VIEWS: 10 PAGES: 45 POSTED ON: 8/31/2011 Public Domain
Can Mutual Fund Families Aﬀect the Performance of Their Funds? Ilan Guedj and Jannette Papastaikoudi∗ MIT - Sloan School of Management First Draft - August 2003 This Draft - January 2004 Abstract We examine whether mutual fund families aﬀect the performance of the funds they manage. From a sample of funds belonging to large families we ﬁnd that last year’s best performing funds outperform last year’s worst performing funds by 58 basis points. We also show that there exists persistence of performance of these funds inside their respective families. This persistent excess performance is related to the number of funds in the family which we interpret as a measure of the latitude the family has in allocating resources unevenly between its funds. Supporting these ﬁndings, we also show that the better performing funds in a family have a higher probability of getting more managers, one of the main resources available. This is consistent with the view that fund families allocate resources in proportion to fund performance and not fund needs. ∗ We would like to thank John Chalmers, Joseph Chen, Denis Gromb, Dirk Jenter, Paul Joskow, Ajay Khorana, Jonathan Lewellen, Stewart Myers, Stefan Nagel, Anna Pavlova, Steve Ross, Antoinette Schoar, Henri Servaes, Peter Tufano, and the participants of the 2003 Transatlantic Conference at LBS for their helpful comments. Corresponding Author: Ilan Guedj, MIT Sloan School of Management, E52-442, 50 Memorial Drive, Cambridge MA, 02142, USA or email: guedj@mit.edu 1 1 Introduction The mutual fund industry has grown at an incredible rate in the past years, showing an ex- plosive increase particularly over the last decade. However, the recent revelations about the expropriation of investors’ trust and wealth by mutual fund families have shaken investors’ conﬁdence and have brought the mutual fund industry under a lot of scrutiny. So far, we have witnessed preferential treatment of certain clients by allowing market timing and after hours trading. A legitimate question arises whether there have been other forms of prefer- ential treatment that aﬀect directly the observable performance of mutual funds. Towards that end, we ask whether mutual fund families can and indeed do aﬀect the performance of their funds in a way that might systematically discriminate certain investors. One of the most extensively researched questions is whether mutual funds have persistent abnormal returns. If so, this provides evidence for the existence of superior managerial investment ability. In fact, persistence is well documented. Lehmann and Modest (1987), Hendricks, Patel and Zeckhauser (1993) and Wermers (1997), among others, have found evidence of persistence in fund performance over short horizons of one to three years. Yet, Brown and Goetzmann (1995) conclude that even if there is some predictability it is quite diﬃcult to detect. Carhart (1997), Daniel, Grinblatt, Titman and Wermers (1997), Wermers (2000) and Pastor and Stambaugh (2002) argue that most of this persistence is due to factors other than managerial ability. In particular they show that positive abnormal returns can be attributed to momentum in stock prices while negative abnormal returns can be attributed to managerial expenses and transaction costs. In this paper we view mutual fund performance from a diﬀerent perspective than the majority of the existing literature. Instead of treating a mutual fund as a completely independent entity, we view it as part of a larger group, the mutual fund family. Given this dependence, diﬀerences might arise between the objectives of the fund and the family it belongs to. Furthermore, these diﬀerences in objectives might potentially translate into discrepancies between expected and observed performance. In particular, if it serves the family’s interest, it could decide to follow a strategy of selectively allocating its limited resources unevenly across its funds. The rationale behind such a strategy follows from both the convexity of the performance-ﬂow relationship and ﬂow spillovers inside the family. Spitz (1970), Chevalier and Ellison (1997) and Sirri and Tufano (1998) have documented that abnormal positive returns generate disproportionately more inﬂows than abnormal 2 negative returns would generate outﬂows. This implies that if the family had a choice between managing two mediocre performing funds or managing one well performing fund and one poorly performing fund, the family would prefer the latter combination. Apart from this empirical ﬁnding, it has also been observed by Khorana and Servaes (2002) and Nanda, Wang and Zheng (2003) that there exists a ﬂow spillover within funds of families that possess at least one fund with an excellent performance record. The implication of this observation is equivalent to that of the convex performance-ﬂow relationship: It is suﬃcient for the family to only have some well-performing funds in order to experience a large inﬂow in its assets under management. We therefore expect families to want to promote their funds selectively1 . However, in order to act along these lines, the family needs to possess the latitude to do so, i.e. it needs to have enough funds to be able to move resources from one to another. Hence, we hypothesize that one should expect larger families to be more capable of aﬀecting the performance of their funds. To test our hypothesis, we use monthly open-end mutual fund data from the Center for Research in Security Prices (CRSP) for the period of 1990-2002. Using a methodology similar to Carhart (1997) we analyze the persistence of the performance of mutual funds that belong to large families, deﬁned either by the number of funds they hold or by their market capitalization. We ﬁnd a short-term persistence in mutual fund performance. The diﬀerence in abnormal returns between a portfolio of funds which were last year’s winners and a portfolio of funds which were last year’s losers is 58 basis points per month (statistically signiﬁcant at the 1% level), an annualized diﬀerence of 7.2%. In addition to these results, we perform the persistence methodology on a relative scale, i.e. when the funds are placed into portfolios with respect to their performance relative to other funds inside the same family. Again, we ﬁnd a statistically signiﬁcant diﬀerence between the top and bottom portfolio of 53 basis points per month, which translates into a diﬀerence of 6.5% per year in abnormal returns. The fact that persistence in fund performance is detected even within a mutual fund family can be viewed as evidence that families are actively intervening in their funds’ performance. 1 Some anecdotal evidence about selective promotion of funds has been provided in a case study by Loeﬄer (2003). He analyzes the case of the creation of a new class of pension funds. The three German mutual fund companies he follows enhanced the performance of the new funds by allocating underpriced IPOs to their portfolios. All that, in order to generate a favorable track record for the funds. 3 If the family is actively promoting some funds over others, it can accomplish this by unequal resource allocation. Managers (analysts) are one of the main resources families have both in actual terms and in the eyes of the investors. Chevalier and Ellison (1999a, 1999b) show that personal characteristics of the fund manager can help predict superior stock picking ability. Families seem to place serious consideration on this fact, since their decisions on managerial turnover is linked to the managers’ past performance. Khorana (1996, 2001) shows that a low performance of at least two years is necessary before a manager is removed from the management of the fund. He also shows that in the post-replacement period there is a signiﬁcant improvement in the performance of the fund. Hu, Hall and Harvey (2000) analyze promotions and demotions of managers and conclude that there exists a positive relationship between the promotion probability and lagged fund returns. All these ﬁndings indicate that mutual fund families take their fund management very seriously and are not reluctant to intervene in their funds’ management, if needed. Hence, we can expect to observe family intervention which can potentially lead to preferential treatment of funds. In order to proxy an an attempt from the family to push certain funds, we use the probability of adding another manager to a fund. If our hypothesis is wrong one would expect that after controlling for fund characteristics, there would no longer be any statistically signiﬁcant diﬀerence in the probability of adding a manager between funds in diﬀerent groups of performance ranking. After using controls such as size and expenses we ask if within rankings, the probability of adding another manager depends on abnormal returns and lagged abnormal returns. We ﬁnd that only funds in the top portfolio have a signiﬁcant positive coeﬃcient for their alpha, which would indicate that the decision to add a new manager depends positively on the fund’s performance only when it is among the top in its family. Second, we ask if the probability of adding a manager given the funds’ past returns depends on the ranking group of the fund, i.e. on its relative performance against its peers in the same family. We ﬁnd that the probability increases when a fund belongs to the top rank and decreases as the rank of the fund decreases. Lastly, we identify cases where only one manager was in charge of a fund and ask what the probability is of adding at least one new manager to such a fund. This question seems to be a more precise proxy for a deliberate action by the family, since the change from a single managed fund to a team managed fund is more substantial than adding another manager to an existing team. We ﬁnd that again, the probability increases when a fund belongs to the top rank and decreases when the fund belongs to the bottom 4 rank. The remainder of the paper proceeds as follows. In section 2 we develop our hypotheses and methodology and describe the data. In section 3 we present our results and provide alternative explanations. In section 4 we perform probit regressions to further enhance our results. We conclude in section 5. 2 Hypotheses, Data and Methodology 2.1 Hypotheses A mutual fund is not a stand alone entity, but belongs to a broader organizational structure, the family. The family could impact the decisions of the fund, and thus could potentially have a signiﬁcant eﬀect on the fund, its performance and the persistence of this performance over time. In fact, there are two main reasons for the family to want to inﬂuence that performance. The ﬁrst reason is the performance-ﬂow relationship. Chevalier and Ellison (1997) and Sirri and Tufano (1998) have documented the existence of a convex relation between lagged fund performance and present fund ﬂows: abnormal positive returns generate disproportion- ately more inﬂows than abnormal negative returns would generate outﬂows. This implies that if the family had the choice between owning two mediocre performing funds or one well performing fund and one poorly performing fund, the family would prefer the latter combination. This is the case because the convexity of the performance-ﬂow relationship would translate into an increase of the net amount of assets under the family’s management. Thus, it is reasonable to believe that faced with some better performing funds than others, the family would consciously choose to oﬀer preferential treatment to the better ones in order to maintain their good track record, even if this would come at the expense of other poorly performing funds in the family. The convex performance-ﬂow relation is not the only reason why a family would want to inﬂuence the performance of its funds. The second reason is that there exists evidence that ﬂows are not independent across funds of the same family, on the contrary, there seems to exist a strong cross-sectional dependence. Khorana and Servaes (2002) ﬁnd that a fund’s market share within an investment objective is not only driven by its family’s policies within that objective, there are important spillover eﬀects from other funds within the same family 5 as well. Nanda, Wang and Zheng (2003) ﬁnd that there is a strong positive spillover from a star performer to other funds within the same family. This spillover results in greater cash inﬂows not only to the star fund but to other funds in the family as well. As such, the cross sectional correlation between fund ﬂows can increase aggregate inﬂows in a non-linear fashion and could be another incentive for the family to ”push” the better performing funds in order to maintain their performance. Therefore, both the convexity of the ﬂow-performance relation and the cross sectional correlation between fund ﬂows imply that maximizing aggregate ﬂows to the entire family is not necessarily equivalent to maximizing ﬂows to each and every fund individually. In fact, and due to the cost of fund management, such a policy could even be sub-optimal. Instead, it seems to be suﬃcient for a family to have a few funds that persist at outperforming their peers; this would lead to increased ﬂows to the entire family while being a more cost eﬃcient strategy. Such a strategy that focuses on improving performance selectively and not to the entire universe of funds within a family might have direct implications on the observed behavior and performance of individual funds. The relevant question is which funds to promote in order to obtain the desired aggregate inﬂow eﬀect to the family. Since investors look at past performance when deciding about the allocation of their wealth between funds2 , a family, if capable, would deliberately attempt to ”push” its currently well performing funds also in the following year (after they had demonstrated good performance) in order for them to further increase their returns and therefore increase ﬂows to the entire family3 . The family has several resources it can use in order to achieve this goal of ”pushing” certain funds at the expense of others. The existence of a centralized research department, the waiving of management fees, and the ability to move managers from one fund to another or even share them among funds are only some of those resources that can be used and allocated not equitably between funds. Yet, even if it may be in the best interest of every family to act along these lines, not all are capable of adopting a fund promoting strategy, the main inhibition being the availability as well as the ﬂexibility and latitude in allocating those resources. Evidently, larger families that oﬀer a larger number and a greater variety of funds, have more ﬂexibility in using their 2 Wilcox (2003) runs an experiment with investors and ﬁnds that performance is one of the most important determinants of investor choice. 3 In fact, Nanda, Wang and Zheng (2003) ﬁnd that the empirically documented spillover eﬀect from a star fund to other funds may induce lower quality families to pursue a star creating strategy. 6 resources in order to promote certain funds. Therefore, these families should exhibit a more pronounced fund-promoting behavior. We thus hypothesize that larger families are better equipped and are expected to act more along these lines than smaller families. Hence we expect to obtain two main ﬁndings. First, that funds belonging to larger families have a more persistent performance than funds belonging to smaller families. And second, given the strategy of the family to promote only a few of its funds, we expect to ﬁnd persistence in fund performance also within a family. For example, even in a family where there isn’t even one fund that performs well when measured in absolute terms, we still expect the family to ”push” its relatively better funds to persist and improve. 2.2 Data Formation and Descriptive Statistics 2.2.1 Data Our data originates from the Center for Research in Security Prices’s ”Survivor-Bias Free US Mutual Fund Database” (CRSP). This database is considered free of survivorship bias since it includes funds that no longer exist4 . Unfortunately it contains information on the fund families only from 1992 onwards. Therefore, although the CRSP mutual fund database is our primary source of data, we use complementary information to complete our study. Since we focus on the mutual fund family, we need to obtain accurate information on the families the funds belong to. Even though CRSP oﬀers names of the fund family after 1992, it is often not consistent in their documentation across time and across funds. The implications thereof might aﬀect adversely an analysis regarding families. As a precautionary measure we revert to the use of an alternative mutual fund database that has been commonly used in other mutual fund studies5 : Morningstar. We use the ”Morningstar Principia Pro”’s CDs (Morningstar) for the years 1990-2002 to cross-check our information on the fund families of our sample. This also allows us to extent the time period to 1990. If a fund does not possess such information it is dropped out of our sample. In a few cases where Morningstar does not provide the identify of the fund family, while CRSP possesses the information, the latter source is used. 4 Though, some studies have disputed whether CRSP is indeed survivorship-bias free. See Elton, Gruber, and Blake (2001) for an extensive analysis of this issue. 5 For example; Brown and Goetzmann (1997) Chevalier and Ellison (1999a, 1999b) and Hu, Hall, and Harvey (2000) among others. 7 In order to compare our results with the outstanding literature we keep only actively managed equity funds, hence funds that have an objective of Growth, Aggressive Growth, Growth and Income, Small Company and Equity Income6 . Therefore, we exclude Sector, International, Balanced, Bond, and Municipal funds. In addition, for a fund to enter our sample an additional requirement of one year of past reported returns is imposed. We use Morningstar as a consistent source for identifying the objectives of the funds. For the time period of interest we record 7,310 funds that meet these criteria. Next, we correct our sample for multiple share classes7 . To construct return series for the funds after share class aggregation, we value weight the returns of each individual class by class size. This reduces our sample to 3,046 funds. In our sample, each fund possesses on average 2.32 classes, 48.4% of the funds only have one share class. From those funds that possess more than one class, they possess on average 3.52 share classes. Our ﬁnal sample incorporates 678 families across the years 1990-2002. Not all of them exist throughout the entire sample. Acquisitions and mergers across families lead to some variation in the number of families appearing each year. 2.2.2 Descriptive Statistics In table 1 we provide annualized summary statistics of our sample. As one can note, the number of funds has more than tripled between 1990 and 2000, while remaining almost at a constant level during the recent years. This might be well due to the large amount of consolidation in the mutual fund arena in the past few years, as well as the recent low performance in the equity markets. The same pattern seems to appear for the average size of assets under management. The average amount of assets under management has more than quadrupled between 1990 and 2000. The table also includes some fund characteristics such as loads, fees and turnover. Loads are fees charged to the investor by the mutual fund, either when the fund is purchased (front-end load), when it is sold (rear-end loads) or depend on the amount of time the investor holds the fund (deferred loads). 12-B1 is an annual distribution charge (12b-1 fee) charged to the investor for marketing needs of the funds. Loads and fees show little to no variation across time in contrast to turnover that 6 Funds with these objectives have been identiﬁed by Morningstar as pure equity funds. 7 Mutual funds very often oﬀer diﬀerent types of shares supported by the same underlying portfolio, that diﬀer only in the fee structure imposed to the investor. These diﬀerent types are labeled as share classes and given that they correspond to the same underlying portfolio, they have highly correlated returns. 8 has been increasing on average over the years. Table 2 contains summary statistics of the family across years. As in the case of indi- vidual mutual funds, there is a steady increase in the number of fund families, yet slightly dropping over the recent years, consistent with family consolidations. Though, in contrast to the observed trend in the number of families, the average number of equity mutual funds per family has been monotonically increasing over time reaching 5 funds per family in 2002. Total net assets on the family level exhibit the same pattern as total assets on the fund level, hence indicating that the recent drop in value is due to the downturn in the ﬁnancial mar- kets. Other parameters such as expense ratios, fees and loads do not exhibit a substantial variation in time. 2.3 Measuring Fund Persistence In order to test our hypotheses we perform a series of tests based on the standard mutual fund persistence methodology8 .We restrict our attention to a sub-sample of funds that belong to families with a larger number of funds (or assets) under management, in order to test our hypothesis that funds belonging to larger families are more persistent. To measure abnormal returns, we use a four factor model. We use the Fama and French (1993) three-factor model augmented with a momentum factor similar to Jegadeesh and Titman (1993). This model has been shown in various contexts to provide explanatory power for the observed cross-sectional variation in fund performance. We therefore apply the following multi-factor model: rit = αi + bi M KT RFt + si SM Bt + hi HM Lt + pi M OMt + it t = 1, 2, ..., T (1) where rit is the excess monthly return for fund i. MKTRF is the excess monthly return on the CRSP value-weighted stock index net of the one year Treasury-Bill. SMB and HML are the Fama-French (1993) factor mimicking portfolios for size and book to market. MOM, the momentum portfolio, is the equal-weighted average of ﬁrms with the highest 30 percent eleven-month returns lagged one month minus the equal-weighted average of ﬁrms with the 8 See for example Grinblatt, Titman, and Wermers (1997), Carhart (1997), and Wermers (2000) among others for a detailed description. 9 lowest 30 percent eleven-month returns lagged one month9 . At the beginning of every year, we form a sample of funds that belong to large families. A family is considered large if the number of funds it owns (or its market capitalization) exceeds a pre-speciﬁed threshold. Using prior twelve month returns, we regress each fund’s monthly returns on the four factor model (equation 1) to obtain each fund’s alpha over the prior year. Given this measure of performance we rank each fund by its alpha and assign the funds to one of 10 portfolios based on these rankings. The composition of these 10 rank- sorted portfolios remains unchanged for the following 12 months. Following the sorting procedure, a value weighted return series is calculated for each portfolio. This process is repeated every year and results in 10 time series which are then regressed on the four factor model. A comparison of the alpha of the top portfolio (the portfolio of mutual funds that had the highest alphas the year prior to the ranking) with the alpha of the bottom portfolio (the portfolio of mutual funds that had the lowest alphas the year prior to the ranking) gives an indication of the persistence of mutual fund performance. To further test our hypothesis that large families choose selectively which of their funds to promote, we devise a second test that is a variation on the standard mutual fund persis- tence methodology. At the beginning of every year, after selecting those funds that belong to large families, we regress their prior twelve month returns on the four factor model (equa- tion 1), thus obtaining each fund’s alpha. Then, we rank each fund by its alpha inside its family and build 10 portfolios based on these rankings; this implies that for every family of at least 10 funds each of its funds is allocated to one of the 10 portfolios. One has to stress that contrary to the standard methodology, the result is not a portfolio of mutual funds ranked by their absolute performance, but a portfolio of funds ranked by their relative performance inside their respective families. We repeat this procedure for every year and consequently obtain 10 time series for the portfolios which are subsequently regressed on the four factor model. We then compare the alpha of the top portfolio (the portfolio of mutual funds that had the highest alpha within their family the previous year) with the alpha of the bottom portfolio (the portfolio of mutual funds that had the lowest alpha within their family the previous year). The statistical and economic signiﬁcance of the diﬀerence of the alphas implies that funds exhibit persistence in their performance within their family. 9 We are grateful to Ken French for providing us with the SMB and HML factors, and to Mark Carhart for providing us with the MOM factor. 10 3 Empirical Results 3.1 Fund Persistence In table 3 we report the estimation results of the regression given in equation 1 for the 10 portfolios as outlined in section 2.3, while restricting ourselves to the sub-sample of funds that belong to large families, i.e. families with 10 funds or more10 . Since in this paper we take the viewpoint of the mutual fund family, we report our results using gross returns, however, the same results hold for net returns too. The same characteristics of multi-factor models regarding mutual fund performance reported in the literature also hold in our results. First, the R-squared are above 90% for almost all portfolios. Second, beta is signiﬁcant at a 1% level for all portfolios, and so is the loading on SMB. The loading on HML is signiﬁcant for most portfolios at the 1% signiﬁcance level, except for portfolios 7 and 8. These are by now standard results outlined by Lehmann and Modest (1987) and many others, that a three factor model explains the main part of the return of mutual funds. Third, the momentum factor is only signiﬁcant for portfolios 9 and 10, also a well documented result. One implication is that funds in portfolios with the highest alphas invest in momentum stocks, which funds in the lesser performing portfolios seem not to follow. However, beyond these standard results, the ﬁndings of our family oriented methodology diﬀer from the standard results when analyzing the alphas of the portfolios. As can be seen in table 3, we ﬁnd that the top portfolio has a positive monthly alpha of 35 basis points, and the bottom portfolio has a negative monthly alpha of -23 basis points. Thus, the portfolio that consists of longing the top portfolio and shorting the bottom portfolio (referred to as the 10-1 spread in all the tables) has a positive monthly alpha of 58 basis points (signiﬁcant at a 1% level). The statistical signiﬁcance becomes even more important when considering the economic signiﬁcance of this diﬀerence, a monthly alpha of 58 basis points is equivalent to an annual abnormal return of 7.2%. In table 4 we give the analogous results using the full sample11 .The spread in alphas between deciles 1 and 10 is estimated at 35 basis points, and is statistically signiﬁcant only at the 10 percent level. These results for persistence are relatively weak. The marginal statistical signiﬁcance of the diﬀerence between winner and loser portfolios of funds has 10 An analysis of diﬀerent deﬁnitions of large families is given in section 3.3. 11 These results are consistent with the standard results in the literature, such as Carhart (1997) and seem to be robust to the diﬀerent time period used in this study. 11 lead the literature to conclude that there is no true managerial ability in actively managed mutual funds after accounting for the known risk factors in the market. Tables 3 and 4 are directly comparable. One notes immediately the larger diﬀerence in alphas between deciles 1 and 10 for the full sample which is estimated at 35 basis points per month with a t-stat of 1.75 compared to 58 basis points per month with a t-stat of 2.86. This statistically signiﬁcant diﬀerence of 23 basis points per month is the result we expected to see given our hypothesis in section 2.1, bigger families seem to be able to maintain a better persistence of their funds’ performance. The loadings on the factors are qualitatively similar when comparing the two tables, although the loadings on the momentum factor are higher in table 3, which implies that the winner funds in this sub-sample seem to hold more momentum stocks than the funds in the full sample. Since these results are quite diﬀerent from the standard persistence results, we start by analyzing the 10 portfolios to see if they could provide us with some insight for these results. Table 5 Panel A reports the descriptive statistics of the 10 portfolios using the full sample and table 5 Panel B reports the descriptive statistics of the 10 portfolios using the sub- sample. As in the case of the regression results, we can make a direct comparison between the composition of the portfolios when considering the full sample and the sub-sample. The ﬁrst thing one can notice is that the average and median fund in each rank of the sub-sample is consistently bigger than the average and median fund in the equivalent rank of the full sample. This observation is consistent with the results of Chen, Hong, Huang, and Kubik (2002) who show that even though fund size can adversely aﬀect performance, family size may actually improve performance. Within each panel and across portfolios one can observe again the fact that the funds in the top and bottom portfolio are smaller than the funds in the other portfolios; again in accordance to the ﬁndings of Chen, Hong, Huang, and Kubik (2002). This is also consistent with the insight of Berk and Green (2002) who claim that performance should deteriorate with an increase in ﬂow and a consecutive increase in fund size. The expense ratio and the loads don’t vary much across portfolios, although the loads are higher in terms of mean and median than the loads on the whole sample. One might suspect that one reason for the better performance of the sub-sample might be the higher loads the funds charge, which deter investors from leaving or entering the fund and hence avoid in- or out-ﬂows that have an adverse eﬀect on fund performance (Edelen (1999)). To answer this, we calculate the average net ﬂow in each portfolio and show that the mean 12 and median net ﬂow of each portfolio doesn’t diﬀer signiﬁcantly across the full sample and the sub-sample as reported in panels A and B of table 5. This is inconsistent with the idea that funds in the sub-sample that also charge higher fees should experience less ﬂows than funds in the entire sample. Another possible alternative explanation to our results could be that the abnormal returns of the top portfolio could be due to higher spending on non observable resources12 that contribute to the abnormal returns. Table 5 Panel B shows that the average expense ratio of all the portfolios are quiet similar. The top portfolio has an average expense ratio of 0.0139 while the average expense ratio of the bottom portfolio is 0.0135. In addition, our persistence results also hold when using net returns. The 10-1 spread portfolio has a positive alpha of 53 basis points, signiﬁcant at a 1% level. This shows that our results are not driven by expenses. To see how long lived this fund persistence is, we perform the same methodology with a varying estimation time period. We conclude that the above described persistence holds only in the short term, observable at a one year horizon. With an estimation period of 2 years the 1-10 spread is reduced to 17 basis points per month.Using a 3 years estimation period the statistical signiﬁcance is lost. One alternative explanation to our results could be the presence of incubated mutual funds in our sample. By incubating funds for a short period, in eﬀect the family builds up a considerable track record for its funds so that they show exceptional performance before being launched to the public13 . There are several reasons why, although appealing, this idea isn’t applicable to our results. First, the number of incubated funds is not signiﬁcant compared to the number of funds in our sample. Evans (2003) researches all the SEC ﬁlings for 1995-2003 and ﬁnds only 60 such funds for the entire 9 years that were eventually launched. To put it in perspective, each of our portfolios includes every year on average more than 150 funds. Second, incubated funds have a very small size. Arteaga, Ciccotello, and Grant (1998) ﬁnd their average size to be $5 million. In our sample, the average size of a fund in our top portfolio is $840 million which counters the notion of small funds entering the portfolios. Third, Arteaga, Ciccotello, and Grant (1998) and Wisen (2002) seem to conclude that incubation lasts one year. Since Evans (2003) ﬁnds that post incubation funds do not perform better than other funds, thus, it seems clear that our predictions 12 Resources such as higher investments in research or in more capable and expensive fund managers. 13 This tactic could allow the discontinuation of poor performing funds and thus capture only the positive side of the ﬂow-performance relationship. 13 wouldn’t be aﬀected since in our methodology the presence of incubated funds would aﬀect only the estimation period and not of the predictive period. According to our hypothesis, only families that have the ability to allocate resources will do so, and by using the number of funds as a proxy for this latitude we ﬁnd persistence in fund performance. However, if we use the market capitalization of the family (although highly correlated with the number of funds in the family) we expect to get weaker results since even if it is a good proxy for the existence of resources we believe it is not that good a proxy for the family’s latitude to allocate them in an unequal way between its funds. Market capitalization is a much cruder measure, since a family with one big fund has no latitude in allocating its resources compared to a family with 5 small funds. In table 6 we report the results of the same methodology but instead of using the number of funds as the proxy for family size we use the market capitalization of the family. We construct a sub-sample based on family size, where the threshold to family capitalization will yield the same sub-sample size as the prior analysis (that used the number of funds within the family), i.e. we use a threshold level such that the resulting sub-sample has the same number of funds as was generated by a threshold of 10 funds or more per family. The results in table 6 corroborate our hypothesis. The 10-1 spread is 35 monthly basis points statistically signiﬁcant only at the 10% level, similar to the full sample. This result reinforces two claims. First, that it is not the amount of resources the family has, but the latitude to allocate them unevenly. And second, that it is not the mere statistical artifact of reducing the sample size that generates the results but indeed the selection criterion. 3.2 Ranking within the family As mentioned in section 2.1, given our hypothesis we not only expect to ﬁnd an increased persistence in larger families but also that this persistence holds inside the family. Since the families have limited resources they are expected to favor certain funds over others. To test this, we thus perform a ”within” analysis as detailed in section 2.3. Before analyzing these results it is important to stress the main diﬀerence between the two methods and their alternative interpretation. This analysis diﬀers from what has been traditionally considered to be an analysis of persistence of mutual fund performance, since we do not look at the previous year’s absolute best performers and ask if they persist at doing so. Instead, we ask whether, by taking into account information about the family of the fund ex-ante, one 14 can make some predictive statements about the performance of the fund ex-post. In table 7 we report the estimation results of these regressions. To be consistent with section 3.1 we use the same sub-sample and report the results for gross fund returns. Again the results hold also for net returns. The characteristics of the multi-factor model are similar to the one described in section 3.1. The loadings on the market, HML, SMB, and momentum are similar in pattern and magnitude. However, beyond these standard results, the ﬁndings of this within family methodology are quite interesting. As can be seen in table 7, we ﬁnd that the top portfolio has a positive monthly alpha of 32 basis points, and the bottom portfolio has a negative monthly alpha of -21 basis points. Thus, the portfolio that consists of longing the top portfolio and shorting the bottom portfolio has a positive monthly alpha of 53 basis points (signiﬁcant at a 1% level), equivalent to an annual abnormal return of 6.5%. These results are quite striking since they show that there is predictability of mutual fund performance based on the ranking of these funds relative to other funds inside their respective families. We believe that these predictability results are even more indicative of the fact that the family has an instrumental inﬂuence on the future performance of its funds than the results of section 3.1. If this were not the case then the 10-1 spread should have been much smaller than the one in table 3. This sorting criterion introduces some randomness into the portfolio ranking, since our methodology removes funds from the top portfolio that performed well on an absolute scale, yet were not the best performers in their families, and in return adds funds that didn’t perform well on an absolute scale yet were the best performers in their family. Thus, it should have weakened the spread. The fact that it didn’t, shows that the ranking criteria is indicative of an important phenomenon. In section 4 we perform several tests to try and corroborate this claim. 3.3 Robustness Tests We perform several robustness tests in order to check that indeed the results are a reﬂection of our hypotheses and not a statistical artifact. 3.3.1 Fund Size Table 5 Panel B includes an interesting result. There is a systematic pattern in the Mean Total Assets across all the portfolios. The funds in the top and bottom portfolio have a substantially lower average amount of assets under management than the funds in all other 15 portfolios. This result is not surprising, it is well documented that size and performance exhibit negative correlation. The only concern one can have while looking at these ﬁgures is whether there could be a way that our results are driven by fund size. In order to investigate this possibility, we perform the standard persistence methodology as described in section 3.1, where instead of sorting by alphas we sort by fund size (market capitalization). Using the same sub-sample and gross returns, at the beginning of every year, we rank each fund by its size and assign the funds to one of 10 portfolios based on these rankings. The composition of these 10 rank-sorted portfolios remains unchanged for the following 12 months. Following the sorting procedure a value weighted return series is calculated for each portfolio. This process is repeated every year and results in 10 time series which are then regressed on the four factor model. The results of this analysis are shown in table 8. None of the 10 portfolios has an alpha that is statistically diﬀerent from zero, neither is the 10-1 spread portfolio. This shows that although our methodology generates portfolios that comprise of diﬀerent size mutual funds, size is not the element driving the positive predictability results, but indeed it is the relative ranking inside the family. 3.3.2 Sample Size The criterion we use for keeping only funds that belong to families that have 10 or more funds reduces our sample to only 7% of all families, although in terms of funds it represents 38% of the full sample. Since the size of our sub-sample might be an issue, we apply the methodology for alternative sizes of the family, i.e. for various numbers of funds within the family, and obtain qualitatively similar results. When requiring that a fund belongs to a family of more than 5 mutual funds, our sample encompasses 22% of all the families, and 63% of all the funds of the full sample. As an illustration one can look at table 9 where we perform the same methodology for all the funds with 5 or more funds. As one can see the results are very similar. Our results also hold when using 5 instead of 10 portfolios. Interestingly, when using 5 portfolios the results of the ”within” analysis gives better persistence results than the standard methodology. We attribute this fact to issues of power of our tests when performing the relative ranking procedure. 16 3.4 Fund Correlations Given our empirical ﬁndings so far, one might legitimately ask whether the observed persis- tence in funds within a family is due to the fact that some families are better than others. In other words, whether some families have better strategies which they employ in order to enhance the performance of their funds. If this indeed were the case, then one would expect to see a large correlation in returns of funds of the same family, as well as correlations of their abnormal returns (alphas), assuming that the family implements the same invest- ment strategy in all funds. To test for that possibility, we calculate the average pairwise correlations of fund returns and abnormal returns within the same family. Correlations of fund returns are high, the average full sample correlation for gross (net) returns is 75.9% (75.8% respectively). This should not necessarily come as a surprise, since there is a lot of co-movement in fund returns generated by the common underlying factors (Market, SMB, HML, MOM). However, when we consider the correlations of the abnormal returns of the funds within one family, we get a diﬀerent picture. The results are presented in table 10. Panel A provides characteristics of the distribution of average correlations of fund alphas within each family. The numbers are provided for the full sample of families, as well as for large and small families, i.e. for families with more (or respectively less) funds than a pre-speciﬁed cutoﬀ. The numbers are quite revealing. The average pairwise correlation of fund alphas is 22.0% for the full sample. Yet, when we examine the correlations for the sub-samples, we note that within larger families, the correlations are substantially lower than within smaller families. The pairwise correlations of funds belonging to families with more than 10 funds have a mean of 13.1%, compared to a mean of 21.9% for the comple- mentary sample of families with less than 10 funds. In addition, the distribution of average pairwise correlations has a higher standard deviation in the sample of smaller families than in larger families. This result is even more striking in Panel B, where the absolute value of the pairwise correlations of fund alphas within the same family is displayed. The average correlation of the full sample is 30.2%, and in accordance with Panel A, the sub-samples of small families has a much larger average pairwise correlation (30.6% for families with less than 10 funds) than the sub-sample of large families (14.1% for families with less than 10 funds). These results do not provide evidence that supports the idea that bigger families main- tain a better persistence of performance by applying a common strategy to all their funds. 17 Instead, we actually note that there is a larger variation in investment styles within larger families14 . This fact points towards the hypothesis of a limitation of resources a family has, and implies that even if families have good investment ideas and strategies, these are not fully scalable and hence cannot be applied to the entire universe of funds within one family. Some selective allocation of these ideas has to occur. What we have claimed so far is that the choice where to allocate them depends on the eﬀects of the performance-ﬂow relationship and the implications thereof to aggregate inﬂows to the family. 4 The Role of the Family in Performance If a mutual fund family decides to promote some of its funds more than others, it will make sure that they exhibit an attractive performance record. One straightforward and observable way to accomplish this is by allocating human resources to those funds. Man- agers are one of the main resources families have both in actual terms and in the eyes of the investors. Therefore, it follows that asking whether the probability of adding another manager to a fund gives us an insight (and a good proxy) to an attempt from the family to ”push” certain funds. If our hypothesis is wrong one would expect that after controlling for fund characteristics, there would no longer be any statistically signiﬁcant diﬀerence in the probability of adding a manager between funds in diﬀerent groups of performance ranking. In this section, we analyze this point. We hypothesize that a fund family promotes its best performing mutual funds by using its main resource, its managers, in a systematic and predictable way. We use a logistic regression framework to estimate the probability of a manager being added to a fund given that it is one of the best (respectively, one of the worst) performing funds within the family. 4.1 Data and Summary Statistics on Managers In order to test our hypothesis we use data on managers of the mutual funds that comprise our database. Our main source on manager data is the ”Morningstar Principia Pro”’s CDs (Morningstar), described in the data section 2.2.1. Although one can obtain manager names at the fund level from CRSP, there exist some severe inaccuracies in the manager names 14 Our sample is comprised of funds that have an objective of Growth, Aggressive Growth, Growth and Income, Small Cap and Equity Income; hence the investment strategies considered here are not very diﬀerent and therefore the sample of funds should not be that diversiﬁed. 18 as well as in the managing period. For instance, CRSP either misreports or leaves out manager names that are active in mutual funds, is not consistent in the documentation of the manager’s name or is not consistent in the managing period, often observing a name of a manager dropping out and then reappearing after some time period. These drawbacks are quite severe and might aﬀect adversely an analysis of managerial turnover. For these reasons we use Morningstar to obtain accurate information on the names of the managers for the funds of our sample. During the period of 1990-2002 we identify 4,150 diﬀerent managers in our sample of 3,046 funds. Table 11 provides some summary statistics of manager characteristics and their evolution over time. In panel A, one can note that the number of managers increases over time although it is clear that the bad market conditions of the past years had an impact in the evolution of the industry since this increase reverted in the last two years. Yet, contrary to the slight decline in the number of managers of actively managed equity mutual funds, the average number of funds under their control has been steadily increasing. This eﬀect is probably purely mechanical and predominately driven by the faster reduction in the number of managers than in the number of mutual funds. The average size of the funds under management also increases over time, although it exhibits a decline in the last three years. Throughout their career, managers worked on average for 1.15 families. This is an indication of managerial turnover, as well as of consolidation within the mutual fund industry15 . In panel B, we look at the managers that manage more than one fund. As we noted earlier, 46% of the families are families of one equity fund16 and therefore their managers can rarely (except for subcontracting with another fund) manage more than one fund. We can see that among those who manage more than one fund the number of funds has been steadily increasing over the years and has always averaged above 2.5 funds per manager. The descriptive statistics indicate that it is very prevalent to have more than one manager per mutual fund. In addition, managers are quite mobile since they switch between funds over time. 15 In a few cases, a merger or acquisition between two diﬀerent mutual fund families results in the manager eﬀectively changing managing company. In our sample, we treat this as a switch of family. 16 Viewed at the fund level, 11% of the funds in our sample are the only funds in their respective family. 19 4.2 The Probability of Adding a Manager In order to study the probability of adding a manager to a fund, we deﬁne a dichotomous dependent variable in a logistic regression framework. The variable equals one when there was at least one net managerial addition to the management team of a fund, and equals zero when no net additions were made to the team17 . This methodology is similar to Khorana (1996). Our main concern in this analysis is an issue of power. The low net managerial turnover has to be analyzed using 10 portfolios, resulting in an even further reduced managerial turnover by portfolio. If for consistency we were to perform our tests on 10 portfolios, there would be not enough managerial movements to get any statistical signiﬁcance. To address this issue, we carry out all the logistic regressions using 5 portfolios instead. As we have mentioned in section 3.3.2, our results on fund persistence within the family hold also in a 5 portfolio framework, hence our hypothesis should also be testable in this case. To test our hypothesis, we run the following regression for each portfolio: P rt+1 (Adding a manager) = Λ [β1 + β2 αt + β3 αt−1 + β4 Log(T N A)t + β5 Log(T N A)t−1 + β6 Expenset ] (2) where Λ [·] is the logistic cumulative distribution function. We estimate the speciﬁcations using maximum likelihood. αt is the alpha of the fund, estimated from the four factor model of equation 1. We account for the size of the fund by using the natural log of the total net assets (TNA). This should control for the situation where the fund’s size increases and a new manager is needed due to the increased workload. We also account for the expenses of the fund in order to justify a situation where a manager is added in order to rationalize to investors the higher fees they are being charged18 . The results of the regression can be seen in table 12. The only portfolio where the alpha is signiﬁcant is the top portfolio (portfolio 5). The coeﬃcient is statistically signiﬁcant at a 1% level. As we have hypothesized, the probability of adding a manager increases when a fund belongs to the top performing funds inside its family. Size is statistically signiﬁcant for portfolios 3 and 4. The result is not surprising, since the largest funds are concentrated in the middle portfolios, hence they are be the ones 17 If the total number of managers in a team does not change even if there was managerial turnover, for example, if one manager was replaced by another one, we do not count this as a net addition. 18 We also performed these regressions including other combinations of controls such as, turnover, change in size and their lags. The main results were unchanged. 20 that would require additional managerial support when experiencing an increase in assets under management19 . Expenses are signiﬁcant for portfolios 3 and 4. Those portfolios are comprised of funds that were relatively mediocre in their respective families, and the only explanation for adding a manager is an increase (or a historical trend of increase) in size and thus workload, or in order to justify the expenses of the funds. To account for the possibility that a family measures the performance of its funds from their returns instead of their alphas, we estimate the probability of adding a manager to a fund given prior returns (net or gross). We run the following regression where Ii is a dummy variable for portfolio i: P rt+1 (Adding a manager) = 5 5 Λ β1 + β2,i Ii · Rett + β3,j Ij · Rett−1 + β4 Log(T N A)t + β5 Log(T N A)t−1 + β6 Expenset i=1 j=1 (3) where we let the dummy Ii interact with past returns. The results for net returns are pre- sented in table 13 (we obtain similar results for gross returns). Interestingly, the interaction terms are signiﬁcant only for the top and bottom portfolios. For funds that belong to the top relative performers, the probability of receiving an additional manager increases with past returns, while for funds belonging to the bottom relative performers, the probability of receiving a manager decreases with past returns. For funds belonging to intermediate portfolios, returns do not matter in the decision whether to allocate an additional manager of not to a fund. Finally, according to Khorana (1996), managerial turnover depends on the objective- adjusted percentage change in a fund’s assets. To calculate this change in a fund’s assets P CASSETt , we determine the average growth rate of other funds within the same invest- ment objective and subtract it from the asset growth rate of the fund. The diﬀerence should capture in- and out-ﬂows from of a fund that are due to managerial performance and not 19 In the same way that we have calculated summary statistics in table 5 for 10 portfolios, the same results hold for 5 portfolios: The middle portfolios possess larger funds on average, and expenses do not signiﬁcantly vary across portfolios. 21 to some aggregate investor sentiment. We run the following regression: P rt+1 (Adding a manager) = 5 5 Λ β1 + β2,i Ii · Rett−1 + β3,j Ij · P CASSETt + β4 Expenset + β5 T urnovert i=1 j=1 (4) The results of the regressions can be seen in table 14. As in table 13, past returns interacted with the dummy variable are signiﬁcant for the top and bottom portfolio. Yet, the change in asset growth rate does not aﬀect the probability of adding a manager to funds belonging to these groups. Families do not consider past objective adjusted asset growth rate when deciding where to allocate more resources, instead they focus only on past per- formance. The only statistically signiﬁcant portfolio is the third. A possible interpretation of this result would be that for mediocre funds it is more important how much inﬂows they receive compared to their peers, since their performance is not outstanding anyways. 4.3 The Probability of Moving from One to Multiple Managers Although we control for fund size and expenses, additions of managers to an already existing team might also be due to other reasons than just the intention to further promote the fund. As such, a manager might be added for a short period to ﬁll in a temporary vacancy or for training purposes, to obtain experience within an existing team. Therefore, we identify cases where only one manager was in charge of a fund and ask what the probability is of adding at least one new manager to such a fund. This question seems to be a more precise proxy for a deliberate action by the family, since the change from a single managed fund to a team managed fund is more substantial than adding another manager to an existing team. Thus, for each portfolio we run the following regression: P rt+1 (Adding a manager to a single managed f und) = Λ [β1 + β2 αt + β3 αt−1 + β4 Log(T N A)t + β5 Log(T N A)t−1 + β6 Expenset ] (5) We restrict ourselves to the subset of funds that are managed by only one manager and regress the probability of adding a manager on each fund’s lagged alpha, controlling for size and expenses. The results of the regression can be seen in table 15. The results are very 22 similar to the ones of equation 2. The only portfolio where the alpha is signiﬁcant is the top portfolio (portfolio 5). The coeﬃcient is statistically signiﬁcant at a 1% level. As we have hypothesized, the probability of adding a manager increases when a fund belongs to the top performing funds inside its family. The only other statistically signiﬁcant parameters are size for portfolios 3 and 4, and expenses for portfolios 3 and 4. We also perform tests similar to equation 3. The results can be seen in table 16. The results are quite similar to table 13. Good past performance increases the probability of converting a single-managed fund to a team-managed fund, only if the fund is one of the top performers in its family. Contrary to table 13, we do not observe the reverse behavior for funds in the bottom portfolio for lag t, instead the signiﬁcance moves to lag t − 1. This might be very well due to power issues, since the criterion of examining funds which were initially single managed reduces our sample by approximately 60%. In this paper, we have provided evidence that mutual fund families use managerial turnover as a means to promote some of their funds more than others. However, managers are not the only resource mutual fund families have that they can allocate unevenly between their funds. Reuter (2002), Gaspar, Massa, and Matos (2003), and Loeﬄer (2003) show that families sometimes distribute their allocations of IPOs unevenly between their funds in order to promote certain funds at the expense of others. Gaspar, Massa, and Matos (2003) also mention the possibility of a cross-subsidization of fund performance based on opposite trades of funds inside the family that would result in shifting bad performance from one fund to the another. However compelling these claims might be, they have a fundamental diﬀerence from our evidence on managerial turnover. Managers are the main and permanent resource families have, hot IPOs could be an added value, but are conditional on them being hot and available20 . The evidence on the managers shows that the families’ preferential treatment does not appear as an exploitation of a temporary opportunity, but a more consistent and reoccurring behavior. 20 Loughran and Ritter (2003) among others show that the astronomical discount in IPO prices occurred only during the 1999-2000 bubble period. 23 5 Conclusion Mutual fund families have an incentive to selectively favor their well performing funds in order for them to continue exhibit abnormal performance and thereby increase the inﬂows accruing to the entire family. In this paper we hypothesize that larger families not only have the incentive but also the means to do so. We show that funds that belong to larger families have a more persistent performance than the entire universe of funds. We show that this persistence is directly related to the number of funds in the family which we interpret as a measure of the latitude the family has in allocating resources unevenly between its funds. We also show that there exists persistence of performance of these funds inside their respective families. This is another indication that the family is actively engaged in aﬀecting the performance of its funds. In order to support our hypothesis that the stronger persistence of performance in larger families is due to a deliberate attempt by the family, we run a series of probit regressions which estimate the probability of assigning additional managers to a fund. We show that even after controlling for changes in size, expenses, and past performance, there still is a higher probability of adding a manager to a fund that belongs to last year’s family’s relative best performers. This implies that the family does not allocate resources (in this case managers/analysts) proportionally according to the funds’ needs but in a way that allows the family to promote certain funds, if this can help increase the inﬂows to the entire family. The approach presented here has broader implications than just showing that contrary to the outstanding literature there exists some persistence in mutual fund performance. In particular, it contributes to the understanding of the organization of mutual funds. Scharfstein and Stein (2000) and Rajan, Servaes and Zingales (2000), among others, have shown that there exists a subsidization across divisions in an organization. When the mutual fund family is viewed as the organization with the funds as its divisions, our empirical ﬁndings would be consistent with a theory where some divisions are awarded more resources than others given their high temporary output. Since managers compete to get more family resources (and thus improve their returns and ﬂows and by that their compensation) it might create an incentive to distort the signal the managers send to the head of the organization. In the context of mutual funds this could translate into risk shifting through the investment in more volatile assets. Chevalier and Ellison (1999a) have shown that such risk shifting 24 could be induced by career concerns of fund managers. However, our results imply that risk shifting could be also induced by a rent-seeking behavior on the part of the fund managers. Our results have also a direct implication on the welfare of mutual fund investors. If mutual fund families consciously promote some of their funds more than others, this will result in a transfer of wealth from one group of investors to another one. By using the management fees collected from all its funds to favor only a smaller sub-group of them, the family is eﬀectively shifting wealth between the majority of investors to those that invest in the funds it promotes. The regulatory implications of such a behavior cannot be ignored since it implies indirect investor discrimination and preferential treatment. 25 References [1] Arteaga, Kenneth R., Ciccotello, C. S., and C. T. Grant, 1998, New Equity Funds: Marketing and Performance, Financial Analysts Journal 6, 43-49. [2] Berk, Jonathan B., and R. C. Green, 2002, Mutual Fund Flows and Performance in Rational Markets, NBER working paper 9275. [3] Brown, Stephen J., and W. N. Goetzmann, 1995, Performance Persistence, Journal of Finance, 50 (2), 679-698. [4] Brown, Stephen J. and W. N. Goetzmann, 1997, Mutual Fund Styles, Journal of Financial Economics 43 (3), 373-399. [5] Carhart, Mark M., 1997, On Persistence in Mutual Fund Performance, Journal of Finance, 52 (1), 57-82. [6] Chen, Joseph, Hong, H., Huang, M., and J. D. Kubik, 2003, Does Fund Size Erode Mutual Fund Performance? The Role of Liquidity and Organization, working paper, Princeton University. [7] Chevalier, Judith and G. Ellison, 1997, Risk Taking By Mutual Funds as a Response to Incentives, Journal of Political Economy 105 (6), 1167-1200. [8] Chevalier, Judith and G. Ellison, 1999a, Career Concerns of Mutual Fund Managers, Quarterly Journal of Economics 114 (2), 389-432. [9] Chevalier, Judith and G. Ellison, 1999b, Are Some Mutual Funds Better than Others? Cross-sectional Patterns in Behavior and Performance, Journal of Finance 54 (3), 875- 899. [10] Daniel, Kent, Grinblatt, M., Titman, S., and R. Wermers, 1997, Measuring Mutual Fund Performance with Characteristic-Based Benchmarks, Journal of Finance 52 (3), 1035-1058. [11] Edelen, Roger, M., 1999, Investor Flows and the Assessed Performance of Open-End Mutual Funds, Journal of Financial Economics 53, 439-466. 26 [12] Elton, Edwin J., Gruber, M. J., and C. R. Blake, 2001, A First Look at the Accuracy of the CRSP Mutual Fund Database and a Comparison of the CRSP and Morningstar Mutual Fund Databases, Journal of Finance 56 (6), 2415-2430. [13] Evans, Richard B., 2003, Mutual Fund Incubation and Termination: The Endogeneity of Survivorship Bias, working paper, University of Pennsylvania. [14] Fama, Eugene F. and K. R. French, 1993, Common risk factors in the returns on stocks and bonds, Journal of Financial Economics 33 (1), 3-56. [15] Gaspar, Jose-Miguel, Massa, M., and P. Matos, 2003, Favoritism in Mutual Fund Families? Evidence on Strategic Cross-Fund Subsidization, working paper, INSEAD. [16] Grinblatt, Mark, Titman, S., and R. Wermers, 1995, Momentum investment strate- gies, portfolio performance and herding: A study of mutual fund behavior, American Economic Review, 85, 1088-1105. [17] Hendricks, Darryll., Patel, J., and R. Zeckhauser, 1993, Hot Hands in Mutual Funds: Short-Run Persistence of Relative Performance, 1974-1988, Journal of Finance, 48 (1), 93-130. [18] Hu, Fand, Hall A. R., and C. R. Harvey, 2000, Promotion or demotion? An empirical investigation of the determinants of top mutual fund manager change, working Paper, Duke University. [19] Jegadeesh, Narasimhan and S. Titman, 1993, Returns to Buying Winners and Selling Losers: Implications for Stock Market Eﬃciency, Journal of Finance 48, 93-130. [20] Khorana, Ajay, 1996, Top Management Turnover - An Empirical Investigation of Mu- tual Fund Manager, Journal of Financial Economics, 40, 403-427. [21] Khorana, Ajay, 2001, Performance Changes Following Top Management Turnover: Ev- idence from Open-End Mutual Funds, Journal of Financial and Quantitatie Analysis, 36, 371-393. [22] Khorana, Ajay, and H. Servaes, 2002, Examination of Competition in the Mutual Fund Industry, working paper, London Business School 27 [23] Lehmann, Bruce N., and D. Modest, 1987, Mutual Fund Performance Evaluation: A Comparison of Benchmarks and Benchmark Comparisons, Journal of Finance, 42, 233-265. [24] Loeﬄer, Gunter, 2003, Anatomy of a Performance Race, working paper, University of Ulm. [25] Loughran, Tim, and J. R. Ritter, 2003, Why Has IPO Underpricing Changed Over Time?, working paper, Univerity of Florida. [26] Nanda, Vikram, Wang, Z. J., and L. Zheng, 2003, Family Values and the Star Phe- nomenon, forthcoming, Review of Financial Studies. [27] Pastor, Lubos, and R. Stambaugh, 2002, Investing in Equity Mutual Funds , Journal of Financial Economics, 63 (3), 351-380. [28] Rajan, Raghuram, H. Servaes, and L. Zingales, 2000, The cost of diversity: The diver- siﬁcation discount and ineﬃcient investment, Journal of Finance 55, 35-80. [29] Reuter, Jonathan, 2002, Are IPO Allocations For Sale? Evidence from the Mutual Fund Industry, working paper, MIT. [30] Scharfstein, David S., and J. C. Stein, 2000, The dark side of internal capital markets: Divisional rent seeking and ineﬃcient investment, Journal of Finance 55, 2537-2564. [31] Sirri, Erik and P. Tufano, 1998, Costly Search and Mutual Fund Flows, Journal of Finance 53 (5), 1589-1622. [32] Spritz, A. E., 1970, Mutual fund performance and cash inﬂow, Applied Economics 2, 141-145. [33] Wermers, Russ, 1997, Momentum Investment Strategies of Mutual Funds, Performance Persistence, and Survivorship Bias, working paper, University of Colorado. [34] Wermers, Russ, 2000, Mutual Fund Performance: An Empirical Decomposition into Stock- Picking talent, Style, Transactions Costs, and Expenses, Journal of Finance, 55 (4), 1655-1695. 28 [35] Wilcox, Ronald T., 2003, Bargain Hunting or Star Gazing: Investors’ Preferences for Stock Mutual Funds, Journal of Business, 76(4), 645-663. [36] Wisen, Craig H., 2002, The Bias Associated with New Mutual Fund Returns, unpub- lished thesis, Indiana University. 29 Table 1: Descriptive Statistics - Funds This table reports summary statistics for the funds of our sample after accounting for the diﬀerent share classes. The number of funds is the number of mutual funds that meet our selection criteria. Total Assets are the total assets under management at the end of the calendar year in millions of dollars. Monthly Return is the cross-sectional average of the annual averages of the monthly returns of the funds in our sample, expressed in percentage. Total Load is the total front-end load, deferred and rear-end charges as a percentage of new investments. 12-B1 is the annual distribution charge (12b-1 fee) in percentage of total assets. Turnover is the minimum of aggregate purchases of securities or aggregate sales of securities, divided by the average Total Net Assets of the fund. Standard deviations are reported in parentheses. Number of Total Monthly Maximum Other Total Year Funds Assets Returns Load Loads Loads 12-B1 Turnover 1990 796 260.33 0.10 2.09 0.00 2.09 0.002 0.74 (758.42) (1.37) (2.62) (0.05) (2.62) (0.003) (0.92) 1991 898 299.31 2.59 2.11 0.00 2.11 0.002 0.70 (901.25) (1.58) (2.58) (0.05) (2.58) (0.003) (0.95) 1992 1,072 343.60 0.84 2.19 0.61 2.80 0.002 0.66 (1,119.79) (1.13) (2.54) (1.41) (3.01) (0.003) (0.97) 1993 1,267 413.67 1.16 2.13 0.79 2.93 0.002 0.65 (1,412.39) (2.15) (2.46) (1.60) (3.25) (0.003) (0.77) 30 1994 1,447 432.38 -0.08 2.05 0.94 2.99 0.002 0.74 1,562.34 0.76 2.42 1.80 3.47 0.003 1.13 1995 1,617 499.40 2.23 2.10 1.20 3.30 0.002 0.71 (1,562.34) (0.76) (2.42) (1.80) (3.47) (0.003) (1.13) 1996 1,885 596.51 1.56 2.07 1.36 3.44 0.002 0.67 (2,351.61) (0.91) (2.44) (2.06) (3.96) (0.003) (1.40) 1997 2,091 722.62 1.73 2.10 1.41 3.51 0.002 0.72 (2,933.48) (1.32) (2.50) (2.09) (4.15) (0.003) (0.86) 1998 2,337 826.75 1.39 2.17 1.56 3.73 0.002 0.69 (3,563.00) (1.97) (2.56) (2.18) (4.36) (0.003) (1.22) 1999 2,471 984.65 2.45 2.28 1.68 3.96 0.002 0.97 (4,415.89) (3.19) (2.61) (2.22) (4.48) (0.003) (3.94) 2000 2,599 1,140.87 -0.06 2.42 1.85 4.27 0.002 1.19 (4,816.08) (2.35) (2.66) (2.27) (4.60) (0.003) (6.69) 2001 2,587 948.89 -0.70 2.47 1.94 4.41 0.002 1.19 (4,028.29) (2.58) (2.68) (2.28) (4.64) (0.003) (2.63) 2002 2,409 905.59 -3.31 2.49 1.98 4.48 0.002 1.21 (3,634.55) (2.09) (2.68) (2.28) (4.65) (0.003) (2.42) Table 2: Descriptive Statistics - Families This table reports summary statistics for the families in our sample. The number of families is the number of fund families whose funds have met our selection criteria. Funds per family is the average number of funds each family has. TNA is the total assets under management of each family at the end of the calendar year in millions of dollars. Expense Ratio is the percentage of the total investment that shareholders pay for the mutual funds operating expenses. Total Load is the total front-end load, deferred and rear-end charges as a percentage of new investments. 12-B1 is the annual distribution charge (12b-1 fee) in percentage of total assets. Turnover is the minimum of aggregate purchases of securities or aggregate sales of securities, divided by the average Total Net Assets of the fund. Standard deviations are reported in parentheses. Number of Funds per TNA per Expense Maximum Other Total Year Families Family Family Ratio Load Loads Loads 12-B1 Turnover 1990 293 2.51 651.92 1.25 1.84 0.00 1.85 0.001 0.62 (3.86) (2449.44) (0.96) (2.45) (0.06) (2.45) (0.002) (0.65) 1991 322 2.70 827.35 0.98 1.92 0.00 1.92 0.001 0.62 (4.34) (3277.09) (0.81) (2.41) (0.06) (2.41) (0.002) (0.93) 1992 366 2.76 969.63 1.44 1.93 0.33 2.25 0.002 0.61 (4.26) (4181.76) (1.09) (2.31) (0.91) (2.55) (0.002) (1.20) 1993 401 2.97 1239.73 1.40 1.87 0.39 2.26 0.001 0.62 (4.62) (5687.50) (0.92) (2.22) (0.96) (2.62) (0.002) (0.65) 31 1994 424 3.27 1470.02 1.40 1.74 0.46 2.20 0.001 0.76 (4.98) (6998.25) (0.76) (2.16) (1.13) (2.69) (0.002) (1.16) 1995 452 3.42 1782.89 1.44 1.75 0.65 2.40 0.001 0.70 (5.14) (9201.39) (0.73) (2.13) (1.35) (2.94) (0.002) (1.41) 1996 494 3.61 2270.75 1.43 1.70 0.73 2.43 0.002 0.67 (5.54) (12330.25) (0.72) (2.12) (1.42) (3.06) (0.002) (1.33) 1997 509 3.87 2961.27 1.44 1.59 0.80 2.39 0.001 0.69 (5.93) (16157.77) (1.20) (2.12) (1.48) (3.16) (0.002) (0.82) 1998 542 4.13 3545.65 1.46 1.59 0.90 2.49 0.002 0.66 (6.38) (19940.71) (1.15) (2.15) (1.59) (3.32) (0.002) (1.02) 1999 554 4.27 4356.66 1.48 1.68 0.97 2.65 0.002 0.82 (6.74) (25367.13) (1.17) (2.25) (1.65) (3.47) (0.002) (1.91) 2000 541 4.62 5461.84 1.47 1.74 1.07 2.81 0.002 1.39 (7.21) (30094.69) (1.08) (2.31) (1.71) (3.61) (0.002) (10.24) 2001 514 4.93 4765.06 1.51 1.67 1.08 2.74 0.002 1.21 (7.78) (25838.39) (1.06) (2.29) (1.69) (3.57) (0.002) (3.54) 2002 479 4.99 4523.02 1.52 1.64 1.13 2.76 0.002 1.24 (7.98) (23766.78) (1.07) (2.28) (1.71) (3.59) (0.002) (3.71) Table 3: Portfolio of Mutual Funds Formed on Lagged 1-Year Alphas - Families With 10 or More Funds Using mutual fund data from 1990 to 2002, we perform a persistence analysis similar to Carhart (1997). At the beginning of the year we drop all the funds that belong to a family that has less than 10 funds. Then, we regress each fund’s monthly excess gross returns a four factor model to ﬁnd each fund’s alpha. We then rank each fund by its alpha. We build portfolios based on these rankings and keep them for a year. The procedure is then repeated every year. This results in 10 time series which we regress on a four factor model. MKTRF is the excess return on the CRSP value-weighted market proxy. SMB and HML are the Fama and French (1993) factor mimicking portfolios for size and book to market equity. MOM is the momentum factor mimicking portfolio similar to Jegadeesh and Titman (1993). The t-statistics are reported in parentheses. Portfolio Alpha MKTRF SMB HML MOM Adj R-Sq 1 -0.0023 1.0707 *** 0.2316 *** 0.2008 *** 0.0080 0.880 -(1.49) (26.13) (5.77) (3.99) (0.35) 2 -0.0014 1.0332 *** 0.1552 *** 0.1665 *** 0.0246 0.939 -(1.35) (38.08) (5.84) (4.99) (1.63) 3 -0.0019 ** 0.9962 *** 0.1088 *** 0.1300 *** 0.0007 0.959 -(2.32) (46.57) (5.20) (4.95) (0.06) 4 -0.0015 ** 0.9693 *** 0.0954 *** 0.1317 *** -0.0129 0.965 -(2.11) (50.19) (5.05) (5.55) -(1.20) 5 0.0002 1.0070 *** 0.0580 *** 0.1321 *** -0.0062 0.970 (0.24) (55.62) (3.27) (5.94) -(0.62) 6 -0.0003 0.9944 *** 0.0622 *** 0.0934 *** 0.0166 * 0.974 -(0.56) (61.47) (3.80) (4.55) (1.82) 7 0.0003 0.9824 *** 0.1116 *** 0.0644 ** 0.0161 0.954 (0.31) (42.61) (4.95) (2.27) (1.26) 8 0.0015 * 0.9785 *** 0.1363 *** -0.0109 0.0227 * 0.964 (1.88) (46.05) (6.55) -(0.42) (1.92) 9 0.0023 ** 1.0012 *** 0.3316 *** -0.1325 *** 0.0893 *** 0.946 (2.07) (33.11) (11.20) -(3.57) (5.31) 10 0.0035 *** 1.0456 *** 0.4562 *** -0.2471 *** 0.1117 *** 0.949 (2.80) (31.05) (13.84) -(5.97) (5.97) 10-1 0.0058 *** -0.0250 0.2246 *** -0.4478 *** 0.1036 *** 0.500 spread (2.86) -(0.46) (4.23) -(6.72) (3.44) Note: *** denotes signiﬁcance at the 1-percent level; ** denotes signiﬁcance at the 5-percent level; * denotes signiﬁ- cance at the 10-percent level. 32 Table 4: Portfolio of Mutual Funds Formed on Lagged 1-Year Alphas Using mutual fund data from 1990 to 2002, we perform a persistence analysis similar to Carhart (1997). At the beginning of every year, we regress each fund’s monthly excess gross returns a four factor model to ﬁnd each fund’s alpha. We then rank each fund by its alpha. The procedure is then repeated every year. This results in 10 time series which we regress on a four factor model. MKTRF is the excess return on the CRSP value-weighted market proxy. SMB and HML are the Fama and French (1993) factor mimicking portfolios for size and book to market equity. MOM is the momentum factor mimicking portfolio similar to Jegadeesh and Titman (1993). The t-statistics are reported in parentheses. Portfolio Alpha MKTRF SMB HML MOM Adj R-Sq 1 -0.0012 1.0433 *** 0.2418 *** 0.1786 *** 0.0059 0.905 -(0.97) (30.85) (7.06) (4.16) (0.31) 2 -0.0019 ** 1.0096 *** 0.1578 *** 0.2057 *** 0.0062 0.942 -(2.09) (41.71) (6.44) (6.70) (0.45) 3 -0.0016 ** 0.9765 *** 0.1152 *** 0.1467 *** 0.0006 0.966 -(2.29) (53.86) (6.28) (6.38) (0.06) 4 -0.0013 ** 0.9913 *** 0.1063 *** 0.1491 *** -0.0017 0.969 -(2.00) (57.04) (6.04) (6.77) -(0.17) 5 -0.0007 0.9690 *** 0.0709 *** 0.1186 *** -0.0015 0.971 -(1.15) (57.96) (4.19) (5.59) -(0.16) 6 0.0003 0.9658 *** 0.0713 *** 0.1068 *** -0.0005 0.975 (0.45) (62.89) (4.59) (5.49) -(0.06) 7 0.0003 1.0032 *** 0.1189 *** 0.0908 *** 0.0099 0.963 (0.47) (50.82) (5.95) (3.63) (0.88) 8 0.0010 0.9698 *** 0.1510 *** 0.0423 * 0.0017 0.967 (1.46) (51.78) (7.96) (1.78) (0.16) 9 0.0015 0.9966 *** 0.3140 *** -0.0521 * 0.0494 *** 0.959 (1.63) (42.02) (13.07) -(1.73) (3.69) 10 0.0023 0.9981 *** 0.4669 *** -0.2832 *** 0.0867 *** 0.933 (1.65) (27.46) (12.68) -(6.14) (4.22) 10-1 0.0035 * -0.0452 0.2251 *** -0.4618 *** 0.0808 *** 0.500 spread (1.72) -(0.83) (4.10) -(6.72) (2.64) Note: *** denotes signiﬁcance at the 1-percent level; ** denotes signiﬁcance at the 5-percent level; * denotes signiﬁ- cance at the 10-percent level. 33 Table 5: Summary Statistics of the Ranks Summary statistics of the ranks based on the two samples. Total Assets are the total amount of assets under management by the fund in millions of Dollars. Expenses is the expense ratio of the fund as a percentage of total net assets. Total Load is the total front-end load, deferred and rear-end charges as a percentage of new investments. Turnover is the fraction of the portfolio that is replaced during a year. The ﬂow is the monthly net ﬂow of the fund as a fraction of the assets under management. Panel A: Ranks based on the full samples - as described in table 4. Rank Mean Median Mean Median Mean Median Mean Mean Median Total Assets Total Assets Expenses Expenses Total Load Total Load Turnover Flow Flow 1 438.52 63.77 0.015 0.013 3.30 0.0 1.048 0.102 0.002 2 777.20 121.00 0.013 0.012 3.61 1.0 0.888 0.050 0.002 3 1025.51 141.13 0.013 0.012 3.79 1.0 0.687 0.050 0.003 4 1211.48 155.85 0.012 0.011 3.68 1.0 0.804 0.048 0.005 5 1115.95 160.45 0.012 0.011 3.62 1.0 0.616 0.045 0.008 6 1221.77 169.23 0.011 0.011 3.49 0.5 0.586 0.047 0.006 7 1065.03 150.63 0.012 0.012 3.73 1.0 0.657 0.045 0.007 8 982.75 132.82 0.013 0.012 3.40 0.5 0.754 0.054 0.009 9 808.86 131.81 0.014 0.013 3.40 0.0 0.820 0.049 0.012 10 395.20 80.10 0.016 0.014 3.14 0.0 0.969 0.146 0.031 Panel B: Ranks based on the sample of funds that belong to families that have 10 or more funds - as described in table 3. Rank Mean Median Mean Median Mean Median Mean Mean Median Total Assets Total Assets Expenses Expenses Total Load Total Load Turnover Flow Flow 1 920.75 184.80 0.014 0.013 5.08 5.0 0.956 0.094 0.002 2 1513.98 293.16 0.012 0.012 5.06 5.0 0.907 0.065 0.003 3 1919.74 318.29 0.012 0.012 5.28 5.5 0.720 0.037 0.003 4 2361.45 360.08 0.011 0.011 4.98 5.0 0.660 0.049 0.006 5 1879.79 391.68 0.011 0.011 4.87 4.5 0.643 0.040 0.010 6 1860.23 320.85 0.011 0.011 4.47 3.0 0.657 0.049 0.007 7 1715.35 331.87 0.012 0.012 4.91 4.8 0.712 0.034 0.008 8 1982.90 328.02 0.012 0.012 4.52 3.0 0.812 0.042 0.009 9 1454.30 292.87 0.013 0.012 4.61 4.0 0.882 0.053 0.011 10 838.39 200.00 0.014 0.013 4.93 4.8 0.911 0.159 0.031 34 Table 6: Portfolio of Mutual Funds Formed on Lagged 1-Year Alphas - Families with a Market Capitalization Above a Threshold Using mutual fund data from 1990 to 2002, we perform a persistence analysis similar to Carhart (1997). At the beginning of the year we drop all the funds that belong to a family that has a market capitalization below a given threshold. The threshold is such that the number of funds is equivalent to the number generated by a threshold of 10 funds per family. Then, we regress each fund’s monthly excess gross returns a four factor model to ﬁnd each fund’s alpha. We then rank each fund by its alpha. We then rank each fund by its alpha inside its family. We build portfolios based on these rankings and keep them for a year. The procedure is then repeated every year. This results in 10 time series which we regress on a four factor model. MKTRF is the excess return on the CRSP value-weighted market proxy. SMB and HML are the Fama and French (1993) factor mimicking portfolios for size and book to market equity. MOM is the momentum factor mimicking portfolio similar to Jegadeesh and Titman (1993). The t-statistics are reported in parentheses. Portfolio Alpha MKTRF SMB HML MOM Adj R-Sq 1 -0.0008 1.0648 *** 0.2276 *** 0.1762 *** 0.0111 0.895 -(0.60) (29.26) (6.18) (3.82) (0.54) 2 -0.0023 ** 1.0489 *** 0.1610 *** 0.2115 *** 0.0183 0.936 -(2.26) (39.47) (5.98) (6.28) (1.22) 3 -0.0016 ** 0.9800 *** 0.1243 *** 0.1478 *** 0.0074 0.966 -(2.36) (53.99) (6.76) (6.42) (0.72) 4 -0.0012 * 0.9911 *** 0.0958 *** 0.1540 *** -0.0050 0.966 -(1.82) (54.55) (5.21) (6.69) -(0.49) 5 -0.0005 0.9732 *** 0.0599 *** 0.1194 *** -0.0040 0.968 -(0.77) (55.50) (3.37) (5.37) -(0.40) 6 0.0001 0.9794 *** 0.0748 *** 0.1122 *** 0.0051 0.973 (0.20) (61.00) (4.60) (5.51) (0.56) 7 0.0006 0.9979 *** 0.1120 *** 0.0627 ** 0.0126 0.960 (0.72) (47.99) (5.32) (2.38) (1.07) 8 0.0012 * 0.9864 *** 0.1706 *** 0.0357 0.0066 0.968 (1.71) (52.35) (8.94) (1.49) (0.62) 9 0.0012 1.0036 *** 0.3432 *** -0.0610 * 0.0550 *** 0.953 (1.22) (38.61) (13.04) -(1.85) (3.75) 10 0.0026 ** 1.0237 *** 0.4673 *** -0.2842 *** 0.0956 *** 0.939 (2.00) (29.08) (13.11) -(6.37) (4.81) 10-1 0.0035 * -0.0411 0.2397 *** -0.4604 *** 0.0845 *** 0.504 spread (1.68) -(0.75) (4.31) -(6.61) (2.73) Note: *** denotes signiﬁcance at the 1-percent level; ** denotes signiﬁcance at the 5-percent level; * denotes signiﬁ- cance at the 10-percent level. 35 Table 7: Portfolio of Mutual Funds Formed on Lagged 1-Year Alphas - Analysis Within the Family Using mutual fund data from 1990 to 2002, we perform a persistence analysis similar to Carhart (1997). At the beginning of the year we drop all the funds that belong to a family that has less than 10 funds. Then, we regress each fund’s monthly excess gross returns a four factor model to ﬁnd each fund’s alpha. We then rank each fund by its alpha. We then rank each fund by its alpha inside its family. We build portfolios based on these rankings and keep them for a year. The procedure is then repeated every year. This results in 10 time series which we regress on a four factor model. MKTRF is the excess return on the CRSP value-weighted market proxy. SMB and HML are the Fama and French (1993) factor mimicking portfolios for size and book to market equity. MOM is the momentum factor mimicking portfolio similar to Jegadeesh and Titman (1993). The t-statistics are reported in parentheses. Portfolio Alpha MKTRF SMB HML MOM Adj R-Sq 1 -0.0021 * 1.0829 *** 0.2104 *** 0.2475 *** 0.0006 0.929 -(1.89) (35.71) (7.09) (6.65) (0.04) 2 -0.0016 1.0171 *** 0.1798 *** 0.1638 *** 0.0123 0.913 -(1.28) (31.30) (5.65) (4.10) (0.68) 3 -0.0014 * 0.9933 *** 0.0944 *** 0.1426 *** -0.0018 0.962 -(1.84) (48.60) (4.72) (5.68) -(0.16) 4 -0.0011 0.9895 *** 0.1222 *** 0.1259 *** 0.0179 * 0.969 -(1.55) (53.72) (6.78) (5.56) (1.75) 5 -0.0009 0.9900 *** 0.0853 *** 0.0959 *** 0.0037 0.976 -(1.39) (60.44) (5.32) (4.77) (0.41) 6 0.0003 1.0019 *** 0.0710 *** 0.0739 *** 0.0009 0.963 (0.34) (50.16) (3.51) (2.92) (0.08) 7 0.0005 0.9831 *** 0.1347 *** 0.0315 0.0220 ** 0.967 (0.65) (49.28) (6.90) (1.29) (1.98) 8 0.0011 0.9903 *** 0.1680 *** -0.0183 0.0461 *** 0.971 (1.50) (50.93) (8.83) -(0.77) (4.26) 9 0.0018 ** 1.0216 *** 0.3074 *** -0.1216 *** 0.0801 *** 0.969 (2.11) (44.73) (13.75) -(4.33) (6.31) 10 0.0032 *** 1.0430 *** 0.4503 *** -0.1740 *** 0.0873 *** 0.949 (2.62) (32.19) (14.20) -(4.37) (4.85) 10-1 0.0053 *** -0.0399 0.2398 *** -0.4215 *** 0.0867 *** 0.626 spread (3.19) -(0.90) (5.50) -(7.71) (3.50) Note: *** denotes signiﬁcance at the 1-percent level; ** denotes signiﬁcance at the 5-percent level; * denotes signiﬁ- cance at the 10-percent level. 36 Table 8: Portfolio of Mutual Funds Formed on Lagged 1-Year Market Capitalization Using mutual fund data from 1990 to 2002, we perform a persistence analysis similar to Carhart (1997). At the beginning of the year we drop all the funds that belong to a family that has less than 10 funds. Then, we ﬁnd the market capitalization of each fund and rank them based on that. We build portfolios based on these rankings and keep them for a year. The procedure is then repeated every year. This results in 10 time series which we regress on a four factor model. MKTRF is the excess return on the CRSP value-weighted market proxy. SMB and HML are the Fama and French (1993) factor mimicking portfolios for size and book to market equity. MOM is the momentum factor mimicking portfolio similar to Jegadeesh and Titman (1993). The t-statistics are reported in parentheses. Portfolio Alpha MKTRF SMB HML MOM Adj R-Sq 1 0.0015 0.9556 *** 0.2392 *** 0.0498 0.0494 *** 0.940 (1.45) (35.45) (9.07) (1.51) (3.30) 2 0.0006 0.9971 *** 0.2132 *** 0.1319 *** 0.0111 0.946 (0.66) (39.30) (8.58) (4.23) (0.78) 3 -0.0002 0.9758 *** 0.2104 *** 0.0751 *** 0.0194 0.959 -(0.22) (44.32) (9.76) (2.78) (1.58) 4 -0.0006 1.0011 *** 0.2732 0.0177 0.0631 *** 0.975 -(0.83) (55.21) (15.39) (0.80) (6.26) 5 0.0001 1.0238 *** 0.2712 *** 0.0721 *** 0.0555 *** 0.973 (0.16) (53.74) (14.54) (3.08) (5.24) 6 -0.0004 1.0247 *** 0.2303 *** 0.1063 *** 0.0330 ** 0.949 -(0.39) (40.17) (9.22) (3.39) (2.33) 7 -0.0006 1.0309 *** 0.1402 *** 0.0346 0.0238 ** 0.968 -(0.74) (50.15) (6.97) (1.37) (2.08) 8 -0.0005 0.9990 *** 0.1336 *** 0.0255 0.0231 ** 0.970 -(0.69) (51.86) (7.09) (1.08) (2.16) 9 -0.0008 1.0115 *** 0.0374 ** 0.0952 *** -0.0012 0.970 -(1.21) (54.09) (2.04) (4.14) -(0.12) 10 0.0008 1.0088 *** 0.0132 -0.0264 0.0072 0.977 (1.28) (58.63) (0.78) -(1.25) (0.76) 10-1 -0.0006 0.0532 -0.2260 *** -0.0762 * -0.0421 ** 0.308 spread -(0.50) (1.57) -(6.80) -(1.83) -(2.23) Note: *** denotes signiﬁcance at the 1-percent level; ** denotes signiﬁcance at the 5-percent level; * denotes signiﬁ- cance at the 10-percent level. 37 Table 9: Portfolio of Mutual Funds Formed on Lagged 1-Year Alphas - Families With 5 or More Funds Using mutual fund data from 1990 to 2002, we perform a persistence analysis similar to Carhart (1997). At the beginning of the year we drop all the funds that belong to a family that has less than 5 funds. Then, we regress each fund’s monthly excess gross returns a four factor model to ﬁnd each fund’s alpha. We then rank each fund by its alpha. We then rank each fund by its alpha. We build portfolios based on these rankings and keep them for a year. The procedure is then repeated every year. This results in 10 time series which we regress on a four factor model. MKTRF is the excess return on the CRSP value-weighted market proxy. SMB and HML are the Fama and French (1993) factor mimicking portfolios for size and book to market equity. MOM is the momentum factor mimicking portfolio similar to Jegadeesh and Titman (1993). The t-statistics are reported in parentheses. Portfolio Alpha MKTRF SMB HML MOM Adj R-Sq 1 -0.0027 ** 1.0738 *** 0.2388 *** 0.2207 *** 0.0070 0.895 -(1.98) (29.74) (6.53) (4.82) (0.34) 2 -0.0018 * 1.0536 *** 0.1672 *** 0.2019 *** 0.0271 * 0.928 -(1.67) (37.11) (5.81) (5.61) (1.69) 3 -0.0016 ** 0.9751 *** 0.1273 *** 0.1329 *** 0.0082 0.965 -(2.38) (53.26) (6.87) (5.72) (0.80) 4 -0.0012 * 0.9976 *** 0.0833 *** 0.1441 *** -0.0002 0.968 -(1.81) (56.26) (4.64) (6.41) -(0.02) 5 -0.0008 0.9645 *** 0.0556 *** 0.1160 *** -0.0041 0.965 -(1.11) (52.83) (3.01) (5.01) -(0.40) 6 -0.0001 0.9830 *** 0.0672 *** 0.1133 *** 0.0065 0.970 -(0.16) (57.70) (3.89) (5.25) (0.68) 7 0.0006 0.9988 *** 0.0964 *** 0.0571 ** 0.0168 0.952 (0.75) (43.88) (4.18) (1.98) (1.31) 8 0.0006 0.9875 *** 0.1767 *** 0.0410 0.0097 0.961 (0.78) (47.36) (8.37) (1.55) (0.82) 9 0.0011 0.9950 *** 0.3333 *** -0.1238 *** 0.0727 *** 0.951 (1.12) (37.01) (12.24) -(3.63) (4.79) 10 0.0031 *** 1.0364 *** 0.4610 *** -0.2707 *** 0.1073 *** 0.946 (2.49) (31.30) (13.75) -(6.45) (5.74) 10-1 0.0058 *** -0.0374 0.2222 *** -0.4914 *** 0.1003 *** 0.561 spread (3.04) -(0.74) (4.33) -(7.64) (3.50) Note: *** denotes signiﬁcance at the 1-percent level; ** denotes signiﬁcance at the 5-percent level; * denotes signiﬁ- cance at the 10-percent level. 38 Table 10: Summary Statistics of Average Pairwise Correlations of Fund Alphas Within Each Family This table reports summary statistics of the average pairwise correlations of fund abnormal returns (alphas) within the funds’ respective families. Within each family we calculate the pairwise correlations between all the fund alphas for their overlapping periods and average the correlations. Thus, all the reported statistics are cross-sectional statistics across families. We report the correlations for the full sample (referred to as Cutoﬀ - Full), and the sub-samples when we split by the number of funds in the family (referred to as Cutoﬀs bigger or smaller than either 5 or 10 funds in a family). Panel B reports the same statistics for the absolute values of the pairwise correlations. Panel A: Summary statistics of average pairwise correlations of fund alphas within each family. Cutoﬀ Mean Median Standard Standard Skewness 95th 80th 75th 60th 50th 35th 25th 15th 5th Error Deviation Percentile Percentile Percentile Percentile Percentile Percentile Percentile Percentile Percentile Full 0.220 0.183 0.017 0.332 -0.115 0.835 0.466 0.391 0.240 0.183 0.096 0.044 -0.042 -0.254 ≥ 10 0.131 0.109 0.014 0.118 0.587 0.366 0.218 0.203 0.148 0.109 0.071 0.056 0.021 -0.041 < 10 0.219 0.180 0.018 0.338 0.055 0.866 0.480 0.404 0.250 0.180 0.083 0.016 -0.060 -0.269 ≥5 0.154 0.127 0.013 0.159 1.504 0.428 0.247 0.217 0.152 0.127 0.085 0.060 0.029 -0.052 39 <5 0.226 0.211 0.021 0.383 -0.033 0.897 0.564 0.480 0.307 0.211 0.063 -0.042 -0.153 -0.360 Panel B: Summary statistics of absolute value of average pairwise correlations of fund alphas within each family. Cutoﬀ Mean Median Standard Standard Skewness 95th 80th 75th 60th 50th 35th 25th 15th 5th Error Deviation Percentile Percentile Percentile Percentile Percentile Percentile Percentile Percentile Percentile Full 0.302 0.209 0.013 0.260 1.070 0.860 0.522 0.426 0.282 0.209 0.147 0.106 0.070 0.023 ≥ 10 0.141 0.114 0.013 0.106 1.012 0.366 0.218 0.203 0.148 0.114 0.080 0.064 0.037 0.012 < 10 0.306 0.223 0.014 0.262 1.000 0.879 0.530 0.446 0.301 0.223 0.144 0.096 0.056 0.019 ≥5 0.168 0.135 0.011 0.144 2.117 0.428 0.247 0.217 0.157 0.135 0.091 0.072 0.049 0.023 <5 0.353 0.277 0.015 0.271 0.788 0.902 0.596 0.531 0.349 0.277 0.193 0.141 0.081 0.029 Table 11: Descriptive Statistics of Managers and the Funds they Manage This table reports the descriptive statistics of the fund managers and the funds they manage. The Total Assets Managed is the average total net assets of the funds each manager manages. Returns are expressed in percentages. Panel A: Descriptive statistics of all the the managers and their funds in the sample set. Year Number of Managers Total Assets Monthly Net Monthly Gross Managers per Fund Managed Return Return 1990 719 1.40 349.23 1.88 1.97 1991 854 1.44 493.53 5.62 5.69 1992 1037 1.46 555.76 0.85 0.88 1993 1230 1.51 633.82 1.24 1.22 1994 1420 1.54 612.49 2.60 2.51 1995 1600 1.59 784.69 0.93 0.95 1996 1799 1.73 953.78 1.62 1.62 1997 2020 1.73 1210.16 3.92 3.73 1998 2251 1.84 1378.11 -0.09 0.01 1999 2438 1.86 1549.20 2.22 2.20 2000 2519 1.89 1462.09 -2.48 -2.31 2001 2420 1.88 1338.84 2.56 2.62 2002 2254 1.86 1054.02 -1.79 -1.67 Panel B: Descriptive statistics of all the the managers that manage more than one fund. Year Number of Managers Managers per Fund 1990 184 2.58 1991 246 2.54 1992 316 2.52 1993 383 2.64 1994 469 2.63 1995 578 2.64 1996 743 2.76 1997 830 2.77 1998 1011 2.87 1999 1088 2.93 2000 1098 3.05 2001 1038 3.05 2002 961 3.03 40 Table 12: Logistic Regression of the Probability of Adding at least One New Manager We use mutual fund data from 1990 to 2002. At the beginning of the year we drop all the funds that belong to a family that has less than 5 funds. Then, we regress each fund’s monthly excess gross returns a four factor model to ﬁnd each fund’s alpha. We then rank each fund by its alpha. We then rank each fund by its alpha. We build portfolios based on these rankings and keep them for a year. We group all the funds by their portfolio and run the following regression by rank: P r(Adding a manager) = Λ [β1 + β2 αt + β3 αt−1 + β4 Log(T N A)t + β5 Log(T N A)t−1 + β6 Expenset ] where α is the fund’s alpha as measured by a 4 factor model. Log(TNA) is the log of the total net assets of the fund, and Expense is the expense ratio of the fund. Standard errors are reported in parentheses. Rank Intercept Alphat Alphat−1 Log(TNA)t Log(TNA)t−1 Expenset 1 3.26 *** -0.63 0.00 -0.06 0.00 8.56 (0.62) (3.81) (1.61) (0.11) (0.11) (24.89) 2 3.42 *** 0.07 1.66 0.01 -0.07 11.90 (0.42) (3.28) (4.04) (0.10) (0.10) (17.32) 3 4.42 *** 4.82 1.97 -0.26 ** 0.09 -11.22 ** (0.30) (3.18) (3.18) (0.10) (0.10) (4.76) 4 3.39 *** -0.73 -0.55 -0.38 *** 0.29 *** 41.13 * (0.48) (0.94) (0.37) (0.12) (0.11) (22.79) 5 3.64 *** 60.84 *** -10.84 0.00 -0.12 0.68 (0.69) (21.45) (13.85) (0.14) (0.14) (24.69) Note: *** denotes signiﬁcance at the 1-percent level; ** denotes signiﬁcance at the 5-percent level; * denotes signiﬁ- cance at the 10-percent level. 41 Table 13: Logistic Regression of the Probability of Adding at least One New Manager - With Rank Dummies We use mutual fund data from 1990 to 2002. At the beginning of the year we drop all the funds that belong to a family that has less than 5 funds. Then, we regress each fund’s monthly excess gross returns a four factor model to ﬁnd each fund’s alpha. We then rank each fund by its alpha. We then rank each fund by its alpha. We build portfolios based on these rankings and keep them for a year. This table reports results of the logistic regression using dummy variables to account for these portfolios. Rett and Rett−1 are the fund’s past returns. All the return terms are interacted with Ii which is a dummy variable for portfolio i. Log(TNA) is the log of the total net assets of the fund, and Expense is the expense ratio of the fund. P rt+1 (Adding a manager) = h P5 P5 i Λ β1 + i=1 β2,i Ii · Rett + j=1 β3,j Ij · Rett−1 + β4 Log(T N A)t + β5 Log(T N A)t−1 + β6 Expenset Variable Estimate StdErr ChiSq Prob Intercept 3.8407 0.1979 376.63 0.0000 Rett I1 0.0424 0.8479 0.00 0.9601 Rett I2 -0.0492 0.6203 0.01 0.9368 Rett I3 -0.1330 0.4870 0.07 0.7848 Rett I4 -0.4539 0.3989 1.30 0.2551 Rett I5 0.1444 0.4820 0.09 0.7645 Rett−1 I1 -0.8277 0.5023 2.71 0.0994 Rett−1 I2 0.8195 0.5403 2.30 0.1293 Rett−1 I3 0.6506 0.4686 1.93 0.1650 Rett−1 I4 0.2996 0.4515 0.44 0.5070 Rett−1 I5 2.1822 0.7782 7.86 0.0050 Log(T N A)t 0.0483 0.0957 0.25 0.6138 Log(T N A)t−1 -0.1582 0.0938 2.84 0.0917 Expensest -2.4479 5.4029 0.21 0.6505 42 Table 14: Logistic Regression of the Probability of Adding at least One New Manager to a Single Managed Fund - With Rank Dummies and Asset Turnover We use mutual fund data from 1990 to 2002. At the beginning of the year we drop all the funds that belong to a family that has less than 5 funds. Then, we regress each fund’s monthly excess gross returns a four factor model to ﬁnd each fund’s alpha. We then rank each fund by its alpha. We then rank each fund by its alpha. We build portfolios based on these rankings and keep them for a year. This table reports results of the logistic regression using dummy variables to account for these portfolios. Rett and Rett−1 are the fund’s past returns. All the return terms are interacted with Ii which is a dummy variable for portfolio i. PCASSET is the asset growth rate of the fund adjusted by the average asset growth rate of the fund’s objective. Turnover is the fund’s average yearly turnover. Expense is the fund’s yearly expense ratio. P rt+1 (Adding a manager) = h P5 P i Λ β1 + i=1 β2,i Ii · Rett−1 + 5 β3,j Ij · P CASSETt + β4 Expenset + β5 T urnovert j=1 Variable Estimate StdErr ChiSq Prob Intercept 3.2547 0.1045 970.11 0.0000 P CASSETt I1 0.0631 0.0567 1.24 0.2660 P CASSETt I2 0.0312 0.0444 0.50 0.4813 P CASSETt I3 -0.0098 0.0045 4.85 0.0276 P CASSETt I4 0.0005 0.0084 0.00 0.9515 P CASSETt I5 0.0087 0.0302 0.08 0.7732 Rett−1 I1 -0.9579 0.5242 3.34 0.0677 Rett−1 I2 0.6091 0.5309 1.32 0.2512 Rett−1 I3 0.4265 0.4462 0.91 0.3391 Rett−1 I4 -0.0655 0.4232 0.02 0.8769 Rett−1 I5 1.9565 0.7536 6.74 0.0094 Expensest 0.2255 5.8227 0.00 0.9691 T urnovert 0.0483 0.0581 0.69 0.4060 43 Table 15: Logistic Regression of the Probability of Moving from a Single Managed Fund to a Team Managed Fund We use mutual fund data from 1990 to 2002. At the beginning of the year we drop all the funds that belong to a family that has less than 5 funds. Then, we regress each fund’s monthly excess gross returns a four factor model to ﬁnd each fund’s alpha. We then rank each fund by its alpha. We then rank each fund by its alpha. We build portfolios based on these rankings and keep them for a year. We group all the funds by their portfolio and run the following regression: P r(Adding a manager to a single managed f und) = β1 + β2 αt + β3 αt−1 + β4 Log(T N A)t + β5 Log(T N A)t−1 + β6 Expenset + εt where α is the fund’s alpha as measured by a 4 factor model. Log(TNA) is the log of the total net assets of the fund, and Expense is the expense ratio of the fund. Standard errors are reported in parentheses. Rank Intercept Alphat Alphat−1 Log(TNA)t Log(TNA)t−1 Expenset 1 3.26 *** -0.65 0.01 -0.06 0.00 8.77 (0.62) (3.83) (1.61) (0.11) (0.11) (24.98) 2 3.42 *** 0.07 1.65 0.01 -0.07 11.87 (0.42) (3.28) (4.03) (0.10) (0.10) (17.30) 3 4.42 *** 4.83 1.96 -0.26 ** 0.09 -11.22 ** (0.30) (3.18) (3.19) (0.10) (0.10) (4.77) 4 3.39 *** -0.73 -0.55 -0.38 *** 0.29 *** 41.19 * (0.48) (0.94) (0.37) (0.12) (0.11) (22.78) 5 3.64 *** 61.34 *** -10.82 0.00 -0.12 0.11 (0.69) (21.51) (13.85) (0.14) (0.14) (24.49) Note: *** denotes signiﬁcance at the 1-percent level; ** denotes signiﬁcance at the 5-percent level; * denotes signiﬁ- cance at the 10-percent level. 44 Table 16: Logistic Regression of the Probability of Adding at least One New Manager to a Single Managed Fund - With Rank Dummies We use mutual fund data from 1990 to 2002. At the beginning of the year we drop all the funds that belong to a family that has less than 5 funds. Then, we regress each fund’s monthly excess gross returns a four factor model to ﬁnd each fund’s alpha. We then rank each fund by its alpha. We then rank each fund by its alpha. We build portfolios based on these rankings and keep them for a year. This table reports results of the logistic regression using dummy variables to account for these portfolios. Rett and Rett−1 are the fund’s past returns. All the return terms are interacted with Ii which is a dummy variable for portfolio i. Log(TNA) is the log of the total net assets of the fund, and Expense is the expense ratio of the fund. P rt+1 (Adding a manager to a single managed f und) = h P5 P i Λ β1 + i=1 β2,i Ii · Rett + 5 β3,j Ij · Rett−1 + β4 Log(T N A)t + β5 Log(T N A)t−1 + β6 Expenset j=1 Variable Estimate StdErr ChiSq Prob Intercept 4.1629 0.2414 297.40 0.0000 Rett I1 -1.9127 1.0711 3.19 0.0741 Rett I2 0.0061 0.8860 0.00 0.9945 Rett I3 -0.4316 0.6137 0.49 0.4819 Rett I4 -0.1286 0.6269 0.04 0.8374 Rett I5 0.6241 0.8185 0.58 0.4458 Rett−1 I1 -0.3922 0.8374 0.22 0.6395 Rett−1 I2 1.3661 0.7967 2.94 0.0864 Rett−1 I3 0.2532 0.6039 0.18 0.6750 Rett−1 I4 0.6109 0.6885 0.79 0.3749 Rett−1 I5 2.2477 1.1609 3.75 0.0528 Log(T N A)t 0.1580 0.1158 1.86 0.1722 Log(T N A)t−1 -0.1830 0.1151 2.53 0.1117 Expensest -7.7294 4.1857 3.41 0.0648 45