Asset Prices in Affine Real Business Cycle Models*
  • 1 0000000404811396https://isni.org/isni/0000000404811396International Monetary Fund
  • | 2 0000000404811396https://isni.org/isni/0000000404811396International Monetary Fund

We develop a tractable way to solve for equilibrium quantities and asset prices in a class of real business cycle models featuring Epstein-Zin preferences and affine dynamics for productivity growth and volatility. The method relies on log-linearization and exploits the log-normality of all the quantities. It is an easy substitute for more involved numerical techniques, such as higher order perturbation methods, and allows for easy implementation and analytical results. We show explicitly the link with perturbation techniques and find that the quantitative difference between the two is insignificant for several models of interest.

Abstract

We develop a tractable way to solve for equilibrium quantities and asset prices in a class of real business cycle models featuring Epstein-Zin preferences and affine dynamics for productivity growth and volatility. The method relies on log-linearization and exploits the log-normality of all the quantities. It is an easy substitute for more involved numerical techniques, such as higher order perturbation methods, and allows for easy implementation and analytical results. We show explicitly the link with perturbation techniques and find that the quantitative difference between the two is insignificant for several models of interest.

I. Introduction

Recursive preferences and time variation in means and volatilities have become important features of consumption-based asset pricing literature. The introduction of these features into real business cycle (RBC) models has allowed the study of the joint behavior of real and financial variables along the business cycle. As the analysis of asset prices requires computing risk adjustments, simple log-linearization is insufficient. Furthermore, numerical methods such as value-function iteration are computationally expensive and ill-suited for problems with a large number of state variables.

In this paper we propose a simple alternative. We describe how to compute log-linearized dynamics and risk adjustments that accurately characterize asset pricing and welfare implications while retaining the computational simplicity of log-linearization methods. More specifically, we consider a standard RBC model augmented along two dimensions—recursive Epstein-Zin preferences (see, Epstein and Zin (1989), Epstein and Zin (1991), and Weil (1989)) and a general affine structure for the exogenous state variables. We show how to solve this class of models by exploiting the joint log-normality of shocks and using log-linearization. The suggested method is an easy and tractable alternative to the more common techniques. Our approach is closest to higher-order perturbation methods, where equilibrium conditions are expanded (perturbed) around a steady state using Taylor series expansions. However, whereas higher-order perturbation methods required for asset prices or models with stochastic volatility are difficult to compute, the linearity and analytical form of our suggested method is convenient and computable by hand. The suggested approximation is equivalent to a perturbation solution where higher-order terms describing the dynamics of the quantities are omitted while the key terms accounting for risk adjustments are retained. We show that the resulting approximations from the two methods are almost identical numerically for some examples of interest. We use the theoretical and implementation framework provided by Schmitt-Grohe and Uribe (2004) to do so.

We see three main advantages to our method. First, the linear structure of the solution makes it easy to describe the time-series properties of the variables of interest and carry out estimation. Second, the analytical form allows us to explicitly inspect the mechanisms behind quantity dynamics and asset prices. In particular, it is interesting to examine the effects of separating relative risk aversion from the inverse of the elasticity of inter-temporal substitution in a production economy compared to the more standard time-additive utility case. Finally, our method can deal with stochastic volatility much more easily compared to standard perturbation methods. With perturbation methods at least a third-order expansion is necessary for studying stochastic volatility. We avoid the complications of higher-order exansions in our method by nesting stochastic volatility in a general affine shock structure.

Our work is related to two others that derive risk adjustments in log-linearized RBC models. Backus, Routledge, and Zin (2007) consider a problem similar to ours but without stochastic volatility. We will argue, based on the work by Caldara and others (2008), that their choice of equilibrium conditions to linearize leads to an inconsistent approximation of the value function, which is of crucial importance for models with recursive preferences.

Uhlig (2010) looks at a similar framework without stochastic volatility. The author computes risk adjustments from asset pricing equations only. We show that it is important to consistently compute the risk adjustment resulting from all the equations in the model. The failure to compute risk adjustments for quantities, in particular for models with recursive preferences, can lead to a bias in prices even if risk adjustment for prices are taken into account. For example, ignoring precautionary savings will result in a biased level of the risk-free rate if there is a preference for early resolution of uncertainty. The main contribution of our paper is to demonstrate how log-linearization and risk adjustment can be applied to solve models with stochastic vololatiliy. Solving models that feature stochastic volatility has attracted special attention. Fernandez-Villaverde and others (2009) argue that a third-order approximation is necessary when using perturbation methods. The same conclusion is reached in Malkhozov and Shamloo (2010): the third-order terms in the approximation obtained using a perturbation method capture the first-order dynamic effects of time-varying volatility. The authors use a fourth-order approximation to capture second-order dynamic effects of volatility changes as a robustness check. In this paper we can directly obtain the first-order effects of stochastic volatility on quantities and asset prices by log-linearizing the model and accounting for risk adjustments appropriately.

Furthermore, we demonstrate that standard perturbation method techniques are perfectly suitable for dealing with recursive preferences. We discuss how approximation techniques such as the one we present can be easily applied to such models. In this respect, our work is close to Rudebusch and Swanson (2008), Swanson, Anderson, and Levin (2006), Binsbergen van and others (2008), and Caldara and others (2008), who all address the specific issue of solving models with recursive preferences using perturbation methods.

The affine dynamics of exogenous state variables is central to our solution method. We build on the work by Duffie, Pan, and Singleton (2000), who introduced continuous time affine processes as a powerful modeling tool that allows finding closed-form solutions for a number of problems in finance and economics.1

We demonstrate the application of our model using two examples. Both are RBC models in which agents work, consume and own assets. The first example features recursive preferences, and the second presents an application of stochastic volatility. The elements in these examples draw on several strands in the asset pricing literature. We will briefly mention a few.

The literature on asset prices in production economies has been developing rapidly. Kaltenbrunner and Lochstoer (2007), Croce (2008), and Malkhozov and Shamloo (2010) study asset prices and macroeconomic quantities in RBC-type models. This paper describes a simple and tractable way to tackle problems in this branch of literature. The two examples in our paper are very closely related to these models.

Shocks to volatility have recently emerged as an important factor in driving the business cycle. Justiniano and Primiceri (2008), Fernandez-Villaverde and others (2009), and Bloom, Floetotto, and Jaimovich (2009) show the importance of these shocks for macroeconomic quantities. While these papers argue that changes in uncertainty are most easily observable in financial markets, they do not address the issue of asset pricing implications specifically. Malkhozov and Shamloo (2010) use perturbation methods to study asset prices in a simple growth model with stochastic volatility in productivity growth. This paper offers an easy way to investigate the role of volatility shocks both for macroeconomic quantities and asset prices.

The remainder of the paper is organized as follows. Section II introduces the setup and describes the log-normal risk adjustment technique. Sections III and IV present our two examples, their calibration and the numerical comparison of the solutions obtained by the log-linear risk adjustment method with the widely used perturbation techniques. The example in Section III is a standard RBC model with Epstein-Zin preferences as well as stationary and nonstationary shocks. The example in Section IV features stochastic volatility. Section V concludes.

II. Setup

In this section we describe our baseline real business cycle model with Epstein-Zin preferences and an affine structure of shocks. The framwework can be further extended along several dimensions, such as flexible labour supply or capital adjustment costs, without changing the results of the following sections of the paper.

A. Preferences

The representative consumer maximises the utility function defined recursively

MaxCtUt

where

Ut=(Ct11/ψ+β(Et(Ut+11γ))11/ψ1γ)111/ψ

Unlike CRRA utility function, Epstein-Zin recursive preferences allow us to separate the elasticity of intertemporal substitution from the coefficient of relative risk aversion (see Epstein and Zin, 1989). The parameter γ controls agents relative risk aversion and ψ his elasticity of intertemporal substitution. The standard power utility can be obtained as a special case by setting γ = 1/ψ. This separation has an important implication for the agents preferences towards the early resolution of uncertainty. In the power utility case investor is indifferent towards the timing of resolution of uncertainty, if γ > 1/ψ, (γ < 1/ψ), an investor prefers early (late) resolution of uncertainty. Intuitively, with γ > 1/ψ agents propensity to smooth consumption across states of the world is greater than propensity to smooth consumption across time.

It is important to include recursive preferences in our analysis. Separating the relative risk aversion parameter (γ) and the elasticity of intertemporal substitution (ψ) has been instrumental in tackling asset pricing puzzles in recent literature (see Bansal and Yaron (2004), Kaltenbrunner and Lochstoer (2007) and Croce (2008)). We show that standard macroeconomic techniques can easily handle models featuring recursive preferences and that the use of computationally expensive approaches, such as value function iterations, is not necessary. Moreover the analytical structure of the solution allows us to explicitly identify and analyse the additional effects on quantities and prices introduced by this preferecnes specification.

B. Technology

The consumption good is produced according to a Cobb-Douglas production function

Yt=ZtAt1αKtα

where Yt denotes output, Zt and At denote the stationary and nonstationary components of the total factor productivity and Kt denotes the capital stock at time t. The law of motion of capital is given by

Kt+1=(1δ)Kt+YtCt

where δ is the rate of depreciation of physical capital. In addition define Rt as the marginal product of capital. It follows that:

Rt=α(AtKt)1α+(1δ)

C. Shocks

The total factor productivity is driven by a vector of exogenous state variables ut. We define the first two elemets of ut to be

ut+11 = lnAt+1lnAtut+12 = lnZt+1

The specification for the vector of exogenous variables is the main ingredient of our setup. Recent work suggests that changes in expectations and uncertainty about the productivity are important drivers of the business cycle (see Bloom (2009) and Bloom, Floetotto, and Jaimovich (2009)) and asset prices fluctuations (see Bansal and Yaron (2004) and Malkhozov and Shamloo (2010)). Higher dimension of the vector xt and a very general specification for its dynamics will allow us to capture a wide range of rich information structures about productivity.

We assume discrete-time affine dynamics for exogenous variables

ut+1=H0+H1ut+Σtεt+1,(1)

where H0 and H1 are (n × 1), n being the number of observable exogenous variables. The vector of innovations εtN(0,Inε). Furthermore, the elements of ΣtΣtT, the conditional variance-covariance matrix of the innovations, evolves as:

(ΣtΣtT)ij=(G0)ij+(G1)ijut,(2)

where G0 is (n × n) and G1 is (n × n × n).2

A few comments about specification (1) are in order. First, it allows for time-varying volatility of shocks as (2). Time-varying, or stochastic, volatility has important implications for asset pricing particularly in delivering time-varying premia (see Bansal and Yaron (2004) and Malkhozov and Shamloo (2010)). It is also increasingly important for characterizing business cycle dynamics (see Bloom (2009) and Bloom, Floetotto, and Jaimovich (2009)).

Second, note that both the conditional expectation and the conditional variance-covariance matrix of ut are affine in vector xt itself. This affine structure is a crucial assumption which allows us to assume—and verify—that up to the first order all variables in the model are normally distributed. We will elaborate on this issue in more detail when discussing the solution method.

Finally, note that specification (1) encompasses higher-order autoregressive (AR) structures. For instance, an AR(2) process in ut, ut+1 = H0 + H1ut + H2ut–1 + ∑tε, can be expanded as u^t+1=H^0+u^tH^1+εt+1Σ^t, where u^t=[utut1]. Additional lags can be added in the same manner to the vector of state variables to account for AR(L) terms, where L > 1.

D. Stationary Version, Equilibrium Conditions and the Solution

The model presented above features a permanent TFP shock (At). In order to transform the model into a stationary version, we define the scaled version of any nonstationary variable Xt as X˜t=XtAt1. Note that the technology-adjusted value function V˜t=VtAt1=V(K˜t,A˜t) since V(Kt, At) is homogeneous of degree 1 in At and Kt.

The stationary equivalent of the model is defined by equations (3) to (5) below (see Appendix A for derivation).

V˜t11/ψ=maxC˜t(C˜t11/ψ+A˜t11/ψβ(Et(V˜t+11γ))11/ψ1γ)(3)
K˜t+1=(1δ)K˜tA˜t1+ZtA˜tαK˜tαC˜tA˜t1(4)
Rt=(1δ)+αZtA˜t1αK˜tα1.(5)

The optimal policy is described by the Euler equation:

Et[βA˜t1/ψ(V˜t+1Et (V˜t+11γ)1/(1γ))1/ψγ(C˜t+1C˜t)1/ψRt+1] =1.(6)

A closed form solution of (6) is often difficult (or impossible) to obtain. However, when there is no uncertainty a closed-form solution is known to exist. This particular case is referred to as the non-stochastic steady state. This is different from the stochastic steady state, where εt has a variance, but the model is evaluated at a point where the particular realizations of εt are equal to their means. Appendix B describes the non-stochastic steady state. As is customary in perturbation methods, we approximate the solution around the non-stochastic steady state.3

It is worth noting that Epstein-Zin preferences do not prohibit using perturbation methods. This point is also emphasized by Uhlig (2010). To deal with the term Et(Vt+11γ), one can define an additional control variable wt=Et(Vt+11γ) and add this identity to the set of equilibrium equations. Replace Et(Vt+11γ) with wt in all the equations and expand the equations as usual.

E. Log-Normal Risk Adjustment: An Approximation Technique

In this section we introduce an approximation technique, which we will refer to as the “log-normal risk adjustment.” Denote with small letters the log of original variables, such that xt = lnXt. Separate the variables in this model into nx pre-determined variables denoted by vector xt and ny non-pre-determined variables by vector yt. Vector yt can include jump variables (choice variables determined at t) and exogenous random variables (innovations to shocks).4 Note that pre-determined variables are not necessarily the set of state variables (the set of variables that uniquely define the position of the system in the state space). For instance, lnÃt and lnZt are state variables; however, in this classification they would be part of the yt vector since they are not pre-determined.

The set of equilibrium conditions for a large group of DSGE models, including the prototype model introduced earlier, can be written in the following form:

f(xt+1,xt,yt,Et[exp(Γyt+1)])=0,(7)

where yt+1 is an ny × 1 vector and Γ is a conformable matrix of constants.5 In the model described above, xt = [kt] and yt = [ct; vt; ut].

We conjecture that the linearized solution will imply that yt is a vector of conditionally normal variables. We linearize the model given this conjecture and obtain a solution for yt which is linear in the state-variables. Given the affine structure assumed for the state variables, vector yt will be conditionally normal, verifying our initial conjecture. We will show this in detail using our two examples.

Based on the conjecture that yt is a vector of conditionally normal variables we can re-write equation (7) as:

f(xt+1,xt,yt,exp(ΓEt[yt+1]+12Γ(Vartyt+1)Γ))=0.(8)

Linearizing equation (8) above, it is clear that by design yt will be linear in exogenous shocks which are assumed to be normal. Therefore, our conjecture regarding conditional normality of yt will bear out. We refer to this approximation as log-normal risk adjustment. We linearize the set of equilibrium conditions defined by f in (8).

We claim that this method captures the risk adjustment in RBC models where the stochastic and non-stochastic steady states are different. Furthermore, unlike perturbation methods this is an easy way to capture stochastic volatility. The method has the advantage that it is computationally simple and semi-analytical equations can be derived, similar to perturbation techniques. In addition, it allows us to capture the risk adjustment in all the variables without going to second-and higher-orders in the Taylor expansion of the terms, thereby making it a useful technique for understanding financial variables such as risk premia and risk-free rates, as well as utility measures that contain second-order terms. We will discuss these implications in detail in the following two examples.

III. Example 1: RBC with Recursive Preferences

In this example we develop the method explained above using a simple RBC model with Epstein-Zin preferences. We assume there is no stochastic volatility. (We will present an example with stochastic volatility in Section IV). We keep the structure of the shocks simple.

ut =[atzt]at+1=(1ρa)μ+ρaat+σaεt+11zt+1= ρzzt+σzεt+12.

We now need to re-write the equilibrium conditions in the same form as equation (8). The value function can be written as:

V˜t11/ψ=maxC˜t(C˜t11/ψ+A˜t11/ψβ(Et(exp(1γ)vt+1))11/ψ1γ)=maxC˜t(C˜t11/ψ+A˜t11/ψβ [exp((1γ)Etvt+1+12(1γ)2Var(vt+1))]11/ψ1γ),(9)

where the last step follows because of our assumption that vt is normal. We will verify this conjecture later on. Note that we have dropped the subscript t from the variance term since this model assumes constant volatility of shocks, and therefore, the variance of all variables will be time-independent.

Define hat variables as deviations from their non-stochastic steady state. Linearizing equation (9) around the non-stochastic steady state we obtain:

ζ1v^t=ζ2c^t+ζ3(a^t+Etv^t+1+12(1γ)2Var(v^t+1)),(10)

where ζ1=V˜11ψ, ζ2=C˜11ψ and ζ3=βA˜11ψV˜11ψ. V˜, C˜ and à denote the non-stochastic steady state values of those variables. Linearizing equations for K˜t and Rt and re-writing the evolution of the shocks as deviations from their respective non-stochastic steady states we obtain:

k^t+1kkk^t+kaa^t+kzz^t+kcc^t(11)r^t rak(a^tk^t)+rzz^ta^t =μ+ρAa^t1+σAεtaz^t =ρZz^t+σZεtz,

where kk, ka, kz and kc are known. 6

Finally, re-write the Euler equation (6) as Et (Mt+1Rt+1) = Et [exp (mt+1 + rt+1)] = 1 where:

Mt+1=βA˜t1/ψ(V˜t+1Et (V˜t+11γ)1/(1γ))1/ψγ(C˜t+1C˜t)1/ψ.

Again, assuming log-normality of Mt+1 and Rt+1, re-write the Euler equation and the definition of the stochastic discount factor as:

Et(mt+1+rt+1)+12Var(mt+1+rt+1)=0(12)
Mt+1=βA˜t1/ψ(V˜t+1exp(Etvt+1+12(1γ)Var(vt+1)))1/ψγ(C˜t+1C˜t)1/ψ.

Taking logs and subtracting the non-stochastic steady state yields:7

m^t+1=(1/ψγ)(v^t+1Etv^t+1+12(1γ)Vart(v^t+1))1/ψa^t1/ψ(c^t+1c^t).

Equations (10)-(12)) give us three equations in three unknowns (v^t, c^t, and k^t) which can be solved. The system of equations can be written as:

[Etv^t+1Etc^t+1k^t+1]=R+W[v^tc^tk^t]+QEt[a^t+1z^t+1].(13)

Note that R is a vector of constants. This is an important difference between our method and the standard log-linearization technique. When approximating a model to the first-order, as in log-linearization, on average variables will be at their non-stochastic steady state. This is because volatility is a second-order effect. Our suggested method on the other hand, explicitly finds the difference between the stochastic and non-stochastic steady states. This “risk-adjustment” term is summarized in vector R.

The vector R might not be important to us if the only objects of interest are the dynamic responses of variables to shocks. However, the constant term for each variable carries important economic intuition about how agents evaluate risk. For instance, the constant term in Etĉt+1 reflects the consumption deficit (compared to a model without risk) due to precautionary savings. More importantly, the size of the constant affects financial variables such as the unconditional mean of the risk-free rate.

The set of equations presented by (13) can be solved to obtain state-evolution and decision rules. The solution is presented in Appendix C. An alternative solution using the method of undetermined coefficients as in Campbell (1994) is presented in Appendix D.

A. Quantities and Prices

The log-linear structure of the model is convenient for deriving and studying the time-series properties of the variables of interest. For illustrative purposes we will focus on just some of them.

The log consumption growth is:

gt+1c=c˜t+1c˜t+a˜t.

The one-period risk-free rate is defined by:

rtf=lnEtMt+1Etmt+112Vart(mt+1),

and has the following expression:

rtflnβ12(1/ψγ)(γ1)Vart(v˜t+1)12Vart(mt+1)+1/ψEt(gt+1c).

Note that a first-order log-linearization would have omitted the two terms involving Vart(v˜t+1) and Vart (mt+1). The risk-free rate is lower in this model (compared to a deterministic model) because there is more risk: movements in realized and expected consumption growth cause a variance in the stochastic discount factor which drives down the interest rate. The intuition for why the term involving Var(v˜t+1) increases the risk-free rate is more subtle and comes about because of the Epstein-Zin preferences. Note that this term is only positive if γ > 1/ψ, or in other words, if the agents have a preference for early resolution of uncertainty. In this case agents would rather consume more today to resolve future uncertainty earlier. By bringing forward consumption they push up interest rates.

Returns on any asset i satisfy Et (exp(mt+1+rt+1i))=1. Up to the first-order, all expected returns are the same Et(rt+1i)=Et(mt+1). However, using log-normality to adjust for risk we can show that the risk premium of any asset i is:

Et(rt+1irtf)=Covt(rt+1i,mt+1)12Vart(rt+1i),

where 12Vart(rt+1i) is a Jensen’s inequality correction term.

B. Log-Normal Risk

Define the entropy of the stochastic discount factor as ln EtMt+1Et lnMt+1. Entropy can be interpreted as measure of market price of risk (see Alvarez and Jermann (2005)). Since Mt+1 is log-normal, the entropy depends only on the second moment and has the following simple expression:

lnEtMt+1EtlnMt+1=12Vart(mt+1).

Note that in each step we exploit the log-linear form of all the expressions of interest and the normality of the innovations to the exogenous variables.

C. Theoretical Comparison with Second-Order Perturbation Methods

The example above can also be analyzed using perturbation methods. For understanding financial variables such as risk premia and risk-free rates it is essential to use a second-order perturbation method.8 This section is intended to compare the results obtained using the log-normal adjustment method with a perturbation method, both theoretically and numerically. We will show that log-normal risk is a truncated version of the second-order perturbation. However, the numerical results show that log-normal risk is a very close approximation to the full second-order approximation. In order to perform a second-order approximation we use the code provided by Schmitt-Grohe and Uribe (2004). We compare the results with those obtained using the log-normal risk adjustment method.

First, we consider our method with the second-order perturbation method theoretically. Consider a generic model, with Y as the set of control variables and S the set of state variables. Denote the solution to this model as:

Yt=g(St).(14)

The solution can be approximated to the second-order around the non-stochastic steady state, denoted by (Y¯,S¯), as follows:

YtY¯+gs(S¯)(StS¯)+12gss(S¯)(StS¯)2.(15)

The log-normal risk solution method keeps only certain terms in the 12gss(S¯)(S¯tS¯)2. In particular all the terms involving variances of state variables are kept, whereas quadratic terms in endogenous state variables (second-order dynamic terms) are omitted. The terms involving the variances of state variables are those summarizing the “risk adjustment” in control variables.9

The accuracy of the log-normal risk adjustment depends on the importance of the second-order dynamic terms. If these terms are negligible, the log-normal risk adjustment method provides an accurate and computationally efficient alternative to second-order perturbation methods. The next section compares a numerical calibration of the model above using second-order perturbation methods and the log-normal risk approximation.

D. Quantitative Comparison with Second-Order Perturbation Methods

In this section we present a calibration of Example 1 and compare the numerical solution found using second-order perturbation methods with the solution obtained by the log-normal risk adjustment method.

Table 1 shows a calibration of the example in Section III. We choose a monthly calibration for the model. Parameter β is chosen so that the annual rate of time preference is 0.98. The capital share in the Cobb-Douglas production is set to the common value of 1/3. The rate of growth of the economy is determined by the trend component in the α shock, and we set μ such that the annual growth rate of the economy is 2 percent. The rate of depreciation of capital, δ, is chosen as 10 percent per annum. We let γ = 5 and the intertemporal elasticity of substitution, ψ, is chosen to be 1.5, which are both well within the range used in the literature (see, for instance, Kaltenbrunner and Lochstoer (2007)).

Table 1.

Calibrating the Benchmark Parameters

This table shows the calibration values for Example 1. The calibration is based on a monthly frequency.

article image

Table 2 shows the values obtained for the first-order and second-order terms using a second-order perturbation method, and the log-normal risk method suggested in this paper. First note that an approximation of the form (15) for the solution to the model finds the first-and second-order derivative with respect to the state variables of the function which links the state to the control variables. These derivatives evaluated at the non-stochastic steady state are shown in the “Perturbation” columns.

Table 2.

Comparing Coefficients in Perturbation Method and Log-normal Risk

This table shows the values for the first-and second-order terms of the approximation to the model in Example 1. The “Perturbation” column shows the value for the terms obtained using a second-order perturbation method. The “LN Risk” column shows the value for the same terms obtained using the log-normal risk method.

article image

The equivalent of the first-order terms in the log-normal risk approximation is the coefficients of the state variables. They are computed exactly the same way, so their values are identical to the second-order perturbation method. However, the log-normal approximation method does not include second-order terms except for the variance of the external shocks. Therefore, the equivalent of the second-order dynamic terms (those involving two state variables) in our suggested method is zero.

Finally, compare the second-order terms involving variances in both methods. These are the “risk-adjustment terms”. The risk-adjustment terms are very close in both methods—though not identical. The difference is related to the fact that second-order dynamic terms are set to zero in our method. We show this fact rigorously in Appendix F.

Figures 1 and 2 show the impulse responses of capital, consumption and risk-free rates in this model to a one-standard-deviation shock in a(t) and z(t). The responses obtained from each method are superimposed. All variables are calculated as percentage deviations from the stochastic steady state. The motivatuion for this is to concentrate on the dynamics in the two models. We will compare the stochastic steady states below.

Figure 1.
Figure 1.

Comparison of Impulse Responses: Perturbation Method vs. Log-normal Risk

Citation: IMF Working Papers 2010, 249; 10.5089/9781455209491.001.A001

This figure shows the impulse response of the model in Example 1 to a one standard deviation shock in z(t) or the stationary technology shock. The responses calculated using the second-order perturbation method and the log-normal risk adjustment method are superimposed.
Figure 2.
Figure 2.

Comparison of Impulse Responses: Perturbation Method vs. Log-normal Risk

Citation: IMF Working Papers 2010, 249; 10.5089/9781455209491.001.A001

This figure shows the impulse response of the model in Example 1 to a one standard deviation shock in a(t) or the non-stationary technology shock. The responses calculated using the second-order perturbation method and the log-normal risk adjustment method are superimposed.

Note that the dynamic responses of capital, consumption and risk-free rates are nearly identical. This implies that the second-order dynamic terms are not quantitatively important, hence the log-normal method does a good job predicting the dynamic responses.

The responses shown in Figures 1 and 2 are the deviations of each variable from its value in the stochastic steady state. Therefore, for a complete comparison we also look at the stochastic steady state values using both approximation methods. From Table 3, we observe the estimated stochastic steady states calculated using a second-order perturbation method and the log-normal risk method are very close for all variables of interest. The percentage difference in the stochastic steady state between the two methods is less than 1 percent for each variable.

Table 3.

Comparison of Stochastic Steady States: Perturbation Method vs. Log-normal Risk

article image
This table shows the values of capital, consumption, risk-free rate and the value function in the stochastic steady state, computed using the second-order perturbation method and the log-normal risk adjustment method. The final column shows the percentage difference in each computed value using the two different methods.

IV. Example 2: Stochastic Volatility in a Simple Growth Model

Now, we will consider the prototype model introduced in Section II and introduce stochastic volatility in productivity growth. We set 1/ψ = γ and assume away stationary (Z) shocks so that we can focus on the effect of stochastic volatility. The shock structure can be described as follows,

a^t+1=μ+σtεt+1aσt+12=(1φ)θ+φσt2+ωεt+1σ.

where Et (εt+1aεt+1σ)=0. Define εt=[εta,εtσ]. In terms of our notation (1)),

ut+1=H0+H1ut+Σtεt+1H0=[μ(1φ)θ]H1=[000φ]Σt=[σt00ω].

Again, the model can be written as:

[Etc^t+1k^t+1]=R+W[c^tk^t]+QEt[a^t+1σt+12].

Note that given the CRRA preferences, ĉt describes all the information in the movements of value function, and therefore, the value function can be eliminated from the set of control variables. As a result, the solution to this model is slightly different from the previous one.

Note that because Et (xt+1)+12Vart(xt+1)=0 we need to know Vart (xt+1) to be able to solve for xt+1. Here is when the affine structure of shocks becomes important. In the previous problem we knew that if the only source of uncertainty are ut shocks, then a linear solution would imply that xt is a linear function of ut and moreover that Vart (xt+1) is a constant. We included this constant in the R matrix above and solved for R once we found the coefficient of optimal xt in response to ut. We repeat the same exercise here. We know the solution is of the form:

c^t=Rc+αa^t+βσt2,

with unknown variables β and Rc. It follows that Vart(c^t+1)=(α2+β2ϕ)σt2+β2ω2. Furthermore, we have an equation linking Et (ct+1) and Vart (ct+1). By equating the coefficients for σt2 and the constant terms we can solve for β and matrix R. As in Example 1, an alternative solution uses the method of undetermined coefficients. The solution using this method is presented in Appendix E.

Note that the conditional variance of consumption growth Vart(gt+1c)=Vart(c˜t+1c˜ta˜t)=ca2σt2+cσ2ω2. It inherits the form of the productivity variance process:

(σt+1c)2=(1φ)(ca2θ+cσ2ω2)+φ(σtc)2+ca2ωεt+1.(16)

This is also the form assumed by several recent consumption-based asset pricing models. See for example Bansal and Yaron (2004), Backus, Routledge, and Zin (2008) and Beeler and Campbell (2009).

A. Qualitative Comparison with Second-Order Perturbation Methods

As mentioned above, dealing with stochastic volatility with perturbation methods is not easy. To capture the full effect of stochastic volatility in perturbation methods we need to approximate the solution to the fourth-order. Therefore, we could not readily use, for instance, the software of Schmitt-Grohe and Uribe (2004). Doing a fourth-order Taylor expansion to the general form of the solution and solving for the coefficients analytically is also extremely tedious. We therefore resort to the dynare++ software, which already has this capability built in.

Programming a model, however complicated in nature, is very simple using dynare++, and it therefore enables an analysis of models that we may not have otherwise had the tools. To be sure, there are also some disadvantages associated with using dynare++. First, it is a black box. The average user—including the authors—may not be familiar with the methods that the software uses to approximate the model and therefore may not be capable of improving on them or judging their suitability. Second, a large number of simulations are needed to obtain accurate approximations. For instance, even using 30,000 simulations our fourth-order approximation to the stochastic volatility model still is not perfectly smooth. Needless to say, running such a large number of simulations is computationally very expensive.

Comparatively, our method suggests that once the coefficients for the state-space system above are found (algebraically tedious as they may be, technically these are just quadratic equations), finding the dynamic response of the model is almost trivial. In the next section we show that this method is also very close to the response we obtain from dynare++ software.

B. Quantitative Comparison with Second-Order Perturbation Methods

There is no standard calibration for the stochastic volatility model presented above. Most stochastic volatility exercises in finance literature (such as Bansal and Yaron (2004), Bansal, Kiku, and Yaron (2006) and Bansal, Kiku, and Yaron (2007)) assume stochastic volatility for the consumption process.10 However, as equation (16) shows there is a direct relationship between the volatility process of consumption and that of productivity. Therefore, if we assume a particular volatility for the consumption process similar to ones in the existing asset pricing studies, the parameters of the productivity volatility process can be backed out such that they result in the desired consumption volatility. Note that our focus in this paper is the solution methodology and its accuracy, and therefore, the suitability of the calibration is of secondary concern. Malkhozov and Shamloo (2010) explore macroeconomic and asset pricing implications of these calibrations further.

Table 4 shows the calibration of the more standard parameters for Example 2. We keep them unchanged relative to the model in Example 1, except that the assumption of CRRA preferences implies that ψ = 1/γ = 0.2. Beeler and Campbell (2009) summarize some popular calibrations of the variance of the consumption growth. They are reproduced in the top panel of Table 5. The bottom panel gives the equivalent parameters for stochastic volatility in productivity. The examples we present in this section are based on the Bansal, Kiku, and Yaron (2006) calibration. The results for the other calibrations are similar but omitted for brevity.

Table 4.

Calibrating the Benchmark Parameters

article image
This table shows the calibration of the standard parameters in the model of Example 2.
Table 5.

Stochastic Volatility in Productivity and Implied Stochastic Volatility in Consumption

This table shows the calibration of the stochastic volatility parameters for the model in Example 2. The top panel shows the calibration of consumption stochastic volatility according to Bansal, Kiku, and Yaron (2006, 2007) and Bansal and Yaron (2004), respectively in rows 1 to 3. The lower panel shows the calibration of the stochastic volatility in productivity which would give rise to the consumption stochastic volatilities displayed in the top panel.

article image

Figure 3 shows the dynare++ results from the 30,000 simulations. We approximate the model up to fourth-order to observe how the effects of stochastic volatility are captured in different orders of approximation. Figure 3 superimposes the different orders of approximations to the model. It is clear that consumption and risk-free rates do not respond to a shock to the volatility of productivity to the first-order. The second-order is purely noise. Note that we expect consumption to drop following a positive shock to volatility since consumers save more as the amount of risk in the economy rises (this is the immediate effect). However, as volatility returns to its steady state value consumption increases; in fact, as agents initially accumulate capital to counter the effect of higher volatility, production capabilities increase and so consumption overshoots before returning to its steady state level.11 The risk-free rate decreases as agents try to save more relative to the pre-shock levels and increases again slowly as consumption returns to its unconditional mean.

Figure 3.
Figure 3.

Stochastic Volatility in Productivity and Implied Stochastic Volatility in Consumption

This figure shows the impulse response of consumption and risk-free rate in the model of Example 2 to a one standard deviation volatility shock. The results are obtained using dynare++ averaged over 30,000 simulations. The model is approximated to first, second, third and fourth-order and the results are superimposed.

Citation: IMF Working Papers 2010, 249; 10.5089/9781455209491.001.A001

Note that the third and fourth-order approximations of the model capture the dynamics of consumption and risk-free rate responses to the shock. Whereas the first-and second-order approximations do not even qualitatively match these results; what we observe is just noise.

Figure 4 compares the fourth-order results obtained using dynare++ along with the log-normal risk responses. We observe that the responses using the log-normal risk method are very close to the fourth-order approximation responses. Evaluating the accuracy of the responses is difficult since there is no unique result obtained from dynare++ (recall that the results are the average over 30,000 simulations). However, compared to the dynare++ output, the log-normal risk results are much smoother.

Figure 4.
Figure 4.

Stochastic Volatility in Productivity and Implied Stochastic Volatility in Consumption

This figure shows the impulse response of consumption and risk-free rate in the model of Example 2 to a one standard deviation volatility shock. The results are obtained using the log-normal risk adjustment method (solid line) and are superimposed over the fourth approximation obtained using dynare++ (dashed line).

Citation: IMF Working Papers 2010, 249; 10.5089/9781455209491.001.A001

C. Stochastic Volatility and Implications for Asset Prices

A strand of papers in the asset pricing literature claims that since stochastic volatility is an extra risk-factor in the economy, modeling it—usually as consumption stochastic volatility—will increase the premia given a certain volatility in consumption (see Bansal and Yaron (2004), Bansal, Kiku, and Yaron (2006) and Bansal, Kiku, and Yaron (2007)). These papers assume that shocks to volatility in consumption (the stochastic volatility shock) and shocks to consumption growth are uncorrelated, and therefore, as volatility of consumption increases, consumption is reduced in order to raise precautionary savings. This suggests a negative correlation between consumption growth and consumption growth volatility.

However, if stochastic volatility in consumption is due to stochastic volatility in productivity, the mechanism suggested in Bansal and Yaron (2004) and the subsequent asset pricing implications are unlikely to emerge in equilibrium. The importance of the correlation between consumption growth and consumption growth variance for asset pricing has been highlighted by Backus, Routledge, and Zin (2008). Our analysis shows that even when innovations to productivity growth and productivity growth variance are uncorrelated, innovation to consumption growth and consumption growth variance are,

Covt(gt+1c,(σt+1c)2)=cσcx2ω2.

Note that this correlation is positive, implying that consumption growth increases when there are shocks to consumption growth volatility. In other words, stochastic volatility in consumption becomes a hedge for expected consumption growth thereby reducing risk premia. The effect is the opposite of what is assumed in endowment economy asset pricing literature.

Figure 5 shows correlations for various values of parameters assuming the level of variance is equal to the long-term mean θ. Higher elasticity of intertemporal substitution implies a stronger negative correlation. Asset pricing implications of stochastic volatility of productivity are studied in detail in Malkhozov and Shamloo (2010).

Figure 5.
Figure 5.

Consumption Growth and Variance of Consumption Growth

This figure shows the correlation between consumption growth and variance of consumption growth for different calibrations and values of ψ.

Citation: IMF Working Papers 2010, 249; 10.5089/9781455209491.001.A001

V. Conclusion

We suggest a way to solve real business cycle models using approximation techniques common in asset pricing literature. Even if eventually we prefer perturbation methods for their generality, we argue that log-linearization as presented in this paper is a very convenient tool that enables correctly capturing not only the dynamics of quantities but also asset pricing and welfare implications. We show precisely how the method is related to a standard higher-order perturbation approach. Furthermore, it suggests a computationally efficient way for solving and estimating models with stochastic volatility.

References

  • Alvarez, F., and U. J. Jermann, 2005, “Using Asset Prices to Measure the Persistence of the Marginal Utility of Wealth,” Econometrica, Vol. 73, No. 6, pp. 19972016.

    • Search Google Scholar
    • Export Citation
  • Backus, D. K., B. R. Routledge, and S. E. Zin, 2007, “Asset Prices in Business Cycle Analysis,”.

  • Backus, D. K., B. R. Routledge, and S. E. Zin, 2008, “The Cyclical Component of US Asset Returns,”.

  • Bansal, R., D. Kiku, and A. Yaron, 2006, “Risks for the Long Run: Estimation and Inference,”.

  • Bansal, R., D. Kiku, and A. Yaron, 2007, “A Note on the Economics and Statistics of Predictability: A Long Run Risks Perspective”.

  • Bansal, R., and A. Yaron, 2004, “Risks for the Long Run: A Potential Resolution of Asset Pricing Puzzles,” The Journal of Finance, Vol. 59, No. 4, pp. 14811509.

    • Search Google Scholar
    • Export Citation
  • Beeler, J., and J. Y. Campbell, 2009, “The Long-Run Risks Model and Aggregate Asset Prices: An Empirical Assessment,”.

  • Binsbergen van, J. H., J. Fernández-Villaverde, R. S. J. Koijen, and J. F. Rubio-Ramírez, 2008, “Likelihood Estimation of DSGE Models with Epstein-Zin Preferences,”.

    • Search Google Scholar
    • Export Citation
  • Blanchard, O. J., and C. M. Kahn, 1980, “The Solution of Linear Difference Models under-Rational Expectations,” Econometrica, Vol. 48, No. 5, pp. 13051311.

    • Search Google Scholar
    • Export Citation
  • Bloom, N., 2009, “The Impact of Uncertainty Shocks,” Econometrica, Vol. 77, No. 3, pp. 623685.

  • Bloom, N., M. Floetotto, and N. Jaimovich, 2009, “Really Uncertain Business Cycles,”.

  • Caldara, D., J. Fernández-Villaverde, J. F. Rubio-Ramírez, and W. Yao, 2008, “Computing Models with Recursive Preferences,”.

  • Campbell, J. Y., 1994, “Inspecting the Mechanism: An Analytical Approach to the Stochastic Growth Model,” Journal of Monetary Economics, Vol. 3, No. 3, pp. 463506.

    • Search Google Scholar
    • Export Citation
  • Croce, M. M., 2008, “Long Run Productivity Risk: A New Hope for Production-based Asset Pricing,”.

  • Dai, Q., A. Le, and K. J. Singleton, 2010, “Discrete-Time Dynamic Term Structure Models with Generalized Market Prices of Risk,” Review of Financial Studies, Vol. 23, No. 5, pp. 21842227.

    • Search Google Scholar
    • Export Citation
  • Duffie, D., J. Pan, and K. Singleton, 2000, “Transform Analysis and Asset Pricing for Affine Jump-Diffusions,” Econometrica, Vol. 68, No. 6, pp. 13431376.

    • Search Google Scholar
    • Export Citation
  • Epstein, L. G., and S. E. Zin, 1989, “Substitution, Risk Aversion, and the Temporal Behavior of Consumption and Asset Returns: A Theoretical Framework,” Econometrica, Vol. 57, No. 4, pp. 93769.

    • Search Google Scholar
    • Export Citation
  • Epstein, L. G., and S. E. Zin, 1991, “Substitution, Risk Aversion, and the Temporal Behavior of Consumption and Asset Returns: An Empirical Analysis,” Journal of Political Economy, Vol. 99, No. 2, pp. 26386.

    • Search Google Scholar
    • Export Citation
  • Fernandez-Villaverde, J., P. Guerron-Quintana, J. F. Rubio-Ramirez, and M. Uribe, 2009, “Risk Matters: The Real Effects of Volatility Shocks,”.

    • Search Google Scholar
    • Export Citation
  • Justiniano, G., and A. Primiceri, 2008, “The Time Varying Volatility of Macroeconomic Fluctuations,” American Economic Review, Vol. 98.

    • Search Google Scholar
    • Export Citation
  • Kaltenbrunner, G., and L. A. Lochstoer, 2007, “Long Run Risk through Consumption Smoothing,”.

  • Malkhozov, A., and M. Shamloo, 2010, “Asset Prices in a News Driven Real Business Cycle Model,”.

  • Rudebusch, G. D., and E. T. Swanson, 2008, “The Bond Premium in DSGE Model with Long-Run Real and Nominal Risks,”.

  • Schmitt-Grohe, S., and M. Uribe, 2004, “Solving Dynamic General Equilibrium Models Using a Second-Order Approximation to the Policy Function,” Journal of Economic Dynamics and Control, Vol. 28, No. 4, pp. 75575.

    • Search Google Scholar
    • Export Citation
  • Swanson, E. T., G. S. Anderson, and A. T. Levin, 2006, “Higher-Order Perturbation Solutions to Dynamic, Discrete-Time Rational Expectations Models,” in Federal Reserve Bank of San Francisco Working Paper Series 2006-01.

    • Search Google Scholar
    • Export Citation
  • Uhlig, H., 2010, “Easy EZ in DSGE,”.

  • Weil, P., 1989, “The Equity Premium Puzzle and the Risk Free Rate Puzzle,” Journal of Monetary Economics, Vol. 24, No. 3, pp. 40121.

    • Search Google Scholar
    • Export Citation

Appendix A. Equilibrium Conditions

The value function:

V(Kt,At)=maxC(Kt,Zt)(Ut)Vt=maxC(Kt,Zt)((1β)Ct11/ψ+β(Et(Vt+11γ))11/ψ1γ)111/ψ.

Define the scaled variables X˜t=XtAt1 and the scaled value function V˜t=V(K˜t,A˜t). Ut, Yt, Kt+1 are homogeneous of degree one in At and Kt; therefore, the value function is homogeneous of degree one in At and Kt as well. In particular we can scale the problem by At–1,

V(Kt,At)=At1V(KtAt1,AtAt1).

In scaled variables the equilibrium conditions can be written as:

VtAt1=maxC(Kt,At)((CtAt1)11/ψ+1At111/ψβEt(Et(Vt+11γ))11/ψ1γ)111/ψVtAt1=maxC(Kt,At)((CtAt1)11/ψ+At11/ψAt111/ψβ(Et(Vt+11γAt1γ))11/ψ1γ)111/ψV˜t=maxC(K˜t,Zt)(C˜t11/ψ+A˜t11/ψβ (Et(V˜t+11γ))11/ψ1γ)111/ψ.

The first-order condition with respect to C˜t and the envelope condition are as follows:

C˜t1/ψ=βA˜t1/ψEt[V˜t+11γ]γ1/ψ1γEt[V˜t+1γV˜K˜t+1]V˜K˜t=βA˜t1/ψV˜t1/ψEt [V˜t+11γ]γ1/ψ1γEt [V˜t+1γV˜Kt+1]Rt.

Iterating the envelope condition one period forward and combining it with the first-order condition, we obtain the Euler equation for consumption and an expression for V˜K˜t:

βEt[A˜t1/ψEt [V˜t+11γ]γ1/ψ1γV˜t+1γ+1/ψC˜t+11/ψC˜t1/ψRt+1]=1V˜K˜tV˜t1/ψC˜t1/ψRt=0.

Notice that Rt can be expressed in terms of scaled variables:

Rt=(1δ)+αZtA˜t1αK˜tα1.

Appendix B. Non-Stochastic Steady State

The following relations define the equilibrium:

Et [βA˜t1/ψEt [V˜t+11γ]γ1/ψ1γV˜t+1γ+1/ψC˜t+11/ψC˜t1/ψRt+1]=1V˜t11ψmaxC(K˜t,Zt)(C˜t11ψ+A˜t11/ψβ(Et(V˜t+11γ))11/ψ1γ)=0K˜t+1(1δ)K˜tA˜t1ZtA˜tαK˜tα+C˜tA˜t1=0.

The non-stochastic steady state can be described by:

u=(IH1)1H0A˜=exp(u1)Z=exp(u2)R=β1A˜1/ψK˜=A˜[β1A˜1/ψ(1δ)αZ]1α1Y˜=ZA˜1αK˜αC˜=(1δ)K˜+Y˜K˜A˜V˜=C˜(11βA˜11ψ)111/ψ.

Appendix C. Difference Equation Solution

Equation (13) can be re-written as:

[Etv^t+1Etc^t+1k^t+1]=R+PΛP1[v^tc^tk^t]+Qu^t,(17)

where Λ and P are the eigenvalues and eigenvectors associated with W. As noted above, we should find that there are m roots larger than one, associated with the jumpy variables. In this case m = 2, the two roots associated with v^t and c^t.

First, note that QEtu^t+1=QH1u^t=Qu^t. Second, note that R includes constants such as Vart(v^t+1) and Vart(m^t+1) which are unknown. However, the stochastic part of the solution is independent of the constant term so we can solve the model above in two stages. First, solve the model without the R,

P1[Etv^t+1Etc^t+1k^t+1]=ΛP1[v^tc^tk^t]+P1Qu^t.

Redefine the system as:

[Etyt+11Etyt+12yt+13]=[λ1000λ2000λ3] [yt1yt2yt3]+P1Qu^t.

For λ1 > 0 and λ2 > 0, the roots can be solved forward and for λ3 it can be solved backwards yielding,

yt1= λ1λ1ρA(P1Q)11a^t(P1Q)12z^tyt2= λ2λ2ρA(P1Q)21a^t(P1Q)22z^tyt3= λ2yt13+(P1Q)31a^t+(P1Q)32z^t.

Then, find v^t, c^t and k^t by:

[v^tc^tk^t]=P[yt1yt2yt3].(18)

Furthermore,

Vart[v^t+1c^t+1k^t]=PVart(yt+1)P,

with solutions for v^t and c^t at hand one can easily calculate the conditional variance of these variables at time t and thus, solve for matrix R.

Appendix D. Example 1 Solution: Method of Unknown Coefficients

The log of the stochastic discount factor is:

mt+1=lnβ+(1/ψγ)(v˜t+11(1γ)lnEte(1γ)v˜t+1)1/ψa˜t1/ψ(c˜t+1c˜t).

Because of the assumed log-linearity of v˜tand the normality of shocks it can be written:12

mt+1=lnβ+(1/ψγ)(v˜t+1Etv˜t+1+12(1γ)Vart(v˜t+1))1/ψa˜t1/ψ(c˜t+1c˜t).

The Euler equation (6) can be written as Et (exp(mt+1 + rt+1) = 1, which again using the log-normal structure of the model implies the following condition:

Et(mt+1+rt+1)+12Vart(mt+1+rt+1)=0.

Define some preliminary expressions. First,

1/ψa^t1/ψ(c^t+1c^t)+rt+1=1/ψ[kcck2+(kk1+ψrakkc)ck+ψrakkk]k^t1/ψ[ι1(I+(ck+ψrak)kaIψrakH1)+ι2((ck+ψrak)kzIψτ2H1)+cx((1+(ck+ψrak)kc)I+H1)]u^t1/ψ[cxtψ(rakι1+rZι2)t]εt+11/ψ[c0(ck+ψrak)kc].

Next,13

Vart(v^t+1)=(vut)(vut)T=vuTvuT=vuG0vu.

Finally,

Vart(mt+1+rt+1)=(lt)(lt)T=lttTlT=lG0lT,

where

l=(γ+1/ψ)vu1/ψ(cuψ(rakι1+rZι2)).

As the value function enters the Euler equation we need to approximate it by log-linearizing its definition (3) to complete the solution. Again using log-normality we can rewrite (3):

e(11/ψ)v˜t=e(11/ψ)c˜t+βe(11/ψ)(a˜t+Etv˜t+1+12(1γ)Vart(v˜t+1)),

and linearize it

ζ1v^t=ζ2c^t+ζ3(a^t+Etv^t+1+12(1γ)Vart(v^t+1)),

where

ζ1=V˜11ψζ2=C˜11ψζ3=βA˜11ψV˜11ψ.

Regrouping terms in the conditions implied by the Euler equation and the linearization of the value function definition give us the following system of equations for (c0, ck, cx, v0, vk, vx):

kcck+(kk1+ψrakkc)ck+ψrakkk=0ι1(I+(ck+ψrak)kaIψrakH1)+ι2((ck+ψrak)kzIψτ2H1)+cx((1+(ck+ψrak)kc)I+H1)12ψ((γ1/ψ)(1γ)vxG1vx+lG1lTx)=0c0(ck+ψrak)kc12ψ((γ1/ψ)(1γ)(vxG0vx+vxG1vxx)+lG0lT+lG1lTx)=0ζ1vkζ2ckζ3vk(kk+kcck)=0ζ1vxζ2cxζ3(ι1+vk(kaι1+kZι2+kccx)+vxH1+12(1γ)vxG1vx)=0ζ1v0ζ2c0ζ3(v0+vkkcc0+12(1γ)(vxG0vx+vxG1vxx))=0

In the most general case this is a system of quadratic equations we have to solve numerically.

In Example 1, the case without stochastic volatility, the solutions for the coefficients are:

ck=(kk1+ψrakkc)±(kk1+ψrakkc)24kckkψrak2kccx=(ι1(I+(ck+ψrak)kaIψrakH1)+ι2((ck+ψrak)kZIψτ2H1))((1+(ck+ψrak)kc)I+H1)1c0=(12ψ((γ1/ψ)(1γ)L1L1T+((γ+1/ψ)L11/ψL2)((γ+1/ψ)L11/ψL2)T))(ck+ψrak)kcvk=ζ2ckζ1ζ3(kk+kcck)vx=(ζ2cx+ζ3(ι1+vk(kaι1+kZι2+kccx)))(ζ1Iζ3H1)1v0=ζ2c0+ζ3(vkkcc0+12(1γ)L1L1T)ζ1ζ3,

where

L1=vuH2L2=(cuψ(rakι1+rZι2))H2ut=[atzt].

Appendix E. Example 2 Solution: Method of Unknown Coefficients

The solution in Example 2, with stochastic volatility is as follows:

ck=(kk1+ψrakkc)±(kk1+ψrakkc)24kckkψrak2kccu=1+(ck+ψrak)ka1(ck+ψrak)kccσ=12ψ((γ1/ψ)(1γ)vx2+(1/ψ(cxψrak)+(γ+1/ψ)vx)2)1+φ+(ck+ψrak)kcc0=12ψ((γ1/ψ)(1γ)(vσ2ω2+vx2θ)+(1/ψcσω+(γ+1/ψ)vσω)2+(1/ψ(cxψrak)+(γ+1/ψ)vx)2θ)(ck+ψrak)kcvk=ζ2ckζ1ζ3(kk+kcck)vu=ζ2cu+ζ3(1+vk(ka+kccx))ζ1vσ=(ζ2+ζ3vkkc)cσ+12ζ3(1γ)vx2ζ1ζ3φv0=ζ2c0+ζ3(vkkcc0+12(1γ)(vσ2ω2+vx2θ))ζ1ζ3.

Appendix F. Relation to Perturbation Methods

In this appendix we show that the standard perturbation method solution and the log-normal risk adjustment approach are closely related. For models without stochastic volatility the log-linear approximation can be written as follows:

[g(xt,σ)]i=[g(x¯,0)]i+[gx(x¯,0)]i(xx¯)+12[gσσ*(x¯,0)]i[σ][σ][h(xt,σ)]i=[h(x¯,0)]i+[hx(x¯,0)]i(xx¯)+12[hσσ*(x¯,0)]i[σ][σ].

Compared to second-order perturbation methods there are two differences. First, we drop the quadratic terms gxx and hxx. Second, gσσ, hσσ and gσσ*, hσσ* are not exactly the same.

Schmitt-Grohe and Uribe (2004) show how to compute gσσ and hσσ from other first-and second-order terms. gσσ* and hσσ* can be computed in exactly the same way except for, again, ignoring a term in gxx.

To illustrate this result, consider a simple model which nevertheless captures all the elements of the result. Control yt is a function of state variables xt and a parameter σ scaling uncertainty. State variables evolution is assumed linear and doesn’t need to be approximated,

yt=g(xt,σ)xt+1=hxxt+σηεt+1.

The equilibrium condition which allows us to approximate g as exponential in yt+1 and xt+1 is

Et(eαyt+1+βxt+1+γ)=1.

At the steady state (σ = 0),

eαy+γ = 1 y= γα.

Taking the first derivative with respect to xt and evaluating it at the steady state we obtain gx :

Et(eαyt+1+βxt+1(αgxhx+βhx))=gx=0β/α.

Similarly taking the second derivative,

Et(eαyt+1+βxt+1((αgxhx+βhx)2+αgxxhx))=gxx=00.

Next, we take the first derivative with respect to σ and are able to verify the general result that gσ is equal to zero:

Et(eαyt+1+βxt+1(αgσ+αgxηεt+1+βηεt+1))=0.

Finally, we take the second derivative with respect to σ,

Et(eαyt+1+βxt+1((αgσ+αgxηεt+1+βηεt+1)2+αgσσ+αgxxη2εt+12))=0.

gσ = 0 and in our particular example gxx = 0 therefore,

gss=(αgx+β)2η2α.

Now consider solving the model using log-normality. We assume:

yt = gxxt+g0xt+1 = hxxt+σηεt+1.

The equilibrium condition implies:

Et(αyt+1+βxt+1+γ)+12Vart(αyt+1+βxt+1+γ)=0(αgxhx+βhx)xt+αg0+γ+12(αgx+β)2η2=0.

Regrouping the terms:

gx  = β/αg0  = γ+12(αgx+β)2η2α.

We verify exactly that:

g0=y+12gss.

This will not hold exactly if gxx 0, which in our setup would have been the case if hxx 0. In other terms we compute gss*=2(g0y) in the same way as gss using standard perturbation methods except for ignoring the second-order terms in gxx.

*

We are grateful to Alex Mourmouras, Brian Routledge, Francisco Ruge-Murcia, and Harald Uhlig for insightful comments and discussions. Caroline Silverman provided excellent research assistance. All errors are ours.

McGill University. Email: aytek.malkhozov@mcgill.ca

IMF Institute. Email: mshamloo@imf.org.

1

Note that our analysis is in discrete time. For the discrete time counterpart to Duffie, Pan, and Singleton (2000) see, for example, Dai, Le, and Singleton (2010).

2

Note that since G1 is an (n ×n ×n) array, (G1)ij is a row vector of dimension (1×n):

3

Backus, Routledge, and Zin (2007) log-linearize a different set of conditions. Namely they choose the firstorder and the envelope conditions of the dynamic programming problem. We argue that the choice of the Euler equation and the definition of the value function is the correct approach. As discussed in Caldara and others (2008) in the so-called Value Function Perturbation approach, where first-order conditions are approximated, the approximation of the value function should be an order higher than the desired approximation of the solution. Intuitively, this is because derivatives of state variables appear in the first-order conditions. Under the Equilibrium Conditions Perturbation approach, one could expand all equilibrium conditions to the same order as the desired solution. Therefore, log-linearizing the envelope condition as in Backus, Routledge, and Zin (2007) is insufficient for finding a first-order approximated solution. For instance, following the Backus, Routledge, and Zin (2007) approach will not result in a value function approximation that is homogeneous of degree one in capital and productivity. We avoid this issue by linearizing the equilibrium conditions (including the Euler equation) instead.

5

Specifically, Γ will be of dimensions n ×ny; where n = nx+ny, i.e. the number of equations in the system.

6kk=(1δ)A˜1+αZA˜αK˜α1ka=(1δ)A˜1αZA˜αK˜α1+C˜K˜1A˜1kz=C˜K˜1A˜1=kkkakc=ZA˜αK˜α1

7

Note that vt+1 and v^t+1 differ only in a constant, that is the value function in the non-stochastic steady state, and as a result Vart(vt+1)=Vart(v^t+1).

8

Note that second-order perturbation method is sufficient for a model without stochastic volatility. However higher order perturbation methods are required for the example with stochastic volatility as we will discuss in Section (IV).

9

For instance, gkk or gkz, second derivatives with respect to two state variables, are omitted whereas gss, the second order derivative with respect to the variance of the shocks, is included.

10

Bloom, Floetotto, and Jaimovich (2009) has a model with stochastic volatility in consumption, but the volatility is calibrated very simply as a binary variable

11

Bloom, Floetotto, and Jaimovich (2009) model stochastic volatility in productivity as a two state Markov switching process and calibrate it to capture the high frequency spikes in uncertainty. In this paper we have in mind lower frequency movements in volatility and we model it as an autoregressive process.

12

Useful result about certainty equivalence under log-normality: if lnx ~ N(μ, V) then 1αlnE(xα)E(lnx)=12αv2.

13

κ2G1κ2T is a (1 × n) vector with kth element equal to ∑i;j κ2;iG1,ijk κ2,j.

Asset Prices in Affine Real Business Cycle Models
Author: Maral Shamloo and Aytek Malkhozov
  • View in gallery

    Comparison of Impulse Responses: Perturbation Method vs. Log-normal Risk

  • View in gallery

    Comparison of Impulse Responses: Perturbation Method vs. Log-normal Risk

  • View in gallery

    Stochastic Volatility in Productivity and Implied Stochastic Volatility in Consumption

    This figure shows the impulse response of consumption and risk-free rate in the model of Example 2 to a one standard deviation volatility shock. The results are obtained using dynare++ averaged over 30,000 simulations. The model is approximated to first, second, third and fourth-order and the results are superimposed.

  • View in gallery

    Stochastic Volatility in Productivity and Implied Stochastic Volatility in Consumption

    This figure shows the impulse response of consumption and risk-free rate in the model of Example 2 to a one standard deviation volatility shock. The results are obtained using the log-normal risk adjustment method (solid line) and are superimposed over the fourth approximation obtained using dynare++ (dashed line).

  • View in gallery

    Consumption Growth and Variance of Consumption Growth

    This figure shows the correlation between consumption growth and variance of consumption growth for different calibrations and values of ψ.