Frank Schorfheide, University of Pennsylvania: Bayesian Methods 1

Bayesian Analysis of DSGE Models

Frank Schorfheide
Department of Economics, University of Pennsylvania
Frank Schorfheide, University of Pennsylvania: Bayesian Methods 2

Comparison to VARs: DSGE-VARs

• Compare fit of DSGE model to that of a VAR based on marginal data densities.

Mechanics are non-trivial. Under a very diffuse prior for the VAR coefficients, the

DSGE model is likely to win the comparison.

• Careful construction of VAR prior is crucial, for instance:

– Minnesota-style prior, Sims-Zha priors for identified VARs.

– DSGE-VARs: Del Negro and Schorfheide (2004, 2005), Del Negro, Schorfheide,

Smets, and Wouters (2004).

• Compare DSGE model dynamics to (identified) VAR dynamics

Frank Schorfheide, University of Pennsylvania: Bayesian Methods 3

Two Views of DSGE-VARs

• Improve VAR estimates by “restricting” its parameter estimates.

• Improve DSGE model by relaxing its restrictions.

Frank Schorfheide, University of Pennsylvania: Bayesian Methods 4

DSGE-VARs: Improving VARs

• Consider a vector autoregressive specification of the form

yt = Φ0 + Φ1yt−1 + . . . + Φpyt−p + ut, IE[utu0t] = Σ. (1)

• Write VAR as Y = XΦ + U , Y is T × n, X is T × k.

• Difficulty: too many parameters which leads to imprecise estimates.

• Solution: tilt estimates toward a point in the parameter space. Example: Minnesota

prior tilts toward random walks.

• Here: tilt toward DSGE model restrictions.

Frank Schorfheide, University of Pennsylvania: Bayesian Methods 5

Example

• n independent draws of yt from N (µ, 1).

• MLE of µ:
n
1X
µ̂M L = yt .
n t=1

• Bayes estimator based on prior µ ∼ N (0, τ 2)

X n
1 n 1/τ 2
µ̂B = yt = µ̂M L + 0
n + 1/τ 2 t=1 n + 1/τ 2 n + 1/τ 2
Frank Schorfheide, University of Pennsylvania: Bayesian Methods 6

Example

• MSE of MLE:
· ¸
1
IE µ (µ − µ̂M L)2 = |{z}
02 + .
n
|{z}
Bias2
Variance
• MSE of Bayes Estimator:
· ¸ µ 2
¶2
1/τ n
IE µ (µ − µ̂B )2 = µ2 + .
n + 1/τ 2 (n + 1/τ 2)2
| {z } | {z }
Bias2 variance

• If µ2 is small, i.e. the discrepancy between the a priori expected value and the “true”

value is small, then the Bayes estimator clearly dominates.

Frank Schorfheide, University of Pennsylvania: Bayesian Methods 7

DSGE-VARs: Improving VARs

• Complication: DSGE model depends on parameters θ.

• Solution: place prior on θ. Use notion of dummy observations to construct priors

conditional on θ. Overall:

p(Φ, Σ, θ) = p(θ)p(Φ, Σ|θ). (2)

• Let’s look at: p(Φ, Σ|θ).

Frank Schorfheide, University of Pennsylvania: Bayesian Methods 8

DSGE-VARs: Improving VARs

• Quasi-likelihood function for artificial observations (sample size T ∗ = λT ) generated

from DSGE model:

p(Y ∗(θ)|Φ, Σu) ∝ (3)

½ ¾
1 0 0 0 0
|Σu|−λT /2 exp − tr[Σu−1(Y ∗ Y ∗ − Φ0X ∗ Y ∗ − Y ∗ X ∗Φ + Φ0X ∗ X ∗Φ)] .
2

• Let IE D
θ [·] be the expectation under DSGE model and define the autocovariance ma-

trices

ΓXX (θ) = IE D 0
θ [xt xt ], ΓXY (θ) = IE D 0
θ [xt yt ].

0 0
• Replace sample moments Y ∗ Y ∗ by IE D ∗ ∗
θ [Y Y ] = λT ΓY Y (θ), etc.
Frank Schorfheide, University of Pennsylvania: Bayesian Methods 9

DSGE-VARs: Improving VARs

• Define

Φ∗(θ) = Γ−1
XX (θ)ΓXY (θ), Σ∗(θ) = ΓY Y (θ) − ΓY X (θ)Γ−1
XX (θ)ΓXY (θ). (4)

• Prior distribution:
µ ¶
Σ|θ ∼ IW λT Σ∗(θ), λT − k, n (5)
à · ¸−1!
1
Φ|Σ, θ ∼ N Φ∗(θ), Σ−1 ⊗ ΓXX (θ) ,
λT
Frank Schorfheide, University of Pennsylvania: Bayesian Methods 10

DSGE-VARs: Relaxing DSGE Restrictions

• Alternative motivation...
Frank Schorfheide, University of Pennsylvania: Bayesian Methods 11

) (T ): Cross-equation
I2 restriction for given value Prior contours for misspecification
of T parameters )'

) (T )+)'

)'

subspace generated by the

DSGE model restrictions

I1
Frank Schorfheide, University of Pennsylvania: Bayesian Methods 12

DSGE-VARs: Relaxing DSGE Restrictions

• There is a vector θ and matrices Φ∆ and Σ∆ such that the data are generated from

the VAR in Eq. (1)

Φ = Φ∗(θ) + Φ∆, Σ = Σ∗(θ) + Σ∆. (6)

• We will construct a prior for Φ∆ and Σ∆

• For now assume Σ∆ = 0.

Frank Schorfheide, University of Pennsylvania: Bayesian Methods 13

DSGE-VARs: Relaxing DSGE Restrictions

• Our prior for Φ∆ has the property that its density is proportional to the expected

likelihood ratio of Φ∗ + Φ∆ versus Φ∗.

• Likelihood ratio:
· ¸
L(Φ∗ + Φ∆, Σ∗u|Y∗, X∗)
ln (7)
L(Φ∗, Σ∗u|Y∗, X)
· µ ¶¸
1 ∆0 0 ∗0 0 ∗0 0
= − tr Σ∗−1u Φ X ∗ X ∗ Φ ∆
− 2Φ X ∗ X ∗ Φ ∆
− 2(Φ ∗
+ Φ ∆ 0 0
) X∗ Y∗ + 2Φ X ∗ Y∗ .
2

• Taking expectations yields

· · ∗ ∆ ∗
¸¸ · µ ¶¸
L(Φ + Φ , Σ |Y
u ∗ , X ∗ ) 1 0
IE D
θ ln ∗ ∗
= − tr Σ∗−1
u Φ∆ λT ΓXX Φ∆ . (8)
L(Φ , Σu|Y∗, X∗) 2
Frank Schorfheide, University of Pennsylvania: Bayesian Methods 14

DSGE-VARs: Relaxing DSGE Restrictions

• We now choose a prior density that is proportional (∝) to the expected likelihood ratio:
½ · µ ¶¸¾
1 0
p(Φ∆|Σ∗u) ∝ exp − tr λT Σ∗−1 u Φ∆ ΓXX Φ∆ . (9)
2

• Transform this prior into a prior for Φ = Φ∗(θ) + Φ∆:

à · ¸−1!
1
Φ|Σ∗, θ ∼ N Φ∗(θ), Σ∗−1 ⊗ ΓXX (θ) . (10)
λT

• Relax the assumption that Σ∆ = 0.

• Again, we obtain:
µ ¶
Σ|θ ∼ IW λT Σ∗(θ), λT − k, n (11)
à · ¸−1!
1
Φ|Σ, θ ∼ N Φ∗(θ), Σ−1 ⊗ ΓXX (θ) ,
λT
Frank Schorfheide, University of Pennsylvania: Bayesian Methods 15

DSGE-VARs: Local Misspecification

f∆, then the prior density

• If we re-scale the misspecification as follows: Φ∆ = T −1/2Φ

becomes independent of the actual sample size:

½ · µ ¶¸¾
1
f∆|Σ∗, θ) ∝ exp − tr λΣ∗−1 Φ g
∆0 Γ f∆
p(Φ XX (θ)Φ (12)
2

• Large values of λ mean small misspecifications.

f∆ is “local” misspecification. DSGE model provides good albeit not perfect approx-
•Φ

imation to reality.
Frank Schorfheide, University of Pennsylvania: Bayesian Methods 16

DSGE-VARs: Posteriors

• The joint posterior density of VAR and DSGE model parameters can be factorized:

pλ(Φ, Σ, θ|Y ) = pλ(Φ, Σ|Y, θ)pλ(θ|Y ). (13)

The λ-subscript indicates the dependence of the posterior on the hyperparameter.

Frank Schorfheide, University of Pennsylvania: Bayesian Methods 17

DSGE-VARs: Posteriors

• The posterior distribution of Φ and Σ is also of the Inverted Wishart – Normal form:
µ ¶
Σ|Y, θ ∼ IW (1 + λ)T Σ̂b(θ), (1 + λ)T − k, n (14)
µ ¶
Φ|Y, Σ, θ ∼ N Φ̂b(θ), Σ ⊗ (λT ΓXX (θ) + X 0X)−1 ,

• where Φ̂b(θ) and Σ̂b(θ) are the given by

Φ̂b(θ) = (λT ΓXX (θ) + X 0X)−1(λT ΓXY + X 0Y )

µ ¶−1 µ ¶
λ 1 X 0X λ 1 X 0Y
= ΓXX (θ) + ΓXY +
1+λ 1+λ T 1+λ 1+λ T
·
1
Σ̂b(θ) = (λT ΓY Y (θ) + Y 0Y ) − (λT ΓY X (θ) + Y 0X)
(1 + λ)T
¸
×(λT ΓXX (θ) + X 0X)−1(λT ΓXY (θ) + X 0Y ) .
Frank Schorfheide, University of Pennsylvania: Bayesian Methods 18

DSGE-VARs: Posteriors

• The marginal posterior density of θ can be obtained by evaluating the marginal likeli-

hood
n (1+λ)T −k
|λT ΓXX (θ) + X 0X|− 2 |(1 + λ)T Σ̂b(θ)|− 2
pλ(Y |θ) = n λT −k
|λT ΓXX (θ)|− 2 |λT Σ∗(θ)|− 2
n((1+λ)T −k) Qn
(2π)−nT /22 2
i=1 Γ[((1 + λ)T − k + 1 − i)/2]
× n(λT −k) Qn
.
2 2 i=1 Γ[(λT − k + 1 − i)/2]

and the prior density p(θ).

• We can also compute the marginal data density

Z
pλ(Y ) = pλ(θ|Y )p(θ)dθ. (15)
Frank Schorfheide, University of Pennsylvania: Bayesian Methods 19

DSGE-VARs: Posteriors

MCMC Algorithm for DSGE-VAR:

1. Use the RWM Algorithm to generate draws θ(s) from the marginal posterior distribu-

tion pλ(θ|Y ).

2. Use Geweke’s modified harmonic mean estimator to obtain a numerical approximation

of p̂λ(Y ).

3. For each draw θ(s) generate a pair Φ(s), Σ(s), by sampling from the IW−N distribution.
Frank Schorfheide, University of Pennsylvania: Bayesian Methods 20

DSGE-VARs: Posterior of θ
• Where does the information about θ come from? Rewrite posterior as

p(Φ, Σ, θ|Y ) = p(Φ, Σ|Y )p(θ|Φ, Σ). (16)

Projection of VAR estimates on DSGE model restriction.

• Consider quasi-likelihood function:

½ · ¸¾
1 −1
p∗(Y |θ) ∝ |Σ∗(θ)|−T /2 exp − tr Σ∗ (θ)(Y − XΦ∗(θ))0(Y − XΦ∗(θ)) . (17)
2

• Maximizing quasi-likelihood function with respect to θ is equivalent to minimizing the

discrepancy between Φ̂mle and Σ̂mle and the restriction functions Φ∗(θ), Σ∗(θ).
T
ln p∗(Y |θ) = − vech(Σ̂mle − Σ∗(θ))0D(Σ̂−1 0 ∗
mle ⊗ Σ̂mle )D vech(Σ̂mle − Σ (θ))
0
2
1
− vec(Φ̂mle − Φ∗(θ))0(Σ̂−1 0 ∗
mle ⊗ X X)vec(Φ̂mle − Φ (θ))
2
+const + small. (18)
Frank Schorfheide, University of Pennsylvania: Bayesian Methods 21

DSGE-VARs: Posterior of θ

• Proposition 1: As λ −→ ∞ (weight of the prior tends to infinity), our procedure

becomes equivalent to making inference based on the quasi-likelihood function p∗(Y |θ).

Information accumulates at rate T .

• Proposition 2: As λ −→ 0, T −→ ∞, and λT −→ ∞ (moderate weight of the

prior, large sample), the marginal log-posterior density of θ is approximately quadratic

in the discrepancy between Φ̂mle and Σ̂mle and the restriction functions Φ∗(θ), Σ∗(θ).

Information accumulates at rate λT .

Frank Schorfheide, University of Pennsylvania: Bayesian Methods 22

DSGE-VARs: Marginal Likelihood of λ

• We will study the fit of the DSGE model by examining the marginal likelihood function

of the hyperparameter λ:
Z
p(Y |λ) = p(Y |θ, Σ, Φ)pλ(θ, Σ, Φ)d(θ, Σ, Φ). (19)

• Maximum / mode:

λ̂ = argmaxλ∈Λ p(Y |λ).

• It is common in the literature to use marginal data densities to document the fit of

DSGE models relative to VARs with diffuse priors. In our framework this corresponds

to comparing

p(Y |λ = small) and p(Y |λ = ∞)

Frank Schorfheide, University of Pennsylvania: Bayesian Methods 23

Prior

Likelihood λ=∞

Φ Φ*
Frank Schorfheide, University of Pennsylvania: Bayesian Methods 24

Prior

Likelihood λ=∞

λ→0

Φ Φ*
Frank Schorfheide, University of Pennsylvania: Bayesian Methods 25

Marginal Likelihood of λ: Example

• Suppose the VAR takes the special form of an AR(1) model:

yt = φyt−1 + ut, ut ∼ iidN (0, 1) (20)

and the DSGE model restricts φ to be equal to φ∗.

• Denote the DSGE model implied autocovariances by γ0 and γ1.

• Let γ̂0 and γ̂1 be sample autocovariances.

• Prior simplifies to
µ ¶
1
φ∼N φ∗, . (21)
λT γ0
Frank Schorfheide, University of Pennsylvania: Bayesian Methods 26

Marginal Likelihood of λ: Example

• Marginal likelihood of λ takes the following form

T 2 1
ln p(Y |λ, φ∗) = −T /2 ln(2π) − σ̃ (λ, φ∗) − c(λ, φ∗). (22)
2 2

• The term σ̃ 2(λ, φ∗) measures the in-sample one-step-ahead forecast error:

2 ∗ 1X 1X
lim σ̃ (λ, φ ) = (yt − φ̂yt−1)2, 2 ∗
lim σ̃ (λ, φ ) = (yt − φ∗yt−1)2,
λ−→0 T λ−→∞ T

• The third term in (22) can be interpreted as a penalty for model complexity and is of

the form
µ ¶
γ̂0
c(λ, φ∗) = ln 1 + .
λγ0

• If an interior maximum of marginal likelihood exists, it is given by

γ0γ̂02
λ̂ = . (23)
T (γ̂0γ1 − γ0γ̂1)2 − (γ0)2γ̂0
Frank Schorfheide, University of Pennsylvania: Bayesian Methods 27

Marginal Likelihood of λ: Example

• As λ approaches zero, the marginal log likelihood function tends to minus infinity.

• Consider the comparison of two models M1 (φ∗(1)) and M2 (φ∗(2)).

– For small values of λ the goodness-of-fit terms are essentially identical. Marginal

likelihoods differentials are due to differences in the penalty terms.

– For large values of λ, marginal likelihood comparison is driven by the relative

in-sample fit of the two restricted specifications.

Frank Schorfheide, University of Pennsylvania: Bayesian Methods 28

DSGE-VARs: Posterior of λ

• Numerical Illustration in An and Schorfheide (2005):

Specification Data Set 1 Data Set 2

DSGE Model -196.66 -279.38

DSGE-VAR λ = ∞ -196.88 -277.49

DSGE-VAR λ = 5.00 -198.87 -270.46

DSGE-VAR λ = 1.00 -206.57 -258.25

DSGE-VAR λ = 0.75 -209.53 -257.53

DSGE-VAR λ = 0.50 -215.06 -258.73

DSGE-VAR λ = 0.25 -231.20 -269.66

Frank Schorfheide, University of Pennsylvania: Bayesian Methods 29

DSGE-VARs: Comparison of DSGE and VAR

• Goal of IRF comparisons is to document in which dimensions the DSGE model dy-

namics are (in)consistent with the data.

• To what extent does the VAR satisfy key structural equations implied by the DSGE?

E.g., is the Phillips curve equation misspecified?

• Examples: Cogley and Nason (1994), Rotemberg and Woodford (1997), Schorfheide

(2000), Boivin and Giannoni (2003), and Christiano, Eichenbaum, and Evans (2004),

to name a few.

• Important issue: estimation and the identification of the VAR that serves as a bench-

mark. Problems: many parameters to estimate, many shocks to identify.

Frank Schorfheide, University of Pennsylvania: Bayesian Methods 30

DSGE-VARs: Comparison of DSGE and VAR

• In our framework: compare (i) DSGE-VAR(∞) and DSGE-VAR(λ̂) IRFs; (ii) DSGE-

VAR(λ̂) and DSGE IRFs.

Frank Schorfheide, University of Pennsylvania: Bayesian Methods 31

DSGE-VARs: Identification

• So far the DSGE-VAR is reduced form. For most applications we would like a mapping

from VAR innovations into structural shocks.

• An (exactly) identified VAR is a triplet: (Φ, Σ, Ω), where Ω is orthonormal:

µ ¶
∂yt
= Σtr Ω.
∂²0t V AR

• The DSGE model is identified: there is a matrix Ω∗(θ) that maps the variance-

covariance matrix of innovations into the portion attributed to each shock:

µ ¶
∂yt
0 = Σ∗tr (θ)Ω∗(θ).
∂²t DSGE

• Identified DSGE-VAR: (Φ, Σ, Ω∗(θ)).

Frank Schorfheide, University of Pennsylvania: Bayesian Methods 32

DSGE-VARs: Identification

MCMC Algorithm for DSGE-VAR:

1. Use the RWM Algorithm to generate draws θ(s) from the marginal posterior distribu-

tion pλ(θ|Y ).

2. Use Geweke’s modified harmonic mean estimator to obtain a numerical approximation

of p̂λ(Y ).

3. For each draw θ(s) generate a pair Φ(s), Σ(s), by sampling from the IW−N distribution.

Moreover, compute the orthonormal matrix Ω(s) as described above.

Frank Schorfheide, University of Pennsylvania: Bayesian Methods 33

DSGE-VARs: IRF Comparisons

• How well is the state-space representation of the linearized DSGE model approximated

by the finite-order VAR?

• For each θ draw compare responses of the state-space version of the DSGE to the

DSGE-VAR(λ = ∞) version.

(insert figures here)

 66

Figure 11: Impulse Responses, DSGE and DSGE-VAR(λ = ∞) – Model M1 (L),

Data D5 (L)

Notes: DSGE model responses computed from state-space representation: posterior mean
(solid); DSGE-VAR(λ = ∞) responses: posterior mean (dashed) and pointwise 90% proba-
bility bands (dotted).
Frank Schorfheide, University of Pennsylvania: Bayesian Methods 34

DSGE-VARs: IRF Comparisons

• How different are the IRFs of the VAR that is estimated subject to the DSGE model

restrictions from the IRFs of the VAR in which restrictions are relaxed?

• For each (Φ, Σ, θ) draw compare responses of the state-space version of the DSGE to

the DSGE-VAR(λ = ∞) version.

• We plot posterior mean responses of DSGE-VAR(λ = ∞).

• Moreover, for each draw we compute the difference between DSGE-VAR(λ) and DSGE-

VAR(λ = ∞). We use these differences to compute a posterior mean and 90% proba-

bility bands.

(insert figures here)

 67

Figure 12: Impulse Responses, DSGE-VAR(λ = ∞) and DSGE-VAR(λ = 1) –

Model M1 (L), Data D5 (L)

Notes: DSGE-VAR(λ = ∞) posterior mean responses (solid), DSGE-VAR(λ = 1) posterior

mean responses (long dashes). Pointwise 90% probability bands (short dashes) signify shifted
probability intervals for the difference between λ = ∞ and λ = 1 responses.
Frank Schorfheide, University of Pennsylvania: Bayesian Methods 35

DSGE-VARs: IRF Comparisons

• Suppose we rewrite the structural equations as follows:

1 ˆ tπt+1] = (1 − ρg )ĝt + IE tẑt+1

ŷt − ŷt+1 + [R̂t − IE
τ
π̂t − βIE t[π̂t+1] − κŷt = −κĝt

R̂t − ρR R̂t−1 − (1 − ρR )ψ1π̂t + (1 − ρR )ψ2ŷt = −(1 − ρR )ψ2ĝt + ²R,t

• For instance, in response to a monetary policy shock, the right-hand-side of the Euler

equation and the Phillips curve equation has to be zero.

• We can check these conditions for the DSGE-VAR(λ) response.

• We overlay the right-hand-side for DSGE and DSGE-VAR.

(insert figures here)

 68

Figure 13: Impulse Responses, DSGE-VAR(λ = ∞) and DSGE-VAR(λ = 1) –

Model M1 (L), Data D5 (L)

Notes: DSGE model responses: posterior mean (solid); DSGE-VAR responses: posterior
mean (dashed) and pointwise 90% probability bands (dotted).
Frank Schorfheide, University of Pennsylvania: Bayesian Methods 36

DSGE-VARs: Some Empirical Results

• Based on Del Negro, Schorfheide, Smets, and Wouters (2006)...

• U.S. data; unless noted otherwise thirty years of observations (T = 120), starting in

QII:1974 and ending in QI:2004. Lag length p = 4.

• Forecasting exercise: beginning from QIII:1954 we construct 58 rolling samples of 120

observations.

• Four parts of the analysis:

– Parameters: priors and posteriors

– Marginal likelihood function

– Model comparison: baseline versus No Habit and No Indexation.

– Some pseudo-out-of-sample forecast error statistics.

Frank Schorfheide, University of Pennsylvania: Bayesian Methods 37

DSGE-VARs: Some Empirical Results

Prior DSGE-VECM(λ̂) Post. DSGE Post.

Mean Interval Mean Interval Mean Interval

ζp 0.60 [ 0.29 , 0.93 ] 0.79 [ 0.72 , 0.86 ] 0.83 [ 0.79 , 0.87 ]

ιp 0.50 [ 0.08 , 0.95 ] 0.75 [ 0.53 , 1.00 ] 0.76 [ 0.57 , 0.97 ]

ζw 0.60 [ 0.29 , 0.94 ] 0.79 [ 0.70 , 0.87 ] 0.89 [ 0.84 , 0.93 ]

ιw 0.50 [ 0.05 , 0.93 ] 0.45 [ 0.04 , 0.80 ] 0.70 [ 0.47 , 0.96 ]

h 0.70 [ 0.62 , 0.78 ] 0.75 [ 0.70 , 0.81 ] 0.81 [ 0.77 , 0.85 ]

ψ1 1.50 [ 0.99 , 2.09 ] 1.80 [ 1.42 , 2.19 ] 2.21 [ 1.79 , 2.63 ]

ψ2 0.20 [ 0.05 , 0.35 ] 0.16 [ 0.09 , 0.22 ] 0.07 [ 0.03 , 0.10 ]

ρr 0.50 [ 0.18 , 0.83 ] 0.76 [ 0.70 , 0.83 ] 0.82 [ 0.78 , 0.86 ]

Frank Schorfheide, University of Pennsylvania: Bayesian Methods 38

DSGE-VARs: Some Empirical Results

• Roughly: λ = ∞ dogmatically imposes DSGE model restrictions; λ = 0 completely

ignores the restrictions.

• Best fit in terms of Bayesian marginal likelihood and out-of-sample forecasting perfor-

mance is obtained for intermediate values of λ – indicating some disagreement between

DSGE and data.

 Frank Schorfheide, University of Pennsylvania: Bayesian Methods 39

Marginal Likelihood of λ: 30-Year Sample: QII:1974 to QI:2004

−1040

−1049

−1060

−1080

−1100

−1118
−1120 −1123

−1140
0.33 0.5 0.75 1 1.25 1.5 2 5 Inf DSGE
λ
 Frank Schorfheide, University of Pennsylvania: Bayesian Methods 40

Marginal Likelihood of λ: 30-Year Sample: QII:1970 to QI:2000

−1060

−1077
−1080

−1100

−1120

−1140
−1148

−1160 −1162

−1180
0.33 0.5 0.75 1 1.25 1.5 2 5 Inf DSGE
λ
 Frank Schorfheide, University of Pennsylvania: Bayesian Methods 41

Monetary Policy Shock IRFs: DSGE-VAR(λ̂) vs. DSGE-VAR(∞)

Y C I H
0.6 0.4 2 0.6

0.4 0.4
0.2 1
0.2 0.2

0 0 0 0

−0.2 −0.2 −1 −0.2

−0.4 −0.4
−0.4 −2
−0.6 −0.6

−0.8 −0.6 −3 −0.8

0 4 8 12 16 0 4 8 12 16 0 4 8 12 16 0 4 8 12 16

W Inflation R
0.3 0.3 1.5

0.2 0.2
1
0.1 0.1

0 0
0.5
−0.1 −0.1

−0.2 −0.2
0
−0.3 −0.3

−0.4 −0.4 −0.5

0 4 8 12 16 0 4 8 12 16 0 4 8 12 16
 Frank Schorfheide, University of Pennsylvania: Bayesian Methods 42

Technology Shock IRFs: DSGE-VAR(λ̂) vs. DSGE-VAR(∞)

Y C I H
2 1.4 4 1

1.2 0.8
3
1.5
1 0.6

0.8 2 0.4
1
0.6 1 0.2

0.4 0
0.5
0
0.2 −0.2

0 0 −1 −0.4
0 4 8 12 16 0 4 8 12 16 0 4 8 12 16 0 4 8 12 16

W Inflation R
1 0.1 0.2

0 0.1
0.8
−0.1 0
0.6 −0.2 −0.1

0.4 −0.3 −0.2

−0.4 −0.3
0.2
−0.5 −0.4

0 −0.6 −0.5
0 4 8 12 16 0 4 8 12 16 0 4 8 12 16
 Frank Schorfheide, University of Pennsylvania: Bayesian Methods 43

Comparing Different DSGE Model Specifications

−1040
−1049
−1058
−1060

−1080

−1100 −1101

−1118
−1120 −1123 Baseline
−1128
−1139
−1140
No Indexation
−1155
−1160

−1180

−1200

−1220
−1230
No Habit
−1240

0.33 0.5 0.75 1 1.25 1.5 2 5 Inf DSGE

λ
 Frank Schorfheide, University of Pennsylvania: Bayesian Methods 44

No Habit Specification: Monetary Policy Shock IRFs

Y C I H
0.6 0.6 1.5 0.4

0.4 0.4 1 0.2

0.2 0.2 0.5 0

0 0 0 −0.2

−0.2 −0.2 −0.5 −0.4

−0.4 −0.4 −1 −0.6

−0.6 −0.6 −1.5 −0.8

−0.8 −0.8 −2 −1
0 4 8 12 16 0 4 8 12 16 0 4 8 12 16 0 4 8 12 16

W Inflation R
0.2 0.3 1.5

0.2
0.1
1
0.1
0 0
0.5
−0.1 −0.1

−0.2
0
−0.2
−0.3

−0.3 −0.4 −0.5

0 4 8 12 16 0 4 8 12 16 0 4 8 12 16
 Frank Schorfheide, University of Pennsylvania: Bayesian Methods 45

No Habit Specification: Technology Shock IRFs

Y C I H
1.4 1 3 0.6

1.2
0.8 2 0.4
1

0.8
0.6 1 0.2
0.6

0.4
0.4 0 0
0.2

0 0.2 −1 −0.2
0 4 8 12 16 0 4 8 12 16 0 4 8 12 16 0 4 8 12 16

W Inflation R
0.8 0.3 0.4

0.2 0.3
0.6
0.1 0.2

0 0.1
0.4
−0.1 0

−0.2 −0.1
0.2
−0.3 −0.2

0 −0.4 −0.3
0 4 8 12 16 0 4 8 12 16 0 4 8 12 16
 Frank Schorfheide, University of Pennsylvania: Bayesian Methods 46

No Indexation Specification: Monetary Policy Shock IRFs

Y C I H
0.6 0.4 1 0.6

0.4 0.4
0.2 0
0.2 0.2

0 0
0 −1
−0.2 −0.2

−0.4 −0.4
−0.2 −2
−0.6 −0.6

−0.8 −0.4 −3 −0.8

0 4 8 12 16 0 4 8 12 16 0 4 8 12 16 0 4 8 12 16

W Inflation R
0.3 0.2 1.5

0.2
0.1
1
0.1

0 0
0.5
−0.1 −0.1
−0.2
0
−0.2
−0.3

−0.4 −0.3 −0.5

0 4 8 12 16 0 4 8 12 16 0 4 8 12 16
 Frank Schorfheide, University of Pennsylvania: Bayesian Methods 47

No Indexation Specification: Technology Shock IRFs

Y C I H
1.6 1 4 0.6

1.4
0.8 3 0.4
1.2

1 0.6 2 0.2

0.8 0.4 1 0
0.6
0.2 0 −0.2
0.4

0.2 0 −1 −0.4
0 4 8 12 16 0 4 8 12 16 0 4 8 12 16 0 4 8 12 16

W Inflation R
1 0.2 0.3

0.1 0.2
0.8
0 0.1
0.6 −0.1 0

0.4 −0.2 −0.1

−0.3 −0.2
0.2
−0.4 −0.3

0 −0.5 −0.4
0 4 8 12 16 0 4 8 12 16 0 4 8 12 16
Frank Schorfheide, University of Pennsylvania: Bayesian Methods 48

Discussion

Question: If we did not have the baseline model at hand, but only the alternative specifi-

cation, can we learn something from our procedure about what is missing?

• No Habit: For money shock (but also technology), clearly something is amiss! Con-

sumption (Y , and H) responds to quickly in DSGE-VAR(∞) relative to DSGE-VAR(λ̂

• . . . however DSGE-VAR(λ̂)’s IRFs are close to those of the Baseline model: Even if

the DSGE model w/o Habit is grossly misspecified, the DSGE-VAR(λ̂) is not too bad

as a benchmark.

• No Indexation: No clear evidence from Money and/or Tech IRFs that a feature that

can lead to improved fit is missing.

 Frank Schorfheide, University of Pennsylvania: Bayesian Methods 49

Forecasting Performance: One-period Ahead Root Mean Square Error Summary

12

10 10 11

10

Baseline 3
No Indexation 3
2

−2
No Habit −2

−4
0 0.33 0.5 0.75 1 1.25 1.5 2 5 Inf DSGE
λ
Frank Schorfheide, University of Pennsylvania: Bayesian Methods 50

Discussion

• Roughly: DSGE model and unrestricted VAR are comparable. DSGE-VAR improves.

• This suggests:

– Unrestricted VARs are not a good benchmark for DSGE evaluations.

– DSGE-VARs are useful tools for central banks

Frank Schorfheide, University of Pennsylvania: Bayesian Methods 51

DSGE-VARs: Extensions

• Start from DSGE with interest-rate feedback rule, allow for deviations from cross-

coefficient restrictions while maintain form of policy rule – leads to a collection of

identified VARs.

• Observables yt: Interest Rate, Inflation, Output Gap.

• Let x0t = [yt−1

0 0
, . . . , yt−p , 1]. Rewrite policy rule in general terms:

y1,t = x0tM1β1(θ) + y2,t 0

M2β2(θ) + ²1,t. (24)
|{z}
Interest Rate
¡ ¢
0 −1 D
• Define Ψ∗(θ) = IE D
θ [x x
t t ] 0
IE θ [xty2,t ] and write remainder of system:

0
y2,t = x0tΨ∗(θ) + u02,t. (25)
|{z}
Inf lation, OutputGap
Frank Schorfheide, University of Pennsylvania: Bayesian Methods 52

• VAR approximation (25) is in general not exact yet quite accurate with four lags in

our application.

• (24) and (25) comprise a partially identified (based on exclusion restrictions) VAR.
Frank Schorfheide, University of Pennsylvania: Bayesian Methods 53

DSGE-VARs: Extensions

• Rewrite Interest Rate equation

y1,t = x0tM1β1(θ) + x0tΨ∗(θ)M2β2(θ) + u1,t, (26)

• and create restricted VAR for yt

yt0 = x0tΦ + u0t, IE[utu0t] = Σ (27)

with

Φ = Φ∗(θ) = B1(θ) + Ψ∗(θ)B2(θ), Σ = Σ∗(θ).

Frank Schorfheide, University of Pennsylvania: Bayesian Methods 54

DSGE-VARs: Extensions

• There is a vector θ and matrices Ψ∆ and Σ∆ such that the data are generated from

the VAR in Eq. (27)

Φ = B1(θ) + (Ψ∗(θ) + Ψ∆)B2(θ), Σ = Σ∗(θ) + Σ∆. (28)

• (Assume Σ∆ = 0) Construct a prior with property that its density is proportional to

the expected likelihood ratio of Ψ evaluated at its (misspecified) restricted value Ψ∗(θ)

versus the “true” value Ψ = Ψ∗(θ) + Ψ∆.

• Same analysis as before... MCMC is a bit more complicated.

Frank Schorfheide, University of Pennsylvania: Bayesian Methods 55

DSGE-VARs: Extensions

• The forecast error u2,t is a function of the structural shocks: u02,t = ²1,tA1 + ²02,tA2.

• After some matrix algebra we can determine A1 and A02A2, which identifies monetary

policy shocks, but does not separate technology from demand shocks.

• We follow the idea in Del Negro and Schorfheide (2004) and decompose the DSGE
0 0
model response AD D ∗
2 (θ) = A2,tr (θ)Ω (θ).

• And then let: A2 = chol(A02A2)Ω∗(θ).

• Now we have a collection of identified VARs. Del Negro and Schorfheide (2005) study

effects of changes in the policy rule in this framework.

Frank Schorfheide, University of Pennsylvania: Bayesian Methods 56

Conclusions and Outlook

• Large body of empirical work on the Bayesian estimation / evaluation of DSGE models.

• An and Schorfheide (2005) paper illustrates techniques based on New Keynesian model.

• Model size / dimensionality of the parameter space pose challenge for MCMC methods.

• Lack of identification of structural parameters is often difficult to detect; creates a

challenge for scientific reporting.

• Model misspecification is and will remain a concern in empirical work with DSGE mod-

els despite continuous efforts by macroeconomists to develop more adequate models.

• Hence it is important to develop methods that incorporate potential model misspecifi-

cation in the measures of uncertainty constructed for forecasts and policy recommen-

dations.