JOURNAL OF

nancial
ELSEVIER Journal of Financial Economics 43 (1997) 373 399
ECONOMICS

Mutual fund styles

Stephen J. Brown", William N. G o e t z m a n n *'b
~Leonard Stern School ol Business, New York Unicersitv. New York, NY 10012, USA
byale School of Management, Yale UniversiO,, New Hacen, CT 06520, USA

(Received September 1995; final version received June 1996)

Abstract

Mutual funds are typically grouped by their investment objectives or the 'style' of their
managers. We propose a new empirical to the determination of manager "style.' This
approach is simple to apply, yet it captures nonlinear patterns of returns that result from
virtually all active portfolio management styles. Our classifications are superior to
c o m m o n industry classifications in predicting cross-sectional future performance, as well
as past performance, and they also outperform classifications based on risk measures and
analogue portfolios. Interestingly, 'growth' funds typically break down into several
categories that differ in composition and strategy.

Key words: Mutual funds; Management style; Style analysis

J E L class!fication: G20: G23; GI 1

1. Introduction

Investment objectives and style classifications are widely used in the financial
industry to characterize differences between money managers. Mutual funds, for

*Corresponding author.
The authors thank Michael Barclay. Ken French, Mark Grinblan, Toshiyuku Otsuki, and Matthew
Spiegel for helpful comments. We also thank participants in the 1995 WFA session on investment
styles, the 1995 Conference on Finance and Accounting at the University of New South Wales, the
1995 Conference on Finance and Accounting, workshops at Berkeley, The City University of Hong
Kong, University of California at lrvine, International University of Japan, University of
Melbourne. Rutgers, Stanford. Princeton, Virginia Polytechnic, Washington University in St. Louis,
the University of Washington, and the University of Wisconsin at Milwaukee, for their suggestions
and comments. We also thank Morningstar, Inc. and lbbotson Associates for providing data
for analysis. All errors are the sole responsibility of the authors.

0304-405X.97/$17.00 ~" 1997 Elsevier Science S.A. All rights reserved

PIt S 0 3 0 4 - 4 0 5 X ( 9 6 ) 0 0 8 9 8 - 7
374 S.J. Brown. I~iN. Goetzmann/Journal qlFinancial Economics 43 (1997) 373 399

instance, are typically grouped according to the type of securities in which they
invest and the "style" of their managers. Equity funds range from 'aggressive
growth' funds holding low-dividend, high-growth stocks to 'income'
funds seeking high-yield equities. Such fund classifications are ubiquitous,
but do they actually tell us anything about the strategies of investment man-
agers'? Do they help explain differences in future returns among funds or even
provide useful benchmarks for evaluating relative past performance? These
fundamental questions about mutual fund classifications are the motivation for
this research.
The definition of standard equity mutual fund categories is generally broad
enough to allow a wide range of different investment policies. The Investment
Company Institute uses a very general description of the largest investment
category:
Growth Funds invest in the common stock of well established com-
panies. Their primary aim is to produce an increase in the value of their
investments (capital gains) rather than a flow of dividends. (Investment
Company Institute, 1991, p. 12)
This definition makes it obvious that the typical growth fund manager has great
latitude in the types of stocks to hold, the timing of purchases and sales, the level
of fund diversification, the industry concentration of the portfolio, and a host of
other factors that go into determining the returns to client investments. Given
this broad latitude, it is not surprising to find widely divergent behavior among
funds pursuing the same objective. As a result, existing classifications do a poor
job of forecasting differences in future performance.
Moreover, the financial press has identified several cases of funds apparently
misclassifying themselves (e.g., Donnelley, 1992). The S.E.C. has a stated
mandate to insure that the composition of a fund does not contradict its
objective, if the objective is included as part of its name. Such governmental
concerns are not unfounded. Recent papers by Witkowski (1994) and Kim,
Shukla, and Tomas (1995) find that the movement of many mutual funds is
better explained by the performance of a style index other than their own. In this
paper, we find some evidence suggesting that such misclassification may be
intentional, in that it works to improve ex post relative performance measures,
on average.
Because management styles are so widely used as the basis for performance
measurement and compensation, there is a great need for style classifications
that are objectively and empirically determined, consistent across managers, and
related to the manager's strategy. Objectivity is important because of the moral
hazard inherent in allowing managers to self-report their styles without objec-
tive verification. Consistency is needed for purposes of performance comparison.
The desirability of such a classification scheme is clear to all participants in the
industry, and industry alternatives to the existing classification procedures have
 S.J. Brown, W.N. Goetzmann/Journal of Financial Economics 43 (1997) 373 399 375

already begun to evolve (see, e.g., Tierney and Winston, 1991; Christopherson,
1995). Beyond the practical need for meaningful styles, there is a fundamental
question about whether any classification system (which, after all, is only
a multinomial statistic) is sufficient to characterize differences in fund manage-
ment.
To examine all of these issues, we develop a 'style classification' algorithm that
is consistent with asset pricing models. The consistency is useful, because
the multinomial statistic represents a 'coarsening' of a fully specified
stochastic model of portfolio returns and it is useful to clarify where and how
this coarsening takes place. Our algorithm groups funds based on the cross-
sectional time series of past returns as well as on the response to exogenously
specified and endogenously determined stochastic variables. Using mutual
fund data from 1976 through 1994, we find that equity funds broadly
fall into some familiar and not-so-familiar patterns of behavior. The familiar
patterns include 'small-cap', 'growth', 'growth and income', 'income', and 'inter-
national' funds. However, as many as half of all currently classified
"growth' funds fall into different categories, according to our procedure.
We also identify some unfamiliar categories that are n o t captured by the
traditional objectives, including 'value' managers, 'trend-chasers', and "glamour"
managers.
Our derived classifications specified ex ante do a better job of predicting
cross-sectional variation in fund returns than do traditional mutual fund classi-
fications. In addition, we find that simple classifications capture major differ-
ences in manager behavior as manifested in the temporal pattern of returns.
While these classifications provide less information about the magnitude of fund
loadings on major macroeconomic factors, they provide a useful means to
identify widespread, common patterns in manager behavior.
The implications of our results are broad. A return-based classification system
such as ours can reduce the incentive to 'game' the styles to improve relative ex
post rankings. More formal classification procedures for mutual funds can help
investors better understand the future behavior of their investments, and can
provide ex post or ex ante performance benchmarks. From the perspective of
researchers interested in understanding investment manager behavior, catego-
ries that incorporate 'value' and 'glamour' funds may better characterize how
managers behave. A by-product of the estimation procedure is the creation of
a parsimonious set of robust factors composed of positive weights on existing
mutual funds, These style factors typically outperform prespecified macroeco-
nomic factors in out-of-sample tests on fund returns, and thus may have further
implications for asset pricing.
The paper proceeds as follows. Section 2 provides background on some
statistical and strategic issues in the paper. Section 3 describes the data and
methodology. Section 4 reports the results of the empirical analysis and Sec-
tion 5 concludes.
376 S.J. Brown, If'iN. Goetzmann/Journa/ o/ Financial Economics 43 (1997) 373 399

2. Statistical and strategic issues

2.1. Statistical issues: Time-vao,ing port[blio weights'

There is a long tradition of characterizing mutual funds according to parameters

estimated via a linear model of returns. The technology of asset pricing was first
applied by Jensen (1968) to grouping mutual funds according to their systematic
risk characteristics. Connor and Korajczyk (1986), Lehmann and Modest (1987),
and Grinblatt and Titman (1988, 1989) all apply linear asset pricing methods to
differentiate mutual funds on the basis of systematic risk characteristics. Quanti-
tative methods of security aggregation first appeared in the finance literature in
the work of Elton and Gruber (1970), who developed a classification algorithm
using a linear model of fundamental characteristics that proved useful for
grouping securities and forecasting cross-sectional differences. Carleton and
McGee (1970) suggest that the related techniques of switching regressions and
hierarchical clustering methods can be used to aggregate financial assets.
There are reasons to expect that linear models might poorly characterize
mutual fund returns. Single-factor and multiple-factor linear models are only
exactly correct when portfolio weights remain fixed through time and when the
systematic risk characteristics of the securities held in the portfolio remain fixed as
well. While it is common to assume that securities change little, we have no basis
for presuming that portfolio weights remain fixed, Indeed, active fund manage-
ment clearly implies the strategic reallocation of portfolio weights across assets.
While typical examples of such active allocation are market timing and portfolio
insurance strategies, a buy-and-hold strategy as well as active security selection
can also result in time-varying systematic risk characteristics. Recent studies of
mutual fund manager behavior report unambiguous evidence of strategic changes
in mutual fund portfolios. Grinblatt, Titman, and Wermers (1993) identify herding
activity by mutual fund managers. Ferson and Schadt (1996) find that managers
rebalance in anticipation of changing economic conditions. Brown, Harlow, and
Starks (1993) find systematic changes in risk conditional on past performance.
Lakonishok, Shleifer, Thaler, and Vishny (1991) find that 'window dressing'
accounts for portfolio rebalancing by pension fund managers.
Active portfolio management affects performance measurement in nontrivial
ways. Dybvig and Ross (1985), for instance, show how linear risk models fail to
properly rank fund managers when they change their asset weights through
time. Connor and Korajczyk (1991) consider how to risk-adjust for nonlinear
portfolio strategies by mutual fund managers. Grinblatt and Titman (1993)
avoid problems posed by nonlinearities by explicitly considering active strat-
egies as the basis for a benchmark-free approach to performance measurement.
While such nonlinearities present problems for style identification as well, our
procedure accommodates nonlinear strategies by allowing factor loadings to
change on a month-by-month basis. This is crucial in light of the fact that many
 S.J. Brown. W.N. G o e t z m a n n / J o u r n a l (?['Financial E c o n o m i c s 43 (1997) 373 399 377

fund managers actively vary their exposure both to the market and to industry
sectors. To the extent that groups of managers change these exposures together
(i.e., they 'herd' into the market, or in and out of sectors), our procedure will
group them together. Although it is a relatively simple technique, when we
compare the style categories formed in the space of past returns to alternate
categorization schemes formed in the space of fixed factor loadings, we find the
former to be superior in explaining the out-of-sample cross-section of mutual
fund returns. Our method, which relies on a low-dimensional multinomial
statistic with intuitive interpretation as a style, compares favorably to the use of
continuous multivariate measures such as factor loadings. We find some evid-
ence, in the form of time-varying factor loadings, that this is due to the presence
of dynamic management styles in the mutual fund universe.

2.2. Strate qic issues." Se!['-misclassi[ication

Thus far we have been concerned with the ability of existing style classifica-
tions to pick up management behavior that is dynamic and not well captured by
static models of investment. Of great concern is the further issue of selecting
a procedure that prevents ex post changes in style in order to improve relative
historical performance. A self-reported fund objective, announced ex post, could
have been chosen to minimize poor relative performance. Anecdotal evidence
(cited above) from the financial press suggests that such misrepresentation
occurs.
We find some empirical evidence to back up the casual observation that funds
can switch to improve their relative historical rankings. Using equity mutual
fund data over the period 1976 through 1992, described in further detail below,
we find 237 cases in which equity mutual funds switched their fund objective.
For each of these, we subtract the average objective return from the fund
return in the year before the switch, using first the old objective and then
the new objective. That is, the net gain for fund I is defined as
(ri. t -- rj,old) -- (ri. t - - r j . . . . ), whereji,ojd is the style from which the fund switched
in period t + 1 and .Ji.... is the style to which the fund switched. Thus, the
difference between these is the net gain or loss in ex post performance of the
previous year. The average net gain in benchmarked returns was 0.098, or 9.8%,
with a t-statistic of 5.47, assuming that all switches are independent.
While this simple test does not prove that fund managers were switching for
strategic purposes during this period, the results are certainly consistent with
such an interpretation. Were we to use self-reported styles for benchmarking,
without checking to see whether the fund recently reclassified itself, we might be
misled regarding the relative performance of the fund. This is also true if we were
to base the style classification of the fund on its current portfolio holdings.
'Window-dressing' is a common end-of-period ploy of fund managers to throw
out poor performers and/or change the apparent strategy of the fund. Since our
378 S.J. Brown, ~ N . Goetzmann,'Journal e~[Fmancial Economics" 43 (1997) 373 399

procedure uses past returns, not portfolio holdings, it is not fooled by window-
dressing. In most cases, even if we knew that the fund switched, mutual fund
data vendors do not provide a historical record of past fund classifications, so
the true benchmarked history is impossible to reconstruct.

3. Data and methodology

Morningstar, Inc. provided monthly returns of equity mutual funds for the
period January 1976 though June 1995, together with a classification into fifteen
categories: equity income, growth and income, growth, small-company, Europe,
foreign, world, Pacific, financial sector, health sector, natural resources sector,
precious metals sector, high technology sector, utilities sector, and an unaligned
sector. These equity categories include funds that invest in bonds as well as
stocks. The distinction between equity funds and bond funds is generally one of
degree. Even all-equity funds typically hold some cash balances. Like most data
sources of mutual fund returns, the Morningstar data are not free of survivor-
ship bias, and the effect of fund attrition has an unknown effect on ex post
classification. In order to address the problem of changes in fund classification,
we merged the Morningstar data with the annual Weisenberger data used by
Brown and Goetzmann (1995). The Weisenberger data are updated through
1992, the last volume in which Weisenberger provides a comprehensive
Panorama section to their mutual fund annual, based on funds that were willing to
report their performance results over the previous year. While not entirely free of
bias, the data identify changes in fund objectives through time. In addition, we use
a third source of mutual fund data that provides rich material for cross-sectional
analysis: the Morningstar 'On-Disc' database. While only available since 1993, this
CD-ROM program provides information on the composition of each fund as well
as summary statistics about the securities in the fund. We cross-index this informa-
tion with the monthly returns and the Weisenberger data to allow an analysis of
our endogenously determined styles by a broad range of characteristics.

3.1. Stochastic specification

The objective of our analysis is to use past returns to determine a natural

grouping of funds that has some predictive power in explaining the future cross-
sectional dispersion in fund returns. Such groupings are referred to as styles. If
there are K styles, the ex post total return in period t for any fund can
be represented as:
Rj~ = ~ + fl'l, + e~, (1)
where fund j belongs to style J. There are several ways of interpreting this
equation. In a traditional financial economics framework, this equation refers to
a multi-factor or a multi-beta model. The factor loadings on the factors I, are
 XJ. Brown, WN. Goetzmann/Journal of Financial Economics 43 (I99D 373 399 379

given by fls,. These loadings are allowed to change t h r o u g h time. As Sharpe

(1992) points out, if we regard the factors I, as returns on index portfolios, the
factor loadings can be thought of as equivalent portfolio weights associated with
a dynamic portfolio strategy that might be associated with the style in question.
In an interpretation closer to that of financial practitioners, fi~, refers to a char-
acteristic of a typical stock in the J t h style classification (size, m a r k e t - t o - b o o k ,
price-earnings multiple, etc.) and l, is the return to that attribute (see e.g.,
Lakonishok, Shleifer, and Vishny, 1994).
Regardless of how we interpret the equation, the style classifications will explain
the cross-sectional dispersion of fund returns. We can write the equation as

R j, =/*Jr + ~:j,, (2)

where/~a, is the expected return for style J conditional on the factor realization
1,. If the idiosyncratic return component ~-j, has a zero mean ex ante and is
uncorrelated across securities, the classification into styles will suffice to explain the
cross-sectional dispersion of fund returns to the extent that/~j, differs across styles.
The task of assigning funds to style categories can be t h o u g h t of as a problem
in endogenously defining regimes (see, e.g., Q u a n d t , 1959, 1960). In this way, it
bears a 'family' resemblance to a switching regression, although unlike the
switching regression, an exact solution to the style classification problem is only
obtained through exhaustive combinatorics. The a p p r o a c h we use finds a local
o p t i m u m via the minimization of a 'within-group' sum-of-squares criterion over
a specific time period, t = 1. . . . . T. The inputs to the procedure are a T-by-N
matrix of monthly returns for a set of N mutual funds. We g r o u p the N funds
together into K styles by minimizing the within-style mean returns for each
period, t = 1. . . . . T. Thus, we are jointly estimating the time series of mean
returns for the styles J = 1. . . . . T (~j,) for t = 1. . . . . T, and the membership to
each style. The benefit of the resulting classification is that groups could result
from either fixed portfolio strategies, such as similar asset compositions, or from
d y n a m i c portfolio strategies, such as portfolio insurance rebalancing.
The classification procedure assumes that we k n o w the n u m b e r of styles.
Conditional u p o n restrictions on the exact n u m b e r of groups, Eq. (2) is perfectly
well-specified and can be used to estimate the style groupings. Eq. (1) gives
further insight into the nature of time-varying portfolio strategies, although the
parameters in this equation are not identified without further restrictions. ~

A restriction sufficientfor identification purposes is to assume that the portfolio strategy is constant
over a number of months greater than the number of factors. This might seem unduly restrictive.
However, for the purposes of characterizing the time-varying strategies of each style it suffices that
we assume a quarterly holding period with two factors given by the return on cash and on equity
investments. Other quarterly factors are captured in the ~s, terms. Monthly data will then suffice to
estimate Eq. (1). This approach can be contrasted with Sharpe's (1992) use of a rolling regression
technology. The monthly updated portfolio shares should be interpreted as the average style-based
portfolio shares for the previous 24 months.
380 S.J. Brown, W.N. Goelzmann/Journal o/Financial Economics 43 (1997) 373 399

In order to implement our style classification (SC) algorithm, we prespecify a

number of styles.
A modification of the basic algorithm is a generalized least squares procedure
(GSC), which allows time-varying and fund-specific residual return variance. By
scaling observations by the inverse of the estimated standard deviation, we
decrease the influence of extreme observations in the classification process.
Amihud, Christiansen, and Mendelson (1992), for instance, find that this shrink-
age improves forecasts of security returns. The details of the GSC procedure are
provided in the Appendix.
It is important to note that the SC and GSC procedures make minimal
demands on the available data. We can estimate Eq. (2) without needing to
know factor loadings or style attributes represented by the vector//j, which may
well change from period to period. We only need to know ex post returns on
individual funds. There is a direct analogy between our estimation technique
and cluster analysis procedures. The criterion being minimized in Eq. (2) via the
SC algorithm is the same criterion applied in the k-means clustering approach
(e.g., Hartigan, 1975). Cluster analysis usually attempts to minimize the squared
differences within groups of k characteristics. In this context, the characteristics
might include risk exposure and the features of the average stock in the fund
portfolio. In our classification procedure, the k characteristics are month-by-
month returns and the group means are the conditional expectations appearing
in Eqs. (1} and (2). These characteristics are explicitly time-variant and capture
not only risk but also dynamic portfolio strategies that are specific to particular
fund styles.
Because we relax the requirement of constant portfolio weights through time,
we would not expect to identify perfect analogues to the categories derived using
a fixed linear time-series model, or even a very rich set of cross-sectional
characteristics, including stock portfolio composition, observed at one point in
time. Nonetheless, after estimating categories based on our classification algo-
rithm, we report cross-tabulations with Morningstar mutual fund data fields,
and also use the Sharpe (1992) procedure for estimating approximate fixed-
positive-weight portfolio analogues using a standard set of wide-spanning asset
class returns provided by Ibbotson Associates. These two procedures give some
intuition about the resulting mutual fund clusters. Our comparisons yield
evidence of different management strategies, which relate to known classifica-
tions such as 'growth' and 'value' management.

3.2. How man>, sO,les?

Because the procedure relies on prespecifying the number of styles, it is

natural to ask what is the right number. To address this question, we use
a likelihood ratio test suggested by Quandt (1960) for each successive decrease in
the number of prespecified styles from nine. The test statistic for K styles (as
 s.J, Brown. W.N. Goetzmann.Journal Of Financial Economics 43 11997) 373 399 381

opposed to K + 1) styles is

ssql,;+l
LR = Tm lnSSqk-ln
Tm Tm ]

where T is the number of time periods, m the number of funds, and SSqKand SSqK + 1
are the appropriate heteroskedasticity-adjusted sum of squared errors. This
statistic should be approximately distributed as ~2 with 2T degrees of freedom.
Applying this measure to successive levels of fund aggregation, we find evidence
for using at least eight separate categories. There is some ambiguity about the
appropriate degrees of freedom, as well as the appropriateness of the ;~2 distribu-
tion in this case (see Quandt, 1960). Nonetheless, the observed test statistics are
very large. F o r k = 8 through k = 3 styles, the test statistic values are 4,682.9,
4,092.1, 32,217.3, 6555.5, 7,106.2, and 10,197.7, respectively. In each case, the
p-values are arbitrarily close to zero, indicating that an increase in the number of
styles is useful in explaining returns. This result is similar to that reported for
Z2 tests for the number of factors, where typically too m a n y factors are identified
(e.g., Brown, 1989). An important caveat is that the Z2 test is sensitive to
departures from normality. 2 Using fewer than five groups, the distribution of the
group returns suggests that the )~2 test is well-specified. For these low numbers of
groups, the algorithm clearly forces disparate funds together, increasing the model
error. When the number of groups is increased beyond five, it is difficult to judge
the relative magnitude of incremental improvement, although the sign of the test
is positive for all values below nine, suggesting that more groups are needed.

3.3. C o m p a r i n g p r o c e d u r e s

A key question is how the G S C classification performs relative to standard

industry classifications by investment objective. As Trzcinka {1995) points out,

ZThere are significant differences in skewness and kurtosis by style category (for a normal distribu-
tion, skewness is 0 and kurtosis is 3):
Group Skewness Kurtosis
1 0.0172 2.65
2 0.0541) 3.45
3 0.1033 3.95
4 0.1481 3.97
5 0.0670 5.37
6 0.2183 5.32
7 0.1697 5.52
8 0.1015 14.45
Entire sample 0.0782 4.11
The last four groups show significant departures from normality. Thus, the Z2 distribution may be
inappropriate for evaluating the unusualness of the test statistic. In other words, gains to increasing
the number of styes above five groups may be overstated.
382 S.J. Brown, I+iN. Goetzmann/Journal q/'Financial Economics 43 (1997) 373 399

there are no generally accepted standards for comparing style classifications,

which are put to a broad range of uses, from developing benchmarks for risk and
return to establishing specifications used in investment management contracts.
Despite this ambiguity, we borrow a natural measure from the asset pricing
literature. Specifically, we compare our empirically determined styles with the
classifications provided by Weisenberger over the period 1976 through 1992 and
Morningstar over the period 1993 through 1995. The reason we use Weisen-
berger for the early period rather than Morningstar is that mutual fund styles
change through time. The Weisenberger style codes were obtained at the end of
each year in the sample period, and thus they have no ex post bias. Funds are
classified using the GSC algorithm applied to data up to and including the
Weisenberger publication date, with the number of styles chosen to match the
number of industry objectives extant in the last month of the estimation
window. Fund returns are then computed over the following year. Results are
qualitatively similar using a one-month test period and using rolling month-by-
month returns for 24, 36, 48, and 60 months to classify funds. However, the
performance of the industry-based style categories is notably inferior to the
other methods. This is not surprising, since the other methods use data sub-
sequent to the publication date of the industry styles to classify funds. For this
reason, we report results using a one-year test period.
We then cross-sectionally regress fund returns on a matrix of dummy vari-
ables that indicate whether each fund belongs to a particular style. If the style
classification contains information about future differences in returns, we would
expect these regressions to explain a significant amount of cross-sectional
variation. This same procedure is performed for the industry classifications.
A comparison of adjusted R2s indicates which has the superior predictive
ability. This procedure resembles classical time-series, cross-section tests of
pricing models, except that the cross-sectional regressors are not loadings but
a matrix of dummy variables.
As an alternative to the classification based on returns, we report a variety of
other reasonable classification schemes. First, we classify funds based on latent
variable factor loadings derived from a principal components analysis applied to
the time-series matrix of fund returns in the estimation period. We apply the
classification algorithm to the fund loadings on these factors. This is analogous
to the principal component reduction in Elton and Gruber (1970), who use
loadings as the inputs to a classification algorithm. We estimate the loadings
following the procedures described in Connor and Korajczyk (1986) and
Lehmann and Modest (1987).
Second, we prespecify factors and use the prespecified factor loadings to apply
the SC algorithm. This approach has extensive precedent in the empirical
literature. Chen, Roll, and Ross (1986) and Berry, Burmeister, and McElroy
(1988), for example, prespecify macroeconomic risk factors to analyze stock
portfolios and industry characteristics. In an application to mutual funds,
 S.J. Brown, WN. Goetzmann/Journal of Financial Economics 43 (1997) 373 -399 383

Lehmann and Modest (1987), Grinblatt and Titman (1989), Elton, Gruber, Das,
and Hvakla (1993), and Hendricks, Patel, and Zeckhauser (1993) all prespecify
'control' portfolios according to factors such as size and dividend yield. There
are numerous other examples. For the purposes of this analysis, we use eight
indexes: gold, the EAFE minus U.S. global equity index, the EAFA European
equity index, the EAFA Pacific equity index, U.S. Treasury bills, commercial
paper, long-term government bonds, long-term corporate bonds, high-yield
bonds, the S&P 500, small stocks, and IPO's, all obtained from Ibbotson
Associates. This approach has several advantages. First, the profile of each
category has some intuitive interpretation - one group may be tilted towards
bonds, while another is tilted towards stocks, for instance. Second, it suffers less
from the difficulty of heteroskedasticity across funds that introduces systematic
error into the endogenously determined principal components. Third, the coeffi-
cients (when properly scaled) have a natural interpretation as portfolios. The
drawbacks are, of course, that the procedure does not allow for temporal
variation in the portfolio weights. Finally, we cluster in the space of 'Sharpe
coefficients' (see Sharpe, t992) estimated on the same capital market indexes as
above. These are estimated via a constrained optimization procedure under the
assumptions that the weights remain fixed over the estimation period, that they
are nonnegative, and that they sum to one. The weights can thus be interpreted
as portfolio weights for passive, investable indexes.
As a benchmark to the performance of these various classification alterna-
tives, we also report cross-sectional regression results for the factor loadings
themselves. In other words, we use as independent variables in the cross-section
regression four separate sets of explanatory variables. First, we use the coeffi-
cients estimated for each fund obtained by regressing the individual fund return
series on the set of SC styles. Second, we use the first k principal components
(where k corresponds to the number of extant industry objectives). Third, we use
the capital market indexes described above. Fourth, we use the Sharpe coeffi-
cients. These four alternate procedures allow us to quantify how much is lost by
reducing the continuous coefficients down to a simple classification scheme.

4. Empirical results

4.1. Summao; o[ GSC categories

In Table 1, we report the cross-tabulation of the GSC categories with the

Morningstar categories. Since Morningstar categories are identified only at one
point in time, i.e., at the end of the sample period, we would not expect a perfect
correspondence, The key feature of Table 1 is that the 'growth' category, which
is the single largest designation for Morningstar, is spread widely across several
different GSC categories, especially categories 1, 3, 4, and 5. While it is common
384 S.J. Brown. W.N. Goelzmann/Journal of Financial Economics 43 (19971 373 399

Table 1
Cross-tabulation of equity funds by Morningstar and GSC categories~ summary of results using
GSC algorithm, January 1976 to December 1994

GSC group 1 2 3 4 5 6 7 8 Total

Equity income 23 3 74 5 0 0 0 0 105

Europe 0 0 2 1 17 () 17 0 37
Foreign 0 0 2 2 137 1 136 0 278
Growth 196 296 54 117 1 77 0 0 741
Growth income 247 34 89 21 0 0 0 0 391
Pacific 0 0 (1 0 42 0 28 0 70
Small-cap 0 16 7 132 0 115 I 0 271
Financial 1 0 5 7 ~ 0 0 0 15
Health 2 7 0 0 0 7 0 0 16
Metal 0 0 (I 0 1 0 0 35 36
Nat.Resource 0 8 1 9 0 6 I 32
Technology 0 7 0 3 1 19 0 0 30
Unaligned 2 5 4 24 1 0 0 0 36
Utilities I 2 75 1 3 1 (1 0 83
World 3 3 12 4 85 7 27 1 142

Total 482 373 332 318 299 227 215 37 2283

The table reports the cross-tabulation of mutual fund GSC categories with Morningstar style
categories. The Morningstar categories are those attributed to the funds as of 1994 by the company
itself, and thus do not take into account style shifts through the sample period. The unaligned group
includes miscellaneous sector funds, such as REIT funds. The GSC procedure is a maximum
likelihood method described in the text. It allows portfolio weights to vary on a quarterly basis, with
eight factors pre-specified. A likelihood ratio test suggested by Quandt and Ramsey {1978) shows
that the cross-section of mutual fund returns are driven by at least eight separate factors, for which
loadings may vary.

t o a p p r o x i m a t e l y c o n t r o l for risk in m u t u a l f u n d s t u d i e s b y f o c u s i n g o n l y o n
g r o w t h f u n d s (see, for e x a m p l e , H e n d r i c k s , P a t e l , a n d Z e c k h a u s e r , 1993; B r o w n
a n d G o e t z m a n n , 1995; a n d I b b o t s o n a n d G o e t z m a n n , 1994), T a b l e 1 i n d i c a t e s
t h a t m a n y d i f f e r e n t p o r t f o l i o s t r a t e g i e s c a n fall u n d e r t h a t b r o a d r u b r i c . I n d e e d ,
the GSC algorithm groups a significant percentage of growth funds with
growth-and-income funds, suggesting that these labels do not provide parti-
c u l a r l y useful d i s t i n c t i o n s for i n v e s t o r s . A l s o n o t e t h a t t h e s m a l l - c a p c a t e g o r y
splits i n t o t w o d i s t i n c t g r o u p s . A p p a r e n t l y , t h e a v e r a g e c a p i t a l i z a t i o n of t h e
s t o c k s in t h e p o r t f o l i o is n o t a s u t t i c i e n t s t a t i s t i c for performance. F o r t h e s e c t o r
f u n d s , t h e M o r n i n g s t a r c l a s s i f i c a t i o n s a n d G S C c l a s s i f i c a t i o n s g e n e r a l l y agree.
T h e h e a l t h , m e t a l s , utilities, a n d u n a l i g n e d ( p o s s i b l y real e s t a t e ) c a t e g o r i e s a r e
unambiguous. The technology sector and natural resource sector (which
i n c l u d e s f o r e s t p r o d u c t s as well as oil a n d gas) a r e split. It is c l e a r f r o m
T a b l e 1 t h a t G S C g r o u p 8 is t h e p r e c i o u s m e t a l s f u n d c a t e g o r y : it i n c l u d e s n o
f u n d s o t h e r t h a n m e t a l s e c t o r funds. It a l s o a p p e a r s t h a t g r o u p 1 is c o m p o s e d
 S.J. Brown, tKN. Goelzmann/Journal o f Financial Economics 43 gl997) 373 399 385

mostly of growth-and-income funds, group 2 is composed mostly of growth

funds, and most utility sector funds fall into group 3, suggesting that this third
group is an equity-income category.
Table 2 provides further insight into the characteristics of the GSC categories.
For each category, we estimate the mean and standard deviation of portfolio
weights, assuming a 24-month (nonoverlapping) return interval. Following
Sharpe (1992), we constrain the coefficients to be positive, so that they can be
interpreted as weights in short-sale constrained analogue portfolios. Groups
1 and 2 have a large average exposure to the S&P 500, groups 4 and 6
have a large exposure to the small-company stock index, and groups 5 and
7 have large exposures to non-U.S, indexes. Group 8 has a large exposure to
gold.
Table 3 provides further evidence on the dynamics of manager strategies. We
decrease the nonoverlapping estimation interval to six months in order to pick
up variations in exposure to key indexes. In addition, we estimate the correla-
tion of the style return to the previous period's index return. Thus, positive
correlations indicate 'trend-chasing', while negative coefficients indicate a 'con-
trarian' approach. Groups 1 and 3 both have negative portfolio weight correla-
tions to lagged S&P 500 return values. Groups 5, 6, and 7 all have positive
portfolio weight correlations to lagged S&P 500 return values. Group 7, an
international style, is most heavily weighted towards the EAFE index and is
little invested in the U.S. market. Group 7 managers tend to buy the EAFE
stocks when returns were low last period. Group 5, the other international style,
has much greater weight on the S&P 500, and appears to be a 'trend chaser' with
respect to the U.S. market.
Table 4 reports cross-tabulations of Morningstar 'On-Disc' categories with
GSC and Morningstar groups. It reveals useful information about fund strat-
egies. For instance, it indicates average P/E ratios, average price-to-book ratios,
and average ex post five-year earnings growth. These measures have been found
to explain differences in security returns. They also appear to explain differences
in style classifications. While these data represent a snapshot of the funds as of
the last date in our data base, we believe they provide an important validation
for the style classification procedure.
There are two styles that are composed of Morningstar 'small-cap' funds.
Group 4 managers invest in stocks with low price-to-book ratios and low
price-earnings ratios. These are 'value' managers. Group 6 managers invest in
companies with high price-to-book ratios and high ex post earnings growth.
These are 'glamour' stock managers who purchase companies that have grown
rapidly in the past. Group 4 managers buy stocks with low betas, and group
6 managers buy stocks with high betas. Group 4 managers buy financials,
cyclicals, and services, and Group 6 managers buy health care stocks and high
technology issues. In the terminology of Lakonishok, Shleifer, and Vishny
(1994), these are "glamour' managers. In view of the behavioral model proposed
386 S.J. Brown, I4qN. Goetzmann/dournal ~ / F i n a n c i a l Economics 43 (1997) 373 399

.= .,-.

.q-

2~

p..

.-z

i~ -

~. =-.
.~_ -Z
p.-

..= ,, ¢.3~" ttn, ttn

Ca.

Q
¢q ~ eq ¢q ,~- =G;
~D

-a

..e g
~. ~ ~., ~. ",2 ;2
 S.J. Brown, W.N. Goetzmann Journal q/'Financial Economics 43 (1997) 373 399 387

Table 3
Mean standard deviation and trading correlations of 6-month (nonoverlapping) Sharpe implied
portfolio weights, December 1978 to December 1994

S&P T bills EAFE-US

Group l
Mean 0.88868 0.07536 0.03596
Std. dev. 0.06872 0.06268 0.04653
Corr. - 0.36285 -- 0.11794 0.08072
Group 2
Mean 0.92767 0.02873 0.04361
Std. dev. 0.14101 0.07967 0.07964
Corr 0.00976 0.02262 0.04144
Group 3
Mean 0.65325 0.27648 0.07027
Std. dev. 0.12637 0.12772 0.07421
Corr 0.33967 - 0.05222 0.07570
Group 4
Mean 0.80581 0.09788 0.09632
Std. dev. 0.21447 0,16608 0.15136
Corr. 0,04370 - 0.08677 0.08074
Group 5
Mean 0.53262 0.13748 0.32990
Std. dev, 0.29663 0.18198 0.19950
Corr. 0.33378 0.05611 0.04871
Group 6
Mean 0.90455 0.02557 0.06988
Std. dev. 0.20093 0.09784 0.13963
Corr. 0.18407 0.04409 0.02801
G~vup 7
Mean 0.10346 0.14319 0.75335
Std. dev. 0.12823 0.15949 0.18585
Corr. (I.25193 0.10338 0.26766
Group 8
Mean 0.17685 0.41080 0.41235
Std. dev. 0.30928 0.41667 0.40264
Corr. 0.25022 - 0.00200 0.06077

This table reports the summary statistics for the time-series of Sharpe coefficients, i.e., implied
portfolio weights calculated using the procedure in Sharpe (1992) for each GSC style over the period
1976 through 1994.
Coefficients are constrained to be constant over rolling 6-month periods. EAFE-US is the EAF'E
index of global equity returns not including the U.S. market. Correlation is between change in
portfolio position and previous period index return. Change in portfolio position is measured as (- 1,
0, + 1~ relative to the previous semiannual portfolio bought and held into the current period.

b y t h e s e a u t h o r s , it is n o t s u r p r i s i n g t o f i n d t h a t t h e s e g l a m o u r m a n a g e r s a r e
a l s o ' t r e n d c h a s e r s ' a s is e v i d e n t f r o m t h e r e s u l t s o f t h e p r e v i o u s t a b l e , e n g a g i n g
in a l m o s t t w i c e t h e a m o u n t of trading of their 'value' counterparts in g r o u p 4.
388 S.J. Brown. ~'I'\N. Goetzmann Journal O/ Financial Economtcs 43 (1997) 373 399

e.--~

i i~ ~ ~
, _:~--

Z ,.d Z~ -- ,-4 ~< .--; e-~

L~

~ ~ ~< z ~ ~ ~: ,.~ ~ - r-: ~ ~,

oq r~, ~ ~ ~o 04 oq

~., ,~-= ~ ~ ~ ~ ~, ~, ~ ~

c-q . . . . .

>, ._

r~~,

E DIJ
,3
 S.J, Brown. W.N. Goetzmann Journal q/'Financial Economics 43 (19971 373- 399 389

ca

0-

cZ

o @ ,.c ,_ E . _o

~ . ~ .~
d E ~

" N ~

O
- - ej ._

~-~ g~_=

~,.c -o~
- ~i ,z-; '~ r,~, r--
d --cK '..-'; ,,'g g & ~ ,",-; ~ ~ ~ d ',~ .,~
: ,'-i
_ ri = {

2~ ,'b

r~ r ~- r ~- ~ ~.o ~ ~1 -- t~, r~, ~ ~ ~ ¢~, oo ~ r-- ~ ~l ~.c ,'-',-, ,.r., =,c r ~-

G; -=

PL ~ 2
~ z= Z ~
L~

:~ o U E ~.:.< --

~" ©
390 S.J. Brown, W.N. Goetzmann/Journal ~/'kTnancial Economics 43 (1997) 373 399

l-aken together, Tables 1, 2, and 3 suggest a categorization that is somewhat

different from the typical industry groupings. The following summarizes our
GSC styles:
Category 1: ~Growth and Income" Comprised primarily of Morningstar
'growth' and 'growth-and-income' funds, Category 1 funds have
the highest positive weights of any category on the S&P 500. They
invest in relatively large companies.
Category 2: "Growth' Comprised primarily of Morningstar 'growth' funds,
Category 2 funds have a major exposure to the S&P 500 and to
a lesser extent to small stocks. Their exposure to debt asset classes
is minor.
Category 3: "Income' Comprised primarily of Morningstar 'equity income',
"growth-and-income', and "utility sector' categories, Category
3 funds have the highest cash balances and the highest exposure to
debt asset classes.
Category 4: "Value' Comprised mostly of "small-cap' funds, this category
seeks stocks with low price-to-book and low price-earnings ratios.
This is consistent with value management.
Category 5: 'Global Timim/ Comprised principally of non-US, equities, this
category nonetheless pursues a dynamic strategy of increasing
exposure to the U.S. market when it rises; the variability of this
U.S. exposure suggests a timing strategy.
Category 6: ~Glamour" Comprised primarily of Morningstar 'growth' and
'small company" funds, Category 6 funds have a major exposure to
small-cap stocks. The equities in the typical portfolio have relatively
high price-to-book and PIE ratios and high ex post five-year
earnings growth. These are also domestic 'trend-chasers', display-
ing positive correlation to preceding S&P index returns.
Category 7: "International' Comprised of funds that are not strongly exposed
to the U.S. market through time, Category 7 funds appear to vary
their exposure to European and Pacific markets considerably.
Category 8: "Metal Funds' Comprised entirely of funds from the ~precious
metals and commodities' Morningstar category.

4.2. Stabilita' o/styles

How consistent are the style classifications'? Because of their statistical nature,
all classification schemes run the risk of misclassification. Funds at the margin
between two styles, for instance, may be difficult to confidently allocate to one
 S.J. Brown, 14,iN. Goetzmann/Journal of Financial Economics 43 (1997) 373 399 391

style or the other. To address the problem of estimation error, we use a boot-
strapping procedure to determine the frequency with which funds 'switch'
classifications. Using a single 24-month window, we repeatedly apply the SC
algorithm to the same data and count the number of changes in the pairwise
associations between each fund in the sample. We find an average 'switching
rate' of 11%.3 We also bootstrap the switching rate under the null hypothesis of
no cross-sectional structure. This null is constructed by forming 24-month
samples via random draws without replacement from actual fund returns. The
typical rate of change under this null is 27.3%.
The bootstrapped null distribution is helpful in examining the question of
style stability through time. To do this, we apply the SC algorithm to rolling
24-month windows of mutual fund returns. As a measure of the style stability
over successive 24-month windows, each year we count the number of changes
in pairwise associations between each fund. The average annual percentage of
fund associations that change each year is 17.6%. This is higher than the 11%
we would expect due solely to statistical variation conditional on no classifica-
tion change, but considerably less than we would expect under the null, i.e., if the
classifications were spurious. In fact, we find that 12 of 16 of the sample years
have percentage changes below the 5% quantile of the bootstrapped null
distribution, allowing us to reject the null for that year. For further details of
bootstrapping association frequencies see Abraham, Goetzmann, and Wachter
(1994) and Goetzmann and Wachter (1995).
Our bootstrapping tests suggest that the GSC and SC algorithms manifest
statistical variation. Like most statistical measures, they are noisy. Despite this
noise, our tests reject the hypothesis that our management styles are spurious.
On the other hand, the bootstrapping tests are silent on the usefulness of the
resulting classification scheme. The true value of style classification rests in the
extent to which it successfully explains out-of-sample fund performance. The
next section addresses the usefulness of the SC and GSC classifications.

4.3. Explanatory power o['styles

How well do estimated fund styles explain cross-sectional variation in out-of-

sample fund returns? Table 5 reports the results of out-of-sample cross-sectional
regressions for each of the categorization methods as well as for the industry

3 Note that this measure overrepresents the percentage of funds that change classification each year.
For example, take a sample of five funds, two of which remain in one category over two iterations,
and two of which remain in the other category over two iterations. However, the fifth fund changes
its association from the first group to the second group. This results is a change of association with
each of the four other funds in the sample out of the possible ten pairwise associations a 40%
switching rate, generated by only 20% of the funds changing classification. When the number of
funds in each of the two groups is equal, the misclassification percentage is half the swtching rate.
392 S.J. B~'own, W.N. G o e t z m a n n J o u r n a l ~?I F i n a n c i a l E ~ o n o m i c s 43 (1997) 373 399

~-z =_

o~

-z .~

-¸~ oi ~

8 8

> ":5
z
 S.J. Brown, W.N. Goetzmann,Journal o['Financial Economics 43 (1997) 373 399 393

~o=~.~
N ° "-- ~ E

?
~ [..- "~. ~= _= ~E N -
Io

~ ~'~,-- ~ ~
= _~
~ ~ ,.~
-~ ~,~&~
-:.-:. o ~ = ~ 5

.~o~

'-=

sN-~;

~g

Eo 7-4

i
394 S.J. Brown, W.N. Goet:mann Journal o/Financial Economics 43 (1997) 373 399

objective classifications. We omit sector funds from the analysis, since there is
relatively little ambiguity about their classification. Instead of using the entire
history of fund returns to form styles, we use the rolling period of 24 months for
estimation purposes. As shown in the preceding section, this results in less
'stable' styles, but it relaxes the assumption that funds belong to the same style
over the entire period and only uses ex ante industry information.
Columns 1, 2, and 3 in each panel show the adjusted R z that results from the
application of the iterative relocation algorithm to different spaces: the space of
returns, the space of 'Sharpe coefficients', and the space of principal component
loadings. Column 4 reports the results based on the industry objective classifica-
tion. Notice that, although adjusted RZs differ for various estimation intervals,
grouping in the space of returns and grouping in the space of factor loadings
typically explain significant amounts of performance. Grouping funds according
to the Sharpe coefficients performs about as well as using the industry codes.
This may be due to the fact that, for any fund, a significant number of coefficients
are zero, due to the nonnegativity constraint. For each estimation period, the
GSC algorithm applied to returns marginally outperforms the algorithm
applied to loadings on principal components. This may be due to the fact that
the model of classification in Eq. (1) is well-specified when loadings change
through time, but principal components rely upon stationary loadings. The
GSC categories explain about one-third of cross-sectional variation of returns,
ex ante. The Weisenberger categories explain, on average, 16% of the variation
in fund returns, while classifying funds according to Sharpe coefficients explains
only about 8%, on average.
The last four columns in each panel of Table 5 report the percentage of
cross-sectional variation explained by the estimated factor loadings themselves.
We would expect these to have greater explanatory power, since they are
continuous rather than dummy variables. This is particularly important for
outlying funds that have extreme exposures to some factor. While the GSC
algorithm will either group this outlier by itself or lump it in with distant
neighbors, the factor loadings themselves may capture the magnitude of its
deviation in cross-section. In addition to using the Sharpe coefficients and
principal component factor loadings as regressors, we also create indexes based
on the SC centers in the space of returns, and estimate unconstrained loadings
on passive indexes. This last column allows us to examine how much explana-
tory power the Sharpe procedure gives up in return for its positivity constraint.
The Sharpe positivity constraint is useful, because it allows the coefficients to be
interpreted as a vector of portfolio weights on investable indexes. Our style
categories do not have this property. Consequently, our GSC procedure is not
intended as a competing procedure to the Sharpe 'style analysis'. In this paper,
we show that the two tools can be used together to identify common strategies
among managers. The GSC procedure identifies aggregate behavior, and the
Sharpe procedure helps interpret it as strategy.
 S.J. Brown, W.N. Goetzmann/Journal o/Financial Economics 43 (1997) 373 399 395

The second panel of Table 5 shows that the loadings themselves all perform
better than the classification indicators. Typically, they explain on average
about 6 percentage points more cross-sectional variation out-of-sample. This
suggests that, in absolute terms, factor loadings, however they are constructed,
are a superior method of risk adjustment. On the other hand, the GSC proced-
ure does not do badly on a relative scale. Grouping by manager style is not an
alternative method for risk-adjusting manager returns. However, given that
benchmarking by style is a common practice, our analysis indicates that there is
not a great deal of information lost by using simple style classifications that are
appropriately chosen.
It is interesting to note that the loadings on the GSC centers typically
outperform the constrained loadings on prespecified financial indexes and do
a little better than the loadings on principal factors. They do almost as well as
the unconstrained loadings on the prespecified factors. It is tempting to conjec-
ture that the 'glamour' vs. 'value' division in the styles is responsible for the
success of the simple multinomial statistic, since this division may capture one of
the fundamental factors found to be superior in out-of-sample tests on U.S.
equities (see Fama and French, 1992; Lakonishok, Shleifer, and Vishny, 1994).

4.4. Interpretation

It is not surprising that the categories based on returns outperform the

standard industry classifications. Categories like 'growth' and 'growth and
income' represent an invitation to fund management gamesmanship. Once
a fund is classified into a particular category, there is little incentive to pursue an
investment strategy that will ensure that future fund performance will be close to
the category average in the future. Sirri and Tufano (1992) and Goetzmann and
Peles (1996) report evidence that mutual fund investors flock to superior
performers in each fund category. Given this information, fund managers are
not rewarded by maintaining strategies consistent with their industry classifica-
tion. As a result, we find that the GSC categories formed via returns and
loadings 'agree to disagree' with the standard industry classifications. It is
evident that the industry classifications have relatively little power to explain
differential fund performance.
It is also not surprising that classification based on returns typically equals or
beats classification based on principal factor loadings for longer holding
periods. The principal component loadings represent a linear projection of the
monthly returns on a reduced space. Both the reduction of dimensionality and
the linearity of the projection represent constraints that 'coarsen' the informa-
tion about fund returns. The advantage of the classification based on scaled
principal components is the natural interpretation of the groups in terms of
systematic risk classes. The disadvantage is that funds are subject to misclassifi-
cation due to nonlinear strategies.
396 S.J. Brown, ~]/v~ Goetzmann'Journal ol Financial Economics 43 f1997) 373 399

It is likewise not surprising to find that loadings on principal components

outperform loadings on prespecified asset series. The principal components were
selected so as to maximally spread returns in the preceding period. While Chen,
Roll, and Ross (1986) show that prespecified factors likewise spread returns,
collinearities among factors appear to increase the standard error of the factor
loadings, and thus make fund classification via the SC procedure more difficult.
The major advantage to estimating positive-constrained coefficients on pre-
specified factors is that it provides some insight into the composition and
behavior of the categories.
Our results provide mixed signals regarding the approach to characterizing
funds according to their profile of loadings on prespecified indexes. Even when
the loadings are not constrained to be positive, we find no evident advantage in
term of explanatory value. When loadings are used to identify styles, they
perform as poorly as the standard industry classifications. Thus, their incremen-
tal advantage obtains in circumstances when the loadings themselves, rather
than a derived style classification, can be used. Their disadvantage is that
loadings on correlated indexes will be estimated with inaccuracy, due to
collinearities.

5. Conclusion

We show that a simple procedure based on the switching regression techno-

logy applied to monthly returns dominates other management style classifica-
tions based on standard investment objectives and also does better than
classification based on observed factor loadings. Given the potential 'category
gaming' by fund managers, it is useful to have a classification method that uses
publicly available and independently verifiable time-series information about
returns.
Why are we interested in manager styles at all'? Our classification algorithm
identifies a few major types of fund strategies. While these may not exhaust the
range of different fund managers, they do provide an overview of the strategies
that differentiate managers. Our results validate the use of traditional, self-
reported categories such as equity income, growth and income, and growth.
However, funds apparently do not always correctly categorize themselves. We
find two somewhat surprising divergences from standard industrial categories.
First, we find evidence of "value+ vs. 'glamour" managers, rather than a mono-
lithic small-cap group. Second, we find evidence that style involves dynamic
strategies, rather than simply fixed portfolio weights.
The focus of this paper is the development of a classification algorithm that is
consistent with commonly used asset pricing models. Such classification is
unlikely to replace the use of continuous, multivariate models for risk adjust-
ment, nor should it. The absolute magnitude of systematic risk exposures will
 S.J. Brown, W.N. G o e z z m a n n : J o u r n a l o['Financial Economics 43:1997/ 373 399 397

always be i m p o r t a n t to portfolio decisions. O u r analysis suggests, however, that

there are a few intentional, recognizable strategies within the population of
investment managers. As the finance profession continues to investigate the
behavioral basis for investment decisions, it will be useful to further identify and
study these c o m m o n patterns. The G S C algorithm is a potentially useful tool for
doing so. Its advantage over heuristic classification is that researchers can use it
to d e c o m p o s e 'styles' into more familiar measures such as time-varying factor
loadings and risk premiums.

Appendix: GSC methodology

The G S C can be thought of as an extension of a standard iterative relocation

algorithm such as k-means. It is motivated by the insight that if rl, = /~i, +
e, and var(e~,) = a:,z~,, where z, is i.i.d, normal for both I and t with mean 0 and
variance 1, then var(ei.) = a/2E(o't) 2 and var(e.,) = o,2 E(at)',"~ where we interpret
a~ and at as independent and identically distributed drawings from a population
of time-series and cross-sectional standard deviations respectively independent
of z~,. This is not inconsistent with a variety of G A R C H or other processes for
returns. As a result, var(e,) is proportional to var(ei.) * var(e,). Therefore, it is
a simple matter to infer the variance of each time and fund residual as p r o p o r -
tional to the marginal time and fund variances in excess of the estimate of t*~,.
Since the efficient (GLS) estimate o f / ~ , depends on a~, we need a second pass
(GLS) for efficient estimation of both #~, and var[ei,).
Computationally, we proceed as follows. For a given definition of the clusters,
we calculate

^ i~l
tZlt - -
c o u n t ( / c I)"
Once this is done, we c o m p u t e

~-'it = R i t - - f l i t .

For all I, we then calculate var(~i.), and for all t we calculate var(~ ,). N u m e r -
ically, these numbers tend to be small, so we normalize them by the average
marginal variances.
We now do a G L S correction for the mean, computing.

fl *, = Z g,, ~i 1
i~i var ( e ; i . ) v a r (3i.)"
We use this updated G L S estimate of the mean to update variance measures. We
also use this formula to update centroid means whenever funds are switched
from one cluster to the next, although for c o m p u t a t i o n a l simplicity we do not
398 S.J. Brown. W.:vi G o e t z m a n n J o u r n a l o / F i n a n c i a l E c o n o m i c s 43 (1997) 3 7 3 - 3 9 9

update variance measures at each switch. Denote the clusters formed at the jth
switch as l(j). Then the criterion function at t h e j t h switch is proportional to

1" ( R i , __ f i , t ) 2
ssOj = ,~=1 y~ y
= lel~iel var (3*)var
• (~*)
.

using the result that var(e.) is proportional to var(ei.)*var(eit).

References

Abraham, Jesse, William N. Goetzmann, and Susan Wachter, 1994, Homogeneous groupings of
metropolitan housing markets, Journal of Housing Economics 3. 186-206.
Amihud. Yakov, Bent Christiansen, and Haim Mendelson, 1992, Further evidence on the risk return
relationship. Working paper (Stern School of Business, New York University, New York, NY).
Berry, Michael, Edwin Burmeister, and Marjorie McElroy, 1988, Sorting out risks using known
APT factors, Financial Analysts Journal 44, 29 42.
Brown, Stephen J.. 1989, The number of factors in security returns, Journal of Finance 44,
1247-1262.
Brown, Stephen J. and William Goetzmann, 1995, Performance persistence, Journal of Finance 50,
679 698.
Brown, Keith C.. W.V. Harlow, and Laura T. Starks, 1993, Of tournaments and temptations:
Analysis of managerial incentives in the mutual fund industry, Working paper (University of
Texas, Austin, TXI.
Carleton, Willard and Victor McGee, 1970, Piecewise regression, Journal of the American Statistical
Association, May, 1109 1124.
Chen. Nai-fu, Richard Roll, and Stephen Ross, 1986, Economic forces and the stock market, Journal
of Business 59, 383 404.
Christopherson, Jon A., 1995, Equity style classifications, Journal of Portfolio Management 2l,
32 43.
Connor, Gregory and Robert Korajczyk, 1986, Performance measurement with the arbitrage
pricing theory: A new framework for analysis, Journal of Financial Economics 15, 373 394.
Connor, Gregory and Robert Korajczyk, 1991. The attributes, behavior and performance of U.S.
mutual funds, Review of Quantitative Finance and Accounting 1, 5 26.
Donnelley, Barbara, 1992, What's in a name'? Some mutual funds make it difficult to judge, Wall
Street Journal, May 5, Section C. p. 1, col. 3.
Dybvig, Philip H. and Stephen A. Ross, 1985, ]'he analytics of performance measurement using
a security market line, Journal of Finance 40, 401 416.
Ellon, Edwin, Martin Gruber, S. Das, and M. Hlavka, 1993, Efficiency with costly information:
A reinterpretation of evidence for managed portfolios, Review of Financial Studies 6, I 22.
Elton. Edwin and Martin Gruber, 1970, Improved forecasting through the design of homogeneous
groupings, Journal of Business 44. 432-450.
Fama. Eugene and Kenneth French, 1992, The cross-section of expected stock returns, Journal of
Finance 47, 427 466.
Ferson. Wayne and Rudi Schadt. 1996, Measuring fund strategy and performance in changing
economic conditions. Journal of Finance 51,425 46l.
Goetzmarln. William N. and Nadav Peles, 1996, Cognitive dissonance and mutual fund investors,
Journal of Financial Research 20.
Goetzmann. William and Susan Wachter. 1995. Clustering methods for real estate portfolios,
Journal of Real Estate Economics 23, 271 310.
 S.J. Brown, WN. Goetzmann/Journal (~f Financial Economics 43 (1997) 373 399 399

Grinblatt, Mark and Sheridan Titman, 1988, The evaluation of mutual fund performance: An
analysis of monthly returns, Working paper (John E. Anderson Graduate School of Manage-
ment, University of California, Los Angeles, CA).
Grinblatt, Mark and Sheridan Titman, 1989, Mutual fund performance: An analysis of quarterly
portfolio holdings, Journal of Business 62, 393 416.
Grinblatt, Mark and Sheridan Titman, 1993, Performance measurement without benchmarks: An
examination of mutual fund returns, Journal of Business 66:1.47 68.
Grinblatt, Mark, Sheridan Titman, and Russ Wermers, 1993. Momentum strategies, portfolio
performance and herding: A study of mutual fund behavior, American Economic Review 85,
1088-1105.
Hartigan, John A., 1975, Clustering algorithms (Wiley, New York, NYi.
Hendricks, D., J. Patel, and R. Zeckhauser, 1993, Hot hands in mutual funds: The persistence of
performance: 1974 88, Journal of Finance 48, 93 130.
lbbotson, Roger and William Goetzmann, 1994, Do winners repeat? Patterns in mutual fund
performance. Journal of Portfolio Management 20, Winter, 9 17.
Jensen, Michael, 1968, The performance of mutual funds in the period 1945 1964, Journal of
Finance 23, 389 416.
Kim, Moon, Ravi Shukla, and Michael Tomas, 1995. Wolf ill sheep's clothing: Do mutual fund
objectives tell the truth'?, Working paper (Syracuse University, Syracuse, NY).
Lakonishok, Josef, Andrei Shleifer, and Robert Vishny. 1994, Contrarian investment, extrapolation
and investment risk, Journal of Finance 49, 1541 1578.
Lakonishok, Josef, Andrei Shleifer, Richard Thaler, and Robert Vishny. 1991, Window dressing by
pension fund managers, American Economic Review 81. 227 231.
Lehmann, Bruce N. and David Modest, 1987, Mutual fund performance evaluation: A comparison
of benchmarks and a benchmark of comparisons, Journal of Finance 42, 233 265.
Quandt, Richard, 1959, The estimation of the parameters in a linear regression system obeying two
regimes, Journal of the American Statistical Association 53. 233 265.
Quandt, Richard, 1960, Tests of the hypothesis that a linear regression system obeys two regimes,
Journal of the American Statistical Association 55, 324 330.
Quandt, Richard and James Ramsey, 1978, Estimating mixture of normal distributions and switch-
ing regressions. Journal of the American Statistical Association 73,730 752.
Sharpe, William F., 1992. Asset allocation: Management style and performance measurement,
Journal of Portfolio Management, Winter, 7 19.
Sirri. Erik and Peter Tufano, 1992, The demand for mutual fund services by individual investors,
Working paper (Harvard Business School, Cambridge, MAI.
Tierney, David E. and Kenneth Winston, 1991, Using generic benchmarks to present manager
styles. Journal of Portfolio Management 17. Summer, 33 36.
Trzcinka, Charles A.. 1995, Equity style classifications: A comment, Journal of Portfolio Manage-
ment 21, Spring, 44 46.
Witkowski, Erik, 1994. Mutual fund misclassification : Evidence based on style analysis, B.A. thesis
(Harvard University, Cambridge, MAt.