieshave documented significant differences in risk-adjusted returns across funds,it became apparent early on (Sharpe (1966)) that those differences are to a largeextent attributable to differences in fund fees. Most research on mutual fundshas thus been aimed at determining whether the cross-sectional variation inperformance that is not explained by fees can be explained by the existenceof managers with superior stock-picking skills (see, e.g., Chevalier and Ellison(1999)). However, little attention has been paid to the relation between before-fee performance and fees. In this paper, we focus on this relation and investigatewhether differences in fees ref lect differences in the value that mutual fundscreate for investors. ∗ Gil-Bazo and Ruiz-Verdu are with the Department of Business Administration, Universidad ´Carlos III de Madrid. The authors thank Campbell Harvey (the editor), an anonymous associateeditor, and an anonymous referee for their comments, which have led to substantial improvementsin the paper. We are also grateful to Laurent Barras, Clemens Sialm, Massimo Guidolin, Alfonso ¨eNovales, Mikel Tapia, Manuel F. Bagu´ s, Fr´ d´ ric Warzynski, and Ralph Koijen as well as sem- e einar participants at Tilburg University, IE Business School, University of the Basque Country,Universidad de Navarra, Universitat Pompeu Fabra, Universidad Carlos III de Madrid, FinancialManagement Association European Conference, and the annual meeting of the European FinancialManagement Association for helpful comments and suggestions. The usual disclaimer applies. Partof the paper was written while Javier Gil-Bazo was visiting the Finance Department of TilburgUniversity. The financial support of Spain’s Ministry of Education and Science (SEJ2005-06655,SEJ2007-67448, and Consolider-Ingenio CSD2006-16), the BBVA Foundation, and the Madrid Au-tonomous Region (CCG07-UC3M/HUM-3413) is gratefully acknowledged. 1 See, for example, Brown and Goetzmann (1995); Gruber (1996); Carhart (1997); Daniel et al.(1997); Wermers (2000); Cohen, Coval, and Pastor (2005); Kacperczyk, Sialm, and Zheng (2005);or Kosowski et al. (2006). 2153
- 2. 2154 The Journal of Finance R Mutual fund fees pay for the services provided to investors by the fund. Be-cause the main service provided by a mutual fund is portfolio management, feesshould ref lect funds’ risk-adjusted performance. It follows that there should be apositive relation between before-fee risk-adjusted expected returns and fees. Incontrast to this prediction, we find a puzzling negative relation between before-fee risk-adjusted performance and fees in a sample of U.S. equity mutual funds:Funds with worse before-fee risk-adjusted performance charge higher fees. Tocheck the robustness of this finding, we use different estimation methods andperformance measures, and we investigate the relation for different subsam-ples. The negative relation between before-fee risk-adjusted performance andfees survives these robustness checks. We then set out to explain this anomalous relation by investigating the roleof performance in the determination of fund fees. Christoffersen and Musto(2002) propose that mutual funds’ fees are set taking into account the elasticityof the demand for their shares, so that funds facing less elastic demand chargehigher fees. These authors argue that funds with worse past performance faceless elastic demand because performance-sensitive investors leave funds fol-lowing bad performance. If performance is persistent for at least the worse-performing funds (as indicated by Carhart (1997)), Christoffersen and Musto’shypothesis could explain our finding of a negative relation between fees and ´before-fee performance. Gil-Bazo and Ruiz-Verdu (2008) provide a related ex-planation. These authors set forth a model of the market for mutual funds inwhich competition among high-performance funds for the money of sophisti-cated (performance-sensitive) investors pushes their fees down and drives thelow-performance funds out of that segment of the market. Low-performancefunds then target unsophisticated investors, to whom they are able to chargehigher fees. We also consider the possibility that funds with low expected per-formance have higher fees because they incur higher marketing costs, whichthey pass on to investors. Underperforming funds will have high distributioncosts or advertising outlays if they target unsophisticated investors, and theseinvestors are more responsive to advertising or more likely to use brokers topurchase mutual funds. We test these strategic fee-setting hypotheses againstan alternative cost-based explanation, according to which the negative relationbetween performance and fees that results from univariate regressions wouldsimply be due to the omission of fund characteristics associated with both loweroperating costs and better performance. To test these hypotheses, we estimate the relation between fund fees andperformance, investors’ sensitivity to fund performance, which we measure bythe estimated slope of the relation between performance and money f lows, anda number of variables that have been previously identified as determinants offunds’ operating costs. Our results are consistent with the strategic explana-tions described earlier: Even after controlling for a host of fund characteris-tics, underperforming funds and funds faced with less performance-sensitiveinvestors charge higher marketing and nonmarketing fees. The apparent inability of mutual fund competition to ensure an adequaterelation between risk-adjusted performance and fees raises the question of
- 3. Price and Performance in the Mutual Fund Industry 2155whether improvements in mutual fund governance can bring fees more intoline with the value that funds generate for investors. To evaluate the poten-tial effects of recent regulatory reforms that impose stricter requirements onmutual fund governance, we analyze the role played by fund governance infee determination. We find some evidence that funds with boards of directorsexpected, a priori, to provide more effective protection of investors’ interestscharge fees that better ref lect their risk-adjusted performance. The paper is organized as follows. In Section I, we describe the mutual fundfee structure and the data set. In Section II, we explain how we estimate fundperformance. In Section III, we estimate the relation between before-fee per-formance and fees and perform several tests to evaluate the robustness of theresults. In Section IV, we discuss and test several explanations for the esti-mated relation between fees and performance. Section V concludes. I. DataA. Mutual Fund Fee Structure Fund management fees are typically computed as a fixed percentage of thevalue of assets under management.2 These fees, together with other operatingcosts, such as custodian, administration, accounting, registration, and transferagent fees—comprise the fund’s expenses, which are deducted on a daily basisfrom the fund’s net assets by the managing company. Expenses are usuallyexpressed as a percentage of assets under management known as the “expenseratio.” Fees paid to brokers in the course of the fund’s trading activity aredetracted from the fund’s assets, but are not included in the expense ratio. Funds often charge “loads,” which are one-time fees that are used to paydistributors. These loads are paid at the time of purchasing (“front-end load”)or redeeming (“back-end load” or “deferred sales charge”) fund shares and arecomputed as a fraction of the amount invested.3 Since 1980, funds may charge so-called “12b-1 fees,” which are included inthe expense ratio and, like loads, are used to pay for marketing and distributioncosts. Since the 1990s, many funds offer multiple share classes with differentcombinations of loads and 12b-1 fees. Among the most common classes are classA shares, which are characterized by high front-end loads and low annual 12b-1fees, and class B and C shares, which typically have no or low front-end loadsbut have higher 12b-1 fees and a contingent deferred sales load. This contingentdeferred sales load decreases the longer the shares are held and is eventuallyeliminated (typically after 1 year for class C shares, and after 6 to 7 years forclass B shares). 2 Some funds allow the percentage to depend on fund performance. Although our data do notallow us to identify these funds, the evidence reported in Elton, Gruber, and Blake (2003), Kuhnen(2005), and Warner and Wu (2006) suggests that for most of our sample period the fraction of fundswith incentive fees is very small. 3 Funds often waive at least a fraction of the loads. Therefore, the loads typically reported indatabases, such as the one we use in this paper, can often overestimate effective loads.
- 4. 2156 The Journal of Finance RB. Sample Description We obtain our data from the Center for Research in Security Prices (CRSP)Survivor-Bias Free U.S. Mutual Fund Database for the period from December1961 to December 2005 (see Carhart (1997), Elton, Gruber, and Blake (2001),and Carhart et al. (2002) for detailed discussions of the data set). The initialsample contains all open-end mutual funds that are active in the 1961 to 2005period. From this initial sample, we exclude all funds that we cannot confi-dently describe as diversified domestic equity mutual funds. Thus, we removemoney market, bond and income, and specialty mutual funds, such as sector orinternational funds. To obtain our sample of diversified domestic equity mutualfunds, we use the information on funds’ investment objectives available in theCRSP database. Unfortunately, this information is not consistent throughoutthe 1961 to 2005 period. To create a homogeneous sample for the full sample pe-riod, we combine all the information available on funds’ investment objectives.Some of our results, however, are derived only for the 1992 to 2005 period, forwhich the information on funds’ investment objectives is precise and consistent. We remove from the sample observations with no information on returnsor expenses, or with zero expenses. The remaining sample contains some ob-servations with extreme values for expenses or returns that are either dataerrors, or that correspond to small funds with unusually high expenses or veryhigh volatility. Given the large size of the data set, we use Hadi’s (1994) out-lier detection method to search for these outliers and remove them from thesample. Finally, to ensure that our results are not driven by differences between indexand actively managed funds or between institutional and retail funds, we iden-tify passively managed (index) and institutional funds and exclude them fromthe sample. In the Internet Appendix,4 we provide a more detailed account ofthe procedure we use to construct our final sample, as well as summary statis-tics of the main variables.C. Fund Governance Data We use the January 2007 Morningstar Principia CD to obtain data on mu-tual funds’ board quality. This data set includes several governance ratings asof December 2006, which Morningstar uses to compute the so-called “Steward-ship Grade” for mutual funds. The measure of board quality that we use inour analysis (Morningstar’s “board quality” grade) is the sum of four equallyweighted components that measure, respectively, the degree to which the boardhas taken action “in cases where the fund clearly has not served investors well”;the significance of independent directors’ investments in the fund; whether theboard is “overseeing so many funds that it may compromise the ability to dili-gently protect the interests of shareholders”; and whether the fund meets theSecurities and Exchange Commission (SEC) requirement for the proportion of 4 The Internet Appendix is available at http://www.afajof.org/supplements.asp.
- 5. Price and Performance in the Mutual Fund Industry 2157 Table I Summary Statistics: Board Quality MeasureThe table shows the distribution of the board quality grade provided by Morningstar for funds inthe governance subsample in year 2005. The first row reports the frequency of each grade; thesecond row reports the relative frequency (in percentage terms); and the third row reports thecumulative frequency (in percentage terms). Very Poor Poor Fair Good Excellent TotalFrequency 1 69 374 486 176 1,106% 0.09 6.24 33.82 43.94 15.91 100Cumulative 0.09 6.33 40.14 84.09 100independent directors, regardless of whether it is subject to the requirement(see Morningstar (2006), pp. 1–2, for a detailed description). Thus, the boardquality grade is closely related to board characteristics that have been the focusof both regulatory reform and academic research.5 Of the 3,677 actively managed, diversified, noninstitutional funds (fundclasses) that are active in 2005 in our sample, there is board quality infor-mation in Morningstar’s January 2007 Principia CD for only 1,106 funds (thegovernance subsample). Although only one-third of the funds in the sample be-long to the governance subsample, these funds manage almost 80% of the totalnet assets managed by all funds in the sample. We note that the governancesubsample is not a random sample from the whole population. Among otherdifferences, funds in the governance subsample perform significantly better onaverage, belong to larger management companies, and are cheaper, larger, andolder than those in the nongovernance subsample. Table I reports the distribution of board quality grades. There are five gradesto which Morningstar assigns a numerical score that ranges from zero to two:Very Poor (0), Poor (0.5), Fair (1), Good (1.5), and Excellent (2). As Table I shows,most funds have Fair or Good grades, and only one obtains a Very Poor grade.The resulting average grade lies between Fair and Good. II. Mutual Fund Performance Estimation We use Carhart’s (1997) four-factor model to estimate before-fee risk-adjustedperformance: rit = αi + βrm,i rmt + βsmb,i smbt + βhml,i hmlt + βpr1 y,i pr1yt + εit , (1)where rit is fund i’s before-expense return in month t in excess of the 30-day risk-free interest rate—proxied by Ibbotson’s 1-month Treasury bill rate;6 rmt is the 5 Morningstar’s Stewardship Grade includes other components. See Wellman and Zhou (2008)for a recent analysis of the Morningstar Stewardship Grade. 6 Because fund returns are reported after expenses, to retrieve monthly before-expense returns,we add annual expenses divided by 12 to reported returns. This measure is only an approximation
- 6. 2158 The Journal of Finance Rmarket portfolio return in excess of the risk-free rate; and smbt and hmlt denotethe return on portfolios that proxy for common risk factors associated withsize and book-to-market, respectively. The term pr1yt is the return differencebetween stocks with high and low returns in the previous year. We includethis term to account for passive momentum strategies by mutual funds.7 Theterm αi is the fund’s alpha and captures the fund’s before-fee risk-adjustedperformance. We also consider Fama and French’s (1993) three-factor model,which uses only rmt , smbt , and hmlt , as well as conditional versions of the four-factor model. As in Carhart (1997), we follow a two-stage estimation procedure to obtain apanel of monthly fund risk-adjusted performance estimates. In the first stage,for every month t in years 1967 to 2005, we regress funds’ before-fee excessreturns on the risk factors over the previous 5 years. If less than 5 years ofprevious data are available for a specific fund-month, we require the fund tobe in the sample for at least 48 months in the previous 5 years, and thenrun the regression with the available data. In the second stage, we estimatea fund’s risk-adjusted performance in month t as the difference between thefund’s before-expense excess return and the realized risk premium, defined asthe vector of betas times the vector of factor realizations in month t.8 Rolling regressions yield a total of 232,386 monthly risk-adjusted before-feereturns corresponding to 3,109 different actively managed retail funds over 468months. Although the average annualized monthly return before expenses inour sample equals 10.52%, subtracting the risk-free rate and the part of fundreturns explained by the portfolio’s exposure to the Fama–French three factorsyields an average annualized monthly alpha of −21 basis points (bp), whichis further reduced to −70.6 bp when we take momentum into account. Thecorresponding annualized standard deviations are 18.13%, 7.33%, and 7.15%,respectively. III. The Relation between Fees and Performance In a well-functioning mutual fund market, mutual fund fees should be posi-tively correlated with expected before-fee risk-adjusted returns. Further, in theabsence of market frictions, all funds should earn zero expected after-fee risk-adjusted returns in equilibrium since, otherwise, there would be excess demand(supply) for funds with positive (negative) expected after-fee risk-adjusted re-turns (Berk and Green (2004)). In this context, if investors know funds’ alphas,because we ignore the compounding effect of the accrual of expenses over the year, and because theactual accrual of expenses may not be completely smooth (Tufano, Quinn, and Taliaferro (2006)). 7 Data are downloaded from Kenneth French’s web site, http://mba.tuck.dartmouth.edu/pages/faculty/ken.french/. 8 We follow Fama and MacBeth (1973) in our choice of a 5-year estimation period, instead ofthe 3-year period used by Carhart (1997). Although a longer estimation period excludes a greaterfraction of funds from the sample, it also reduces sampling error in betas and mitigates the effectof two forms of selection bias that affect mostly the subset of young funds in the CRSP database:omission bias and incubation bias (Elton, Gruber, and Blake (2001) and Evans (2009)).
- 7. Price and Performance in the Mutual Fund Industry 2159then equilibrium requires that αi − fi = 0 for every fund i, where fi denotes fundi’s fees, expressed as a fraction of the fund’s assets. This equilibrium conditioncan be equivalently written as αi = fi , for every fund i. Therefore, a graph depict-ing the equilibrium relation between fees and before-fee performance shouldyield an increasing linear relation with a slope of one. If, on the other hand,investors do not know funds’ alphas, then αi in the equilibrium condition isreplaced by investors’ expectation of fund i’s risk-adjusted returns. Equilibrium in the mutual fund market can be achieved through fee adjust-ment if funds with higher expected before-fee risk-adjusted returns increasetheir fees or underperforming funds lower theirs. However, Berk and Green(2004) show that in the market for mutual funds, market clearing can also beachieved via quantity adjustment: If there are decreasing returns to scale infund management, then f lows of money into funds that are expected to performbetter will reduce those funds’ expected performance until expected after-feerisk-adjusted returns are equalized across all funds in equilibrium. Whethermarket clearing takes place via fees, quantities, or a combination of both is,however, not material for the definition of market equilibrium. In any case,equilibrium requires that expected after-fee risk-adjusted returns be zero forall funds, and thus implies the linear relation (with a slope of one) between feesand before-fee performance described earlier. To investigate the relation between fund fees and before-fee risk-adjustedperformance, we first estimate by pooled ordinary least squares (OLS) the re-gression equation αit = δ0t + δ1 f it + ξit , ˆ i = 1, . . . , N , t = 1, . . . , T , (2)where fit is the fund’s expense ratio and αit is its risk-adjusted before-fee per- ˆformance measured according to Carhart’s (1997) model. The first row of Table II reports the slope coefficient and White’s (1980)heteroskedasticity-robust standard error estimated using the whole sampleof diversified actively managed retail equity funds. The regression includesmonth dummies to ensure that the estimated slope coefficient captures thecross-sectional relation between fees and risk-adjusted returns, not the effectof potentially correlated trends in those variables. The estimated slope coef-ficient is −0.63 and we can reject the null hypothesis of a unit slope at anyconventional significance level. Thus, estimation of equation (2) yields resultsthat are in stark contrast with the implications of a frictionless competitivemarket for equity mutual funds. In a market with frictions, it is not clear whether a priori we should expectδ 1 to be greater or smaller than one. In one plausible scenario, better fundscharge higher fees, but those fees are not high enough to fully compensate forthe differences in before-fee performance. In this scenario, funds with higherfees offer a higher after-fee performance and the estimated δ 1 is greater than one(δ1 > 1 implies that increases in fees are matched by larger increases in before-fee performance). In another plausible scenario, better funds overcharge fortheir ability to generate returns, which leads to differences in fees that exceed
- 8. 2160 The Journal of Finance R Table II Before-Fee Risk-Adjusted Performance and Expense RatiosThe table shows estimated slope coefficients for the OLS regression of funds’ monthly before-feerisk-adjusted performance on monthly expense ratios in the period from January 1962 to December2005. Betas are estimated using Carhart’s four-factor model (rows 1–3) or Fama–French’s three-factor model (row 4) with a 5-year estimation period. Risk-adjusted performance in month t isestimated as the difference between the fund’s monthly before-expense return in month t and theproduct of betas and the factor realizations for that month. All regressions include dummies formonths. Standard errors are reported in parentheses and adjusted R2 statistics in percentage.∗ , ∗∗ , ∗∗∗ indicate statistical significance at the 10%, 5%, and 1% levels, respectively. Superscriptsa, b, and c denote that the null hypothesis of a unit coefficient is rejected at the 10%, 5%, and 1%significance levels, respectively. The number of observations is 232,386.Risk-adjusted StandardPerformance Errors Coefficient Adj. R2Carhart White −0.6284∗∗∗,c 10.07 (0.1055)Carhart Clustered by Time −0.6284∗∗,c 10.07 (0.2529)Carhart Fama–MacBeth −1.4077∗∗∗,c 0.05 (0.3352)Fama–French Clustered by Time −0.2076c 9.43 (0.2599)differences in performance and to an estimated δ1 ∈ (0, 1). In this context, fundswith higher fees should exhibit better before-fee performance but worse after-fee performance. Finally, fees could be completely unrelated to funds’ before-fee performance, leading to δ1 = 0. However, the estimated slope coefficient isnegative and significantly different from zero, which suggests an a priori muchless plausible scenario in which funds with worse before-fee performance chargehigher fees. To account for cross-sectional correlation of residuals, we follow Petersen(2009) and Thompson (2006) and compute robust standard errors clustered bymonth. The second row of Table II shows that the robust standard error clus-tered by month (0.25) is more than twice as large as the White standard error(0.11), which suggests the presence of cross-sectional correlation in residuals(Petersen (2009)). However, further clustering by both fund and month (to alsoaccount for serially correlated residuals) barely changes the standard error(0.27). Therefore, unless otherwise noted, throughout the rest of this sectionwe report robust standard errors clustered by time in all pooled OLS regres-sions. We also estimate the relation between fees and performance using theFama–MacBeth two-step approach (Fama and MacBeth (1973)), which is de-signed to correct for cross-sectional correlation of residuals: First, we estimatemonthly regressions, and then we use the resulting monthly slope estimatesto compute the average slope for the whole sample and its standard error. Thethird row in Table II shows that the Fama–MacBeth method yields a coefficient
- 9. Price and Performance in the Mutual Fund Industry 2161of −1.4 and a standard error of 0.34. (Weighting the monthly slope coefficientsby the number of observations in each month yields a coefficient of −0.76 witha standard error of 0.23.) Therefore, when we take into account the possibilityof correlated residuals in our estimation of standard errors, we also reject, atany conventional significance level, the hypothesis that the slope of the fee-performance relation is one. Moreover, our estimate of the slope coefficientis negative and significantly different from zero at the 5% (1%) significancelevel when we use clustered (Fama–MacBeth) standard errors to perform ourtests. A potential problem with our results is that estimated alphas contain funds’true abnormal performance, but they also contain estimation error from twosources, the residuals of the performance attribution model (1) and estimationerror in the realized risk premium. Estimation error in the dependent vari-able in (2) may affect inference in several ways. First, this estimation errorincreases the variance of the residuals and thus the standard errors of parame-ter estimates. Therefore, estimation error in alphas decreases the likelihood offinding a significant relationship between alphas and expense ratios. Second,if extreme alphas are more likely to include a large estimation error and havea large inf luence on the estimated coefficient, then estimation error in alphasmay affect the estimated slope coefficient. More generally, even in the absenceof estimation error in alphas, our results could be driven by a relatively smallnumber of funds with extreme alphas or expenses. To explore the monotonicity and linearity of the fee-performance relation,Figure 1 shows conditional expected risk-adjusted performance as a nonpara-metric function of expenses. We estimate the conditional expectation by us-ing the Nadaraya–Watson estimator with a Gaussian kernel (see, for example, ¨Hardle (1990)). We also plot conditional expected alphas as a linear function ofexpenses as implied by the estimated OLS coefficient of the linear regression. Toaccount for time effects, we de-mean expenses and risk-adjusted performanceby subtracting the month’s average. For values of expenses below the sample’s99th percentile, Figure 1 shows that before-expense risk-adjusted performancedecreases monotonically with expenses and that the relation may be well de-scribed by a linear function. However, for funds with expenses in the top samplepercentile, the relation appears far from monotonic, although the large confi-dence intervals in this low density region suggest that inference on the meanrisk-adjusted performance of very expensive funds is problematic. Therefore,Figure 1 shows that the presence of some funds with both extreme expenseratios and extreme risk-adjusted performance does not seem to affect the OLSslope coefficient. Finally, although all funds are affected by estimation error in betas, the stan-dard errors of estimated betas may vary across funds in systematic ways. Thus,although White (1980) and clustered standard errors already account for het-eroskedasticity, we also run several generalized least squares regressions, ob-taining results (see the Internet Appendix) that are essentially identical tothose obtained by OLS.
- 10. 2162 The Journal of Finance R 0.04 0.03 0.02 0.01 0 −0.01 −0.02 4-Factor Alpha −0.03 −0.04 −0.05 −0.06 −0.07 −0.08 −0.09 −0.1 −0.11 −0.12 0 0.005 0.01 0.015 0.02 0.025 0.03 0.035 0.04 Expense Ratio 0.05 0.04 4-Factor Alpha (Cross-sectionally Demeaned) 0.03 0.02 0.01 0 −0.01 −0.02 −0.03 −0.04 −0.05 −0.06 −0.07 −0.08 −0.09 −0.1 −0.11 −0.015 −0.01 −0.005 0 0.005 0.01 0.015 0.02 0.025 Expense Ratio (Cross-sectionally Demeaned)Figure 1. Nonparametric Regressions. The figure shows mutual fund conditional expectedrisk-adjusted performance as a nonparametric function of expense ratios (solid thick line). Risk-adjusted performance in month t is estimated as the difference between the fund’s monthly returnin month t and the product of betas and the factor realizations for that month using Carhart’sfour-factor model. Both monthly risk-adjusted performance and expense ratios are annualized. Theconditional expectation has been estimated using the Nadaraya–Watson estimator with a Gaussiankernel. Dotted lines show upper and lower bounds of the 95% pointwise confidence interval. Thefigure also plots conditional expected alphas as a linear function of expenses as implied by theestimated OLS coefficient of the linear regression (solid thin line). Dashed vertical lines correspondto the 1st and 99th percentiles of the expense ratio sample distribution. The bottom panel displaysthe results obtained when we de-mean expenses and risk-adjusted performance by subtracting themonth’s sample average.
- 11. Price and Performance in the Mutual Fund Industry 2163 To further account for potential estimation error in alphas stemming fromthe estimation of betas, we also report results obtained using other performanceevaluation models. The last row of Table II shows the results that we obtainwhen we estimate alpha using the Fama–French three-factor model. The es-timated coefficient is −0.2 with a standard error of 0.26. Therefore, we canreject the null hypothesis that the slope of the fee-performance relation is oneat the 1% significance level. However, the estimated coefficient is not signif-icantly different from zero, in contrast to the results obtained with Carhart’sfour-factor alpha. This finding suggests that on average, more expensive fundsexhibit greater exposure to the momentum factor. Further, in addition to thethree- and four-factor unconditional models, we use several conditional versionsof Carhart’s four-factor model (see, for example, Ferson and Schadt (1996) andKosowski et al. (2006)) and obtain results consistent with those from the uncon-ditional model. In the Internet Appendix, we describe in detail the specificationand results of the conditional models. Another possible concern about our results is that they may be due to theinf luence of funds with small market share. Such funds may exhibit both lowperformance and high expense ratios. However, our requirement that fundshave at least 48 months of return information to be included in the samplealready filters out the effect of unsuccessful funds that are terminated beforereaching that threshold. To evaluate the inf luence of small funds that have sur-vived for at least 5 years, we reestimate equation (2) for a sample that excludesobservations with relatively low values of assets under management in eachmonth. Panel A of Table III shows that the negative relation between expenseratios and before-expense risk-adjusted performance holds when the lowestsize decile is excluded each month. Although the estimated coefficient is signif-icantly different from one at the 1% significance level, it is smaller in absolutevalue than our estimate for the whole sample and only marginally significantlydifferent from zero. Excluding further deciles leads to similar coefficients andstandard errors (with coefficients that are either not statistically significant oronly marginally so). Thus, fund size appears to play a role in explaining therelation between risk-adjusted performance and fund expenses. In the analysis above, we consider expense ratios as the only explicit costof delegated portfolio management. However, investors often pay loads at thetime of purchasing and/or redeeming mutual fund shares. Hence, the previousregressions could be capturing a negative relation between performance anda specific component of total fund share ownership cost, but not necessarilya negative relation between performance and the total fees paid by investors.In particular, if more expensive funds (when only expenses are considered)charged lower loads, then after-fee performance (when all fees are considered)could still be equalized across funds. One way to circumvent this problem is to focus exclusively on funds for whichannual operating expenses account for 100% of all fees. In Table III (Panel B),we estimate equation (2) for no-load funds only. The estimated slope coefficientof −0.91, which is significant at the 1% level, indicates that total ownershipcost and performance are negatively correlated for no-load funds.
- 12. 2164 The Journal of Finance R Table III Regressions by SubsamplesThe table shows estimated slope coefficients for the OLS regression of funds’ monthly before-feerisk-adjusted performance on monthly fees in the period from January 1962 to December 2005 forPanels A to C, and the period from January 1992 to December 2005 for Panel D. Betas are estimatedusing Carhart’s four-factor model with a 5-year estimation period. Risk-adjusted performance inmonth t is the difference between the fund’s monthly before-expense return and the product ofbetas and the factor realizations in t. Monthly fees are defined as the annual expense ratio dividedby 12, except for Panel B, where monthly fees are annual expense ratios divided by 12 plus thesum of front-end and back-end loads divided by the assumed holding period in months. In Panel Athe sample does not include for each month the decile of fund-month observations with the lowesttotal net assets among all actively managed retail funds. In Panel B No-Load Funds are defined asthose charging no front- or back-end loads. All regressions include dummies for months. Standarderrors (in parentheses) are clustered by time. Adjusted R2 statistics are reported in percentage.∗ , ∗∗ , ∗∗∗ indicate statistical significance at the 10%, 5%, and 1% levels, respectively. Superscriptsa, b, and c denote that the null hypothesis of a unit coefficient is rejected at the 10%, 5%, and 1%significance levels, respectively.Subsample Coefficient Adj. R2 Obs. Panel A: Effect of Small FundsDeciles 2–10 −0.4602∗,c 10.28 225,450 (0.2693) Panel B: Other FeesNo-load funds −0.9099∗∗∗,c 8.51 81,323 (0.3220)Load funds (2-year holding period) −0.2660∗∗∗,c 11.22 150,148 (0.1003)Load funds (7-year holding period) −0.5186∗,c 11.22 150,148 (0.2823) Panel C: Regressions by Subperiods1967–1976 −0.9938b 14.26 16,504 (0.9190)1977–1986 −0.9581∗,c 8.81 26,591 (0.5625)1987–1996 −0.8129∗∗∗,c 5.65 39,567 (0.2560)1997–2005 −0.5384c 10.43 149,724 (0.3347) Panel D: Regressions by Investment ObjectiveAggressive growth funds −0.4304a 17.44 13,419 (0.7244)Growth MidCap funds −0.0183a 27.99 15,366 (0.6181)Growth and income funds −0.5840∗∗∗,c 12.79 39,221 (0.1619)Growth funds −0.6557∗∗,c 8.48 70,277 (0.2757)Small company growth funds −0.6931∗,c 21.88 35,065 (0.3974)
- 13. Price and Performance in the Mutual Fund Industry 2165 Because load funds constitute two-thirds of the sample, we also estimate therelation between performance and a measure of total fund ownership cost forthese funds. Following Sirri and Tufano (1998), we compute total annual owner-ship costs by adding annuitized total loads (total loads divided by the number ofyears, τ , that investors keep their money in the fund) to annual expense ratios.Although previous studies typically set τ = 7, redemption rates for equity fundsfor more recent periods suggest a shorter average holding period in the range of2.5 to 5 years. Therefore, we perform the analysis for τ = 2 and 7 years. Becausethe analysis in Table III is conducted at a monthly frequency, the independentvariable is total monthly ownership cost, defined as total annual ownershipcost divided by 12. Panel B of Table III shows that total ownership cost is neg-atively and significantly associated with before-fee risk-adjusted performancefor both holding periods. Further, the unit slope hypothesis is rejected at anyconventional significance level. As a final test of the robustness of our regression results, we estimate equa-tion (2) for different subperiods and mutual fund categories. Panel C of Table IIIshows that the perfectly competitive equilibrium condition is clearly violatedin all subperiods considered. Moreover, the relation between before-fee risk-adjusted performance and expenses is negative in all the subperiods, althoughnot significantly different from zero in the 1967 to 1976 and 1997 to 2005 sub-periods. Lack of significance in the 1967 to 1976 subperiod could be due to therelatively low number of observations, which results in a large standard errorfor the slope coefficient. However, failure to reject the null hypothesis of a zerocoefficient in the last subperiod appears to happen because the fee-performancerelation becomes f latter in the last years of our sample: Although the estimatedslope coefficient lies between −0.81 and −0.99 in the pre-1997 years, it is −0.54in the last subperiod considered. Finally, for the 1992 to 2005 period, for which the classification is detailedand consistent, we divide the sample into subsamples according to the Standard& Poor’s detailed objective code as reported by CRSP, and then run the regres-sion for each subsample. Panel D of Table III shows that expense ratios arenegatively related to performance for all five investment objectives, althoughthe relation is not statistically significant for Aggressive Growth and GrowthMidCap funds. For these investment objectives, the unit slope hypothesis canbe rejected, but only at the 10% significance level. When we replace expenseratios with total ownership cost with τ = 2 or 7 years, we obtain results thatare similar to those of Panel D (see the Internet Appendix for results). IV. Explaining the Relation between Fees and Performance The negative relation between before-fee performance and fees that we un-cover in the previous section is at odds with the intuitive expectation that feesshould, at least to some extent, ref lect the value that funds create for investors.In this section, we set forth and test different explanations for this apparentlyanomalous relation.
- 14. 2166 The Journal of Finance RA. Cost-Based Explanations According to the first explanation, fees simply ref lect the costs of operatingthe fund. If low costs are associated with better before-fee risk-adjusted perfor-mance, then a univariate regression would result in a negative relation betweenfees and performance. Fund performance could be positively associated with fund costs if highercosts ref lect higher salaries to attract more talented managers or a larger in-vestment in research tools, but there are also arguments for a negative corre-lation between costs and performance. For instance, there might be economiesof scale that lower operating costs for larger funds. In addition, larger size maybe associated with better performance if a fund’s size ref lects its past perfor-mance, and performance is persistent. Similarly, older funds might benefit fromlearning economies, which could be passed on to investors in the form of lowerfees. If fund longevity is related to good performance, as would be the case iflow-performance funds were more likely to close down, we could observe a neg-ative relation between costs and performance. Finally, higher managerial skillmay be associated with both better investment decisions and more efficientmanagement of fund operations, which would translate into lower operatingcosts.B. Strategic Explanations The second explanation views the negative relation between before-fee per-formance and fees as the result of strategic fee-setting by mutual fund manage-ment companies or other service providers to the fund. One such explanationhas been proposed and empirically tested for money market mutual funds byChristoffersen and Musto (2002). On the basis of empirical studies on mutualfund f lows (e.g., Sirri and Tufano (1998)) and survey data on mutual fund in-vestors’ behavior (Capon, Fitzimmons, and Prince (1996), Alexander, Jones, andNigro (1997)), Christoffersen and Musto argue that mutual fund investors differin their performance sensitivity. They also argue that funds with a worse per-formance history will have a less performance-sensitive clientele because theperformance-sensitive investors will have f led those funds following bad perfor-mance. Therefore, funds with a greater proportion of performance-insensitiveinvestors will charge higher fees because for these funds the reduction in after-fee performance caused by an increase in fees will not translate into a large f lowof money out of the fund. It follows that funds with bad past performance willfind it optimal to charge higher fees. Christoffersen and Musto’s explanationcan be tested by using a measure of the performance sensitivity of each fund’sf lows. ´ Gil-Bazo and Ruiz-Verdu (2008) provide a related strategic explanation forthe negative relation between before-fee performance and fees. These authorsdevelop an asymmetric information model of the mutual fund market in whichmutual funds differ in their expected performance and investors differ in their ´performance sensitivity. Gil-Bazo and Ruiz-Verdu show that competition for
- 15. Price and Performance in the Mutual Fund Industry 2167the money of performance-sensitive investors leads to an equilibrium in whichfunds that expect to earn higher returns (“good” funds) reduce their fees up tothe point at which they effectively price funds that expect lower returns (“bad”funds) out of the performance-sensitive segment of the market. Good funds areable to price bad funds out of the market because the revenues of managementcompanies are determined as a fraction of assets under management. Therefore,for any given fee (expressed as a fraction of asset value), good funds, whichcan be expected to achieve a larger increase in the value of their assets, willearn higher expected revenues. As a result, there is a fee level at which goodfunds break even in expectation, and low-performance funds incur an expectedloss. Unable to compete for performance-sensitive investors, bad funds raisetheir fees to extract rents from performance-insensitive investors. Gil-Bazo and ´Ruiz-Verdu’s predictions can also be tested by using a measure of a fund’s risk-adjusted expected performance. A related explanation for our results is that low-performance funds incurhigher marketing costs and those costs are passed on to investors in the form ofhigher fees. If low-performance funds target performance-insensitive investorsand these investors purchase mutual fund shares mostly through brokers,then low-performance funds will incur higher marketing costs than funds soldthrough more direct distribution channels (e.g., from mutual fund supermar-kets or directly from the management company). Bergstresser, Chalmers, andTufano (2009) provide evidence that supports this hypothesis. These authorsreport that on average funds sold through the direct or fund supermarket chan-nels perform better and have lower fees than those sold through the brokerchannel. The marketing costs of low-performance funds could also be higher ifintermediaries have to be compensated for the higher effort or potential loss ofreputation associated with selling low-performance funds.9 Alternatively, if un-sophisticated investors are more responsive to advertising and low-performancefunds target those investors, then low-performance funds will spend more onadvertising because the marginal return of their advertising investment willbe higher. These strategic marketing explanations imply that underperformingfunds will have higher marketing fees, an implication that we can test with ourdata. ´ Both Gil-Bazo and Ruiz-Verdu’s (2008) explanation and the strategic mar-keting hypothesis assume that fund management companies form expectationsabout the future performance of the funds they manage, and that they condi-tion their funds’ fees on those expectations. This assumption seems reasonablebecause management companies are able to observe all publicly available infor-mation about funds’ portfolio choices and returns, they have access to a wealthof data not available to outsiders (such as high-frequency data on portfolioholdings), and they have the skills to analyze all that information. Further,management companies themselves may strongly inf luence the performanceof their funds because they decide how to allocate scarce resources (staff, re-search analysis, underpriced IPOs) among them (Gaspar, Massa, and Matos 9 We thank an anonymous referee for suggesting this possibility.
- 16. 2168 The Journal of Finance R(2006), Guedj and Papastaikoudi (2005)). However, a limitation of these hy-potheses is that they do not explain why, rather than adjusting fees, manage-ment companies do not try to improve the expected performance of their fundsby replacing underperforming managers or changing their investment strate-gies. In particular, funds with low expected performance could be turned intocloset indexers, guaranteeing a level of performance close to the benchmark.Addressing these limitations is beyond the scope of this paper, but allowingfor managers’ replacement or changes in investment strategies may not sub-stantially change the predictions of the strategic explanations. First, not allcompanies will be able to hire only the managers with the highest expectedperformance, and mutual funds with less-than-top managers may be able tosurvive, at least in the medium run, especially in the presence of unsophisti-cated investors. Second, even if closet indexing guarantees that returns do notfall too much below a fund’s benchmark, closet indexers will still underperformfunds that have managers who are able to generate positive alphas. Further,a strategy of closet indexing also entails some less-obvious costs. In particular,sophisticated investors will leave a fund that they identify as a closet indexer be-cause closet indexers are dominated in terms of after-fee returns by index funds. Another potential limitation of the strategic explanations is that althoughmany studies document the existence of a significant pool of unsophisticatedinvestors, it is an open question whether unsophistication can persist in themedium or long run. In particular, cheaper or better-performing funds maywant to educate performance-insensitive investors to avoid expensive funds.However, Gabaix and Laibson (2006) show that firms may not have incentivesto educate investors to avoid “shrouded” costly attributes (product attributesthat are hidden by firms, even though they could be almost costlessly revealed).Even though most of Gabaix and Laibson’s analysis focuses on add-ons (productattributes that the consumer can substitute away at a cost), the authors alsodiscuss factors that might limit firms’ incentives to educate unsophisticatedinvestors about unavoidable shrouded costs.C. Fund Governance The strategic explanations discussed earlier implicitly assume that mutualfund fees are set by management companies so as to maximize fee revenues.However, U.S. mutual funds are legal entities that are independent of the com-panies managing their portfolio. Control over the fund is delegated by fundshareholders to a board of directors (or trustees), which is responsible for con-tracting the management of the fund’s portfolio with a management company.Thus, the management fee is not set unilaterally by the management company;rather, it is negotiated with the board of directors. Similarly, the fund’s direc-tors negotiate the fees paid to other service providers, such as distributors ortransfer agents. Boards of directors have the fiduciary duty to ensure that thosefees ref lect the value for fund investors of the services they are paying for. Despite legal provisions imposing rigorous governance requirements on mu-tual funds, there remain important conf licts of interest that may interfere with
- 17. Price and Performance in the Mutual Fund Industry 2169fund directors’ fiduciary duty. For instance, management companies select themembers of a fund’s initial board of directors. In practice, contract renegotia-tions and changes in the fund’s management company are infrequent (Kuhnen(2005), Warner and Wu (2006)), suggesting that directors’ interests may be morealigned with those of management companies than with those of fund investors. Prior research is not fully conclusive as to whether certain mutual fund gov-ernance structures are able to mitigate these conf licts of interest and lead tolower fees. Although Tufano and Sevick (1997) provide evidence for 1992 thatfunds with smaller boards and funds with boards that have a higher fractionof independent directors have lower fees, more recent studies (Meschke (2007),Ferris and Yan (2007)) obtain mixed results. We hypothesize that the boardsof better-governed funds will approve “fair” or “reasonable” fees in fulfillmentof their fiduciary duty. This does not necessarily mean lower fees (althoughbetter governance could also result in lower fees), but rather that fees are morein line with the fund’s performance. Therefore, the relation between fees andperformance should be positive, or at least f latter, for better-governed funds.Similarly, boards of higher quality may resist more strongly any attempts bymanagement companies and other service providers to charge higher fees infunds with more performance-inelastic investors. If this were the case, therelation between fund fees and performance sensitivity would be f latter forbetter-governed funds. Although with limitations imposed by the nature of ourdata, we test these hypotheses using Morningstar’s board quality grade as ameasure of fund governance.D. Empirical Strategy To test the empirical validity of the proposed explanations for the negativerelation between fees and before-fee risk-adjusted performance, we investigatehow fees vary with fund characteristics, f low-to-performance sensitivity, andperformance. We assume that fund i’s fee at time t, fit , is a linear function of avector xit−1 of lagged values of variables that are likely to determine the fund’soperating costs, the performance-sensitivity of the fund’s f lows, Sit , and thefund’s expected before-fee performance in period t, αit : f it = γ xit−1 + λ S Sit + λα αit + νit , (3)where νit is a generic error term. Because data on most variables are availableyearly during most of the sample period, the time index t in equation (3) refersto calendar years. To test the potential effects of fund governance on fund fees, we also esti-mate an extended version of equation (3) that allows both the intercept and thecoefficients on performance and performance sensitivity to depend on boardquality. We build on the literature on mutual fund fee determinants, which mostlyconsiders fund fees as a ref lection of operating costs, to select the variables that
- 18. 2170 The Journal of Finance Rmay inf luence the costs of operating a fund.10 For every fund-year observationwe consider the following variables: size, which we define as the log of the year-end total net asset value; age, computed as the log of the number of years sincethe fund’s organization; size of the complex and number of funds in the complex,which we define as the log of the total net asset value of the funds managed bythe company that manages the fund, and the total number of funds managed bythat company, respectively; reported annual turnover; volatility, computed asthe standard deviation of the fund’s monthly returns in the year; and dummyvariables for the fund’s investment objective. We also include a dummy variableto identify single-class load funds and dummies for the main share classes.Doing so makes it possible for us to correct for the potential distortions inducedby using a homogeneous holding period for all funds because we can expectinvestors with different holding periods to select different share classes. Weinclude time dummies in all regressions. We use αit , which we define as the sum of estimated monthly alphas in year t,as our proxy for the fund’s expected before-fee performance.11 Estimated alpha(αit ) is a good measure of expected performance (αit ) as long as the measure-ment error in αit is not correlated with the level of fees. If no such correlationexists, the result of including estimated, rather than expected, performanceas a regressor reduces to the well-known attenuation bias in the presence ofmeasurement error. Thus, the performance coefficient estimates are likely tobe biased toward zero. To obtain a measure of the f low-to-performance sensitivity, Sit , we proceedin two steps. First, we estimate a model of money f lows into mutual funds.Based on prior studies of fund f low determinants, we allow the sensitivity off lows to past performance to depend on both the level of performance and fundcharacteristics. In the second step, we estimate f low-to-performance sensitivityfor each fund and year as the first derivative of conditional expected f low withrespect to the previous year’s performance. The Appendix provides a detaileddescription of the construction of this variable.E. The Determinants of Fund Fees To investigate the relation between before-fee performance and the total feespaid by investors, we first use total ownership cost as our dependent vari-able. We compute total ownership cost as the expense ratio plus total loads di-vided by seven. Since we consider a 7-year holding period, to account for usual 10 Different aspects of mutual fund fee determination have been studied, among others, by Fer-ris and Chance (1987), Tufano and Sevick (1997), Latzko (1999), Malhotra and McLeod (1997),Chalmers, Edelen, and Kadlec (2001), Luo (2002), Deli (2002), and Golec (2003). 11 In the remainder of the paper, we focus on unconditional Carhart’s alpha exclusively. We useαit as a measure of the alpha expected by the manager of fund i at the beginning of period t underthe assumption that fees are set at the beginning of period t. If fees were set in the middle ofperiod t, our measure of expected performance would thus aggregate performance observed priorto setting fees with expected performance. We have estimated the fee equation using αit+1 as ameasure of expected returns and obtained identical results.
- 19. Price and Performance in the Mutual Fund Industry 2171 Table IV Mutual Fund Fee DeterminantsThe table reports estimated coefficients for yearly regressions of funds’ fees on selected fund char-acteristics in the 1993 to 2005 period. The dependent variable in columns (1) to (6) is total annualownership cost (TOC), computed as total loads divided by seven plus the annual expense ratio.In columns (7) and (8), the dependent variable is marketing fees (Mark.), defined as total loadsdivided by seven plus 12b-1 fees, and nonmarketing fees (N-Mark.), computed as the expense ratiominus 12b-1 fees, respectively. Back-end loads are assumed to be zero for share classes B and C.The coefficients in columns (1) to (3) and (7) to (8) are estimated by pooled OLS. Column (4) reportsestimated coefficients for a regression with management company fixed effects. In column (5) allvariables are asset-weighted averages at the management company level. Column (6) reports es-timated coefficients for a regression with fund fixed effects. The size of the management companyand the number of funds in the management company are denoted by Co. Size and # funds, respec-tively. σt is the standard deviation of the fund’s monthly returns in year t. St denotes the slope of theestimated f low-to-performance relation. αt is the year t four-factor alpha. All regressions includeyear dummies and dummy variables for the different investment objectives and share classes. Allfees are expressed in bp. The table also reports robust standard errors (in parentheses), which areclustered by fund in columns (1) to (3) and (6) to (8), and by management company in columns (4)to (5). The total number of observations and the adjusted R2 of the regression (in percentage) arereported at the bottom of the table. ∗ , ∗∗ , ∗∗∗ indicate statistical significance at the 10%, 5%, and1% levels, respectively. TOC Mark. N-Mark. (1) (2) (3) (4) (5) (6) (7) (8)Sizet−1 −9.38∗∗∗ −9.77∗∗∗ −9.76∗∗∗ −7.04∗∗∗ −7.25∗∗∗ −6.35∗∗∗ −3.40∗∗∗ (0.70) (0.82) (0.82) (1.02) (0.80) (0.60) (0.54)Aget−1 −4.96∗∗∗ −8.65∗∗∗ −8.63∗∗∗ −12.05∗∗∗ −0.18 −9.91∗∗∗ −2.99∗∗∗ −5.61∗∗∗ (1.55) (1.73) (1.73) (1.69) (0.19) (0.00) (1.10) (1.29)Co. Sizet−1 −1.06 −1.17 −1.14 −3.75∗∗ −12.82∗∗∗ −1.39∗ 7.46∗∗∗ −8.64∗∗∗ (0.73) (0.81) (0.80) (1.75) (1.58) (0.84) (0.64) (0.55)#fundst−1 −0.09 −0.14 −0.14 0.16 1.44∗∗∗ −0.05 −0.66∗∗∗ 0.52∗∗∗ (0.09) (0.10) (0.10) (0.19) (0.53) (0.10) (0.09) (0.06)Turnovert−1 3.29∗∗∗ 3.27∗∗∗ 3.19∗∗∗ 0.14 9.84∗∗∗ 1.93∗∗∗ −0.70 3.92∗∗∗ (0.73) (0.86) (0.86) (1.04) (2.34) (0.82) (0.70) (0.60)σ t−1 42.53∗∗∗ 34.70∗∗∗ 28.18∗∗ 13.09∗ 58.53∗∗ −12.22∗∗∗ 17.89∗∗ 16.33∗∗ (10.45) (11.49) (11.34) (7.62) (27.78) (6.01) (8.73) (7.92)αt −14.65∗∗∗ −18.56∗∗∗ −19.48∗∗∗ −11.74∗∗ −39.16∗∗ −5.10∗ −11.79∗∗∗ −6.94∗ (5.33) (5.64) (5.68) (5.00) (17.77) (2.92) (4.41) (3.67)αt 2 75.84∗∗∗ (27.38)St −7.06∗∗∗ −7.01∗∗∗ −3.81∗∗∗ −6.54∗∗∗ −0.70 −1.18∗ −5.87∗∗∗ (1.11) (1.11) (0.96) (2.10) (0.55) (0.70) (0.82)Obs. 12,709 10,290 10,290 10,290 1,580 10,290 10,284 10,353Adj. R2 52.11 55.19 55.23 74.67 54.83 94.22 59.17 42.69practices we assume that effective back-end loads are zero for classes B and Cshares. Column (1) in Table IV presents the results of estimating equation (3) withoutthe performance-sensitivity measure. If the negative relation between expectedperformance and fees were the consequence of the omission of variables (for
- 20. 2172 The Journal of Finance Rexample, size, age, or turnover) that are likely to determine operating costs andare related to performance, then we would expect the coefficient on expectedperformance to change sign, or at least to become statistically insignificant oncewe include these variables in the regression. As column (1) shows, this is notthe case: The coefficient on expected performance remains negative and sig-nificantly different from zero. Therefore, cost-based arguments cannot explainwhy fees and performance are negatively related. Column (2) reports the results of estimating the full model (3), in which weinclude both performance and performance-sensitivity as regressors. The neg-ative (and statistically significant at the 1% level) coefficient on performancesensitivity suggests that equity mutual funds strategically exploit a low elastic-ity of demand with respect to net performance to increase their fees. Therefore,our results extend the findings of Christoffersen and Musto (2002), which wereobtained for a cross-section of money market mutual funds, to the market foractively managed equity mutual funds, for a much larger sample, and witha more precise measure of performance sensitivity. However, the inclusion ofperformance sensitivity does not eliminate the negative association betweenexpected before-fee risk-adjusted performance and fees. On the contrary, theestimated coefficient in column (2) is not only negative and statistically signif-icant, but higher (in absolute value) than the estimated coefficient in column(1). Thus, elasticity of demand appears to be an important determinant of fees,but does not in itself explain why underperforming funds set higher fees. To clarify the economic significance of these results, the estimated coefficientsin column (2) imply that a one-standard deviation increase in annual before-feerisk-adjusted performance is associated with a decrease of 1.38 basis points inannual total ownership cost, while a one-standard deviation increase in perfor-mance sensitivity is associated with a 4.22-basis point decrease in fees per year.To put these figures in perspective, increases of one-standard deviation in fundvolatility and turnover are associated with increases in total ownership costof 1.99 and 4.79 basis points, respectively, while an increase of one standarddeviation in fund size is associated with a reduction in total ownership cost of23.21 basis points. To account for possible nonlinearities, in column (3) we include alpha squaredin the regression. The associated positive coefficient suggests that the relationbetween alpha and fees becomes f latter for funds with higher alphas, althoughit remains negative for plausible values of alpha. These results indicate that worse-performing funds and those whose in-vestors have a lower performance sensitivity charge higher fees. Because man-agement companies typically manage many funds, the results could be due todifferences between management companies, differences within managementcompanies, or a combination of both. Column (4) of Table IV reports the resultswhen we estimate equation (3) including a management company fixed effectto capture time-invariant differences between management companies. The es-timated coefficients indicate that differences in fees within management com-panies are negatively related to differences in alpha or performance sensitiv-ity. Thus, the results are consistent with management companies strategicallysetting the fees of their different funds to match the funds’ expected
- 21. Price and Performance in the Mutual Fund Industry 2173performance or the performance sensitivity of their investors.12 The large in-crease in the adjusted R2 of the regression with respect to the R2 obtained incolumns (1) to (3) also suggests that there are management company charac-teristics beyond the company’s total size and number of funds that are relatedto fees. (The joint hypothesis that all management company fixed effects arezero can be rejected at any reasonable significance level.) We also generate ob-servations at the level of the management company by taking asset-weightedaverages of all variables, except for complex size and the number of funds in thecomplex, for which we retain the management company totals. Column (5) ofTable IV reports the results of estimating equation (3) at the managementcompany level. For each management company we compute the asset-weightedaverage of each variable and exclude management companies for which totalnet assets of funds with information on the variable are less than 75% of thetotal net assets of the management company. We also exclude managementcompanies with more than one-third of assets under management in index orinstitutional funds or more than 10% of assets in funds categorized as outliers.The estimated coefficients for alpha and sensitivity indicate that the resultsobtained at the fund level extend to the management company level. To evaluate whether marketing or nonmarketing fees are responsible for thenegative relation between total fees and performance, we reestimate equation(3), replacing the dependent variable with marketing and nonmarketing fees,alternatively. We define marketing fees as the sum of front- and back-end loadsdivided by seven (except for B and C share classes, for which we only add front-end loads) and 12b-1 fees. We define nonmarketing fees as the expense ratiominus 12b-1 fees. Columns (7) and (8) of Table IV show that both marketing andnonmarketing fees are negatively related to f low-performance sensitivity andto risk-adjusted performance.13 Therefore, the results are consistent with thethree strategic explanations discussed in Section IV.B. In particular, the resultsobtained for marketing fees suggest that marketing variables, such as distribu-tion channel or advertising expenditures, play a significant role in determiningmutual fund fees and contribute to explaining the negative fee-performancerelation. Although our results are consistent with the strategic fee-setting hypothe-ses, the negative relation between before-fee performance and fees could alsobe caused by unobserved fund characteristics that have correlations of oppo-site signs with fees and before-fee risk-adjusted performance. To control fortime-invariant unobserved heterogeneity and to shed some light on the deter-minants of fee changes, we also estimate the fee equation (3) with fund fixed 12 Our estimates with management company fixed effects might also pick up longitudinal vari-ation at the management company level. To isolate the cross-sectional variation, we also estimateyearly cross-sectional regressions with management company fixed effects and obtain coefficientestimates and standard errors (see the Internet Appendix) for the whole sample period, using theFama–MacBeth procedure. The resulting coefficients for alpha and performance sensitivity are neg-ative, statistically significant, and only slightly smaller (in absolute value) than the correspondingcoefficients obtained without management company fixed effects. 13 Given the large incidence of funds with zero marketing fees, we also estimate a Tobit model andobtain very similar coefficients and standard errors (results available in the Internet Appendix).
- 22. 2174 The Journal of Finance Reffects. Column (6) of Table IV shows that the signs of the coefficients on perfor-mance and performance sensitivity remain negative in the fixed effects specifi-cation. Thus, funds appear to alter their fees over time in response to changesin performance in the same direction found in the absence of fund fixed ef-fects.14 However, the coefficients on performance and performance sensitivityare smaller in absolute value than those in column (2) and the latter coefficientis not statistically different from zero at conventional significance levels. Thesmaller value and reduced statistical significance of the estimated coefficientson performance and performance sensitivity in the fund fixed effects specifi-cation could be due to fees being largely a matter of long-term strategy, or tothe existence of adjustment costs that create substantial inertia in fees. Thedifferences between the specifications could also be due to the fact that ourperformance and performance sensitivity measures are estimated variablesthat are likely to contain substantial measurement error. This measurementerror may have a larger impact on coefficient estimates when we use only time-series variation to estimate them. For example, in the extreme case in whichfund managers’ skills are constant over time, the entire time-series variation inalphas would be due to estimation error. At the same time, the cross-sectionalvariation in estimated performance would allow us to pick up at least part ofthe effect of the true alpha. We note that there is a possible alternative explanation for the negative rela-tion between fees and before-fee performance that we do not explicitly considerin our analysis. Higher fees could be paying for better tax management orother fund services, such as check writing, web or telephone services, or bettershareholder statements, that compensate investors for differences in after-feeperformance. Although our data do not enable us to rule out this explanation,there are reasons to believe that its ability to explain our findings is limited.First, the explanation is valid only if the value of these services is negatively re-lated to fund performance. Moreover, if better fund services or more efficient taxmanagement fully compensated investors for lower after-fee performance, thenthere would be no reason to expect investors’ money to f low into funds withhigher after-fee performance. The explanation also finds little support fromthe empirical evidence on both the perceived and actual value of fund services,which seem to be limited (Capon, Fitzimmons, and Prince (1996), Elton, Gruber,and Busse (2004), Christoffersen, Evans, and Musto (2005), Bergstresser,Chalmers, and Tufano (2009)). The fact that we find a negative relation be-tween fees and performance when we include fund fixed effects also casts doubton the plausibility of this interpretation, unless the level of services changesover time in a direction opposite to performance.F. Fund Governance and Fees To analyze the role played by fund governance in the determination of fees, wecreate dummy variables corresponding to each one of the Morningstar board 14 In tests available in the Internet Appendix, we also estimate the fee equation in differences,as well as logit and probit regressions for the probability of fee changes, and obtain similar results.
- 23. Price and Performance in the Mutual Fund Industry 2175quality grades. Because only one fund in our sample received Morningstar’slowest board quality rating, we create a single dummy for the Very Poor andPoor categories, which we refer to as “Poor.” We also create interaction variablesof board-quality dummies with before-fee risk-adjusted performance and withour f low-to-performance sensitivity measure. In Table V, we present our regression results. Because of the limitationsof our governance subsample, discussed in Section I, these results should beinterpreted with caution. First, the fact that our governance information cor-responds to the end of the sample period makes a causal interpretation of thegovernance coefficients problematic and implies that the results may be sub-ject to survivorship bias. Further, the governance subsample is not a randomsample of the population. As we describe in Section I, funds in the governancesubsample tend to have lower-than-average fees and higher-than-average per-formance. Keeping these caveats in mind, the results in columns (1) to (5) ofTable V provide some support for the hypothesis that better board quality leadsto fees that better ref lect the value generated by mutual funds, and that betterboard quality also limits the ability of management companies to extract rentsfrom performance-insensitive investors. Column (1) presents the results of estimating the fee equation when we ex-clude the performance and performance sensitivity terms, but allow the inter-cept to depend on the board quality grade. In the table, we omit the dummyvariable for the Poor grade, so coefficients associated with Fair, Good, or Ex-cellent grades represent differences with respect to funds with a Poor grade.Funds with Good and Excellent grades are associated with lower total own-ership costs than Poor and Fair funds, although the difference between thecoefficients associated with Excellent and Poor grades is not statistically dif-ferent from zero. Column (2), which reports the results of estimating the fullmodel, also implies that for most values of before-fee alpha or performance sen-sitivity in the sample, funds with Good or Excellent grades are predicted to becheaper, all other things equal, than funds with Poor or Fair grades (althoughthis conclusion cannot be directly inferred from the table because the differencebetween the fees of two funds identical in every respect except for their boardquality grade is not just captured by the intercept, but depends on the levelsof alpha and performance sensitivity). Thus, columns (1) and (2) suggest thatbetter board quality is associated with lower total ownership costs. However,although in no case are the total ownership costs of a fund with a higher boardquality grade significantly higher than those of a fund with a lower grade, someof the estimated differences are not statistically significant and the levels of theestimated coefficients are not strictly monotonic in board quality. If better board quality brings fees more in line with performance, then thecoefficient on before-fee performance should be negative for poorly governedfunds and the interaction coefficients should be positive and increasing in boardquality. The estimated coefficients on performance in column (2) are consistentwith this hypothesis. The coefficients on performance sensitivity point in thesame direction, but in this case none of the differences is statistically significantat conventional significance levels.
- 24. 2176 The Journal of Finance R Table V Fund Governance and FeesThe table reports estimated coefficients for the pooled OLS regression of funds’ fees on selected fundcharacteristics in the 1993 to 2005 period, except for columns (6) and (7), where the sample periodis 2003 to 2005. In columns (1) and (2) the dependent variable is fund total annual ownershipcost (TOC), computed as total loads divided by seven plus the annual expense ratio. In column(3) the dependent variable is the asset-weighted average at the management company level oftotal annual ownership costs. All regressors in column (3) are also asset-weighted averages at themanagement company level of the corresponding fund-level variables, except for the size of themanagement company (Co. Size) and the number of funds in the management company (# funds),which are management company totals. In column (4) the dependent variable is marketing fees(Mark.), defined as total loads divided by seven plus 12b-1 fees. Back-end loads are assumed tobe zero for share classes B and C. In columns (5) and (7) the dependent variable is nonmarketingfees (N-Mark.), computed as the expense ratio minus 12b-1 fees. Column (6) shows results formanagement fees (Mgmt.). σt is the standard deviation of the fund’s monthly returns in year t. αtis year t’s four-factor alpha. St denotes the slope of the estimated f low-to-performance relation.Fair, Good, and Excell. are dummy variables that take a value of one if the observation has Fair,Good, or Excellent board quality grade, respectively, and zero otherwise. All regressions includeyear dummies and dummy variables for the different investment objectives and share classes. Allfees are expressed in bp. The table also reports robust standard errors (in parentheses) clusteredby fund in all columns except column (3), in which they are clustered by management company. Thetotal number of observations and the adjusted R2 of the regression (in percentage) are reported atthe bottom of each column. ∗ , ∗∗ , ∗∗∗ indicate statistical significance at the 10%, 5%, and 1% levels,respectively. Mgmt. N-Mark. TOC TOC TOC Mark. N-Mark. (2003–2005) (2003–2005) (1) (2) (3) (4) (5) (6) (7)Sizet−1 −6.87∗∗∗ −7.93∗∗∗ −6.14∗∗∗ −1.83∗∗∗ −0.46 −1.85∗∗∗ (0.61) (0.98) (0.97) (0.55) (0.44) (0.53)Aget−1 −2.98∗∗∗ −10.50∗∗∗ −0.53∗∗ −1.43 −9.07∗∗∗ −6.01∗∗∗ −7.21∗∗∗ (1.01) (1.69) (0.26) (1.64) (1.21) (1.19) (1.41)Co. Sizet−1 −1.16 0.53 −3.25∗ 7.87∗∗∗ −7.33∗∗∗ −4.40∗∗∗ −8.37∗∗∗ (0.77) (1.05) (1.72) (0.95) (0.63) (0.72) (0.82)# fundst−1 −0.16∗ −0.37∗∗∗ −0.20 −0.71∗∗∗ 0.34∗∗∗ −0.02 0.44∗∗∗ (0.09) (0.12) (0.33) (0.12) (0.07) (0.07) (0.08)Turnovert−1 −0.58 −2.64∗ −2.05 −5.21∗∗∗ 2.58∗∗ 1.09 1.29 (1.12) (1.48) (4.19) (1.52) (1.06) (1.03) (1.25)σ t−1 −5.51 3.82 41.91 −5.52 9.25 29.99∗ 38.60∗∗ (10.95) (16.12) (35.98) (15.77) (10.46) (17.11) (17.45)Fair 4.77 5.03 −2.33 19.75∗∗∗ −14.69∗∗∗ −12.98∗∗∗ −17.06∗∗∗ (4.50) (5.81) (21.28) (4.52) (4.39) (3.06) (4.38)Good −9.83∗∗ −17.12∗∗∗ −19.03 −1.94 −15.05∗∗∗ −9.40∗∗∗ −19.06∗∗∗ (4.40) (5.93) (19.17) (4.73) (4.19) (3.17) (4.22)Excell. −7.95 −13.74∗ −38.63∗ 5.72 −19.33∗∗∗ −10.69∗∗∗ −19.52∗∗∗ (5.11) (7.09) (21.36) (5.88) (4.88) (3.76) (5.00)αtˆ −50.70∗∗ −104.83 −2.06 −48.57∗∗∗ 9.12 −17.50 (22.70) (116.06) (15.41) (15.08) (24.14) (40.89)αt Fairˆ 28.15 35.53 −42.82∗∗ 71.05∗∗∗ −3.01 67.03 (24.52) (116.67) (18.42) (16.81) (27.19) (44.59)αt Goodˆ 57.35∗∗ 157.18 1.90 54.74∗∗∗ 2.62 61.96 (24.27) (114.03) (17.24) (15.72) (28.92) (43.64) (continued)
- 25. Price and Performance in the Mutual Fund Industry 2177 Table V—Continued Mgmt. N-Mark. TOC TOC TOC Mark. N-Mark. (2003–2005) (2003–2005) (1) (2) (3) (4) (5) (6) (7)αt Excell.ˆ 61.63∗∗ 106.99 19.90 41.32∗∗ −72.74∗∗ −41.67 (26.42) (112.55) (18.60) (17.54) (35.48) (51.16)St −8.59 −7.71 6.01∗ −14.91∗∗∗ −6.08 −17.22∗∗∗ (5.37) (18.35) (3.52) (4.08) (3.98) (4.37)St Fair 3.60 −12.43 −9.15∗∗ 12.99∗∗∗ 6.14 17.52∗∗∗ (5.94) (19.30) (4.37) (4.36) (4.19) (4.69)St Good 7.31 6.28 −3.09 10.56∗∗ 4.11 12.90∗∗∗ (5.70) (18.05) (4.00) (4.23) (4.33) (4.58)St Excell. 4.97 15.06 −12.21∗∗ 17.14∗∗∗ 10.31∗ 16.08∗∗∗ (6.50) (19.28) (4.81) (4.93) (5.43) (5.74)Obs. 7,767 3,933 545 3,933 3,971 1,734 1,734Adj. R2 58.58 67.39 68.64 65.73 46.59 39.51 39.40 Because as mentioned earlier, funds managed by the same management com-pany frequently share the same board, we also reestimate the fee regressionat the management company level by taking asset-weighted averages of thevariables. Column (3) in Table V reports that the signs of the coefficients ofinterest are the same as in the previous column, although no coefficient (otherthan the intercept for funds with an Excellent grade) is significantly differentfrom zero (a result that is not due to a reduced magnitude of the coefficientsbut to much larger standard errors). Tufano and Sevick (1997) argue that the effect of fund governance shouldbe most visible for management fees, since for these fees we can expect theconf lict of interest between the management company and shareholders to bemost severe. However, these authors find little difference in how board char-acteristics are related to marketing and nonmarketing fees. We replicate theanalysis carried out in columns (1) and (2) with marketing and nonmarketingfees as dependent variables, although we report only the estimated coefficientsof the full model in columns (4) and (5) of Table V, respectively. The resultsfor marketing fees are mixed. Funds with Good and Excellent grades chargelower fees (although in column (4) the intercept for funds with an Excellentgrade is marginally higher than that for funds with a Poor grade, the formerare cheaper on average, other things equal) and have the highest coefficients onbefore-fee performance. However, the funds with a Fair grade—not those witha Poor grade—are the ones that charge the highest fees and have the lowestcoefficient on before-fee performance. Therefore, there is no monotonicity at thebottom in the relation between board quality and marketing fees. Further, thedifferences between funds with a Poor grade and funds with a Good or Excel-lent grade are not statistically significant. Finally, the sensitivity coefficientsfollow no clear pattern: The lowest (and negative) coefficient corresponds tofunds with the highest board quality grade. The results for nonmarketing fees are more in line with those in columns(1) and (2). When we estimate the same specification as in column (1) (see the
- 26. 2178 The Journal of Finance RInternet Appendix for the results), we find that other things equal, funds witha Poor grade charge the highest nonmarketing fees, and funds with Good andExcellent grades charge the lowest nonmarketing fees (which is the same or-dering as the one implied by the intercepts in column (5)). The relation betweenperformance sensitivity and nonmarketing fees becomes f latter as board qual-ity improves: The sensitivity coefficient is −14.91 for funds with a Poor gradeand the estimated differences imply that the same coefficient for funds with anExcellent grade is 2.23. However, further tests indicate that this coefficient isnot significantly different from zero. Finally, funds with a Poor grade have anegative coefficient on performance. This coefficient is significantly lower thanthe coefficients for all other grades, although the coefficients for the highergrades are not increasing in board quality. Therefore, the results are consistentwith the hypothesis that better-governed funds resist attempts by manage-ment companies to exploit investors’ low performance sensitivity. The resultsalso provide weaker evidence for the hypothesis that better board quality bringsnonmarketing fees more in line with fund performance. Since CRSP started providing data on management fees as of 2003, we canperform the analysis with management fees as the dependent variable for the2003 to 2005 period. When we estimate a specification such as the one in col-umn (1), we again find that the worst-governed funds charge the highest fees,as suggested by the ordering of the intercept coefficients reported in column(6). Further, the differences between the sensitivity coefficients are similar tothose in column (5), although the statistical significance of the coefficients ismarginal at best. However, the performance coefficients follow no clear pattern:The coefficient for the best-governed funds is not only negative but lower thanall other coefficients, a result that is in stark contrast with those in columns(1) to (5). To check whether this result is due to a different behavior of manage-ment fees or to the different sample period, we reestimate the equation withnonmarketing fees, which are the closest approximation to actual managementfees, as the dependent variable for the 2003 to 2005 period, and report the re-sults in column (7) of Table V. Although the size and statistical significance ofthe performance coefficients vary, the ordering of these coefficients is the samein columns (6) and (7), showing that the best-governed funds have the lowestperformance coefficient in both columns. We cannot explain this result. It maybe due to an effort by the best-performing among the best-governed funds tolower their fees to compensate mutual fund investors for the decline in the stockmarkets during the 2000–2002 period, or to a response by these funds to theseries of mutual fund scandals that emerged in 2003 and 2004. However, theshort sample period, the noise inherent in our performance measure, and thefact that the only significant differences between the full sample and the 2003to 2005 period have to do with the coefficients on performance suggest cautionwhen interpreting the differences as indicating a structural change. V. Conclusion In this paper we show that there is a negative relation between funds’ before-fee performance and the fees they charge to investors. Since this evidence is at
- 27. Price and Performance in the Mutual Fund Industry 2179odds with economic intuition, we subject it to a series of robustness tests, andfind that it survives all of them. Next, we propose two explanations for this anomalous result. According tothe first explanation, the negative relation is the consequence of factors, whichare omitted in univariate regressions, that are both positively correlated withreturns and negatively correlated with funds’ operating costs, and thus alsowith fund fees. According to the second explanation, in contrast, the nega-tive relation is the result of funds that strategically set fees as a function oftheir past or expected performance. We consider three related rationales forthis strategic behavior. The first, proposed by Christoffersen and Musto (2002),argues that funds with worse past performance have a pool of investors thatare less sensitive to fund performance. Faced with an inelastic demand fortheir shares, underperforming funds optimally increase fees. The second expla- ´nation, proposed by Gil-Bazo and Ruiz-Verdu (2008), argues that funds withlower expected performance optimally set higher fees and target performance-insensitive investors, since these funds anticipate that they will not be ableto compete with better-performing funds in the market for sophisticated in-vestors. On the other hand, better-performing funds keep fees low because ofcompetition among them for the money of performance-sensitive investors. Thethird explanation argues that funds with different expected performance choosedifferent marketing strategies. Funds with low expected performance are mar-keted to performance-insensitive investors and have higher distribution costs,which translate into higher fees. The empirical analysis finds support for all strategic-pricing explanations.Even though funds’ operating costs are important determinants of fees, theydo not explain away the negative relation between before-fee performance andfees. When we control for those cost determinants, we find that funds withlower expected before-fee performance and funds with less elastic demandscharge higher marketing and nonmarketing fees. Therefore, it appears thatmutual fund competition and regulation have not been sufficient to ensure thatfees ref lect the value that funds create for investors. However, we find someevidence that better fund governance may be associated with fees that are morein line with performance: Among the best-governed funds, worse performanceneed not mean higher fees. Appendix: Estimation of Flow-to-Performance Sensitivity We define annual net f low to fund i in year t, Flowit , as the relative growthof the fund’s total net assets (TNA) adjusted for returns net of expenses, Rn :15 it TNAit − TNAit−1 1 + Rit n Flowit = . (A1) TNAit−1 15 Elton, Gruber, and Blake (2001) report a number of errors associated with mutual fund mergersand splits in the CRSP sample. Huang, Wei, and Yan (2007) argue that these errors could lead toextreme values of f lows. We deal with this problem by eliminating the 1% of observations with thelowest and highest f lows in each year.