Está en la página 1de 16

JOURNAL OF FINANCIAL AND OUANTITATIVE ANALYSIS

VOL. 29, NO. 1, MARCH 1994

Tests of Conditional Asset Pricing with Time-Varying Moments and Risk Prices
Harry Turtle, Adolf Buse, and Bob Korkie*

Abstract
This paper tests conditional capital asset pricing models in a multivariate GARCH framework employing both constant and time-varying prices of market and bond risk. Depending on the interpretation of the market portfolio, the ICAPM with one hedge portfolio or the two-factor myopic CAPM are supported using weekly data from July 1983 to December 1989. In contrast, we reject the myopic single-factor CAPM under a constant price of market risk. We find that interest rate risk is highly significant, which suggests that previous rejections of the conditional CAPM using only stock market data may be due to omitted hedge terms or an incomplete market factor.

I.

Introduction

The predictability of asset retums has led to a large body of literature that examines time variability in expected retums and variances. Support for time variability in expected retums has been foutid in the works of Fama and French (1989), Ferson (1989), Fama (1990), and Harvey (1991), for example. Predictability of conditional variances has been studied using generalized autoregressive conditional heteroskedastic (GARCH) processes.' Schwert (1989) provides strong evidence of the predictability of conditional variance and, subsequently, Schwert and Seguin (1990) conclude that portfolio research should incorporate predictable volatility changes. In this paper, we use the word "conditional" to imply that the investment opportunity set shifts over time and that conditional moments are employed. Because all asset pricing models considered here are conditional, the adjective is generally omitted. We define a "myopic" CAPM, or simply a CAPM, to be those conditional
* Turtle. Faculty of Management, University of Manitoba, Winnipeg, Manitoba, R3T 5V4 Canada; Buse, Department of Economics, and Korkie, Faculty of Business, University of Alberta, Edmonton, Alberta, T6G 2R6 Canada, respectively. The authors thank Steve Beveridge, Elias Shiu, Prem Talwar, and JFQA Referee and Associate Editor Wayne Ferson for their comments. Provision of some FORTRAN code by Robert Stambaugh is acknowledged. Financial support was provided to Turtle by the SSHRC and to Korkie by the A. F. Collins professorship. Earlier versions of the paper were presented at the 1991 Northern Finance Association Meetings, the 1992 French Finance Meetings, and the Universities of Alberta, Manitoba, and Waterloo. 'Bollerslev, Chou, and Kroner (1992) provide an extensive survey of the use of GARCH models in finance.

15

16

Journal of Financial and Quantitative Analysis

asset prieing models that do not eontain portfolios hedgitig shifts iti the opportunity set. ConditionaJ asset pricing models that contain hedge portfolios are called an ICAPM. Both models may contain one or more factor portfolios, which, in the ICAPM case, are interpreted as hedge portfolios or, in the CAPM case, are called completing market factors. Tests of asset pricing models differ widely in their assumptions about the evolution of retums. The finance literature has modeled changes in conditional means, often in models that admit unspecified heteroskedasticity. Shanken (1990) rejects a two-factor model, while Ferson (1990) examines a multifactor asset pricing model and finds support for two to three unspecified time-varying risk premia.^ Subsequently, Ferson and Harvey (1991) do not reject the hypothesis that the predictable components of retum are due to five time-varying betas and risk premia. GARCH models permit the explicit specification of conditional asset pricing relations in the context of time-varying second moments. The GARCH methodology models variances as deterministic functions of previous periods' squared disturbances and conditional variances. Although this allows a myriad of factors to affect conditional variances, it suggests that the infomiation filtration is generated only by previous retums. The GARCH process is intended to "fit" variances, rather than to strictly model variances in an economic sense. Multivariate GARCH models have provided strong evidence against the myopic CAPM in numerous contexts (c.f., Bodurtha and Mark (1991) and Ng (1991), who use a U.S.-based equity index as a market proxy; Bollerslev, Engle, and Wooldridge (1988), who employ govemment bills and common stocks as a market proxy; or Giovannini and Jorion (1989), who measure the market using an international index). Ng (1991) provides support for the CAPM for beta-sorted portfolios, but rejects the CAPM using size-sorted portfolios. The rejection of the myopic CAPM, using a wide variety of market proxies, is consistent with the ICAPM hypothesis that investors hedge against intertemporal shifts in the investment opportunity set, unless restrictive assumptions are met.^ Engle, Ng, and Rothschild (1990) and Ng, Engle, and Rothschild (1992) examine a multifactor model by generating explanatory factors and subsequently using these factors in asymptotic tests. In the 1992 paper, the value-weighted equity index is rejected as a sufficient factor to explain the dynamic behavior of 10 size portfolios. The authors conclude that, among three zero-beta market factors explaining excess retum dynamics, one factor is related to long-term bond risk. No explicit allowance is made in either paper for the statistical distortions caused by generated regressors. In this paper, we estimate a conditional asset pricing model with time-varying moments using a multivariate G ARCH-M model. Two specifications are employed to examine the sensitivity of asset pricing relations to time-vary ing market and bond
^The Shanken study rejects a two-factor CAPM under the null hypothesis of a three-factor CAPM; therefore, the rejeciion of the null hypothesis restriction is to be expected because two factors will not span the same space as three factors. It is difficult to say what would result in the Shanken study if another hedge portfolio were included or one of the state variables were omitted in order to make the test compatible with the null hypothesis. Respectively, it is possible that neither the three-factor nor the two-factor CAPM would be rejected. Potentially then, the Shanken results are consistent with the results of this study and other studies. ^See Fama (1970), Merton (1971), (1973), or Rubinstein (1976) for a discussion of the restrictions.

Turtle, Buse, and Korkie

17

risk prices. First, a model that restricts the prices of market and bond risk to be constant but allows variation in moments over time is estimated. Next, the model is respecified to permit both risk prices and moments to time vary. In both cases, interest rate risk is highly significant and the ICA.PM is not rejected. The definition of the market proxy is critically important in defining the linear restrictions that exist under a true null hypothesis. The importance arises because the market portfolio is mean-variance efficient in a CAPM but it is not efficient in an ICAPM. In fact, the ICAPM market portfolio is a combination of the CAPM efficient market portfolio and inefficient hedge portfolios. If the ICAPM is correct, there is no observable portfolio with weights equal to the equilibrium weights of the CAPM market portfolio that can be used for testing purposes. Without loss of generality, most tests use the ICAPM market portfolio and the hypothesized hedge portfolios because they span the same space. If there is measurement error in the market portfolio, the use of a proxy for the true market portfolio makes it difficult to distinguish the CAPM from the ICAPM . For example, in a CAPM market, an equity index may be mean-variance inefficient due to mismeasurement of the market portfolio and, hence, a long-term bond portfolio might be a market completing factor instead of a hedge portfolio. Therefore, the results of any tests must be interpreted with care because measurement error may prevent a clear distinction between the myopic CAPM and the ICAPM. The CAPM results of Ferson, Kandel, and Stambaugh (1987), with timevarying risk premia and market betas, illustrate the preceding discussion. While an unspecified single-risk premium model is not rejected, the conditional meanvariance efficiency of a value-weighted stock index is rejected. That is, regardless of the true model, any /T-factor model can be collapsed to an unspecified one-factor model that may not be rejected. For ease of interpretation, we consider the market proxy portfolio as exogenous and the null hypothesis tests the joint specification of the market index portfolio and the ICAPM. The remainder of the paper is organized as follows. In Section II, the models are developed. The data and estimation methodology are presetted in Section III. Section IV has the primary empirical results of the paper. In Section V, diagnostic analysis is performed to verify the integrity of the statistical assumptions as well as to ascertain the success of the model in capturing moment dynamics. Concluding comments are offered in Section VI.

II.

The Models

We examine asset retums based upon the ICAPM that posits an additional risk premium over the myopic Sharpe (1964)-Lintner (1965) CAPM, hedging shifts in the investment opportunity set.** Using Merton ((1973), Eq. 34), (1) aj, - Rf, = pjMt [aM, - Rft) + PjBt {otB, - Rft), j = \,2,...,N,

where Oy, is assety's conditional expected rate of retum at time /, Rf, is the riskless rate.
'For other examples of conditional models that introduce terms other than the market, see Long (1974), Breeden (1979), (1986), and Cox, IngersoU, and Ross (1985).

18

Journal of Financial and Quantitative Analysis

M and B identify the market and long-term govemment bond portfolios, respectively, N is the total number of assets in the multivariate system under consideration,
cTyi ipjM - PJBPBM) . ^, 1 . 1u .

GM,

(1 -

IS the market nsk beta.

^ ^ y ' ( ^ ^ - f t ^ M is the bond risk beta,


( ^ [ ^ P )

(jj, is asset y's conditional standard deviation, and pij is the conditional correlation coefficient between the retums for asset i and asset/ Equation (1) shows that, in equilibrium, the expected excess retum on any security i is given by a market risk beta multiplied by an expected market risk premium, plus a bond risk beta multiplied by an expected bond risk premium. In specification (1), all moments are time dependent except for the correlations that, consistent with Schwert and Seguin (1990) and Ng (1991), are assumed constant. Ex ante long-term bond portfolio excess retums are employed in the asset pricing relation (1) because of the sensitivity of long-term bond retums to changes in the short rate (cf., Merton (1973)). Shanken (1990) also uses excess retums on the market and the long-term-bond portfolio as regressors in a multifactor model; regressor coefficients are assumed to be linear functions of Treasury bill retums. Treasury bill volatility, and a January dummy. Our multifactor model is similar except that our regression coefficients are derived directly from the asset pricing relation, using second moment dynamics, rather than from an assumed relation. Two versions of the ICAPM are nested in (1) to permit empirical testing. The first specification restricts risk prices to be constant while the second allows time-varying risk prices. A. The Constant Risk Price Model The price of market risk is defined as the tradeoff between ex ante market excess retum and market variance, 8UMJ = (oiMiPfi)/crlMi' ^"^ similarly, the bond risk price is defined as 6UBJ = (/?< - f^fd/f^lB,, j = \,2,...,N.^ Substituting these definitions into the ex ante relation (1), adding an intercept, and allowing a random disturbance term, up, to affect ex post excess retums results in,
(2a) rj, = Cuj + 0ujMSuMj(^ujt<^uMl + 6ujB8uBj(^uj,CruBt + Uj,,
(PujM PUJBPUMB

where

r,,

Rj, - Rf,, ^UJM -

^ i

V , and

PuMB

a f P"JB ~ PujMPuMB VujB = I ' 2


V ' " PuMB

'To distinguish between the constant and time-varying risk price models, an additional subscript related to the disturbance term (either u or e) is appended to distinguish the models. For notational convenience, we omit the additional subscript when the context is clear.

Turtle, Buse, and Korkie

19

The constant risk price model posits constant prices for both market and bond risks; i.e., 6UMJ = SUM and^uBj = 6UB, j= 1,2,... ,N. Adding the restriction Cuj ~ 0, for all assets7, we obtain the ICAPM with constant prices of market and bond risk,^ (2b) rj, = (

Altemative restrictions on the values of the coefficients, Cuj, SUMJ^ and SUBJ. j = 1,2, ...,N, allow tests of the hypothesized asset pricing relation undei varying altemative hypotheses. A sufficient set of conditions, which reduces Equation (2a) to the CAPM with a constant market risk price, is Cuj = 0, SUMJ = ^UM, and 6UBJ - 0 for all assets J. Retums then collapse tor;, = ^uMPujM^uitOuMf^^jt- In this case, huM can be interpreted as a scalar related to the aggregate relative risk aversion of the representative agent (c.f., Ferson, Kandel, and Stambaugh (1987) or Ng (1991)). All model disturbances are assumed to be multivariate normal with zero mean and a time-varying covariance matrix specified by a multivariate GARCH( 1,1) process. We choose a GARCH(1,1) process because of the strong support found for this process in recent empirical work, and the tractability provided by low order variance processes in a multivariate setting (c.f., Bollerslev, Chou, and Kroner (1992)). Additionally, we find strong support for the (1,1) process based upon a consistent search methodology over second moment processes reported in an earlier draft of this paper. Therefore, second moments are described by the multivariate GARCH(1,1) process. (3)
.2 _

%i<^ljt-\+Kjul-x,

0 0 where 0 0

and u, = {uj,} is distributed as a multivariate normal random vector (0, (Ju,Pu(Tut) with pu = {puij}NxN- We compute the conditional correlation between any asset a n d the m a r k e t as, pujM = J2k=\ ^ktPujk((^ukt/cruMi), J = l,2,...,N, w h e r e tj/ct is the weight of portfolio k in the market portfolio for period /, and Pujt is the pairwise unconditional correlation between the disturbances for the excess retums on portfolios y and k.^ The standard deviation of the market disturbance at time t
is given by auMt = [!^',{<^utPuCru,)'(^,f^^, where zu, = {wk,)N^\^Constancy of S^M has been required in previous CAPM studies, such as Bollerslev, Engle, and Wooldridge (1988), and Giovannini and Jorion (1989), which use multivariate GA-RCH techniques. Ng (1991) reports results under the assumption of constancy of i5,^, although time variability in Che ratio of ex ante market retum to variance is admitted. 'To be notationally precise, correlations with the market should be written as functions of r because market weights change with time. However, for notational simplicity and because portfolio weights are relatively stable over the sample period, the t subscript is omitted.

20 B.

Journal of Financial and Quantitative Analysis Time-Varying Risk Price Model

To develop a time-varying risk price specification, consider the general linear model for N assets,
(4a) rj, Cej, + l3ejM,rMi + PejBirBi + ej,, j = l,2,...,N,

where (3ejMi is the sensitivity to market excess retums,/?,,;^, is the sensitivity to bond excess retums, E(ej,) = 0, and E(ej,rM,) = E(ej,rB,) = 0 for all j and t. Because (1) is conditional upon time, so is the preceding retum-generating model and the variables are random over the conditional state space at any point in time, t. It is known from Roll (1977) (see also Jobson and Korkie (1989) for a survey) that Ceji = Cej = 0, for each period t, implies and is implied by the asset pricing relation (1). Using the constant correlation assumption, the retum model (4a) may be written for each period as, (4b) r,, =

where 6..

{ ^^JLE^iEMl] ^-p'eMB

and J

^ '

I P'^'IT PejMPeMB V 1-PeMB

The prices of market and bond risk, (aM, - Rfi)/^^^, ^"d (^B; - i^f,)/(^eBi' respectively, are implicitly allowed to time vary along with the conditional variances.^ Because the long-term bond is a hedging portfolio under the null hypothesis, it trivially satisfies both Equations (2) and (4), irrespective of the values of its conditional moments. Therefore, an exogenous model of conditional bond excess retums is required. To maintain tractability and to avoid misspecification due to omitted variables, we estimate the bond retum process as a univariate time series model. Identification analysis of long-term bond portfolio excess retums reveals significant spikes in low-order terms of the sample partial autocorrelation function and a decay in the sample autocorrelations.^ Because this is consistent with a loworder autoregressive process, we describe the evolution of ex ante bond excess returns using the first-order autoregressive process, (5) rs, = db + (f)rB,-i+eB,.

Conditional variances, a^j,, are then fit for all portfolios, including the bond, using a GARCH(1,1) process analogous to Equation (3).
*In Equation (4), ex post market and bond portfolio excess retums enter as regressors for ex post excess retums on asset y. One can view the constant risk price model as being derived from the returngenerating model in (4) by replacing r^, with (ayjf, /f/,)+^M( ^nd ''fir *i'h (ag, W/-,)+?B, -This shows that the disturbances are related according to Uj, = ej, - {Cej,/(TeMt)^ejMiMi - (.'^eji/'^eBiWejB^Bi' which implies that the time-varying risk price model has smaller residual variance than the constant risk price model. 'Specifically, the first eight sample autocorrelations, with standard errors in parentheses, are 0.22 (0.05),0.25 (0.06),0.15 (0,06),0.14(0.06),0.16(0.06),0.11 (0.06),0.049 (0.06),and0.071 (0,06). In contrast, only the first two sample partial autocorrelations of 0,22 and 0,21 are significantly different from zero. These sample results suggest a low-order autoregressive process in a Box-Jenkins univariate model, ignoring GARCH effects.

Turtle, Buse, and Korkie

21

III.
A.

Data and Methodology


Data

Weekly retums over the period July 14, 1983, to December 14, 1989, were used in the estimation. The decision to study this period was driven by the availability of Treasury bond and bill retums, which were obtained from Reuters. Weekly stock retums were computed by compounding the total daily retums reported by CRSP from Thursday close until Thursday close, an interval motivated by the desire to avoid exchange closures. If a retum is missing on Thursday, Wednesday's retum is used. If a retum is missing on the Wednesday, Friday's retum is used. If a retum is missing on all three days, the stock is excluded. The 1419 available stocks were combined to form three size portfolios where each size portfolio contains its respective third of all firms ranked by market capitalization at the beginning of each week. For each week. Treasury bill retums were computed for maturities one week hence. At each point in time, the long-term bond portfolio contains a single U.S. Treasury bond with a maturity as close to 20 years as possible; retums are computed as in Ibbotson Associates ((1986), pp. 1819). Because market weights for the long-term bond portfolios are unavailable, the value-weighted average of the three size portfolios is defined as the market proxy. Table 1 presents descriptive statistics for the three size-based portfolios and the long-term U.S. Treasury bond portfolio. Panel A contains maximum likelihood estimates of the unconditional mean and standard deviation for each portfolio. The large retum on the long-term Treasury bond portfolio reflects the decline in shortterm interest rates over the sample period. Small firms show markedly poorer performance than large firms, which Reinganum (1992) ascribes to a predictable reversal of the size effect approximately every five years. We speculate that small and large firms differ with respect to interest rate risk as found in Ng, Engle, and Rothschild (1992); however, an examination of the small-firm anomaly is not of concem in this paper. The next four columns display four extrema for each excess retum series and their associated dates. Notice that each size portfolio has a very large negative extremum just prior to the stock market crash of 1987. Contrarily, the long-term bond portfolio displays its largest retum at the same time. The reported Box-Pierce Portmanteau Q(12) statistics indicate serial correlation in all excess retum series. Panel B reports estimates of the mean and standard deviation of the squared excess retum series along with the Box-Pierce Portmanteau 2(12) test statistics. These 2(12) statistics demonstrate a significant amount of serial correlation in second moments, which we attempt to remove with a multivariate GARCH process. B. Estimation Methodology

The mean specification given by Equation (2) in conjunction with secondmoment dynamics as described by Equation (3) determines the constant risk price version of the model. The time-varying risk price model has mean relations generated by (4) and (5), with second moments given by the equation equivalent to (3) for the disturbance vector e,, rather than ,. In either case, the data allow the

22

Journal of Financial and Quantitative Analysis


TABLE 11
Summary Statistics for Weekiy Excess Returns and Squared Excess Returns on Four Portfolios

Extrema Portfolio Mean Dev. Lowest Highest O(12)

Panel A. Portfolio Excess Returns Largest Size Portfolio Midsize Portfolio Smallest Size Portfolio LT Bond Portfolio 0.153 0.111 0.024 0.324 2.134 2.087 2.099 1.782 -17.16 10/22/87 -19.16 10/22/87 -20.45 10/22/87 -4.16 04/24/86 -7.55 12/03/87 -8.01 10/29/87 -11.35 10/29/87 -4.15 05/08/86 6.06 01/07/88 7.01 01/08/87 8.48 01/07/88 11.26 10/22/87 5.98 01/08/87 6.67 11/05/87 7.36 01/08/87 6.40 02/27/86 31.51** 44.90** 51.37** 67.39**

Panel B. Sauared Excess Returns Largest Size Portfolio Midsize Portfolio Smallest Size Portfolio LT Bond Portfolio 0.046 0.044 0.044 0.033 0.171 0.211 0.246 0.087 21.87* 26.37** 46.80** 49.60**

Summary statistics for the 336 effective weekly excess returns (multiplied by 100) from July 14, 1983, to December 14, 1989, are shown in Panel A. Weekly long-term bond portfolio excess returns, rg,, are constructed from U.S. Treasury bond data provided by Reuters. Panel B sfiows summary statistics for the squared excess return series. Mean and Std. Dev. are the unconditional maximum likelihood estimates of the mean and standard deviation. The extrema reported are the two smallest and two largest observations for each series. Dates are reported below extrema in month/day/year format. The 0(12) statistic reported is the Box-Pierce Portmanteau test statistic to determine if the first 12 autocorrelations are significantly different from zero as a group. 0(12) is distributed as a x random variable with significance at the 1-percent and 5-percent ievels denoted by ** and *, respectively.

computation of each model's multivariate error term, which is subsequently used to maximize the likelihood function. The estimation procedure follows Engle (1982), Bollerslev (1986), Giovannini and Jorion (1989), and Ng (1991). Excess retum disturbances are assumed to follow a multivariate normal distribution. Then, for any particular specification of the model with A ^ variates, the log of the likelihood function is L(6) = J2^^\ ''(^)' where 1,(6) is the log of the normal likelihood for any particular observation, t, a,(6)p(0)a,(6) is the covariance matrix of the disturbance vector e,(9) where ,(6) is either e, OTU,,9 is the (^ x 1) parameter vector, and g represents the number of parameters estimated in any particular model. For example, in the constant risk price model under the null hypothesis of ICAPM, the parameter vector includes the prices of market and bond risk, all pairwise correlations, and all GARCH(1,1) parameters; hence, g is 20. To estimate an altemative hypothesis with nonzero intercepts, we add four intercepts to the parameter vector. As evident in Section II, the likelihood function depends on ^ in a highly nonlinear manner. We maximize (8) using numerical derivatives in conjunction with the method of Bemdt et al. (1974). The convergence criterion is set at 0.0001. Estimation of all models involved extensive investigation at the optima to ensure maxima are not simply local peaks in the likelihood function. Nonnegativity of variances is ensured by estimating the square root of all second moment

Turtle, Buse, and Korkie

23

parameters."^ For any particular model, the likelihood is maximized under the altemative hypothesis and then under the restricted null hypothesis. Finally, likelihood ratio statistics are employed to determine if the asset pricing constraints are binding.

IV.

Empirical Results

The results are presented first for the constant risk price model under both the CAPM and the ICAPM, and then for the ICAPM under time-varying risk prices. A. Constant Risk Price Model

Table 2 reports the estimation resuUs foT the constant risk price versions of the CAPM and ICAPM models using Equations (2) and (3) and assuming constant correlations. Panel A shows that under the CAPM, the estimated market price of risk is 6.2 and falls within literature standards." Panel B reports likelihood ratio statistics for specific alternatives to the CAPM that admit nonzero intercepts and cross-sectional differences in the prices of market and bond risk. Altemative hypotheses that include a constant, which is the same for all assets, may be viewed as Black CAPM tests. Cross-sectional variation in the price of market risk may be due to failure of the constant market risk price assumption, misspecification of the index, failure to include bond risk, or other market imperfections such as taxes and transactions costs. For example, in thefirstrow of Panel B, the test statistic value of 8.33 rejects, at the 5-percent level, the CAPM against the specific altemative of unrestricted market risk prices. Therefore, we conclude that a constant risk price, single-factor CAPM does not explain cross-sectional differences in size portfolio retums. The second row of Panel B shows that the multivariate intercept restriction is not binding as the ^ statistic is insignificant at even the 10-percent level: the intercept restriction appears to be the least stringent of the null's restrictions. The third row has unreported f)M and f>B estimates (r-statistics in parentheses) of 4.30 (1.47) and 8.30 (2.41), respectively. These estimates indicate failure of the CAPM and motivate the inclusion of risk premia related to bond risk and/or time-varying risk prices. The remaining rows of Panel B show the myopic CAPM model is consistently rejected for a number of specific altematives. Panels C and D of Table 2 report the estin\ation results for the ICAPM under constant prices of market and bond risk. The price of market risk remains significant and largely unchanged at 6.1, compared to the CAPM estimate in Panel A; the estimated bond risk price of 9.6 is statistically and economically significant. The significance of both risk prices provides support for the ICAPM or a completing market factor in a myopic CAPM. In Panel D, we report x^ test statistics for two specific altematives to the ICAPM. In both cases, the reported test statistics are insignificant at the 10-percent
'"Preliminary estimation results, without positivity restrictions on the GARCH parameters, led to nonpositive-definiteness of the covariance matrix. Previous research has made extensive use of parameter restrictions to ensure positive variances, especially in GARCH models (c.f.. Ng (1991)). '' For example, Ferson (1989) estimates the market price of risk to be approximately seven.

24

Journal of Financial and Quantitative Analysis


TABLE 2 Hypotheses and Tests' Results of the Myopic CAPM and the ICAPM with Constant Risk Prices Myopic CAPM ICAPM

Null Hypothesis: = 0, y = 1.2 4 General Alternative Hypottiesis: Ho/;Cu; = 0, SuM / = 1,2 4

UjM ~ PujB * PuMB

^ - PIMB
PujB ~ PujM * PuMB

Second Moment Dynamics:


^u^t 2

0
''u2t

0 0 0
Pu =

0 0
^U5t

0 0

0 0 0

" I = {"iti

and u, is distributed normal [0, (

Panei A. Myopic CAPM Tests Horn : Cui = 0, 6uMi = SUM- and 6uBi = 0 Pane) B. Specific Myopic CAPM Alternative Hypotheses Tests M' and S^Bi = 0 M. and ue/ = 0 i y u; = 0. uM; i* uj + 0' *uM/ + Panel C. iCAPM Tests HOI : c^j - 0. SijMi = * 8.33** 10.49 9.11'** 15.89** 18.98* 6.21'** (2.33) d.f. 3 7 1 7 11

M- and o

and

6.09** (2.29)

9.62*** (2.98) d.f 10 6

Panel D. Specific tCAPM Alternative Hypotheses Tests y ^ 0, y = 0.

x^
8.53 5.44

Excess returns on three size based portfolios Ci,, (2,. f3,, and a long-term bond porlfolio, rg,, are used in estimation. Three size portfolios are aggregated to form the market portfolio where the weight of each size portfolio is time varying and based upon its market capitalization at the beginning of each period. The other parameters in the model include: I5UM/, a parameter related to market risk; SUBJ. a parameter related to bond risk; p^ji, the constant unconditional correlation between u,, and Ujt\ and t/f, the vector of^model disturbances, which is assumed to be distributed as a multivariate normal with zero mean vector and conditional covariance matrix given by [<Tuiit]' where ^r^y^ = Puqi^uit^uit' Panel A reports the coefficient estimate of the market price of risk, S^^jf with its asymptotic (-statistic (in parentheses) under the null hypothesis of the myopic CAPM. Panel B reports x^ lest statistics for various specific myopic CAPM alternative hypotheses. H^/K/, against the null hypothesis, HQJM, with * * * , * * , and * denoting significance at the 1-percent, 5-percent, and 10-percent levels, respectively. Panel C reports estimates and /-statistics for market and bond risk prices, ^UM and UB. respectively, under the ICAPM null hypothesis. Panel D reports likelihood ratio test results for specific ICAPM alternative hypotheses.

level. Further support for the ICAPM is provided by the f-statistics of the unrestricted intercepts in the constant risk price model with Cuj 4 0. ^UMJ 4 ^UM, and ^uBj 7^ ^uB- The (unreported) intercept f-statistics are 0.106, -0.726, -0.718, and 0.529 for the largest size, mid-size, smallest size, and long-term bond portfolios, respectively. The consistent insignificance of these ^statistics is supportive of the likelihood ratio inferences. Thus, if one is willing to accept that the portfolio, M, is a reasonable proxy for the market portfolio in an ICAPM equilibrium, the results of Table 2 strongly support the inclusion of a hedge term in the asset pricing model. Alternatively, the CAPM is not rejected if the portfolio M is not representative of the true market

Turtle, Buse, and Korkie

25

portfolio in a CAPM equilibrium; rather, some combination of M and the bond portfolio, 6, is a better proxy for the true market ponfolio. Because of the omission of the bond factor portfolio, previous studies using equity-based market proxies are misspecified in an empirically important manner. Previous research that has led to the rejection of the CAPM using similar asymptotic test statistics, includes the work of Giovannini and Jorion (1989) and Ng (1991). The overall results and the strong significance of risk prices presented in Table 2, in conjunction with numerous rejections of the myopic CAPM in earlier literature, lends support to the conclusion of a missing hedge portfolio or a completing market factor portfolio. B. Time-Varying Risk Price Model In this section, we report the estimation results for the ICAPM with timevarying risk prices and time-varying betas of market and bond risk, which are nonlinear functions of correlations, portfolio weights, and time-varying standard deviations. As discussed in Section II, bond portfolio excess retums are described as an autoregressive GARCH process. Additionally, we consider the case where conditional bond excess retum means are constant, due to the insignificance of the bond's AR(1) coefficient in both the null and altemative versions of the model. Imposing the multivariate zero intercept restriction for the three size portfolios results in insignificant x^ test statistics of 2.10 and 2.16 for the bond process under an autoregressive mean process or simply a constant, respectively. Therefore, we conclude that the ICAPM is not rejected in the time-varying risk price model. These asymptotic inferences are consistent with the insignificance of the intercept f-statistics regardless of which bond process is used. In fact, the largest absolute /-statistic for intercepts across both versions of the model is only 0.85. In summary, the ICAPM model is supported under the assumption of both constant and time-varying risk prices. Insignificant x^ values for all tests of the ICAPM model, in conjunction with significant risk price terms for the constant risk price model, lend support to the ICAPM paradigm with time-varying moments or to a CAPM with time-varying moments and a completing market factor. In the following diagnostic section, these conclusions are examined further.

V.

Diagnostic Analyses

The purpose of this section is to examine the dynamics of the standardized residuals and squared standardized residuals from the estimated models in relation to the original excess retum series, shown in Table 1. Standardized residuals are computed as model residuals divided by contemporaneous standard deviations, Vj,/sj,, for each portfolio y. Under normality, standardized and squared standardized residuals should be distributed with means of zero and unity, respectively. Table 3 reports the mean, standard deviation, extrema, and the Box-Pierce Portmanteau statistic for the first 12 sample autocorrelations, 2(12). For brevity and to focus on shortcomings of the asset pricing restrictions, we only report diagnostic statistics for each of the ICAPM models under the null hypotheses. In Panel A, for constant risk prices, we note that the standardized residuals tend to have negative means; nevertheless, squared standardized residuals appear close

26

Journal of Financial and Quantitative Analysis

to the hypothesized value of unity. Panel B shows that, for the time-varying risk price model, residual means are negative for two portfolios and positive for two portfolios and squared standardized residuals appear larger than unity for the midsize and the small portfolio.
TABLE 3 Diagnostic Statistics for the ICAPM Residuals under Constant and Time-Varying Risk Prices Diagnostic Statistics Std. Dev. Extrema Lowest Highest

Portfolio

Mean

0(12)

Panel A. ICAPM under Constant Risk Prices Standardized Residuals Largest Size Portfolio Midsize Portfolio Smallest Size Portfolio Long-Term Bond Portfolio Squared Standardized Residuals Largest Size Portfolio Midsize Portfolio Smallest Size Portfolio Long-Term Bond Portfolio Panel B. ICAPM with Time-VarvinQ Standardized Residuals Largest Size Portfolio Midsize Portfolio Smallest Size Portfolio Long-Term Bond Portfolio Squared Standardized Residuals Largest Size Portfolio Midsize Portfolio Smallest Size Portfolio Long-Term Bond Portfolio 0.715 1.75 1.92 4.71 1.90 4.77 1.03 2.28 112.14** 114.27** 86.32** 8.76 0.087 0.842 -0.129 1.38 -2.07 3.71 -2.49 4.61 12/03/87 01/08/87 10/29/87 01/07/88 -7.13 -6.55 4.02 3.39 10/29/87 01/15/88 12/03/87 01/08/87 -8.00 -5.20 3.62 3.55 10/29/87 01/15/88 01/05/84 01/08/87 -2.73 -2.35 5.62 3.96 04/02/87 04/24/86 10/22/87 02/27/86 15.13 15.89 13.10 13.84 1.01 3.54 1.01 3.67 1.01 4.09 0.998 1.99 Risk Prices 5.74 4.02 4.63 11.49 -0.048 -0.054 -0.064 -0.018 1.01 1.01 1.01 1.00 -7.79 -3.72 2.95 2.61 10/22/87 09/11/86 01/08/87 08/02/84 -7.83 -3.67 4.09 2.59 10/22/87 09/11/86 01/08/87 08/02/84 -8.28 -3.20 4.61 3.18 10/22/87 09/11/86 01/08/87 01/07/88 -2.81 -2.66 4.96 3.71 04/02/87 04/24/86 10/22/87 02/27/86 25.15* 36.05* 40.67* 12.44

-0.158 1.37 0.040 1.02

Standardized residuais are computed as model residuals divided by the contemporaneous conditional standard deviation, Vji/Sji. Squared standardized residuals are computed as (.Vjt/Sj,)^. All panels report maximum likelihood estimates of the mean and standard deviation (Std. Dev.) of each series. The extrema reported are the two smaliest and two iargest observations for each data series. Dates are reported below extrema in month/day/year format. O(12) is the Box-Pierce Portmanteau statistic for the first 12 sample autocorrelations. Panel A reports results for the ICAPM model under constant risk prices. Panel C reports summary measures for the ICAPM model with time-varying risk prices and a constant bond excess return mean relation. Significance at the 1-percent and 5-percent levels for the O statistics is denoted by '" and *, respectively.

Comparison of the 2(12) statistics in Table 1 and Panel A of Table 3 reveals that the constant risk price mode) effectively eliminates serial correlation in squared standardized residuals; however, standardized residuals continue to display dependence. Although there is clear improvement in the dynamic properties

Turtle, Buse, and Korkie

27

of the standardized residuals, the dependence of pricing errors on information that is readily available to investors suggests that the constant risk price model does not adequately capture the behavior of conditional expected retums over the sample. Contrarily, Panel B shows that the time-varying risk price model's standardized residuals appear free of serial correlation; however, squared standardized residuals continue to display serial correlation for all portfolios but the long-term bond portfolio. This suggests that the multivariate GARCH(1,1) model and time-varying risk prices do not fully capture the dynamics of second moments, but that the evolution of expected retums is adequately explained.'^ It is interesting to note that modeling the time variability in long-term bond portfolio second moments eliminates the serial correlation in long-term bond portfolio standardized residuals, even in the case of the bond's constant mean specification. Thus, the autocorrelation in bond retums found in Table 1 may arise mostly from variance dynamics, rather than conditional mean dynamics. Comparison of unreported nonstandardized residual variances, from the constant and time-varying risk price models, shows a marked decrease in residual variances in the time-varying risk price models. Under the null of the ICAPM, the time-varying risk price model residual Variances are (multiplied by 1,000) 0.012, 0.067,0.124, and 0.323, for the largest size, midsize, smallest size, and long-term bond portfolios, respectively. The constant risk price model, which uses only lagged variables in the information set, results in much larger residual variances of 0.440,0.390,0.410, and 0.334, as expected (fn. 8). Table 3 shows that the lowest extrema for all size portfolios is due, in all models, to the stock market crash of 1987. Similarly, bond portfolio standardized residuals display their largest extremum at the same point in time. Because of the influence of outliers on sample autocorrelations and, hence, our Q statistics, we performed subperiod analysis excluding the 1987 stock market crash. After splitting the sample into two subperiods that exclude the crash (July 1983-September 1987 and December 1987-December 1989), serial correlation in model residuals and squared residuals is virtually eliminated according to the Q statistics.'^ Therefore, we conclude that the correlation in asset pricing errors is primarily due to the stock market crash of 1987. Further analyses, linking asset pricing standardized residuals or squared standardized residuals to market variance, found little support for an additional risk factor related to aggregate market volume.

VI. Conclusions
In this paper, we examine conditional asset pricing models with time-varying moments and risk prices and assuming constant conditional correlations. We interpret all tests as joint tests of the market proxy and the asset pricing paradigm. A value-weighted equity portfolio is used as the market proxy and a long-term
'^The reported Q statistics detect serial correlation, conditional upon past standardized residuals, or past squared standardized residuals, respectively. It is likely that a richer conditioning set tnight find further evidence of remaining predictability. '^Because these subsample tests suffer a loss in power due to a reduced sample size, we performed additional tests that included the crash period in each of the subsamples and we rejected the null hypothesis of zero autocorrelations in the subsamples.

28

Journal of Financial and Quantitative Analysis

bond portfolio is used as the hedging portfolio or the completing market factor. GARCH(1,1) variance processes specify evolving conditional variances that impact ex ante portfolio mean retums. We posit two specific formulations of the pricing model, one with constant risk prices and the other with time-varying risk prices. Using weekly data from July 1983 to December 1989, we reject the singlefactor myopic CAPM with a constant risk price but we do not reject the ICAPM, particularly when time-varying risk prices are employed. We find empirical evidence that investors price long-term bond risks, which hedge changes in the investment opportunity set. AJtematively, the hedge term may be viewed as a completing market factor hecause the market portfolio is subject to measurement error. Thus, the strong significance of risk prices supports the notion of hedging terms in the ICAPM, or completing market factors in a myopic CAPM. Equity index proxies for the hedge portfolios and the market portfolio are clearly inadequate. The time-varying risk price models capture much of the serial correlation in standardized residuals and greatly reduce the residual variance for all portfolios, as expected. Diagnostic analysis reveals some remaining serial correlation in standardized residuals, for the constant risk price model, and in squared standardized residuals, for the time-varying risk price model, suggesting that these models do not entirely capture portfolio retum dynamics. Subperiod diagnostics show that the correlation in standardized residuals and squared residuals is associated with the stock market crash of 1987 and it is not due to a missing state variable related to total market volume. Exclusion of the crash period results in an estimated time-varying risk price model with little remaining correlation in residuals or squared residuals. Our results are consistent with the idea that investors are surprised by high variance periods from unknown sources of risk. The GARCH variance processes do not capture these large and unanticipated shocks to the system. In general, the ICAPM with a long-term bond hedge ponfolio and timevarying means, risk prices, and variances modeled by a GARCH-M seems to do a reasonably good job of capturing cross-sectional differences in our portfolio retums except when there are extreme changes in the moments.

References
Bemdt. E. K.; B. H. Ha)); R. E. Ha)); and J. A. Hausman. "Estimation and Inference in Non)inear Structural Models." Annals of Economic and Social Measurement, 4 (1974), 653-665. Bodurtha, J. N., and N. C. Mark. "Testing the CAPM with Time-Varying Risks and Retums." Journal of Finance, 46 (1991), 1485-1505. Bollerslev. T. "Generalized Autoregressive Conditiona) Heteroscedasticity." Journal of Econometrics, 31 (1986). 307-327. Bollerslev, T; R. Y. Chou; and K. F. Kroner. "ARCH Modelling in Finance: A Review of the Theory and Empirical Evidence." Journal of Econometrics, 52 (1992), 5-60. BoIIersiev, T.; R. F. Engle; and /. M. Wooidddge. "A Capital Assel Pricing Model with Time-Varying Covariances." Journal of Political Economy, 96 (1988).l 16-131. Breeden, D. "An tntertemporal Asset Pricing Model with Stochastic Consumption and Investment Opportunities." Journal of Einancial Economics, 1 (1979), 265-296. "Consumption, Production, Inflation, and Interest Rates: A Synthesis." Journal of Financial Economics, 16 (1986), 3-39. Cox, J.; J. Ingersoll; and S. Ross. "An Intertemporal General Equilibrium Model of Asset Prices." Econometrica, 53 (1985), 463-484.

Turtle, Buse, and Korkie

29

Engle, R. F. "Autoregressive Conditional Heteroscedasticity with Estimates of the Variance of U.K. Inflation." Econometrica, 20 (1982), 83-104. Engle, R. E; V. K. Ng; and M. Rothschild. "Asset Pricing with a Factor-ARCH Covariance Structure: Empirical Estimates for Treasury Bills." Journal of Econometrics, 45 (1990), 213-237. Fama, E. F. "Multiperiod Consumption-Investment Decisions." American Economic Review, 60 (1970), 163-174. "Stock Retums, Expected Retums, and Real Activity." Journal of Finance, 45 (1990), 1089-1108. Fama, E. F., and K. R. French. "Business Conditions and Expected Retums on Stocks and Bonds." Journal of Financial Economics, 25 (1989), 23-50. Ferson, W. E. "Changes in Expected Security Retums, Risk, and the Level of Interest Rates." Journal of Finance, 44 (1989), 1191-1217. "Are the Latent Variables in Time-Varying Expected Retums Compensation for Consumption Risk?" Journal of Finance, 45 (1990), 397-429. Ferson, W. E., and C. Harvey. "The Variation of Economic Risk Premiums." Journal of Political Economy, 99 (1991), 385^15. Ferson, W. E.; S. Kandel; and R. F. Stambaugh. "Tests of Asset Pricing with Time-Varying Expected Risk Premiums and Market Betas." Journal of Finance, 42 (1987), 201-220. Giovannini, A., and P. Jorion. "The Time Variation of Risk and Retum in the Foreign Exchange and Stock Markets." Journal of Finance, 44 (1989), 307-325. Harvey, C. "The World Price of Covariance Risk." Journal of Finance, 46 (1991), 111-157. Ibbotson Associates. Stocks, Bonds, Bills, and Inflation: 1986 Yearbook. Chicago, IL (1986). Jobson, J. D., and R. Korkie. "A Performance Interpretation of Multivariate Tests of Asset Set Intersection, Spanning, and Mean-Variance Efflciency." Journal of Financial and Quantitative Analysis, 24(1989), 18-204. Lintner, J. "The Valuation of Risky Assets and the Selection of Risky Investments in Stock Portfolios and Capital Budgets." Review of Economics and Statistics, 47 (1965), 13-37. Long, J. "Stock Prices, Inflation, and the Term Structure of Interest Rates." Journal of Financial Economics, 1 (1974), 131-170. Merton, R. C. "Optimum Consumption and Portfolio Rules in a Continuous-Time Model." Journal of Economic Theory, 3 (1971), 373^13. "An Intertemporal Capital Asset Pricing Model." Econometrica, 41 (1973), 867-887. Ng, L. "Tests of the CAPM with Time-Varying Covariances: A Multivariate GARCH Approach." Journal of Finance, 46 (1991), 1507-1521. Ng, V; R. F. Engle; and M. Rothschild. "A Multi-Dynamic-Factor Model for Stock Retums." Journal of Econometrics, 52 (1992) 245-266. Reinganum, M. "A Revival of the Small-Firm Effect." The Journal of Portfolio Management, 18 (1992), 55-62. Roll, R. "A Critique of the Asset Pricing Theory's TestsPart 1: On Past and Potential Testability of the Theory." Journal of Financial Economics, 4 (1977), 129-176. Rubinstein, M. "The Valuation of Uncertain Income Streams and the Pricing of Options." Bell Journal of Economics, 7 (1976), 407-425. Schwert, G. W. "Why Does Stock Market Volatility Change over Time?" Journal of Finance, AA (1989), 1115-1153. Schwert, G. W., and P. J. Seguin. "Heteroskedasticity in Stock Retums." Journal ofFinance, 45 (1990), 1129-1155. Shanken, J. "Intertemporal Asset Pricing: An Empirical Investigation." Journal of Econometrics, 45 (1990), 99-120. Sharpe, W. F. "Capital Asset Prices: A Theory of Market Equilibrium under Conditions of Risk." Journal of Finance, 19 (1964), A25-A42.

También podría gustarte