Chapter 1 nonlinear

Why Nonlinear/Non-Gaussian DSGE Models?

Jesús Fernández-Villaverde

University of Pennsylvania

July 10, 2011

Jesús Fernández-Villaverde (PENN) Nonlinear/Non-Gaussian DSGE July 10, 2011 1 / 38

Introduction

Motivation

These lectures review recent advances in nonlinear and non-gaussian
macro model-building.

First, we will justify why we are interested in this class of models.

Then, we will study both the solution and estimation of those models.

We will work with discrete time models.

We will focus on DSGE models.


Introduction

Nonlinearities

Most DSGE models are nonlinear.

Common practice (you saw it yesterday): solve and estimate a
linearized version with Gaussian shocks.

Why? Stochastic neoclassical growth model is nearly linear for the
benchmark calibration (Aruoba, Fernández-Villaverde, Rubio-Ramírez,
2005).

However, we want to depart from this basic framework.

I will present three examples.

Three Examples

Example I: Epstein-Zin Preferences

Recursive preferences (Kreps-Porteus-Epstein-Zin-Weil) have become
a popular way to account for asset pricing observations.

Natural separation between IES and risk aversion.

Example of a more general set of preferences in macroeconomics.

Consequences for business cycles, welfare, and optimal policy design.
Link with robust control.

I study a version of the RBC with in‡ation and adjustment costs in
The Term Structure of Interest Rates in a DSGE Model with
Recursive Preferences.

Three Examples

Household

Preferences:
2 31 θγ
6 1 γ
1 γ
1
7
Ut = 6 ctυ (1 lt )1 Et Ut +1 7
υ θ θ
4 +β 5
| {z }
Risk-adjustment operator

where:
1 γ
θ= 1
.
1 ψ

Budget constraint:
bt +1 1 bt
ct + it + = rt kt + wt lt +
pt Rt pt
Asset markets.

Three Examples

Technology

Production function:
yt = kt (zt lt )1
ζ ζ

Law of motion:

log zt +1 = λ log zt + χσε εzt +1 where εzt N (0, 1)

Aggregate constraints:

ct + it = kt (zt lt )1
ζ ζ

kt +1 = (1 δ) kt + it


Three Examples

Approximating the Solution of the Model
b b
De…ne st = kt , log zt ; 1 where kt = kt kss .
Under di¤erentiability conditions, third-order Taylor approximation of
the value function around the steady state:
1 1
V kt , log zt ; 1 ' Vss + Vi ,ss sti + Vij ,ss sti stj + Vijl ,ss sti stj stl ,
b
2 6
Approximations to the policy functions:
1 1
var kt , log zt ; 1 ' varss + vari ,ss sti + varij ,ss sti stj + varijl ,ss sti stj stl
b
2 6
and yields:

b
Rm kt , log zt , log π t , ω t ; 1 ' Rm,ss + Rm,i ,ss sat
1 i j 1 i j l
+ Rm,ij ,ss sat sat + Rm,ijl ,ss sat sat sat
2 6

Three Examples

Structure of Approximation

1 The constant terms Vss , varss , or Rm,ss do not depend on γ, the
parameter that controls risk aversion.

2 None of the terms in the …rst-order approximation, V.,ss , var.,ss , or
Rm,.,ss (for all m) depend on γ.

3 None of the terms in the second-order approximation, V..,ss , var..,ss ,
or Rm,..,ss depend on γ, except V33,ss , var33,ss , and Rm,33,ss (for all
m). This last term is a constant that captures precautionary behavior.

4 In the third-order approximation only the terms of the form V33.,ss ,
V3.3,ss , V.33,ss and var33.,ss , var3.3,ss , var.33,ss and Rm,33.,ss , Rm,3.3,ss ,
Rm,.33,ss (for all m) that is, terms on functions of χ2 , depend on γ.


Three Examples

Example II: Volatility Shocks

Data from four emerging economies: Argentina, Brazil, Ecuador, and
Venezuela. Why?

Monthly data. Why?

Interest rate rt : international risk free real rate+country spread.

International risk free real rate: Monthly T-Bill rate. Transformed
into real rate using past year U.S. CPI in‡ation.

Country spreads: Emerging Markets Bond Index+ (EMBI+) reported
by J.P. Morgan.
EMBI data coverage: Argentina 1997.12 - 2008.02; Ecuador 1997.12 -
2008.02; Brazil 1994.04 - 2008.02; and Venezuela 1997.12 - 2008.02.

Three Examples

Data

Figure: Country Spreads and DSGE Real Rate
Jesús Fernández-Villaverde (PENN) Nonlinear/Non-Gaussian T-Bill July 10, 2011 10 / 38

Three Examples

The Law of Motion for Interest Rates
Decomposition of interest rates:

rt = |{z} +
r εtb,t + εr ,t
|{z} |{z}
mean T-Bill shocks Spread shocks

εtb,t and εr ,t follow:

εtb,t = ρtb εtb,t 1 + e σtb,t utb,t , utb,t N (0, 1)
εr ,t = ρr εr ,t 1 + e σr ,t ur ,t , ur ,t N (0, 1)
σtb,t and σr ,t follow:

σtb,t = 1 ρσtb σtb + ρσtb σtb,t 1 + η tb uσtb ,t , uσtb ,t N (0, 1)

σr ,t = 1 ρσr σr + ρσr σr ,t 1 + η r uσr ,t , uσr ,t N (0, 1)
I could also allow for correlations of shocks.

Three Examples

A Small Open Economy Model I
Risk Matters: The Real E¤ects of Volatility Shocks.

Prototypical small open economy model: Mendoza (1991), Correia et
al. (1995), Neumeyer and Perri (2005), Uribe and Yue (2006).

Representative household with preferences:

∞ Ct ω 1H ω 1 v 1
∑β t t
E0 .
t =0 1 v

Why Greenwood-Hercowitz-Hu¤man (GHH) preferences? Absence of
wealth e¤ects.

Three Examples

A Small Open Economy Model II

Interest rates:

rt = r + εtb,t + εr ,t
εtb,t = ρtb εtb,t 1 + e σtb,t utb,t , utb,t N (0, 1)
σr ,t
εr ,t = ρr εr ,t 1 +e ur ,t , ur ,t N (0, 1)
σtb,t = 1 ρσtb σtb + ρσtb σtb,t 1 + η tb uσtb ,t , uσtb ,t N (0, 1)

σr ,t = 1 ρσr σr + ρσr σr ,t 1 + η r uσr ,t , uσr ,t N (0, 1)

Household’ budget constraint:
s
Dt + 1 Φd
= Dt Wt Ht Rt Kt + Ct + It + ( Dt + 1 D )2
1 + rt 2
Role of Φd > 0 (Schmitt-Grohé and Uribe, 2003).


Three Examples

A Small Open Economy Model III

The stock of capital evolves according to the following law of motion:
!
2
φ It
Kt + 1 = ( 1 δ ) Kt + 1 1 It
2 It 1

Typical no-Ponzi condition.

Production function:
1 α
Yt = Ktα e X t Ht

where:
Xt = ρx Xt 1 + e σx ux ,t , ux ,t N (0, 1) .
Competitive equilibrium de…ned in a standard way.

Three Examples

Solving the Model
Perturbation methods.
We are interested on the e¤ects of a volatility increase, i.e., a positive
shock to either uσr ,t or uσtb ,t , while ur ,t = 0 and utb,t = 0.
We need to obtain a third approximation of the policy functions:
1 A …rst order approximation satis…es a certainty equivalence principle.
Only level shocks utb,t , ur ,t , and uX ,t appear.
2 A second order approximation only captures volatility indirectly via
cross products ur ,t uσr ,t and utb,t uσtb ,t ,. Thus, volatility only has an
e¤ect if the real interest rate changes.
3 In the third order, volatility shocks, uσ,t and uσtb ,t , enter as
independent arguments.
Moreover:
1 Cubic terms are quantitatively important.
2 The mean of the ergodic distributions of the endogenous variables and
the deterministic steady state values are quite di¤erent. Key for
calibration.

Three Examples

Example III: Fortune or Virtue
Strong evidence of time-varying volatility of U.S. aggregate variables.

Most famous example: the Great Moderation between 1984 and 2007.

Two explanations:

1 Stochastic volatility: fortune.

2 Parameter drifting: virtue.

How can we measure the impact of each of these two mechanisms?

We build and estimate a medium-scale DSGE model with:

1 Stochastic volatility in the shocks that drive the economy.

2 Parameter drifting in the monetary policy rule.

Three Examples

The Discussion

Starting point in empirical work by Kim and Nelson (1999) and
McConnell and Pérez-Quirós (2000).

Virtue: Clarida, Galí, and Gertler (2000) and Lubik and Schorfheide
(2004).

Sims and Zha (2006): once time-varying volatility is allowed in a
SVAR model, data prefer fortune.

Follow-up papers: Canova and Gambetti (2004), Cogley and Sargent
(2005), Primiceri (2005).

Fortune papers are SVARs models: Benati and Surico (2009).

A DSGE model with both features is a natural measurement tool.

Three Examples

The Goals

1 How do we write a medium-scale DSGE with stochastic volatility and
parameter drifting?

2 How do we evaluate the likelihood of the model and how to
characterize the decision rules of the equilibrium?

3 How do we estimate the model using U.S. data and assess model …t?

4 How do we build counterfactual histories?


Three Examples

Model I: Preferences
Household maximizes:
( )
∞
mjt ljt+ϑ
1
E0 ∑ β dt t
log (cjt hcjt 1 ) + υ log
pt
ϕt ψ
1+ϑ
t =0

Shocks:
log dt = ρd log dt 1 + σd ,t εd ,t

log ϕt = ρ ϕ log ϕt 1 + σ ϕ,t ε ϕ,t

Stochastic Volatility:
log σd ,t = 1 ρσd log σd + ρσd log σd ,t 1 + η d ud ,t

log σ ϕ,t = 1 ρσ ϕ log σ ϕ + ρσ ϕ log σ ϕ,t 1 + η ϕ u ϕ,t


Three Examples

Model II: Constraints
Budget constraint:
Z
mjt bjt +1
cjt + xjt + + + qjt +1,t ajt +1 d ω j ,t +1,t =
pt pt
mjt 1 bjt
wjt ljt + rt ujt µt 1 Φ [ujt ] kjt 1 + + Rt 1 + ajt + Tt + zt
pt pt
The capital evolves:
xjt
kjt = (1 δ) kjt 1 + µt 1 V xjt
xjt 1

Investment-speci…c productivity µt follows a random walk in logs:
log µt = Λµ + log µt 1 + σµ,t εµ,t

log σµ,t = 1 ρσµ log σ + ρ
µ σµ log σµ,t 1 + η µ uµ,t

Three Examples

Model III: Nominal Rigidities

Monopolistic competition on labor markets with sticky wages (Calvo
pricing with indexation).

Monopolistic intermediate good producer with sticky prices (Calvo
pricing with indexation):

1 α
α d
yit = At kit 1 lit φzt
log At = ΛA + log At 1 + σA,t εA,t


log σA,t = 1 ρσA log σA + ρσA log σA,t 1 + η A uA,t


Three Examples

Model IV: Monetary Authority
Modi…ed Taylor rule:
0 0 1γy ,t 11 γR
ytd
B Πt
γR γΠ,t
Rt Rt 1 @ ytd 1
A C
= @ A exp (σm,t εmt )
R R Π exp Λy d


log σm,t = 1 ρσm log σm + ρσm log σm,t 1 + η m um,t

Parameter drifting:

log γΠ,t = 1 ργ log γΠ + ργ log γΠ,t 1 + η Π επ,t
Π Π

log γy ,t = 1 ργ log γy + ργ log γy ,t 1 + η y εy ,t
y y


More About Nonlinearities

More About Nonlinearities I

The previous examples are not exhaustive.

Unfortunately, linearization eliminates phenomena of interest:

1 Asymmetries.

2 Threshold e¤ects.

3 Precautionary behavior.

4 Big shocks.

5 Convergence away from the steady state.

6 And many others....



More About Nonlinearities II

Linearization limits our study of dynamics:

1 Zero bound on the nominal interest rate.

2 Finite escape time.

3 Multiple steady states.

4 Limit cycles.

5 Subharmonic, harmonic, or almost-periodic oscillations.

6 Chaos.



More About Nonlinearities III

Moreover, linearization induces an approximation error.

This is worse than you may think.

1 Theoretical arguments:

1 Second-order errors in the approximated policy function imply
…rst-order errors in the loglikelihood function.

2 As the sample size grows, the error in the likelihood function also grows
and we may have inconsistent point estimates.

3 Linearization complicates the identi…cation of parameters.

2 Computational evidence.



Arguments Against Nonlinearities
1 Theoretical reasons: we know way less about nonlinear and
non-gaussian systems.

2 Computational limitations.

3 Bias.

Mark Twain
To a man with a hammer, everything looks like a nail.

Teller’ Law
s
A state-of-the-art computation requires 100 hours of CPU time on the
state-of-the art computer, independent of the decade.

Solution

Solving DSGE Models

We want to have a general formalism to think about solving DSGE
models.

We need to move beyond value function iteration.

Theory of functional equations.

We can cast numerous problems in macroeconomics involve
functional equations.

Examples: Value Function, Euler Equations.


Solution

Functional Equation

Let J 1 and J 2 be two functional spaces, Ω <l and let H : J 1 ! J 2
be an operator between these two spaces.

A functional equation problem is to …nd a function d : Ω ! <m such
that

H (d ) = 0

Regular equations are particular examples of functional equations.

Note that 0 is the space zero, di¤erent in general that the zero in the
reals.


Solution

Example: Euler Equation I
Take the basic RBC:

∞
max E0 ∑ β t u ( ct )
t =0
ct + kt +1 = e zt ktα
+ (1 δ) kt , 8 t > 0
zt = ρzt 1 + σεt , εt N (0, 1)
The …rst order condition:

u 0 (ct ) = βEt u 0 (ct +1 ) 1 + αe zt +1 ktα+1
1
δ

There is a policy function g : <+ < ! <2 that gives the optimal
+
choice of consumption and capital tomorrow given capital and
productivity today.

Solution

Example: Euler Equation II

Then:

u 0 g 1 (kt , zt ) = βEt u 0 g 1 (kt +1 , zt +1 ) 1 + f g 2 (kt , zt ) , zt +1 δ

or, alternatively:

u 0 g 1 (kt , zt )
βEt u 0 g 1 g 2 (kt , zt ) , zt +1 1 + f g 2 (kt , zt ) , zt +1 δ =0

We have functional equation where the unknown object is the policy
function g ( ).

More precisely, an integral equation (expectation operator). This can
lead to some measure theoretic issues that we will ignore.

Solution

Example: Euler Equation III

Mapping into an operator is straightforward:

H = u0 ( ) βEt u 0 ( ) (1 + f ( , zt +1 ) δ)
d =g

If we …nd g , and a transversality condition is satis…ed, we are done!


Solution

Example: Euler Equation IV
Slightly di¤erent de…nitions of H and d can be used.

For instance if we take again the Euler equation:

u 0 ( ct ) βEt u 0 (ct +1 ) 1 + αe zt +1 ktα+1
1
δ =0

we may be interested in …nding the unknown conditional expectation
Et u 0 (ct +1 ) 1 + αe zt +1 ktα+1 δ .
1

Since Et is itself another function, we write

H (d ) = u 0 ( ) βd = 0

where d = Et fϑ g and ϑ = u 0 ( ) (1 + f ( , zt +1 ) δ) .

Solution

How Do We Solve Functional Equations?
Two Main Approaches

1 Perturbation Methods:

n
d n (x, θ ) = ∑ θ i (x x0 )i
i =0

We use implicit-function theorems to …nd coe¢ cients θ i .

2 Projection Methods:

n
d n (x, θ ) = ∑ θ i Ψi (x )
i =0

We pick a basis fΨi (x )gi∞ 0 and “project” H ( ) against that basis.
=

Solution

Relation with Value Function Iteration
There is a third main approach: the dynamic programing algorithm.

Advantages:

1 Strong theoretical properties.

2 Intuitive interpretation.

Problems:

1 Di¢ cult to use with non-pareto e¢ cient economies.

2 Curse of dimensionality.

Estimation

Evaluating the Likelihood Function

How do we take the model to the data?

Usually we cannot write the likelihood of a DSGE model.

Once the model is nonlinear and/or non-gaussian we cannot use the
Kalman …lter to evaluate the likelihood function of the model.

How do we evaluate then such likelihood? Using Sequential Monte
Carlo.


Estimation

Basic Estimation Algorithm 1: Evaluating Likelihood
Input: observables Y T , DSGE model M with parameters
γ 2 Υ.

Output: likelihood p y T ; γ .

1 Given γ, solve for policy functions of M.

2 With the policy functions, write the state-space form:

St = f ( St 1 , Wt ; γ i )
Yt = g (St , Vt ; γi )
3 With state space form, evaluate likelihood:
T
p y T ; γi = ∏p yt jy t 1
; γi
t =1

Estimation

Basic Estimation Algorithm 2: MLE
Input: observables Y T , DSGE model M parameterized by
γ 2 Υ.

Estimates: b
γ

1 Set i = 0. Fix initial parameter values γi .

2 Compute p y T ; γi using algorithm 1.

3 Is γi = arg max p y T ; γ ?

1 Yes: b
Make γ = γi . Stop.

2 No: Make γi γ i +1 . Go to step 2.


Estimation

Basic Estimation Algorithm 3: Bayesian
Input: observables Y T , DSGE model M parameterized by
γ 2 Υ with priors π (γ).
Posterior distribution: π γj Y T

1 Fix I . Set i = 0 and chose initial parameter values γi .

2 Compute p y T ; γi using algorithm 1.

3 Propose γ = γi + ε where ε N (0, Σ) .

p (y T ;γ )π (γ )
4 Compute α = min p (y T ;γi )π (γi )
,1 .

5 With probability α, make γi +1 = γ . Otherwise γi +1 = γi .

6 If i < M, i ı + 1.
¯ Go to step 3. Otherwise Stop.


Chapter 1 nonlinear

More Related Content

What's hot (18)

Viewers also liked (7)

Similar to Chapter 1 nonlinear (20)

More from NBER (20)

Recently uploaded (20)

Chapter 1 nonlinear