You are on page 1of 35

Components of market risk and return

John M. Maheu and Thomas H. McCurdy∗
1st Draft: November 2004; This version: June 2007
Forthcoming: Journal of Financial Econometrics

Abstract
This paper proposes a flexible but parsimonious specification of the joint dynamics of market risk and return to produce forecasts of a time-varying market equity
premium. Our parsimonious volatility model allows components to decay at different
rates, generates mean-reverting forecasts, and allows variance targeting. These features
contribute to realistic equity premium forecasts for the U.S. market over the 1840-2006
period. For example, the premium forecast was low in the mid-1990s but has recently
increased. Although the market’s total conditional variance has a positive effect on
returns, the smooth long-run component of volatility is more important for capturing
the dynamics of the premium. This result is robust to univariate specifications that
condition on either levels or logs of past realized volatility (RV), as well as to a new
bivariate model of returns and RV.

JEL classification: C53; C22; G10
Keywords: volatility components; long-run market risk premium; Realized Volatility

Maheu, Department of Economics, University of Toronto, 100 St. George Street, Toronto ON M5S 3G3,
416-978-1495, jmaheu@chass.utoronto.ca; McCurdy, Joseph L. Rotman School of Management, University
of Toronto, 105 St. George Street, Toronto, ON M5S 3E6, 416-978-3425, and Associate Fellow, CIRANO,
tmccurdy@rotman.utoronto.ca. The authors thank the editor, Ren´e Garcia, two anonymous referees, Hakan
Bal, Harjoat Bhamra, Chris Jones, Mark Kamstra, Raymond Kan, Benoit Perron, Daniel Smith, participants at the CIREQ-CIRANO Financial Econometrics Conference, the NFA annual meetings, and seminar
participants at Simon Fraser University and McMaster University for very helpful comments, as well as Bill
Schwert for providing U.S. equity returns for the period 1802-1925. They are also grateful to SSHRC for
financial support. An earlier draft of this paper was titled “The long-run relationship between market risk
and return”.

1

1

Introduction

The expected return on the market portfolio is an important input for many decisions in
finance. For example, accurate measures or forecasts of the equity premium are important
for computing risk-adjusted discount rates, capital budging decisions involving the cost-ofequity capital, as well as optimal investment allocations.
The simplest approach to measuring the market premium is to use the historical average
market excess return. Unfortunately, this assumes that the premium is constant over time.
If the premium is time varying, as asset pricing theory suggests, then an historical average
will be sensitive to the time period used. For example, if the level of market risk were
higher in some subperiods than others, then the average excess return will be sensitive to
the subsample chosen.
A better approach to estimating the premium is to directly incorporate the information
governing changes in risk. For example, the Merton (1980) model implies that the market
equity premium is a positive function of market risk, where risk is measured by the variance
of the premium. Under certain conditions discussed in the next section, intertemporal asset
pricing models reduce to a conditional version of Merton (1980). That is, if the conditional
variance of the market portfolio return is larger, investors will demand a higher premium to
compensate for the increase in risk.1
This positive risk-return relationship for the market portfolio has generated a large literature which investigates the empirical evidence. Historically, authors have found mixed
evidence concerning the relationship between the expected return on the market and its
conditional variance. In some cases a significant positive relationship is found, in others it
is insignificant, and still others report it as being significantly negative.2
Recent empirical work investigating the relationship between market risk and return offers some resolution to the conflicting results in the early literature. Scruggs (1998) includes
an additional risk factor implied by the model of Merton (1973), arguing that ignoring it in
an empirical test of the risk-return relationship results in a misspecified model. Including a
second factor, measured by long-term government bond returns, restores a positive relationship between expected return and risk. Campbell and Hentschel (1992), Guo and Whitelaw
(2006) and Kim, Morley, and Nelson (2004) report a positive relationship between market
risk and return when volatility feedbacks are incorporated. Using a latent VAR process to
1

There are many other models of asset premiums, many of which can be thought of as versions of a
multi-factor approach, such as the three-factor model of Fama and French (1992), or the arbitrage pricing
theory of Ross (1976). For example, Claus and Thomas (2001), Fama and French (2002), and Donaldson,
Kamstra, and Kramer (2004) use earnings or dividend growth to estimate the market premium. Pastor,
Sinha, and Swaminathan (2007) use implied cost of capital as a measure of expected return.
2
Early examples include Campbell (1987), Engle, Lilen, and Robins (1987), French, Schwert, and Stambaugh (1987), Chou (1988), Harvey (1989), Turner, Startz, and Nelson (1989), Baillie and DeGennaro (1990),
Glosten, Jagannathan, and Runkle (1993) and Whitelaw (1994). Table 1 of Scruggs (1998) summarizes that
empirical evidence.

2

model the conditional mean and volatility of stock returns, Brandt and Kang (2004) find
a negative conditional correlation between innovations to those conditional moments but a
positive unconditional correlation due to significant lead-lag correlations. Lundblad (2007)
reports a positive tradeoff over a long time period, Engle and Lee (1999) find a positive relationship between return and the permanent volatility component, and Ghysels, Santa-Clara,
and Valkanov (2005) find a positive tradeoff using a mixed data sampling (MIDAS) approach
to estimate variance. Pastor and Stambaugh (2001) find evidence of structural breaks in the
risk-return relationship.3
Our paper investigates a conditional version4 of the risk-return specification. We exploit improved measures of ex post variance and incorporate them into a new component
forecasting model in order to implement a time-varying risk model of the equity premium.
How do we achieve better forecasts of a time-varying market equity premium? Firstly,
we use a nonparametric measure of ex post variance, referred to as realized volatility (RV).
Andersen and Bollerslev (1998) show that RV is considerably more accurate than traditional
measures of ex post latent variance. Due to data constraints for our long historical sample,
in this paper we construct annual RV using daily squared excess returns.5
Secondly, as in Andersen, Bollerslev, Diebold, and Labys (2003), French, Schwert, and
Stambaugh (1987) and Maheu and McCurdy (2002), our volatility forecasts condition on
past realized volatility. Since RV is less noisy than traditional proxies for latent volatility, it
is also a better information variable with which to forecast future volatility.
Thirdly, we propose a new volatility forecasting function which is based on exponential
smoothing. Our model inherits the good performance of the popular exponential smoothing
filter but allows for mean reversion of volatility forecasts and targeting of a well-defined
long-run (unconditional) variance. This feature adds to the parsimony of our forecasting
function, which is important in our case given the relatively low frequency data necessary to
allow estimation over a long time period. It also allows for multiperiod forecasts.
Fourthly, motivated by the component-GARCH approach of Engle and Lee (1999) applied
to squared returns, we extend our conditional variance specification, which conditions on past
RV, to a component forecasting model. This flexible conditioning function allows different
decay rates for different volatility components. We also investigate whether or not total
market risk or just some component of it is priced, that is, we allow our risk-return model to
determine which components of the volatility best explain the dynamics of the equity risk
3

There is also a literature which investigates a nonlinear relationship between market risk and return,
for example, Pagan and Hong (1990), Backus and Gregory (1993), Whitelaw (2000) and Linton and Perron
(2003).
4
As discussed below, our conditional parameterization is motivated by the intertemporal asset pricing
models of Campbell (1993) and Merton (1973), as well as the component model of Engle and Lee (1999).
5
Early empirical applications of this measure at low frequencies, for example using daily squared returns
to compute monthly volatility, included Poterba and Summers (1986), French, Schwert, and Stambaugh
(1987), Schwert (1989), Schwert and Seguin (1990) and Hsieh (1991).

3

Again. Building on Engle and Lee (1999). the conditional variance of excess returns is obtained as the conditional expectation of the RV process. long samples of historical information are useful as conditioning information and contribute to improved estimates of the time-varying market premium. equity market. our models can generate multiperiod premium forecasts. The 2-component conditional variance specification provides a superior variance forecast. Exploiting our mean-reverting multiperiod variance forecasts. For example. The equity premium varies considerably over time and confirms that the average excess return associated with subperiods can be misleading as a forecast. We analyze a long. we focus on a multiple component formulation of our new volatility forecasting function in order to allow components of volatility to decay at different rates and to investigate which component is priced. Section 3 details our results on the significance of the riskreturn relationship for several model specifications. we generalize the univariate risk-return model for the market equity premium by estimating a bivariate stochastic specification of annual excess returns and the logarithm of RV. in one of our parameterizations. Nevertheless. produces precise estimates of the risk-return relationship. This involves several new contributions. Furthermore. our Figure 9 shows how well our forecasts captured the declining equity premium in the mid-1990s. which only requires one parameter per volatility component. one component tracks longrun moves in volatility while another captures the short-run dynamics. The paper is organized as follows. low-frequency dataset and show that our models produce realistic time-varying premium forecasts over the entire 1840-2006 time period. Our empirical results show that for 167 years of the U.premium. there is a significant positive relationship between market risk and the market-wide equity premium. and the range of implied premiums that the models produce. The forecasts of a time-varying premium match important features of the data. Finally. We focus on the dynamics of the premium over the 1840-2006 period. 4 . Section 2 introduces the models that motivate our empirical study. and show how mean reversion can be imposed in the model to target the unconditional mean of RV. we use improved measures of volatility in a parsimonious forecasting model which allows components of volatility with different decay rates to be priced in a conditional risk-return model. We introduce a new weighting function on past RV. We discuss the importance of volatility components. In this case. it is the long-run component in the variance that provides a stronger risk-return relationship. In our 2-component specifications of the conditional variance. Our volatility specification.S. In summary. Finally. multiperiod forecasts are available from the assumed dynamics of the bivariate process. and discusses the importance of the measurement and modeling of the variance of market returns. Section 4 summarizes the results and future work.

6 Section III of Campbell (1993) provides conditions that produce this conditional market risk-return relationship. Sufficient conditions are if the investment opportunity set is essentially constant. the intertemporal model implies a market risk-return relationship with no additional factors. market risk is captured by the variance of the market portfolio. then the latter will be a sufficient statistic for market risk. Under some assumptions.t 6 (2. Campbell (1993) provides a discrete-time intertemporal model which substitutes-out consumption. 5 . As in Merton (1973). empirically plausible. The IAPM of Merton (1973) relates the expected market return and variance through a representative agent’s coefficient of relative risk aversion and also allows sensitivity of the market premium to a vector of state variables (or hedge portfolios) which capture changing investment opportunities. most of the explained cross-sectional variation in returns is explained by cross-sectional variation in the assets’ covariances with the market return. or if the representative investor has logarithmic utility.2 The Risk-Return Model 2. if the coefficient of relative risk aversion is equal to 1 or if the investment opportunity set is constant or uncorrelated with news about future market returns. or if the change in consumption in response to a change in wealth is much larger than that associated with a change in other state variable(s). Merton (1980) argues that this case will be a close approximation to the intertemporal asset pricing model in Merton (1973) if either the variance of the change in wealth is much larger than the variance of the change in the other factor(s).t ) = γ1 σM. However. the Campbell (1993) derivation provides an alternative.1 Background Both static and intertemporal models of asset pricing imply a risk-return relationship. In this case. that is. the expected market premium is a function of its variance as well as its covariance with news (revisions in expectations) about future returns on the market. the expected market premium will only be a function of the market return variance. condition under which that market risk-return relationship obtains.1) Campbell (1996) reports empirical evidence in support of an analogous condition in a cross-sectional application for which the restriction is that covariances of all asset returns with news about future returns on invested wealth are proportional to their covariances with the current return on wealth. In this case. and also the conditional CAPM. Examples of intertemporal models which do not require consumption data are the intertemporal asset pricing models proposed by Merton (1973) and Campbell (1993). If the covariance of the market return with news about future investment opportunities is proportional to the variance of the market return. This motivates a risk-return model 2 Et−1 (rM.

this measure of volatility is very noisy and of limited use in assessing features of volatility such as its time-series properties. termed realized volatility (RV). Merton (1980) argues that one should impose the prior that the expected equity premium is non-negative. Bollerslev. 6 . For ease of notation. and Diebold (2004) and Barndorff-Nielsen.1) were the risk-return model. As shown by Andersen. we employ a nonparametric measure of volatility. Bollerslev. and Shephard (2004). A traditional proxy for ex post latent volatility has been squared returns or squared residuals from a regression model. RV is computed as the sum of squared returns over this time interval. Graversen. developed in a series of papers by Andersen. and Labys (2001). and σM. This model implies a proportional relationship between the market equity premium and its conditional variance. Thus RV is a consistent estimate of ex post variance for that period. and whether we are referring to a variance or a standard deviation. Recent reviews of this growing literature are Andersen.1 Measuring Volatility In this paper. Bollerslev. we discuss how we measure and then forecast the volatility which drives the time-varying risk premiums. the forecast of the variance conditional on time t − 1 information). as the sampling frequency is increased. In this paper. If Equation (2. we also drop the subscript M on the market excess return and its conditional 2 variance so that henceforth rt ≡ rM. we use the term volatility to refer generically to either the variance or standard deviation. The asymptotic distribution of RV has been studied by Barndorff-Nielsen and Shephard (2002a) who provide conditions under which RV is also an unbiased estimate. 2. we refer specifically to whether it is an ex post (realized) measure or a conditional estimate (forecast). As discussed in Section 1 above.t ) is the conditional expectation of the excess return on the market. Note that. and Barndorff-Nielsen and Shephard. throughout the paper. As shown by Andersen and Bollerslev (1998).t and σt2 ≡ σM. The increment of quadratic variation is a natural measure of ex post variance over a time interval. Diebold.t . this would be satisfied as long as γ1 > 0. Better measures of ex post latent volatility are available.2. we use a measure of ex post variance. Diebold and co-authors. the sum of squared returns converges to the quadratic variation over a fixed time interval for a broad class of models.2 where Et−1 (rM.2 Measuring and Forecasting Volatility In this section.t is the conditional variance of the market excess return (that is. 2. Where necessary for clarity.

2. (2. According to our test equations. Our volatility forecasts condition on past RV.Defining RV of the market return for year t as RVt .4) . estimates can be derived from the recursion 2 σt2 = (1 − α)RVt−1 + ασt−1 . We analyze whether computing annual RV prior to 1886 from lower frequency intra-period observations has any effect on our results. Given that RV has a superior signal-to-noise ratio for measuring latent volatility. conditional on information at time t − 1.2) j=1 where rt. and σt2 is the conditional variance. except for the fact that the standard practice is to smooth on lagged squared returns rather than on lagged RV. α is the smoothing parameter.3) implies the following weighting function on past RV. parsimony is an obvious objective. that is 2 . This allows us to use a longer time period. A small value of α puts more weight on the the most recent observable value of RV. Therefore. we construct annual RV as RVt = Dt X 2 rt. Prior to 1886.j (2. it should be a superior conditioning variable for forecasting future volatility. The recursion in Equation (2. σt2 = (1 − α) ∞ X j=0 in which the weights sum to one. RVt is computed from squared monthly continuously compounded returns since we only have monthly data for that early part of our sample. One candidate is the infinite exponential smoothing function. Conversely.3) in which 0 ≤ α < 1. and Dt is the number of days in year t for which market equity returns were available. 2. (2. In this case. and less weight on the past forecast σt−1 on recent observations and more weight on past forecasts which smooths the data. the latter is measured by the conditional variance of market excess returns. That is.2 Forecasting Volatility Our time-varying risk model of the equity premium is forward looking. This model is analogous to the popular RiskMetrics filter. 7 αj RVt−j−1 . What functional form should be used to summarize information in past RV? Given our limited data measured at annual frequencies. the expected market equity premium is a function of market equity risk.j is the continuously compounded return for the j th day in year t. we need a forecast of the time t volatility. an α close to 1 puts less weight RVt−1 .

8 Meddahi (2002) and Barndorff-Nielsen and Shephard (2002b) also discuss the theoretical relationship between integrated volatility and RV. 8 . 10 To make the mean reversion in variance forecasts clear. RVt = ω + t Pτ 2 j−1 .6) This unconditional variance leads to our strategy of variance targeting by setting σ 2 from the data and using 2 ω = σ [1 − (1 − α) τ −1 X αj ]. note that with the assumption Et−1 (RVt ) = 2 σ . This is problematic for maximum-likelihood estimation of the smoothing parameter.Although infinite exponential smoothing provides parsimonious estimates. where vt = RVt − σt . we assume that the conditional expectation of annual quadratic variation (QVt ) is equal to the conditional variance of annual returns. (2. and. Given our long time period and low sampling frequency. this new specification is similar in spirit to exponential smoothing but allows 7 In the empirical applications below. To circumvent these problems. the dynamics of RV implied by Equation (2. filters with no intercept and whose weights sum to one are degenerate in the asymptotic limit in the sense that the distribution of the variance process collapses to zero variance asymptotically.5) j=0 where we truncate the expansion at the finite number τ . The smoothing parameter α can be estimated using an objective criterion. and βj = (1 − α)α 2 forecasts can be derived from Et σt+i = Et RVt+i .5) follow the ARCH-like AR representation. Et−1 (QVt ) = Vart−1 (rt ) ≡ σt2 . as Nelson (1990) has shown in the case of squared returns or squared innovations to returns. as i → ∞. Bollerslev. we have Equation (2. we fixed τ = 40 due to presample data requirements.9 Assuming that RV is a unbiased estimator of quadratic variation it follows that Et−1 (RVt ) = σt2 . Diebold. σt2 = ω + (1 − α) τ −1 X αj RVt−j−1 .7) j=0 In summary.6).7 In this specification the weights sum to less than one allowing mean reversion in volatility forecasts as shown below. 9 We assume that any stochastic component in the intraperiod conditional mean is negligible compared to the total conditional variance. we prefer to estimate the smoothing parameter α rather than taking it as given as is common practice with filtering approaches such as RiskMetrics. Conditional variance j=1 βj RVt−j + vt . and Labys (2003).5) as10 E(σt2 ) = ω P −1 j .8 that is. and we can derive the unconditional variance associated with specification (2. 1 − (1 − α) τj=0 α (2. (2. We provide some robust checks for alternative choices of τ . but still retain the parsimony and accuracy of exponential smoothing. it possesses several drawbacks. However. such as maximum likelihood. with Et−1 vt = 0. we propose the following volatility specification. Based on Corollary 1 of Andersen.

mixtures of regimes. Relative to component-GARCH models. k..12 This can be achieved with a component volatility function which estimates the conditional variance as the average of two or more components. and Tauchen (2003). Alizadeh. past data receive an exponentially declining weight. for example. 2 i 2 can be written as σt = ω/(1 − β) + α i=0 β ²t−1−i which requires an extra parameter compared to (2. Santa-Clara.1) σt2 = ω + α²2t−1 + βσt−1 . Indeed.8) where 2 σt. and Valkanov (2005) use a mixed data sampling (MIDAS) approach to estimate volatility. 12 Several literatures. Santa-Clara. Maheu and McCurdy (2000). k i=1 t. Chacko and Viceira (2003) and Ghysels. Gallant. besides being an efficient estimate of ex post volatility. our modeling assumptions are on the annual conditional variance process given observations on RV. As we will see below. long memory.. and Valkanov (2006).1) model for short-run volatility dynamics which mean revert to a non-parametrically estimated time-varying unconditional volatility. Related work on volatility modeling includes the component model of Engle and Lee (1999) and Ghysels.9) j=0 Note that the conditional variance components are projections on past RV. and Diebold (2002). i = 1. (2. . using data from 1928-2000. have highlighted that a single exponential decay rate is inadequate to capture volatility dynamics over time. our parameterization only requires 1 parameter per component rather than two. Another difference is that we smooth on past annual RV. Our specification is also more parsimonious that the covariance-stationary GARCH(1. Engle and Rangel (2004) develop a two-component volatility specification with a GARCH(1. 11 2 The covariance-stationary GARCH(1. Formally. A simple approach to providing a more flexible model is to allow different components of volatility to decay at different rates. P∞ model. Examples include Engle and Lee (1999). the finite unconditional variance allows for variance targeting which means that only one parameter needs to be estimated.(k) 1X 2 =ω+ σ . 9 .i = (1 − αi ) τ −1 X αij RVt−j−1 . Therefore.5). Chernov.1) model. As is evident from Equation (2. Santa-Clara. and Valkanov (2005).11 As discussed further below. etc.5). Ghysels. is that it will be a consistent estimate of volatility for a very large class of empirically realistic models. In addition. this is not flexible enough to capture the time-series dynamics of RV. define the k-component volatility model as k 2 σt. the monthly conditional variance of returns is modelled using a flexible functional form to estimate the weight given to each lagged daily squared return. Bollerslev and Zhou (2002). with ²t−1 ≡ rt−1 − Et−2 rt−1 . one of the attractions of using RV.for mean reversion in volatility forecasts. In that paper.. Brandt. at least for our sample of annual data the more parsimonious specification is critical for precision of the estimates of the risk-return relationship and also for generating reasonable premium estimates. mixtures of jumps and stochastic volatility. Ghysels.. We do not specify the high-frequency dynamics of spot volatility.i (2.

The conventional univariate test equation for (2.10) For example. our empirical models based on Equation (2. We allow for alternative components of volatility with different decay rates. wj (κ1 . which might account for a constant equity premium unrelated to risk. rather than the full conditional variance. (2. Scruggs (1998) motivates the addition of an intercept to account for market imperfections. In the next section. This joint stochastic specification of returns and RV allows for multiperiod forecasts of the premium. that is. Our objective is to have a parsimonious and flexible function that summarizes information in past RV that might be useful for forecasting changes in the market equity risk premium. This allows us to use variance targeting which may be important to gain precision in our application. and Valkanov (2006) also estimate market risk based on 14 10 . In contrast. Sinko. and Valkanov (2005) interact (multiplicatively) in a way that makes it difficult to isolate their separate effects on pricing. in our case with two components.14 Our conditional variance specification maintains the parsimony of smoothing models but allows mean reversion. In this subsection we introduce two alternative empirical specifications of the risk-return relationship. Each of our models jointly estimate the conditional mean and conditional variance parameters using maximum likelihood. That is.8). allowing us to separate out and price short-lived versus slower-decaying components. Each of the empirical models implies a time-varying equity premium which is a function of its own conditional second moment. rt = γ0 + γ1 σt2 + ²t . the coefficient on RVt−j−1 is (1 − α1 )α1j /2 + (1 − α2 )α2j /2. κ2 ) = exp(κ1 j+κ2 j 2 ) P251 i=0 exp(κ1 i+κ2 i2 ) . We label the first specification univariate since it fits the stochastic excess return process by conditioning on variance forecasts which are estimated using a projection on past RV as in Equation (2.15 The second specification 13 Their monthly variance is Vt = 22 P251 j=0 2 wj rt−j . 2.13 Our decay rates are additive.1) are motivated as special cases of an IAPM. is more important in driving the market premium.3 The Empirical Risk-Return Models As discussed in Section 2.They find that a two parameter filter works well. we extend the existing literature to investigate a bivariate risk-return specification. such as differential tax treatment of equity versus T-bill returns.1. but it also allows us to investigate whether or not a particular component. Not only is this a more flexible way to capture the time-series dynamics of volatility. 15 The risk-return regressions in Ghysels. 2 Note that their weight on past squared returns is proportional to exp(κ1 j + κ2 j 2 ) = κ˜1 j κ˜2 j with κ˜1 = exp(κ1 ) and κ˜2 = exp(κ2 ). Santa-Clara.1) is a linear model. the smoothing coefficients in Ghysels. a forecast of the equity premium’s time t variance conditional on time t − 1 information.

On the other hand. in which q denotes the number of volatility components that affect the market premium when the conditional variance follows a k-component model.. we set q = k. zt ∼ N (0.(q) )2 1 1 2 . k = 2. 2. lt = − log(2π) − log(σt.9). .(k) zt . 1). for models in which q = 1. or applied to log(RV) as in k 2 log σt. α1 .(q) in the conditional mean.(k) 2 log σt. Note that q ≤ k. The loglikelihood with respect to the parameters γ0 .14) PT t=1 lt is maximized Bivariate Risk-Return Specification In this case.3.(k) 2 where σt. γ1 .(·) depends on the levels or log specification. 2. i = 1. That is. (2.12) αij log RVt−j−1 ...(k) zt ..is bivariate since we estimate a bivariate stochastic specification of annual excess returns and log(RV).3. 2 rt = γ0 + γ1 σt. 1) past RV. ²t = σt.i 1X 2 = ω+ log σt. as in Equations (2. zt ∼ N (0.. The t-th contribution to the loglikelihood is 2 (rt − γ0 − γ1 σt. . ²t = σt. where we want the total variance to enter the mean.(q) + ²t .1 Univariate Risk-Return Specifications The conditional mean is. k i=1 = (1 − αi ) τ −1 X (2. (2. k.(k) ) − 2 2 2 2σt. The parameterization has conditional mean 2 rt = γ0 + γ1 σt. 11 (2. we specify σt.2 (2. For example.8) and (2.15) . we estimate a bivariate stochastic specification of annual excess returns and log(RV). αk .i ..11) and the conditional variance is either applied to levels of RV. the conditional variance of excess returns is obtained as the conditional expectation of the RV process.(q) + ²t .13) j=0 where k indexes the total number of variance components and q indexes the number of 2 variance components that affect the conditional mean. In that case. we let the maximum likelihood estimator determine which component of the variance is optimal in the mean specification.

(k) ) − log(φ2 ) 2 2 P 2 2 2 2 (rt − γ0 − γ1 σt.. As noted in Section 2.i + ηt .18) j=0 2 ≡ Et−1 (RVt |q).16) k log RVt = ω + 1X 2 log σt. Second. as shown by Merton (1980) for i. sampling more frequently (for example.(k) ≡ Et−1 (RVt |k) = exp(Et−1 (log RVt |k) + . However. 1 1 2 log(σt.16). i = 1. 2 2σt. k i=1 2 log σt. In other words. collecting daily instead of monthly data) will not improve the precision of the market premium estimates.(k) 2φ2 lt = − log(2π) − (2.5Vart−1 (log RVt |k)) (2. 12 .. lt consists of contributions from the return and volatility equation.17) (2. we will need data at a higher frequency than our estimation frequency in order to compute realized volatility (RV) as in Equation (2.i ) − − . 3 3.(q) ) (log RVt − ω − k1 ki=1 log σt. (2. represents the conditional variance component(s) Again.1 above. since we use a nonparametric measure of realized volatility which is constructed as the sum of squared intraperiod returns. we require data on a broad equity market index and on a riskfree security so that we can construct returns on the equity index in excess of the riskfree rate.i. . k. q ≤ k. we use monthly returns from 1802-1885 and daily returns from 1886-2006 to compute annual RV.2).19) In this case. First. There are two considerations. For this reason we want our historical sample to be as long as possible..d. we explicitly model the annual log(RV) process as stochastic given the most recent information. returns. we can only increase the precision of expected return estimates by increasing the length of the time period (calendar span) covered by our historical sample.1 Results Data and Descriptive Statistics We are evaluating the risk-return relationship associated with equity for the market as a whole. by parameterizing the joint density of annual excess returns and log(RV). making a conditional log-normal assumption for RV allows for a convenient calculation of the conditional variance as in Equation (2. σt.2.and the following conditional variance specification 2 σt.(q) that affect the conditional mean. Therefore. φ2 ) αij log RVt−j−1 .i = (1 − αi ) τ −1 X ηt ∼ N (0.

Notice how much smoother the square root of RV is than the absolute value of annual excess returns. Starting in 1840 provides presample values to condition our time-varying volatility model. we use monthly bid yields associated with U. capital budgeting versus asset pricing. Annual continuously compounded equity returns are computed as the sum of the monthly continuously compounded returns.18 Figure 1 plots annual market excess returns for the entire sample from 1803-2006. our focus is on forecasting the market equity premium over a (approximately) riskfree rate rather than over an alternative risky asset. The average excess return is 4. for example. 16 Schwert adjusts for the time averaging in the original series for 1863-1885 and adds an estimate of dividends for the period 1802-1870.S. The equity return data are converted to continuously compounded returns by taking the natural logarithm of one plus the monthly or daily equity return.The U. Annual excess returns are then computed by subtracting the annual riskfree rate. Henceforth. The annual riskfree yield was obtained by summing the 12 monthly yields. 18 We use continuously compounded returns to conform with our estimate of RV.S. market equity returns from 1802-1925 are from Schwert (1990). It is also clear from these plots that the data for the period 1803-1834 have a very different structure than that for the remainder of the sample. The Schwert equity index was constructed from various sources: railway and financial companies up to 1862 from Smith and Cole (1935). 3-month T-Bills from the Fama monthly riskfree file provided by CRSP. unless otherwise indicated. 13 . we use returns (including distributions) for the value-weighted portfolio of NYSE. Table 1 reports summary statistics for the time period. NASDAQ and AMEX stocks compiled by the Center for Research in Security Prices (CRSP). 1840-2006. our analyses focus on the time period 1840-2006. Booth (1999) compares several approaches and shows that for certain periods the risk associated with long-term Treasuries was almost as large as that associated with market equity. the Macaulay (1938) railway index (1863-1870). Figures 2 and 3 plot two alternative measures of ex post volatility. respectively. Both Schwert (1989) and Pastor and Stambaugh (2001) drop this period. 17 The market excess return is often measured with respect to a long-term (for example.17 For the period 1926-2006. As noted in Schwert (1990). data from only a small number of companies was available for that subperiod. Although this choice depends on investment horizon and the purpose of the calculation. the absolute value of the excess returns and the square root of RV. Siegel (1992) describes how these annual yields were constructed to remove the very variable risk premiums on commercial paper rates in the 19th century. and the Dow Jones index of industrial and railway stocks for 1885-1925.27%. returns and excess returns refer to continuously compounded rates. For the same reason. the value-weighted market index published by the Cowles Commission for the time period 18711885. Annual bill yield data (riskfree rate) for the 1802-1925 period are from Jeremy Siegel.16 For the 1926-2006 period. 30-year) Treasury yield rather than the T-Bill yield used in this paper.

On the other hand. which implies that it is more influenced by the most recent RV. Note that our models with an intercept all have a positive slope parameter so that restricting the statistically insignificant intercept to zero does not force the slope parameter to be positive but rather allows us to estimate it with more precision.7). σ 2 . U(2) and U(2s ).20 Models U(2) and U(2s ) have the same volatility specification but U(2s ) allows the risk-return model to determine which volatility component has the most explanatory power for the dynamics of the equity premium. based on results showing low power of a Wald test when an unnecessary intercept is included. as expected from the smoothing coefficient estimate α ˆ 1 = .9) for k = 1.396. For linear parameterizations of the premium. as in Equation (2.3. We find that. To avoid influential observations in our small sample. when given the choice. It is clear from Figure 4 that the 2-component volatility specification tracks RV better than the 1-component version.12) and (2.21 Figure 5 plots the individual components of volatility from the U(2) specification. and the natural logarithm of RV. we report both the linear parameterization of the risk-return relationship. to the sample median of annual RV over the period 1802-2006.19 Due to the importance of the length of the time period necessary for efficient conditional mean forecasts. as in Equations (2. √ 21 Note that these volatility plots are square roots of the volatility forecasts versus RVt . See Inoue and Kilian (2004) for evidence in favor of using in-sample forecasts for model evaluation.11). Note in particular how long it takes the smooth component to decay from its high level in the 1930s. that is. 3. and Ebens (2001) show that for equity returns the natural logarithm of RV is more normally distributed than the level of RV. For example. 20 See Lanne and Saikkonen (2006) and Bandi and Perron (2006) for statistical arguments in favor of estimating the proportional model when the estimated intercept is insignificant. This shows that the smooth component is persistent. Bollerslev.950 reported in Table 2. U(1). Equation (2. 2.1).2. the univariate specifications fit the stochastic excess return process by conditioning on variance forecasts which are estimated using a projection on past RV. as motivated by Equation (2.8) and (2. Note that ω is computed. Equation (2. the second component has much lower persistence. 14 . as in Equations (2. For each model.3.1 Using RV Levels Table 2 reports parameter estimates and likelihood ratio test (LRT) results associated with the risk-return relationship estimated using the alternative component volatility models applied to levels of RV.11) and the proportional parameterization with γ0 = 0.1. Lanne and Saikkonen (2006) conclude with a recommendation to always restrict the intercept to zero if it is implied by the theory being tested. the maximum likelihood estimator always chooses to price the long-run or smooth component. we use all of our data and thus forecasts are in-sample.11).2 Univariate Risk-Return Results As discussed in Section 2. we set the long-run target volatility. the LRT results reveal a statistically significant 19 Andersen. Table 2 reveals that the risk-return relationship is positive for all specifications. α ˆ 2 = . Diebold. We report results using both the levels of RV.

the regression is RVt. as do the LRT statistics which are more than twice as large for the proportional as opposed to the linear model. pricing the smooth component of volatility (the U(2s ) model) results in a lower peak which decays more slowly.06) for the U(2) case. Note that the U(2s ) model. In this case. The LRT statistics in the final column of the table all reject the null hypothesis that γ1 = 0.5 for the proportional model for all of the alternative volatility specifications. 15 . the more parsimonious proportional premium model cannot be rejected by the data. As noted above. also has the best premium fit whether measured from the perspective of MAE. there is a high peak in the forecasted market equity premium which lasts for 2 years.22 The results in Table 3 support the 2-component models U(2) and U(2s ) over the more restrictive 1-component version U(1). as well as the R2 from Mincer and Zarnowitz (1969) forecast regressions for a particular model’s premium and volatility forecasts.64. Table 3 reports model diagnostics and statistics for volatility and premium fit. That is. The t-statistic for γ1 also supports this conclusion in that it is greater than 3. 22 For example for the volatility forecasts. which had the highest log-likelihood. is highest (2. This result is particularly strong for the proportional parameterizaton of the risk premium in which case the test statistics are all greater than 12 so that the p-value associated with the null hypothesis is very small. the premium forecast ranges from 1. Figure 6. In contrast. That version also has the highest log-likelihood. γ1 . for all volatility specifications.2% with the average forecasted premium over the 1840-2006 sample equal to 4.624) for the U(2s ) model.29 percent. plots the market equity premium forecast for the proportional risk premium parameterization. the tstatistic on the risk-return slope coefficient. although it is marginal (p-value = 0.7% to 9. RMSE. For example. the precision of the estimate γˆ1 improves. or R2 . and a LRT statistic of 7. The one exception is the U(2s ) model which reveals a strongly positive market risk-return relationship for both the linear and the proportional parameterization of the premium.54 which rejects the hypothesis that γ1 = 0. both from the perspective of t-tests on γ0 and from the LRT reported in the 2nd-last column of Table 2. When total volatility from the 2-component volatility specification is priced (the U(2) model). the intercept in the conditional mean is statistically insignificant. the R2 increases from about 7% to 26% by adding a second volatility component. the t-statistics associated with the hypothesis γ1 = 0 increase. These include mean absolute error (MAE) and root mean squared error (RMSE).5 = a + bˆ σt + ut in which σ ˆt is the square root of the particular model’s volatility forecast for time t given information up to time t − 1. as suggested by Figure 4.relationship between excess market returns and their conditional variance. For instance. 41. 1933 and 1934. However. by restricting the intercept to be zero. For these linear parameterizations.

15) and (2.23 The parameter estimates using log(RV). has the best overall fit. reported in Table 4. Again.42.05 respectively for the LR test of the restriction γ1 = 0. Figure 7 23 The log transformation reduced the outliers present in the levels of RV.3 Bivariate Risk-Return Results In this section.16) to (2. MAE. Table 5 shows that the 2-component parameterizations of volatility dominate the 1component version from the perspective of all three criteria. The LRT statistics in the final column range from 16. are the linear parameterizations of the premium with 1-component and 2-component volatility specifications. with p-values of 0. Equations (2. 24 16 . The two exceptions. This Figure also shows that these premium forecasts are similar to the univariate case in which volatility forecasts were estimated from past levels of RV rather than log(RV). the Ulog (2s ) model.2. RMSE. Figure 8 shows that the smooth proportional premium forecasts using the univariate model that conditions on log(RV). the proportional parameterizations of the premium do result in very statistically significant postive risk-return relationships for all cases. we generalize our risk-return model for the market equity premium by estimating a joint stochastic specification of the conditional mean and annual log(RV) of market excess returns.11) using variance forecasts which are estimated from a projection on the natural logarithm (rather than levels) of past RV.08 and 0.12) and (2. as well as Figures 7 to 8. γˆ1 . As in the levels case.33% over the period 1840 to 2006. reveal that the risk-return relationship is again positive for all specifications and. range from 1. for volatility fit. However. the Ulog (2s ) model. that is.18). The plots of the premium forecasts for all specifications show that the premiums are increasing from a low level over the first few years of the estimation period.5% with an average of 4. even larger than in Table 2. The t-statistics associated with the risk-return slope parameter estimates. with two exceptions. which prices the smooth component of volatility.13). The results are very comparable to the univariate results reported above.6% to 9.3. is very significant – even more so than when projecting on the levels of past RV. that is. The system of test equations is (2.00 to 18. ω is computed to target long-run volatility. Tables 4 and 5 report these results. and the results are summarized in Tables 6 and 7.97 to 4. or R2 .24 3. range from 3.15 in this case. The logarithmic specification of RV ensures that volatility is non-negative during estimation of the joint stochastic process. except that in the log(RV) case the peaks are slightly higher with faster decay. This is partly due to the low volatility of the presample data from 1802-1834 which the initial part of our estimation period require as conditioning information when τ = 40. for which log σ 2 is set to the 1802-2006 average of log(RV).2 Using Log(RV) We also fit the univariate risk-return model (2.

that the average realized excess return is not a very reliable forecast of the market equity premium since it will be sensitive to the subsample chosen. once again.15% in 1997 to 5.13) for k = 1. The premium forecasts from this bivariate stochastic model range from 1. This suggests that the maximum-likelihood estimator is reliable for our sample which addresses one of the issues raised by Stambaugh (1999) in our context. or R2 . Equation (2. As displayed in Figure 8. the forecasted continuously compounded premium has been increasing from the recent low of 3. MAE. we investigated potential small-sample bias for the maximum-likelihood estimator by simulating 1000 bootstrap samples using the parameter estimates for the B(2s ) model in Table 6 as the DGP. the proportional premium for the preferred B(2s ) specification is similar to that for the univariate case which conditions on past log(RV). that is. they are considerably lower than the 1980-2006 average continuously compounded excess return of 6. Finally.27%. as opposed to −144. the t-statistic associated with γ1 is smaller for the linear risk premium parameterizations B(1) and B(2). either by t-tests or by the LR tests reported in the 2nd-last column of the table.6% to 9.09 for the B(1) model which has a 1-component volatility specification. Table 7 shows that the 2-component volatility specification fits log(RV) better.2%. 2. For example.25 As in the univariate cases.4% in 2004 then back down to 4. The time-varying premium estimates for the period 1990 to 2000 were below the long-run average of 4. 26 We do not have enough annual data to get reliable estimates from a conventional GARCH-in-Mean parameterization using squared returns.77 for the B(2s ) model. This graphically illustrates the point.31%. the LRT results in the final column indicate that the relationship between excess market returns and their conditional variance is very strong for proportional parameterizations of the risk premium.plots the individual components of log volatility from the B(2) specification. all of the premium fit statistics. 17 . reaching a low of 3. Table 7 reports that the preferred model is. Note that all of our specifications deliver reasonable estimates of the market premium. support this conclusion. as does the log likelihood which is −135.26 Figure 9 illustrates the premium for the 1990-2006 period generated by our bivariate specification B(2s ). However.9% in 2006. the R2 for volatility fit increases from about 23% to 28% by adding the 2nd volatility component. that is. In addition. as in Tables 2 and 4. discussed in Section 1. the risk-return relationship is significantly positive for all volatility specifications. The bias for all of the parameters was small.7% with an average of 4. the data support a proportional parameterization of the risk premium. the 2s specification. As in the log(RV) univariate case. since we are unable to reject that γ0 = 0. 25 Given our relatively small number of annual observations. Again. Table 6 reports that the risk-return relationship is positive for all models and statistically significant for all of the models except two. In that case. RMSE. 2 volatility components with the smooth component being priced in the conditional mean.15% in 1997.

5Vart−1 (log RVt ) + ²t . zt ∼ N (0.1) (3.4) j=0 The results for this model for the sample 1840-2006 are repeated in the second panel of Table 8.i 2 i=1 2 log σt. φ2 ) log RVt = ω + log σt. This specification specializes the bivariate system in Section 2. These cases were discussed in the previous subsection and Table 6 as models B(2) and B(2s ) respectively. labelled Model B(2a ). That is.3.3) (3.2 to the case with 2 volatility components (k = 2). These results are summarized in Table 8. The third and fourth panels of Table 8 report results for the B(2s ) model with alternative values of τ . the number of lags used in the exponential smoother components.1 + (1 − γ2 ) log σt. While it is useful to see how the two volatility components are separately priced. less persistent volatility component is not significantly different from zero.5) are reported in the first panel of Table 8. 1) 2 1X 2 + ηt . As compared to the default value.5) Imposing γ2 = 1/2 corresponds to the total variance being priced. the t-statistic on the slope parameter γ1 increases. we can reject that it is zero (the t-statistic is 5. Therefore.i = (1 − αi ) τ −1 X αij log RVt−j−1 i = 1.201. ηt ∼ N (0.035 with a standard error of 0. The estimate for parameter γ2 is 1. To check that our parsimonious B(2s ) specification is able to adequately capture the potential differential effect of the smooth long-run volatility component on the dynamics of the premium.2) (3. On the other hand.4 Robustness Analyses We now check the robustness of the above results to alternative parameterizations and various subsamples.5Vart−1 (log RVt ) + ²t (3.2 + . The bivariate model with the most empirical support is labelled B(2s ). while restricting γ2 = 1 is one component being priced.3.1) with £ ¤ 2 2 rt = γ0 + γ1 exp ω + γ2 log σt. The results for the less parsimonious specification in Equation (3.1 + . 2 (3.(2) zt . the coefficient on the second. As τ is increased. we introduce a new bivariate specification that replaces Equation (3. We have found that 18 . τ = 40. labelled Model B(2s ).149) which confirms the results from all of our earlier 2s specifications that the persistent volatility component has a positive and very significant effect on the dynamics of excess returns. £ ¤ 2 rt = γ0 + γ1 exp ω + log σt. allowing one of those components to affect the conditional mean (q = 1). ²t = σt. notice that our more parsimonious model B(2s ) has an almost identical fit which confirms the robustness of the 2s parameterizations reported earlier. the τ = 5 case results in an insignificant risk-return relationship.

although that indicator variable was significantly different from zero indicating that annual RV was slightly lower over that first part of the sample. the univariate specifications favoured higher values of τ . However. equity over the period 1840-2006 using a time-varying market premium for equity risk. the last panel of Table 8 reports results for model B(2s ) for which annual RV is computed using monthly squared returns for the entire 1840-2006 sample. note that the results for the 1886-2006 subsample refer to the period over which annual RV is always computed from daily squared returns. rather than daily squared returns when they became available in 1886. the B(2s ) parameterization still always dominates a 1-component volatility specification. The next three panels of Table 8 report subsample results for model B(2s ). Therefore. Our conditional variance specification maintains the parsimony of exponential smoothing functions but allows mean reversion in forecasts. Finally. We begin with a univariate specification of the risk-return relationship. Note that as the sample gets shorter. but it also allows us to investigate whether or not a particular component. Also. the positive and significant risk-return relationship still obtains.S. the persistence. we chose a common value that was high enough to capture the smooth volatility component with our limited number of data.the optimal τ is somewhat model and criterion specific. Again. 4 Summary and Future Directions This paper evaluates the market risk-return relationship for U. This application models the stochastic market excess returns by conditioning on variance forecasts which are estimated by projecting onto past RV. α2 . a τ = 25 produces a significantly better loglikelihood (−131. Our default model which uses the latter has a higher loglikelihood but otherwise the results are similar. 19 .57) and higher t-statistic (3. 1928-2006 and 1945-2006.153) associated with γ1 than our base B(2s ) which has τ = 40. rather than the full conditional variance. the risk-return results were similar. This specification which prices the smooth volatility component of a two-component volatility model was able to deliver a positive and statistically significant risk-return relationship for subsamples such as 1886-2006. We assess the robustness of those results by also estimating a univariate version which projects onto past log(RV). is more important in driving the market premium. We propose a parsimonious and flexible function that summarizes information in past RV that might be useful for forecasting changes in the market equity risk premium. For example. We also tried using an indicator variable in the volatility function for the 1840-1885 part of the sample for which daily data were not available. We allow for alternative components of volatility with different decay rates. Not only is this a flexible way to capture the time-series dynamics of volatility. Nevertheless. of the second volatility component goes to zero. Again. and targets the implied long-run variance.

. the preferred model from the perspective of overall as well as equity premium fit..” Econometrica. and Y. 885–905. Labys (2003): “Modeling and Forecasting Realized Volatility. M.” in Handbook of Financial Econometrics. Andersen.In addition. Secondly. and T. T. Diebold. In summary. T. First. in addition to our volatility specification. Bollerslev. and Sohn (2006).. Standard Volatility Models Do Provide Accurate Forecasts. All of the empirical specifications support a conclusion that the relationship between risk and return for the market equity premium is positive. Ghysels. 27 A persistent component in also shown to be important for asset pricing in Bansal and Yaron (2004). Andersen. Bollerslev. 1047–1092. T. we generalize the risk-return model for the market equity premium by estimating a bivariate stochastic specification of annual excess returns and log(RV). see Bansal. Brandt. The higher is the expected market risk. F. by L.” Journal of Finance. the higher the market equity premium. 39(4). and F.” International Economic Review. X. and F. 71. Diebold (2004): “Parametric and Nonparametric Volatility Measurement. and P. S. we allow for increased flexibility by explicitly modeling more than one volatility component. we extend the analysis to a bivariate risk-return model of returns and RV. X. ed. G. we consider two additional contributions to the literature. This relationship is strongest for a two-component specification of volatility with the long-run smooth component being priced in the conditional mean. G. Diebold (2002): “Range-Based Estimation of Stochastic Volatility Models. T. the conditional variance of excess returns is obtained as the conditional expectation of the RV process. T.27 In fact. and Kiku (2007) who investigate the implications for risk and return of risks in the long run versus the short run. we investigate whether or not one volatility component is more important than total volatility in driving the dynamics of the equity premium. In future work. 57. 529–626. Hansen. Finally. Bollerslev (1998): “Answering the Skeptics: Yes. X. In this case. 20 . unlike the univariate specifications that dominate the literature. Andersen. Dittmar.. as in Engle and Rangel (2004) and Engle. Also. is the 2s model which uses a 2-component volatility specification and prices the smooth component of volatility in the conditional mean. Ait-Sahalia. for both the univariate and bivariate specifications. North Holland. References Alizadeh. we plan to explore potential sources of this smooth component of volatility.

. T. 11. and H. DeGennaro (1990): “Stock Returns and Volatility. Zhou (2002): “Estimating Stochastic Volatility Diffusion using Conditional Moments of Integrated Volatility. G. Yaron (2004): “Risks for the Long Run: A Potential Resolution of Asset Pricing Puzzles. Gregory (1993): “Theoretical Relations Between Risk Premiums and Conditional Variances. Shephard (2002a): “Econometric Analysis of Realised Volatility and its Use in Estimating Stochastic Volatility Models. 487–512. Campbell. 41A. 42–55. and N. 253–280.. Bansal. (1996): “Understanding Risk and Return. (1987): “Stock Returns and the Term Structure. Baillie. E. Diebold. P. and H. R. 59(4). T. Bollerslev. L. Backus. 298–345. 12.” American Economic Review. O.” Journal of Finance. 61. 203–214. (1993): “Intertemporal Asset Pricing without Consumption Data. Kiku (2007): “Cointegration and Consumption Risks in Asset Returns. 25. 72. and P. J. Barndorff-Nielsen. O. and B. Kang (2004): “On the relationship between the conditional mean and volatiity of stock returns: A latent VAR approach. R. 373–399.” Journal of Econometrics. 133–143. and N. Shephard (2004): “Power Variation and Stochastic Volatility: a Review and some New Results.” Journal of Applied Probability. E. Series B. Dittmar.. M. and Q. Bollerslev. Bansal.. 109.” Journal of Political Economy. and A. R. Perron (2006): “Long-Run Risk-Return Trade-Offs. (2002b): “Estimating Quadratic Variation Using Realized Variance.” Journal of Financial Economics. G. Barndorff-Nielsen. Ebens (2001): “The Distribution of Realized Stock Return Volatility.” Journal of Financial Economics.. 1481–1509. E. (1999): “Estimating the Equity Risk Premium and Equity Costs: New Ways of Looking at Old Data. T.” Journal of Applied Econometrics. R. X. F. Diebold. F. W. and D. D.Andersen.. Review of Financial Studies. F.. 83. T.” Journal of Financial Economics. T. 177–185.” Journal of Applied Corporate Finance. and R. 217–257. 21 .” Forthcoming. Y. 96.. 43–76.. T. Graversen.” manuscript. X. W..” Journal of Financial and Quantitave Analysis.” Journal of the Royal Statistical Society. 457–477. K. 33–65. Brandt. 17.. Bandi. Labys (2001): “The Distribution of Exchange Rate Volatility. S.” Journal of the American Statistical Association. 18. 100–112. and A. Booth. Andersen. Bollerslev. 64. 104.” Journal of Business & Economic Statistics. M.

F. Y. R. and G. 279–294. 225–257. Hentschel (1992): “No News is Good News: An Asymmetric Model of Changing Volatility in Stock Returns. 3–29..Campbell. Viceira (2003): “Spectral GMM estimation of continuous-time processes. Robins (1987): “Estimating Time Varying Risk Premia in the Term Structure: The ARCH-M Model. 22 ...” Manuscript. F. R. M. L. E. 116. 55. Claus. Engle.. Chernov. D.. 3. F.” NYU and UCSD. Tauchen (2003): “Alternative models for stock price dynamics. ed. (2006): “Predicting Volatility: How to Get the Most Out of Returns Data Sampled at Different Frequencies. R. by Engle.. and R. and R. F.” Journal of Econometrics. G. and R.” Journal of Applied Econometrics. Kamstra. A. Manuscript. Engle. Y.” in Cointegration.” Journal of Finance. 637–659. E. G. 1629–1666. E. Engle. R. Sinko.. Stambaugh (1987): “Expected Stock Returns and Volatility. Valkanov (2005): “There is a Risk-Return Tradeoff After All. and J. 76. 56(5).” Journal of Finance. Causality. 391–407. and L. and White. G. A. R. 57(2). Sohn (2006): “On the Economic Sources of Stock Market Volatility. P. 427–465. Lee (1999): “A Long-Run and Short-Run Component Model of Stock Return Volatility. (2002): “The Equity Premium..” Econometrica. Fama. 131. E.” Journal of Finance. and L. and Forecasting. F.. R. R. 47. Gallant.” Journal of Econometrics... M. Ghysels. (1988): “Volatility Persistence and Stock Valuations: Some Empirical Evidence Using GARCH. and R. and L. Lilen. Valkanov (2006): “MIDAS Regressions: Further Results and New Directions. and G.” Journal of Financial Economics. E. Kramer (2004): “Stare Down the Barrel and Center the Crosshairs: Targeting the Ex Ante Equity Premium. 31. P. Ghysels. Ghysels. 281–318. Chacko. and K. Thomas (2001): “Equity Premium as Low as Three Percent? Evidence from Analysts’ Earnings Forecasts for Domestic and International Stock Markets. 116. Ghysels.. and B.” Journal of Financial Economics. 259–292. R.” Forthcoming in Econometric Reviews. French. Santa-Clara. W. Rangel (2004): “The Spline GARCH model for Unconditional Volatility and its Global Macroeconomic Causes. York University. 59–95. K. 19. G.. Donaldson. Chou. F. M.” Journal of Financial Economics. G. 509–548. and J. Oxford University Press. J. Schwert. French (1992): “The Cross-Section of Expected Stock Returns. R. Engle.” manuscript.” Journal of Econometrics. J.

. and M. 118–125. C. (1980): “On Estimating the Expected Return on the Market: An Exploratory Investigation. Kilian (2004): “In-sample or out-of-sample tests of predictability: Which one should we use?. Saikkonen (2006): “Why is it so difficult to uncover the risk-return tradeoff in stock returns?. and B.-J..” Economics Letters.. NBER. Ljung. E. and T. Inoue. Hsieh. R. 867– 888.. C. 21(3). 61(3). and R. O. Meddahi.” Journal of Financial Economics. Kim. 85. Jagannathan. R. 7(3-4). Nelson (2004): “Is there a positive relationship between stock market volatility and the equity premium?. Kiefer. 323–361. Lundblad.” Journal of Applied Econometrics. (2002): “Nonlinear Features of FX Realized Volatility.. Linton. (1938): The Movements of Interest Rates. 23 .” Review of Economics and Statistics. Guo. 479–508. Merton. McCurdy (2000): “Volatility Dynamics under DurationDependent Mixing.. and Banking. Lanne. F..” Journal of Financial Economics. (1991): “Chaos and Nonlinear Dynamics: Applications to Financial Markets.” Journal of Finance. L. 354–367. and D. Whitelaw (2006): “Uncovering the Risk-Return Relation in the Stock Market. 36(3).” Journal of Empirical Finance.” Biometrika. Salmon (1983): “Testing Normality in Econometric Models. C. 17. H.” Journal of Finance. Perron (2003): “The Shape of the Risk Premium: Evidence from a Semiparametric Generalized Conditional Heteroscedasticity Model?. R. 339–360. 345–372. Runkle (1993): “On the Relation between the Expected Value and the Volatility of the Nominal Excess Return on Stocks. 84(4). (2007): “The risk return tradeoff in the long-run: 1836-2003. 668–681. 23(4). A. L. R. 371–402..” Economics Letters. (1973): “An Intertemporal Asset Pricing Model. 297–303. Maheu. R. M.” Econometrica.” Journal of Money. 67. J. and C. Macaulay. D. Harvey. 41. 289–317.” Journal of Financial Economics. 123–127.” Journal of Business & Economic Statistics. 8.” Econometric Reviews. C. 1779–1801. M. Box (1978): “On a Measure of Lack of Fit in Time-Series Models. 46. 92. N. 48(5).” Journal of Finance. (2002): “A Theoretical Comparison between Integrated and Realized Volatility. Morley. H. R. J.Glosten. 123–150. Credit. C. and P. and L. 24.. Bond Yields and Stock Prices in the United States since 1856. N. and G. 1433–1463. 11. 1839–1877. (1989): “Time-varying conditional covariances in test of asset pricing models.

. and Learning in the Stock Market. Whitelaw. 1790-1860. by Mincer. J. National Bureau of Economic Research. F.” Forthcoming. 49..” Journal of Finance. J. (1990): “ARCH Models as Diffusion Approximations. R. Stambaugh.Mincer. 341–360. Poterba. G. ed. 4. Scruggs.” Journal of Financial Economics. Summers (1986): “The persistence of volatility and stock market fluctuations. Journal of Finance. 45. 515–541. J. 13. and B. (1999): “Predictive Regressions... J. Whitelaw.” Journal of Econometrics. 1142–1151. Pastor. ed. Nelson (1989): “A Markov Model of Heteroskedasticity. Seguin (1990): “Heteroskedasticity in Stock Returns. Hong (1990): “Non-Parametric Estimation and the Risk Premium. 24 . G. 7–39. W.” Journal of Finance. R. Powell. (1989): “Why Does Stock Market Volatility Change Over Time?.” in Nonparametric and Semiparametric Methods in Econometrics and Statistics. Stambaugh (2001): “The Equity Premium and Structural Breaks. A. and R. by W. Swaminathan (2007): “Estimating the intertemporal risk-return tradeoff using the implied cost of capital. Risk. 53. Schwert. Schwert.” Journal of Finance. (2000): “Stock Market Risk and Return: An Equilibrium Approach.” in Economic Forecasts and Expectations. 227–252. 54.K. Tauchen. R. and V. 45. Sinha. Zarnowitz (1969): “The Evaluation of Economic Forecasts and Expectations. (1998): “Resolving the Puzzling Intertemporal Relation between the Market Risk Premium and Conditional Market Variance: A Two-Factor Approach. B. and L... Ross. Harvard University Press.. Barnett. R. F. 25. W.” American Economic Review. L. Pastor. F. and the U. Pagan. L. and G.” Journal of Economic Theory. 3–22. 76. 29. 13.S. 63(3). 1207–1231. W. and Y. Schwert. 44. S. J.S. T. G.” Journal of Financial Economics. 375–421. D.. J. Turner. Startz. 575–603.” Review of Financial Studies. (1976): “The Arbitrage Theory of Capital Asset Pricing.” Journal of Business.. 1115–1154. Smith. (1990): “Indexes of U. Stock Prices from 1802 to 1987. Cole (1935): Fluctuations in American Busines.” Journal of Monetary Economics. (1992): “The Real Rate of Interest from 1800-1990: A Study of the U. (1994): “Time variations and covariations in the expectation and volatility of stock market returns. and P. and A. and C. Cambridge University Press. C.” Journal of Finance. 1129–1155. Siegel. 399–426. M. 521–547. Nelson.” Journal of Finance. New York.

0293 .0250 0.Table 1: Summary Statistics Sample Mean Excess Return StdDev Min 1840-2006 0.6017 25 Max 0.1824 -0.0427 0.4490 Realized Volatility Mean StdDev 0.

01) 12.i = (1 − αi ) k i=1 t.338 [2. U(2s ) ≡ univariate risk-return model with 2-component volatility but the conditional mean prices the smooth volatility component.282) .438 (.027) . zt ∼ N (0.0009 0.0003 -0.930] U(2) (q = k = 2) -0.052 [-1.02) LRTγ1 =0 Notes: Conditional mean coefficient estimates of γi with t-statistics in brackets. volatility function coefficient estimates αi have standard errors in parenthesis.64 39.0009 0.479] 0.00) 5.314 [1. L is the log-likelihood function. σt.768 [3.794 [3.59 39.020) .24 (0. 1) Table 2: Parameter Estimates: Univariate Risk-Return using RV Levels .027] γ1 3.012 [-0.294) α2 40.408 (..019) .(q) + ²t .122 [2.26 τ −1 k X 1X 2 2 αij RVt−j−1 .14 L 2.36 (0.00) 3.71 37.58 (0.986 [3.0009 2.543] Proportional Model ω 0.54 (0. i = 1.032 [-1. ²t = σt..950 (.573] ..68 (0.910 (. ω is the volatility function intercept which is consistent with targeting the median of RV.242) .903 (.624] 1.06) 12.021) .(k) 2 rt = γ0 + γ1 σt.557] 4.950 (.406 (.0009 1.24 (0.76 (0. 2 σt. k = ω+ σ .256) .396 (.39 41.950 (.0002 1.i j=0 0.(k) zt . .580] Proportional Model U(2s ) (q = 1.024) α1 .50 (0.11) 0.00) 7. and LRT are likelihood ratio test statistics with p-values in parenthesis.08 (0.197] U(1) (q = k = 1) γ0 Model Labels: U(k) ≡ univariate risk-return model with k-component model of RV .30) LRTγ0 =0 14. Proportional Model 0.62) 1.021) . k = 2) -0.60 38.951 (.

²t = σt.0250 .0147 . RMSE ≡ square root of the mean-squared error.71 41.(q) + ²t .63 U(1) (q = k = 1) U(2) (q = k = 2) U(2s ) (q = 1.1807 .0162 .0272 .1410 .i = (1 − αi ) αij RVt−j−1 . We report p-values for these diagnostic tests. i = 1.0251 .14 39.19) (0.57) (0. MAE ≡ mean absolute error.1405 .09) (0...87) (0.15) (0.(k) = ω+ 2 rt = γ0 + γ1 σt. k k i=1 j=0 38.(k) zt .i . 1) Table 3: Model Comparisons and Diagnostics: Univariate Risk-Return using RV Levels . σt.23) (0..1800 . and KS is the Kiefer and Salmon (1983) test for departures from the normal distribution for standardized residuals.03) KS .1795 . U(2s ) ≡ univariate risk-return model with 2-component volatility but the conditional mean prices the smooth volatility component.2639 . L Model 2 Diagnostics LB (10) LB(10) Model Labels: U(k) ≡ univariate risk-return model with k-component model of RV . R2 from Mincer-Zarnowitz regressions applied to the variance and premium forecasts. .16) (0. 2 σt.0284 .0727 Volatility Fit MAE RMSE R2 . k = 2) (0.0165 .27 k τ −1 X 1X 2 2 σt.0224 Premium Fit MAE RMSE R2 L is the log-likelihood function. zt ∼ N (0. LB 2 (10) is the Ljung and Box (1978) heteroskedastic-robust portmanteau test statistic for serial correlation in the squared standardized residuals up to 10 lags.1410 . LB(10) is the same for conditional mean residuals.2656 .0147 .54) (0.

1688 -0.051) .12 (0. Ulog (2s ) ≡ univariate risk-return model with 2-component volatility but the conditional mean prices the smooth volatility component. and LRT are likelihood ratio test statistics with p-values in parenthesis.00) 3.021) .921 (.938 (.28 τ −1 k X 1X 2 2 = (1 − αi ) αij log RVt−j−1 .(q) + ²t .04 20.966] Proportional Model ω -0.931] Ulog (2) (q = k = 2) -0.0796 2.60 (0..0018 2.921 (. ²t = σt.09 26.734 [1.235) α2 24.287 [3.0812 -0. volatility function coefficient estimates αi have standard errors in parenthesis.230) .940] .78 20.355 [4.81) LRTγ0 =0 18.00) 3.42 (0.311 (.029) .231] γ1 2.236 (.96 (0.348] -0.(k) zt .294] 2..08) LRTγ1 =0 Notes: Conditional mean coefficient estimates of γi with t-statistics in brackets. log σt..149] Proportional Model Ulog (2s ) (q = 1. zt ∼ N (0. k .98 (0.0025 -0.032) .032) .04 (0.631 [3. L is the log-likelihood function. i = 1.(k) 2 rt = γ0 + γ1 σt.20 (0.924 (.238) . Proportional Model -0.98 23. k = 2) -0.259 [3.i k i=1 j=0 -0.987] 5.00) 10.01 (0. 1) Table 4: Parameter Estimates: Univariate Risk-Return using Log(RV) .007 [-0.830 (. ω is the volatility function intercept which is consistent with targeting the mean of log(RV ).823 (.056) α1 .i log σt.05) 16.207) .07 22.576] Ulog (1) (q = k = 1) γ0 Model Labels: Ulog (k) ≡ univariate risk-return model with k-component model of the natural logarithm of RV .617 [1.73) 0. .00 (0.309 (.308 (.008 [-0.05) 0.067 [-1.81 L 3.0931 2. 2 = ω+ log σt.00) 16.

1179 Volatility Fit MAE RMSE R2 .2737 .i log σt.19) (0.10) KS .118) (0. k .1809 .07 Ulog (1) (q = k = 1) Ulog (2) (q = k = 2) Ulog (2s ) (q = 1.(k) zt .19) (0.1414 .(k) 2 + ²t ..69) (0.0263 .0263 .(q) Table 5: Model Comparisons and Diagnostics: Univariate Risk-Return using Log(RV) . R2 from Mincer-Zarnowitz regressions applied to the variance and premium forecasts. MAE ≡ mean absolute error. Ulog (2s ) ≡ univariate risk-return model with 2-component volatility but the conditional mean prices the smooth volatility component.0145 . log σt. and KS is the Kiefer and Salmon (1983) test for departures from the normal distribution for standardized residuals. LB 2 (10) is the Ljung and Box (1978) heteroskedastic-robust portmanteau test for serial correlation in the squared standardized residuals up to 10 lags.0199 . i = 1.0135 . 1) rt = γ0 + γ1 σt.0119 . RMSE ≡ square root of the mean-squared error. LB(10) is the same for the conditional mean residuals. L Model 2 Diagnostics LB (10) LB(10) Model Labels: Ulog (k) ≡ univariate risk-return model with k-component model of the natural logarithm of RV . .23) (0.1415 .15) (0. zt ∼ N (0.26) (0. ²t = σt.81 23..2608 . k = 2) (0.0133 Premium Fit MAE RMSE R2 L is the log-likelihood function. We report p-values for these diagnostic tests.05 26.1807 .29 τ −1 k X 1X 2 2 = (1 − αi ) αij log RVt−j−1 .i k i=1 j=0 20.36) (0.0136 . 2 = ω+ log σt.1406 .1803 ..0281 .

239] -0.94 (0.175 (. B(2s ) ≡ bivariate model with 2-component volatility but the conditional mean prices the smooth volatility component.33) LRTγ1 =0 Notes: Conditional mean coefficient estimates of γi with t-statistics in brackets.861 [0.027) .0653 1.2e-7 0.72 L 2. k = 2) -0.019 [0.60 -144.046) 0.30 -136.006 [-0.(k) zt .03) 10.39) LRTγ0 =0 13.141) .765 (0.00) 4.08 (0.08 (0.045) 0.915 (.272] Proportional Model ω -0.09 -143.30 = (1 − αi ) j=0 τ −1 X αij log RVt−j−1 . ω is the volatility function intercept which is consistent with targeting the mean of log(RV ).140) .78 (0. . φ2 ) -0..645] Proportional Model B(2s ) (q = 1. volatility function coefficient estimates αi have standard errors in parenthesis..031) . σt.171 (.745 (0.874] γ1 0.176 (.09) 0.791 [3. and LRT are likelihood ratio test statistics with p-values in parenthesis.954 [1.533 [3.648] . i = 1.045) 0.045) 0.77 -134.0612 1. zt ∼ N (0.(q) + ²t .145) ..i log RVt 1X 2 = ω+ log σt.693] B(2) (q = k = 2) -0.079) α1 .48 (0.00) 9.046) φ -135.034) .04 (0.00) 12.289 [2.859] 1.7e-7 1.638 (.745 (0.763 (0. 2 log σt. L is the log-likelihood function.140) α2 0.5Vart−1 (log RVt |k)). ²t = σt.1429 -0. k ηt ∼ N (0.i + ηt .440] 4.045) 0.62 -136.914 (.84) 0.078) . q ≤ k 2 rt = γ0 + γ1 σt.934 (. k i=1 k 2 2 σt.(k) ≡ Et−1 (RVt |k) = exp(Et−1 (log RVt |k) + .88 (0. Proportional Model -0.034) .066 [-1.00) 0.621 (.(q) ≡ Et−1 (RVt |q).963] B(1) (q = k = 1) γ0 Model Labels: B(k) ≡ bivariate model of returns and the natural logarithm of RV with k-component volatility.159 (.74 (0.0596 -0. 1) Table 6: Parameter Estimates: Bivariate Return and Log(RV) Model .94 (0.706 [3.745 (0.916 (.744 (0.

2 log σt.0196 .0248 .2809 .1810 . LB 2 (10) is the Ljung and Box (1978) heteroskedastic-robust portmanteau test for serial correlation in the squared standardized residuals up to 10 lags. 1) Table 7: Model Comparisons and Diagnostics: Bivariate Return and Log(RV) Model .0268 . R2 from Mincer-Zarnowitz regressions applied to the variance and premium forecasts.60 -134. k = 2) (0.1413 . We report p-values for these diagnostic tests.21) (0. B(2s ) ≡ bivariate model with 2-component volatility but the conditional mean prices the smooth volatility component. ²t = σt.. φ2 ) -143.34) (0.31 = (1 − αi ) j=0 τ −1 X αij log RVt−j−1 .(q) ≡ Et−1 (RVt |q). i = 1.1415 .0133 Premium Fit MAE RMSE R2 L is the log-likelihood function.31 B(1) (q = k = 1) B(2) (q = k = 2) B(2s ) (q = 1. q ≤ k 2 rt = γ0 + γ1 σt.5Vart−1 (log RVt |k)).(q) + ²t .2771 .0115 .20) (0.1806 .0165 .67) (0. k ηt ∼ N (0.. log σt. RMSE ≡ square root of the mean-squared error.26) (0. MAE ≡ mean absolute error.i log RVt 1X 2 = ω+ + ηt .0149 . . zt ∼ N (0. and KS is the Kiefer and Salmon (1983) test for departures from the normal distribution for standardized residuals.i k i=1 k 2 2 σt.1804 . L Model 2 Diagnostics LB (10) LB(10) Model Labels: B(k) ≡ bivariate model of returns and the natural logarithm of RV with k-component volatility.(k) zt .2330 Volatility Fit MAE RMSE R2 ..(k) ≡ Et−1 (RVt |k) = exp(Et−1 (log RVt |k) + .72 -136. LB(10) is the same for the conditional mean residuals. σt.0148 .0248 .15) (0.17) (0.12) (0.1407 .33) KS .

043) 0. zt ∼ N (0.758] 0.771 (0.066 [-1.2e-8 (.148) α2 .745 (0. L is the log-likelihood function.957 (. ²t = σt.924 [2.045) -139.960 (.745 (0.127] 7.257] -0.437 -0.289 [2.(2) zt .677 (0.365] τ =5 Model B(2s ) Model B(2a ) γ1 4. 2 2 i=1 j=0 ¤ £ 2 = γ0 + γ1 exp ω + log σt.071 [-1.934 (.826 [1.32 4.658 (0.314] 4.365) .188 [2. ²t = σt.035 [5.1e-8 (. zt ∼ N (0.5Vart−1 (log RVt ) + ²t .028) .636 [2.958 (.29 L Notes: Conditional mean coefficient estimates of γi with t-statistics in brackets.744 (0.269 (.i + ηt .149 [-1. i = 1.1325 ω .225) .97 -34.020) .149] -0.741 (0.483 [2.073 [-1.994] τ = 15 -0.064) 0. ηt ∼ N (0.185 (.159 (.868] 1.224 [2.000] -0.859] γ0 -0. volatility function coefficient estimates αi have standard errors in parenthesis.804 [2.1429 -0.932 (.741 (0.141) .043) α1 . RV from monthly (1840-2006) 3.080 [-0.67 -87.042] -0.603] 4.1 + . log σt.067) 0.56 -65.047) 0.161 (.046) φ 0.932 (.045) 0.648] ¤ £ 2 2 + .184 -1.154] -0.(2) zt .167) .i = (1 − αi ) αij log RVt−j−1 .30 -134.2 Model B(2a ) : rt = γ0 + γ1 exp ω + γ2 log σt.5Vart−1 (log RVt ) + ²t .024 [-0. 1) + (1 − γ2 ) log σt. 1) Table 8: Robustness Checks: Bivariate Return and Log(RV) Model .3689 -1.94 -134.920 (.56 -135.87 -134.140) .4059 -0.229 (.045) 0.018) .033) .1e-7 (. φ ).088 [-1.027) .959 (.3882 -0.140) 0.1 Model B(2s ) : rt log RVt 2 τ −1 X 1X 2 2 2 = ω+ log σt.1281 -0.007] -0.142) .064 [-1.028) .736] 1886-2006 1928-2006 1945-2006 γ2 1.025) . ω is the volatility function intercept which is consistent with targeting the mean of log(RV ).569] -0.

1 0 1803 1840 1880 1920 33 .2 0.2 0 -0.2 -0.5 0.4 0.4 0.1 0 1803 1840 1880 1920 Figure 3: Realized Volatility Measure: √ 1960 2000 1960 2000 RV 0.4 0.6 0.4 -0.5 0.3 0.Figure 1: Annual Realized Excess Returns: 1803-2006 0.6 0.2 0.6 0.3 0.6 1803 1840 1880 1920 1960 2000 Figure 2: Annual Absolute Value Measure of Ex Post Volatility 0.

15 0.08 0.2 0.18 0.25 0.i ): U(2) Model 0.35 0.4 0.45 0.05 1840 1860 1880 1900 1920 1940 1960 1980 2000 2 Figure 5: Volatility Component Forecasts (σt.2 0.1 0.5 sqrt(RV) 1C Volatility Forecast 2C Volatility Forecast 0.06 0.16 0.05 0 1840 1860 1880 1900 1920 1940 1960 1980 2000 Figure 6: Proportional Premium Forecasts: U(2s ) Model Pricing Smooth Vol Component Pricing Total Volatility 0.Figure 4: Volatility Forecast vs Realized: U(1) and U(2) Models 0.3 0.2 Volatility C1 Volatility C2 0.15 0.14 0.1 0.04 0.1 0.12 0.02 1840 1860 1880 1900 1920 34 1940 1960 1980 2000 .

035 0.06 0.04 0.2 ): B(2) Model Figure 7: Log-Volatility Component Forecasts (log σt.5 Log-Volatility C1 Log-Volatility C2 -2 -2.5 -3 -3.045 0.03 1990 1992 1994 1996 1998 35 2000 2002 2004 2006 .1 0.5 -4 -4.08 0.06 Average 1840-2006 Bivariate Model Time-Varying Premium 0.04 0.12 Bivariate Model Univariate Model using LogRV Univariate Model using LevelRV 0.02 1840 1860 1880 1900 1920 1940 1960 1980 2000 Figure 9: Proportional Premium Forecasts: 1990-2006 0.055 0.5 -5 1840 1860 1880 1900 1920 1940 1960 1980 2000 Figure 8: Proportional Premium Comparisons 0.05 0.i -1.