Further Topics

Frank Schorfheide
University of Pennsylvania, CEPR, NBER

June 19, 2013

Further Topics

• Additive Hybrid Models

• Hierarchical Hybrid Models

• Model Mixtures

Frank Schorfheide Further Topics

Additive Hybrid Models

• Augment state-space model by latent process zt :

yt = Ψ0 (θ) + Ψ1 (θ)t + Ψs (θ)st +Λ0 + Λ1 t + Λz zt (1)
st = Φ1 (θ)st−1 + Φ (θ)t , zt = Γ1 zt−1 + ηt .

• Identification and commonly-used restrictions:

• Measurement error set-up
• Model-free detrending
• Frequency-domain restrictions
• ”Reverse minimum-distance” estimation

• Basic idea is old, but some recent papers add new twist.

References: Altug (1989, “Time-to-Build and Aggregate Fluctuations: Some New

Evidence”), Ireland (2004, “A Method for Taking Models to the Data”)

Frank Schorfheide Further Topics

Additive Hybrid Model to Capture Un-Modeled Trends

• Augment DSGE model by latent process trend process zt :

yt = Ψs (θ)st + Λ0 + zt
st = Φ1 (θ)st−1 + Φ (θ)t

zt = zt−1 + z̄t−1 + ηt
z̄t = z̄t−1 + νt

• zt captures a variety of trend processes (orthogonal to st )

• Joint estimation ensures that “trend uncertainty” is reflected in

parameter estimates.

References: Canova (2009, “Bridging Cyclical DSGE Models and the Raw Data”)

Frank Schorfheide Further Topics

Additive Trend Hybrid Model: Application

• Back to the MIU model...

• Data: output (linearly detrended), inflation, interest rates, M1

inverse velocity, target inflation

• Roughly: our target inflation series removes most of low freq

movements from inflation and interest rates; however, output and
velocity retain low freq variation.

• Add zt processes to output and velocity, using a prior variance of

one for the initial states. (User Manual suggests “diffuse” prior.)

Frank Schorfheide Further Topics

Additive Trend Hybrid Model: Parameter Estimates

W/o Trend Process W Trend Process

Name Mean 90% Intv Mean 90% Intv
Households
ν 31.8 [24.8, 38.1] 35.0 [34.3, 35.8]
Firms
ζ 0.76 [0.73, 0.78] 0.75 [0.72, 0.78]
ι 0.04 [0.00, 0.07] 0.07 [0.01, 0.11]
Central Bank
ψ2 1.03 [0.85, 1.22] 2.15 [1.79, 2.40]
Structural Shocks
ρχ 0.98 [0.97, 0.99] 0.96 [0.92, 0.99]
σχ 1.30 [1.17, 1.42] 0.57 [0.45, 0.62]
ρz 0.80 [0.72, 0.89] 0.83 [0.78, 0.91]
σZ 2.10 [1.45, 2.70] 0.81 [0.57, 1.09]

Frank Schorfheide Further Topics

Additive Trend Hybrid Model: Trend Process Parameter
Estimates
Trend Process
zt = zt−1 + z̄t−1 + ηt , and z̄t = z̄t−1 + νt .

W/o Trend Process W Trend Process

Name Mean 90% Intv Mean 90% Intv
Trend Shocks for Output
ση 0.34 [0.24, 0.43]
σν 0.27 [0.21, 0.34]
Trend Shocks for Velocity
ση 0.35 [0.25, 0.46]
σν 0.37 [0.28, 0.48]

Frank Schorfheide Further Topics

Additive Trend Hybrid Model: A Decomposition

• Fix parameters at posterior mean.

• Run Kalman filter forward, which generates:

yt = Ψ0 (θ̄) + Ψs (θ̄)st|1:t + Mz zt|1:t

• Construct two counterfactual series:

(z)
yt = Ψ0 (θ̄) + Mz zt|1:t
(s)
yt = Ψ0 (θ̄) + Ψs (θ̄)st|1:t

Frank Schorfheide Further Topics

KF Prediction (Trends Only) (Blue), HP-Trend (Green),
Actual (Red)

Frank Schorfheide Further Topics

KF Prediction (Structural States Only) (Blue), Deviations
from HP (Green)

Frank Schorfheide Further Topics

Additive Hybrid Models for Large Data Sets

• Macroeconomists have access to a large set of aggregate variables,

but it is neither feasible nor advisable to model all of these variables
in a DSGE model.

• xt is large cross section of non-modeled macro variables.

• DSGE-DFM: The joint law of motion of yt and xt is given by

yt = Ψ0 (θ) + Ψ1 (θ)t + Ψs (θ)st + ηy ,t
st = Φ1 (θ)st−1 + Φ (θ)t

xt = Λ0 + Λ1 t + Λs st + ηs,t

References: Boivin and Giannoni (2006, “DSGE Models in a Data Rich Environment”),
Kryshko (2010, “Data-Rich DSGE and Dynamic Factor Models”), Schorfheide, Sill,
Kryshko (2010, “DSGE Model-Based Forecasting of Non-modelled Variables”)

Frank Schorfheide Further Topics

DSGE & DFM Factors – Source: Kryshko (2010)
Figure D4. Do Empirical Factors and DSGE Model State Variables Span the Same Space?
Pure DFM (iid errors): Estimated and Predicted FACTORS
Factor 1 Factor 2 Factor 3
3 4 8

2 3 6

1 2 4

0 1 2

-1 0 0

-2 -1 -2

-3 -2 -4
posterior mean

-4 -3 -6
84 86 88 90 92 94 96 98 00 02 04 84 86 88 90 92 94 96 98 00 02 04 84 86 88 90 92 94 96 98 00 02 04

Factor 4 Factor 5 Factor 6

6 8 6

FACTOR5
4 6 FACTOR5_F 4

2 4 2

0 2 0

-2 0 -2

-4 -2 -4

-6 -4 -6
84 86 88 90 92 94 96 98 00 02 04 84 86 88 90 92 94 96 98 00 02 04 84 86 88 90 92 94 96 98 00 02 04

Notes: The figure plots the actual empirical factors extracted by the DFM (54)-(56) (blue line) and the empirical factors predicted by the data-rich DSGE model state
variables using (73) in the main text (red line).

72

Frank Schorfheide Further Topics

Hierarchical Hybrid Models: DSGE-VARs

• Idea: construct a prior distribution for parameters of a VAR that is

centered at DSGE model restrictions.

• This relaxes potentially misspecified DSGE model restrictions.

Frank Schorfheide Further Topics

DSGE-VARs

• VAR(p):

yt = Φ1 yt−1 + . . . + Φp yt−p + Φ0 + ut , E[ut ut0 ] = Σ.

• Identification: ut = Σtr Ω

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

• Create hierarchical model:

p(Y , Φ, Σ, Ω, θ) = p(Y |Φ, Σ)pλ (Φ, Σ|θ)p(Ω|θ)p(θ),

where pλ (Φ, Σ|θ) is a prior distribution of the VAR parameters given

the DSGE model parameters θ.

Frank Schorfheide Further Topics

Specifying a Prior p(Φ, Σ|θ)

) (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 Further Topics

Specifying a Prior p(Φ, Σ|θ)

• Quasi-likelihood function for artificial observations (sample size

T ∗ = λT ) generated from DSGE model:
p(Y ∗ (θ)|Φ, Σ) ∝
 
−λT /2 1 −1 0 ∗0 ∗ 0 ∗0 ∗
|Σu | exp − tr [Σ (−2Φ X Y + Φ X X Φ)] .
2

• Let ED
θ [·] be the expectation under DSGE model and define the
autocovariance matrices
0 0
ΓXX (θ) = ED
θ [xt xt ], ΓXY (θ) = ED
θ [xt yt ].

0 0
• Replace sample moments Y ∗ Y ∗ by ED ∗ ∗
θ [Y Y ] = λT ΓYY (θ), etc.

Frank Schorfheide Further Topics

Specifying a Prior p(Φ, Σ|θ)

• Define
Φ∗ (θ) = Γ−1
XX (θ)ΓXY (θ)
Σ∗ (θ) = ΓYY (θ) − ΓYX (θ)Γ−1
XX (θ)ΓXY (θ)

• Prior distribution:
 
Σ|θ ∼ IW λT Σ∗ (θ), λT − k
 −1 !
∗ 1 −1
Φ|Σ, θ ∼ N Φ (θ), Σ ⊗ ΓXX (θ) ,
λT

Frank Schorfheide Further Topics

Prior for Ω
• According to the DSGE model, the one-step-ahead forecast errors ut
are functions of the structural shocks t : ut = Σtr Ωt .
• Let A0 (θ) be the contemporaneous impact of t on yt according to
the DSGE model and use the factorization
 
∂yt
= A0 (θ) = Σ∗tr (θ)Ω∗ (θ), (2)
∂0t DSGE

where Σ∗tr (θ) is lower triangular and Ω∗ (θ) is an orthogonal matrix.

• Initial impact of t on yt in the VAR:
 
∂yt
= Σtr Ω. (3)
∂0t VAR

• Replace the rotation Ω in (3) with the function Ω∗ (θ) in (2).

Frank Schorfheide Further Topics

Prior for θ
• See DSGE model estimation.

Frank Schorfheide Further Topics

Posterior Distribution

• The following factorization is useful for MCMC

pλ (θ, Φ, Σ, Ω|Y ) = pλ (θ|Y )
×pλ (Φ, Σ|Y , θ)
×p(Ω|Y , Φ, Σ, θ)

Frank Schorfheide Further Topics

Posterior p(θ|Y )
• The marginal posterior density of θ can be obtained by evaluating
the marginal likelihood
pλ (Y |θ)
n (1+λ)T −k
|λT ΓXX (θ) + X 0 X |− 2 |(1 + λ)T Σ̂b (θ)|− 2
= (2π)−nT /2 n λT −k
|λT ΓXX (θ)|− 2 |λT Σ∗ (θ)|− 2

n((1+λ)T −k) Qn
2 2
i=1 Γ[((1 + λ)T − k + 1 − i)/2]
× n(λT −k) Qn .
2 i=1 Γ[(λT − k + 1 − i)/2]
2

and the prior density p(θ).

• Draws from this posterior can be obtained in the same manner as
draws in the regular Bayesian estimation of a DSGE model, e.g.
with RW Metropolis algorithm.
• Based on the MCMC output the marginal data density
Z
pλ (Y ) = pλ (θ|Y )p(θ)dθ

can be approximated.
Frank Schorfheide Further Topics
Posterior pλ (Φ, Σ|Y , θ)
• The posterior distribution of Φ and Σ is of the Inverted Wishart –
Normal form:  
Σ|Y , θ ∼ IW (1 + λ)T Σ̂b (θ), (1 + λ)T − k
 
Φ|Y , Σ, θ ∼ N Φ̂b (θ), Σ ⊗ (λT ΓXX (θ) + X 0 X )−1 ,

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

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

1
Σ̂b (θ) = (λT ΓYY (θ) + Y 0 Y ) − (λT ΓYX (θ) + Y 0 X )
(1 + λ)T

0 −1 0
×(λT ΓXX (θ) + X X ) (λT ΓXY (θ) + X Y ) .

Frank Schorfheide Further Topics

Posterior p(Ω|Y , Φ, Σ, θ)

• Recall joint distribution:

p(Y , Φ, Σ, Ω, θ) = p(Y |Φ, Σ)pλ (Φ, Σ|θ)p(Ω|θ)p(θ).

• Deduce:

p(Ω|Y , Φ, Σ, θ) ∝ p(Ω|θ)

• Here the conditional prior of Ω does not get updated because Ω

does not enter the likelihood function.

• Also note that

pλ (θ|Y , Φ, Σ) ∝ pλ (Φ, Σ|θ)p(θ)

• But: marginal posterior pλ (θ|Y ) gets updated because we learn

about (Φ, Σ) from the data.

Frank Schorfheide Further Topics

Posterior Draws for DSGE-VAR

1 Use RWM Algorithm to generate a sequence of draws θ(s) ,

s = 1, . . . , nsim , from the posterior distribution of θ, given by
pλ (θ|Y ) ∝ pλ (Y |θ)p(θ). Moreover, compute Ω(s) = Ω∗ (θ(s) ).

2 For s = 1, . . . , nsim : draw a pair (Φ(s) , Σ(s) ) from its conditional

MNIW posterior distribution given θ(s) . 2

Frank Schorfheide Further Topics

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). (4)

• Maximum / mode:

λ̂ = argmaxλ∈Λ p(Y |λ).

Frank Schorfheide Further Topics

Illustration: Marginal Likelihood of λ

Prior

Likelihood λ=∞

Φ Φ*
Frank Schorfheide Further Topics
Illustration: Marginal Likelihood of λ

Prior

Likelihood λ=∞

λ→0

Φ Φ*
Frank Schorfheide Further Topics
DSGE-VARs: An Application

• Recall the estimated monetary DSGE model...

• Monetary Policy Rule:

1−ρR ρR
Rt = R∗,t Rt−1 exp{σR R,t }
 ψ1  ψ2
πt Yt
R∗,t = (r∗ π∗,t )
π∗,t γYt−1
• Agents forecast target inflation according to:

π∗,t = π∗,t−1 + π,t .

Frank Schorfheide Further Topics

Evaluating a Monetary DSGE Model
• Suppose we compute the posterior odds of the DSGE model versus a
VAR of the form
y1,t = Φ0 + Φ1 yt−1 + . . . + Φp yt−p + Ψ∆y2,t + u1,t
y2,t = y2,t−1 + σπ∗ π∗ ,t ,
• y1,t : output, inflation, interest rates, inverse velocity
• y2,t : target inflation rate constructed from inflation expectations and
low-freq band-pass filtered inflation.
• u1,t ∼ N (0, Σ11 ) and is independent of π∗ ,t . This identifies π∗ ,t .
• For now the VAR is equipped with Minnesota prior (see Del Negro
and Schorfheide, 2010).
• Posterior odds of VAR versus DSGE model

πV ,T πV ,0 p(Y |MV )
=
πD,T πD,0 p(Y |MD )
| {z }
e 25

Frank Schorfheide Further Topics

DSGE-VAR Estimation
• Now let’s construct a prior from the DSGE model...

• Two modifications for the benchmark setup:

1 We combine the DSGE prior with a Minnesota prior such that the
prior remains proper as λ −→ 0

2 We make adjustments to account for the unit root in the DSGE

model

Frank Schorfheide Further Topics

Combining Minnesota Prior and DSGE Model Prior
• The VAR can be written as Y = X Φ + U

• Recall that priors can be represented through dummy observations

Y ∗ and X ∗ .

• Combine dummy observations that represent Minnesota prior with

dummies that represent DSGE prior, e.g.:
0
X∗ (θ)0 X∗ (θ) = (λD T )ΓD M M
XX (θ) + X∗ X∗

• Similarly for Y∗ (θ)0 X∗ (θ) and Y∗ (θ)0 Y∗ (θ).

Frank Schorfheide Further Topics

Allowing for Unit Roots
• DSGE model has state-space representation

yt = Ψ0 + Ψs st , st = Φ1 st−1 + Φ t .

• If DSGE model has unit roots then autocovariances ΓD

XX (θ) are not
time-invariant.
• Assume s−τ = 0 and t ∼ iidN(0, Σ ). Iterate state-transition
equation forward to obtain joint distribution of y0 , . . . , yp .
• Define matrices ΓD D D
XX , ΓXY , and ΓYY based on covariance matrix of
y0 , . . . , yp .
• If some of the elements of st are non-stationary and others are
stationary, the stationary ones can be initialized in period −τ
through their ergodic distribution, and the non-stationary ones with
a pointmass at zero.
• In our application, st contains one non-stationary element, namely
the target inflation rate, and we set τ = 40.

Frank Schorfheide Further Topics

Application

Frank Schorfheide Further Topics

Effect of a Change in Target Inflation (as Function of λ)

Frank Schorfheide Further Topics

Model Mixtures

• DSGE-VAR provide a way of combining a VAR and a DSGE model

• Bayesian model averaging can be used to combine different DSGE

models (or DSGE model and VAR)

• We briefly review an alternative method recently proposed by

Geweke and Amisano (2011)

References:
• Geweke, J. and G. Amisano (2011): “Optimal Prediction Pools,” Journal of
Econometrics
• Waggoner, D. and T. Zha (2011): “Confronting Misspecification in
Macroeconomics,” Working Paper, FRB Atlanta.

Frank Schorfheide Further Topics

Model Mixtures
• Recall Bayesian Model Averaging... based on posterior model
probabilities

πi,0 p(Y |Mi )

πi,T = P2
j=1 πj,0 p(Y |Mj )

• Marginal data density

T
X
ln p(Y |M) = ln p(yt |Y1:t−1 , M)
t=1

• Notice that unless two models forecast equally well on average, the
posterior probability that delivers more accurate forecasts will
converge to one as sample size increases.

• If model space is incomplete, that is, the relevant model M∗ is

omitted, then one could do better by constructing a mixture of M1
and M2 .

Frank Schorfheide Further Topics

Model Mixtures
• Option 0:
p(Y1:T |λ) = λp(Y1:T |M1 ) + (1 − λ)p(Y1:T |M2 ), λ ∈ [0, 1]
• Option 1:
T
Y
p(Y1:T |λ) = p(yt |Y1:t−1 , λ)
t=1
T
Y  
= λp(yt |Y1:t−1 , M1 ) + (1 − λ)p(yt |Y1:t−1 , M2 )
t=1
• Option 2:
p(Y1:T |λ1:T )
T
Y
= p(yt |Y1:t−1 , λ1:t )
t=1
T
Y  
= λt p(yt |Y1:t−1 , M1 ) + (1 − λt )p(yt |Y1:t−1 , M2 )
t=1
• Sequential estimation of model parameters and λ versus
simultaneous estimation of model weights and parameters.
Frank Schorfheide Further Topics