Professional Documents
Culture Documents
Computing Economic
Equilibria Using
Projection Methods
Access provided by Arizona State University on 10/21/21. For personal use only.
Annu. Rev. Econ. 2020.12:317-353. Downloaded from www.annualreviews.org
317
1. INTRODUCTION
Modern research in economics and finance is dominated by the analysis of dynamic models. In dy-
namic economics, key economic objects change over time and economic agents react to changes
in economic conditions. In many applications, the analysis of dynamic economic models leads
to the mathematical problem of determining an unknown function. Examples of such problems
arising in economic and financial modeling include solving differential equations in continuous-
time finance and economics, computing value functions that solve the Bellman equation in a dy-
namic programming problem, computing price and policy functions from Euler equations and
market-clearing conditions in dynamic equilibrium models, determining price/dividend ratios in
consumption-based asset-pricing models, and many other problems. All too often, the solutions
to these problems have no closed form. Economists must then resort to approximation techniques
in the analysis of dynamic models.
Traditionally, economists solving difficult models often ignored standard numerical methods
Access provided by Arizona State University on 10/21/21. For personal use only.
in favor of idiosyncratic and ad hoc methods with poorly understood properties that produced
Annu. Rev. Econ. 2020.12:317-353. Downloaded from www.annualreviews.org
approximate solutions of unknown quality. Judd (1998, p. 19) described the state of the art in
the late 1990s as follows: “Many ‘innovations’ in computational economics are straightforward
applications of standard numerical methods, inferior reinventions of earlier work, or ad hoc pro-
cedures with little mathematical foundation.” While Judd’s criticism has remained valid well into
the twenty-first century, the innovations of the information age during the first two decades of
the new century have begun to engulf economics and finance. Economists increasingly use so-
phisticated numerical methods as well as the latest available hardware and software in the analysis
of their models. This rapid growth in sophistication has also been apparent in the function ap-
proximations necessary in dynamic economics. Among several classes of useful methods applied
by economists, one particular class of methods stands out due to its flexibility and robustness:
projection methods.
Projection methods were developed by applied mathematicians seeking numerical solutions
to differential equations during the 1960s and 1970s (see Reddien 1980 and the many citations
therein). Although such methods had also been applied in some economic work in an ad hoc fash-
ion (see, for example, Taylor & Uhlig 1990), Judd (1992) formally introduced projection methods
to economics. Importantly, Judd observed that these methods are so flexible that they can also
be applied to many other economic problems, such as solving discrete-time dynamic economic
problems in which equilibria are characterized by functions on a continuous domain.
Applying projection methods requires that we formulate economic equilibrium conditions as a
functional equation N ( f ) = 0, where the operator N is a mapping between function spaces. The
equation N ( f ) = 0 could, for example, be a differential equation, a Bellman equation of a dynamic
programming problem, or a system of necessary first-order conditions and market-clearing equa-
tions in a recursive equilibrium problem. The solution function f is then the solution function to
the differential equation, the value function of the dynamic programming problem, or the vector
of equilibrium policy and price functions defined on an appropriate state space, respectively. For
most dynamic economic models, the functional equation does not have a closed-form solution.
Because it is impossible to represent an exact solution on a continuous domain on a computer,
all we can hope for is to find a finite representation of an approximate solution fˆ that satisfies
N ( fˆ ) ≈ 0. For this purpose, we restrict the search to a set of functions spanned by a finite set of
basis functions, {φ j }nj=0 . Any candidate solution fˆ = nj=0 a j φ j is then a linear combination of the
basis functions. Naturally, we want to choose the weights {a j }nj=0 such that some chosen norm of
the residuals N ( fˆ ) = N ( n a j φ j ) over the domain of interest for the function f is sufficiently
j=0
small in a chosen sense—for example, a norm. The minimization of some error norm can often be
Macroeconomists enthusiastically adopted perturbation methods because they are a natural ex-
tension of linearizations, which have a long history of applications in macroeconomics since the
emergence of work by Hall (1971) and Magill (1977). Perturbation methods construct an approx-
imation of an unknown function around a particular point in the domain at which the function
value is known exactly. For example, a stochastic economic model may have a deterministic steady
state for the degenerate case of no uncertainty. Perturbation methods then build an approximation
around this point relying on implicit function theorems and several orders of derivatives at this
point. In contrast to the global projection approach, economists refer to perturbation techniques
as local methods.
Naturally, the question arises of which of the two sets of techniques works better on dynamic
economic models. In their pioneering paper, Gaspar & Judd (1997) compare these methods across
a variety of dynamic economic problems with respect to computational speed and accuracy. They
conclude that perturbation methods are much faster than projection methods (on the comput-
ers available at the time of their research) and are capable of producing good approximations if
the ergodic distribution of the dynamic model has a small support. In the presence of substantial
nonlinearities, however, they conclude that projection methods should be used. About a decade
later, Aruoba et al. (2006) compare the performance of perturbation and projection methods in
the case of the neoclassical growth model. These authors strongly argue that economists should
abandon linear and log-linear approximations whenever possible in favor of higher-order pertur-
bation methods and projection methods. They prefer perturbation methods, considering them
a good compromise between accuracy, speed, and programming burden, although they do ac-
knowledge the excellent performance of projection methods with finite elements or Chebyshev
polynomials. Caldara et al. (2012) compare computational methods for solving dynamic stochas-
tic general equilibrium (DSGE) models with recursive utility and stochastic volatility and find
that perturbation methods are competitive in terms of accuracy with Chebyshev polynomials but
offer significantly faster speed. They also conclude that when numerical errors are the key con-
cern, projection with Chebyshev polynomials dominate. Fernández-Villaverde & Levintal (2018)
compare computational methods for solving DSGE models with rare disasters and find that even
higher-order perturbation methods sometimes produce unacceptably large numerical errors.
The general conclusion of comparison studies in the economic literature is clear. For compar-
atively simple models, perturbation methods can produce approximations of sufficient accuracy
faster than projection methods. For more complex models, in most cases we must abandon local
methods and apply projection methods. Such cases include, first of all, models in which economic
constraints (for example, borrowing or collateral constraints, or zero-lower-bound constraints)
and finance, respectively, and in Section 6, we describe applications in other economic fields such
as microeconomics, econometrics, and climate-change economics. In addition, we describe some
exciting new endeavors on the current research frontier. We conclude in Section 7.
2. FIRST EXAMPLE
For a first look at projection methods, we apply two such methods to a simple ordinary differential
equation (ODE). We choose this particular example because the ODE has a simple analytical
solution, which helps us illustrate the analysis of numerical errors in a projection method. We
choose these two particular methods because they are rather straightforward and intuitive.
∂f
− r f = 0, 1.
∂t
f (T ) = 1, 2.
for positive constants r and T. This ODE has the analytical solution
f ∗ (t ) = er(t−T ) . 3.
Our task is now to solve the ODE in Equation 1 numerically, subject to the boundary condition
in Equation 2. Following Judd (1992), the first step of a projection method requires us to choose
a set of basis functions with which we want to approximate the (unknown) function f ∗ . A simple
set of basis functions is a set of ordinary polynomials, {1, t, t2 , . . . , tn }, for some positive integer n.
The approximation fˆ is then defined as
n
fˆ (t; a) = a jt j , 4.
j=0
∂ fˆ
R(t; a) = − r fˆ ,
∂t
which leads to
n
n
R(t; a) = a j jt j−1 − r a jt j . 5.
j=1 j=0
Access provided by Arizona State University on 10/21/21. For personal use only.
Annu. Rev. Econ. 2020.12:317-353. Downloaded from www.annualreviews.org
n
fˆ (T ; a) = a j T j = 1. 6.
j=0
Ideally, we should like to find a coefficient vector a∗ so that the residuals satisfy R(t; a∗ ) ≡ 0 on
the domain [0, T ], but since the true function is not polynomial, such coefficients do not exist.
Instead, the approximation will typically lead to errors |R(t; a)| > 0. Therefore, in the next step
of a projection method, we need to choose the projection conditions—that is, conditions on the
residual function R that a satisfactory approximation fˆ must satisfy.
Next, we must solve the constrained minimization problem in Equation 8 in the unknown
coefficients a ∈ Rn+1 . In our specific instance, the residual function is linear in the elements of a.
Consequently, the problem in Equation 8 is a straightforward quadratic programming problem,
which even allows for a closed-form solution. For our purposes here, a numerical solution, denoted
by â, suffices.
0 5 × 10–7
f̂(t;â) – f*(t) –5 × 10–7
R(t;â) 0
f*(t)
–1 × 10–6
–5 × 10–7
–1.5 × 10–6
–1 × 10–6
2 4 6 8 10 2 4 6 8 10
t t
Figure 1
fˆ (t;â)− f ∗ (t )
(a) A plot of the relative errors, f ∗ (t )
, in the approximated ODE solution function on the entire
domain [0, T]. (b) The residuals, R(t; â). Abbreviation: ODE, ordinary differential equation.
Access provided by Arizona State University on 10/21/21. For personal use only.
Annu. Rev. Econ. 2020.12:317-353. Downloaded from www.annualreviews.org
We can now compute an approximate solution for given constants r, T > 0 and parameters m
[0,10] = {0, 1, 2, . . . , 10} and obtain
and n. For r = 0.05, T = 10, m = 10, and n = 4, we have G10
fˆ (t; â) = 0.606530 + 0.0303258t + 7.59394 · 10−4 t 2 + 1.22365 · 10−5 t 3 + 2.03659 · 10−7 t 4 .
In the final step of a projection method, we want to verify that our solution yields a reason-
able approximation. Since we have an exact solution in Equation 3 for the model, we can easily
determine the relative approximation errors,
fˆ (t; â) − f ∗ (t )
∀ t ∈ [0, T ]. 9.
f ∗ (t )
Figure 1a shows a plot of these relative errors. Due to the constraint fˆ (T ; a) = 1, the relative
error in the solution function fˆ (t; â) is zero at t = T = 10. The maximal relative error is less than
2 × 10−6 . In typical applications of projection methods, however, we do not know an exact solution
f ∗ , so we cannot determine the relative errors of Equation 9 in the approximation of the unknown
function. Instead, all we can do is to compute the residuals R(t; â). Figure 1b shows a plot of the
residuals. We observe that the maximal absolute residual is less than 10−6 . If we consider this error
to be too large, we could now try to increase the degree n of the approximating polynomial in the
hope of reducing the numerical errors.
n
j−1
n
j
a j jti −r a j ti = 0 ∀ ti ∈ Gn−1
[0,T ] , 10.
j=1 j=0
n
a j T j = 1. 11.
j=0
The system is solved for given positive constants r and T and parameters m + 1 = n. For r =
0.05, T = 10, and m = 5 (so n = 6), the linear system has seven equations and unknowns. We have
G5[0,10] = {0, 2, 4, 6, 8, 10} and obtain
fˆ (t; â) = 0.606531 + 0.0303265t + 7.58165 · 10−4 t 2 + 1.26348 · 10−5 t 3 + 1.58317 · 10−7 t 4
+ 1.52554 · 10−9 t 5 + 1.69504 · 10−11t 6 .
Figure 2a depicts a plot of the relative errors in Equation 9. Recall that without knowing the
exact solution we would not be able to determine these errors. Instead, all we can typically do is
to compute the residuals R(t; â) (see Figure 2b for a plot of the residuals). We observe that the
residuals are indeed zero at the six elements of G5[0,10] = {0, 2, 4, 6, 8, 10}. In addition, the relative
error in the solution function fˆ (t; â) is also zero at t = T = 10 due to the constraint fˆ (T ; a) = 1.
3.1.1. Step 1: identifying the functional equation describing the economic problem. For
an application of projection methods, we need to formulate the economic problem as finding a
function f : D ⊂ RK → RM that must satisfy a functional equation
N ( f ) = 0, 12.
where N is a mapping (i.e., an operator) from a function space B1 to a function space B2 . Note that
f ࢠ B1 and N ( f ) ∈ B2 . Equation 12 is satisfied if and only if N ( f (x)) = 0 for all x ࢠ D. The
Therefore, we must restrict attention to functions fˆ that have a finite representation. And so, for
some finite n ∈ N, we choose an (n + 1)-dimensional vector space of functions V n that is a subset
n
of B1 . Denote by n = φ j j=0 a basis of V n , so that we obtain
⎧ ⎫
⎨
n ⎬
Vn = g | g(x; a) = a j φ j (x), x ∈ D, a ∈ Rn+1 . 14.
⎩ ⎭
j=0
The projection method then restricts attention to functions fˆ ∈ V n ⊂ B1 in the search for an
approximate solution N ( fˆ ) ≈ 0.
For the ODE in Section 2, the basis n is the set {1, t, t2 , . . . , tn } of the first (n + 1) monomials,
which spans the (n + 1)-dimensional vector space of polynomials of degree less than or equal to
n. Hence, we obtain
⎧ ⎫
⎨
n ⎬
fˆ ∈ g | g(t; a) = a j t , t ∈ [0, T ], a ∈ R
j n+1
.
⎩ ⎭
j=0
3.1.3. Step 3: defining a residual function. Projection methods return a function fˆ ∈ V n that
yields a reasonable approximation of the true f. For the evaluation of possible approximations, we
define the residual function R : D × Rn+1 → RM by R(·; a) = N ( fˆ ) with the term
⎛ ⎞
n
R(x; a) = N ⎝ a j φ j (x)⎠. 15.
j=0
The residual R(x; a) is the error in the equation N ( fˆ (x; a)) = 0. Note that we do not approximate
the operator N but rather search for an approximation fˆ of the unknown function f. Judd (1992)
points out that in many economic applications the operator N cannot be represented exactly on
a computer, and thus it requires an approximation as well.
For the ODE in Section 2, we obtain the residual R by replacing f (t) by fˆ (t; a) in
Equation 13:
n n
j=1 ja j t j−1 − r j=0 a jt j
R(t; a) = n . 16.
j=0 ajT j − 1
R(x; a) = 0 ∀ x ∈ G ⊂ D. 17.
Yet another alternative approach to determining the unknown coefficients a ∈ Rn+1 is to min-
Access provided by Arizona State University on 10/21/21. For personal use only.
Annu. Rev. Econ. 2020.12:317-353. Downloaded from www.annualreviews.org
min
R(x; a)
2G such that a ∈ Rn+1 . 19.
3.1.5. Step 5: evaluating the approximate solution. We evaluate the quality of the approxi-
mate solution by determining a close upper bound on the residuals R(x; â) on the domain D. If we
are assured that
for some predefined error tolerance (for example, = 10−8 ), then we accept the computed solu-
tion fˆ (x; â) as an acceptable approximation for the true function term f ∗ (x). If, however, the max-
imal residuals exceed the error tolerance, then we change the function representation for fˆ (x; a),
for example, by increasing the number of basis functions. We return to Step 4 and compute a new
approximation. We repeat these steps until the condition in Equation 20 is satisfied.
The least-squares approach in Section 2.2 delivered an upper bound on the residuals of 10−6 .
This bound exceeds = 10−8 . So, if we require the maximal residual to fall below this bound, we
may want to increase n, repeat the optimization, and check whether we can reduce the error. The
uniform collocation approach in Section 2.3 delivers an upper bound on the residuals below 10−9
and thus satisfies our error requirement.
for some weighting function w(x). Using inner products, we can now formally state all methods
in Step 4 in a standardized fashion. All methods determine the unknown coefficients a such that
with the weighting function w(x) = 1 for all x ࢠ D. In econometrics these conditions are often
formulated in discrete form on a finite set of data points G ⊆ D, so that we have
∂R(xi ; a)
R(xi ; a) = 0, i ∈ {0, 1, . . . , |G| − 1}.
i
∂a j
with δ denoting the Dirac delta function [and the weighting function w(x) = 1 for all x ࢠ D].
In sum, although the notation via projections allows for mathematically elegant statements,
these are not directly helpful to an explanation or in the implementation of projection methods.
x−L U −L U +L U −L
y=2 − 1 and x= y+ = (y + 1) + L. 22.
U −L 2 2 2
The resulting grid for the problem is then given by
⎧ ⎫
⎨U − L ⎬
U + L 2i − 1
G= y+ y = cos π , i = 1, 2, . . . , n + 1 , 23.
⎩ 2 2 2(n + 1) ⎭
where cos 2i−1
2(n+1)
π is a convenient representation of the Chebyshev zeros. The approximation
to f (x) is
n
x−L
fˆ (x; a) = a j Tj 2 −1 , 24.
j=0
U −L
where we retransform the grid G to ensure the evaluation of Chebyshev polynomials on [−1, 1].
After replacing the unknown term f (x) by fˆ (x; a) in the residual R, we must then solve the resulting
square system of residual equations.
An alternative approach, which closely follows the language of projection, is Galerkin’s method.
The conditions to pin down the coefficients are orthogonality conditions on the residual function
and the basis functions. On the domain [−1, 1], these conditions are
1
1
R(y; a)Tj (y) dy = 0 for j = 0, 1, . . . , n.
−1 1 − y2
Since we need these conditions typically on a different domain [L, U], we apply again the linear
change of variables in Equation 22. For a general residual R(x; a) on [L, U], the Galerkin conditions
are
U
x−L 1
R(x; a)Tj 2 −1
2 dx = 0 for j = 0, 1, . . . , n. 25.
U − L
L
1 − 2 Ux−L
−L
−1
Before we can solve the system in Equation 25 we must evaluate the integral in each equation. Due
to the inclusion of the particular weight function, a great choice for the numerical evaluation of
U −L
Applying the transformation in Equation 22 including dx = 2
dz and this quadrature rule to the
integral in Equation 25, we obtain the system of equations
Access provided by Arizona State University on 10/21/21. For personal use only.
Annu. Rev. Econ. 2020.12:317-353. Downloaded from www.annualreviews.org
U −Lπ
k
U −L
R (zi + 1) + L; a Tj (zi ) = 0 for j = 0, 1, . . . , n. 26.
2 k i=1 2
We must now solve this square system of equations in the unknown coefficient vector a to obtain
the desired approximation fˆ (x; â) of f (x).
Observe that both the Chebyshev collocation method and Galerkin’s method require us to
solve a square system of equations. Even though in some rare cases the system may be linear,
we typically face a nonlinear system of equations, and such a nonlinear system may have several
solutions.
a grid, while CGA further clusters these simulated points to represent each such cluster with one
central grid point. Besides, based on the very structure of the model, the optimal policy is split
into the optimal consumption and labor (static, intratemporal policy) and the optimal rule for
capital (dynamic, intertemporal policy). Solving the intratemporal problem in isolation—that is,
conditioned on the choice for capital—greatly improves the accuracy of the overall solution to
the model. It is therefore a smart choice of the approximation, integration, and grid construction
techniques—together with an informed approach to the problem itself—that delivers the best
resolution of the trade-off between computational costs and the accuracy of the obtained solution.
keeps the number of points where the model is evaluated moderately (polynomially) growing. The
challenging feature of their model is that the production shocks change the wealth distribution;
therefore, the future wealth distributions are part of the state space of an agent’s problem. The
algorithm can nevertheless handle the model with up to 30 agents.
The method pioneered by Krueger & Kubler (2004) enables the assessment of the welfare
effects of social security systems in stochastic OLG models. Sánchez-Marcos & Sánchez-Martín
(2006) use the method to solve an OLG model with demographic shocks and a pay-as-you-go
pension system; this system appears to have a limited ability of smoothing the welfare effect of
the fluctuations in cohort sizes. The introduction of the pension system to the economy has an
insurance effect of intergenerational risk sharing. The consequent welfare gains, however, do not
outweigh the reduction in per-capita income in comparison to the crowding out of private savings
in the absence of pensions. Gottardi & Kubler (2011) subsequently show that the introduction of a
social security system can in fact be Pareto-improving, given a few characteristics of the economy.
Their model with production shocks features ex-ante welfare evaluation and an infinitely lived
asset (land). These two conditions appear crucial for a pay-as-you-go pension system to prove
beneficial: The former carries the potential of a Pareto-improvement even when the markets are
complete, while the latter can compensate for the welfare loss from decreasing capital stock.
Heterogeneous agent models in general share the challenge of a state space expanded by the
cross-sectional distributions of state variables. In addition to the aforementioned tools, this chal-
lenge can be tackled with hybrid approaches that combine the best features of projection and
perturbation methods. Reiter (2009) uses such a hybrid approach to solve a heterogeneous agents
model with both idiosyncratic and aggregate shocks and liquidity constraints. The projection part
of the algorithm parameterizes the stationary solution of the model with idiosyncratic shocks only,
while the perturbation part consequently linearizes the solution to the model with added aggre-
gate shocks. Reiter (2015) solves a large OLG model with an inverse technique—it starts with a
solution linearized very close to the steady state and gradually builds a global, projection-based
solution as the aggregate shocks grow.
2 Sections 6.2 and 6.3 provide more details on efficient grid construction.
apply Chebyshev collocation to resolve a trade-off between debt and taxes in an economy with
conflicting views of different generations on the provision of public goods. In this conflict, only a
strong concern with public goods can induce fiscal discipline and avoid the otherwise inevitable
debt accumulation and deterioration of public goods. Niemann et al. (2013) use the collocation
method to analyze the interplay between fiscal and monetary policy when, in contrast to the cen-
tral bank, the fiscal authority is impatient. The central bank’s choice between money growth rate
and nominal interest rate as a policy instrument turns out to determine the equilibrium allocation.
in this strand of literature (see Lan 2018 for a review of accuracies for different calibrations of
Annu. Rev. Econ. 2020.12:317-353. Downloaded from www.annualreviews.org
the DMP model). Petrosky-Nadeau et al. (2018) show that—embedding the stylized labor market
into a general equilibrium economy and accounting for nonlinearities introduced by the match-
ing process—a series of negative disturbances can depress the economy into disasters. Particularly,
Petrosky-Nadeau & Zhang (2017) show that, for the correct solution, recent calibrations of the
DMP model as presented by Hagedorn & Manovskii (2008) understate unemployment volatility
by a factor of two and strongly overstate negative correlation of vacancies and unemployment.
In the following, we lay out the key equations of the DMP model and demonstrate how we can
solve them with a projection algorithm (for more details see Petrosky-Nadeau & Zhang 2017).
Yt = Xt Nt , 27.
where Nt is the level of employment and Xt denotes marginal productivity. Logarithmic produc-
tivity, xt , follows the AR(1) autoregressive process
with εt + 1 being a random shock with standard normal distribution, σ being its volatility, and
0 ≤ ρ < 1 being the persistence of log productivity. Family members face the risk of being either
employed or unemployed, for which they grant each other insurance. Employment dynamics in
the economy are described by
with s denoting the job separation rate. The amount of vacancies posted is given by Vt , and the
job-filling rate q(θ t ) is defined as
G(Vt , Ut ) 1
q(θt ) = =
1/ι 30.
Vt 1 + θtι
where κ t is the cost of posting one vacancy. The term on the left-hand side of Equation 31 is the
difference between the marginal cost of one hire and the Lagrange multiplier λt ; the multiplier
denotes the marginal cost of not being able to post negative vacancies—that is, of not actively
firing employees. The accompanying Kuhn–Tucker conditions are
Access provided by Arizona State University on 10/21/21. For personal use only.
Annu. Rev. Econ. 2020.12:317-353. Downloaded from www.annualreviews.org
q (θt ) Vt ≤ 0, λt ≥ 0, λt q (θt ) Vt = 0.
The right-hand side of Equation 31 depends on two terms. First, the cash-flow term, Xt + 1 −
Wt + 1 , denotes the difference between a worker’s marginal output, Xt + 1 , and their wage, Wt + 1 .
The wage follows from Nash bargaining between the firm and the worker,
Wt = η (Xt + κt θt ) + (1 − η) b,
where b denotes the cash flow in case of unemployment. The second term of the right-hand side
of Equation 31 accounts for the continuation value of having one worker employed, discounted by
the probability of separation. The firm’s inability to actively fire a worker also lowers the continu-
ation value, just as on the left-hand side of the equation. Due to the linear production technology,
the level of employment does not alter the marginal productivity of labor, and so employment
does not show up in the Euler Equation 31. Solving for an equilibrium now requires us to solve
this Euler equation by approximating unknown functions for θ and λ.
4.7.2. The projection solution. To apply projection methods, we view Equation 31 not as a
sequence of conditions in infinite discrete time, t = 0, 1, 2, . . . , but instead as a functional equation.
The exogenous stochastic process xt has as support the real line, R. The two unknown functions
θ, λ : R → R++ must satisfy the functional equation
! "
κ (x) κ (x )
− λ(x) = E β ex − W (x ) + (1 − s) − λ(x
) x , 32.
q (θ (x)) q (θ (x ))
with
x = ρx + ε, ε ∼ N (0, σ 2 ).
In the formulation of the functional equation, we use the fact that the variables W, κ, and q are
explicitly given functions of x.
Compared to the introductory example in Section 2, this equation entails a nontrivial complica-
tion. The functional equation has to be solved subject to the nonnegativity constraint on vacancies.
We make sure that our solution satisfies this condition using the Kuhn–Tucker conditions below.
Annu. Rev. Econ. 2020.12:317-353. Downloaded from www.annualreviews.org
∞ ρx+ε
κ (ρx + ε) 1 − ε
2
β e − W (ρx + ε) + (1 − s) − λ (ρx + ε) √ e 2σ 2 dε,
−∞ q (θ (ρx + ε)) σ 2π
where we replace the expectations operator from Equation 31 with an integral including the nor-
mal density. We must approximate this integral. For this purpose we apply Gauss–Hermite quadra-
ture (see Judd 1998). This approximation replaces the integral by a finite sum. And so we obtain
an expression that includes values of the unknown functions at a few points in the domain,
#
n √ √
wiGH β eρx+ 2σ εi
− W (ρx + 2σ εi )
i=1 ⎛ √ ⎞⎞ 33.
κ (ρx + 2σ ε ) √
+ (1 − s) ⎝ − λ ρx + 2σ εi ⎠⎠ ,
i
√
q θ ρx + 2σ εi
with the Gauss–Hermite quadrature nodes ε i and weights wiGH . Note that here we replace the
right-hand side in the exact functional Equation 32, involving an integral, by the approximation
in Equation 33, involving a finite sum. This completes Step 1, the identification of the functional
equation. However, before moving to Step 2, we need to take the time to specify how we move
from this approximation to obtaining θ and λ. Denote the term in Equation 33 by E (x). We need
to differentiate between two cases—namely, whether the constraint is binding or not.
Case 1 (not binding). To obtain θ, we begin by assuming that the constraint on vacancies
is not binding. Therefore, λ = 0 and Equation 32 simplifies to
κ (x)
q∗ (θ (x)) = , 34.
E (x)
with the job filling probability (see Equation 30) being weakly between 0 and 1.
Case 2 (binding). Once κ (x)/E (x) > 1, which would mean that the probability of filling
a vacancy exceeds 1, the algorithm ought to set q∗ (θ(x)) = 1 and calculate the Lagrange
multiplier as
κ (x)
λ (x) = − E (x). 35.
1
m
x−L
Ê (x; a) = a j Tj 2 −1 ,
j=0
U −L
where we need to specify the interval [L, U] of interest for our application. In Step 3, we replace
the function terms E (x) in Equation 33 by the polynomial approximation Ê (x; a) and then obtain
the residual R(x; a) for this model. Observe that here we approximate only one function with a
Access provided by Arizona State University on 10/21/21. For personal use only.
polynomial, namely Ê (x; a). The other two variables, λ̂(x, â) and θ̂ (x, â), follow from the case-by-
Annu. Rev. Econ. 2020.12:317-353. Downloaded from www.annualreviews.org
case analysis and are, once we have approximated E, given in closed form.
For Step 4 we choose the Galerkin projection method with a Chebyshev basis. Therefore, we
need to solve the system of equations
k
U −L
R (zi + 1) + L; a Tj (zi ) = 0 for j = 0, 1, . . . , n
i=1
2
5. MACRO-FINANCE
Numerical methods in the macroeconomics literature are of use to approximate policy functions
describing agents’ optimal reactions to changes in their economic environments. These policy
functions can, for instance, determine the amount of labor deployed, the utilization of production
facilities, or the investments in the capital stock. A projection—that is to say, a global solution—is
well suited when accounting for extreme curvature, changes of curvature, and even kinks intro-
duced by, for example, borrowing or nonnegativity constraints on investments. Contrary to this
focus on the real economy, the finance literature prevailingly focuses on the impact of macroe-
conomic quantities on financial markets. Here, projection methods are used to approximate vari-
ables such as the risk-free rate, the price/dividend ratio, or the conditional expected return of an
asset.
0.3 1.5
0.1 0.5
0 0
–0.05 0 0.05 –0.05 0 0.05
x x
4
–4
0 –10
–0.05 0 0.05 –0.05 0 0.05
x x
Figure 3
(a) The approximated Lagrange multiplier, λ̂(x; â). (b) Labor market tightness, θ̂ (x; â). (c) The marginal value
of a worker, Ê (x; â). (d) The log10 of the relative Euler equation error, |Ê (x;â)−E
E (x)
(x)|
. All four graphs depict
functions on the domain [−0.0852, 0.0852]. The economic parameters underlying the graphs are the
time-preference parameter, β = 0.991/12 ; the persistence of log productivity, ρ = 0.9895; the volatility
parameter, σ = 0.0034; the workers’ bargaining weight, η = 0.052; the workers’ flow value of unemployment
activities, b = 0.955; the job separation rate, s = 0.0081; and the elasticity of the matching function, ι =
0.407. The graphs show numerical errors using Gauss–Hermite quadrature with 15 nodes and weights and
Chebyshev polynomials of degree 15. The approximations are obtained for the Galerkin method using 25
points. Productivity is represented by x. The polynomial approximation, Ê, of the marginal value of a worker,
E, has as arguments the productivity x and the coefficient vector â.
explanation for the extended period of low growth after the financial crisis of 2007. Again, due to
the constraint on leverage, the model is solved using a projection algorithm.
Another class of market frictions that has received much attention after the 2007 financial cri-
sis is borrowing constraints resulting from collateral requirements. In many circumstances, bor-
rowers must put up substantial amounts of collateral assets to receive loans from lenders. Brumm
et al. (2015a) investigate such collateral constraints and their impact on asset returns in an infinite-
horizon general equilibrium model with heterogeneous agents. They show that borrowing against
collateral substantially increases the return volatility of long-lived assets such as stocks. To solve
their model, they apply the projection method of Brumm & Grill (2014), which is explicitly de-
signed for occasionally binding constraints and uses piecewise linear functions as bases. Brumm
et al. (2015b) use the same approach to explain why the active management of margin require-
ments on US stock markets from 1947 until 1974 had little if any measurable effect on market
volatility.
Ct+1 −γ
Pt = Et β (Pt+1 + Dt+1 ) for t = 0, 1, 2, . . . .
Ct
Here, β ࢠ (0, 1) denotes the agent’s discount factor and γ > 0 denotes the agent’s coefficient of
relative risk aversion. Under the assumption that dividends are the only source of income, Ct =
Dt , we can rewrite the Euler equation as
$
Pt Ct+1 1−γ Pt+1
= Et β +1 .
Access provided by Arizona State University on 10/21/21. For personal use only.
Ct Ct Ct+1
Annu. Rev. Econ. 2020.12:317-353. Downloaded from www.annualreviews.org
Following Burnside (1998), we define the price/consumption ratio v t = Pt /Ct and obtain
$
Ct+1 1−γ
vt = Et β (vt+1 + 1) . 36.
Ct
with the persistence parameter ρ ࢠ (−1, 1). Since the exogenous process provides log consump-
tion growth xt instead of actual consumption or consumption growth, it is helpful to rewrite the
equation for the price/consumption ratio as
$
1−γ
vt = Et β ext+1 (vt+1 + 1) . 37.
Burnside (1998) derives a necessary and sufficient condition that a finite solution v t of Equation
37 exists given a value for the exogenous process xt . The price/consumption ratio Burnside (1998,
equation 19) is
∞
vt = β i eai +bi (xt −μ) , 38.
i=1
! "
1 σ2 ρ 21−ρ
2i
ai = (1 − γ )iμ + (1 − γ )2 i − 2 (1 − ρ i
) + ρ
2 (1 − ρ )2 1−ρ 1 − ρ2
and
ρ
bi = (1 − γ ) (1 − ρ i ),
1−ρ
5.3.2. Projection solutions. To apply projection methods, we view Equation 37 not as a se-
quence of conditions in infinite discrete time, t = 0, 1, 2, . . . , but instead as a functional equa-
tion. The exogenous stochastic process xt has as support the real line, R. The unknown function
v : R → R++ must satisfy the functional equation
1−γ
v(x) = E β ex (v(x ) + 1) x 39.
Access provided by Arizona State University on 10/21/21. For personal use only.
Annu. Rev. Econ. 2020.12:317-353. Downloaded from www.annualreviews.org
with
x = (1 − ρ )μ + ρx + , ∼ N (0, σ 2 ).
The conditional expectation of x given x is just an expected value based on the normal distribution
of the shock . So, we can simply write the functional equation as
∞ 1−γ 1 2
−
v(x) = β e(1−ρ )μ+ρx+ v((1 − ρ )μ + ρx + ) + 1 √ e 2σ 2 d. 40.
−∞ σ 2π
To obtain a numerical approximation of the integral on the right-hand side, we apply Gauss–
Hermite quadrature (see Judd 1998). This approximation replaces the integral by a finite sum.
And so we obtain an expression that includes values of the unknown function v at a few points in
the domain, namely
with the Gauss–Hermite quadrature nodes i and weights wiGH . Note that here we replace the
exact functional Equation 40, involving an integral, by an approximation in Equation 41, involving
a finite sum. This completes Step 1, the identification of the functional equation.
In the following we explain how we can approximate the unknown price-consumption func-
tion v by a sum of Chebyshev polynomials. We employ our two favorite projection methods,
Chebyshev collocation and Galerkin projection with Chebyshev polynomials, which we described
in Section 3.3.
In Step 2, we use Chebyshev polynomials (of the first kind) as a basis of V n . So, the polynomial
approximation term v̂(x) is of the form
m
x−L
v̂(x; a) = a j Tj 2 −1 ,
j=0
U −L
where we need to specify the interval [L, U] of interest for our application. In Step 3, we replace
the function terms v(x) in Equation 41 by the polynomial approximation v̂(x; a). The resulting
left-hand side is the residual R(x; a) for this model.
–5 × 10–6 6 × 10–7
–0.000010 4 × 10–7
–0.000015 2 × 10–7
v̂ (x;â) – v*(x)
–0.000020 R(x;â) 0
v*(x)
–0.000025 –2 × 10–7
–0.000030 –4 × 10–7
0 –6 × 10–7
–0.05 0 0.05 0.10 –0.05 0 0.05 0.10
x x
Figure 4
∗
(a) A plot of the relative errors, v̂(x;â)−v
v ∗ (x)
(x)
, in the approximated solution function on the domain [−0.06, 0.1]. (b) The residuals, R(x; â).
The economic parameters underlying the two graphs are β = 0.95, γ = 5, ρ = 0.8, μ = 0.02, and σ = 0.0205. The graphs show
numerical errors using Gauss–Hermite quadrature with 15 nodes and weights and Chebyshev polynomials of degrees 0 to 8. The state
Access provided by Arizona State University on 10/21/21. For personal use only.
Annu. Rev. Econ. 2020.12:317-353. Downloaded from www.annualreviews.org
variable, log consumption growth, is denoted by x. The polynomial approximation, v̂, of the price/consumption ratio, v ∗ , has as
arguments the variable x and the coefficient vector â.
In Step 4 we choose Chebyshev collocation and use the m + 1 zeros of the Chebyshev polyno-
mial of degree m + 1 as collocation nodes. We obtain a linear square system of m + 1 equations—
one residual equation R(xj ; a) = 0 for each grid point xj ࢠ G, j = 0, 1, . . . , m. The collocation
points are the zeros of the (m + 1)-degree Chebyshev polynomial transformed to the interval
[L, U] = [−0.06, 0.1] for the state variable x. We sum the first 1,000 elements of the infinite sum
in Equation 38 to obtain (an excellent approximation of ) the exact solution to the model.
Figure 4a shows a plot of the relative errors in Equation 9. Recall that without knowing the
exact solution we would not be able to determine these errors. Instead, all we can typically do is
to compute the residuals R(x; â) (see Figure 4b for a plot of these residuals on [−0.06, 0.1]). We
observe that the absolute relative errors are smaller than 4 × 10−5 , while the absolute residuals
are smaller than 7 × 10−7 . The residuals closely resemble the equi-oscillation property, which we
expect from a good approximation via Chebyshev polynomials.
Lorenz et al. (2020) compare the results of the workhorse model of this literature presented
by Drechsler & Yaron (2010) with an accurate solution obtained by a projection algorithm using
complete polynomials. They show that the additions to state variable dynamics severely worsen
the inaccuracies introduced by log-linearization of returns, up to the point that their qualitative
and quantitative results are largely driven by numerical errors.
Another possible extension to long-run risk models is to modify the agent’s preferences beyond
the recursive time aggregation of utility. Gallant et al. (2019) use a Bayesian method to estimate
consumption-based asset-pricing models featuring smooth-ambiguity preferences. All their esti-
mations rely on a model solution obtained with collocation projection using Chebyshev polyno-
mials. They show in a model comparison that models with ambiguity, learning, and time-varying
volatility are preferred to the standard long-run risk model. Lorenz & Schumacher (2018) study
the role of generalized disappointment aversion (cf. Routledge & Zin 2010) in explaining asset-
pricing quantities, using insights from behavioral economics. The type of preferences introduces
large nonlinearities, which the authors account for by using a Galerkin projection approach to
solve the model.
While all of the above build on a single-agent economy, Pohl et al. (2020) add a second agent
to the model to establish trade. The agents disagree on the persistence of the long-run growth
rate of consumption. Adding a second agent introduces an endogenous state variable—the wealth
share—complicating the solution procedure. The wealth share evolves over time and influences
the optimal consumption decision of agents. The value functions of the two agents are a function
of both exogenous and endogenous state variables. The value functions are approximated with a
two-dimension cubic spline. On the boundaries, with the wealth share being 0 and 1, the solution
is replaced by the solution of an otherwise equal representative-agent economy. Pohl et al. (2020)
use a discrete-time framework and a full nonlinear solution. Branger et al. (2016) use a continuous-
time setup, featuring disagreement on jump intensities and a solution method inspired by Benzoni
et al. (2011) and outlined above. The final solution is obtained with a partial differential equation
(PDE) solver from the Numerical Algorithms Group (NAG) (http://nag.com). Again, for the lim-
iting case and a boundary condition of the PDE, the authors use a representative-agent economy
solution.
DSGE models with asset-pricing implications, higher-order perturbation methods are ill suited.
They conduct an analysis for economies with transitory and permanent shocks to total factor pro-
ductivity. With calibrations featuring high volatility and risk aversion, asset-pricing quantities tend
to not be accurately matched by the perturbation method. They explain this by the fact that a local
method does very poorly at capturing the curvature of the value function on the whole domain of
the state variables.
Gräber & Schumacher (2019) extend this analysis to the effects of persistent growth risks
on the solution accuracy of DSGE models as introduced by Croce (2014).3 They show that the
presence of persistent growth risks severely worsens the performance of a perturbation-based so-
lution. In a high-volatility calibration, the risk-free rate is understated by more than 100% and the
wealth/consumption ratio is approximately 65% below the true value. Also, macroeconomic values
such as the volatility of investment growth are wrongly approximated by almost 10%. They fur-
ther show that even a second-order collocation projection performs significantly better than any
solution obtained by perturbation. Boldrin et al. (2001) add habit formation and a two-sector setup
to a standard RBC model to improve asset-pricing performance. The model’s equilibrium condi-
tions are solved utilizing a projection algorithm. The authors display impulse response functions
to analyze differences in reactions to exogenous disturbances in both sectors. Petrosky-Nadeau
et al. (2013) use early insights of Petrosky-Nadeau et al. (2018) and study the asset-pricing impli-
cations of matching frictions in labor markets. The solution method to this framework is discussed
in detail in Sections 4.7.1 and 4.7.2. Their economy featuring a DMP-matching process generates
a sizable equity risk premium together with high stock market volatility while also producing a
low risk-free rate. Excess returns are countercyclical and predictable.
explain joint phenomena of the real economy and asset prices. They all use perturbation-based solutions for
their analyses. Favilukis & Lin (2013) also use a setup featuring shocks to the growth rate of productivity but
solve their model with a value function algorithm. They use this framework to explain several salient properties
of investment in a production economy.
Macroeconomics and finance are certainly the two fields in economic modeling that have seen
the most applications of projection methods. However, it would be an omission not to mention
at least briefly a few applications in microeconomics and climate-change modeling, which we do
next. Moreover, we survey other numerical methods that involve projection techniques. Finally,
we briefly describe recent progress on solving models with large-dimensional state spaces and on
taking advantage of parallel computing on high-performance computing architectures.
Bansal et al. (2016) further associate a climate-economic model and the long-run risk frame-
work of finance. In their joint model, solved by a projection-based algorithm, the damage induced
by rising temperatures to economic growth transfers into a positive risk premium on equity mar-
kets. The information from these markets is thereby used to estimate the costs of optimal climate
policy; in particular, it leads to a high optimal carbon tax.
The dynamic stochastic integrated model of climate and economy (DSICE) model used by Cai
et al. (2017) is an impressive example of both advanced computational methods and high-power
computing applied to a climate-economic problem. The model has nine state variables, is enriched
by both climate and economic uncertainty, and spans centuries with annual time steps. Solving a
problem of this size is only feasible on a supercomputer, yet the algorithm builds on projection
techniques in that it approximates the value function with multivariate orthogonal polynomials
and solves the Bellman equation with value function iteration. The ultimate finding is that the
climate and economic risks might well raise the optimal carbon tax tenfold.
and selects the points that represent these clusters. The accuracy of this clustering method corre-
sponds to the density of the points, with better fit in high-probability areas. The second method
selects only a subset of points based on their distance from each other. This distance can be further
flexibly adjusted to provide a more dense grid in the areas that are harder to fit.
a model’s ergodic set from simulations makes it grid-free and flexible enough to handle irregular
shapes. The power of the method is demonstrated on a large OLG model à la Krueger & Kubler
(2004)—expanded further with borrowing constraints in addition to aggregate stochastic shocks
and agents living up to 60 periods. The structure overall leaves the model with around 120 dimen-
sions, successfully processed by the neural networks. Seemingly unorthodox, the algorithm can be
generally classified as a projection method in which neural networks approximate the equilibrium
conditions and the loss function essentially minimizes mean squared errors in these conditions.
7. CONCLUSION
In this review we have discussed the impact of the use of projection methods on different strands
of economic research. We demonstrated that these methods are an excellent choice of numer-
ical techniques to address the nonlinearities arising from constraints, other potentially smooth
frictions, or extreme dynamics of state variables. Not only can these methods deliver accurate ap-
proximations of nonlinear solution functions, but—because they incorporate information from
several points in the domain—they can also do so on large domains. In recent years, projection
methods have delivered new insights into prominent models in economics and finance, which less
accurate solutions, obtained in older studies using simpler techniques, had missed.
As dynamic economic models continue to evolve, so must the numerical methods to solve them.
In light of the recent advances in projection methods, such as new adaptive-grid methods and
the implementation of such methods on high-performance computer architectures, we strongly
believe that projection methods will be the methods of choice for many years to come.
DISCLOSURE STATEMENT
The authors are not aware of any affiliations, memberships, funding, or financial holdings that
might be perceived as affecting the objectivity of this review.
ACKNOWLEDGMENTS
We are grateful to an anonymous referee, Ken Judd, and Harry Paarsch for detailed comments on
an earlier draft. We thank Thomas Cosimano, Uli Doraszelski, Gregor Reich, Simon Scheidegger,
and Ole Wilms for helpful suggestions.
Atolia M, Awad B, Marquis M. 2011. Linearization and higher-order approximations: How good are they?
Comput. Econ. 38:1–31
Atolia M, Chatterjee S, Turnovsky SJ. 2010. How misleading is linearization? Evaluating the dynamics of the
neoclassical growth model. J. Econ. Dyn. Control 34:1550–71
Azacis H, Gillman M. 2010. Flat tax reform: the Baltics 2000–2007. J. Macroecon. 32:692–708
Azinović M, Gaegauf L, Scheidegger S. 2019. Deep equilibrium nets. Work. Pap., Univ. Zurich, Zurich, Switz.
Bajari P. 2001. Comparing competition and collusion: a numerical approach. Econ. Theory 18:187–205
Bansal R, Kiku D, Ochoa M. 2016. Price of long-run temperature shifts in capital markets. NBER Work. Pap.
22529
Bansal R, Yaron A. 2004. Risks for the long run: a potential resolution of asset pricing puzzles. J. Finance
59:1481–509
Becker RA. 1980. On the long-run steady state in a simple dynamic model of equilibrium with heterogeneous
households. Q. J. Econ. 95:375–82
Benzoni L, Collin-Dufresne P, Goldstein RS. 2011. Explaining asset pricing puzzles associated with the 1987
market crash. J. Financ. Econ. 101:552–73
Bocola L. 2016. The pass-through of sovereign risk. J. Political Econ. 124:879–926
Boldrin M, Christiano LJ, Fisher JD. 2001. Habit persistence, asset returns, and the business cycle. Am. Econ.
Rev. 91:149–66
Branger N, Konermann P, Schlag C. 2016. Optimists, pessimists, and the stock market: the role of preferences and
market (in)completeness. Work. Pap., Münster Univ., Münster, Ger.
Brumm J, Grill M. 2014. Computing equilibria in dynamic models with occasionally binding constraints.
J. Econ. Dyn. Control 38:142–60
Brumm J, Grill M, Kubler F, Schmedders K. 2015a. Collateral requirements and asset prices. Int. Econ. Rev.
56:1–25
Brumm J, Grill M, Kubler F, Schmedders K. 2015b. Margin regulation and volatility. J. Monet. Econ. 75:54–68
Brumm J, Scheidegger S. 2017. Using adaptive sparse grids to solve high-dimensional dynamic models.
Econometrica 85:1575–612
Bungartz HJ, Griebel M. 2004. Sparse grids. Acta Numer. 13:147–269
Burnside C. 1998. Solving asset pricing models with Gaussian shocks. J. Econ. Dyn. Control 22:329–40
Cai Y, Judd KL, Lontzek TS. 2017. The social cost of carbon with economic and climate risks. Work. Pap. 18113,
Hoover Inst. Econ., Stanford Univ., Stanford, CA
Caldara D, Fernández-Villaverde J, Rubio-Ramírez JF, Yao W. 2012. Computing DSGE models with recursive
preferences and stochastic volatility. Rev. Econ. Dyn. 15:188–206
Calvo GA. 1983. Staggered prices in a utility-maximizing framework. J. Monet. Econ. 12:383–98
Campbell JY. 2018. Financial Decisions and Markets: A Course in Asset Pricing. Princeton, NJ: Princeton Univ.
Press
Colacito R, Croce M, Ho S, Howard P. 2018. BKK the EZ way: international long-run growth news and
capital flows. Am. Econ. Rev. 108:3416–49
Croce MM. 2014. Long-run productivity risk: a new hope for production-based asset pricing? J. Monet. Econ.
66:13–31
Cuñat A, Maffezzoli M. 2004. Neoclassical growth and commodity trade. Rev. Econ. Dyn. 7:707–36
Cuñat A, Maffezzoli M. 2007. Can comparative advantage explain the growth of US trade? Econ. J. 117:583–
602
de Boor C. 1978. A Practical Guide to Splines. New York: Springer Verlag
Den Haan WJ, Judd KL, Juillard M. 2010. Computational suite of models with heterogeneous agents: incom-
plete markets and aggregate uncertainty. J. Econ. Dyn. Control 34:1–3
Den Haan WJ, Judd KL, Juillard M. 2011. Computational suite of models with heterogeneous agents II:
multi-country real business cycle models. J. Econ. Dyn. Control 35:175–77
Den Haan WJ, Marcet A. 1990. Solving the stochastic growth model by parameterizing expectations. J. Bus.
Econ. Stat. 8:31–34
Den Haan WJ, Marcet A. 1994. Accuracy in simulations. Rev. Econ. Stud. 61:3–17
Den Haan WJ, Rendahl P. 2010. Solving the incomplete markets model with aggregate uncertainty using
explicit aggregation. J. Econ. Dyn. Control 34:69–78
Dergunov I, Meinerding C, Schlag C. 2019. Extreme inflation and time-varying consumption growth. Discuss.
Pap. 16/2019, Deutsche Bundesbank, Frankfurt
Devereux MB, Siu HE. 2007. State dependent pricing and business cycle asymmetries. Int. Econ. Rev. 48:281–
310
Diamond PA. 1982. Aggregate demand management in search equilibrium. J. Political Econ. 90:881–94
Doraszelski U. 2003. An R&D race with knowledge accumulation. RAND J. Econ. 34:20–42
Doraszelski U. 2004. Innovations, improvements, and the optimal adoption of new technologies. J. Econ. Dyn.
Control 28:1461–80
Dou W, Fang X, Lo AW, Uhlig H. 2020. Macro-finance models with nonlinear dynamics. Work. Pap., Becker
Friedman Inst. Econ., Univ. Chicago, Chicago
Drechsler I, Yaron A. 2010. What’s vol got to do with it. Rev. Financ. Stud. 24:1–45
Duffy J, McNelis PD. 2001. Approximating and simulating the stochastic growth model: parameterized ex-
pectations, neural networks, and the genetic algorithm. J. Econ. Dyn. Control 25:1273–303
Epstein LG, Zin SE. 1991. Substitution, risk aversion, and the temporal behavior of consumption and asset
returns: an empirical analysis. J. Political Econ. 99:263–86
Favilukis J, Lin X. 2013. Long run productivity risk and aggregate investment. J. Monet. Econ. 60:737–51
Favilukis J, Lin X. 2015. Wage rigidity: a quantitative solution to several asset pricing puzzles. Rev. Financ.
Stud. 29:148–92
Fernández-Villaverde J, Gordon G, Guerrón-Quintana P, Rubio-Ramirez JF. 2015. Nonlinear adventures at
the zero lower bound. J. Econ. Dyn. Control 57:182–204
Hagedorn M, Manovskii I. 2008. The cyclical behavior of equilibrium unemployment and vacancies revisited.
Am. Econ. Rev. 98:1692–706
Hall RE. 1971. The dynamic effects of fiscal policy in an economy with foresight. Rev. Econ. Stud. 38:229–44
Heer B, Maussner A. 2018. Projection methods and the curse of dimensionality. J. Math. Finance 8:317–34
Hubbard TP, Paarsch HJ. 2009. Investigating bid preferences at low-price, sealed-bid auctions with endoge-
nous participation. Int. J. Ind. Organ. 27:1–14
İmrohoroğlu A, Tüzel Ş. 2014. Firm-level productivity, risk, and return. Manag. Sci. 60:2073–90
Jermann UJ. 1998. Asset pricing in production economies. J. Monet. Econ. 41:257–75
Jermann UJ, Quadrini V. 2012. Macroeconomic effects of financial shocks. Am. Econ. Rev. 102:238–71
Jones LE, Manuelli RE, Siu HE. 2005. Fluctuations in convex models of endogenous growth, II: business cycle
properties. Rev. Econ. Dyn. 8:805–28
Judd KL. 1992. Projection methods for solving aggregate growth models. J. Econ. Theory 58:410–52
Judd KL. 1996. Approximation, perturbation, and projection methods in economic analysis. In Handbook of
Computational Economics, Vol. 1, ed. K Schmedders, KL Judd, pp. 509–85. Amsterdam: Elsevier
Judd KL. 1998. Numerical Methods in Economics. Cambridge, MA: MIT Press
Judd KL, Guu SM. 1993. Perturbation solution methods for economic growth models. In Economic and Finan-
cial Modeling with Mathematica® , ed. HR Varian, pp. 80–103. New York: Springer
Judd KL, Guu SM. 1997. Asymptotic methods for aggregate growth models. J. Econ. Dyn. Control 21:1025–42
Judd KL, Kubler F, Schmedders K. 2003. Computational methods for dynamic equilibria with heteroge-
neous agents. In Advances in Economics and Econometrics: Theory and Applications, Eight World Congress, ed.
M Dewatripont, LP Hansen, SJ Turnovsky, pp. 243–90. Cambridge, UK: Cambridge Univ. Press
Judd KL, Maliar L, Maliar S. 2010. A cluster-grid projection method: solving problems with high dimensionality.
NBER Work. Pap. 15965
Judd KL, Maliar L, Maliar S. 2011. Numerically stable and accurate stochastic simulation approaches for
solving dynamic economic models. Quant. Econ. 2:173–210
Kaltenbrunner G, Lochstoer LA. 2010. Long-run risk through consumption smoothing. Rev. Financ. Stud.
23:3190–224
Kollmann R, Maliar S, Malin BA, Pichler P. 2011. Comparison of solutions to the multi-country Real Business
Cycle model. J. Econ. Dyn. Control 35:186–202
Krueger D, Kubler F. 2004. Computing equilibrium in OLG models with stochastic production. J. Econ. Dyn.
Control 28:1411–36
Kung H, Schmid L. 2015. Innovation, growth, and asset prices. J. Finance 70:1001–37
Lan H. 2018. Comparing solution methods for DSGE models with labor market search. Comput. Econ. 51:1–34
Leith C, Liu D. 2016. The inflation bias under Calvo and Rotemberg pricing. J. Econ. Dyn. Control 73:283–97
Lemoine D, Traeger CP. 2016. Economics of tipping the climate dominoes. Nat. Climate Change 6:514–19
Levintal O. 2018. Taylor projection: a new solution method for dynamic general equilibrium models. Int. Econ.
Rev. 59:1345–73
Malin BA, Krueger D, Kubler F. 2011. Solving the multi-country real business cycle model using a Smolyak-
collocation method. J. Econ. Dyn. Control 35:229–39
Mortensen DT. 1982. Property rights and efficiency in mating, racing, and related games. Am. Econ. Rev.
72:968–79
Ngo PV. 2014. Optimal discretionary monetary policy in a micro-founded model with a zero lower bound on
nominal interest rate. J. Econ. Dyn. Control 45:44–65
Niemann S, Pichler P, Sorger G. 2013. Central bank independence and the monetary instrument problem.
Int. Econ. Rev. 54:1031–55
Papageorgiou C, Perez-Sebastian F. 2004. Can transition dynamics explain the international output data?
Macroecon. Dyn. 8:466–92
Papageorgiou C, Perez-Sebastian F. 2007. Is the asymptotic speed of convergence a good proxy for the tran-
sitional growth path? J. Money Credit Bank. 39:1–24
Petrosky-Nadeau N, Zhang L. 2017. Solving the Diamond–Mortensen–Pissarides model accurately. Quant.
Econ. 8:611–50
Petrosky-Nadeau N, Zhang L, Kuehn LA. 2013. Endogenous disasters and asset prices. Work. Pap., Carnegie
Mellon Univ., Pittsburgh, PA
Petrosky-Nadeau N, Zhang L, Kuehn LA. 2018. Endogenous disasters. Am. Econ. Rev. 108:2212–45
Pichler P. 2011. Solving the multi-country Real Business Cycle model using a monomial rule Galerkin method.
J. Econ. Dyn. Control 35:240–51
Pichler P, Sorger G. 2009. Wealth distribution and aggregate time-preference: Markov-perfect equilibria in
a Ramsey economy. J. Econ. Dyn. Control 33:1–14
Pissarides CA. 1985. Short-run equilibrium dynamics of unemployment vacancies, and real wages. Am. Econ.
Rev. 75:676–90
Pohl W, Schmedders K, Wilms O. 2018. Higher order effects in asset pricing models with long-run risks.
J. Finance 73:1061–111
Pohl W, Schmedders K, Wilms O. 2020. Asset pricing with heterogeneous agents and long-run risk. J. Financ.
Econ. In press
Reddien GW. 1980. Projection methods for two-point boundary value problems. SIAM Rev. 22:156–71
Reich G, Wilms O. 2015. Adaptive grids for the estimation of dynamic models. Work. Pap., Univ. Zurich, Zurich,
Switz.
Reiter M. 2009. Solving heterogeneous-agent models by projection and perturbation. J. Econ. Dyn. Control
33:649–65
Reiter M. 2010. Solving the incomplete markets model with aggregate uncertainty by backward induction.
J. Econ. Dyn. Control 34:28–35
Reiter M. 2015. Solving OLG models with many cohorts, asset choice and large shocks. Work. Pap., Inst. Adv. Stud.,
Vienna, Austria
Renner P, Scheidegger S. 2018. Machine learning for dynamic incentive problems. Work. Pap., Univ. Lancaster,
Lancaster, UK
Trefethen LN. 2013. Approximation Theory and Approximation Practice. Philadelphia: SIAM
Tuzel S. 2010. Corporate real estate holdings and the cross-section of stock returns. Rev. Financ. Stud. 23:2268–
302
van Zandweghe W, Wolman AL. 2019. Discretionary monetary policy in the Calvo model. Quant. Econ.
10:387–418
Weil P. 1989. The equity premium puzzle and the risk-free rate puzzle. J. Monet. Econ. 24:401–21
Werner M. 2016. Occasionally binding liquidity constraints and macroeconomic dynamics. Work. Pap., Univ. Zurich,
Zurich, Switz.
Young ER. 2010. Solving the incomplete markets model with aggregate uncertainty using the Krusell–Smith
algorithm and non-stochastic simulations. J. Econ. Dyn. Control 34:36–41
Annual Review of
Economics
v
EC12_FrontMatter ARI 12 June 2020 12:50
vi Contents