Bayesian Analysis of Dynamic Stochastic General

Equilibrium Models

Frank Schorfheide
University of Pennsylvania, CEPR, NBER

June 19, 2013

Bayesian Inference

• Goals: conditional on prior information and sample information Y

we want to
• characterize uncertainty about θ, f (θ), or future y ’s;
• make optimal decisions that reflect our uncertainty about model
parameters
• Our models (VAR, DSGE, etc.) deliver a likelihood function p(Y |θ).
• We represent our ignorance/knowledge about θ using a prior
probability distribution with density p(θ).
• In view of the data Y , we update the prior

p(Y |θ)p(θ)
p(θ|Y ) = R (Bayes Theorem)
p(Y |θ)p(θ)dθ
• In practical work, we use algorithms to generate draws from the
posterior distribution p(θ|Y ), say, {θ(s) }ns=1
sim
.
• Once we have these draws we can
• compute point estimates, credible intervals, or plot
histograms/densities of posterior distributions;
• compute decisions that minimize posterior expected loss.
Frank Schorfheide Bayesian Analysis of DSGE Models
VAR vs. DSGE
VAR DSGE Model
yt = Ψ(θ)st
Representation yt = Byt−1 + ut
st = Φ1 (θ)st−1 + Φ (θ)
ut ∼ N(0, Σ) t ∼ N(0, I )
Need to solve the DSGE
numerically
to obtain Ψ(θ), Φ1 (θ), Φ (θ)
Likelihood Fcn Available in analytical form Has to be evaluated numerically:
p(Y |θ) - Kalman filter
- particle filter
Prior p(θ) e.g. Minnesota prior see empirical applications
Posterior p(θ|Y ) Analytical characterization No analytical characterization
Sampling from - Direct Sampling - Metropolis-Hastings Algo.
Posterior: {θ(s) }ns=1
sim
- Gibbs Sampling - (other methods)
Reporting Use {θ(s) }ns=1
sim
draws to compute mean, std, percentiles,
histograms, density estimates, etc.
Impulse Responses, Convert {θ(s) }ns=1
sim
draws into IRFs {f (θ(s) )}ns=1
sim
,
etc. compute mean, percentiles, etc. from {f (θ(s) )}ns=1 sim
,
Frank Schorfheide Bayesian Analysis of DSGE Models
Household
• Preferences:
∞
" #
(Ht+s /Bt+s )1+1/ν
X 
Et β t+s ln Ct+s − (1)
s=0
1 + 1/ν

• Budget constraint:

Ct + It ≤ Wt Ht + Rt Kt .

• Capital accumulation:

Kt+1 = (1 − δ)Kt + It , (2)

• First-order conditions:
   1/ν
1 1 1 1 Ht
= βE (Rt+1 + (1 − δ)) and Wt = .
Ct Ct+1 Ct Bt Bt
(3)

Frank Schorfheide Bayesian Analysis of DSGE Models

Firms

• Technology:

Yt = (At Ht )α Kt1−α . (4)

• First-order conditions from profit maximization:

Yt Yt
Wt = α , Rt = (1 − α) . (5)
Ht Kt

• Market clearing:

Yt = Ct + It . (6)

Frank Schorfheide Bayesian Analysis of DSGE Models

Exogenous Processes
• Log Technology:

ln At = ln A0 + (ln γ)t + ln A
et , ln A
e t = ρa ln A
e t−1 + σa a,t (7)

where a,t ∼ iidN(0, 1).

• Preference shifter:

ln Bt = (1 − ρb ) ln B∗ + ρb ln Bt−1 + σb b,t (8)

where b,t ∼ iidN(0, 1).

• Initialzation:

ln A
e −τ = 0 and ln B−τ = 0.

Frank Schorfheide Bayesian Analysis of DSGE Models

Model Solution
• The solution to the rational expectations difference equations
determines law of motion for Yt , Ct , It , Kt , Ht , Wt , and Rt .
• The technology process ln At induces a common trend in output,
consumption, investment, capital, and wages.
• It is useful to detrended the model variables as follows:

et = Yt , C
Y et = Ct , eIt = It , K
et+1 = Kt+1 , W
ft = Wt . (9)
At At At At At

Frank Schorfheide Bayesian Analysis of DSGE Models

Equilibrium Conditions, Rewritten

" #  1/ν
1 1 −at+1 1 f 1 Ht
= βE e (Rt+1 + (1 − δ)) , Wt = (10)
Ct
e C
et+1 Ct
e Bt Bt
Y
et Yet
W
ft = α , Rt = (1 − α) e at
Ht Kt
e
 1−α
Y
et = Htα K et e −at , Y
et = C
et + eIt , K et e −at + eIt .
et+1 = (1 − δ)K

• The process at is defined as

At
at = ln = ln γ + (ρa − 1) ln A
e t−1 + σa a,t . (11)
At−1

• This log ratio is always stationary, because if ρa = 1 the ln A

e t−1
term drops out.

Frank Schorfheide Bayesian Analysis of DSGE Models

Steady State, Etc.
• Steady state:

  e
γ K
e∗ (1 − α)γ eI∗ 1−δ K ∗
R∗ = − (1 − δ), = , = 1− .
β Y∗
e R∗ Y∗
e γ Y∗
e
(12)

• If ρa = 1, the model generates cointegration relationships which are

obtained by taking pair-wise differences of ln Yt , ln Ct , ln It , ln Kt+1 ,
and ln Wt
• Parameters are stacked in vector θ:

θ = [α, β, γ, δ, ν, ρa , σa , ρb , σb ]0 . (13)

Frank Schorfheide Bayesian Analysis of DSGE Models

Loglinearization

 
R∗
C
bt = Et C at+1 −
bt+1 + b Rt+1
b (14)
R∗ + (1 − δ)
H
bt = ct − ν C
νW bt + (1 + ν)Bbt , W ct = Y bt − Hbt ,

bt+1 = 1 − δ K bt + I∗ bIt − 1 − δ b
e
R
bt = bt − K
Y bt + b
at , K at ,
γ Ke∗ γ

bt = C∗ C bt + I∗ bIt ,
e e
Y
bt = αHb t + (1 − α)Kbt − (1 − α)b at , Y
Y
e∗ Y
e∗
A
bt = ρa A
b t−1 + σa a,t , b bt − A
at = A b t−1 , B
bt = ρb B
bt−1 + σb b,t .

Log-linearization of f (x):
z
1 write f (x) = f (e );

2 conduct a first-order Taylor approximation around x0 in terms of z:

f (e ln x ) ≈ f (x0 ) + x0 f (1) (x0 )(ln x − ln x0 ).

Frank Schorfheide Bayesian Analysis of DSGE Models

State-Space Representation
• State-Transition:

st = Φ1 (θ)st−1 + Φ (θ)t . (15)

• Measurement:

yt = Ψ0 (θ) + Ψ1 (θ)t + Ψ2 (θ)st (16)

Frank Schorfheide Bayesian Analysis of DSGE Models

Examples
• Observations on GDP and Hours:
      " #
ln GDPt ln Y0 ln γ Y
bt + Abt
= + t+ ,
ln Ht ln H∗ 0 Ht
b

• Observations on GDP and Investment:

      " #
ln GDPt ln Y0 ln γ A
bt + Y
bt
= + t+ b .
ln It ln Y0 + (ln eI∗ − ln Y
e∗ ) ln γ At + bIt

• Cointegration relationship for ρa = 1:

   
  ln GDPt (1 − α)(γ − 1 + δ)
−1 1 = ln + bIt − Y
bt .
ln It γ/β − 1 + δ

Frank Schorfheide Bayesian Analysis of DSGE Models

Bayesian Estimation – Prior

Name Domain Density Para (1) Para (2)

α [0, 1) Beta 0.66 0.02
ν R+ Gamma 2.00 1.00
4 ln γ R Normal 0.00 0.10
ρa R+ Beta 0.95 0.02
σa R+ InvGamma 0.01 4.00
ρb R+ Beta 0.80 0.10
σb R+ InvGamma 0.01 4.00
ln H∗ R Normal 0.00 10.0
ln Y0 R Normal 0.00 100

Frank Schorfheide Bayesian Analysis of DSGE Models

Random-Walk Metropolis (RWM) Algorithm for DSGE
Model
1 Use a numerical optimization routine to maximize the log posterior,
which up to a constant is given by ln p(Y |θ) + ln p(θ). Denote the
posterior mode by θ̃.
2 Let Σ̃ be the inverse of the (negative) Hessian computed at the
posterior mode θ̃, which can be computed numerically.
3 Draw θ(0) from N(θ̃, c02 Σ̃) or directly specify a starting value.
4 For s = 1, . . . , nsim : draw ϑ from the proposal distribution
N(θ(s−1) , c 2 Σ̃). The jump from θ(s−1) is accepted (θ(s) = ϑ) with
probability min {1, r (θ(s−1) , ϑ|Y )} and rejected (θ(s) = θ(s−1) )
otherwise. Here
p(Y |ϑ)p(ϑ)
r (θ(s−1) , ϑ|Y ) = . 2
p(Y |θ(s−1) )p(θ(s−1) )

Frank Schorfheide Bayesian Analysis of DSGE Models

Posterior

Det. Trend Stoch. Trend

Name Mean 90% Intv. Mean 90% Intv.
α 0.65 [0.62, 0.68] 0.65 [0.63, 0.69]
ν 0.42 [0.16, 0.67] 0.70 [0.22, 1.23]
4 ln γ .003 [.002, .004] .004 [.002, .005]
ρa 0.97 [0.95, 0.98] 1.00
σa .011 [.010, .012] .011 [.010, .012]
ρb 0.98 [0.96, 0.99] 0.98 [0.96, 0.99]
σb .008 [.007, .008] .007 [.006, .008]
ln H∗ -0.04 [-0.08, 0.01] -0.03 [-0.07, 0.02]
ln Y0 8.77 [8.61, 8.93] 8.39 [7.93, 8.86]

Frank Schorfheide Bayesian Analysis of DSGE Models

The Next Steps
1 Solving DSGE models
2 Computation of likelihood function
3 Bayesian estimation of DSGE models
4 A detailed application illustrating the methods and highlighting
challenges

Some readings (see my website):

• An, S. and F. Schorfheide (2007): “Bayesian Analysis of DSGE Models,”
Econometric Reviews, 26(2-4), 113-172.
• Del Negro, M. and F. Schorfheide (2011): “Bayesian Macroeconometrics,” in J.
Geweke, G. Koop, and H. van Dijk (eds.) The Oxford Handbook of Bayesian
Econometrics, Oxford University Press, 293-389.
• Del Negro, M. and F. Schorfheide (2012): “DSGE Model-Based Forecasting,”
prepared for G. Elliot and A. Timmermann (eds.) Handbook of Economic
Forecasting, vol 2, Elsevier North Holland, forthcoming.
• Schorfheide, F. (2010): “Estimation and Evaluation of DSGE Models: Progress
and Challenges” Invited Lecture at 2010 Econometric Society World Congress in
Shanghai, available as NBER Working Paper, 16781.

Frank Schorfheide Bayesian Analysis of DSGE Models

Solving DSGE Models

A Primer on DSGE Model Solutions

Frank Schorfheide Bayesian Analysis of DSGE Models

Recall the Equilibrium Conditions

" #  1/ν
1 1 −at+1 1 f 1 Ht
= βE e (Rt+1 + (1 − δ)) , Wt =
Ct
e C
et+1 Ct
e Bt Bt
Y
et Yet
W
ft = α , Rt = (1 − α) e at
Ht Kt
e
 1−α
Y
et = Htα K et e −at , Y
et = C
et + eIt , K et e −at + eIt
et+1 = (1 − δ)K

At
at = ln = ln γ + (ρa − 1) ln A
e t−1 + σa a,t .
At−1

Frank Schorfheide Bayesian Analysis of DSGE Models

DSGE Model Solutions

• Find solution to nonlinear rational expectations difference equations

that satisfy additional optimality conditions for underlying
optimization problems (e.g. transversality)

• Useful to distinguish between state and control variables based on

the underlying dynamic programming problem. Solution can be
represented as a law of motion for the state variables and a mapping
from the states into controls:
• Exogenous states: at , ln A e t , Bt
• Endogenous states: K et
• Controls: C
et , H
e t , eIt

• In general, state transition as well as mapping from states to

controls is nonlinear.

• We can either use a local Taylor-type approximation near a steady

state; orPa global approximation, say, via Chebyshev polynomials:
N
f (x) ≈ n=0 αn Tn (x).

Frank Schorfheide Bayesian Analysis of DSGE Models

DSGE Model Solutions
• We will focus mostly on log-linear (or first-order perturbation)
approximations.
• In a nutshell... consider the backward-looking model

yt = f (yt−1 , σut ).

• Guess that the solution is of the form

(0) (1)
yt = yt + σyt + o(σ).
(0)
• Expand f (·) around σ = 0 assuming that steady state yt = 0:

f (yt−1 , σut ) = fy yt−1 + fu σut + o(|yt−1 |) + o(σ)

• Now plug-in conjectured solution:

(1) (1)
σyt = fy σyt−1 + fu σut + o(σ)
(1) (1)
• Deduce that yt = fy yt−1 + fu ut

Frank Schorfheide Bayesian Analysis of DSGE Models

DSGE Model Solutions
• A second-order perturbation essentially amounts to:
(0) (1) (2)
yt = yt + σyt + σ 2 yt + o(σ 2 )
(0)
• Expand f (·) around σ = 0 assuming that yt = 0:
f (yt−1 , σut ) ≈ fy yt−1 + fu σut
2
+fyu yt−1 σut + 0.5fyy yt−1 + 0.5fuu σ 2 ut2 + small
• Plug-in conjectured solution:
(1) (2) (1) (2)
σyt + σ 2 yt ≈ fy (σyt−1 + σ 2 yt−1 ) + fu σut
(1) (2)
+fyu (σyt−1 + σ 2 yt−1 )σut
(1) (2)
+0.5fyy (σyt−1 + σ 2 yt−1 )2
+fuu σ 2 ut + o(σ 2 )
• Equate similar terms (similar with respect to σ) on the left and the
right to obtain,
(1) (1)
yt = fy yt−1 + fu ut
(2) (2) (1) (1)
yt = fy yt−2 + fyu yt−1 ut + 0.5fyy [yt−1 ]2 + 0.5fuu ut2

Frank Schorfheide Bayesian Analysis of DSGE Models

Next Steps

• Notice that:
• (log)linear solutions lead to linear state-space models;
• nonlinear solutions lead to nonlinear state-space models

• We will now review some of the details for linear solution techniques
and discuss multiplicity issues

• Later on, we also discuss the estimation of nonlinearly solved DSGE

models (but without reviewing solution techniques in detail)

Frank Schorfheide Bayesian Analysis of DSGE Models

Solving Linearized DSGE Models
• Linearized DSGE leads to linear rational expectations (LRE) system:

Γ0 (θ)st = Γ1 (θ)st−1 + Ψt + Πηt

where
• st is a vector of model variables, t is a vector of exogenous shocks,
• ηt is a vector of RE errors with elements ηtx = x̂t − Et−1 [x̂t ], and
• st contains (among others) the conditional expectation terms
Et [e
xt+1 ].
• Solution methods for LREs: Blanchard and Kahn (1980), King and
Watson (1998), Uhlig (1999), Anderson (2000), Klein (2000),
Christiano (2002), Sims (2002).
• Typically, the solution in terms of st is of the form

st = Φ1 (θ)st−1 + Φ (θ)t .

Frank Schorfheide Bayesian Analysis of DSGE Models

Simple Example
• Simple model:

1
yt = Et [yt+1 ] + t ,
θ
where t ∼ iid(0, 1) and θ ∈ Θ = [0, 2].
• Introduce conditional expectation ξt = Et [yt+1 ] and forecast error
ηt = yt − ξt−1 .
• Thus,

ξt = θξt−1 − θt + θηt .

• Canonical LRE model (Sims, 2002)

Γ0 (θ)st = Γ1 (θ)st−1 + Ψ(θ)t + Π(θ)ηt .

Frank Schorfheide Bayesian Analysis of DSGE Models

Simple Example
• Determinacy: θ > 1. Then only stable solution:

ξt = 0, ηt = t , yt = t

• Indeterminacy: θ ≤ 1 the stability requirement imposes no

restrictions on forecast error:

ηt = M
e t + ζt .

• For simplicity assume now ζt = 0. Then

yt − θyt−1 = M
e t − θt−1 .

Frank Schorfheide Bayesian Analysis of DSGE Models

Simple Example
• Re-parameterization M
e = 1 + M. The indeterminacy region of the
parameter space is labelled ΘI = [0, 1].
• Then

(1 − θL)yt = (1 − θL)t + Mt

• Or

yt = t + M(1 − θL)−1 t

Frank Schorfheide Bayesian Analysis of DSGE Models

Generalization
• Canonical form:

Γ0 (θ)st = Γ1 (θ)st−1 + Ψ(θ)t + Π(θ)ηt ,

• The system can be rewritten as

st = Γ∗1 (θ)st−1 + Ψ∗ (θ)t + Π∗ (θ)ηt .

• Replace Γ∗1 by JΛJ −1 and define wt = J −1 st .

• To deal with repeated eigenvalues and non-singular Γ0 we use

Generalized Complex Schur Decomposition (QZ) in practice.

• Let the i’th element of wt be wi,t and denote the i’th row of J −1 Π∗
and J −1 Ψ∗ by [J −1 Π∗ ]i. and [J −1 Ψ∗ ]i. , respectively.

Frank Schorfheide Bayesian Analysis of DSGE Models

Generalization
• Rewrite model:

wi,t = λi wi,t−1 + [J −1 Π∗ ]i. t + [J −1 Ψ∗ ]i. ηt .

• Define the set of stable AR(1) processes as

 
Is (θ) = i ∈ {1, . . . n} |λi (θ)| ≤ 1

• Let Ix (θ) be its complement. Let ΨJx and ΠJx be the matrices
composed of the row vectors [J −1 Ψ∗ ]i. and [J −1 Π∗ ]i. that
correspond to unstable eigenvalues, i.e., i ∈ Ix (θ).
• Stability condition:

ΨJx t + ΠJx ηt = 0

for all t.

Frank Schorfheide Bayesian Analysis of DSGE Models

Generalization
• Solve system of equations ΨJx t + ΠJx ηt = 0 for ηt , i.e.

η t = η 1 t + η 2 ζ t .

• Then plug solution for ηt into

Γ0 (θ)st = Γ1 (θ)st−1 + Ψ(θ)t + Π(θ)ηt .

Frank Schorfheide Bayesian Analysis of DSGE Models