Likelihoods For DSGE Models

Frank Schorfheide
University of Pennsylvania, CEPR, NBER

June 19, 2013

Computing the DSGE Model Likelihood
• DSGE model solution leads to a (possibly nonlinear) state-space
model.

• We use a so-called filter to evaluate the likelihood function

• State-space representation of linearized DSGE model

yt = Ψ0 (θ) + Ψ1 (θ)t + Ψ2 (θ)st (+ut ) measurement
st = Φ1 (θ)st + Φ (θ)t state transition

Frank Schorfheide Likelihoods For DSGE Models

Filtering - General Idea
• Likelihood function:
T
Y
p(Y1:T |θ) = p(yt |Y1:t−1 , θ)
t=1

• A filter generates a sequence of conditional distributions st |Y1:t . (“It

removes the noise ut from the data yt to extract the signal st ”).
• Iterations:
• Initialization at time t − 1: p(st−1 |Y1:t−1 , θ)
• Forecasting t given t − 1:
1 Transition equation:
R
p(st |Y1:t−1 , θ) = p(st |st−1 , Y1:t−1 , θ)p(st−1 |Y1:t−1 , θ)dst−1
2 Measurement equation:
R
p(yt |Y1:t−1 , θ) = p(yt |st , Y1:t−1 , θ)p(st |Y1:t−1 , θ)dst
• Updating with Bayes theorem. Once yt becomes available:

p(yt |st , Y1:t−1 , θ)p(st |Y1:t−1 , θ)

p(st |Y1:t , θ) = p(st |yt , Y1:t−1 , θ) =
p(yt |Y1:t−1 , θ)

Frank Schorfheide Likelihoods For DSGE Models

Kalman Filter (Linear+Gaussian) and
Particle Filter (Fully Nonlinear)

• If the DSGE model is log-linearized and the errors are Gaussian, then
the Kalman filter can be used to construct the likelihood function.

• We will review the Kalman filter; many textbook treatments are

available.

• Alternatively, one can compute the likelihood by sequential

numerical integration which is done for DSGE models that have
been solved nonlinearly. The algorithm is called particle or sequential
Monte Carlo filter:
• Early contributions: Gordon, Salmond, and Smith (1993) and
Kitagawa (1996)
• Stochastic volatility models: Pitt and Shephard (1999) and Kim,
Shephard and Chib (1998).
• DSGE models: Fernández-Villaverde and Rubio-Ramı́rez (2007)

Frank Schorfheide Likelihoods For DSGE Models

Kalman Filter
• We’ll consider the state-space model
yt = Ψst + ut
st = Φst−1 + t
where t ∼ iidN(0, Σ) and ut ∼ iidN(0, H).

• Initialization: if
1 st is stationary we can initialize the filter with the unconditional
distribution of st . Covariance matrix:

E[st st0 ] = ΦE[st st0 ]Φ0 + Σ;

2 otherwise, could assume that st = 0 for t = −τ or treat s0 as

parameter.

• In linear Gaussian state-space model all distributions are Gaussian.

Thus, Kalman filter only tracks means and covariance matrices.

Frank Schorfheide Likelihoods For DSGE Models

Kalman Filter – Initialization
yt = Ψst + ut , st = Φst−1 + t where t ∼ N(0, Σ) and ut ∼ N(0, H).

• Write E[s0 ] = ŝ0|0 and V[s0 ] = P0|0 .

• Prior distribution for initial state s0 : s0 ∼ N (ŝ0|0 , P0|0 ).

Frank Schorfheide Likelihoods For DSGE Models

Kalman Filter – Forecasting
yt = Ψst + ut , st = Φst−1 + t where t ∼ N(0, Σ) and ut ∼ N(0, H).

• At (t − 1)+ , that is, after observing yt−1 , the belief about the state
vector has the form st−1 |Y1:t−1 ∼ N (ŝt−1|t−1 , Pt−1|t−1 ).

• “Posterior” from period t − 1 turns into a prior for (t − 1)+ .

• Since st−1 and t are independent multivariate normal random

variables, it follows that

st |Y1:t−1 ∼ N (ŝt|t−1 , Pt|t−1 )

where
ŝt|t−1 = Φŝt−1|t−1
Pt|t−1 = ΦPt−1|t−1 Φ0 + Σ

Frank Schorfheide Likelihoods For DSGE Models

Kalman Filter – Forecasting and Likelihood Function
yt = Ψst + ut , st = Φst−1 + t where t ∼ N(0, Σ) and ut ∼ N(0, H).

• The conditional distribution of yt |st , Y1:t−1 is of the form

yt |st , Y1:t−1 ∼ N (Ψst , H)
• Since st |Y1:t−1 ∼ N (ŝt|t−1 , Pt|t−1 ), we can deduce that the
marginal distribution of yt conditional on Y1:t−1 is of the form
yt |Y1:t−1 ∼ N (ŷt|t−1 , Ft|t−1 )
where ŷt|t−1 = Ψŝt|t−1 and Ft|t−1 = ΨPt|t−1 Ψ0 + H.

• Likelihood Function:
p(Y1:T |Ψ, Φ, Σ, H)
T
!−1/2
Y
= (2π)−nT /2 |Ft|t−1 |
t=1
( T
)
1X 0 −1
× exp − (yt − ŷt|t−1 ) Ft|t−1 (yt − ŷt|t−1 )
2 t=1
Frank Schorfheide Likelihoods For DSGE Models
Kalman Filter – Updating
yt = Ψst + ut , st = Φst−1 + t where t ∼ N(0, Σ) and ut ∼ N(0, H).

• To obtain the posterior distribution of st |yt , Y1:t−1 note that

st = ŝt|t−1 + (st − ŝt|t−1 )
yt = ŷt|t−1 + Ψ(st − ŝt|t−1 ) + ut

• and the joint distribution of st and yt is given by

Pt|t−1 Ψ0
     
st t−1 ŝt|t−1 Pt|t−1
Y ∼N , 0
yt ŷt|t−1 ΨPt|t−1 Ft|t−1

Frank Schorfheide Likelihoods For DSGE Models

Kalman Filter – Updating
yt = Ψst + ut , st = Φst−1 + t where t ∼ N(0, Σ) and ut ∼ N(0, H).

• Applying Bayes theorem, i.e., calculating a conditional distribution

based on a joint...

st |yt , Y1:t−1 ∼ N (ŝt|t , Pt|t )

where
−1
ŝt|t = ŝt|t−1 + Pt|t−1 Ψ0 Ft|t−1 (yt − ŷt|t−1 )
−1
Pt|t = Pt|t−1 − Pt|t−1 Ψ0 Ft|t−1 ΨPt|t−1

• The conditional mean and variance ŷt|t−1 and Ft|t−1 were given
above.

• This completes one iteration of the algorithm. The posterior st |Y1:t

is the prior for the next iteration.

Frank Schorfheide Likelihoods For DSGE Models

Particle Filter
• We consider the same state-space model as before:
yt = Ψst + ut
st = Φst−1 + t
where E[t 0t ] = Σ and E[ut ut0 ] = H.

• But now we will use Monte Carlo integration to average out the
latent states.

• This procedure can be easily generalized to nonlinear/non-Gaussian

state space models.

Frank Schorfheide Likelihoods For DSGE Models

Particle Filters - Initialization
yt = Ψst + ut , st = Φst−1 + t where t ∼ N(0, Σ) and ut ∼ N(0, H).

• Suppose we start out by drawing N particles from an initial

distribution p(s0 ), which might correspond to the unconditional
(i)
distribution associated with the state transition equation: {s0 }Ni=1 .

(i)
• By induction, in period t we start with particles {st−1 }N
i=1 which
approximate p(st−1 |Y1:t−1 ).

Frank Schorfheide Likelihoods For DSGE Models

Particle Filters - Forecasting
yt = Ψst + ut , st = Φst−1 + t where t ∼ N(0, Σ) and ut ∼ N(0, H).

(i)
• Draw one-step ahead forecasted particles {s̃t }N
i=1 from p(st |Y1:t−1 )
(i) (i)
as follows: for each s̃t−1 draw an t ∼ N (0, Σ) and let
(i) (i) (i)
s̃t = Φst−1 + t

• This simulation step approximates

Z N
1 X (i)
p(st |Y1:t−1 ) = p(st |st−1 )p(st−1 |Y1:t−1 )dst−1 ≈ p(st |st−1 )
N
i=1

Frank Schorfheide Likelihoods For DSGE Models

Particle Filter - Likelihood Function
yt = Ψst + ut , st = Φst−1 + t where t ∼ N(0, Σ) and ut ∼ N(0, H).

• The distribution yt |Y1:t−1 is approximated by

N
1 X (i)
p(yt |Y1:t−1 ) ≈ p(yt |s̃t ).
N
i=1

• Define the un-normalized weights

 
−n/2 −1/2 1 (i) 0 −1 (i)
π̃ti = (2π) |H| exp − (yt − Ψs̃t ) H (yt − Ψs̃t )
2
• The log likelihood function can be approximated as follows
ln p(Y1:T |Ψ, Φ, Σ, H)
T
X
= ln p(yt |Y1:t−1 , |Ψ, Φ, Σ, H)
t=1
T N
!
X 1 X (i)
≈ ln π̃t .
t=1
N
i=1

Frank Schorfheide Likelihoods For DSGE Models

Particle Filters - Updating
yt = Ψst + ut , st = Φst−1 + t where t ∼ N(0, Σ) and ut ∼ N(0, H).

• The goal is to approximate

p(yt |st ) p(st |Y1:t−1 )

p(st |Y1:t ) = ,
p(yt |Y1:t−1 )
(i)
• Note that for each particle s̃t we have:
(i)
p(yt |s̃t ) : equals π̃ti
(i)
p(s̃t |Y1:t−1 ) : point mass 1/N
N
X
p(yt |Y1:t−1 ) : approximated by π̃ti
i=1
(i)
• Thus, we assign to each particle s̃t the normalized weights:
(i)
(i) π̃
πt = PN t (j)
.
j=1 π̃t

Frank Schorfheide Likelihoods For DSGE Models

Particle Filters - Effective Number of Particles and
Resampling
yt = Ψst + ut , st = Φst−1 + t where t ∼ N(0, Σ) and ut ∼ N(0, H).

• At every step t we can calculate an estimate of the effective number

of particles as follows
b eff = P 1
N N (i) 2
i=1 (πt )

 N
• We can generate a new set of particles sti i=1 by re-sampling with
replacement N times from an approximate discrete representation of
 N
p (st |Y1:t t, θ) given by s̃ti , πti i=1 so that
Pr sti = s̃ti = πti , i = 1, . . . , N.

• The resulting sample is in fact an iid sample from the discretized

density of p(st | Y t ), and hence is equally weighted.

• This completes the iteration of the filter.

Frank Schorfheide Likelihoods For DSGE Models
Illustration
Consider the linear Gaussian state-space model

yt = Γ + Ψst + ut
st = Φst−1 + t

where E [ut ut0 ] = H and E [t 0t ] = Σ. Let the dimension of yt be 2 × 1

and st be 1 × 1.

• Fix parameters at

0.472
     
0.7 1 0
Γ= , Ψ= , H=
0.8 0.65 0 0.622
Φ = 1, Σ = 0.752

• 50 observations
• Kalman Filter delivers the exact likelihood.
• Particle Filter with 10, 100, 1000 particles.

Frank Schorfheide Likelihoods For DSGE Models

State Estimates

Figure: Kalman Filter Figure: Particle Filter (N = 10)

RMSE

KF N = 10 N = 100 N = 1000
0.37 0.41 0.37 0.36

Frank Schorfheide Likelihoods For DSGE Models

State Estimates

Figure: Particle Filter (N = 100) Figure: Particle Filter (N = 1000)

RMSE

KF N = 10 N = 100 N = 1000
0.37 0.41 0.37 0.36

Frank Schorfheide Likelihoods For DSGE Models

Likelihood Estimates
Likelihood estimates based on the particle filter.

Figure: # of Particles = 10 Figure: # of Particles = 100

0 0

−1 −1

−2 −2

−3 −3

−4 −4

−5 −5

−6 −6

−7 −7
KF KF
PF (N = 10) PF (N = 100)
−8 −8
10 20 30 40 50 10 20 30 40 50

Log-likelihood

KF N = 10 N = 100 N = 1000
-99.85 -99.52 -99.61 -99.87

Frank Schorfheide Likelihoods For DSGE Models

Likelihood Estimates
Likelihood estimates based on the particle filter.

Figure: # of Particles = 100 Figure: # of Particles = 1000

0 0

−1 −1

−2 −2

−3 −3

−4 −4

−5 −5

−6 −6

−7 −7
KF KF
PF (N = 100) PF (N = 1000)
−8 −8
10 20 30 40 50 10 20 30 40 50

Log-likelihood

KF N = 10 N = 100 N = 1000
-99.85 -99.52 -99.61 -99.87

Frank Schorfheide Likelihoods For DSGE Models

Effective Number of Particles

Figure: Effective Number of Particles

PF (N = 10)
800 PF (N = 100)
PF (N = 1000)

600

400

200

0
5 10 15 20 25 30 35 40 45 50

Mean

N = 10 N = 100 N = 1000
5.39 51.43 511.54

Frank Schorfheide Likelihoods For DSGE Models

(Sequential) Importance Sampling

• A particle filter is a sequential importance sampler, which also

involves some resampling steps.

• To understand the particle filter, it is important to understand

importance sampling.

Frank Schorfheide Likelihoods For DSGE Models

Importance Sampling
• Let π(θ) = f (θ)/Z . Monte Carlo approximation of
Z Z
1
Eπ [h(θ)] = h(θ)π(θ)dθ = h(θ)w (θ)g (θ)dθ,
Z
where
f (θ)
w (θ) = .
g (θ)
• Define
1
PN i i
N i=1 h(θ )w (θ )
h̄N = 1
PN ,
i
N i=1 w (θ )

where the “particles” θi ’s are drawn from the distribution with

density g (·).
• Notation: Eπ and Eg for expectations under g (·) and π(·).

Frank Schorfheide Likelihoods For DSGE Models

Importance Sampling: Consistency
• Note that
Z
Eg [hf /g ] = hZ πdθ = Z Eπ [h].

• Likewise,
Z
Eg [f /g ] = Z πdθ = Z Eπ [1] = Z .

• Provided that

Eg [|hf /g |] < ∞ and Eg [|f /g |] < ∞

• we can deduce from the Strong Law of Large Numbers and the
Continuous Mapping Theorem that
a.s.
h̄N −→ Eπ [h].

Frank Schorfheide Likelihoods For DSGE Models

Back to the Particle Filter - A Refined Updating Step
• Consider
(i)
(i) p(yt |st )p(st |st−1 )
p(st |Y t , st−1 ) = (i)
p(yt |Y t−1 , st−1 )

• Suppose we want to compute

R (i)
f (st )p(yt |st )p(st |st−1 )dst
Z
t (i)
f (st )p(st |Y , st−1 )dst = R (i)
p(yt |st )p(st |st−1 )dst

• Recall idea of importance sampler: we generate draws from some

normalized density g (st ) and approximate
(j) (j) (i)
1
PJ (j) p(yt |st )p(st |st−1 )
Z
J j=1 f (st ) (j)
(i) g (st )
f (st )p(st |Y t , st−1 )dst ≈ (j) (j) (i)
1
PJ p(yt |st )p(st |st−1 )
J j=1 (j)
g (st )

Frank Schorfheide Likelihoods For DSGE Models

Particle Filters - A Refined Updating Step
• Optimal yet often infeasible choice:

(j) (j) (i)

g (st ) = p(st |Y t , st−1 ).

which leads to the constant importance weight of

Z
(j) (j) (j) (i) (j)
π̃t = p(yt |st )p(st |st−1 )dst .

Frank Schorfheide Likelihoods For DSGE Models