## Are you sure?

This action might not be possible to undo. Are you sure you want to continue?

**Return-based style analysis with time-varying exposures
**

L. Swinkels, P.J. van der Sluis January 2006-2006/03 ISSN 1871-2665

*Views expressed are those of the individual authors and do not necessarily reflect official positions of ABP

Return-based style analysis with time-varying exposures¤

Laurens Swinkelsy Robeco Group

Pieter J. van der Sluisz Free University Amsterdam ABP Investments January 5, 2005

We would like to thank Dan DiBartolomeo, Jenke ter Horst, Siem Jan Koopman, Roderick Molenaar, Theo Nijman, Geert Rouwenhorst, Marno Verbeek, and participants from the 7th International Conference of the Society for Computational Economics at Yale University, the AFFI Conference in Namur, SIRIF Conference on the Performance of Managed Funds in Edinburgh, ERIM Conference in Rotterdam, North…eld’s 15th Annual Research Conference in Yosemite, 1st European Quantitative Forum in Geneva, APFA Conference in Tokyo, EFA Tutorials in Berlin, the University of Amsterdam, and Forecasting Financial Markets 2003 in Paris for their helpful comments. The views expressed in this paper are those of the authors and do not necessarily re‡ect those of our employers and colleagues. y l.swinkels@robeco.nl z Corresponding author. Dr. Pieter Jelle van der Sluis, ABP Investments, Research Department, WTC Schiphol Airport, Tower C 5th Floor, P.O. Box 75753, NL-1118 ZX Schiphol, The Netherlands, phone +31 20 405 4343, fax +31 20 405 9809, e-mail: pj.van.der.sluis@abp.nl

¤

**Return-based style analysis with time-varying exposures
**

Abstract

This paper focuses on the estimation of mutual fund styles by return-based style analysis. Often the investment style is assumed to be constant through time. Alternatively, time variation is sometimes implicitly accounted for by using rolling regressions when estimating the style exposures. The former assumption is often contradicted empirically, and the latter is ine¢cient due to its ad hoc chosen window size. We propose to use the Kalman …lter to model time-varying exposures of mutual funds explicitly. This leads to a testable model and more e¢cient use of the data, which reduces the in‡uence of spurious correlation between mutual fund returns and style indices. Several stylized examples indicate that more reliable style estimates can be obtained by modeling the style exposure as a random walk, and estimating the coe¢cients with the Kalman …lter. The di¤erences with traditional techniques are substantial in our stylized examples. The results from our empirical analyses indicate that the structural model estimated by the Kalman …lter improves style predictions and in‡uences results on performance measurement.

Keywords: Dynamic models, Kalman …lter; Mutual funds; Style analysis

JEL classi…cation: C22; C61; C63; G11; G20; G23

1

1

Introduction

The investment style of a mutual fund is not always clear for investors not acquainted with its manager or the philosophy of the fund family it belongs to. Due to the large number of mutual funds these days, it is almost impossible for an investor to grasp all information regarding their investment styles. Although the investment style of a mutual fund potentially has many dimensions, in our view its sensitivities to risk factors are the most relevant for investors.1 These sensitivities, or exposures, gauge the e¤ect of the return on the style or risk factor on the return of the mutual fund. For a potential investor in mutual funds, a tool introduced by Sharpe (1992) might be appropriate to get a …rst impression of the historical exposures of the fund. These can be used to determine the bene…ts of investing in the fund, for example by predicting style exposures, or analyzing the manager’s skills. Sharpe’s method for analyzing mutual funds is known as return-based style analysis (RBSA). Basically, RBSA is constrained regression of the returns of the mutual funds on relevant style indices. In academic research, variants of RBSA are used in several applications. These applications include the classi…cation of mutual funds, performance evaluation conditional on investment style, and optimal portfolio choice for fund of funds.2 In this paper, we aim to improve the accuracy of the estimation or prediction of style or risk exposures by using fund and index returns only and a valid statistical model. Whereas stylized examples indicate that our method is more advantageous in several circumstances, the empirical applications are somewhat less convincing. A major drawback of RBSA in its original form is the basic assumption that the investment style of a fund remains …xed over the sample period. The use of so called rolling regressions alleviates this drawback to a certain extent. In a rolling regression the investment style is not …xed over the entire sample period but over a given estimation window. In 2

There is no theoretical argument that defends the use of rolling estimation windows when styles vary over time. the length of this window is often chosen to lie somewhere between 24 and 60 months. In order to estimate this dynamic model.empirical work. Furthermore. we use the Kalman …lter approach. The main advantage for this application is the more e¢cient use of available information. which has several advantages over standard regression techniques. we distinguish between the Kalman …lter and Kalman smoother. no choice of window length has to be made. The Kalman …lter optimally determines the weights of each of the observations in determining the exposure. and the estimation technique relies solely on the data to gauge the importance of time-variation for each of the mutual funds. The di¤erence between the two is the conditioning information set. the model speci…cation can be tested and con…dence bounds on the 3 . In this paper. Funds that change their exposures frequently require a shorter window. while the smoother is conditional on the entire sample. while for funds with …xed exposures the use of a longer window generates more precise exposure estimates. given the imposed model structure. In practice the use of rolling windows causes a sub-optimal use of the data by picking an ad hoc window size. This window is shifted month by month over the entire sample period. The Kalman …lter is a method that endogenizes the weighting scheme. Throughout this paper. and hence more suited for descriptive purposes. we contribute to the discussion by explicitly modeling time variation in the investment style of a mutual fund. while allowing for time variation in exposures. We examine several applications and indicate which of the two conditioning sets seems to be the most appropriate. A method with an endogenously determined weighting scheme for historical observations alleviates this problem of choosing a window size. The structural model allows for time-variation. The …lter is conditional on information up to time t and thus more appropriate for prediction. Consequently.

While temporary deviations from the style are often observed. The speci…c needs of investors are re‡ected by a variety of funds with di¤erent investment objectives. Chen & Lakonishok (2002). For a large part of the funds the fund’s name describes its style rather adequately. It is not obvious why mutual funds are unclear about their investment policy. and Chan. Despite the generally poor performance of mutual fund managers. Finally. The structure of the remainder of this paper is as follows. These issues are not straightforward for the traditional techniques. Section 6 concludes. However. see e. the model is described and the similarities between the traditional models and the Kalman …lter approach are indicated. since potential investors need detailed fund information to construct an optimal portfolio of mutual funds. or pursue a di¤erent style than advertised. In section 3. 2 Mutual fund misclassi…cation The number of mutual funds has increased rapidly over the last decade. Consequently. a substantial part of the funds have misleading names. individual investors have increased their demand for investment management. In Section 4. vague investment objectives. A possible reason for vagueness in the stated objective of the mutual fund is lawsuit avoidance. Section 5 contains empirical applications in which we analyze the investment style of two samples of mutual funds. Gruber (1996). the o¢cial investment objective is rarely changed. Section 2 motivates the use of style analysis. Uncertainty about the style of the mutual fund obscures investor’s decision and may be a hindrance for them.g. To illustrate 4 .exposures are readily obtained by applying the Kalman …lter. the stated objectives of the funds should not be too stringent and possess a certain degree of vagueness. three stylized examples are presented in order to compare our approach with traditional rolling window regressions.

g. Its main investment objective is stated: Under normal market conditions. percent stocks) than risk and return measures (e. Sirri & Tufano (1998) indicate that mutual funds with a high rank in the performance lists of magazines attract more money from the investing public. It is our task to be the judge in this contest and identify the large tomatoes from the painted cantaloupes.this we quote from the 2001 prospectus of the Templeton World Fund. At least 65% of its total assets will be invested in issuers located in at least three di¤erent countries (including the U.g. Another possible reason for misleading fund names or objectives is to blur the investor’s notion of the riskiness of the strategy. These studies do not take into account style changes.. DiBartolomeo & Witkowski 5 . Taking on more (undiversi…able) risk usually leads to higher expected returns. In other words. For instance.). poor performing funds are not punished by massive investor out‡ows.S. There is ample evidence of the misclassi…cation of mutual funds. standard deviation. identi…cation of exposures to relevant style or risk factors is of primary importance for investors. For example. Their results suggest that up to 40 percent of mutual funds are in one way or another misclassi…ed. These results might be an incentive for a mutual fund manager to indulge in more risky asset categories.3 As DiBartolomeo & Witkowski (1997) phrase it: The easiest way to win a contest for the largest tomato is to paint a cantaloupe red and hope the judges do not notice. the Fund invests mainly in the equity securities of companies listed anywhere in the world. income ratio.. including emerging markets. both DiBartolomeo & Witkowski (1997) and Brown & Goetzmann (1997) use the realized fund returns as inputs for their analysis. Kim. or use only a limited amount of data to estimate the investment style of a mutual fund. CAPM-beta). Shukla & Tomas (2000) report misclassi…cation up to 50 percent when also taking into account other fund attributes (e. Apparently.

and 6 . 3 Determination of the investment style The RBSA introduced by Sharpe (1992) concentrates on estimating a portfolio that could have been tracked by an investor at relatively low cost. Studies focussing on the performance evaluation of mutual funds with an international investment objective are relatively scarce. We argue later that the short-sale restrictions are in general irrelevant. The importance of incorporating time-variation in style exposures before and after a change in fund manager is investigated by Gallo & Lockwood (1999). They claim that 65 percent of the funds experience a shift in investment exposures after a management change.(1997) and Brown & Goetzmann (1997) take 60 and 24 months. or value factors. We follow DeRoon. Brown & Van Harlow (2002) …nd that mutual funds with low tracking error relative to their style benchmark outperform funds with style drift or high tracking error. Therefore. RBSA where both sets of restrictions are relaxed is named weak RBSA. investors tend to invest more in stocks they are familiar with. Second. Managers from international investment funds might have the same bias and therefore their funds could be misclassi…ed because they have or had a home bias. short sales are prohibited. respectively. we analyze the regional risk exposures of funds with an international equity investment objective. size. Kuo & Satchell (2001) report that country factors provide more diversi…cation opportunities for investors than industry.5 An intermediate form with the portfolio restriction but without the short-sales restrictions is labeled semistrong RBSA. in order to interpret the exposures as portfolio holdings. This tracking portfolio is constrained in two ways. their sum should equal one. Nijman & TerHorst (2004) in their terminology in which they name RBSA subject to both constraints strong RBSA. First.4 According to French & Poterba (1991) and Huberman (2001).

Usually however. This (equality) restriction is K X i=1 ¯ i = 1: (2) The equations (1) and (2) together are called semi-strong RBSA.should be imposed in certain special cases only. The interest from our style analysis is not the actual holdings of a mutual fund. two restrictions are imposed. T. for all important purposes. (1) f indexi the return on style index i. but the exposures to certain style categories. which imposes that all estimated portfolio holdings should be long 7 . This is relevant information for an investor. t t t = 1. the exposures of the mutual fund are obtained by minimizing the sum of squared errors of the equation f index1 Rt und = ® + ¯ 1 ¢ Rt + : : : + ¯ K ¢ RindexK + "f und . This means that a portfolio of US stocks which are also sensitive to the European market results in an exposure towards both these markets. Sharpe’s famous “Duck theorem” applies here: If it walks like a duck and talks like a duck. : : : . it is a duck. Our empirical examples are based on the semi-strong form of RBSA. So. The …rst restriction is the portfolio restriction which requires that the estimations for the parameters ¯ i can be interpreted as portfolio holdings in style i. A second restriction is the short-sales restriction. since she is now also exposed to risks associated with the European market conditions. The standard assumption that the error terms are t independent from the style indices is made. In weak RBSA. and Rt and "f und the fund speci…c error terms. where Rt und denotes the total return of the fund.

: : : . and Kim. when a 36-month window is used. It states that short sales in style categories are not allowed. These (inequality) restrictions are ¯ i ¸ 0. ¯ K . Implicitly. The window moves forward one month.positions. Attempts to resolve this problem have been made by Lobosco & DiBartolomeo (1997). the coe¢cients are estimated over the …rst three years of the sample. Strong RBSA is obtained by equation (1) together with restrictions (2) and (3). Stone & White (2000). the use of sub-samples instead of the entire sample means the implicit introduction of time-varying exposures in an ad hoc manner. In this way. For instance. Especially in case of a change in fund management 8 . time-varying exposures are obtained. ¯ K are not estimated over the entire sample period but over windows (sub-samples). Because of the inequality constraints. K: (3) This does not mean that short sales in general are prohibited. the coe¢cients cannot be obtained by applying ordinary least squares (OLS). con…dence regions are not readily obtained when using QP. rolling regressions are often reported. : : : . However. the exposures ¯ 1 . : : : . i = 1. One of the implicit assumptions of the original RBSA is that the exposures stay constant over the sample period. this creates a contradiction unless the exposure is constant over the entire sample period. This is the Sharpe (1992) method of determining a fund’s investment style. amongst others. Quadratic programming (QP) algorithms solve for the exposures ¯ 1 . Therefore. deleting the …rst observation and adding the observation of the next period. and that is exactly the assumption it should relax. as opposed to OLS. In this way. However. This is highly unlikely in practice. This approach still assumes that style exposures stay constant over the 36 months estimation period.

t » NID(0.» j = 1. K where NID indicates an independent sequence of normally distributed random numbers. The exposures ¯ i. and IK+1 . our approach is able to introduce time-variation by using the return information both before and after that particular period. Another shortcoming of a rolling window approach is that estimates rely on historical returns only at each point in time. : : : . First. so without the short sale and portfolio constraint. the error terms are "t » NID(0. We propose an approach di¤erent from the traditional rolling window regressions to model time variation in fund exposures. · ¢ IK+1 ).t ¢ Rt + : : : + ¯ K. (®. (4) (5) (6) ®t+1 = ®t .7 More speci…cally. This means that the model allows the 9 . which states that the exposure in this period is the exposure in the previous period plus a style shock.t ¢ Rt + "t . with · a large but …nite number.t+1 . using the Kalman …lter. " » j. ¾2 ). The initial conditions for the exposures are (proper) di¤use priors. ¾ 2 ). we explain this method in the weak RBSA form. T: Furthermore.t+1 = ¯ i. Time variation is introduced to the model in equation (1) by the following set of equations f index1 indexK Rt und = ®t + ¯ 1. : : : . K. j. see Gallo & Lockwood (1999) for a discussion on changing style exposures after management changes. for i = 1.6 When the particular application allows this.t are non-stationary in this model. the identity matrix of dimension K + 1: We decide to model the time variation in exposures as in equation (6). ¯ 0 )0 » N(0. :::.t + » i. and t = 1. ¯ i.the assumption of constant exposures within the 36 months window is restrictive.

we have modeled the manager ability in equation (5). 408. The variance of the disturbances can be estimated e¢ciently by using maximum likelihood. and (b) 10 . Benson & Eger (1982). to be constant over time.exposures to evolve in such a way that fundamental changes in the investment style can be accommodated. Applications modeling time-varying market exposures by the Kalman …lter include Alexander. see Harvey (1993) p. Models in state space form have been used extensively in engineering. while the ¯ t -s are derived recursively. Thus. The time-invariant weak RBSA regression model can be obtained as a special case of the models presented above which allow for time-variation. these models have found their way into econometrics and …nance. Extensive treatments of the theory and applications of the Kalman …lter in the …eld of econometrics are Harvey (1993).8 Model (4)–(6) is known as the random walk model. amongst others. transition equation (6) is replaced by ¯ i. Modeling the time-varying exposures explicitly leads to (a) a testable model.t+1 = ¯ i. with equation (4) being the measurement equation. which returns the parameters of the model. This enables us to directly apply the Kalman …lter. However. which is essentially the same).t : The disturbance term » has disappeared (or has zero variance. while the hyper-parameters (¾2 and ¾2 ) are estimated to obtain the model structure. and (5) and (6) being the transition equations.» unobservable exposure coe¢cients ¯ t are the variables of interest. including the ¯ t -s. and Black. More recently. Fraser & Power (1992). The same could be done for the manager ability. Lockwood & Kadiyala (1988). Fisher & Kamin (1985). the " j. When next to the estimates for ® also the ¯ i -s are constant over time. and Durbin & Koopman (2001). The model presented above is in state space form. since they represent the unknown sensitivity of the mutual fund to the style or risk factor.

g. Such test can be important. since misspeci…cation may lead to erroneous inferences. Results from analysis based on either method will be described in the empirical sections of this paper. This is a fairly simple operation. We argue in line with DeRoon et al. When our model is appropriately chosen. our empirical results are based on semi-strong RBSA. and has little impact on the estimation procedure. the arbitrage pricing theory (APT) developed by Ross (1976). using these restrictions may lead to inconsistent parameter estimates when short positions are allowed in practice.9 In accordance with the existing literature on performance evaluation. In fact. the optimal use of the available data is guaranteed by applying the estimation technique we propose. The advantage of this approach is the direct link to asset pricing models like e. a typical growth fund 11 . all exposures sum to one). in the state space model. we incorporate this portfolio restriction by modeling the returns of the mutual funds as well as the returns on the style indices in excess of the risk free rate. A direct statistical comparison with the results from rolling window regressions is complicated. This is equivalent to the inclusion of cash as a style index and incorporating the portfolio restriction. In the empirical section of this paper we test the model with time-variation in exposures relative to a model without and …nd compelling statistical evidence in favor of a model speci…cation with time-varying exposures.e¢cient use of the data because of the structure that is imposed by the model. the betas should be reparameterized. to say the least. The possibility to test the validity of our model assumptions against alternatives makes this technique attractive. For example. The investment styles serve as the risk factors relevant for the potential investor in the mutual fund.e. Thus. In order to incorporate the portfolio restriction (i. (2004) that imposing the non-negativity restrictions is in general not necessary. as is used in the (semi-) strong RBSA literature.

Agarwal & Naik (2000) suggest omitting the restrictions when applying RBSA to hedge funds.will have a negative exposure for the value (HML) factor in the Carhart (1997) model. The use of inappropriate short-sale restrictions may lead to biased or even inconsistent estimates of the exposures. While mutual fund managers are often not allowed to take (large) short positions. a comparison can be made between the estimated exposures by the Kalman …lter approach we advocate 12 . Applying these models to more dynamic hedge funds is left for further research.10 Furthermore. This way. This is clearly at odds with the goal of determining the true underlying investment style of the mutual fund. When the indices representing the investment styles are inappropriately chosen. the exposures might become negative even when there is no reason to believe this is the case in reality. the necessity for non-negative exposures is less clear. We suggest a careful selection of the style indices before applying RBSA. This means that style indices should have low correlation and describe as much as possible from the investment opportunity set of the mutual fund manager. An example of an application in which the non-negativity constraints are useful is when the estimated investment style should be replicable for an investor with short-sale constraints. high correlation (or near-multicollinearity) or the omission of a relevant asset class. 4 Stylized examples The performance of the method described in the previous section can be analyzed by stylized examples in which we a priori …x the exposures of the managed fund. which is the focus in our paper.g. given that they have the same (non-negative) style exposures. The performance of this constrained investor can be used for comparison with the performance of the mutual fund. In most other applications. e. this is common practice for hedge fund managers who have almost no limitations in this respect.

We start by describing the stylized examples. Again. It takes three years for this fund to move from a portfolio mimicking the MSCI USA to a portfolio 13 . In the previous two stylized examples the exposures change dramatically at a certain point in time. we present three examples.or the rolling window regressions that are used in many other empirical applications. We analyze how this sudden change is re‡ected in predicted exposures and style estimates by the methods described above. January 1976 to May 2002. and analyze the prediction and explanation quality of the models in the subsections below. we assume a tracking error of one percent as in the previous example. The style of this fund is to mimic the MSCI USA in even years and in the MSCI Europe in odd years. these examples might provide more insight in the pros and cons of the di¤erent methods when estimating the style exposures of mutual funds. In this section. In the third example. The three arti…cial funds we present are a fund that radically changes its exposure each decade. the fund only slowly changes its exposures. and a fund that changes its exposure very slowly. The second example consists of a mutual fund with highly volatile exposures. However. The sample period we use in our example corresponds to the sample period of our empirical applications in the next section. we estimate the random walk model as described by (4)–(6). The deviations from the benchmark are independently and identically and normally distributed. The …rst example is a fund that alternatingly mimics the MSCI USA and the MSCI Europe at the turn of each decade. Over the entire sample period. a fund that radically changes its exposure each year. These examples are stylized and most likely do not resemble any of the strategies of existing mutual funds. we assume a tracking error of one percent per month. This stylized example shows how the methods behave when investigating a mutual fund with highly volatile exposures. For these three stylized examples.

This fund also has a tracking error of one percent per month. 4.g.g.12 This method uses information from month t ¡ 35 to month t in order to predict the exposure for month t + 1: The estimated exposure over the window is used as an estimator for the future exposure. The conditioning information is used to predict the exposures for period t + 1: The benchmark case is the 36-month rolling window estimator. which can be used for. e. we make a distinction between predicting exposures. Ferson & Schadt (1996)) which may spuriously generate good in-sample predictions.mimicking the MSCI Europe. MSCI USA and MSCI Paci…c11 ) and the (excess) fund return up to time t. In contrast to other studies. performance measurement. risk management purposes. In contrast with the empirical examples investigated in the next section. e. which can be used for.. e. The essential di¤erence between predicting and explaining is the set of conditioning information which is used to estimate the exposure at a particular point in time. The use of future information could be incorporated either explicitly or implicitly. After three years of increasing European exposure. we know the 14 .g.1 Predicting regional exposures Their is widespread agreement on the bene…ts of hindsight for prediction purposes. we do not use other conditioning information such as macro variables (see.. In this subsection. In the two subsections below. The information used to predict the regional exposure of the funds are the returns (in excess of the riskfree rate) on the three regional style indices (MSCI Europe.. we carefully mention which information we use in order to make predictions about the future regional exposures of the stylized funds as described above. the reverse change takes place at the same pace. and explaining returns.

t+1 of our model.S. the MAD and MSD from (7) and (8) are displayed. In Table 1.3 percent for the rolling window estimator. ¯¯ T t=0 p.t+1 where we let t = 0 correspond with the …rst period with an estimate for the rolling window method. The …rst is called the mean absolute deviation (MAD). The average style forecast residual is more than three times as large for the rolling window predictor. the rolling window estimator performs worst regardless of the types of exposure change in the examples. the MAD equals 19. 15 .0 percent. the criteria are T ¡1 ¯ 1 X ¯bU S ¯ ¯ ¯¯ p. In the …rst example.t+1 ¡ ¯ U S ¯ . exposure (¯ p. Panel A.t+1 ) with the actual U. as the MSD puts more weight on large prediction errors. while the Kalman …lter does better with 6. exposure (¯ U S ). As can be seen from the table. We investigate the accuracy bU S of the estimation method by comparing the predicted U. on average.actual exposures of the mutual funds in these stylized examples. We compare these measures of forecasting the style of the mutual fund with the di¤erent prediction methods. The MSD measure is the sample version of the variance. However. which has appealing statistical properties in the sense that under commonly-made assumptions it is a better measure for dispersion than any other.0 percent absolute error. p. In such cases the MAD is a better measure since all prediction errors are equally represented.S. in practice.t+1 T t=0 MADp = MSDp = (7) (8) T ¡1 ¯2 1 X ¯bU S ¯ US ¯ ¡ ¯ p.t+1 ¯ . for a single (…nite) sample this is of less importance. Formally. it is highly in‡uenced by the largest prediction errors. and the second the mean squared deviation (MSD). We use two distance measures to evaluate the forecast performance p. This means that the exposure is predicted with only 6.

MSCI USA. ® ¯b t+1 + Xt+1 ¯ ¯ T t=0 where we recall that in our example Xt+1 is a vector of size 3.t+1 ¯ . T t=0 The measures from (7) and (8) change to d MADp = (9) (10) d MSDp = T ¡1 ¯2 1 X¯ ¯ bp. with 31. The relative improvement is largest when the styles change slowly over time. and MSCI Paci…c (including Japan). In the stylized examples.1. In the third example.t+1 ¡ Rp. These measures are also calculated for the stylized funds and the results are displayed in Panel B of Table 1.t+1 ¯ .t+1 ) times the actual style returns at t+1 (Xt+1 ): ® T ¡1 ¯ 1 X¯ ¯ ¯ b b ¯®t+1 + Xt+1 ¯ p. which is substantially 16 . the average forecast error of both methods are substantially higher. this is obviously not possible anymore. The MAD-gain from using the Kalman …lter in this example is reduced relative to the rolling window estimator. The predicted returns are de…ned as the predicted intercept b (b t+1 ) and the predicted style exposures (¯ p. A comparison between the Kalman …lter and rolling window predictions are again in favor of the former. While the absolute gain is more than 20 percentage points. we are able to compare the estimated style exposures with the real exposures. the relative improvement of 40 percent is a little less than in example one. we choose to compare the predicted returns from our style exposures with the actual returns (Rp.[Insert Table 1 here] Since the second example has highly volatile style changes. with slow moving style exposures.t+1 ). For the empirical examples analyzed below.1 versus 52.t+1 ¡ Rp. The gain in MAD for the third example is over 20 percent. Therefore. containing the returns on the regional benchmarks MSCI Europe. and more than four times as accurate when using the MSD-measure. the Kalman …lter predicts the style twice as accurate when using the MAD-measure.

where at each point in time each style estimate is weighted by its corresponding style return. 4. In fact. such as Carhart (1997). using both information from the past as well as the future. Thus. In order to use the entire sample and still introduce time-variation in exposures.2 Describing regional exposures When the aim of the analysis is a description of regional exposures rather than predicting future exposures. where the style exposures are modeled as random walks.lower than the improvement of 50 percent reported on the basis of the test directly on the style exposure itself. this is standard practice for performance evaluation studies. is frequently used to measure the bene…ts of investing in the mutual fund. actual improvements of the style estimates on basis of this derived statistic might be larger than the percentage di¤erence in this derived statistic. The estimation of Jensen’s alpha conditional on risk exposures to an asset pricing model. This basically is the Kalman …lter used twice. Applications for which the inclusion of future information is harmless is performance evaluation conditional on the style. the prediction is improved in each of the three circumstances. The information used to estimate the risk exposure is based on a regression over the entire sample. we use the Kalman smoother. These examples suggest that using the Kalman …lter to forecast exposures is bene…cial for applications in mutual fund style analysis. …rst forward in time. and when the end of the sample is reached the …lter is applied backwards through time. the conditioning information set increases. In these stylized examples. but in contrast to the existing models. When the quality 17 . This is the optimal way of using information given the imposed model. it allows these exposures to vary within the estimation sample. Our structural model also uses return information from the future to determine the style.

the accuracy of the smoother seems higher than for the …lter or the rolling window estimator. In Figure 2 Panel A the rolling regression yields an exposure that is around 0. This is supported by more quantitative measures. the MAD and MSD are presented between the real exposure and the estimated exposure. As we will see in the examples discussed below. Panel B shows that the exposure already starts increasing a year before the actual change is taking place.5 where in fact the exposure is switching between 0 and 1. it becomes clear that the smoother has more trouble capturing sudden changes in exposure than the …lter. the model might be misspeci…ed.S. Panel B displays the …lter. the smoother. Figure 1. As expected. as can be seen in Table 1. because the smoother anticipates the style change by using future information. and the true exposure. By visual inspection. We also note that the Panels A of Figures 1-3 show the drawbacks of the rolling regression. The smoother is almost 18 . In Figure 1 and 2. Each …gure corresponds to the above three examples.e. 36-month rolling regression and the true exposure. In each of the …gures Panel A displays the …lter. In Figure 1 Panel A we clearly see the e¤ect of dragging the whole 36 month history in the rolling regression method. [Insert Figures 1 to 3 here] In Figures 1 to 3 we display the estimated exposures to the U. However.of the …ltered style estimates is better than the smoothed estimates. Figure 1 corresponds to example 1 and so forth. i. in the case of jumpy exposures this misspeci…ciaton can be observed. structural shifts in the exposures can be captured well by our model. and only when the last observation from the old regime leaves the regression window after 36 months the estimates fully re‡ect the new regime. In that table. The exposure only slowly adapts to the new regime. In Figure 2 we see that the smoother has some di¢culty capturing the dynamic exposures of the fund. given the slight misspeci…cation (the stylized exposures are not random walks)..

or (b) an ad hoc method to account for time-variation. we apply the techniques introduced in the previous section to real mutual fund exposure prediction and performance attribution. the same disadvantages that we mentioned in the previous sections apply to this rolling window estimator. which are both displayed in Panel B. One could think of an ad hoc solution to incorporate valuable future information by using the 18 months before and 18 months after a certain period to obtain the 36-months rolling window estimator. The lines with true exposures and estimated exposures deviate only close to the turning point. we observe that when exposures change smoothly over time. The evidence presented in this section is suggestive in the sense that existing style estimates are inaccurate either due to (a) the lack of accounting for time-variation. we conclude that the use of the smoother is superior to that of the …lter when exposures are varying slowly over time. and hence we propose the use of the Kalman smoother instead. but in both other examples it outperforms the …lter and rolling window estimator in capturing the exposures. The results from this section show that the rolling window approach which is used in many empirical applications can be improved upon when the coe¢cients are changing over time. In Figure 3. However. the smoother is able to capture the dynamics accurately. Comparing the smoother to the …lter. In the next section. it is certainly not standard when time-variation is introduced. Not surprisingly.equal to the …lter in the …rst example as far as MAD is concerned. 19 . Whereas the use of future information is common for performance evaluation with …xed exposures (see Carhart (1997)). the use of an expanded information set renders results closer to the true results. Panel B.

fund managers might be home biased (see e. we examine the regional exposures of a sample of mutual funds with an international or foreign investment objective. based international mutual 20 .13 5. others choose to follow a value weighted international benchmark.g.S. The second empirical example analyzes the exposures to the market. we have analyzed the di¤erences between the traditional rolling window estimators and the Kalman …lter by three stylized examples. i. These issues warrant more insight in regional exposures of international mutual funds which motivates our sample selection criterion.5 Empirical applications with time-varying exposures In the previous section. The funds in our sample are based on prior research on U. These results suggest that the use of this technique may improve existing results in the mutual fund literature. we apply our technique to the prediction of styles. and size factor of a sample of asset allocating funds.1 Regional exposures of international mutual funds In this section.e. market timers. However. We start this section by a short description of the data. Next. we examine the ability of mutual fund managers to change their exposures towards markets with high returns. In addition. Heston & Rouwenhorst (1994) and Kuo & Satchell (2001) report that the largest diversi…cation bene…ts for investors can be obtained by investing internationally. French & Poterba (1991) or Huberman (2001)) and invest disproportionally in domestic stocks. Finally. value. research on international mutual funds is scarce compared to domestic equity funds. we examine two empirical applications of our method. While some international funds keep their regional exposures …xed over time. In this section. The …rst empirical example is the regional exposures of a sample of mutual funds with an international or foreign investment objective.

35 percent per month. The lowest monthly return is obtained by the Paci…c.funds.18 percent. The average returns of the funds range from 0. We have monthly total return data from January 1976 until May 2002. [Insert Table 2 here] The data on the regional indices are from Morgan Stanley Capital International (MSCI).15 Descriptive statistics on the fund from our sample can be found in Table 2. The standard deviations of the returns are between 2. which lost more than 40 percent of its value in one month. The predicted returns are calculated as the predicted exposure times the realized return on the benchmark indices. and contain reinvested dividends. we need another measure to check for the quality of our predictions. Kolodny & Resnick (1991) with an international or foreign investment objective. We take the funds used in Cumby & Glen (1990) and Eun. We chose The data on the risk free rate are obtained from the website of Kenneth French.87 to 1. [Insert Table 3 here] The analysis below focuses on the prediction of regional exposures for the mutual funds in our sample.75 and 5. The correlation between the four indices is in general modest. Descriptive statistics of the regional indices can be found in Table 3. almost 25 percent. The US has had the highest average returns and the lowest volatility over the sample period. Europe and the US top the list with just over 63 percent.14 We collect our mutual fund data from Morningstar. Since the real exposures16 are not available. We measure the di¤erence between the realized return and the predicted return by the rolling window and Kalman …lter method. The sharpest increase within a month is in Japan. For reasons of brevity we do not provide graphs of the regional exposures for each of the 21 .

as can be seen in Figure 4. the smoother. [Insert Figure 4 here] The prediction quality between the …lter and the rolling window estimator is analyzed in Table 4. the economic magnitude of the improvement is generally not substantial. with a p-value of 0:002. while our method statistically improves the style estimates in most cases. the Vanguard International Growth Fund. This is again con…rmed by the BusinessWeek interview where the manager links the problems of a weakening yen in Japan to other Asian countries. In 11 out of the 12 mutual funds from our sample. This is con…rmed in the same interview where the fund manager expressed his views on a weakening Yen and restructurings. An interview in BusinessWeek in the spring of 1999 with the fund manager of this Vanguard fund provides additional information to the exposures17 .7 percent. we expect no exposure to the US market. the …lter. This result is obtained from the rolling window estimator. The European exposure of the Vanguard fund increases over the period 1995-2002. 22 . while the average MSD is reduced by 9.funds in our sample. and provide tabulated results for the remaining funds. the predicted returns improve when the Kalman …lter is used. Since this fund is a foreign investment fund. and holdings information. The exposure to the Paci…c index (Panel D) is also decreasing from 1995-2002. Thus. The exposure to Japan (Panel B) has decreased from 20 percent to 10 percent. In the interview the fund manager reveals his bias towards Europe at that time.0 percent. Panel A. Instead. we examine only one fund in detail with graphs. Panel C shows that there is a slightly negative exposure over the latter part of the sample. The average gain in MAD is 3. A (pairwise) Wilcoxson signed rank test rejects the null hypothesis of equal medians for both methods. where the most notable change takes place in 1998.

Vanguard. A more detailed analysis investigating regional timing is beyond the scope of this paper. When the fund sticks to its benchmark. we use the entire data set to estimate the exposure in a certain period. This timing ability can be analyzed by the correlation between the exposure and the returns. the method used to calculate timing ability has an impact on the …nal results on timing. In particular. the regional exposures change over time due to the changing weights of the regions in the index.[Insert Table 4 here] We stress that a shift in regional exposures need not be a shift in investment style of the mutual fund. the signi…cant positive timing for Japan in the rolling window estimator for American Funds. Rowe Price. the signi…cant negative timing for Europe for T. Since we are not concerned with prediction in this application. Nevertheless. The use of this future information is in similar spirit to the current literature on mutual fund manager ability. On the other hand. The table shows that inference on regional timing is in‡uenced by di¤erent style estimation methods. Rowe Price and Vanguard also vanish when the smoothed analysis is used. This allows us to use the Kalman smoother instead of the Kalman …lter. The correlation analysis is displayed in Table 5. Hence. for example the MSCI World. It is relevant for potential investors to investigate whether international mutual funds are able to anticipate abnormal high (low) regional returns by increasing (decreasing) their exposure towards these regions. A study investigating the regional timing performance of mutual funds requires an accurate description of the regional exposures. Such change in exposure should not be confused with a change in investment style. T. 23 . when time variation of exposures is allowed the use of rolling window estimators (which do not use future information) is common practice. see Carhart (1997) amongst others. and Waddell & Reed is not found when the Kalman smoother is used.

The funds in this section explicitly state in their objectives that they are changing their exposure to risk factors. The …gure shows that for this type of funds. This is done. We investigate whether the exposures to the risk factors in Fama & French (1993) can be predicted by the Kalman …lter method. by investigating the di¤erence between the actual fund return and the predicted factor exposures times the realized factor returns. Figure 5 contains the sorted percentage improvement of the MAD measure for the sample of funds. This sample consists of 87 mutual funds. The economic di¤erences between the alphas are modest in magnitude. our results indicate that the relative performance for international mutual funds is positive rather than negative. The sample period ranges from June 1982 until May 2002. For reasons of brevity.[Insert Tables 5 and 6 here] The Jensen’s alphas of the funds are in Table 6. For 66 out of 87 cases. Nevertheless. we do not report results for each of the funds separately. indicating that the median alpha of the Kalman smoother is signi…cantly larger than that of the rolling window estimator. We select the mutual funds with an asset allocation objective and inception date before 1995. our method is an 24 .18 5. Averaging the rolling window alpha estimates results in a negative alpha of 31 basis points per annum. A (pairwise) Wilcoxson signed rank test rejects the null hypothesis with a p-value below 0:001. such as the market. the Jensen’s alpha measure is higher when analyzed by the structural model and estimated by the Kalman smoother. For eleven out of twelve funds in our sample. the reduction in MAD is more than 5 percent for half of the sample. as is indicated by the rolling window estimators. as in the previous section.2 Three-factor exposures for asset allocators The time-variation in exposures for the funds in the previous section is not obvious.

4 percent worse than the rolling window. with a p-value below 0:001.7 percent for the MAD. generic. the assumption that styles are constant within each of the estimation windows is inconsistent by itself. Improvements for the latter measure are as large as 42 percent. A (pairwise) Wilcoxson signed rank test rejects the null hypothesis of equal medians for both methods. In addition. In this paper. 6 Conclusions Return-based style analysis is a useful. the Kalman …lter approach is used to explicitly model time variation of exposures for return-based style analysis. There is ample evidence that the investment style of managed funds is time-varying. the results from style analysis are often used for performance measurement. We introduce an alternative statistical model and estimation technique which alleviates these problems while explicitly incorporating style changes. while in the worst case. while 24 funds account for an improvement larger than 7.improvement over the rolling window predictor. In contrast with rolling window regressions. while it is 9. Even in the worst case. the Kalman …lter method is just 7. time-varying exposures are incorporated implicitly by using rolling window estimators. unless the style is constant over the entire sample period. which complicates the style estimation in a standard regression framework. Often. the entire sample period can be used to obtain e¢cient estimates of the exposures at each point 25 .6 percent for the MSD. only an 18 percent loss is observed. Moreover. These results indicate that for prediction purposes. The average reduction is 4. and does not depend on arbitrarily chosen window sizes. the use of the random walk model can improve the style or risk exposures of mutual funds. The length of this window is generally chosen ad hoc and not motivated by statistical theory. and quickly applicable tool for investors to get a …rst impression about the investment philosophy of a mutual fund or the family it belongs to.4 percent when using the Kalman …lter.

whereas the rolling window analysis suggests that the mutual funds from our sample exhibit a negative performance relative to their style benchmark. reducing the in‡uence of spurious correlation between style indices and mutual fund returns in small samples. Due to the unobservability of the actual style exposures. which have the change of exposures as a stated objective in their prospectus. Since the exposures to style factors are unobservable. We …nd that our method improves style predictions for 66 out of 87 funds. In addition. Second. The empirical results suggests that. Our results indicate that for 11 out of 12 funds the return predictions are better by using our method. in addition to the statistical justi…cation. we also apply our method to a sample of asset allocators. the regional exposure of a sample of US-based international mutual funds are analyzed.in time by the Kalman smoother. albeit a modest improvement only.4 percent is obtained for 24 of the mutual funds. We also report di¤erences in inference when timing ability is measured using rolling window estimators and the Kalman smoother. The statistical model we advocate in this paper is applied on two sets of mutual funds. From these stylized examples it follows that the Kalman …lter and smoother are much closer to the true underlying investment style than the rolling window alternative. our model 26 . the Kalman smoother approach …nds weak evidence for positive selectivity performance. we analyze their use in predicting the mutual fund returns. Even the worst case increases the mean absolute deviation by only 7. We present three stylized examples in order to demonstrate the huge di¤erences that may be obtained by rolling window regressions instead of the Kalman …lter approach. these style estimates change smoothly over time. Our stylized examples suggest that the improvements in the style exposures might be substantially higher. A reduction in MAD larger than 7.4 percent. First. Finally. we cannot infer the increased level of precision of the style predictions.

is also of practical use. the magnitude of the improvements obtained are modest and can be seen as a re…nement on current practices. 27 . However. it is clear that this is not a cure-all.

Shephard & Doornik (1999). 8 We use the Ox-based computer package Ssfpack to evaluate the system. See for example Lhabitant (2000) and Crowley & Stutzer (2001) for a discussion how mutual fund managers may game their Sharpe ratio or Morningstar rating by including simple option strategies in their portfolio.Notes 1 Other style dimensions include the fee structure. the Kalman …lter has next to a classical also a Bayesian interpretation. See www. Eun et al.ssfpack. 3 Alternatively.g.. 5 Alternatively.. the discussion in Belden & Waring (2001) on the disadvantages of the use of historical returns only in a performance measurement context. Agarwal & Naik (2000) label RBSA without restrictions generalized RBSA. e. (1991). 9 In fact. 2 An application in which an investor chooses an optimal portfolio of mutual funds using RBSA can be found in DeRoon et al. Droms & Walker (1994). 4 Notable exceptions are Cumby & Glen (1990). see. It is known that identifying non-linear asset structures by applying RBSA is complicated. 7 As this terminology suggests. Ssfpack is described by Koopman.com for more details. and Kao.g. the investment process. Cheng & Chan (1998).oxmetrics.com or www. e. For a detailed description of Ox see Doornik (2001). Agarwal & Naik (2000). the self-declared investment objective. the K random walks for each ¯ i are replaced by a series of K ¡ 1 random 28 . the use of derivative products. (2004). 6 See. non-linear assets could be used to game their investment style or perfor- mance measures. to say the least.

1). while the June 1982 – May 2002 are from the Morningstar Principia Pro Plus for Mutual Funds June 2002. but the results are very similar to the 36 months we use throughout this paper. and (0.j ° j. our analysis focuses on exposures rather than holdings. 29 . While holdings information might also provide insight in the fund’s style. (1. from which the ¯ i -s are obtained as ¯ i.g. Shleifer.. The Transatlantic Fund is not in the Morningstar database with unknown reason. Lakonishok. Thaler & Vishny (1991)) and the use of futures might blur risk measurements based on this information. see Table 1. However. 10 Like walks for the newly de…ned parameters ° j . In Figures 4a-d the holdings are presented next to the exposure estimates. The Kemper International Fund merged into the Scudder International Fund. 11 MSCI 12 We Paci…c is here including Japan have experimented with rolling windows between 24 and 60 months. 16 The Morningstar data some data on the regional holdings of the mutual funds. known problems such as window dressing (see.t+1 = PK¡1 1 j=1 ! i. with ! j representing the weights. 15 The total return data from January 1976 – May 1982 are from the Morningstar Principia March 1995 CD-ROM.t . e.0). For instance. 13 Unreported Chow and CUSUMSQ tests indicate that the hypothesis of constant expo- sures over the entire sample period is rejected for almost all funds in these applications. 14 Several mutual funds from these papers have changed names. the weights K + Lobosco & DiBartolomeo (1997) we argue that indices that are highly correlated for a longer period of time are essentially capturing the same style and should be treated as one and the same style. The Invesco GT Paci…c Fund and Merrill Global Equity Fund are not in our sample because they are not listed in the US anymore.-1).are (-1. when K = 3.

because we have ignored the potential survivorship bias. since most of the funds used in studies in the early 90-s still exist. 18 This statement holds for our sample of funds. 1999. In any case. 30 . we expect the in‡uence of survivorship bias to be low. and not for international mutual funds in general. "Vanguard’s Richard Foulkes: How to win – cautiously – in Asia" BusinessWeek March 19.17 see Robert Barker.

. (2002). N. & Van Harlow. C. C. Brown. Mutual fund styles. Benson. (2002). Journal of Asset Management 1(1): 93–109. E. Brown. Journal of Investing 10(4): 66–72.-L.. S. V. D. H. (2000). P. Journal of Financial Economics 43(3): 373–399. Chen. Generalized style analysis of hedge funds. P. Fraser. J. M. & Eger. Journal of Finance 31 . & Goetzmann. & Waring. S. P. W. Staying the course: The impact of investment style consistency on mutual fund performance. C. Journal of Banking and Finance 16(5): 1015–1033. (1997). On mutual fund investment styles. M. B. (2001). (1997). G. Alexander.References Agarwal. Journal of Financial and Quantitative Analysis 17(4): 579– 601. & Naik. Timing decisions and the behavior of mutual fund systematic risk. M. UK unit trust performance 1980-1989: A passive time-varying approach. Improve your Morningstar-rating using options. J. J.. Black. On persistence in mutual fund performance. Crowley. (1992). & Stutzer. University of Texas Working Paper SSRN(306999). L. Compared to what? A debate about picking benchmarks. Y. (2001). & Power. 52(1): 57–82. (1982). Belden. & Lakonishok. W. Journal of Investing 10(4): 73–87. A. Carhart. Chan. Review of Financial Studies 15(5): 1407–1437. G. N. K.

Eun. & Koopman. S. & French.S. (1997). London. E. G. R. Journal of Portfolio Management 17(3): 88–94.. Journal of Empirical Finance 11: 29–53. D. E. E. (1994). R. W. B. F. J. DeRoon. (1985). D. A. J. (2004). Journal of Financial Research 17(1): 1–14. Droms. (1991). R. Doornik. Ox 3.. & Glen. (2001). J. T. Journal of Finance 45(2): 497–521. (2001).Cumby. R. H. Fisher. Oxford. Forecasting systematic risk: Estimates of "raw" beta that take into account the tendency of beta to change and the heteroskedasticity of residual returns. Fama. Timberlake Consultants Press. & Resnick. Common risk factors in returns on stock and bonds. J. Kolodny. Investment performance of international mutual funds. U. DiBartolomeo. G. Evaluating style analysis. Journal of Finance 51(2): 425–461. (1996). Measuring fund strategy and performance in changing economic conditions. & Witkowski. J. K. C. & Kamin. E. Durbin. L. S. E. Time Series Analysis by State Space Methods. D. W.-based international mutual funds: A performance evaluation.0: An Object-Oriented Matrix Programming Language. Evaluating the performance of international mutual funds. Oxford University Press. & Walker. (1993). J. (1990). Nijman. A. Mutual fund misclassi…cation: Evidence based on style analysis. Journal of Financial Economics 33(1): 3–56. & TerHorst. A. R. W. Ferson. Financial Analysts Journal 53: 32–43. Journal of Financial and Quantitative Analysis 20(2): 127–149. & Schadt. 32 . F.

(1998). Kim. (1991). L. (2000). M. University of California Working Paper . American Economic Review 81(2): 222–226. (2001). (1993). Huberman. Financial Review 33(2): 127–144. Shephard. S. C. Stone. Time Series Models. H. Journal of Finance 51(3): 783–810. Journal of Financial Economics 36(1): 3–27. Investor diversi…cation and international equity markets. Cheng. & White. M.2. & Lockwood. W. Heston. London. G.. K. Fund management changes and equity style shifts. Another puzzle: The growth in actively managed mutual funds. & Rouwenhorst. (1996). International mutual fund slectivity and market timing during up and down market conditions. K. Shukla. Statistical algorithms for models in state space using Ssfpack 2. A. D. R. A. Harvey. & Tomas. (1994). S. Review of Financial Studies 14(3): 659– 680. T. G.. (2000). Gruber.. Familiarity breeds investment. J.-H. Kao. M.French. Journal of Economics and Business 52(4): 309–323. T. K. Econometrics Journal 2(1): 1–55. G. W. Does industrial structure explain the bene…ts of international diversi…cation?. & Chan. Asymptotic and Bayesian con…dence intervals for Sharpe style weights. & Doornik. L. Koopman. C. Gallo. J. M. 2nd edn. G. Kim. N. (1999). Mutual fund objective misclassi…cation. & Poterba. R. J. L.. Financial Analysts Journal 55(5): 44–52. 33 . J. J. (1999). J. Prentice Hall/Harvester Wheatsheaf.

Journal of Economic Theory 13(3): 343–362. (1998). Asset allocation: Management style and performance measurement. & Kadiyala. E. R. (2001). 34 . A.-S. Lobosco. Journal of Banking and Finance 12(3): 457–467. (1988). Derivatives Quarterly 7(2): 39–46. Approximating the con…dence intervals for Sharpe style weights. D. Thaler. Window dressing by pension fund managers. American Economic Review 81(2): 227–231. Journal of Portfolio Management 18(2): 7–19. Sirri. Journal of International Financial Markets.Kuo. J. Institutions and Money 11: 1–28. F. Derivatives in portfolio management: Why beating the benchmark is easy?.. K. Financial Analysts Journal 53(4): 80–85. Lockwood. (1976). S. Sharpe. The arbitrage theory of capital asset pricing.. F. & Tufano. A. Global equity styles and industry e¤ects: The pre-eminence of value relative to size. & Vishny. Lakonishok. Costly search and mutual fund ‡ows. R. W. P. E. (1991). A. W. J. Lhabitant. R. S. & Satchell. Ross. (2000). (1997). Journal of Finance 53(5): 1589–1622. L. Shleifer. & DiBartolomeo. (1992). Measuring investment performance with a stochastic parameter regression model. R.

255 MSD 0.162 0.595 MSD 6.318 0.302 0.521 0.592 0.236 5.033 0.060 0.635 35 .828 0.883 1.311 0. Reported numbers are averages over 100 simulations.152 0. Example 1 MAD Rolling window Kalman …lter Kalman smoother 1.004 Panel B: Comparison of predicted mutual fund returns.193 0.018 Example 2 MAD 0.291 1.135 0. and are not used to compute the distance measures.654 MSD 3.069 0.613 Example 3 MAD 1.131 0.Table 1: Comparison of style prediction methods.057 0.205 0. Example 1 MAD Rolling window Kalman …lter Kalman smoother 0.759 2.090 Example 3 MAD 0.622 MSD 3.380 1. The …rst 36 months are used as initialization period. Panel A: Comparison of exposure estimates for the US.060 MSD 0.028 0. This table contains the mean absolute deviation (MAD) en mean squared deviation (MSD) of the US exposure for the three examples.042 MSD 0.338 1.701 Example 2 MAD 1.668 0.

66 36 .10 3. The last two columns contain the average realized total returns. Name Alliance International American Funds New Perspective Evergreen International Growth First Eagle SoGen Global Oppenheimer Global Putnam Global Growth Scudder International Templeton Growth Templeton World T. which denotes the tax or fee structure of the fund. The second column displays the share class.01 4.24 1. the First Eagle SoGen Global Fund was the SoGen International. Foreign Stock World Stock Foreign Stock Multi-Asset Global World Stock World Stock Foreign Stock World Stock World Stock Foreign Stock Foreign Stock Foreign Stock Average 0.98 4.20 1. and the …fth the prospectus objective.Table 2: Descriptive statistics mutual funds.May 2002 for ease of comparison.37 2. Although we use return data starting in January 1976 in the empirical analysis.06 1.35 1.07 1.12 4.98 4. The American Funds New Perspective Fund was previously named the New Perspective Fund. the fourth the ticker code. The …rst column contains the name of the fund in May 2002.62 4.20 1. June 1982 – May 2002.75 5.75 4. and the standard deviation of the total return. In the third column we …nd the the fund’s inception date.15 1.08 Stdev 5. and standard deviation are calculated over the period June 1982 . the arithmetic average monthly return.91 1.87 1. the Evergreen International Growth Fund was the Keystone International Fund. Waddell & Reed Adv International was the United International Growth.11 1. Rowe Price International Stock Vanguard International Growth Waddell & Reed Adv International Class A A B A A A S A A A A A Inception 1981-06 1973-03 1954-10 1970-04 1969-12 1967-09 1953-06 1954-11 1978-01 1980-05 1981-09 1970-06 Ticker ALIFX ANWPX EKZBX SGENX OPPAX PEQUX SCINX TEPLX TEMWX PRITX VWIGX UNCGX Prospectus object.18 4.26 0.69 4.

The average US dollar returns.92 20.24 0. The Paci…c index is excluding Japan.00 37 .42 4. June 1982 – May 2002.55 -19.94 St.52 7.50 1.53 0. Average USA Europe Paci…c Japan 1.63 0.32 1.00 0.31 1.28 11.00 0.28 1.96 0.58 0. The average and standard deviation are in percentages per month. The regional indices used are those of Morgan Stanley Capital International (MSCI).00 0.38 Max 13.76 24.dev.98 -43.Table 3: Descriptive statistics regional style factors.67 6.22 -18.26 Correlation 1. standard deviations. and correlations of the style factors are presented.12 Min -21. 4.

Rowe Price International Stock Vanguard International Growth Waddell & Reed Adv International Total MAD 1.06 1.61 1.07 4.34 8.26 1.5 -6.32 1.7 -7.75 2.8 -4.72 Percentage MAD -7.6 -4.0 38 .43 4.38 2.5 -3.97 3.8 -3.22 1.2 -3.40 2.55 1.02 1.7 -0.21 1.50 2.3 -8.34 2.8 -4. The last two columns show the perentage gain obtained when using the Kalman …lter relative to the 36 month rolling window estimator.1 -6.3 -2.73 1.80 5.34 1.84 2.15 8.1 1.7 MSD -22.07 Kalman …lter MAD 1.19 1.0 -20.72 2. The last row contains averages over our sample.48 1.4 -0.37 4. The mean absolute deviation (MAD) and mean squared deviation (MSD) between predictions generated by the 36-month rolling window estimator and the Kalman …lter are displayed for our sample of international (world or foreign) mutual funds.85 2.27 1. Since the actual style exposures are unknown.60 1.06 1.4 -8.44 MSD 6.97 6.0 0.5 -8.35 2.3 -13.07 1.26 4.24 2.81 1. Rolling window Name Alliance International American Funds New Perspective Evergreen International Growth First Eagle SoGen Global Oppenheimer Global Putnam Global Growth Scudder International Templeton Growth Templeton World T. The …rst 36 months are used as initialization period and not used to measure the di¤erences.15 2.38 MSD 4.7 -5.11 2.0 -9.3 -0.31 1.28 1.19 2.37 3.1 -6.67 3.Table 4: Comparison of style forecasting performance for international mutual funds.16 4.47 1.75 1.13 1. the MAD and MSD measure the di¤erence between the predicted style times the realized style return and the realized return of the mutual fund.5 -8.11 1.

008 -0.023 -0.014 -0. Rowe Price Vanguard Waddell & Reed USA ¤ 0. Smoother PAC 0.077 ¤¤ 0.026 -0.061 0.015 -0.019 -0.078 -0.036 0.115 EUR ¤¤ -0.013 0.024 0.124 -0.023 -0.076 ¤¤ -0.028 0.000 0.031 -0.017 0.029 0.074 JAP -0.004 0.057 -0.028 39 .018 0.040 0.030 -0.012 0.044 ¤¤ -0.004 -0.000 0.061 -0.092 0.015 -0.043 -0.097 EUR ¤¤ -0.050 0.021 -0.019 -0.012 -0.110 0.007 0.061 -0.120 USA 0.177 ¤¤ 0.159 ¤ 0. The signi…cance of the estimated correlations is indicated by ¤ and ¤¤ for the 90 and 95 percent signi…cance level.034 0.013 0.019 0.027 -0.Table 5: Comparison of correlations between changes in style and style returns.034 0.017 -0.034 -0.039 0.020 0.088 -0.018 -0.000 -0.005 0.020 -0.041 ¤ -0.025 0.056 -0.136 PAC 0.140 ¤¤ -0.010 0.036 0.000 0.029 -0.013 0.004 -0.011 0.174 0.055 -0.071 0.136 Rolling window Name Alliance American Funds Evergreen First Eagle SoGen Oppenheimer Putnam Scudder Templeton Growth Templeton World T.006 -0.030 0.007 -0.091 0.000 0. respectively. The correlation between the estimates from the rolling window (and Kalman …lter) and the realized return in the same period are used to obtain insight in the additional value provided by the mutual fund by changing the regional exposure.070 0.036 -0.017 ¤¤ 0.014 0.030 -0.017 0.042 0.035 0.045 0.005 0.042 -0.033 JAP 0.032 0.043 0.

31 t-val -3.59 3.04 2.49 -0.60 -0.Table 6: Comparison of selectivity coe¢cent (alpha) for international mutual funds.07 -0.09 -0.22 -1.57 -1.84 1.33 1.99 3. The last row contains the average over our sample of funds.42 0. The standard errors are computed using the Newey-West correction for autocorrelation. The average value alpha from the rolling window estimator are used to measure the selectivity or micro-forecasting skills of the mutual fund manager.56 1.15 -2.37 -0.34 -0.29 0.62 0.48 -2.18 -0.29 -0.49 0.06 -0.90 -0.60 2.48 3.48 -0.64 t-val -0.18 1.48 -0.35 -1.47 40 .36 -1.93 0.75 1.85 0. The smoothed alphas with t-values are displayed in the last two columns.68 1.38 -3.67 1.09 1.83 Kalman smoother Alpha -1.32 2.79 -0.22 -1.86 3.89 2.65 -2.68 2. Rolling window Name Alliance American Funds Evergreen First Eagle SoGen Oppenheimer Putnam Scudder Templeton Growth Templeton World T.62 -1. Rowe Price Vanguard Waddell & Reed Total Alpha -3.

1. The true exposure alternates each decade between 100 percent US and 100 percent Europe. (true exposure) 1.4 0.Figure 1: Exposure to the U. the predicted exposures by using the 36-month rolling window (circles) and the Kalman …lter (squares) are displayed.4 0.2 0.0 1989 1990 1991 1992 41 .0 1990 1991 1992 1993 1.8 0.change exposure each decade U. The estimation is performed over the full sample (January 1976 – May 2002).8 0.S. In Panel A.0 0. together with the true exposure.S.S.2 Example 1: Prediction . Panel B contains the Kalman …lter (squares) and Kalman smoother (triangles). (Kalman filter) U.S.0 0. for Example 1.S. (36m rolling window) U.6 0.S.S.change exposure each decade U. The true exposure is depicted for ease of reference.6 0. (true exposure) U.2 0. (Kalman smoother) 1. (Kalman filter) U.2 Example 1: Explaining . but in this …gure subsamples are shown for reasons of clarity.

In Panel A.2 Example 2: Prediction . The estimation is performed over the full sample (January 1976 – May 2002). (true exposure) 1.6 0.6 0.2 Example 2: Explaining . together with the true exposure.0 0.change exposure each year U.S.S.S. (Kalman smoother) U. 1.change exposure each year U. but in this …gure subsamples are shown for reasons of clarity. The true exposure is depicted for ease of reference.S.2 0.S.4 0. the predicted exposures by using the 36-month rolling window (circles) and the Kalman …lter (squares) are displayed.4 0. (36m rolling window) U. The true exposure alternates each year between 100 percent US and 100 percent Europe.8 0.0 1990 1991 1992 1993 1.0 0.S. (Kalman filter) U.8 0. Panel B contains the Kalman …lter (squares) and Kalman smoother (triangles). for Example 2.Figure 2: Exposure to the U.2 0. (true exposure) 1.0 1990 1991 1992 1993 42 .S. (Kalman filter) U.

slowly changing exposure U. (true exposure) 1.S. together with the true exposure.8 0. (36m rolling window) U.S. In Panel A. 1.6 0.S.2 0.2 Example 3: Explaining .4 0. The estimation is performed over the full sample (January 1976 – May 2002).0 1990 1991 1992 1993 43 .0 0. (Kalman filter) U. for Example 3.0 0. The true exposure is depicted for ease of reference. It takes three years for a fully invested position in one region is exchanged for the other. (Kalman filter) U.S.S. (true exposure) U.8 0.S. the predicted exposures by using the 36-month rolling window (circles) and the Kalman …lter (squares) are displayed.2 0.4 0.0 1990 1991 1992 1993 1. but in this …gure subsamples are shown for reasons of clarity.slowly changing exposure U. The true exposure slowly alternates 100 percent US and 100 percent Europe.Figure 3: Exposure to the U.2 Example 3: Prediction . (Kalman smoother) 1.6 0.S. Panel B contains the Kalman …lter (squares) and Kalman smoother (triangles).

MSCI US. The four panels of this table display the exposures to the MSCI Europe. the Kalman …lter.55 0.Exposure to MSCI Japan Kalman filter Rolling window (36 months) Holdings (annual report) Kalman smoother 0.45 Rolling window (36 months) Kalman filter Kalman smoother Holdings (annual report) 1995 1996 1997 1998 1999 2000 2001 2002 0.00 1995 1996 1997 1998 1999 2000 2001 2002 44 .80 0. Vanguard International Growth .75 0.65 0.Exposure to MSCI Europe 0. The diamonds indicated the holdings fromt the annual reports. but in this …gure subsamples are shown for reasons of clarity.70 0. the Kalman smoother are presented. MSCI Japan.20 0. and MSCI Paci…c excluding Japan.15 0.05 0.60 0. The estimation is performed over the full sample (January 1976 – May 2002).50 0. In panels A–D the exposure estimates from the traditional 36month rolling window.Figure 4: Exposures for the Vanguard International Growth Fund.25 0.35 Vanguard International Growth .10 0.30 0.

05 0.05 -0.20 1995 1996 1997 1998 1999 2000 2001 2002 Vanguard International Growth .15 0.00 -0.20 Vanguard International Growth .00 1995 1996 1997 1998 1999 2000 2001 2002 45 .15 -0.Exposure to MSCI Pacific ex Japan 0.10 -0.Exposure to MSCI USA Kalman smoother Rolling window (36 months) Kalman filter Holdings (from annual report) 0.20 0.25 0.10 0.10 0.05 Kalman filter Rolling window (36 months) Kalman smoother Holdings (annual report) 0.15 0.Figure 4 (continued): 0.

05 0. and goes down to a 30 percent gain in prediction accuracy for the funds on the left on the horizontal axis.20 -0. sorted on the percentage gain of the mean absolute deviation (MAD) between the predicted fund return and the realized fund return.10 -0. Percentage decrease in MAD of predicted returns for asset allocators 0.05 -0. The horizontal grid starts at a 5 percent loss in prediction accuracy at the top of the …gure. The predicted fund return is the forecasted style exposure mulitplied by the realized factor return.30 0 10 20 30 40 50 60 70 80 90 46 .25 -0.00 -0.Figure 5: Improvements from estimating three-factor model for asset allocators. This table contains the 87 asset allocation funds.15 -0.

- Decision Making
- Vortrag5
- Master Time Factor trading
- Law Reading
- michellehanlon
- chance and probability program
- darwin's finches lab
- naman
- fl200_e
- Jbptitbpp Gdl Ekapermana 34220 8 2009ts A
- science stage 3 scope
- The Ultimate Book of Investment Quotations
- ANNEXURE
- CMMI High Maturity Best Practices HMBP 2010
- New Microsoft Word Document
- WMTC17 Presentation Ben Carlson
- 14.pdf
- Tomorrow's Investor - Towards Accountable and Responsible Financial Services
- Mean
- Breaking Out of the Black-Box
- [IJETA-V4I6P9]:Jigar V. Pandya, Dr.Vikram M. Patel
- FM300 Syllabus Section a 2012-13
- Download BioMed Solution
- Preqin Report Infrastructure Funds
- Descent of Fetal Head (Station) During 1st Stage of Labor
- SSRN-id272180
- ch21.srm
- xianjunliang kong resume ppdu
- 02.2014PortfolioSpotlight--DiversifyingStrategies
- Joint Report by Jp Morgan and Oliver Wyman Unlocking Economic Advantage With Blockchain a Guide for Asset Managers

Sign up to vote on this title

UsefulNot usefulClose Dialog## Are you sure?

This action might not be possible to undo. Are you sure you want to continue?

Loading