Академический Документы
Профессиональный Документы
Культура Документы
Abstract. Despite the wide acceptance of return-based style analysis, the method has several limitations. One important drawback is the assumption that style exposures are time invariant. We apply results on break tests established in Andrews and Ploberger (1994), Hansen (1997) and Bai and Perron (1998; 2003) to examine profoundly the possibility of style breaks. We find strong evidence against the hypothesis of constant time exposures in time in daily return data of European equity funds. All funds exhibit at least one break, while 60% of the funds exhibit even more than one break. We investigate the importance of (i) conditional investment strategies based on predictive information variables and volatility estimates and (ii) variables related to the organization of mutual funds. We conclude that the principal economic rational behind (the majority of) style breaks is the reliance of mutual fund managers on conditional investment strategies.
Key words: Mutual funds, performance evaluation, structural breaks, style analysis. JEL classification: C22, C52, G11, G20.
() We are grateful to the editor and the anonymous referee for their insightful comments.
Return-based style analysis (RBSA hereafter) has become a popular tool in analyzing mutual fund returns and investment objectives. The method was introduced by Sharpe (1988; 1992) as a tool to determine the effective asset mix of a mutual fund. Originally, RBSA boils down to estimating a constrained regression in order to determine the exposure of the historical mutual fund returns to different factors (returns on asset classes). The principal goal of RBSA is to find the best mimicking strategy that is in accordance with the investment style of the mutual fund.
RBSA has become a widely accepted analytic tool, both by academics and practitioners. Besides problems related to misclassification, RBSA has been employed to address issues concerning performance evaluation and object gaming of mutual funds (see e.g. Sharpe, 1992; de Roon, et al., 2004), construction of diversified portfolios or efficient portfolios of mutual funds with specified factor loadings (de Roon, et al., 2004), short-term risk assessment of a mutual fund manager (Sortino, et al., 1997) and hedge fund style and risk evaluation (e.g. Fung and Hsieh, 1997; Schneeweis and Spurgin, 1998; Lhabitant, 2002). Object gaming refers to the practice of deliberately deviating from the stated objective to attain higher relative performance.
Despite this wide acceptance, RBSA has in its original form a number of drawbacks and limitations. The main problems can be traced to the choice of the asset classes, estimation of confidence intervals of style exposures, the purpose-dependent restrictions to be imposed in RBSA, and the underlying hypothesis of constant style exposures in time. While we briefly discuss the former potential pitfalls in Section 2, the main focus of this paper is related to the
assumption that style exposures are constant over the period studied. Indeed, a number of papers have documented that style changes over time do occur (see e.g. Chan, et al., 2002; Kim, et al., 2000; Swinkels and Van Der Sluis, 2001). Style breaks may be motivated by economic reasons like market or volatility timing, herding or portfolio alteration in anticipation of changing economic conditions. They may also be induced by changes in the mutual fund management structure or by flow dynamics.
In this article we contribute to the existing discussion on RBSA-methodology by conducting a profound analysis on the nature of the time variation in exposures. Contrary to previous literature, we conduct formal and econometrically powerful tests to examine the occurrence of style breaks. Besides the econometric advantages, this approach renders it easy to assess the performance of extended style models that are introduced subsequently to capture this time variation. In a straightforward way, we can compare the number of breaks still present in the extended models against the occurrence of breaks in the basic model. We also investigate if we can identify classes of funds that are subject to style breaks. By doing so, we relate our finding on style breaks to other strands of the mutual fund literature that focus on mutual fund organization and the practical implementation of investment strategies by mutual fund managers. We return to this issue later on.
As preliminary descriptive statistics indicate that changes in style exposures of European equity funds may be important, we formally test for breaks in the estimated style exposures by employing the testing framework described in Bai and Perron (1998; 2003). This framework has the advantage that it enables testing for multiple breaks at a priori unknown break dates, as well as investigating breaks in a subset of variables. For every fund in our sample we find evidence of at least one style break, and for many funds (60%) even multiple breaks. Therefore, we try to gain insight in the economic behavior that drives these breaks.
Hence, in the second part of the empirical research we try to model the time variation. In addition, we compare the number of breaks in the basic model and extended models that are put forward to account for the time variation. By doing so, we can assess the performance of the extended models. In these models we assume that mutual fund managers use conditional investment strategies whereby they rebalance their portfolios based on publicly available information like dividend yield or volatility estimates. We also relate our findings to
organizational features of the mutual fund industry and test whether, for instance, fund family membership or mutual fund tournaments are relevant determinants of the documented style breaks.1 We test this by examining whether the number of breaks differs between fund groupings based on fund family membership, historical performance or the level of industry concentration. We show that the first approach, where we look at conditional investment strategies of mutual fund managers, is the most fruitful approach despite its limitations. Indeed, the formulation of the extended models is probably too unsophisticated to be realistic provided that we express style exposures as linear functions of publicly available information variables and we only include only one information variable at a time due to data limitations. Yet, we do find that when exposures are allowed to vary in such a simple way, a significant reduction in the number of breaks per fund is noticeable. This clearly indicates that an adequate style analysis should allow for time-varying exposures, a result consistent with the performance evaluation literature.
Besides the profound analysis of the style exposures, there are two noticeable features of our data sample that are worth mentioning. Firstly, the set-up of the empirical part is distinct from
Due to a lack of data on mutual fund managers and fund flows, we can regrettably not test the hypothesis that the style breaks are induced by fund-specific factors (like changes in the fund management or flow dynamics) empirically. Nevertheless, it is unlikely that a switch of managers occurs frequently. Therefore, a high number of significant breaks over the sample period is at least indirect evidence against the hypothesis that the style break(s) is (are) only motivated by substitutions in mutual fund management structure.
other existing studies in that it uses daily input instead of generally applied monthly data. The higher frequency of the data opens up the opportunity to grasp a more detailed picture of the time-varying nature in style exposures, given that we are able to pick up intra-month style variation. Secondly, we focus on funds with 'European equity' as investment objective which are distributed in Belgium. By studying a sub-market of the European mutual fund industry we contribute to the sparse but growing empirical evidence on this industry. The European mutual fund industry is by far the second largest market after the American mutual fund market.
The remainder of this paper is organized as follows. Section 2 gives a short overview of RBSA and its pitfalls. In Section 3 the testing procedure for structural breaks is described. Our dataset and selection procedures are introduced in Section 4. The empirical results are presented in Section 5. Finally, we conclude.
2.
The interest in RBSA was fuelled by difficulties, faced by investors and the management companies distributing funds, in determining the actual mutual fund investment objective and investment behavior. Empirical evidence on misclassification reported by Brown and Goetzmann (1997), Kim, et al. (2000), or diBartolomeo and Witkowski (1997) corroborate that deviations between the actual investment style and the one reported in the prospectus or by the data vendor do occur in a substantial number of funds. As indicated in the introduction, RBSA has subsequently been employed to a number of mutual fund issues. A more detailed discussion of the various applications of RBSA is stated in Van Campenhout (2002).
In comparison to characteristic-based style analysis, RBSA has a number of advantages if we want to find out the actual investment policy. As the former relies on actual portfolio holdings
and thus analyses the stocks held in mutual fund portfolios (e.g. Daniel, et al., 1997; Chan, et al., 2002; Wermers, 2000), its main drawback is that mutual fund holdings are in most cases not readily available. Furthermore, they are generally only available on a quarterly basis and timely information on holdings may be difficult to obtain. Moreover, if mutual fund managers commit window dressing practices, inferences from reported portfolio holdings might be misleading. In contrast, by regressing fund return on returns of selected passive style indices, RBSA relies only on abundantly available historical return data, which makes it an attractive tool from a practical point of view. Chan, et al. (2002) compare the two approaches and conclude that in general they give similar results on the fund's style. However, in the cases where the two approaches yield different results, the characteristics-based approach is better in predicting future fund performance.
The econometric model proposed by Sharpe is part of the general class of factor models and is often written as:
(1)
(1)
where R ft is the return on fund f for the period ending at t, Rit denotes the return on asset class i, K is the number of asset classes, ft is the idiosyncratic noise term, and fi is fund fs factor exposure w.r.t. factor i. A feature of the model is that it decomposes a fund's return into a component that is attributable to style, given by
K i =1
component, given by f + ft . In addition, the portfolio weights are restricted to be positive and to add up to one in the original model. This could result in biased estimates when not all asset classes the fund invests in are accounted for, or when the investments are inaccurately taken into account by the adopted asset classes (de Roon, et al., 2004). Therefore, we prefer in this paper not to impose the restrictions. By estimating the exposures freely, we want to avoid
any interference with the focus in this article, namely the potential time-variation of style exposures.
The proper implication of the original model asks for asset classes that are mutually exclusive, linearly independent and exhaustive (in the sense that the asset factor model should be able to span the fund's portfolio asset mix). (Sharpe, 1992; Lobosco and diBartolomeo, 1997; de Roon, et al., 2004). Highly correlated asset classes should be avoided, since they capture essentially the same style characteristics. Ideally, we would apply ordinary style analysis only if the fund's assets are indexed to well-prescribed indices (Buetow, et al., 2000), create fund-specific asset classes (Bailey, 1992a; b) or combine RBSA with fundamental information on the securities in the fund's portfolio (Lucas and Riepe, 1996).
Gallo and Lockwood (1997) demonstrate that the data vendor's classification scheme could influence the outcome of the style analysis. To ensure that our results are not determined by the data vendor's classification scheme, we employ the classification put forward by two different data vendors. The main conclusions remain unaltered, so the results of the sensitivity analysis are not presented to save space.
Another potential drawback of the standard RBSA-model which is directly related to the main issue in this article is the implicit assumption that the sensitivities to the factors remain fixed over the sample period. In reality, style changes in time have been documented (see e.g. Chan, et al., 2002; Kim, et al., 2000; Swinkels and Van Der Sluis, 2001). To visualize the timevarying property of style coefficients, it is common practice to estimate Sharpe's model over rolling windows (Sharpe, 1992; Lucas and Riepe, 1996; Buetow, et al., 2000). It is then possible to graph the style exposures through time. This approach might help to establish a first impression of the time variation of style coefficients. Nevertheless, one still assumes that the sensitivities are constant over the arbitrarily chosen length of the rolling window. Besides,
the approach of overlapping rolling windows and the fact that every observation is equally weighted cause a delay in the recognition of a style change. A paper that is closely related to our research is Swinkels and Van Der Sluis (2001) who decide to model the time-varying exposures explicitly by applying a Kalman smoother. The approach is suited to model smooth changes in time. Nonetheless, there is no a priori reason to assume that only smooth changes take place. If the Kalman smoother technique is applied, the complete data sample is used to estimate the model, instead of a subset of historical returns (which is used in the rolling window approach). Note that the empirical evidence is derived from a limited sample of funds with a global investment strategy. In support of their time-varying model they also refer to the
CUSUM (CUSUMQ)
and Chow tests for coefficient stability which do not constitute a real
CUSUM (CUSUMQ)
tests
are based on residuals from recursive estimation and have been criticized of having low statistical power (e.g. Andrews, et al., 1996; Greene, 2000). The implementation of the Chow test requires that the sample is split into two. Subsequently, the parameters are estimated for each sub-period and the equality of the two sets of parameters is tested using an F-test. A major drawback of this procedure is that it is only applicable if the break date is known a priori which is not the case in style analysis. Moreover, the procedures cannot be used to test simultaneously for multiple breaks. In the following section, we will discuss a test procedure that allows testing for multiple breaks at unknown break dates. We will also introduce how we will model time-variation in style exposures.
3.
The literature with respect to structural breaks is comprehensive, primarily aimed at the treatment of a single break (see e.g. Krishnaiah and Miao (1988) for a survey). Recent papers derive procedures to deal with a single break or multiple breaks at (an) unknown date(s) in various models.
In general, the research design of our principal structural break tests can be summarized as follows. We want to test whether the coefficient matrix that consists of the fund's style exposures to the various benchmarks is constant in time, or alternatively is subject to one or more structural breaks.
The standard unrestricted return-based style model (Equation (1)) is summarized in matrix notation by:
(2)
r = R + e , where
r = R f 1,..., R fT
(2)
)' , R
)' , and
e = f 1 ,..., fT .
Now consider the standard (unrestricted) RBSA model with m breaks in the coefficient matrix that captures the fund's style exposures to the various benchmarks. Hence, we rewrite model (2). There are m+1 regimes during which fund style exposures have values
( j ) , j = 1, 2,..., m + 1 :
(3)
r = R * * + z + e. (3)
where * = (1) ,..., ( m +1) and R * diagonally partitions R at the break dates ( T1 ,..., Tm ) , i.e.
R * = diag ( R 1 , ..., R m +1 ) with R i = rTi 1 +1 ,..., rTi . Note that by definition we set T0 = 0 and
to change is , the parameter vector is not subject to change. In our setting the factor style exposures to the benchmarks are stacked in , while the intercept is taken up in .
The null hypothesis that will be investigated is that is constant through time, against the alternative hypothesis that (multiple) breaks occur. Depending on the specific test statistic used, there are two related alternative hypotheses. We return to this issue later on. In the main, we test for multiple breaks by applying the break test methodology outlined in Bai and Perron (1998; 2003). Bai and Perron (1998) focus on estimating multiple structural changes in a linear model that is estimated by least squares. They derive the limiting distribution of the estimators as well as the test statistics. The practical issues concerning the implication of the procedure are outlined in Bai and Perron (2003). The general idea is to estimate the model under the alternative hypothesis for every m-partition of the time period studied. The estimates of the parameters and the break dates are obtained for the partition that minimizes the sum of squared residuals. Bai and Perron use insights from dynamic programming to avoid the computational burden to actually estimate the regression for every m-partition.
The break tests are estimated for every individual fund. The empirical results are presented in Section 5. We employ two test procedures. First, we verify if at least one break is present. The first procedure tests the null hypothesis of no structural break against an unknown number of breaks, given some upper bound M using the UDmaxFT ( M, K + 1) test:
(4)
( j = 1,..., m ) .
(4)
The estimates of the break dates T j are obtained from the global minimization of the sum of
squared residuals. In our empirical applications, we set the upper bound to 5. Second, if we find evidence of at least one break based on the UDmaxFT ( M, K + 1) test, we examine the number of breaks based on the sequential test for m versus m+1 breaks. The procedure boils down to applying m+1 tests of the null hypothesis of no structural change versus the alternative hypothesis of a single change. The test is applied to each segment containing the
observations Ti 1 to Ti
10
the model with m breaks if the overall minimal value of the sum of squared residuals is sufficiently smaller than the sum of squared residuals from the model with m breaks. Bai and Perron denote this test by SupFT ( m + 1 m ) .
If we do find evidence of style breaks, the question arises which economic behavior could account for these style break patterns? One possible explanation that we explore in section 5.4 is that the style breaks are induced by the fact that mutual fund managers alter their market exposures conditioned on publicly available information such as macro-economic information variables or volatility shocks. Our research method is inspired by the methodology applied in the conditional performance evaluation literature (Ferson and Schadt, 1996; Christopherson, et al., 1999). They argue that performance measures like Jensens alpha are biased when fund managers change their portfolio composition in response to predictable changes in the investment opportunity set. They model these changes by making the funds exposures to asset classes linear functions of public information variables. The way in which we implement the insights of their approach can best be illustrated as follows. Take one observation from model (2):
(5)
R ft = rt t + ft ,
(5)
where now the parameter vector has an explicit time index. We assume that time variation in the parameters is determined by an L1 information vector zt:
(6)
t = zt ,
(6)
where is a (K+1)L coefficient matrix and the first information variable is by construction a constant. Eq. (5) can then be rewritten as follows:
(7)
R ft = vec ( ) ( rt zt ) + ft ,
(7)
where vec() stacks columns vertically and denotes the Kronecker product. Alternatively, as both xt and zt contain a constant, a reformulation is:
11
(8)
R ft = rt (1) + x t ( 2 ) + ft ,
(8)
where xt rt z t , and z t is zt without the constant term. In a similar vein, we extend the traditional style equation by factors representing economic motives. If the style breaks are fuelled by economic considerations, the standard style coefficients in the extended model should be free of breaks. This hypothesis can be tested by estimating a partial break model in which only the standard coefficients (1) are allowed to vary. The testing procedure outlined above can be adopted in a straightforward manner if we want to test the stability of a subset of coefficients. However, a practical problem that arises is that the number of regressors in the extended style analysis goes up quickly, due to the interaction terms taken up in the regression. Consequently, the multiple break tests by Bai and Perron (1998; 2003) are no longer estimable because the degrees of freedom of the regression model become insufficient when the sample is sequentially subdivided at the various break dates. Therefore we present several extended style analyses where only one extra economic variable and the interaction terms with the indices rt are added.
In view of the literature on volatility timing (Busse, 1999), we also explore the extent to which this type of timing behavior of mutual fund managers could account for possible style breaks. We use a similar approach as outlined above, but the interaction terms of the factors with the information variables is replace by an interaction term capturing the daily volatility shock ( mt +1 mt ) and the broad market index. To model daily market volatility we estimate a GARCH(1,1) specification with t-distributed conditional errors:
(9)
2 = 0 + 1 t21 + 1 2 , t t 1 ~ t ( 0, 2
mt mt 1
mt
) where
12
4.
DATA DESCRIPTION
Our empirical research on the time variation in style exposures is conducted on European mutual fund data. In doing so we complement the growing body of empirical studies on European mutual fund markets (see, e.g., Walter and Weber, 2006 (Germany); Casarin, et al., 2005 (Italy); Otten and Bams, 2000 (UK); Annaert, et al., 2003 (France, Benelux and Offshore Territories); Otten and Schweitzer (2002) and Otten and Bams (2002) who study the European mutual fund sector in general). Notwithstanding that the European mutual fund sector is the second largest (measured by assets under management), the understanding of the European mutual fund industry is still scanty compared to the vast amount of research carried out on the American fund industry. Moreover, to our knowledge we are the first study on the European mutual fund industry to use daily data. All other studies rely on lower-frequency data. The daily frequency of our sample data makes it possible to document intra-month time variation in style exposures and allows for more efficient estimates over a fixed time interval. In the United States, daily data have been used in studies on mutual fund timing (see, e.g., Chance and Hemler, 2001; Bollen and Busse, 2001) or Busse (1999) who documents volatility timing in mutual funds. Other studies that rely on daily data include Edelen and Warner (1999) and Goetzmann and Massa (2003) who study the relationship between mutual fund flows and market returns at high frequencies.
The mutual fund data are taken from the Vision-database. This is a comprehensive database containing daily data on funds distributed in Belgium and the Netherlands. We opt to focus on European equity mutual funds. The initial selections are made based on data available on 17 July 2001. The selection of mutual funds labeled 'European equity' is founded on the classification applied by two different data vendors. The first selection is taken from the 'Financieel Economische Tijd (Fondsen-Tijd)', a leading Belgian financial newspaper. To ensure that our results are not driven by differences in data vendor's classifications (Gallo and Lockwood, 1997), we construct a second selection that follows the classification employed by
13
Efund (an internet mutual fund market specialized in funds distributed in Belgium at that time). Furthermore, we impose the following selection rules: a selected mutual fund should be a capitalization fund; have a daily quotation frequency during the total sample period; have at least one year of daily data; have an adjusted R and exposure coefficient of at least 50% in preliminary regression of mutual fund returns on the return of a general equity index (namely Stoxx Total Market Index (TMI) Europe).
In short, we impose the first criterion to avoid blowing up our sample. For instance, managers often set-up capitalization and distribution funds to appeal to different types of investors. The investment policy of both types of funds is, however, identical (and most likely entrusted to the same fund manager). If breaks would occur, they will be present for both the distribution and the capitalization fund. Note that restricting our sample to capitalization funds biases our sample against finding style breaks. The imposition of the second criterion allows us to fully exploit the daily frequency of the data. Besides funds that report daily, also semi-weekly, weekly and semi-monthly frequencies occur in the category of funds with investment objective 'European equity'. The daily data ensures that we have enough data points to estimate the style equation over shorter time periods than the windows that are usually applied. Moreover, we can embark upon investigating intra-month variation. The third argument ensures that we have enough data points to arrive at sensible results by estimating our model and conducting the statistical tests that were discussed in Section 3. For instance, the sequential procedure splits up the sample at the break date and the model is re-estimated if the length of the split segment is sufficient. Finally, previous studies (Brown and Goetzmann, 1997; Kim, et al., 2000; diBartolomeo and Witkowski, 1997) document that fund misclassification exists. This could be problematic to our tests on time variation in exposures. To prevent that our results are driven by misclassified funds, we preliminary regress the individual fund returns on the return of the general equity index Stoxx TMI Europe and eliminate all funds with an adjusted R 2 and equity exposure beneath 50%. Let us take the
14
example of balanced funds. We expect a priori that this type of mixed funds will alter more rapidly from predominant asset class, because their investment policy is defined more broadly. This kind of investment behavior will result in time-varying exposures. Consequently, if balanced funds are misclassified in the European equity category we will tend to overestimate the time-variation of exposures and probably the number of breaks associated with equity funds. A second argument is that, if the selected funds are not European equity funds, the set of factors is misspecified, and will no longer satisfy the properties outlined in Section 2.
A preliminary analysis of the available data reveals us that the raw data are not error free. Three types of problems occur. Firstly, the mutual fund data contain a number of faulty data points resulting in improbable daily returns. Secondly, it happens that data points are entered on the wrong date (e.g. one day later), causing a shift in the data series from this point onwards. Both errors can be traced efficiently by performing RBSA over very narrow rolling windows (e.g. 30 days). Problematic data points were subsequently spot checked against the daily quotes reported in the newspaper and corrected in the final sample. Thirdly, despite the fact that only funds that report daily quotes are withheld, the data samples are not balanced. In case of missing values, mutual fund returns are calculated based upon the available net asset values enclosing the missing value(s) (see Equation (10)), and appropriate benchmark returns are calculated as well. Since mutual funds are typically well-diversified portfolios of stocks, mutual funds are not as susceptible to problems associated with nonsynchronous trading as individual stocks.2
In a recent paper, Chalmers, et al. (2001) demonstrate that with respect to the U.S. markets non-synchronous trading problems in relation with NAV calculations given cause to an option to trade these assets indirectly at stale prices. Previous research (e.g. Bhargava, et al., 1998) has documented similar issues with respect to foreign mutual funds. It is unclear, if (and to what extent), these results carry over to the European mutual fund markets that have different institutional characteristics.
15
(10)
R ft =
NAV ft + D ft NAV ft 1
1,
(10)
where Rft is the return of fund f over period t, is the number of trading days with missing values of the return period under consideration, NAVft the net asset value of mutual fund f at time t, and Dft the cash distributions over period ending at t. Given that our sample only consists of capitalization funds, no cash distributions take place. We end up with a final sample of 62 funds for the Fet-selection, while there are 53 funds in the Efund-selection meeting the selection criteria. 15 funds are only taken up in the Fet-selection, while six funds are solely part of the Efund-selection. 47 funds are in both selections.
Data are available for the period 3 June 1991 up to 25 June 2001. Summary statistics for both samples are taken up in Table 1. Readers interested in the result of the selection rules may consult Appendix I.
The mutual fund returns, as revealed in panel A and B of Table 1, are comparable for both selections. In general, the average daily fund return is positive, in three years the average daily yearly return is negative. The substantial positive return in the second half of the nineties mirror the bull market in that period. The last column of panel B shows that, likely due to the increased volatility of the underlying markets, the cross-sectional standard deviation increases substantially over time.
The overall aim of our study is to investigate the time variation in the fund's exposures to general asset classes. For this purpose we decide to employ general equity and bond indices as factors in our style analysis. More specifically, in our basic model we employ the following
16
total return indices: the general and small cap equity indices labeled Stoxx Broad Europe Market Index (In2) and the Stoxx Small Europe Market Index (In4) respectively. The broad market index (In2) covers the largest 600 companies measured by market capitalization. The small cap index (In4) is constructed by ranking the stocks along the free float market capitalization. The companies between the 90th and 95th percentile make up the small cap index. If European equity funds are exposed to the government bond market, they might prefer the Belgian bond index. Before the European unification, this behavior could be driven by home bias. After the unification, a motivation often encountered is that a premium on the Belgian bond index could be earned. In a preliminary analysis, we estimate the RBSA-model with the Belgian or the German bond index. The specification with the Belgian bond index yields a bond exposure that is both significant and higher than the bond exposure to the German bond index, as well as a better goodness-of-fit. Thus, it appears from this preliminary analysis that mutual funds are indeed sensitive to the premium. As a consequence, we run the basic model with the Belgian total return government bond index (In5). The bond indices are taken from Datastream. In the sensitivity analysis, we have also run alternative regressions where the Stoxx Broad Europe Market Index is replaced with the Stoxx Total Market Index (In3) or the Euro Stoxx Blue Chip Market Index (In1). The total market index contains the largest 95% of companies sorted by free float market capitalization, while the blue chip market index consists of 50 stocks covering the market sector leaders in the Euro Stoxx index. Since results where not fundamentally altered, we do not report the results of these regressions to save space. Other studies, like Chan, et al. (2002), have selected the Fama and French (1993) 3-factor model or Carhart (1997) 4-factor model. This choice is appropriate in style analysis when one is interested in evaluating relative excess performance of mutual funds consistent with an asset-pricing model constituted of three or four risk factors. The descriptive statistics are given in Table II. The characteristics of the general equity market indices are quite similar, which is also reflected in the high correlation between these three indices. If we compare the characteristics of the two bond indices, we note that the average
17
daily return on the Belgian government bond index is higher than the one of the German bond index, but this goes hand in hand with a higher standard deviation and higher correlations with the equity indices.
We also estimate the correlation of fund returns with the returns on sector indices to assess whether the level of industry concentration of fund portfolios is of importance in explaining time variation in style exposures. We take the Stoxx Supersector indices that have complete data coverage over the sample period. The 15 sector indices are Oil & Gas, Chemical, Basic Resources, Construction & Materials, Industrial Goods & Services, Automobiles & Parts, Food & Beverage, Health Care, Media, Telecommunications, Utilities, Banks, Insurance, Financial Services and Technology. Descriptives are available upon request because it follows from subsequent analysis that industry concentration is not a significant determinant of time variation in style exposures. So, we will not elaborate on this issue.
Finally, a set of extra variables information variables to be used in our extended style analysis has to be selected. We only know that these proxies should be related to future investment opportunities. An extensive literature studies the predictability of both expected returns and future volatility of asset classes. In our empirical analysis we use variables that are among the most powerful predictors. More specifically, we add a January dummy variable, as well as a lagged measure of the dividend yield of the European stock index, a lagged measure of the slope of the term structure, and a lagged measure of the short-term interest rate level. We do not include the quality spread in the corporate bond market because it is not readily available for the European market over the complete data sample. The slope of the term structure is proxied by the difference between the redemption yield of the 10-year German Government index and the one-month German interbank rate. The latter is used as the short-term interest
18
rate. All data are taken from Datastream. In addition to these generally accepted information variables, we add a number of variables that prove to be sensible in our framework. More specifically, the daily data make it possible to pick up intra-month variation. We add a dummy variable to test if the exposure is different in the last week of the month. The results with respect to this dummy variable can be related to the documented turn-of-the month effect (Ariel, 1987). The effect refers to the fact that the stock returns at the turn of the month are higher. Furthermore, if mutual funds are influenced by the previous actions of other mutual funds, we expect that mutual fund returns are positively correlated with lagged returns on an equally weighted portfolio of all existing funds. We test this proposition by adding the average portfolio fund return from the previous week. If we introduce the lagged dividend yield and the lagged fund return with one lag we could run into simultaneity problems. To avoid this problem we add the dividend yield with a lag of five, and measure the lagged average fund return from t-6 up to t-2. Finally, we also add the square of the (contemporaneous) index returns to the zt vector. Of course, these variables are no information variables, but as in Ferson and Schadt (1996) we include them to capture market timing in the spirit of Treynor and Mazuy (1966).
In our empirical analysis we estimate Equation (3) with the three indices used throughout the analysis and a constant in rt. In view of the break tests (and accompanying data limitations) we estimate various models (Equation (8)) where only a single (information) variable is introduced at a time.
5. 5.1
In this section we investigate the possibility of style breaks in detail. We begin by presenting the results of the RBSA over rolling windows. In studies with monthly data a three-year
19
window is often chosen. There are no guidelines on the window to use, so the chosen window length is function of the kind of research questions addressed. Swinkels and Van Der Sluis (2001) report that in empirical work the window length lies mostly between 24 and 60 months. In Figure 1 the exposures are displayed using 180-days rolling regressions. Panel A is based on average cross-sectional data. Apparently, the portfolio restrictions that are imposed in the standard model would have been binding if included in the RBSA. Panel A reveals that substantial variation of the exposures in time is present. Both smooth and abrupt changes do occur. Given that the mutual fund sample does not remain fixed over the sample period, the observed variation may be induced by variation in the sample selection (composition of the sample and number of funds). To shed some light on this issue, panel B provides the average factor exposures for the rolling regression of five funds that exist during the complete sample period. We obtain broadly the same picture. Consequently, we rule out the possibility that the observed exposure variation is due to alternation in the sample.
The important message to remember from panel B is that variation of exposures in time continues to be present. A more rigid method of conduct is a decomposition of the variance due to Mundlak (1978). We can introduce the following types of variance (up to the constant
N. T
). Let the total variance V of a generic fund characteristic (e.g. its bond exposure) vft be
defined as
N
(v
N T f =1 t =1
2
ft
Vbf = T f =1 ( v f . v.. ) and to variance across time Vbt = N t =1 ( v.t v.. ) . The remaining part
is residual variance Vwft = f =1 t =1 ( v ft v.t v f . v.. ) . The average characteristic over time
N T 2
and funds is denoted by v.., whereas the average over funds (time) is vf. (v.t).
20
Table III reports this decomposition with respect to the exposures to the broad equity index (In2), the small cap index (In4) and the Belgian bond index (In5) for the five funds with complete data coverage.
Despite the fact that we use a moving window to compute exposures, time-series variance is the main source of the total variance for exposures with respect to the general equity index (In2) and the bond index (In5) (with 42.2% and 70.7% respectively). Time-series variance with respect to the small cap index remains high (36.6%), but is no longer the prominent source. The decomposition provides corroborating evidence that there is variation over time in the factor exposures. Therefore, there is reason to doubt the underlying hypothesis made in RBSA of constant exposures over the entire sample period. We shall investigate this possible variation more formally by applying the results of the break tests established in Bai and Perron (1998; 2003).
5.2
In this section we present the results of the structural break tests.3 Since there is no a priori reason to presume that a fund's exposure to one or more benchmarks is constant in time, we test for breaks in the matrix of factor exposures to the various benchmarks.4 We first turn to the results of the UDmax-test that examines the null hypothesis of no breaks against an unknown number of breaks that are reported in Table IV.
We would like to thank P. Perron for making the Gauss code for the structural break tests available through the Internet. Because the structural break test could be sensitive to outliers we conduct an analysis of the presence of outliers and influential observations following the procedures and guidelines developed in Belsley, et al. (1980). Chapter 2 of this book provides an excellent discussion on the topic, so we refer interested readers to this chapter for more details. In our analysis we concentrate on the statistics dfbetas, dffits, Hat-matrix and studentized residuals as defined in Belsley, et al. (1980).
21
It is clear from Panel B that all funds in the sample have at least one break during the sample period based on the UDmax test (5%-level). As a consequence, we investigate whether more than one break is present. About sixty percent of those funds have also more than one break (5%-level). Fifteen funds exhibit two breaks, nine funds have three breaks, and thirteen funds even exhibit four breaks. This result is even more striking if we take into consideration that the number of breaks is limited by the time series length of the individual funds.
5.3
In this section we take a closer look at the distribution of style breaks in time. The main question is whether style breaks are scattered in time, or in contrast are clustered in certain time periods. The principal results are summarized in Table V. Style breaks clearly occur more frequently in the second part of the sample, particularly in the last quartile. About two thirds of all style breaks are clustered in this last quartile of the sample. Panel B of Table V provides further evidence on the clustering. We indicate the degree of clustering by listing the relative number of breaks that are clustered with breaks of v other funds ( is listed in the first row). Hence, the second column of this panel indicates the relative number of funds having a break that is not clustered, the third column denotes then the relative number of funds that share a breakdate (or breakperiod) with one other fund and so on. These statistics are reproduced for daily, weekly and monthly periods. In short, this panel provides corroborating evidence that style breaks are clustered. At the daily level, the clustering is not pronounced, but this is of course a very stringent level to analyze clustering. The number of clustered funds augment rapidly when we move to a weekly or monthly level. At the weekly level, more than half of the breaks (55%) demonstrate a form of clustering, while 7% are clustered
22
with four other breaks. At a monthly level, 87% of the breaks are clustered, while about 20% of the breaks are clustered with eight other breaks. As expected from these results, 2 tests reject the null hypothesis that the style breaks at daily, weekly and monthly intervals are uniformly distributed in time (p-values lower than 1%-level in every instance).
All results indicate that clustering of style breaks is present in our sample. Clustering of style breaks could indicate that the style breaks are induced by the reaction of funds to a common factor. In Section 5.4 we investigate this possibility by analyzing if mutual fund managers make use of publicly available information in their investment strategy or whether all funds within a fund family display the same time-varying style behavior.
5.4
In this section we attempt to gain insight in the driving factors behind the time-varying investment style of mutual funds. We explore two general sources for the documented style patterns. Firstly, we explore the importance of publicly available information. In this instance, we investigate whether the style breaks are induced by mutual fund managers engaging in conditional investment strategies based on (i) predictive information variables and (ii) volatility estimates. In the second approach we focus on the organizational aspect of mutual funds and the mutual fund industry. In this case, we question whether the occurrence of style breaks is different between fund groupings where funds are divided into groups based on fund family membership, type of fund group (bank-related versus independent group), investment strategy (i.c. industry concentration) and past performance. We now turn to the discussion of the different variables and their relationship with the retrieved style breaks.
23
We have documented that 60% of the funds have multiple style breaks. More than 40% funds have even more than two style breaks during the sample periods. It appears unreasonable that a funds style breaks occur because the fund changes from manager three or four times during the sample period. A more likely hypothesis is that the style breaks are induced by economic motives, such as herding, timing, or that style breaks result from anticipating changing economic conditions. Therefore we estimate the extended style analysis given by Equation (8). Because of lack of degrees of freedom in subsamples, we will only add one information variable a time. We cannot contrast our results with the most closely related literature on conditional portfolio management given that regression coefficients with respect to the information variables are not reported in this strand of literature (e.g. Ferson and Schadt, 1996). Hence, we also take into consideration other relevant studies on asset pricing and variables related to documented anomalies. In Table VI we state the estimates of our extended style model.
As expected, the mutual funds are mainly exposed to the general equity index. With respect to the additional economic variables the results are somewhat mixed. In general, the results with respect to the effect of the extra economic variables on the general index (which is the main investment class of the mutual funds) are in line with expectations. Note that the additional terms do not necessarily result in a higher R . Because all funds in the sample have 'European equity' as investment objective, our main interest in the discussion goes out to the equity interaction terms. In the relevant literature a positive correlation between the lagged dividend yield and the stock market is put forward. The interaction term with In2 is positive but not significant, while the interaction term for the small cap index is significantly negative. The negative relation with the small cap interaction term seems somewhat odd. A possible explanation is that the mutual funds, as other institutional investors, are reluctant to invest
24
heavily in stocks with low liquidity. The expected negative relation for the short-term interest rate is retrieved for the interaction term of the general index (see e.g. Ferson, 1995). The short-term interest rate may be closely related to expected inflation. The results for the January dummy are puzzling. From previous studies (e.g. Keim, 1983; Reinganum, 1983) one would expect that mutual funds would try to benefit from the 'small-cap-in-January' effect. One line of thought consistent with the results is that mutual funds receive considerable cash inflows from mutual fund investors at the start of the year, and that the strong influx force mutual funds to invest in more liquid stocks. However, without flow data the results remain a matter of conjecture. The dummy for the last week of the month provides evidence that there exists intra-month variation in style exposure. The significant positive effect on the general index interaction term is in line with the documented turn-of-the month effect (Ariel, 1987). Fama and French (1989) document that the term spread predicts bond and stock returns. Fama and French (1993) mention that, if anything, the common variation captured by the term spread is stronger for stocks than for bonds. Hence, it makes sense that mutual funds that change their exposure conditional on the term spread focus on their main investment category. Consistent with previous literature (e.g. Kon, 1983; Chang and Lewellen, 1984; Henriksson, 1984) we find no evidence of significant market timing ability. For a discussion on the causes of the generally retrieved negative coefficient, we refer to Henriksson (1984) and Jagannathan and Korajczyk (1986). Finally, we also find that mutual funds are influenced by the average behavior undertaken by other mutual funds in the previous period.
More interesting to our central research question is whether the style breaks are induced by economic motives. If the answer is affirmative, no style breaks in the set of standard style exposures should show up if we conduct a partial break test on the extended model, where only the set of standard style exposures are allowed to vary. As mentioned before, the number of regressors becomes too large to conduct the break test procedure, if we estimate an extended model that includes all variables proxying economic motives. Therefore, we conduct
25
the partial break test based on an extended model with one additional information variable. The test could be conducted for 54 funds. Given the high explanatory power of the model with the lagged dividend yield, the results reported in Table VII fall back on this particular model.
We compare the number of breaks in a fund's style exposures in the traditional and extended style analysis. Because we only use one additional economic variable, it is unlikely that this variable captures all style breaks. Nevertheless, the reduction in the number of breaks is substantial. To investigate whether there is a significant reduction in the number of breaks, we apply the Wilcoxon (1945) matched-pairs signed-ranks test. We test the null hypothesis that the median of the population of difference (between pairs of breaks with respect to the standard and extended model) is larger or equal to zero. The null hypothesis is rejected (pvalue (0.00)). The number of funds that exhibited a high number of breaks (3 or 4) is reduced from 7 (13) to 3 (1), respectively. Overall, the results are supportive for the hypothesis that a substantial number of style breaks are induced by the fact that mutual fund managers alter their portfolio composition based on information variables. Given that our model does not include all variables proxying economic motives simultaneously, it is likely that the number of breaks will decrease even more if more economic variables are introduced.5
To test this hypothesis we ran specifications where we do include more than one economic variable (to limit the number of regressors we include only one intercept and no bond index). The preliminary results stemming from this approach are consistent with the reasoning outlined above. Nevertheless, no definitive conclusions can be drawn from these experimental tests because we only have results for a very limited group of funds for which a sufficient number of observations were present.
26
Besides the reliance on macro-economic information variables, there is another way for mutual funds to exploit publicly available information. Busse (1999) demonstrates that mutual fund managers alter their market exposure in function of market volatility. This strategy results in higher risk-adjusted performance. To investigate the importance of volatility timing we re-examine our results on style breaks. We argue that if volatility timing accounts for the observed style breaks than, by analogy with the previous tests, we should observe a reduction in the number of breaks in a partial structural break model that includes the volatility component and where only the standard benchmark exposures are allowed to vary. We opt to model conditional volatility based on the GARCH(1,1) specification with t-distributed conditional errors represented by Equation (9).6 It follows from the partial structural break model with the volatility component that volatility timing accounts for a substantial number of breaks detected in the standard model. More than half of the breaks (56%) are explained away by this extended model, compared to 39% in case of the extended model with dividend yield. Thirteen funds are free of style breaks, while there are 32, 12 and 9 funds with one, two and three style breaks, respectively. This evidence is supported by the Wilcoxon signed-ranks test. The test statistic is statistically significant when we compare the number of breaks in (i) the standard model and the model with volatility timing (-5.31, p-value (0,00)) and (ii) the model with dividend yield and the volatility model (-2.92, p-value 0.00)). We conclude that a substantial number of style breaks are induced by the fact that fund managers engage in volatility timing. This concludes our discussion on the role of publicly available information.
6
A similar approach is applied in for instance Corhay and Tourani (1994) who investigate the statistical properties of daily stock indices for five European countries. To model the conditional volatility in this paper we estimate a number of GARCH specifications, including GARCH, exponential GARCH and Treshold GARCH models. We also consider Gaussian, generalized Gaussian and t-distribution for the error terms and estimate GARCH(1,1) as well as GARCH(1,2), GARCH(2,1) and GARCH(2,2)-models. The AIC and BIC test statistics are with -13581,1 an -13552,9 amongst the lowest values. With respect to the distribution of the errors the t-distribution is selected above the normal distribution based on the log-likelihood test. The generalized Gaussian distribution is rejected based on the Kolmogorov-Smirnov test (0.0422, p-value 0.001)). The model fits the data well which is supported by several tests on model misspecification. For instance, the ARCH-test Engle (1982) with one lag equals 0.09 with a p-value of 0.76.
27
We now examine the possible impact of organizational features of mutual funds and the mutual fund industry.
Recent research has demonstrated that it is important to examine mutual fund behavior at the fund family level in order to explain mutual fund investment strategies and the organization of mutual funds. Nanda, et al. (2004), for instance, document the effect of a star performing fund in a fund family, while Siggelkow (2003) retrieves that more focused fund families achieve better performance thanks to the narrowness of product choice. In the same spirit, Elton, et al. (2004) investigate the effect on investors risk if investors opt to confine their mutual fund investments to a single fund family. They argue that funds within a fund family are more likely to behave in the same fashion than funds across families. Investment strategies of funds within a family could be similar because (i) fund managers rely on the same company research, (ii) funds are endorsed to follow a prescribed investment style set at the fund family level or (iii) fund managers use the same macro-economic information in their security selection process. Elton, et al. (2004) document that fund return correlations are significantly higher within fund families than across fund families. For instance, in case of stock funds the intra (inter) family return correlation equals 0.774 (0.734), p-value (0.000). A decomposition of the correlation reveals that the increased within-fund family correlations is primarily due to the fact that funds within families have the same stock holdings. The effect of a common response to factors like general market movements, and industry and sector factors is rather limited.
We relate these findings to our research on style breaks by investigating the style patterns at the fund family level. If funds within the same fund family rely on the same macro-economic information to alter there investment exposures we hypothesize that the documented reduction in style breaks after accounting for publicly available information (dividend yield or volatility) is not independent of the fund family variable. In other words, all funds within a
28
family experience a reduction in style breaks or not. The results are summarized in Table VIII. In our sample we have 14 different fund families that have more than one fund that meet our data requirements. The overall picture is similar for the analysis based on the extended model with dividend yield or the extended volatility model. If we concentrate on the extended model where the dividend yield variable is added, we notice that seven fund families manage funds that predominantly have a reduction in the number of style breaks after accounting for macro-economic variables. These fund families manage the majority of funds examined (29 out of 44 funds (66%)). The style exposures of four fund families appear not to be grafted on a general policy that is captured by the dividend yield variable. A first conclusion that can be reached is that not all fund families condition their investment policy on the dividend yield variable. In this instance, all funds within a family should have a reduction in the number of style breaks after accounting for dividend yield. The Wilcoxon signed rank test is -1.107 (p-value of 0.268) in case of the dividend yield model. Results for the volatility model are similar. Nevertheless, this is a rather limited test to conclude that the investment policy is not set at the fund family given that we have only tested for two variables. We can only conclude that not all fund families apply strategies based on dividend yield or volatility timing. In fact, evidence that all funds within a fund family react in a similar manner (reduction in style breaks or no reduction in style breaks) is commensurable with the fact that the investment strategy is defined at the fund family level. To examine this proposition, we perform a Wilcoxon signed rank test and a sign test on the absolute difference of funds with a reduction and without a reduction in style breaks per fund family. The Wilcoxon signed rank test equals -2.966 with a p-value of (0.000) indicating that the median value is significantly different from 0. The sign test is also equally significant. In this instance, our style break analysis indicates that the investment policy of a mutual fund is largely determined at the fund family level, although the results should be interpreted very cautiously due to the asymmetric distribution introduced in the statistical tests. Nevertheless, we believe that this issue warrants further investigation.
29
Finally, the findings presented above do not depend upon the fact that funds are part of bankrelated or independent fund groups. The 2 -value for the contingency table testing this hypothesis equals 0.189 with a p-value of 0.664. If we examine volatility timing at the fund family level we arrive at similar conclusions, so we do not the discuss the results in detail here.
We investigate the relation between the style break patterns and the investment style of a fund further by relating our results to two other documented features on mutual fund investment strategies. Brown, et al. (1996) and other studies on mutual fund tournaments argue that mutual funds alter their investment strategy during the year conditioned upon their performance in the previous period: mid-year losers have the tendency to increase fund risk.7 This kind of investment behavior is explainable in view of the relative performance contracts that are common in the mutual fund industry. In the same spirit, we could raise the question whether (some of) the style breaks are commensurable with tournament behavior. For instance, funds that want to decrease their risk could decrease their exposure to the equity factors and increase their exposure to the bond factor. To investigate whether this type of investment strategy is related to the observed style breaks we test the null hypothesis that the average performance of funds in the pre-break period (one month or six months) is equal to the average performance of funds that do not exhibit any style break in that period. However, we do not document any evidence that relates this type of tournament investment strategy to the style breaks. For instance, the mean difference in performance in the month preceding a style break between funds that have a style break and funds that do not have a style break equals -0.002 (p-value of 0.211).
A view challenged by Busse (2001) based on daily estimates of fund risk instead of monthly estimates.
30
Kacperczyk, et al. (2005) document that industry concentration is a determinant of the investment policy and performance of mutual funds. Based on portfolio holdings for a sample of U.S. equity funds they show that funds with more concentrated portfolios achieve superior performance. Concentrated funds overweigh growth and small stocks, while portfolios of diversified funds more closely mirror the total market portfolio. We investigate whether the occurrence of style breaks is related to the level of industry concentration of a fund. Regrettably, we do not have data on mutual fund portfolio holdings, so we proceed in an alternative manner. We calculate the correlation of mutual fund returns and 13 sector indices taken from the Stoxx-website ranging from banking, chemical over media to gas & oil indices. We sort funds based on the sum of the squared deviation of the correlation between the fund return and the return on the sector indices and the broad equity index return and the return on the sector indices. If funds follow the same industry allocation as the broad index this measure should be low, and vice versa. Subsequently, we compare the number of breaks between two portfolios consisting of the ten funds with the highest and lowest concentration measure. If industry concentration is of importance to the time variation in style exposures, significant differences in the number of breaks should certainly be evident if we compare these two extreme groups. It follows from Table IX that we do not find any relationship between industry concentration and the susceptibility of funds to style breaks.
6.
CONCLUSIONS
Despite the wide acceptance of RBSA, both by academics and practitioners, the method has several limitations. One important drawback is the underlying assumption that the style exposures do not vary over time. We have applied results on break tests established in Bai and Perron (1998; 2003) to examine profoundly the possibility of style breaks. An advantage of these test procedures is that we can test for multiple breaks at a priori unknown breakdates,
31
and test for breaks in a (sub)set of variables. We have applied these test procedures to a set of daily return data of European equity funds distributed in Belgium. As such, we have extended the empirical literature on the European mutual fund market, and the research relying on highfrequency data. The results did confirm the indications on time-varying style exposures that surfaced by the style analysis over rolling windows and the decomposition of variance. We have found strong evidence against the hypothesis of constant style exposures in time. All funds exhibit at least one break, while 60% of the funds exhibited even more than one break. An analysis of the distribution of breaks in time revealed that breaks are clustered, especially in the second half of the sample. This could indicate that the style breaks are induced by the reaction of funds to a common factor. The documented style patterns are mainly induced by mutual fund managers reliance of conditional investment strategies based on publicly available information variables and particularly volatility shocks. Other variables like, for instance, industry concentration or past performance appear not to be related to the occurrence of style breaks. Finally, we believe that research on the extent that investment style strategy is set at the fund family level would be welcome.
32
Selection Criteria All funds labeled 'Europe equity' Deletion of distribution funds Deletion of TAK23-funds Deletion of funds with non-daily data Deletion of funds with nobs. < 1 year Deletion of non European equity funds based on RBSA Final sample
# Funds in Efund-selection 260 167 137 106 64 53 (-93) (-30) (-31) (-42) (-11)
# Funds in Fet-selection 292 202 162 129 82 62 (-90) (-40) (-33) (-47) (-20)
53
62
33
REFERENCES
Andrews, D.W.K., Lee, I. and Ploberger, W. (1996), Optimal Changepoint Tests for Normal Linear Regressions, Journal of Econometrics 70, 9-38. Annaert, J., van den Broeck, J. and Vander Vennet, R. (2003), Determinants of Mutual Fund Performance: A Bayesian Stochastic Frontier Approach, European Journal of
influential data and sources of collinearity, NY: John Wiley & Sons Inc., New York.
Bhargava, R., Bose, A. and Dubofsky, D.A. (1998), Exploiting International Stock Market Correlations with Open-end International Mutual Funds, Journal of Business, Finance
34
Brown, S.J. and Goetzmann, W.N. (1997), Mutual fund styles, Journal of Financial
Chalmers, J.M.R., Edelen, R.M. and Kadlec, G.B. (2001), On the perils of financial intermediaries setting security prices: the mutual fund wild card option, Journal of
35
Daniel, K., Grinblatt, M., Titman, S. and Wermers, R. (1997), Measuring Mutual Fund Performance with Characteristic-Based Benchmarks, Journal of Finance 52, 10351058. de Roon, F.A., Nijman, T.E. and Ter Horst, J.R. (2004), Evaluation Style Analysis, Journal of
36
Goetzmann, W.N. and Massa, M. (2003), Index funds and stock market growth, Journal of
37
Otten, R. and Bams, D. (2000), Statistical Test for Return-Based Style Analysis, working paper, Maastricht University. Otten, R. and Schweitzer, M. (2002), A Comparison between the European and the US Mutual Fund Industry, Managerial Finance 28, 14-35. Otten, R. and Bams, D. (2002), European Mutual Fund Performance, European Financial
Management 8, 75-101.
Reinganum, M. (1983), The Anomalous Stock Market Behavior of Small Firms in January: Empirical Tests for Tax-Loss Selling Effects, Journal of Financial Economics 12, 89104. Schneeweis, T. and Spurgin, R. (1998), Multifactor Analysis of Hedge Funds, Managed Futures and Mutual Fund Return and Risk Characteristics, Journal of Alternative
Investment 1, 1-24.
Sharpe, W.F. (1992), Asset allocation: management style and performance measurement,
38
Wermers, R. (2000), Mutual Fund Performance: An Empirical Decomposition into Stock Picking Talent, Style, Transactions Costs, and Expenses, Journal of Finance 55, 16551695. White, H.A. (1980), Heteroskedasticity-Consistent Covariance Matrix and a Direct Test for Heteroskedasticity, Econometrica 48, 721-746. Wilcoxon, F. (1945), Individual Comparisons by Ranking Methods, Biometrics 1, 80-83.
39
The yearly statistics are calculated from 3 June of year t until 2 June of year t+1, except for
the period 2000-01 where the sample runs from 3 June 2000 until 25 June 2001. The Efund-selection and Fet-selection consist of 53 and 62 mutual funds respectively, meeting the criteria outlined in Appendix 1. All mutual fund data are taken from the Vision database. Panel A reports, for each selection, the average return for the equally-weighted portfolio of all existing funds over the period. Panel B reports the cross-sectional statistics for all funds with a full track record during the reported period (the fund's starting date should fall before the 30th of the first month, and the ending date should be after the 1st of the last month in the reported period).
Panel A: Descriptive Statistics for Daily Mutual Fund Return Data. Period (*) Selection 1991-01 1991-92 1992-93 1993-94 1994-95 1995-96 1996-97 1997-98 1998-99 1999-00 2000-01 # Funds Fet Efund 62 7 8 10 12 14 16 27 37 61 62 53 7 8 9 11 13 15 23 31 51 53 Average Fund Return Fet 0.061 0.024 -0.018 0.077 0.003 0.095 0.126 0.178 0.036 0.158 -0.070 Efund 0.061 0.024 -0.018 0.077 -0.003 0.092 0.133 0.182 0.039 0.143 -0.062
40
Panel B: Cross-sectional Descriptive Statistics for Daily Mutual Fund Return Data. Period (*) # Funds Fet Efund 1991-92 1992-93 1993-94 5 7 8 5 7 8 9 12 13 15 25 31 Average Fund Return Fet 0.022 -0.023 0.074 -0.004 0.096 0.128 0.182 0.036 0.160 Efund 0.022 -0.020 0.074 -0.010 0.094 0.129 Min Fet -6.565 -4.520 -4.733 -4.186 -4.654 -6.779 Efund -6.56 -4.52 -4.73 -4.19 -4.65 -6.78 Max Fet 7.838 3.929 5.128 3.452 4.866 7.292 Standard Deviation Efund 7.838 3.929 5.128 3.452 4.866 7.292 6.931 13.27 7.668 Fet 0.257 0.380 0.443 0.626 0.527 0.528 0.761 0.978 0.890 Efund 0.257 0.377 0.443 0.521 0.449 0.480 0.713 0.962 0.814
41
42
43
44
SupF (2|1) test: 45.55 (16.19) SupF (3|2) test: 58.78 (18.11) SupF (4|3) test: 23.46 (18.93)
1 break
62 62
2 breaks 15
45
Quartile 1 2
% 4 4
Quartile 3 4
% 19 74
Panel B: Descriptive Statistics regarding the Relative Number of Clustered Style Breaks (in %).
Clustered with breaks of v funds Daily clustering Weekly clustering Monthly clustering
0 75 45 13
1 21 19 10
2 4 22 22
3 6 6
4 7 4
5 13
6 5
7 6
Panel C: Kolmogorov-Smirnov Test. Monthly Clustering Weekly Clustering Daily Clustering Z-value Significance 7.23 0.00 19.12 0.00 58.76 0.00
TIME VARIATION IN MUTUAL FUND STYLE EXPOSURES Panel D: Monthly Clustering of Style Breaks.
46
47
In2
In4
In5
69% -0.001 0.736 0.201 0.084 (0.26) (13.75 (3.81) (2.03) Div.Yieldt-5 Term Spreadt-1 Interestt-1 DJanuary (In2) Dlast week 77% -0.080 0.776 0.320 -0.022 0.035 (1.30) (8.04) (3.01) (0.11) (1.19) 72% -0.010 0.703 0.108 0.149 0.005 (1.96) (33.53 (3.30) (2.61) (1.12) 62% 0.290 0.800 0.169 -0.491 -0.073 (2.67) (21.62 (1.40) (1.24) (2.65) 64% -0.004 0.762 0.076 0.159 -0.003 (1.08) (32.60 (2.85) (3.77) (0.23) 63% -0.002 0.754 0.091 0.148 -0.002 (0.29) (33.80 (3.64) (3.69) (1.08)
of month
63% -0.007 0.747 0.096 0.095 0.011 (1.52) (30.64 (3.34) (2.56) (1.24) 64% -0.009 0.759 0.103 0.144 0.024 (1.75) (32.80 (3.71) (4.41) (1.41)
48
Table VII. Number of Breaks for Individual Funds in the Different Models.
Number of Breaks Standard RBSA Extended RBSA With Dividend Yield 1 2 3 4 21 13 7 13 40 10 3 1 Median Reduction in # of Breaks 0 1 2 2 Missing Funds 4 2 2 0
49
Fund Family
# Funds
8 5 4 3 3 2 4
1 2 0 1 0 0 1
7 3 4 2 3 2 3
2 2 2
1 1 1
1 1 1
1 2 0
1 0 2
3 2 2 2
3 2 2 2
0 0 0 0
3 0 2 2
0 2 0 0
Total
44
17
27
15
29
50
Table IX. Industry Concentration and Number of Funds. We form two portfolio of ten funds with the most extreme industry concentration measure. This measure is calculated as the sum of the squared deviation of the correlation between the fund return and the return on the sector indices and the broad equity index return and the return on the sector indices. We test the null hypothesis that the median number of breaks for these two portfolios is equal by means of the Mann-Whitney U test. Model with Information Variable 2 1 .00 1 3 1 .00 1 33 (0.356) Model with Volatility Timing 4 1 .50 1 4 2 .00 1 32 (0.315)
Standard model