Ferson Schadt Methodology
Ferson Schadt Methodology
Ferson Schadt Methodology
Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at http://www.jstor.org/page/
info/about/policies/terms.jsp
JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of content
in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms of scholarship.
For more information about JSTOR, please contact [email protected].
Wiley and American Finance Association are collaborating with JSTOR to digitize, preserve and extend access to The Journal
of Finance.
http://www.jstor.org
This content downloaded from 14.139.157.28 on Sat, 11 Jul 2015 03:14:07 UTC
All use subject to JSTOR Terms and Conditions
in
THEPROBLEM
OFACCURATELY
measuring the performance of managed portfolios
remains largely unsolved after more than 30 years of work by academics and
practitioners. Standard measures of performance, designed to detect security
selection or market timing ability, are known to suffer from a number of biases.
Traditional measures use unconditional expected returns as the baseline. For
example, an "alpha" may be calculated as the past average return of a fund in
excess of a risk-free rate, minus a fixed beta times the average excess return
of a benchmark portfolio. However, if expected returns and risks vary over
time, such an unconditional approach is likely to be unreliable. Common time
variation in risks and risk premiums will be confused with average performance.
* Ferson is from the University of Washington School of Business Administration,Department
of Finance and Business Economics DJ-10, Seattle, Washington 98195, and Schadt is from the
Edwin L. Cox School of Business, Southern Methodist University, Dallas, Texas 75275-0333. We
thank numerous colleagues for comments and helpful discussions, including Connie Becker, David
P. Brown, Stephen Brown, Jon Christopherson,John Cochrane,William Goetzmann,Mark Grinblatt, CampbellHarvey, Ravi Jagannathan, Debbie Lucas, Jianping Mei, Ed Rice, the editor Rene
Stulz, Bill Sharpe, Andy Siegel, Karl Snow, and an anonymousreferee. Vincent Wartherdeserves
special recognition for providing some of the empirical analysis described herein, and David
Modest and Peter Knez deserve thanks for supplying some of the data. This paper was presented
in workshops at the Universities of Alberta, Arizona State, British Columbia, California at Los
Angeles, Chicago, Notre Dame, Oklahoma, Washington, and at J. P. Morgan Investment Management, and at the following conferences: the 1993 European Finance Association, the 1994
National Bureau of EconomicResearch Summer Institute, the third annual Osaka Conferenceon
Finance, the 1994 Southwestern Finance Association and the 1994 Western Finance Association.
Ferson acknowledges financial support from the Pigott-Paccarprofessorshipat the University of
Washington.
425
This content downloaded from 14.139.157.28 on Sat, 11 Jul 2015 03:14:07 UTC
All use subject to JSTOR Terms and Conditions
426
The problem of confounding variation in mutual fund risks and risk premia
*has long been recognized (e.g. Jensen (1972), Grant, (1977)), but previous
studies interpreted it as reflecting superior information or market timing
ability. We emphasize a different interpretation. This paper takes the view
that a managed portfolio strategy that can be replicated using readily available public information should not be judged as having superior performance.
Traditional, unconditional models can ascribe abnormal performance to an
investment strategy that is based only on public information. (See Breen,
Glosten, and Jagannathan (1989) for an example.) Using instruments for the
time-varying expectations, it is possible to control commonvariation caused by
public information and reduce this source of bias.
Recent studies have documented that the returns and risks of stocks and
bonds are predictable over time, using dividend yields, interest rates, and
other variables. If this predictability reflects changing required returns in
equilibrium, then measures of investment performance should accommodate
the time variation.
There is reason to think that predictability using predetermined instruments represents changing required returns. For example, standard beta
pricing models of expected returns can capture a substantial fraction of the
predictability in passive portfolios (Ferson and Harvey (1991), Evans (1994),
Ferson and Korajczyk(1995)). Also, conditional versions of simple asset pricing
models may be able to explain the cross-section of returns better than unconditional models (e.g. Chan and Chen (1988), Cochrane (1992), Jagannathan
and Wang, (1996)).
We advocate conditional performance evaluation, using measures that are
This content downloaded from 14.139.157.28 on Sat, 11 Jul 2015 03:14:07 UTC
All use subject to JSTOR Terms and Conditions
427
I. The Models
The appeal of a conditional model for performance evaluation can be illustrated with the following hypothetical scenario. Suppose that the expected
market excess return and its volatility move together proportionately over
time with economic conditions, as in the model of Merton (1980). Consider a
mutual fund that wishes to keep its volatility relatively stable over time. Based
on economic conditions, the fund will lower its beta when the market is more
volatile and raise it in less volatile markets. The beta of the fund will be
negatively correlated with the market return, so the average excess return of
the fund will be less than the average beta of the fund applied to the average
market premium. The use of an unconditional model would lead to the conclusion that the fund has a negative alpha. However, this does not reflect
performance, but the fact that the fund takes more risk when the premium for
This content downloaded from 14.139.157.28 on Sat, 11 Jul 2015 03:14:07 UTC
All use subject to JSTOR Terms and Conditions
428
beta risk is low. A conditional model that controls for the time-varying betas
and market premium shows that the fund has neutral performance.
The rest of this section describes our basic assumptions and develops a
simple model for conditional performanceevaluation. We also discuss how our
models are related to the traditional, unconditional approach.
A. The Basic Assumptions
The first assumption is the form of an asset pricing model that describes the
conditional expected returns of the assets available to portfolio managers. We
use a conditional version of Sharpe's (1964) Capital Asset Pricing Model
(CAPM)to illustrate the approach. Of course, studies have rejected the CAPM
for conditional returns (e.g. Ferson, Kandel, and Stambaugh (1987), Bollerslev,
Engle, and Wooldridge(1988), and Harvey (1989)), and Roll (1978) shows that
the inferences about performance can be sensitive to the specification of an
inefficient benchmark. We therefore extend the analysis to a multiple-factor
asset pricing model, and find that the results are robust.
The second important assumption is a notion of market efficiency. The
traditional performance evaluation literature assumes that the user of a performance measure holds unconditional expectations. With this assumption the
use of any information by managers, including public information, may lead to
measured abnormal performance. In contrast, we assume that market prices
fully reflect readily available, public information. We hypothesize that managers may use this information to determine their portfolio strategies. The use
of public information should not imply abnormal performance, under semistrong form market efficiency, because investors can replicate on their own any
strategy which depends on public information.2
The third assumption required for our approach is a functional form for the
betas, or factor sensitivities of a managed portfolio. Time-variation in a managed portfoliobeta may come from three distinct sources. First, the betas of the
underlying assets may change over time. Second, the weights of a passive
strategy such as buy-and-hold, will vary as relative values change. Third, a
manager can actively manipulate the portfolio weights by departing from a
buy-and-hold strategy.
We model the combined effect of these factors on the risk exposures indirectly, as a "reduced form." We use a linear function, which is a natural
extension of traditional approaches. For example, Admati and Ross (1985), and
Admati, et al. (1986) assume that managers act as if they maximize a constant
absolute risk aversion expected utility function, defined over normally distributed variables. In this case the portfolio weights are linear functions of the
information, and if the betas of the underlying assets are fixed over time, the
2 Of course, the argument that investors can replicate or undo
managers'trades that are based
on public information assumes that investors can infer the trades. It also ignores any cost
advantages in trading that funds may have over investors and assumes that managers do not
waste resources by "churning"their clients' portfolios at cost.
This content downloaded from 14.139.157.28 on Sat, 11 Jul 2015 03:14:07 UTC
All use subject to JSTOR Terms and Conditions
429
rit+l =3im(Zt)
rmt+l + ui,t+
t - O,...
T-
1,
(la)
E(ui,t+llZt) = 0,
(lb)
E(ui,t+lrmt+llZt) = 0,
(lc)
where Rit+1 is the rate of return on the investment asset i between times t and
t + 1, rit = Rit - Rft is the excess return, Rft is the return of a one-month
Treasury bill, Zt is a vector of instruments for the information available at time
t, and rmt+1 is the excess return of the market factor. The /im(Zt) are the time
t conditional market betas of the excess return of asset i. Equation (lb) follows
from the market efficiency assumption and equation (ic) says that the f3im(Zt)
are conditional regression coefficients.
Equation (1) implies that any unbiased forecast of the difference between the
return of a security and the product of its beta and the excess return on the
market factor which differs from zero must be based on an information set that
is more informative than Zt. Using only the information Zt the forecast of this
difference is zero. A portfolio strategy that depends only on the public information Zt will satisfy a similar regression. The intercept, or "alpha" of the
regression should be zero, and the error term should not be related to the
public information variables.3
Because we hypothesize that the manager uses no more information than Zt
the portfolio beta, I3pm(Zt), is a function only of Zt. Using a Taylor series
3That is, if Rp,t+l = x(Zt)'Rt+1, where x() is an N-vector of weights and Rt+1 is the N-vector of
the available risky security returns, then the portfolio excess return will satisfy equation (1), with
I3pm(Zt) = x(Zt)'13m(Zt),where 13m(Zt) is the vector of the securities' conditional betas. The error term
in the regression for the portfolio strategy is upt+l=
x(Zt)'ut+i, where ut+1 is the vector of the
uit,+1's, and therefore
E(up,t+jjZt)
= E(x(Zt)'ut+jjZt)
=x(Zt)'E(ut+jjZt)
= 0.
This content downloaded from 14.139.157.28 on Sat, 11 Jul 2015 03:14:07 UTC
All use subject to JSTOR Terms and Conditions
430
= bop+ B,zt,
(2)
up,t+i,
(3)
+ spt+i
(4)
Taking the relevant expected values in (4) and comparing the result with (3)
shows that the model implies ap = 0, 61p = bop,and 82p = Bp.5
C. Interpreting the Conditional Model
Our approach may be interpreted as a special case of a general asset pricing
framework based on the expression E(mt+l Rt+llZt) = 1, where mt+l is a
stochastic discount factor and Rt+l is the vector of the gross returns of the
primitive assets available to portfoliomanagers. Our version of the conditional
CAPM implies that the stochastic discount factor is a linear function of the
market excess return, where the coefficients may depend linearly on Zt.
The regression (4) may also be interpreted as an unconditional multiple
factor model, where the market index is the first factor and the product of the
market and the lagged information variables are additional factors. The additional factors may be interpreted as the returns to dynamic strategies, which
hold zt units of the market index, financed by borrowing or selling zt in Trea4 This interpretation is an approximation,as it ignores the higher order terms in the Taylor
expansion. The informationvariables are demeaned for ease of exposition.
5 OLS estimation of the regression model imposes the same moment conditions as does Hansen's (1982) GMM estimator. Consider a linear conditional beta P,_- = b + B'zt-, in a linear
regression model Yt = x1t4-l + Et. The moment conditions:
Ut = XtYt - (xtx')(b
+ Bzt-),
E(utlzt-1)
= 0
would be the basis of the GMMestimation. Typically, the implementation of the GMMwould use
the implication:E(ut 0 zt-1) = 0. Considerthe OLS regression estimator of the linear model which
results from substituting the beta equation into the regression and note that the error terms are
related as Etxt = ut. It is easy to verify that the two sets of moment conditions are the same.
This content downloaded from 14.139.157.28 on Sat, 11 Jul 2015 03:14:07 UTC
All use subject to JSTOR Terms and Conditions
431
This content downloaded from 14.139.157.28 on Sat, 11 Jul 2015 03:14:07 UTC
All use subject to JSTOR Terms and Conditions
432
D. Multiple-Factor Models
It is easy to extend the analysis to use a conditional multiple-beta or exact
arbitrage pricing (APT) model for the expected returns on the assets available
to managers:6
K
i=O, ...
t
NAT
T-1
(5)
j=1
where the bi1(Zt),.. ., biK(Zt)are the time t conditional betas or factor loadings,
which measure the systematic risk of asset i relative to the K risk factors. The
Aj(Zt),j = 1, ... , K are the market prices of systematic risk or the expected
risk premiums.
In our multiple factor models, we replace equation (2) with a similar equation for each of the K factor-betas of a managed portfolio. That is, we model
each of the conditional betas as a linear function of the information. Our
unconditional K-factor model regression is a multiple regression of the excess
returns on a constant and the K factor-portfolios, and the intercept is the
unconditional alpha. In our conditional K-factor model, the regression equation has (L + 1)K + 1 regressors. The regressors are a constant, the K factorportfolios, and the products of the L information variables in Zt with the
K factor-portfolios.
E. Traditional Measures Revisited: Jensen's (Unconditional) Alpha
A traditional approach to measuring performance is to regress the excess
return of a portfolio on the market factor. Assuming that the market beta is
constant, the slope coefficient is the market beta and the intercept, ap, is the
unconditional alpha coefficient, which measures the average performance(e.g.
Jensen (1968)):
rPt+1 =
ap + bp rmt+1 + Vpt+l.
(6)
(7)
This content downloaded from 14.139.157.28 on Sat, 11 Jul 2015 03:14:07 UTC
All use subject to JSTOR Terms and Conditions
433
where E(@),Cov( ), and Var( ) denote the unconditional mean, covariance and
variance, respectively. The notation z rm refers to the vector of the products of
the lagged information variables with the market excess return.
The first term in the expression for plim(ap) reflects the fact that the OLS
slope-coefficient in equation (6) is not a consistent estimate of even the average
conditional beta, bop. The second term reflects the covariance between the
conditional beta of the fund and the future market return. Equation (7) may be
interpreted as a missing-variable bias in equation (6). The effect depends on
the regression coefficients of the omitted variables, zrm, on the included variable, rm. WhenBp = 0, the managed portfolio beta is not a function of the
public information, the conditional and unconditional betas are the same, and
the probability limit of the intercept is zero.
Evaluating equation (7) at the sample moments and using the OLS estimates from (4), we can obtain bias-adjusted estimates of alpha. The adjusted
alpha is numerically identical to the OLS estimate of ap in equation (4). We can
also use our conditional K-factor models to adjust for bias in the alphas that
results from omitting the conditioning information in those models. The adjustment is a straightforward extension of the adjustment to alpha in the
CAPM.7
Equation (7) is similar to expressions derived by Jensen (1972), Grant
(1977), and Grinblatt and Titman (1989b) in an unconditional setting. However, the interpretation is different. Traditional studies view the covariance
between beta and the future market return as a result of portfolio managers'
superior information. Equation (7) is developed under the hypothesis that
managers do not have superior information. In our model the hypothesis of no
abnormal performance allows a covariance between beta and the future market return, because of their common dependence on the public information
variables.
7 Standard analysis shows that the slope coefficient in the unconditional factor model regression for asset i converges in probability to:
of the expected values of the K conditional betas for asset i, and BP, is
the beta response coefficients for asset i, and F is the vector of the
true alpha is zero, the intercept in the unconditional factor model
to:
This content downloaded from 14.139.157.28 on Sat, 11 Jul 2015 03:14:07 UTC
All use subject to JSTOR Terms and Conditions
434
ap + bprmt+l +
'ytmu[rm,t+i]2 + Vpt+l,
(8)
where the coefficient ytmu measures market timing ability. Admati et al. (1986)
describe a model in which a manager with constant absolute-risk aversion in
a normally distributed world observes at time t the private signal, rmt+l + 7t,
equal to the future market return plus noise. The manager's response is to
change the portfolio beta as a linear function of the signal. They show that the
Ytmu coefficient in regression (8) is positive if the manager increases beta when
This content downloaded from 14.139.157.28 on Sat, 11 Jul 2015 03:14:07 UTC
All use subject to JSTOR Terms and Conditions
435
the signal about the market is positive. The hypothesis of no abnormal performance implies that ytmu is zero.
Using essentially the same analysis as Admati et al. (1986), we propose a
conditional version of the Treynor-Mazuy regression. Assume that the manager observes the vector (zt, rmt+1 + -rj)at time t, and the question is how to
allocate funds between the market portfolio and a risk-free asset. With exponential utility and normal distributions, the demand for the risky asset is a
linear function of the information. In a two-asset model, the portfolio weight
on the market index is the portfolio beta, and it is a linear function of zt and
(rmt+? + 71j).Replacing equation (2) with this linear function and letting 'Tijoin
the regression error term, we have a conditional version of the Treynor-Mazuy
regression:
rpt+1
Vpt+1
(9)
where the coefficient vector Cp captures the response of the manager's beta to
the public information, Zt. The coefficient ytmc measures the sensitivity of the
manager's beta to the private market timing signal. The term Cp(ztrmt+?)in
equation (9) controls for the public information effect, which would bias the
coefficients in the original Treynor-Mazuy model of equation (8). The new term
in our model captures the part of the quadratic term in the Treynor-Mazuy
model that is attributed to the public information variables. In the conditional
model, the correlation of mutual fund betas with the future market return,
which can be attributed to the public information, is not considered to reflect
market timing ability.
C. The Merton-Henriksson Model
Merton and Henriksson (1981) and Henriksson (1984) describe an alternative model of market timing. In their model a manager attempts to forecast
when the market portfolio return will exceed the risk-free rate. When the
forecast is for an up market, the manager adjusts the portfolio to a higher
target beta. When the market forecast is pessimistic, a lower target beta is
used. Given this model, Merton and Henriksson show that if the manager can
time the market, the coefficient yu in the following regression is positive:
rpt+1= ap + bprmt+i+ 'y[rm,t+i]+ + vpt+i,
(10)
8 Merton and Henriksson proposed the regression (10) to separate market timing from security
selection ability. However, just as with the Treynor-Mazuy model, this separation is problematic,
as illustrated by Jagannathan and Korajczyk (1986). Glosten and Jagannathan (1994) provide
conditions under which the sum of the timing and selectivity components of performance can
correctly estimate the average (unconditional) value added by a manager. Their arguments can be
This content downloaded from 14.139.157.28 on Sat, 11 Jul 2015 03:14:07 UTC
All use subject to JSTOR Terms and Conditions
436
(1)
where
r*t+1=
A =Bup-Bd.
{*}is the indicator function. The null hypothesis of no market timing ability
implies that -y and A are zero. The alternative hypothesis of positive
market timing ability is that y, + A' zt > 0, which says that the conditional
beta is higher when the market is above its conditional mean, given public
information, than when it is below the conditional mean. This implies that
E(y, + A'zt) = y, > 0, which says that market timing is on average positive.9
This content downloaded from 14.139.157.28 on Sat, 11 Jul 2015 03:14:07 UTC
All use subject to JSTOR Terms and Conditions
437
series could be updated through December 1990 using total returns from
Morningstar, Inc. Funds were matched using the overlapping time series.10
Table I records the names of the funds, along with summary statistics for the
1968-1990 period. The funds are grouped by their Wiesenberger type as of
1982, which is roughly the middle of our sample period.1" The funds are
primarily equity funds, with objectives as classified by Wiesenberger varying
from "maximum capital gains" to "income." The excess returns are net of the
monthly return of investing in a Treasury bill with maturity greater than or
equal to one month. The bill data are from Ibbotson & Associates. We place
four of the funds in a "special" category. In two cases (Scudder International
and Templeton Growth) the funds had sizable holdings in foreign equity
markets. Century Shares Trust holds primarily securities issued by financial
firms. Fidelity Capital Trust merged into Fidelity Trend in 1982, so the two
series differ only before that date. One fund, Scudder Income, is a bond fund
that we assign to the income fund group.
Our sample of funds has the potential for survivorship bias, as it contains
only surviving funds. Survivorship may be expected to bias the relative performance measures upwards (e.g. Grinblatt and Titman (1988), Brown, Goetzmann, Ibbotson and Ross (1992), Brown and Goetzmann (1995), Malkiel
(1995)). If survivorship bias is important, our estimates of the overall performance of the mutual funds is too optimistic. However, we find that the
traditional measures suggest poor performance for the funds in our sample and
that the poor performance is removed when we control for public information
variables. It seems unlikely that survivorship bias can explain these results.
B. The Predetermined Information Variables
We use a collection of public information variables that previous studies
have shown are useful for predicting security returns and risks over time. The
variables are (1) the lagged level of the one-month Treasury bill yield, (2) the
lagged dividend yield of the CRSP value-weighted New York Stock Exchange
(NYSE) and American Stock Exchange (AMEX) stock index, (3) a lagged
measure of the slope of the term structure, (4) a lagged quality spread in the
corporate bond market, and (5) a dummy variable for the month of January.
We use the 30-day annualized Treasury bill yield from the CRSP RISKFREE
files to predict the future returns. This is based on the bill that is the closest
to one month to maturity at the end of the previous month, using the average
of bid and ask prices on the last trading day of each month. The dividend yield
is the price level at the end of the previous month on the CRSP value-weighted
index of NYSE firms, divided into the previous 12 months of dividend pay10 We are grateful to David Modest for making these data available and to Peter Knez for help
with matching and updating the data.
" Elton et al. (1992) evaluate the effects of classification errors that may arise from grouping a
sample of funds at the beginning versus at the end of the period used by Ippolito (1989), and they
found that only 12 of 143 funds changed categories. We conduct an analysis of the information in
the categories below.
This content downloaded from 14.139.157.28 on Sat, 11 Jul 2015 03:14:07 UTC
All use subject to JSTOR Terms and Conditions
438
Primary Investment
objective
policy
Standard
Mean Deviation
Minimum Maximum
Panel A
Boston Foundation F.
Financial Industrial In. F.
Franklin Custodian F.-In. Series
Keystone Income F. (K1)
Nation-Wide Secu:rities
Northeast Invest. Trust
Provident F. for In.
Putnam Income Fund
Scudder Income Fund
Security Investment F.
Sentinel Balanced Fund
United Income Fund.
Value Line Income Fund
Wellington Fund
Income Group Average
sgi
ig
i
i
isg
i
i
i
is
i
igs
i
i
sig
bal
flex
flex
flex
bal
flex
flex
flex
bonds
flex
bal
flex
flex
bal
0.19
0.49
0.20
0.15
0.29
0.04
0.27
0.06
0.00
0.26
0.24
0.27
0.31
0.19
3.98
4.16
3.41
3.04
3.80
2.12
4.49
2.63
3.08
4.13
2.85
4.43
4.20
3.50
-22.20
-18.60
-11.51
-14.10
-22.40
-6.01
-25.00
-8.81
-11.11
-18.04
-12.10
-13.10
-16.96
-13.80
11.15
14.17
14.17
9.03
11.62
9.09
13.22
10.19
.11.69
11.41
9.46
15.26
15.24
11.89
0.21
3.56
-15.27
11.97
This content downloaded from 14.139.157.28 on Sat, 11 Jul 2015 03:14:07 UTC
All use subject to JSTOR Terms and Conditions
439
Table 1-Continued
Wiesenberger
Classification
Mutual Funds by
1982 Name
Primary Investment
objective
policy
Standard
Mean Deviation Minimum Maximum
Panel A: -Continued
Colonial Fund
Composite Fund
Delaware Fund
Eaton and Howard Stock
Fidelity Fund
Financial Industrial Fund
Founders Mutual Fund
Guardian Mutual Fund
Investment Company of America
Investment Trust of Boston
Keystone High-Grade Common
Stock Fund (S1)
National Industries Fund
Philadelphia Fund
Pine Street Fund
Pioneer Fund
Safeco Equity Fund
Selected American Shares
Sentinel Common Stock Fund
Wall Street Fund
Washington Mutual Investors
gi
gis
gis
gi
gi
gi
gi
gi
gi
gi
gi
cs
cs
cs
cs
cs
cs
cs
cs
cs
cs
cs
0.11
0.16
0.31
0.08
0.32
0.23
0.09
0.49
0.41
0.13
0.08
3.71
4.37
4.67
4.54
4.57
4.79
4.20
4.99
4.44
4.63
4.80
-16.60
-14.50
-18.60
-17.60
-25.30
-25.00
-15.10
-24.70
-18.00
-20.20
-24.70
9.32
12.96
14.42
20.70
15.27
16.03
15.15
14.20
13.25
13.23
22.80
gi
gi
gi
gi
gi
gi
gi
gis
gi
cs
cs
cs
cs
cs
cs
cs
cs
cs
0.03
0.27
0.26
0.38
0.21
0.10
0.38
-0.03
0.46
4.99
5.01
4.37
4.71
5.00
4.30
4.09
5.32
4.49
-20.80
-21.20
-21.70
-25.00
-21.90
-17.90
-19.10
-29.90
-18.50
15.49
17.58
15.27
15.17
15.37
13.93
12.24
15.78
15.53
0.22
4.60
-20.82
15.18
Growth-Income Average
Axe-Houghton Stock Fund
David L. Babson Investment
Boston Company Capital
Appreciation Fund
Colonial Growth Shares
Country Capital Growth
The Dreyfus Fund Inc.
Fidelity Trend Fund
Keystone Growth Fund (K2)
Keystone Growth Common
Stock Fund (S3)
Lexington Research Fund
Penn Square Mutual Fund
g
g
g
cs
cs
cs
0.13
0.20
0.19
5.70
4.80
4.77
-34.90
-23.60
-22.20
20.48
19.26
18.45
g
g
g
g
g
g
cs
cs
cs
cs
cs
cs
0.09
0.18
0.20
0.09
0.04
0.21
5.36
4.86
4.42
5.71
5.32
6.69
-25.30
-18.80
-17.20
-30.20
-25.20
-27.80
17.07
14.88
16.62
20.48
18.42
24.40
g
g
flex
cs
0.20
0.31
5.08
4.78
-22.60
-19.00
13.87
15.55
This content downloaded from 14.139.157.28 on Sat, 11 Jul 2015 03:14:07 UTC
All use subject to JSTOR Terms and Conditions
440
Table 1-Continued
Wiesenberger
Classification
Mutual Funds by
1982 Name
Primary Investment
policy
objective
Standard
Mean Deviation
Minimum Maximum
Panel A: -Continued
g
g
g
g
g
cs
cs
cs
cs
cs
0.22
0.08
0.37
0.28
0.30
5.36
5.23
5.22
5.91
6.56
-20.60
-24.20
-22.70
-25.50
-27.80
17.55
19.69
17.09
18.07
19.07
cs
0.22
5.61
-27.30
19.24
g
g
g
g
cs
cs
cs
cs
0.20
0.09
0.34
0.49
4.73
5.36
6.49
4.76
-17.70
-22.80
-25.50
-17.20
19.24
20.46
22.21
18.09
0.21
5.37
-23.72
18.58
0.12
0.19
0.09
0.66
0.03
0.25
0.75
6.34
5.18
7.96
4.17
5.70
5.75
8.05
-31.60
-22.50
-34.47
-19.40
-27.30
-28.80
-28.80
19.17
11.98
21.43
13.26
15.33
16.97
25.30
7.76
-31.60
30.40
0.26
6.36
-28.06
19.23
0.45
0.07
0.40
0.82
5.89
4.93
4.64
4.29
-15.50
-18.58
-26.50
-23.90
26.25
22.49
14.04
11.62
0.43
4.94
-21.12
18.60
Overall Average
0.25
5.26
-23.78
18.24
Pilgrim Fund
Price (T. Rowe) Growth
Putnam Investors Fund
Security Equity Fund
Stein, Roe and F. Capital
Opportunities Fund
Stein, Roe and Farnham
Stock Fund
United Accumulative Fund
United Science and Energy
Value Line Fund
Windsor Fund
Growth Average
Financial Dynamics Fund
Founders Growth Fund
Keystone Speculative
Mutual Shares Corporation
Oppenheimer Fund
Scudder Special Fund
Twentieth Century
Growth Investors
Value Line Special
Situations Fund
mcg
mcg
mcg
mcg
mcg
mcg
mcg
cs
cs
cs
flex
cs
cs
cs
mcg
cs
g
g
g
g
spec
cs
c&i
cs
-0.01
This content downloaded from 14.139.157.28 on Sat, 11 Jul 2015 03:14:07 UTC
All use subject to JSTOR Terms and Conditions
441
Primary Investment
objective
policy
Mean
Standard
Deviation
0.60
0.29
0.22
4.74
0.25
-22.23
1.35
16.29
0.30
0.08
0.11
0.47
4.65
3.26
2.59
6.68
-22.12
-9.28
-8.97
-29.79
16.06
13.97
11.74
27.09
Minimum Maximum
Panel B
Benchmark Excess Returns
1-Month T-bill (Gross)
Value-Weighted CRSP
index
S&P 500
Government Bonds
Junk Bonds
Small cap
Wiesenberger Classifications
Primary Objective
Investment Policy
mcg
g
cs
bal
i
s
income
stability
bonds
c&i
tf
tax-free municipal
bond
flex
spec
ments for the index. The term spread is a constant-maturity 10-year Treasury
bond yield less the 3-month Treasury bill yield. The corporate bond defaultrelated yield spread is Moody's BAA-rated corporate bond yield less the AAArated corporate bond yield. The bond yields are the weekly average yields for
the previous month, as reported by Citibase.
The issue of data mining arises in connection with the information variables.
Data mining refers to the fact that many researchers in finance use the same
data, and a chance correlation of future returns with a predictor variable is
likely to be discovered as an "interesting" phenomenon. (See Lo and MacKinlay
(1990) and Foster and Smith (1992) for analyses of data mining biases in asset
pricing studies.) Extending the study of predictability to mutual funds is
interesting, as the fund returns represent a new data set. However, to the
extent that the assets held by funds are similar to those used in previous
studies of predictability, the correlation implies that a data mining bias may be
inherited by the funds. Also, the market index and factor returns that we use
have been prominent in the predictability literature.
This content downloaded from 14.139.157.28 on Sat, 11 Jul 2015 03:14:07 UTC
All use subject to JSTOR Terms and Conditions
442
We assume that the lagged variables are readily available, public information over our entire sample period, which starts in 1968. Most of the academic
studies of these variables appeared in the literature after 1968. Using our
approach, a manager who knew in 1968 that dividend yields, interest rate
levels, and yield spreads could be used to predict stock returns gets no credit
for using this knowledge before it was promoted in the academic literature.
However, similar variables were discussed as stock market indicators as early
as Dow (1920) and Graham (1965), which suggests that investors knew about
them long before they became prominent in the academic literature. Pesaran
and Timmermann (1995) cite a number of additional studies from the 1930s to
the early 1960s that emphasize stock market predictability based on interest
rates, dividend yields, and other cyclical indicators. In a model-selection experiment designed to avoid hindsight, they confirm the importance of predictable components in stock returns.
C. The Risk Factors
We use the value-weighted CRSP index for all stocks listed on the NYSE as
the market factor. We also examine a four-factor model. This model uses large
stocks, small stocks, government bonds, and low-grade corporate bonds. Related factor models are examined by Elton et al. (1992), who use three factors;
Sharpe (1988, 1991), who uses 10 to 12 factors; and the investment firm
BARRA, which uses as many as 68 factors in their model. We chose a relatively
parsimonious factor model because our application is mutual funds, as opposed
to individual common stocks. Also, our conditional model requires that we
estimate more parameters than an unconditional model, so parsimony becomes important.
In the four-factor model, the S&P 500 total return is used to represent large
market capitalization (cap) equities. The small cap index from Ibbotson Associates represents stocks whose market values correspond to the ninth and
tenth decile of market values on the NYSE. Starting in 1982, this small cap
index corresponds to the performance of Dimensional Fund Advisors' nine-ten
portfolio. The third factor is the return to a long-term (approximately 20-year)
U.S. Government bond from Ibbotson Associates. Finally, low-grade corporate
bond returns are based on the return series in Blume, Keim, and Patel (1991),
updated using the Merrill Lynch High Yield Composite Index return.12
D. Naive Strategies
If naive strategies appear to have abnormal performance, it implies that our
benchmarks are inefficient. This would call into question the measures of
performance for the managed portfolios. We therefore construct three strategies to provide a basis of comparison. Each naive strategy enters 1968 with an
initial set of weights: 65 percent large stocks, 13 percent small stocks, 20
12 The Blume, Keim, and Patel series is used for 1968:01-1990:01, and the Merrill Lynch series
is used for the last eleven months of 1990.
This content downloaded from 14.139.157.28 on Sat, 11 Jul 2015 03:14:07 UTC
All use subject to JSTOR Terms and Conditions
443
percent government bonds, and 2 percent low-grade bonds. The first strategy
is a monthly rebalancing strategy, for which these weights are held fixed over
the sample. The second is a buy-and-hold strategy, whose weights change over
time as the relative values of the four asset classes evolve. A buy-and-hold
strategy is passive, but its weights depend on relative values over time.13 Third
is an annual rebalancing strategy, which evolves as buy-and-hold unless the
month is a January, in which case the weights are reset to the initial weights.
13
+ R,t)/
Xx,t-l(l
+ RJ,t)1
We do not include a monthly rebalancingstrategy for the four-factormodel, because with constant
weights a regression of this strategy's returns on the factors producesa perfect fit by construction.
This content downloaded from 14.139.157.28 on Sat, 11 Jul 2015 03:14:07 UTC
All use subject to JSTOR Terms and Conditions
444
Table II
where rpt+1 is the excess return of a fund and rmt+1 is the excess return of the CRSP valueweighted market index. Heteroskedasticity-consistent t-ratios are shown as t(.). For the conditional CAPMmodels, the regressions are:
rpt+l
=ap
+ 6lprmt+l
+ 8'p(ztrmt+i)
(4)
+ Ept+j,
where zt is the vector of predeterminedinstruments, consisting of the dividend yield of the CRSP
index, a Treasury yield spread (long- minus short-termbonds), the yield on a short-termTreasury
bill, a corporate bond yield spread (low- minus high-grade bonds), and a dummy variable for
Januarys. Rsq are the R-squares of the regressions and pval(F) is the right-tail probabilityvalue
of the F-test for the marginal significance of the term including the predeterminedvariables. The
funds are groupedby their Wiesenbergertype as of 1982. Our "Special"categoryincludes Scudder
International and. Templeton Growth, which have sizable holdings in foreign equity markets,
Century Shares Trust, which holds primarily securities issued by financial firms, and Fidelity
Capital Trust, which merged into Fidelity Trend in 1982. The data are monthly from 1968-1990,
a total of 276 observations. The units are percentage per month.
Types
Conditional CAPM
Unconditional CAPM
Fund
ap
t(ap)
bp
t(bp) Rsq
ap
t(ap)
5lp
Rsq pval(F)
t(51p)
0.1980
-0.0307
0.64
0.93
1.03
1.18
22.7
37.6
34.2
25.5
0.64
0.93
1.04
1.19
26.6
42.4
39.6
30.5
0.727
0.895
0.863
0.784
0.078
0.066
0.058
0.070
0.928 0.81
15.6 0.619
0.2670
1.300 0.82
22.2 0.651
0.005
0.92
33.4 0.817
0.0220
0.124 0.93
38.4 0.831
0.060
-0.304
0.1980
-0.0331
1.93
-0.73
0.64 41.5
0.93 118.
1.03 76.5
1.18 35.4
0.151
0.008
0.004
0.001
0.81
29.7 0.837
0.2670
2.530 0.82
33.0 0.845
0.027
0.92
80.5 0.971
0.0186
0.421 0.93
90.5 0.974
0.001
445
14 We also construct Bayesian odds ratios to examine the importance of the lagged variables.
Using an uninformedprior, the odds ratios confirmthat the January dummy and default spread
variables are not important predictors, and that the other variables are jointly and individually
important.
This content downloaded from 14.139.157.28 on Sat, 11 Jul 2015 03:14:07 UTC
All use subject to JSTOR Terms and Conditions
446
Table III
CAPM
Minimum t-statistic
Bonferronip-value
-2.326
-1.960
-1.645
-1.282
<
<
<
<
0.0 <
1.282 <
1.645 <
1.960 <
t
t
t
t
t
t<
t<
t<
t<
t<
-2.326
-1.960
-1.645
-1.282
0.0
< 1.282
< 1.645
< 1.960
< 2.326
> 2.326
Maximum t-statistic
Bonferronip-value
Unconditional
Conditional
Unconditional
Conditional
-2.78
0.194
-2.71
0.238
-4.08
0.002
-3.80
0.006
5
4
4
4
26
5
1
3
5
20
8
4
5
4
25
3
3
4
3
25
15
1
3
1
4
18
5
4
2
4
13
0
4
2
2
19
0
3
2
2
3.89
0.004
4.90
0.000
3.80
0.006
6.40
0.000
t-statistic for the hypothesis that 50 percent of the alphas are positive is
-2.32.15
This content downloaded from 14.139.157.28 on Sat, 11 Jul 2015 03:14:07 UTC
All use subject to JSTOR Terms and Conditions
447
This content downloaded from 14.139.157.28 on Sat, 11 Jul 2015 03:14:07 UTC
All use subject to JSTOR Terms and Conditions
448
ate in view of the fat tails of the distributions and the assumption that betas
may vary with information in the conditional models. The restrictions are
tested using the standard GMMchi-square goodness-of-fit statistic.
The tests reject the hypothesis that the unconditional alphas are jointly zero
in the CAPM(and four-factormodel), the test producinga right-tail probability
value equal to 0.07 (0.02). This confirms our previous observation that there
are more negative unconditional alphas than would be attributed to chance.
The tests for zero conditional alphas also reject the joint hypothesis, producing
right-tail probability values of 0.01 (0.04). However, inspecting the coefficients
for the equally-weighted portfolios suggests that the rejection of zero alphas for
the conditional model is driven by the large positive alpha of the special group,
which includes the international funds. When we repeat the tests with only
four portfolios, excluding the special group, we find that no test can reject the
hypothesis that the alphas are jointly zero.
D. Market Timing: Results for the Naive Strategies
Table IV presents an initial analysis of the market timing models, using the
naive strategies. If the models are well-specified, the naive strategies should
not show evidence of abnormal performance. In the unconditional version of
the Treynor-Mazuymodel, the table shows that the buy-and-hold strategy has
a negative timing coefficient, with a t-statistic of -3.76, and a positive alpha,
with a t-statistic of 2.01. In the unconditional Merton-Henriksson model, the
buy-and-hold strategy has a negative timing coefficient, with a t-statistic of
-1.84, and a positive alpha, with a t-statistic of 1.86. These results are similar
to those of Jagannathan and Korajczyk(1986), who show that naive strategies
may exhibit option-like characteristics and hence have timing coefficients and
alphas with opposite signs. These results indicate a clear misspecification of
the unconditional market timing models.
In the conditional version of the Treynor-Mazuy model, none of the naive
strategies have significant alphas or timing coefficients, and the same is true
in the conditional Merton-Henriksson model. This is interesting, as it suggests
that conditional timing models can control misspecification in the unconditional models. The conditional models may therefore be more informative
about the performance of fund managers than the unconditional models.
E. Market Timing: Results for Mutual Funds Using the Treynor-Mazuy
Model
449
Table IV
Models
t(a)
Timing
Coefficient
t-Statistic
0.082
0.047
0.004
2.01
1.09
0.08
-0.004
-0.001
0.001
-3.76
-0.82
0.74
Conditional models
Buy-and-hold
Annual rebalance
Fixed weights
0.057
0.037
0.001
1.39
0.82
0.02
-0.001
0.000
0.001
-0.93
0.12
1.10
0.121
0.045
-0.021
1.86
0.72
-0.32
-0.065
-0.010
0.026
-1.84
-0.33
0.77
Conditional models
Buy-and-hold
Annual rebalance
Fixed weights
0.055
0.028
-0.011
1.03
0.50
-0.20
-0.010
0.009
0.028
-0.33
0.29
0.94
smaller than -1.90. A chi-square test across the five groups strongly rejects
the hypothesis that the coefficients are jointly zero (p-value less than 0.001).
Thus, the evidence is consistent with Grinblatt and Titman (1988) and Cumby
This content downloaded from 14.139.157.28 on Sat, 11 Jul 2015 03:14:07 UTC
All use subject to JSTOR Terms and Conditions
450
t-stat
Conditional
Timing
Coefficient
Unconditional
Alphas
t-stat
Coefficient
t-stat
Conditional
Alphas
Coefficient
t-stat
0.0005
-0.0014
-0.0023
-0.0096
0.086
-0.452
-0.496
-2.180
0.0039
0.0010
0.0009
-0.0018
1.16
0.54
0.40
-0.36
Special
-0.0103
-2.160
-0.0121
-2.23
Overall
-0.0028
-0.810
0.0003
0.31
0.015
-0.026
-0.025
0.137
0.216
-0.198
-0.291
0.848
-0.000
-0.041
-0.052
0.112
-0.019
-0.434
-0.545
0.726
0.432
2.090
0.454
2.310
0.036
0.211
0.018
0.002
0.0005
-0.0014
-0.0023
-0.0096
0.38
-2.33
-1.90
-2.70
0.0039
0.0010
0.0009
-0.0018
1.72
1.07
0.57
-0.44
Special
-0.0103
-4.37
-0.0121
-3.52
Overall
-0.0028
-2.47
0.0004
0.31
0.015
-0.026
-0.025
0.137
0.230
-0.703
-0.408
1.030
-0.000
-0.041
-0.052
0.112
-0.005
-1.140
-0.870
0.870
0.432
4.020
0.454
4.170
0.030
0.618
0.012
0.247
451
excepting for the special group of funds. We do not believe the results of the
unconditional models to be reliable. The evidence in the preceding section
shows that a negative timing coefficient may arise in an unconditional model,
even if the manager follows a buy-and-hold strategy, as the unconditional
model is misspecified.
Using our conditional version of the Treynor-Mazuy regressions, we find
very different results. Of the 67 point estimates of the conditional timing
coefficients for the individual funds, only 27 are negative. There are 13 with
t-ratios larger than two in absolute magnitude, of which only two are negative.
The t-statistic for the timing coefficient of an equally-weighted portfolio of the
funds is +0.31. There is no evidence of systematic nonzero alphas in these
models, excepting the special fund group. Overall, except for the special group,
incorporating conditioning information has essentially removed the findings of
negative timing coefficients in the unconditional model.
The special funds stand out as being different. This group has a strong
positive alpha and a significant negative timing coefficient, even in the conditional Treynor-Mazuy model. This may not be surprising in view of the fact
that the results for the special group are driven largely by the Templeton
Growth Fund and the Scudder International Fund. Glosten and Jagannathan
(1994) also find positive (unconditional) alphas for the Templeton Growth
Fund. The pattern of the coefficients is consistent with a portfolio strategy
similar to writing covered call options. However, these funds concentrate in
international investments, and studies such as Ferson and Harvey (1993) and
Schadt (1995) show that different factors are needed to explain international
equity returns. These results suggest that even the conditional Treynor-Mazuy
model is probably not appropriate for evaluating the special group.
F. Market Timing: Results for Mutual Funds Using the Merton-Henriksson
Model
Table VI summarizes results for the Merton-Henriksson (1981) model, using
equations (10) and (11). In the unconditional model, 46 of the 67 estimates of
the timing coefficients are negative, including all of the estimates for the
maximum gain funds. The binomial test of the hypothesis that 50 percent of
the coefficients are positive produces a t-statistic of -3.05. Of the seven
coefficients larger than two standard errors, six are negative. The coefficient
for an equally-weighted portfolio of the funds has a t-statistic of -2.38. The
chi-square test strongly rejects the hypothesis that the coefficients are jointly
zero (p-value = 0.008). Based on the unconditional model, the mutual funds'
market timing again appears to be of the "wrong" sign. This evidence is
consistent with the results of Chang and Lewellen (1984), Henriksson (1984),
and Glosten and Jagannathan (1994), and similar to our results for the
Treynor-Mazuy regressions.
The results for the conditional model are quite different. There are 25
negative point estimates for the 67 funds, which produces a t-statistic for the
hypothesis that 50 percent are negative, equal to +2.08. The chi-square test of
This content downloaded from 14.139.157.28 on Sat, 11 Jul 2015 03:14:07 UTC
All use subject to JSTOR Terms and Conditions
452
t(a)
Conditional Model
t(y)
t(a)
t(y)
0.0587
0.0261
0.0119
-0.0576
0.894
0.500
0.344
-0.496
-2.200
0.009
-0.001
0.031
0.391
0.114
0.003
0.010
1.560
0.0099
-0.0308
-0.0594
-0.2590
0.11
-0.32
-0.46
-1.97
Special
0.742
2.290
-0.2990
Overall
0.109
0.465
-0.0758
-0.006
-0.057
-0.053
0.179
-0.114
-0.508
-0.550
0.880
-1.77
0.641
2.780
-0.2610
-0.68
0.033
0.008
-0.0014
0.216
0.009
-0.001
0.031
0.391
0.106
-0.030
0.389
2.100
0.0099
-0.0308
-0.0594
-0.2590
0.22
-1.43
-1.59
-2.84
Special
0.742
Overall
0.102
-0.006
-0.057
-0.053
0.179
-0.065
-1.230
-0.669
1.050
5.470
-0.2990
-4.86
0.641
4.910
1.560
-0.0746
-2.38
0.025
0.397
0.0587
0.0261
0.0119
-0.0576
1.200
1.110
0.293
-0.626
-0.2610
-3.450
0.0013
0.039
the hypothesis that the coefficients are jointly zero produces a right-tail pvalue of 0.004. Of the seven coefficients more than two standard errors from
zero, only three are negative. The t-statistic for the equally-weighted portfolio
of the funds is +0.04. Based on the conditional model, and excepting the
This content downloaded from 14.139.157.28 on Sat, 11 Jul 2015 03:14:07 UTC
All use subject to JSTOR Terms and Conditions
453
special group, we no longer find the mutual funds' market timing to be of the
"wrong" sign. If anything, the timing coefficients reveal some weak evidence of
positive market timing ability, once we control for the predetermined information variables.
G. Conclusions About the Market-Timing Models
Table VII summarizes the cross-sectional distributions of the t-ratios associated with the key timing coefficients in the four market timing models. The
Bonferroni p-values indicate that the extreme t-ratios are significantly different from zero, except for the maximum values in the unconditional timing
models, but the tails of the distributions are thicker than the normal. In both
of the timing models the distribution of the t-ratios shifts to the right when the
conditioning information is introduced. The right tails of the distributions are
thicker and the left tails are thinner than in the unconditional models. It is
clear that introducing the conditioning information makes a dramatic impact
on the results of the timing models.
Both the Treynor-Mazuy and the Merton-Henriksson models are motivated
using strong assumptions about how mutual fund managers use any superior
information that they might have. When these strong assumptions fail, the
models will not separate timing and selectivity. As the assumptions are unlikely to be true, the models may be viewed as approximations to a more
complicated relation between the portfolio weights of the managers and the
future market return. In the unconditional versions of the models, the underlying asset betas are assumed to be constant. The Treynor-Mazuy model
approximates the relation between managers' weights and the future market
return by a linear function, while the Merton-Henriksson model uses an
indicator function (the weight is either zero or one, depending on the forecast
of the market return). In our conditional versiohs of the models, the assumptions are simple extensions of the assumptions in the original models.
While our extensions of these models are adequate to illustrate that the use
of conditioning information is important, we do not advocate using them to
evaluate managers in practice. For example, even the conditional timing
models are likely to be misspecified when applied to funds with sizable holdings of non-U.S. stocks. We believe that the development of more sophisticated
and realistic market timing models in the presence of conditioning information
is an important problem for future research.
H. Interpreting the Empirical Results
The evidence shows that the use of conditioning information makes the
performance of the funds in our sample look better. We can interpret these
results using a simple specification analysis. In Table II we found that the
unconditional betas are typically slightly smaller than the average conditional
betas. From equation (7), if the conditional alphas are larger than the unconditional alphas for a typical fund, it implies that the term BpCov(z; rm) is
negative for the typical fund. (We examine the sample values of this term and
This content downloaded from 14.139.157.28 on Sat, 11 Jul 2015 03:14:07 UTC
All use subject to JSTOR Terms and Conditions
454
Table VII
Treynor-Mazuy
Minimum t-statistic
Bonferroni p-value
t < -2.326
-2.326 < t < -1.960
-1.960 < t < -1.645
-1.645 < t < -1.282
-1.282 < t < 0.0
0.0
1.282
1.645
1.960
<
<
<
<
t
t
t
t
t
<
<
<
<
>
1.282
1.645
1.960
2.326
2.326
Maximum t-statistic
Bonferroni p-value
Unconditional
Conditional
Unconditional
Conditional
-5.98
0.000
6
2
4
12
20
-4.01
0.003
2
1
3
5
16
-4.15
0.002
5
2
4
2
33
-3.74
0.008
2
2
1
1
19
16
2
2
2
1
25
4
0
3
8
15
5
0
1
0
31
4
3
2
2
2.43
0.524
4.16
0.001
2.05
1.38
3.33
0.033
confirm that this is the case.) In other words, the component of the correlation
of mutual fund betas with the future market return that can be attributed to
the predetermined information tends to be negative. Since the future market
return can be written as rm = E(rmIZ) + E, and E is uncorrelated with Z, we can
This content downloaded from 14.139.157.28 on Sat, 11 Jul 2015 03:14:07 UTC
All use subject to JSTOR Terms and Conditions
455
This content downloaded from 14.139.157.28 on Sat, 11 Jul 2015 03:14:07 UTC
All use subject to JSTOR Terms and Conditions
456
conditional betas. The net new money for a group of mutual funds is defined by
Warther as the net sales (the dollar value of new shares sold, excluding
automatically-reinvested dividends), normalized by the lagged aggregate stock
market value. Warther finds that changes in net new money are negatively
related to changes in our estimated mutual fund betas. In monthly regressions
for 1976:03 to 1990:12, of beta changes on new money changes, the t-statistics
for the slope coefficients are between -2.8 and -3.9, depending on the fund
group. This is consistent with our conjecture that funds' betas are lower when
more new money flows into the funds.
Warther finds that, on average, funds invest about 62 cents of each new
dollar in the concurrent month, while 38 cents goes into cash. He also analyzes
the relation between new money inflows and the portfolio weight in cash, and
he finds a significant positive relation between the two. When inflows are
large, cash balances tend to increase. While Warther's analysis of our conditional betas is preliminary, the results are consistent with our conjecture that
fund betas change in response to larger cash positions associated with net
inflows of new money. A more comprehensive analysis of these relations should
be interesting. (We are grateful to Professor Warther for contributing this
analysis.)
457
yield spread than are the growth funds.18 In the four-factor models, the
coefficients measuring the sensitivity of the factor betas to the slope of the
term structure are significant for half of the income funds. However, the signs
of the coefficients differ across the funds, which suggests that different income
funds may adopt different strategies in response to changes in the term
structure. There is more variation of the beta response coefficients than of the
average exposures, within a fund group. This makes sense if managers within
a fund group adopt similar long run investment policies, but may use different
short run strategies. We believe that a more in-depth analysis of mutual funds'
conditional betas should be interesting.
Many of the response coefficientsfor the dividend yield (and 23 of the 25 significant ones) are
negative, which suggests that managers reduce their market exposure when equity yields are
high. The passive, buy-and-holdstrategy has a t-statistic of -2.4. Since the dividendyield goes up
when stock prices go down, the buy-and-holdstrategy may automatically reduce its stock market
exposure. Such a "priceeffect"can induce a negative response coefficient.The monthly rebalancing
strategy does not display a dividend yield effect.
19 We find that 40 of the 67 unconditional CAPMalphas are negative in the second subperiod
and all seven with absolute t-ratios larger than two are negative. Thirty-eight of the conditional
alphas are negative, but only five t-ratios are significant, and only two of these are negative. The
Bonferroni p-values for the hypothesis of a negative alpha are less than 0.10 for all of the
unconditionalmodels, with the single exception of the unconditional CAPMin the second subperiod (p-value = 0.102). Only 28 of the conditionalCAPMalphas in the first subperiodare negative,
five t-ratios are smaller than -2.0, and ten t-ratios are larger than +2.0.
This content downloaded from 14.139.157.28 on Sat, 11 Jul 2015 03:14:07 UTC
All use subject to JSTOR Terms and Conditions
458
period alphas use coefficients estimated with only the last 60 months of data.
We find that the mass of the distributions of both of these estimated alphas
shifts to the right when the conditioning information is introduced.
B. The Persistence of Performance
We calculate the cross-sectional correlations of the individual alphas in the
first and second halves of the sample. The correlation of the unconditional
CAPM alphas is 0.26 and the correlation of the alphas in the conditional CAPM
is 0.29, so this evidence of persistence in the alphas is similar in both models.
However, previous studies have found that persistence in performance may be
concentrated in the extreme-performance funds. We delete roughly the top and
bottom 10 percent (seven each) of the alphas in the first subperiod and
calculate the correlations on the remaining sample. The correlation of the
alphas in the conditional CAPM is only 0.14 when the tails are removed, while
using the unconditional CAPM the correlation on the subsample is 0.32. This
suggests that persistence concentrated in the extreme performers may be more
easily detected using conditional methods. We believe that future research
should use conditional models to further study the persistence in mutual fund
performance.20
This content downloaded from 14.139.157.28 on Sat, 11 Jul 2015 03:14:07 UTC
All use subject to JSTOR Terms and Conditions
459
these models are an improvement. Using the conditional market timing models for U.S. equity funds, the evidence of perverse market timing for the typical
fund is removed.
The relatively pessimistic results of the traditional measures is attributed to
common time-variation in the conditional betas and the expected market
return. WVhenthis predictability is ignored, fund managers as a group show
spurious inferior performance. This "inferior" performance is primarily due to
a negative covariance between mutual fund betas and the conditional expected
market return. When the common variation is controlled using lagged instruments, the conditional models make the performance of the funds in our
sample look better. Our results suggest that there is much more interesting
work to be done. Incorporating public information variables into the analysis
of investment performance is an important area for future research.
REFERENCES
Admati, Anat R., and Stephen A. Ross, 1985, Measuring investment performance in a rational
expectations equilibrium model, Journal of Business 58, 1-26.
Admati, Anat R., Sudipto Bhattacharya, Stephen A. Ross, and Paul Pfleiderer, 1986, On timing
and selectivity, Journal of Finance 41, 715-730.
Bollerslev, Tim, Robert F. Engle, and Jeffrey M. Wooldridge, 1988, A capital asset pricing model
with time-varying covariances, Journal of Political Economy 96, 116-131.
Blume, M., Donald B. Keim, and S. A. Patel, 1991, Returns and volatility of low-grade bonds,
1977-1989, Journal of Finance 46, 49-74.
Breen, William, L. R. Glosten, and Ravi Jagannathan, 1989, Economic significance of predictable
variations in stock index returns, Journal of Finance 44, 1177-1190.
Brown, Stephen, and William Goetzmann, 1995, Performance persistence, Journal of Finance, 50,
679-698.
Brown, Stephen, William Goetzmann, Roger Ibbotson, and Stephen A. Ross, 1992, Survivorship
bias in performance studies, Review of Financial Studies 5, 553-580.
Chan, K. C., and Nai-fu Chen, 1988, An unconditional asset-pricing test and the role of firm size
as an instrumental variable for risk, Journal of Finance 43, 309-325.
Chang, Eric C., and Wilbur G. Lewellen, 1984, Market timing and mutual fund performance,
Journal of Business 57, 73-96.
Chen, Zhiwu, and Peter Knez, 1992, A nonparametric evaluation approach to performance evaluation, Working paper, University of Wisconsin.
Chen, Zhiwu, and Peter Knez, 1994, Mutual fund performance: A nonparametric empirical
investigation, Working paper, University of Wisconsin.
Christopherson, Jon A., Wayne Ferson, and Debra A. Glassman, 1994, Conditioning manager
alphas on economic information: Another look at the persistence of performance, Working
paper, University of Washington, Seattle.
Cochrane, John, 1992, A cross-sectional test of a production-based asset pricing model, Working
paper, University of Chicago.
Cumby, Robert, and Jack Glen, 1990, Evaluating the performance of international mutual funds,
Journal of Finance 45, 497-521.
Dow, 1920, Scientific stock speculation, The Magazine of Wall Street, New York.
Dybvig, Philip H., and Ross, Stephen A. 1985, Differential information and performance measurement using a security market line, Journal of Finance 40, 383-399.
Elton, E., M. Gruber, S. Das, and M. Hlavka, 1992, Efficiency with costly information: A reinterpretation of evidence for managed portfolios, Review of Financial Studies 6, 1-22.
Evans, Martin D. D., 1994, Expected returns, time-varying risk and risk premia, Journal of
Finance 49, 655-680.
This content downloaded from 14.139.157.28 on Sat, 11 Jul 2015 03:14:07 UTC
All use subject to JSTOR Terms and Conditions
460
Fama, Eugene F., 1970, Efficient capital markets: A review of theory and empirical work, Journal
of Finance 25, 383-417.
Fama, Eugene F., and G. William Schwert, 1977, Asset returns and inflation, Journal of Financial
Economics 5, 115-146.
Ferson, W., and C. R. Harvey, 1991, The variation of economic risk premiums, Journal of Political
Economy 99, 385-415.
Ferson, W., and C. R. Harvey, 1993, The risk and predictability of international equity market
returns, Review of Financial Studies 6, 527-566.
Ferson, Wayne E., Shmuel A. Kandel, and Robert F. Stambaugh, 1987, Tests of asset pricing with
time-varying expected risk premiums and market betas, Journal of Finance 62, 201-220.
Ferson, W., and Robert A. Korajczyk, 1995, Do arbitrage pricing models explain the predictability
of stock returns? Journal of Business 68, 309-349.
Foster, Douglas, and Tom Smith, 1992, Assessing the goodness-of-fit of asset pricing models,
Working paper, Duke University.
Glosten, Lawrence, and Ravi Jagannathan, 1994, A contingent claims approach to performance
evaluation, Journal of Empirical Finance 1, 133-166.
Graham, Benjamin, 1965, The Intelligent Investor: A Book of Practical Counsel, 3rd ed. (New
York).
Grant, Dwight, 1977, Portfolio performance and the "cost" of timing decisions, Journal of Finance
32, 837-846.
Grinblatt, Mark, and Sheridan Titman, 1988, The evaluation of mutual fund performance: An
analysis of monthly returns, Working paper, University of California, Los Angeles.
Grinblatt, Mark, and Sheridan Titman, 1989a, Mutual fund performance: An analysis of quarterly
holdings, Journal of Business 62, 393-416.
Grinblatt, Mark, and Sheridan Titman, 1989b, Portfolio performance evaluation: Old issues and
new insights, Review of Financial Studies 2, 393-422.
Grinblatt, Mark, and Sheridan Titman, 1994, A study of monthly mutual fund returns and
performance evaluation techniques, Journal of Financial and Quantitative Analysis 29, 419444.
Hansen, Lars P., 1982, Large sample properties of the generalized method of moments estimators,
Econometrica 50, 1029-1054.
Hansen, Lars P., and Ravi Jagannathan, 1991, Implications of security market data for models of
dynamic economies, Journal of Political Economy 99, 225-262.
Hansen, Lars P., and Scott Richard, 1987, The role of conditioning information in deducing
testable restrictions implied by asset pricing models, Econometrica 50, 1269-1286.
Harvey, Campbell, 1989, Time varying conditional covariances in tests of asset pricing models,
Journal of Financial Economics 24, 289-317.
Henriksson, Roy D., 1984, Market timing and mutual fund performance: An empirical investigation, Journal of Business 57, 73-96.
Ippolito, R., 1989, Efficiency with costly information: A study of mutual fund performance 19651984, Quarterly Journal of Economics 104, 1-23.
Jagannathan, Ravi, and Robert A. Korajczyk, 1986, Assessing the market timing performance of
managed portfolios, Journal of Business 59, 217-236.
Jagannathan, Ravi, and Zhenyu Wang, 1996, The conditional CAPM and the cross-section of
expected returns, Journal of Finance 51, 3-53.
Jensen, Michael C., 1968, The performance of mutual funds in the period 1945-1964, Journal of
Finance 23, 389-416.
Jensen, Michael C., 1972, Optimal utilization of market forecasts and the evaluation of investment
performance, in G. P. Szego and K. Shell, Eds.: Mathematical Methods in Investment and
Finance. (North-Holland, Amsterdam).
Lehmann, Bruce N., and David M. Modest, 1987, Mutual fund performance evaluation: A comparison of benchmarks and benchmark comparisons, Journal of Finance 42, 233-266.
Lo, Andrew W., and A. Craig MacKinlay, 1990, Data snooping biases in tests of financial asset
pricing models, Review of Financial Studies 3, 431-468.
This content downloaded from 14.139.157.28 on Sat, 11 Jul 2015 03:14:07 UTC
All use subject to JSTOR Terms and Conditions
461
Longstaff, Francis, 1989, Temporal aggregation and the continuous-time CAPM, Journal of
Finance 44, 871-888.
Malkiel, Burton G., 1995, Returns from investing in equity mutual funds 1971-1991, Journal of
Finance 50, 549-572.
Mayers, David, and Edward Rice, 1979, Measuring portfolio performance and the empirical
content of asset pricing models, Journal of Financial Economics 7, 3-28.
Merton, Robert C., 1973, An intertemporal capital asset pricing model, Econometrica 41, 867-887.
Merton, Robert C., 1980, On estimating the expected return on the market: An exploratory
investigation, Journal of Financial Economics 8, 323-362.
Merton, Robert C., and Roy D. Henriksson, 1981, On market timing and investment performance
II: Statistical procedures for evaluating forecasting skills, Journal of Business 54, 513-534.
Newey, Whitney, and Kenneth D. West, 1987, A simple, heteroskedasticity and autocorrelation
consistent covariance matrix, Econometrica 55, 703-708.
Pesaran, M., and Allan Timmermann, 1995, Predictability of stock returns: Robustness and
economic significance, Journal of Finance 50, 1201-1228.
Roll, Richard, 1978, Ambiguity when performance is measured by the security market line,
Journal of Finance 33, 1051-1069.
Ross, S., 1976, The arbitrage theory of capital market asset pricing, Journal of Economic Theory
13, 341-360.
Schadt, Rudi, W., 1995, Testing international asset pricing models with mutual fund data,
Working paper, Southern Methodist University.
Shanken, Jay, 1990, Intertemporal asset pricing: An empirical investigation, Journal of Econometrics 45, 99-120.
Sharpe, William F., 1964, Capital asset prices: A theory of market equilibrium under conditions of
risk, Journal of Finance 19, 425-442.
Sharpe, William F., 1988, Determining a fund's effective asset mix, Investment Management
Review (December), 59-69.
Sharpe, William F., 1991, Asset allocation: Management style and performance measurement,
Working paper, Stanford University.
Stambaugh, Robert F., 1983, Arbitrage pricing with information, Journal of Financial Economics
12, 357-370.
Sirri, Eric, and Peter Tufano, 1992, Buying and selling mutual funds: Flows, performance, fees and
service, Working paper, Harvard University.
Treynor, Jack, and Kay Mazuy, 1966, Can mutual funds outguess the market? Harvard Business
Review 44, 131-136.
Warther, Vincent A., 1995, Aggregate mutual fund flows and security returns, Journal of Financial Economics 39, 209-236.
This content downloaded from 14.139.157.28 on Sat, 11 Jul 2015 03:14:07 UTC
All use subject to JSTOR Terms and Conditions