testing the capm
TRANSCRIPT
-
8/2/2019 Testing the CAPM
1/32
Testing the CAPM:
A Simple Alternative to Fama and MacBeth
(1973)
Paper Number: 06/04
Cherif Guermat, George Bulkley*, Mark C. Freeman* and RichardD.F. Harris*
Department of Economics and *Xfi Centre for Finance and Investment
University of Exeter
Author for CorrespondenceXfi Centre for Finance and Investment
University of ExeterExeter EX4 4STEngland
Tel: +44-(0) 1392-263155
Fax: +44-(0) 1392-262525
Email: [email protected]@exeter.ac.uk
R.D.F. [email protected]
I.G. [email protected]
-
8/2/2019 Testing the CAPM
2/32
Testing the CAPM:A Simple Alternative to Fama and MacBeth
(1973)
Abstract
Most studies that use the method of Fama and MacBeth (1973) to test
the Capital Asset Pricing Model (CAPM) are unable to reject the null hy-
pothesis that beta and expected returns are uncorrelated. This paper in-
troduces a simple new test which has the CAPM as its null hypothesis. We
show theoretically, by simulation and by using market data that our pro-
posed test has greater power than that of Fama and MacBeth. The new test
helps to establish whether the findings of previous studies result from the
low power of the Fama and MacBeth test or from a failure of the CAPM.
JEL classifi
cation: G10; G12; C12; C15Keywords: Capital Asset Pricing, Fama and MacBeth Test.
1. Introduction
Empirical tests of the Capital Asset Pricing Model (CAPM) have been
unable to find a statistically significant correlation between stock returns and
beta (see, for example, Reinganum, 1982; Lakonishok and Shapiro, 1986;
Ritter and Chopra, 1989; Fama and French, 1992). These tests typically
employ the method of Fama and MacBeth (1973, hereafter FM), which
involves estimating a series of monthly cross-section regressions of individual
stock returns on beta, and then testing whether the average slope coefficient
2
-
8/2/2019 Testing the CAPM
3/32
in these regressions is statistically different from zero using the time series
variation of the estimated slope coefficient in order to calculate the standard
error of the average slope coefficient.
Typically these empirical studies do not reject the null hypothesis of
the FM test that beta risk is not priced. A question raised by Chan and
Lakonishok (1993) is whether the null hypothesis is true or whether the
tests just lack the power to reject it in finite samples. They show that,
for realistic levels of market return volatility, the FM test is likely to have
low power in samples of the size typically employed in practice. The test
cannot reject the null that beta is not priced but it also cannot reject the
null of the CAPM assuming any plausible value for the expected return on
the market. A researcher who has strong prior beliefs in the CAPM would
not be compelled to infer from the point estimate and associated standard
error that the CAPM is false.
The low power of the FM test arises from the fact that, under both the
null and alternate hypotheses, a component of the slope coefficient in each
monthly cross-section regression is the realized excess market return . This
is very volatile, which yields a highly noisy estimated series of monthly slope
coefficients. It is this noise, which is common to both null and alternative
hypotheses, that is responsible for the low power of the FM test.
In this paper, we present a simple test in the spirit of Fama and Mac-
Beth (1973) that addresses this problem of low power. Our modified FM
test (hereafter Modified FM) involves subtracting the realized excess market
return each month from the estimated FM slope coefficient. The variance
of the test statistic is then significantly reduced and this results in a sub-
3
-
8/2/2019 Testing the CAPM
4/32
stantial increase in power. This is demonstrated analytically, by simulation
and by using real market data.
A natural consequence of subtracting the observed excess market return
from the estimated FM slope coefficient is that the CAPM becomes the null
hypothesis of the Modified FM test. This has a major advantage. Hav-
ing the null as the hypothesis that beta is unpriced means that rejecting
the null using the FM test is sufficient to establish that beta helps explain
cross-sectional differences in expected return. It is, though, not directly
informative about whether the price of beta risk is equal to the equity pre-
mium. In contrast, by having the CAPM as its null, finding a Modified FM
test statistic that is significantly different from zero is sufficient to reject the
model. The Modified FM test is therefore a more specific test of the CAPM
than FM.
There are other recent alternative methods to FM for testing the CAPM,
such as Hansens (1982) generalized method of moments (GMM) and the
semi-parametric method of Hodgson et al. (2002). GMM based models
are theoretically superior to FM as they relax both the normality and the
conditional homoscedasticity assumptions (Jagannathan and Wang, 2002).
However, GMM based methods do not generally lead to fully efficient es-
timates (Vorkink, 2003). In addition, Harvey and Zhou (1993) find little
difference between OLS and GMM based tests, while Ferson and Foerster
(1994) provide evidence suggesting that GMM methods lead to asset pricing
tests with aberrant properties. The semi-parametric method of Hodgson et
al. (2002) seems promising, but the evidence is scant and is based only on
a small-scale simulation study (Vorkink, 2003).
4
-
8/2/2019 Testing the CAPM
5/32
Thus, although these alternative methods may have more attractive the-
oretical properties, there is still a lack of empirical evidence in their favor.
From a practical point of view, both methods suffer from the fact that they
are difficult to implement, and that no standard software is available for car-
rying out estimation and testing using these methods. In contrast, both the
FM method and our modification to it combine simplicity with robustness
to cross-sectional correlation. They are also easily implemented using stan-
dard statistical and econometric software, which makes them potentially
attractive methods for practitioners. The simplicity of the FM method-
ology has resulted in its continued popularity. Among the recent studies
that have employed the FM methodology are Coval and Moskowitz (2001),
Lettau and Ludvingson (2001), Gomes, Kogan and Zhang (2003), Menzly,
Santos and Veronesi (2004), and Gompers and Metrick (2001). Variants of
the FM method are also used by Gompers, Ishii and Metrick (2003) and
Baker, Stein and Wurgler (2003).
The outline of the paper is as follows. In the following section, we present
the FM test in detail and discuss its small sample properties. In Section
3, we introduce the Modified FM test and compare it with the FM test.
Section 4 reports the results of simulation experiments to ascertain the size
and power of both the FM test and the Modified FM test. In section 5,
we run the FM and Modified FM tests on US stock market data. We test
the CAPM using the FM and Modified FM tests over several subperiods
of US market data since 1950. These confirm that the Modified FM test
statistic has lower standard error than the FM test statistic. Even with
this increased power, though, it is in many cases not possible to reject the
5
-
8/2/2019 Testing the CAPM
6/32
CAPM under the Modified FM test for beta ranked portfolios. Section 6
offers some concluding remarks.
2. The Fama-MacBeth Test
Suppose that there are N securities, with the return to asset i {1,...,N}between times t 1 and t being denoted by rit. The expected excess returnto any asset is given by
E[Rit] = i + i (1)
where Rit = rit rf,t1 is the excess return between period t1 and periodt for security i, rf,t1 is the one-period risk free rate at time t 1. i =Cov(Rit, Rmt)/V ar(Rmt) is the CAPM beta of security i which is assumed to
be constant across time and Rmt is the excess return on the market portfolio.
is the price of beta risk. Under the CAPM, = E[Rmt] while, if beta is
not priced, = 0. i incorporates the elements of the expected return that
are not captured by beta. This may include size effects, book-to-market
effects or additional Arbitrage Pricing Theory effects. It is assumed that
i does not vary across time. Again, if the CAPM is true, i = 0 for all i.
Now consider the realized excess return to each asset minus its expecta-
tion; it = Rit E[Rit]. This can, without loss of generality, be linearlyprojected against mt, the realized excess return to the market portfolio
minus its expectation:
it = proj(it|mt) + vit (2)
6
-
8/2/2019 Testing the CAPM
7/32
where the projection term is defined by
proj(it|mt) =E[itmt]
E[2mt]mt = imt (3)
i, defined above, is the true, or population, CAPM beta. For ease of
exposition, it will be assumed that this is always estimated without error.
It is easily established that E[vit] = E[vitmt] = 0.1 We will further impose
the restriction that E[vitvj] = 0 for all i, j including i = j and 6= t. This,
essentially, is restricting asset excess returns to display zero autocorrelation.
There is no requirement here that the disturbance terms are cross-sectionally
independent as they may, for example, within an Arbitrage Pricing Theory
context, include co-movement with other macroeconomic factors. This
general framework emphasizes that betas are informative about the ex-post
relationship between individual asset returns and the realized return to the
market portfolio. This, though, is not the central element of the CAPM,
which makes predictions about the ex-ante, pricing, effects of beta captured
by = E[Rmt] and i = 0. Equation 2 can be re-expressed as:
Rit = E[Rit] + i(Rmt E[Rmt]) + vit (4)
Consider the following cross-section regression between excess returns
and beta, which is commonly used in empirical studies of the CAPM:
Rit = at + ti + it (5)1 To establish these results (i) take expectations of equation 2 and (ii) multiply both
sides of this equation by mt and then take expectations. Here we are invoking a standard
asset pricing technique that is often used within a stochastic discount factor setting. See,
for example, Cochrane (2001, p.18).
7
-
8/2/2019 Testing the CAPM
8/32
The OLS estimate oft is given by
bt = Covi(Rit,i)V ari(i)
=Covi(E[Rit],i)
V ari(i)+ Rmt E[Rmt] + t
(6)
where t = S1
Pi(i )vit, S =
Pi(i )2 and = (1/N)
Pi i.
The notations Covi and V ari denote the sample covariance and variance,
respectively, evaluated over the cross-section dimension. By substituting in
for equation 1, it is clear that:
bt = + + Rmt E[Rmt] + t (7)where = S1
Pi(i)(i) and = (1/N)
Pi i. Empirical tests
of the CAPM can therefore be formulated as tests of hypotheses about E[t]
in regression (5). In particular, most tests of the CAPM involve testing the
null hypothesis that E[t] = 0, which is consistent with = 0, against the
alternative that E[t] 6= 0.2
If the CAPM holds, then E[t] = E[Rmt] and
the null hypothesis (E[t] = 0) should be rejected. The slope coefficient t
can be estimated straightforwardly by the usual OLS estimator, t, defined
above. However, the hypothesis that E[t] = 0 cannot be tested using a
conventional OLS t-statistic because the error term is cross-sectionally cor-
related under the null hypothesis. Consequently, the usual OLS estimate
of the standard error of t will be both biased and inconsistent.3 Indeed,
2 Notice that E(t) = Pt S1 Pi(i
)E[vit] = 0. In Fama-MacBeth type tests it is
also usually assumed that = 0.3 Both the pooled OLS regression and the Black, Jensen and Scholes (1972) pure cross
section regression suffer from this shortcoming. See Cochrane (2001) for a discussion.
8
-
8/2/2019 Testing the CAPM
9/32
Cochrane (2001) reports that the errors are so highly cross-sectionally cor-
related that standard errors can often be underestimated by a factor of ten
or more.
In a seminal paper, Fama and MacBeth (1973) propose a solution to this
problem that exploits the fact that, while the estimated standard error in
the regression will be biased, the estimated slope coefficient will be unbi-
ased. Fama and MacBeth propose repeatedly estimating the cross section
regression equation 5. The point estimate is the time-series mean of the
estimated slope coefficient in each of the cross-section regressions:
FM =1
T
TXt=1
t (8)
Under the null hypothesis that E[t] = 0, E[FM] = 0, while under the
alternative hypothesis that E[t] = E[Rmt], E[bFM] = E[Rmt]. In order totest the null hypothesis that E[t] = 0, Fama and MacBeth propose the test
statistic
tFM =FM
SE(FM)(9)
where SE(FM) = SD(t)/
T 1 and SD(t) is the sample standarddeviation of the T cross-section regressions estimates t. Under the null
hypothesis that E[t] = 0, the FM test statistic has a t-distribution with
T1 degrees of freedom. The FM approach has been used by many empiricalstudies, including Reinganum (1982), Lakonishok and Shapiro (1986), Ritter
and Chopra (1989), Fama and French (1992). These studies have invariably
9
-
8/2/2019 Testing the CAPM
10/32
found that the null hypothesis that E[t] = 0 cannot be rejected.4
Since the null hypothesis is that expected returns and beta are uncor-
related, the ability of the FM test to reject non-pricing depends upon the
power of the test to detect the correlation between beta and expected re-
turns. This in turn depends upon the standard error of FM, and hence,
from equation 8, on the variance of t. In the Appendix, we show that the
variance of t is given by
V ar(FM) =2mT
+2
T S+
T(10)
where 2m is the variance of Rmt and
= 2S2
Xi
Xj6=i
(i )(j )Cov(vit, vjt) (11)
. The power of the FM test therefore depends on the variance of the realized
market return, 2m. The higher the variance of2m, the higher the standard
deviation oft, the higher the standard error ofFM and the lower the power
of the FM test. Chan and Lakonishok (1993) show that, for realistic levels
of market return volatility, the FM test is likely to have low power in data
sets of the lengths typically employed in practice, and is therefore unlikely to
reject the null hypothesis that the average slope coefficient between returns
and beta is zero, even if the CAPM is true. We provide more evidence on
the small sample properties of the FM test in Section 4.
Note that since the OLS estimator t is unbiased, from equation 7:
4 Fama and MacBeth (1973) is one of the very few studies that found support for the
CAPM.
10
-
8/2/2019 Testing the CAPM
11/32
E[t] = E[t] = + (12)
The null hypothesis in the FM test, therefore, is whether + = 0,
whereas the alternative is that + 6= 0. Thus, the FM test does not
specifically test the CAPM model which makes three specific predictions;
= E[Rmt], = 0 and that there are no sources of expected return that
are orthogonal to beta. captures the effects of sources of risk that affect
expected returns and are correlated with beta. This might include, for
example, size effects. Usually the rejection of the Fama-MacBeth null is
seen to support the view that beta risk is priced. Again, though, this need
not be the case if 6= 0. The proposed Modified FM test focuses specifically
on testing the CAPM hypothesis.
3. The Modified Fama-MacBeth Test
The low power of the FM test comes from the fact that the slope coeffi-
cient in each monthly regression contains the realized excess market return.
The high volatility of the realized market return translates into a high volatil-
ity of the estimated monthly slope coefficient, and consequently into a high
standard error of the average estimated slope coefficient. In this section, we
propose a modification of the FM test that leads to a substantial increase
in power to discriminate between the CAPM and non-CAPM hypotheses.
To motivate the new test, note that from equation 7, the slope coefficient
in the FM test has a common component, Rmt, under any hypothesis on
the relationship between expected returns and beta (i.e. for any value of
). Our proposal, therefore, is to modify the FM test by first subtracting
11
-
8/2/2019 Testing the CAPM
12/32
the common component, Rmt, from the estimated slope coefficient, t, each
month. If the CAPM holds, = E[Rmt] and hence
t = t Rmt = + E[Rmt] + t (13)
Now consider the null hypothesis that the CAPM is true; = E[Rmt]
and = 0. In this case E[bt] = 0. Rejection of this null is sufficient toreject the CAPM. It will be rejected either because 6= E[Rmt] or because
6= 0.
The Modified FM test is given by
tMFM =MFM
SE(MFM)(14)
where MFM =1
T
Pt t, SE(MFM) = SD(bt)/T 1 and SD(bt) is the
sample standard deviation of
bt over the T cross-section regressions esti-
mates. Under the null hypothesis that the CAPM holds, the Modified FM
test statistic has a t-distribution with T 1 degrees of freedom. By sub-tracting the realized market return each month, the adjusted slope coefficient
estimate has lower variance but the same signal, and hence the Modified FM
statistic should have higher power to discriminate between the null and al-
ternative hypotheses. In the Appendix, we show that the variance of the
modified statistic, MFM, is equal to
V ar(MFM) =2
T S+
T(15)
Thus the variance of MFM is lower than the variance ofFM by 2m/T.
12
-
8/2/2019 Testing the CAPM
13/32
Equivalently, we can note that under their respective nulls:
SE2(bFM) SE2(bMFM) = 2mT (16)In the next section we evaluate the power of the modified FM test to
distinguish between CAPM and the alternative hypothesis that returns and
beta are uncorrelated using simulation experiments.
4. Simulation Experiments
In this section, we evaluate the small sample properties of both the
FM test and the Modified FM test using simulated data. In particular, we
generate data under two distinct hypotheses. Under the first hypothesis, the
CAPM holds, while under the second hypothesis, expected returns and beta
are uncorrelated. The two models used to generate the simulated data are
calibrated using realized monthly returns on 1000 stocks randomly drawn
from stocks in the CRSP database that had as least 24 observations over
the period January 1968 to December 2000.
For each randomly drawn stock i, we use the available time series data
from January 1968 to December 2000 to estimate i as the slope of the time
series regression of the monthly excess return of each stock on the stock
market excess return. The variance of the residuals from this regression 2i
is saved. The excess market return is the monthly CRSP market return
minus the US Treasury bill rate (IMF series USI60C).
For the first model, under which the CAPM holds, returns, Rit, are
generated as follows. At each date t, excess market returns, Rmt, are drawn
from a normal distribution with mean E(Rmt) and variance 2m. We use
13
-
8/2/2019 Testing the CAPM
14/32
E(Rmt) = 0.007 and m = 0.055. These values closely match the sample
moments of the CRSP data and are similar to those used in Jagannathan
and Wang (2002). In order to examine the impact of the volatility of the
market return on the small sample properties of the two tests, we also use
half the standard deviation (m = 0.0275). Four time-series sample sizes,
T, are considered (60, 120, 240, and 360 months). These time horizons are
commonly used in testing the CAPM (see, for example, Fama and French,
1992, and Jagannathan and Wang, 1996). We also consider some cases with
a large sample size (T = 1000).
Individual excess returns under the null hypothesis of the CAPM are
generated by
R1,it = iRmt + vit vit N(0, 2i ) (17)
where i is randomly drawn, with replacement, from the vector of 1000
estimates ofi, and it is randomly drawn from a normal distribution with
variance 2i .
For the second model, where beta risk is not priced, individual excess
returns are generated by
R2,it = E(Rmt) + i(Rmt E(Rmt)) + vit vit N(0, 2i ) (18)
The simulation exercise is based on 1000 replications. In each replica-
tion, equations 17 and 18 are used to generate individual stock return data.
In each replication, and for each of the return-generating models, j = 1, 2,
we estimate a series of cross-section regressions Rj,it = j,t + j,ti + ej,it,
14
-
8/2/2019 Testing the CAPM
15/32
t = 1, , T, to yield two series of estimated slope coefficients, 1,t, and
2,t, t = 1, , T. For each of the two models, j = 1, 2, the FM statistic
is the t-statistic of the sample mean of j,t, given by equation 9, while the
Modified FM test the t-statistic of the sample mean of j,t = j,t Rmt,given by equation 14. The FM statistic is then used to test the null hypoth-
esis that expected returns and beta are uncorrelated against the alternative
hypothesis that beta is priced, and the Modified FM test is used to test
the null hypothesis that the CAPM holds against the alternative hypothe-
sis that 6= E[Rmt]. In order to compute the size of the FM and Modified
FM tests, we calculate the proportion of times each test rejects its respective
null hypothesis, when the respective null hypothesis is true. To compute the
power of the two tests, we calculate the proportions of times each test rejects
its respective null hypothesis, when the respective alternative hypothesis is
true. The tests are conducted at the 5% and 10% two-sided significance
levels.
The simulation results for the FM and Modified FM tests are reported in
Table 1. For all three significance levels, the empirical size of both tests un-
der their respective null hypotheses is close to the nominal significance level.
However, the Modified FM test clearly has greater power to discriminate be-
tween the two hypotheses. Even at the smallest sample size, the power of
the Modified FM test is very close to unity regardless of the volatility of the
market return. For the FM test, however, the power is around 20 percent
for the smallest sample size when m = 0.055. Even when the variance of
the market return is halved, the power is still low, at most 56 percent at the
10% significance level. To achieve reasonable power, the FM test requires
15
-
8/2/2019 Testing the CAPM
16/32
at least 1000 months of data, and even then, the power is still marginally
lower than that of the Modified FM test.
[Insert Table 1 around here]
Some interesting properties of the two tests are shown in Table 2, which
reports the mean, standard deviation, and skewness and excess kurtosis
coefficients of the simulated FM and Modified FM statistics. The two
tests show little diff
erence under their respective null hypotheses. But thedifference is striking under their respective alternatives. The means of the
two statistics have the expected opposite sign, but the difference in their
variance is substantial. Owing to its reduced variance, the Modified FM
statistic is around four and a half times greater at all sample sizes. The
distribution of the Modified FM statistic under the alternative is to the left
of zero as suggested by its 5th and 95th percentiles. The FM statistic is
far closer to its value under the null hypothesis, with the 5th percentile
becoming negative at lower sample sizes.
[Insert Table 2 around here]
4.1. Robustness Check
In this subsection, in order to establish the robustness of our simula-
tion experiments, we consider how the FM and Modified FM tests perform
against some alternative data generating processes for returns. As an alter-
native to the CAPM, Jagannathan and Wang (2002) suggest a model of the
form
16
-
8/2/2019 Testing the CAPM
17/32
Rit = i + iRmt + vit (19)
where i = E(Rit) iE(Rmt). In relation to our general specification,this is the case where i +i is unconstrained. This model allows for other
firm-specific determinants of expected returns, such as market capitalization
and the ratio of book to market value of equity. In the simulation, i is
calibrated using the 1000 randomly drawn stocks from the CRSP database,
and is calculated as
i = Ri iRm (20)
Kan and Zhou (1999) suggest an alternative model that adds noise to
the market return. In particular, one of their suggestions is to use
Rit = ift + vit vit
N(0,2i ) (21)
where ft = (Rmt + nt).p
1 + 2n , nt is a zero mean measurement error
with finite variance 2n, and which is uncorrelated with vit. Since, E(ft) =
E(Rmt).p
1 + 2n , this is consistent with the CAPM. In the simulation we
use equation 21 to generate the data, using nt N(0, 0.01).Table 3 presents the rejection rates when the data are generated by the
Jagannathan and Wang (2002) and the Kan and Zhou (1999) models. For
a well specified test, the CAPM should be accepted for the Kan and Zhou
model and rejected for the Jagannathan and Wang model. Therefore the
FM (Modified FM) test should reject (fail to reject) the null for the Kan and
17
-
8/2/2019 Testing the CAPM
18/32
Zhou process. For the Jagannathan and Wang process the FM (Modified
FM) test should fail to reject (reject) the null.
[Insert Table 3 around here]
Panels A and B of the table show that the noisy CAPM of Kan and
Zhou is not identified by the FM test. The Modified FM rejects it with the
correct size for both levels of market volatility and both significance levels.
Panels C and D of the table show the rejection rate of the two tests whenthe data are generated by the Jagannathan and Wang noisy Non-CAPM
model. While the additional noise does affect the Modified FM test for very
small sample sizes, this effect virtually disappears once T = 120. On the
other hand, the FM test yields a surprising result. Contrary to expectation,
the rejection rate increases with the sample size. Typically, rejection rates
of the FM test should remain close to the nominal sizes of 5% and 10%.
Indeed, as can be seen from Table 1, the rejection rates for the Non-CAPM
data generated by equation 18 are all close to their respective nominal sizes,
irrespective of the sample size. What is different in the Jagannathan and
Wang model is that the Non-CAPM model contains additional noise. This
appears to distort the size of the FM test, especially for large values of T.
Thus, noise not only seems to worsen the power of the FM test when the
null hypothesis is false, but also seems to distort the size of the FM test
when the null is true. This suggests that in the presence of additional noise,
the FM test may not be able to detect the presence (or lack) of correlation
between beta and average returns. Thus, the FM test can potentially fail
completely under certain alternative models. In contrast, the Modified FM
18
-
8/2/2019 Testing the CAPM
19/32
test performs reasonably well in both cases examined here.
5. Testing the CAPM
In this section, we test the CAPM on US data using the FM and Modi-
fied FM tests. We randomly selected 9,000 stocks from the CRSP dataset
of 22,716 companies. Stocks with less than 60 consecutive monthly observa-
tions throughout the period January 1968 to December 2000 were removed
from the sample. This left slightly more thanfi
ve thousand stocks available,of which the first five thousand were selected for the test.
In the first stage, the estimated beta of each stock, bi, i = 1,..., 5000,was calculated. Following standard procedures, these were calculated using
a time series regression over all the available observations for each stock
Rit = i + iRmt + eit (22)
On the basis of these estimated individual stock betas, 100 beta-ranked
portfolios were formed with 50 stocks in each. Again, following standard
Fama-MacBeth procedure, the beta of each portfolio was re-estimated using
the 60 months of data prior to the test period in order to remove potential
measurement error bias. This provided estimates of portfolio betas bp,p = 1,..., 100. Then t was estimated from the cross section regression
Rpt = at + tp + vpt (23)
The FM and Modified FM tests are then simply t-tests on the sample
mean of the time series t and (t Rmt), respectively. The results, shown
19
-
8/2/2019 Testing the CAPM
20/32
in Table 4, contrast the conclusions of the two tests for a selection of starting
dates and sample sizes, T.
[Insert Table 4 around here]
As expected, the Modified FM test statistic has a substantially lower
standard error than the FM test, potentially giving it much greater power.
According to equation 16:
2m = ThSE2(bFM) SE2(bMFM)i (24)Using this equation to estimate 2m from the standard errors in Table 4,
the average estimated annualized standard deviation of the market is 14.6%
with a minimum and maximum estimate of 11.6% and 16.9% respectively,
which is broadly consistent with observed returns.
Despite the increase in power, the Modified FM test statistic is statis-
tically significant at the 10% level only twice. From Table 1, it has been
established that the null hypothesis that this test statistic is equal to zero is
rejected almost 100% of the time if beta is unpriced and if market returns
in the test period are normally distributed with a mean of 0.7% per month.
These results do, then, give some support to the view that beta can help
explain cross-sectional variations in average returns. This is backed by the
FM test statistics, which are always positive, of similar value to the esti-
mated equity premium and where, despite the tests low power, the p-value
is less than 10% in three out of the nine cases.
20
-
8/2/2019 Testing the CAPM
21/32
6. Conclusion
It is known from a large number of studies in empirical asset pricing that
it is very difficult to reject the null hypothesis that the correlation between
beta and expected returns is zero. This has led some to infer that the
CAPM beta is not priced. However, others have noted that, given the low
power of the Fama and MacBeth test and the short test periods employed,
the inability to reject the null would be unsurprising even if the CAPM held.
In this paper, we note that the key factor that lies behind the low power
of the standard FM test is the volatility of realized market returns. The
ex-post excess return to the market is a common component of the test sta-
tistic under both the null and alternate hypotheses. We therefore construct
an alternate test, the Modified FM, where the realized return to the market
in each period is removed from the FM statistic. It is shown theoretically,
by simulation and through an empirical study using market data that the
Modified FM has greater power than the FM statistic. In principle, there-
fore, the Modified FM is better able to distinguish between the CAPM and
alternative hypotheses.
In the final section of the paper, we have conducted FM and Modified
FM tests on various samples of US data. Using beta ranked portfolios, the
decreased standard errors of the Modified FM test is clearly demonstrated
and is of the expected order of magnitude. It is interesting to see in theseresults that, when the modified test with substantially enhanced power is
applied, in most cases it is not possible to reject the null hypothesis that
the CAPM is true.
21
-
8/2/2019 Testing the CAPM
22/32
Appendix
As in equation 4, the data generating process is given by:
Rit = E[Rit] + i (Rmt E[Rmt]) + vit (25)
where Rmt = E[Rmt] + ut, ut iid(0,2m), E(vit) = 0, E(vitvjs) = 0 fort 6= s, i,j. We also assume V ar(vit) = 2v, and E(utvis) = 0, i,t,s.
The OLS estimator oft, as in equation 7, is given by
t = + + Rmt E[Rmt] + t (26)
The variance of t is given by
V ar(t) = V ar(Rmt) + V ar(t) + 2Cov(Rmt, t) (27)
Consider the covariance term. Since E[t] = 0 this can be re-expressedas E[Rmtt]. Substituting in for t this is equal to E[RmtS
1
Pi(i)vit].
Rearranging gives S1P
i(i )E[Rmtvit] = 0 from the properties of vitgiven above. It follows that
V ar(t) = V ar(Rmt) + V ar(t) (28)
As asset excess returns are assumed to be zero autocorrelated, Cov(t, s) =
0, t 6= s, the variance of FM is given by
V ar(FM) =1
TV ar(t) (29)
We have V ar(Rmt) = 2m and, assuming i fixed,
22
-
8/2/2019 Testing the CAPM
23/32
V ar(t) = V ar
1
S
Xi
(i )vit!
=
=1
S2V ar
Xi
(i )vit!
=2vS
+1
S2
Xi
Xj6=i
(i )(j ) Cov(vit, vjt)
=2vS
+
=2vP
i(i )2+
where = S2P
i
Pj6=i(i )(j ) Cov(vit, vjt). Hence
V ar(FM) =2mT
+2v
T S+
T(30)
The estimated variance of FM is defined as
Var(FM) = 1T2FM (31)
where 2FM =1
T1
Pt(t FM)2.
To compute E[Var(FM)] rewrite t and FM using (26)
t FM = Rmt Rm + t (32)
= ut u + 1S
Xi
(i )(vit vi). (33)
Note that E[2FM] = 1T1Pt E[(t FM)2] = 1T1Pt V ar(t FM),
23
-
8/2/2019 Testing the CAPM
24/32
since E[t
FM] = 0. Hence,
V ar(t FM) = E
(ut u)2
+1
S2
Xi
(i )2E
(vit vi)2
+1
S2
Xi
Xi6=j
(i )(j ) E[(vit vi)(vjt vj)]
=
1 1
T
2m +
1 1
T
1
S2
Xi
(i )22v
+
1 1
T
1
S2Xi Xi6=j(i )(j ) Cov(vit, vjt)
=T 1
T
2m +
1
S2v + .
Therefore,
E[2FM] =1
T 1T 1
T
2m +
1
S2v + T =
2
m +1
S2v +
and
E[Var(FM)] =2mT
+2v
T S+
T. (34)
Thus, the sample variance of the FM estimator is an unbiased estimator of
its true variance.
Now turn to the variance of the modified FM statistic, which is given by
MFM =1
T
Xt
t =1
T
Xt
(t Rmt) (35)
Using (26), t is given by
t = + E[Rmt] + t (36)
Repeating the same steps used above, it follows trivially that the true
variance of MFM is given by
24
-
8/2/2019 Testing the CAPM
25/32
V ar(MFM) =2v
T S+
T
and that the sample variance is unbiased, E[Var(MFM)] = V ar(MFM).
25
-
8/2/2019 Testing the CAPM
26/32
References
Baker M., Stein J.C. and Wurgler J. (2003) When does the market mat-
ter? Stock prices and the investment of equity-dependent firms, Quarterly
Journal of Economics, 118, 969-1005.
Chan L. and Lakonishok J. (1993) Are the reports of betas death prema-
ture?, Journal of Portfolio Management, 19, 5162.
Cochrane J.H. (2001), Asset Pricing, Princeton University Press.
Coval J.D. and Moskowitz T.J. (2001) The geography of investment: in-
formed trading and asset prices, Journal of Political Economy, 109, 811-841.
Fama E. F. and French K. R. (1992) The cross-section of expected stock
returns, Journal of Finance, 47, 427465.
Fama E. F. and French K. R. (1993) Common Risk Factors in the Returns
on Stocks and Bonds, Journal of Financial Economics, 33, 356.
Fama E. F. and MacBeth J.D. Risk, Return and Equilibrium: Empirical
Tests, Journal of Political Economy, 81, 607636.
Hansen L. (1982) Large Sample Properties of Generalized Method of Mo-
ments Estimators, Econometrica, 50, 1029-1054.
Ferson W. and Foerster S. (1994) Finite Sample Properties of the Gener-
alized Method of Moments in Tests of Conditional Asset Pricing Models,
Journal of Financial Economics, 36, 29-55.
Gomes J., Kogan L. and Zhang L. (2003) Equilibrium Cross Section of
Returns, Journal of Political Economy, 111, 693-732.
26
-
8/2/2019 Testing the CAPM
27/32
Gompers P.A., Ishii J. and Metrick A. (2003), Corporate governance and
equity prices, Quarterly Journal of Economics, 118, 107-155.
Gompers P.A. and Metrick A. (2001), Institutional investors and equity
prices, Quarterly Journal of Economics, 116, 229-259.
Harvey C. and Zhou G. (1993) International Asset Pricing with Alternative
Distributional Specifications, Journal of Empirical Finance, 1, 107-131.
Hodgson D. J., Linton O. and Vorkink K. (2002), Testing the capital as-
set pricing model efficiently under elliptical symmetry: a semiparametric
approach, Journal of Applied Econometrics, 17, 617639.
Jagannathan R. and Wang Z. (2002) Empirical evaluation of asset pricing
models: a comparison of the SDF and beta methods, Journal of Finance,
57, 23372367.
Kan R. and G. Zhou (1999) A Critique of the Stochastic Discount Factor
Methodology, Journal of Finance, 54, 1221-1248.
Lakonishok J. and Shapiro A.C. (1986) Systematic Risk, Total Risk and
Size as Determinants of Stock Market Returns, Journal of Banking and
Finance, 10, 115-132.
Lettau M. and Ludvingson S. (2001) Resurrecting the (C)CAPM: A Cross-
Sectional Test When Risk Premia Are Time-Varying, Journal of Political
Economy, 109, 1238-1287.
Menzly L., Santos T. and Veronesi P. (2004) Understanding Predictability,
Journal of Political Economy, 112, 1-47.
27
-
8/2/2019 Testing the CAPM
28/32
Reinganum M.R. (1982) A Direct Test of Rolls Conjecture on the Firm
Size Effect, Journal of Finance, 37, 27-35.
Ritter J.R. and Chopra N. (1989) Portfolio Rebalancing and the Turn-Of-
The-Year Effect, Journal of Finance, 44, 149-166.
Vorkink K. (2003), Return Distributions and Improved Tests of Asset Pric-
ing Models, Review of Financial Studies, 16, 845-874.
28
-
8/2/2019 Testing the CAPM
29/32
29
Table 1. Simulation Results for the Fama-MacBeth and Modified FM tests.
Fama-MacBeth Modified FM Test
T 5% 10% 5% 10%
Panel A: Size of tests (m=0.0275)60 0.043 0.090 0.057 0.111
120 0.053 0.106 0.042 0.086
240 0.044 0.086 0.054 0.102
360 0.049 0.112 0.053 0.098
Panel B: Size of tests (m=0.055)
60 0.047 0.100 0.045 0.100
120 0.049 0.097 0.049 0.093
240 0.055 0.093 0.049 0.092
360 0.036 0.082 0.058 0.105
1000 0.051 0.091 0.044 0.097
Panel C: Power of tests (
m=0.0275)60 0.404 0.556 0.988 0.998
120 0.711 0.806 1.000 1.000
240 0.958 0.981 1.000 1.000
360 0.991 0.996 1.000 1.000
Panel D: Power of tests (m=0.055)
60 0.162 0.244 0.987 0.996
120 0.266 0.380 1.000 1.000
240 0.498 0.611 1.000 1.000
360 0.651 0.762 1.000 1.000
1000 0.971 0.987 1.000 1.000
-
8/2/2019 Testing the CAPM
30/32
30
Table 2. Sample Properties of Simulated Statistics (m=0.055).SampleMean
SampleVariance
5thPercentile
95thPercentile Skewness
ExcessKurtosis
JB*Statistic
p-value(JB)
T=60
Fama-MacBeth (H0) 0.020 1.031 -1.647 1.681 -0.095 0.071 1.706 0.426
Modified FM (H0) -0.031 1.017 -1.718 1.644 -0.040 -0.109 0.763 0.683
Fama-MacBeth (HA) 0.995 1.031 -0.666 2.696 -0.053 0.021 0.492 0.782
Modified FM (HA) -4.377 1.216 -6.227 -2.630 -0.200 0.111 7.192 0.027
T=120
Fama-MacBeth (H0) -0.012 1.014 -1.661 1.625 0.015 0.084 0.328 0.849
Modified FM (H0) -0.040 1.010 -1.614 1.650 0.109 -0.050 2.067 0.356
Fama-MacBeth (HA) 1.356 1.028 -0.276 2.987 0.052 0.100 0.865 0.649
Modified FM (HA) -6.155 1.195 -7.925 -4.357 -0.099 -0.003 1.625 0.444
T=240
Fama-MacBeth (H0) 0.016 0.984 -1.669 1.527 -0.163 -0.067 4.621 0.099
Modified FM (H0) -0.055 0.973 -1.644 1.566 -0.033 0.084 0.474 0.789
Fama-MacBeth (HA) 1.938 0.990 0.224 3.472 -0.144 -0.086 3.741 0.154
Modified FM (HA) -8.679 1.135 -10.519 -6.998 -0.163 0.185 5.859 0.053
T=360
Fama-MacBeth (H0) -0.027 0.938 -1.554 1.583 0.146 0.412 10.639 0.005
Modified FM (H0) -0.013 1.032 -1.686 1.640 -0.005 0.357 5.305 0.070
Fama-MacBeth (HA) 2.334 0.938 0.821 3.916 0.159 0.437 12.169 0.002
Modified FM (HA) -10.557 1.166 -12.383 -8.781 -0.031 0.175 1.437 0.487
T=1000
Fama-MacBeth (H0) -0.070 0.968 -1.607 1.513 0.040 0.048 0.361 0.835Modified FM (H0) 0.018 0.993 -1.595 1.684 0.026 -0.116 0.673 0.714
Fama-MacBeth (HA) 3.859 0.973 2.339 5.478 0.043 0.047 0.407 0.816
Modified FM (HA) -17.534 1.190 -19.338 -15.776 -0.076 -0.131 1.670 0.434
-
8/2/2019 Testing the CAPM
31/32
31
Table 3. Simulation Results for the FM and Modified FM tests.
Fama-MacBeth Modified FM Test
T 5% 10% 5% 10%
Panel A: Rejection rate (Kan&Zhou, CAPM) (
m=0.0275)60 0.072 0.140 0.042 0.090
120 0.093 0.173 0.041 0.087
240 0.164 0.256 0.048 0.096
360 0.243 0.365 0.050 0.096
Panel B: Rejection rate (Kan&Zhou, CAPM) (m=0.055)
60 0.073 0.137 0.058 0.105
120 0.108 0.166 0.049 0.096
240 0.167 0.256 0.051 0.109
360 0.183 0.287 0.038 0.092
1000 0.497 0.616 0.056 0.103
Panel C: Rejection rate (Jag&Wang, Non-CAPM) (m=0.0275)
60 0.134 0.220 0.593 0.726
120 0.223 0.335 0.871 0.925
240 0.409 0.552 0.992 0.998
360 0.566 0.681 1.000 1.000
Panel D: : Rejection rate (Jag&Wang, Non-CAPM) (m=0.055)
60 0.094 0.154 0.580 0.700
120 0.102 0.165 0.876 0.933
240 0.141 0.221 0.995 0.998
360 0.202 0.306 1.000 1.000
1000 0.459 0.610 1.000 1.000
-
8/2/2019 Testing the CAPM
32/32
Table 4. Estimation Results for the FM and Modified FM using US market data
FM test Modified FM test
T Sample St. Sample St.
(months) Mean Error Mean Error
Start (%) (%) t-statistic p-value (%) (%) t-statistic p-value
Jan-73 120 0.553 0.535 1.035 0.303 0.527 0.298 1.769 0.08
240 0.497 0.351 1.415 0.158 0.13 0.193 0.671 0.503
336 0.583 0.305 1.915 0.056 0.058 0.183 0.319 0.75
Jan-78 120 0.577 0.451 1.278 0.204 0.026 0.235 0.112 0.911
240 0.512 0.297 1.728 0.085 -0.234 0.168 -1.392 0.165
Jan-83 120 0.440 0.457 0.964 0.337 -0.267 0.243 -1.099 0.274
216 0.600 0.370 1.621 0.107 -0.202 0.230 -0.875 0.383
Jan-88 120 0.448 0.387 1.158 0.249 -0.494 0.239 -2.072 0.04
156 0.771 0.449 1.718 0.088 -0.09 0.294 -0.306 0.76