Time-Varying Identification

of Structural Vector Autoregressions

Annika Camehla,∗, Tomasz Woźniakb
a
Erasmus University Rotterdam, Burgemeester Oudlaan 50, 3062 PA Rotterdam, The Netherlands
b
University of Melbourne, 111 Barry St., 3053 Carlton, VIC, Australia
arXiv:2502.19659v1 [econ.EM] 27 Feb 2025

Abstract

We propose a novel Bayesian heteroskedastic Markov-switching structural vector

autoregression with data-driven time-varying identification. The model selects among
alternative patterns of exclusion restrictions to identify structural shocks within the
Markov process regimes. We implement the selection through a multinomial prior
distribution over these patterns, which is a spike’n’slab prior for individual parameters.
By combining a Markov-switching structural matrix with heteroskedastic structural
shocks following a stochastic volatility process, the model enables shock identification
through time-varying volatility within a regime. As a result, the exclusion restrictions
become over-identifying, and their selection is driven by the signal from the data. Our
empirical application shows that data support time variation in the US monetary policy
shock identification. We also verify that time-varying volatility identifies the monetary
policy shock within the regimes.
Keywords: Identification through Heteroskedasticity, Markov Switching, Time-varying
Parameters, Verification of Identification
JEL classification: C11, C12, C32

∗
Corresponding author. Email address: [email protected].
1
1. Introduction

Most researchers assume that restrictions identifying shocks in structural vector

autoregressions remain constant over time, as proposed by Sims (1980). This
time-invariance in identification is imposed in the literature regardless of the
identification strategy or time variation in the parameter values (see, e.g., Primiceri,
2005; Sims and Zha, 2006). However, time-invariant identification strategies might be at
odds with the data. We relax this time-invariance assumption and introduce a novel
approach that implements time-varying identification (TVI) through a data-driven
search mechanism.
We propose a new model, a Bayesian Markov-switching (MS) structural vector
autoregression (SVAR) with time-varying identification and heteroskedasticity
following stochastic volatility (SV). This framework allows structural shock
identification to adapt to the data over time. Our paper has three contributions. As the
first contribution, we introduce time-varying identification through a data-driven
search. This allows our model to select alternative exclusion restrictions across regimes.
Two key features of our model implement this selection: first, we introduce a new prior
distribution to estimate a TVI indicator which alters the identification pattern in each MS
regime. Second, we let the structural matrix be regime-dependent as in Sims and Zha
(2006). Together, these features let both the structural parameters as well as different
exclusion restrictions vary over regimes. We do not rule out time-invariant identification
as no restriction pattern is associated with a specific regime. Instead, we rely on the data
to select which exclusion restrictions are supported within a regime. We implement
the data-driven search process for regime-specific identifying exclusion restrictions for a
specific shock, hence, providing partial identification.
As the second contribution, we introduce identification through heteroskedasticity
within regimes. The data-driven search of regime-dependent exclusion restriction
patterns is only feasible if these restrictions are overidentifying within the regime. We
achieve this by partially identifying a particular structural shock through

2
heteroskedasticity within a regime. To that end, we rely on three features combining
them in a novel way in our model. First, we use that our structural matrix follows a MS
process while the structural shocks’ variances instead evolve according to a SV process.
That way, the conditional variances can change at each period and shocks can thus be
heteroskedastic within a regime. Second, we adopt a non-centered representation of the
SV process as in Kastner and Frühwirth-Schnatter (2014) in our model and let the SV
prior distribution concentrate on a homoscedastic model. This specification ensures that
identification through heteroskedasticity is driven by the data rather than by the prior.
Moreover, the non-centered representation simplifies the process of verifying
heteroskedasticity, reducing it to a single parameter. This allows us to introduce as the
third novel feature, an easy way to confirm within-regime heteroskedasticity.
Specifically, we define partial identification through heteroskedasticity within regimes
extending the results in Lütkepohl, Shang, Uzeda, and Woźniak (2024) who use SVARs
without MS. We then verify identification in each regime by checking that a
shock-specific regime-dependent log-volatility standard deviation parameter implies
heteroskedastic shocks within a regime. As a new feature, we allow this volatility
parameter to be regime-dependent. That way, confirming within-regime
heteroskedasticity reduces to an easy check: verifying that this parameter is different
from zero.
As the third contribution, we provide new empirical insights by showing that the data
strongly support time variation in US monetary policy shock identification. We show that
these shocks are partially identified through heteroskedasticity. Our results are based
on a model with industrial production growth, consumer price index inflation, federal
funds rates, term spreads, M2 money, and stock prices for January 1960 to September
2024. We include four alternative restriction patterns which allow interest rates to move
simultaneously with (1) inflation and output (Leeper, Sims, Zha, Hall, and Bernanke,
1996), (2) inflation, output, and term spread (Vázquez, Marı́a-Dolores, and Londono,
2013; Diebold, Rudebusch, and Boraǧan Aruoba, 2006), (3) inflation, output, and money
(Andrés, López-Salido, and Vallés, 2006; Belongia and Ireland, 2015), and (4) only money
3
(Leeper and Roush, 2003; Sims and Zha, 2006). Additionally, we allow for two alternatives
in the fourth row: no restriction or a zero on the simultaneous movement of term spread
with interest rates, implying that monetary policy interventions narrow the term spread
while the short term interest rate does not move (Baumeister and Benati, 2013; Feldkircher
and Huber, 2018).
Our results show that in the first regime, which dominates before 2000 and in more
tranquil economic conditions, data strongly favor identifying shocks based on
simultaneous movements in interest rates, inflation, output, and term spreads. This
suggests that the monetary policy shock is best identified in relation to the entire bond
yield curve, reflecting adjusted expectations on future economic developments in spirit
of Diebold et al. (2006). In this regime, an expansionary shock increases inflation,
stimulates long-term economic growth, flattens the yield curve, features a liquidity
effect, and raises stock prices. In contrast, the second regime, which mostly occurs after
2000, particularly during the global financial crisis, the COVID-19 pandemic, and
periods of unconventional monetary policy, supports identifying monetary policy
shocks through simultaneous movements in interest rates, inflation, output, and money
supply. This aligns with the results in Belongia and Ireland (2015) who show that after
2008, monetary policy shocks are best identified by the high volatility of monetary
aggregates. In this regime, a shock that lowers interest rates significantly stimulates
economic activity after a year, increases the money supply, and steepens the bond yield
curve.
Our model contributes to the literature on SVAR identification by introducing
data-driven, regime-dependent identification patterns which extends existing methods
that impose such patterns a priori. For example, Kimura and Nakajima (2016) analyse
changes in Japan’s monetary policy using two exogeneously given regimes, based on
interest rates or bank reserves, embedded within a latent threshold time-varying
parameter SVAR. Bacchiocchi, Castelnuovo, and Fanelli (2017) impose two identification
patterns with an additive relationship in a heteroskedastic SVAR and apply them to
study the effects of monetary policy. They report differences in impulse responses
4
during the Great Moderation compared to previous periods. Arias, Rubio-Ramı́rez,
Shin, and Waggoner (2024) impose sign restrictions on impulse responses to identify US
monetary policy shocks and add restrictions on contemporaneous structural parameters
in a priori-defined periods when the Fed used the federal funds rates as the main policy
instrument. Pagliari (2024) identifies unconventional monetary policy shocks of the ECB
by restricting the expectation component spread’s reaction to the shock to be positive
before and zero after June 2014. Our approach challenges the common assumption of
time-invariant shock identification, allows the data to decide on the identification
pattern for various periods, and estimates the probabilities of regime occurrences.
Moreover, our paper relates to the literature acknowledging uncertainty around
identifying restrictions in SVAR models (such as Baumeister and Hamilton, 2018;
Giacomini, Kitagawa, and Volpicella, 2022) looking at it through the angle of uncertainty
over time.
In the following, we introduce our new model in Section 2, and report the empirical
evidence on time-varying identification of US monetary policy shocks in Section 3.

2. Data-driven time-varying identification

To study time variation in the identification of structural shocks, we estimate the structural
vector autoregressive model:

p
X
yt = Al yt−l + Ad dt + εt (1)
l=1

B (st , κ(st )) εt = ut (2)

where the vector yt of N dependent variables at time t follows a vector autoregressive

model in equation (1) with p lags, a d-vector of deterministic terms dt , an error term εt ,
and N × N autoregressive matrices Al , for lag l = 1, . . . , p, and the N × d matrix Ad of
deterministic term slopes. The structural equation (2) links the reduced form residuals to
the structural shocks ut in a time-varying linear relationship through the N × N structural

5
matrix B (st , κ(st )).
Without further restrictions the structural parameters in B cannot be recovered from
the reduced form. To solve the identification problem, exclusion restrictions can be used
(e.g., Leeper et al., 1996; Primiceri, 2005; Sims and Zha, 2006; Wu and Xia, 2016, impose
various exclusion restrictions to identify the monetary policy shock). These restrictions
impose that some contemporaneous relations in B are zero based on theoretical arguments
and are commonly assumed to hold for the whole sample. However, which variables are
contemporaneously related might change over time, and, hence, how the structural shocks
are identified could be time-varying.
Our model allows for such time-variation in the identifying exclusion restrictions by
modelling the relationship between the reduced- and structural-form shocks in
equation (2) regime-dependent. First, the parameter estimates in our structural matrix
B (st , κ(st )) depend on a regime indicator st of a discrete Markov process (see Hamilton,
1989) with M states. We assume that the MS process, st , is stationary, aperiodic, and
irreducible, with a M × M transition matrix P and an N-vector of initial values π0 ,
denoted by st ∼ Markov(P, π0 ).
Second and new to the literature, exclusion restrictions on the contemporaneous
relations to identify structural shocks can vary over time, denoted by the dependence on
a regime-specific collection of TVI indicators, κ(st ). The TVI indicator selects amongst K
patterns of the exclusion restrictions imposed on the nth equation for each regime. Thus,
we can select the nth structural shocks’ identification patterns which are specific to the
regimes identified by the MS process. Hence, we do not only allow for time-variation in
the exclusion restrictions but we endogenously determine which restrictions are
supported over time. We do not associate a priori an identifying pattern with a specific
regime but estimate for each of the K exclusion restriction schemes the occurrence
probability within each regime. In effect, a MS model without TVI might be estimated
and the TVI will only occur if data supports it. Moreover, irrespective of TVI, MS in the
structural coefficients enables time-varying impulse responses, which Lütkepohl and
Schlaak (2022) find to be important in heteroskedastic models.
6
 While in general the exclusion restrictions itself are sufficient to achieve partial
identification of the nth shock – assuming that the researcher sets them according to the
identification results in Rubio-Ramı́rez, Waggoner, and Zha (2010) – just imposing them
would be not enough to endogenously select which are supported over time. In fact, the
data-driven selection among the exclusion restriction patterns is possible if these are
overidenifying. We achieve this by identification of the nth shock through
heteroskedaticity within a regime. To that end, our model includes heteroskedastic
structural shocks that follow SV. We specify that the structural shocks at time t are
contemporaneously and temporarily uncorrelated and jointly conditionally normally
distributed given the past observations on vector yt , denoted by Yt−1 , with zero mean
and a diagonal covariance matrix:

  
ut | Yt−1 ∼ NN 0N , diag σ 2t (3)

where σ 2t is an N-vector of structural shocks’ conditional variances at time t filling in the

main diagonal of the covariance matrix, and 0N is a vector of N zeros.
Notably, our model now uses two different specifications to model time-variation:
we model regime-dependence of structural parameters and exclusion restrictions with
an MS process while we model heteroskedasticity with SV. This way, we can achieve
identification through heteroskedasticity within a regime which we discuss in detail in
section 2.2. Broadly speaking, a percentage of the observations is allocated to a specific
regime. This subset of observations can exhibit heterosekcastic residuals within the
regime as the residuals follow a SV process allowing for different conditional variances at
each point in time. Importantly, our specification implies that the MS process is driven by
changes in the identification pattern and the values of structural parameters. The former
are extreme as either a zero is imposed or not. Hence, only overwhelming support of
the data for an exclusion restriction would lead to a sharp probability to be in a specific
regime.

7
2.1. Time-varying identification

To search for changes in the identification pattern of the nth shock structural shock across
regimes, we develop the TVI mechanism and implement it through a new hierarchical
prior distribution. We set a multinomial prior distribution for the TVI component
indicator κ(m) = k ∈ {1, . . . , K} with flat probabilities equal to K1 , denoted by

 
κ(m) ∼ Multinomial K−1 , . . . , K−1 . (4)

The indicator values, κ(m) = k ∈ {1, . . . , K}, are associated with a specific restriction
pattern. Hence, this prior setup allows us to estimate posterior probabilities for each TVI
component in each regime and provides data-driven evidence in favour of the underlying
identification pattern. The indiscriminate multinomial prior reflects our agnostic view of
which identification pattern applies.
We combine the prior in equation (4) with a 3-level equation-specific local-global
hierarchical prior on the structural matrix. The prior guarantees high flexibility and avoids
arbitrary choices as the level of shrinkage is estimated within the model. For simplicity,
denote by Bm.k the matrix B (st , κ(st )) for given realisations of the MS process st = m, where
m denotes one of the M regimes, and the TVI indicator κ(m) = k, where k stands for one
out of K exclusion restrictions pattern. To implement different restriction patterns, we
follow Waggoner and Zha (2003) and decompose the nth row of the structural matrix,
[Bm.k ]n· , into a 1 × rn.m.k vector bn.m.k , collecting the unrestricted elements to be estimated,
and an rn.m.k × N matrix Vn.m.k , containing zeros and ones placing the elements of bn.m.k
at the appropriate spots, [Bm.k ]n· = bn.m.k Vn.m.k , where rn.m.k is the number of parameters
to be estimated for the restriction pattern k. The restriction matrix Vn.m.k reveals which
parameters are estimated and on which an exclusion restriction is imposed for regime m
and restriction pattern k. We have K such restriction matrices for the nth row and the mth
regime, each associated with a TVI component indicator κ(m) = k ∈ {1, . . . , K}.
Given the fixed regime st = m and sampled TVI component indicator, κ(m) = k, we
assume that bn.m.k is zero-mean normally distributed with an estimated shrinkage. This
8
equation-specific level of shrinkage for γB.n follows an inverted gamma 2 distribution
with scale sB.n and shape νB . The former hyper-parameter has a local-global hierarchical
gamma-inverse gamma 2 prior, with the global level of shrinkage determined by sγB :

b′n.m.k | γB.n , κ(m) = k ∼ Nrn.m.k 0rn.m.k , γB.n Irn.m.k , with (5)
     
γB.n | sB.n ∼ IG2 sB.n , νB , sB.n | sγB ∼ G sγB , νγB , sγB ∼ IG2 ssB , νsB , (6)

where Irn.m.k is the identity matrix. The subscript n denotes an equation-specific parameter.
In our empirical application we set νB = 10, νγB = 10, ssB = 100, and νsB = 1 to allow a wide
range of values for the structural matrix elements and show extraordinary robustness
of our results that applies as long as the shrinkage towards the zero prior mean is not
too imposing. This conditional normal prior is equivalent to that by Waggoner and Zha
(2003) for a time-invariant model and to a variant of the generalized–normal prior of
Arias, Rubio-Ramı́rez, and Waggoner (2018).
Our TVI prior specification is a multi-component generalisation of Geweke (1996)’s
spike-and-slab prior. We show this in the following Corollary.

Corollary 1. Let [Bm ]n.i denote the (n, i)th element of the regime-specific structural matrix
B(m, κ(m)) for the mth regime with indicator κ(m) = k, where k stands for one out of K exclusion
restrictions pattern. Consider the multinomial prior distributions for κ(m) = k as in (4) and the
3-level equation-specific local-global hierarchical prior on bn.m.k as in (6)–(5), where bn.m.k collects
the unrestricted elements of the nth row of Bm.k . Let KR denote the components in which the
element [Bm ]n.i is restricted to zero.
Then, the prior distribution for [Bm ]n.i , marginalized over κ(m) is

K − KR KR
[Bm ]n.i | γB.n ∼ N(0, γB ) + δ0 . (7)
K K

Proof. The element [Bm ]n.i is restricted to zero in KR components, whereas it stays
unrestricted and normally distributed in K − KR of them. Therefore, its prior distribution
KR
is a Dirac mass at value zero, δ0 , with probability K
, and normal with probability
9
K−KR
K
. □

We can easily calculate which set of exclusion restrictions is best supported by the
data by estimating the posterior probability for each restriction pattern. Having obtained
S posterior draws using a Gibbs sampler, this probability is estimated by a fraction of
posterior draws of the TVI component indicator, κ(m)(s) , for which it takes a particular
value k, for each of its values from 1 to K:

S
X
b [κ(m) = k | YT ] = S−1
Pr I(κ(m)(s) = k). (8)
s=1

Focusing on the identification of the nth structural shock, we implement the TVI
mechanism for the nth structural equation and, thus, provide partial identification of the
nth structural shock. In general, the TVI mechanism can also be applied to multiple
rows in the structural matrix and, hence, can search for TVI of multiple structural shocks
making it adaptable to a broad set of applications. It is then straightforward to calculate
the conditional posterior distribution of the nth equation specification given a particular
specification of the ith row. Also, if a specific economic interpretation is given to a particular
combination of exclusion restrictions for the structural matrix, one can estimate the joint
posterior probability of such a specification by computing the fraction of the posterior
draws for which this combination of TVI indicators holds.

2.2. Within-regime identification through heteroskedasticity

Our search of the identification pattern is feasible only if data can distinguish between the
different identifying patterns. Our TVI mechanism in the nth row of the structural matrix
requires that the nth structural shock is identified within each regime, absent the potential
additional restrictions we are searching for. We partially identify the nth structural shock
and the nth row of the structural matrix within a regime through heteroskedasticity. The
partial identification through heteroskedasticity focuses on the identification of a specific
shock and all parameters of the corresponding row of the structural matrix within a
regime without the necessity of imposing additional restrictions.
10
 Identification through heteroskedasticity within regimes is feasible because the
volatility of the structural shocks evolves according to a SV process. The volatility can
have different values at each point in time. Instead, the structural matrix follows a MS
process where time-variation occurs via allocating a subset of observations to a specific
regime. Hence, structural residuals can still be heteroskedastic within the regime (and
not only across regimes). Once the identification of a shock through heteroskedasticity is
verified in each of the regimes, any set of exclusion restrictions overidentify the shock,
which enables the data to discriminate between the alternative restriction patterns.
We obtain partial identification of the nth row of the regime-specific structural matrix,
denoted by Bm , where the dependence on k is neglected, within each regime under the
following corollary.

Corollary 2. Consider the SVAR model with the structural matrix following MS from equations
(1)–(3). Denote by t1(m) , . . . , t(m)
TM
the time subscripts of the Tm observations allocated to the
PM
Markov-switching regime st = m ∈ {1, . . . , M}, where m=1 Tm = T. Let
σ 2n.m = (σ2 (m) , . . . , σ2 (m) ) be the vector containing all conditional variances σ2n.t of the nth row
n.t1 n.tT
M
associated with the observations in the mth regime. Then the nth row of Bm is identified up to a
sign if σ 2n.m , σ 2i.m for all i ∈ {1, . . . , N} \ {n}.

Proof. The proof proceeds by the same matrix result as in Corollary 1 by Lütkepohl et al.
(2024) set for a heteroskedastic SVAR model without time-variation in the structural
matrix. In line with Theorem 1 by Lütkepohl and Woźniak (2020), the matrix result holds
even if the inequality above is true for one period. □

According to Corollary 2, the nth row of Bm.k is identified through heteroskedasticity in

the mth regime if changes in the conditional variances of the nth structural shocks evolve
non-proportionally to those of other shocks within that regime. This result assures that
the corresponding impulse responses, the nth column of B−1
m.k
, are identified as well. Global
identification of Bm through heteroskedasticity is achieved when Corollary 2 holds for all
n ∈ {1, . . . , N}.

11
 It is essential to verify identification through heteroskedasticity within each regime
(see the argument in ?) and, to that end, we introduce as a new feature of our model
that the SV process of the structural shocks has a regime-dependent standard deviation
parameter. Each of the N conditional variances, σ2n.t , follows a non-centered SV process
with regime-dependent standard deviation parameter that decomposes the conditional
variances into

σ2n.t = exp{ωn (st )hn.t } with (9)

hn.t = ρn hn.t−1 + vn.t and vn.t ∼ N(0, 1) (10)

where ωn (st ) is a standard deviation parameter of the log-conditional variances log σ2n.t .
Equation (9) implies that both hn.t and ωn (st ) are identified up to a simultaneous sign
change without affecting the conditional variance σ2n.t . The log-volatility of the nth
structural shock at time t, hn.t , follows an autoregressive model with initial value hn.0 = 0,
where ρn is the autoregressive parameter and vn.t is a standard normal innovation. Note
that the unconditional expected value of the conditional variances σ2n.t depends on ωn (st )
and ρn . To fix it, we normalise the structural shocks’ conditional variances around one
by a prior that shrinks the model towards homoskedasticity which we explain in section
2.3. The specification extends the SV process by Kastner and Frühwirth-Schnatter (2014)
by the regime-dependence in the parameter ωn (st ). Therefore, the Markov process st
drives the time variation in both B (st , κ(st )) and ωn (st ).
We can verify the hypothesis of within-regime homoskedasticity of each of the shocks
by checking the restriction (see Chan, 2018, for time-invariant heteroskedastic models)

ωn (m) = 0. (11)

If it holds, the nth structural shock is homoskedastic in the mth regime. If it does not, then
the structural shock’s conditional variance changes non-proportionally to those of other
shocks with probability 1, rendering the shock identified through heteroskedasticity.

12
Consequently, two cases guarantee the identification of the nth structural shock. First,
this shock, and no other, is homoskedastic within a specific regime. Hence, the condition
in equation (11) only holds for the nth shock. Second, the nth structural shock is
heteroskedastic in that regime implying that restriction (11) is not satisfied for the nth
shock.

2.3. Priors, Normalisation, and Posterior Sampler

In this section, we discuss the prior specification of the remaining parameters of the model
and their estimation. We use the following notation: en.N is the nth column of the identity
matrix of order N, ın is a vector of n ones, p is a vector of integers from 1 to p, and ⊗ is the
Kronecker product.
 
Let the N×(Np+d) matrix A = A1 . . . Ap Ad collect all autoregressive and constant
parameters. Each row of matrix A is independent and follows conditional multivariate
normal distribution with a 3-level global-local hierarchical prior on the equation-specific
shrinkage parameters γA.n :

 
[A]′n· | γA.n ∼ NNp+d m′A.n , γA.n ΩA , with (12)
     
γA.n | sA.n ∼ IG2 sA.n , νA , sA.n | sγA.n ∼ G sγA , νγA , sγA ∼ IG2 ssA , νsA , (13)

   ′
where mA.n = e′n.N 0′N(p−1)+d , ΩA is a diagonal matrix with vector p −2′
⊗ ı′N 100ı′d on the
main diagonal, hence, incorporating the ideas of the Minnesota prior of Doan, Litterman,
and Sims (1984). In our empirical application we set νA , νγA , ssA , and νsA all equal to 10,
which facilitates relatively strong shrinkage that gets updated, nevertheless. Providing
sufficient flexibility on this 3-level hierarchical prior distribution occurred essential for a
robust shape of the estimated impulse responses.
Each row of the transition probabilities matrix P of the Markov process follows
independently a Dirichlet distribution, [P]m· ∼ Dirichlet(ıM + dm em.M ), as does the initial
regime probabilities vector π0 , π0 ∼ Dirichlet(ıM ). To assure the prior expected regime
duration of 12 months, we set dm = 11 for models with M = 2 in our application, and

13
show the robustness of our results to alternative prior expected regime durations.
The prior distribution for the regime-specific standard deviation parameter, ωn (st ),


follows a conditional normal distribution, ωn (m) | σ2ω.n ∼ N 0, σ2ω.n , with a gamma
distribution on the shrinkage level, σ2ω.n ∼ G(1, 1). Notably, the marginal prior for ωn (m)
combines extreme prior probability mass around the restriction for homoskedasticity
ωn (m) = 0 and fat tails. The latter enables efficient extraction of the heteroskedasticity
signal from data assuring that a shock is heteroskedastic within a regime only if the data
favour this outcome. Our empirical results remain unchanged subject to variation of the
prior scale of σ2ω.n ranging from values 0.1 to 2. Finally, the prior distribution for the
autoregressive parameter of the log-volatility process is uniform over the stationarity
region, ρn ∼ U(−1, 1).
This prior setup following Lütkepohl et al. (2024) implies that the structural shocks’
conditional variance, σ2n.t , have a log-normal-product marginal prior implied by the
normal prior distributions for the latent volatility, hn.t , and for the standard deviation
parameters, ωn (m). The log-normal-product marginal prior has a pole at value one,
corresponding to homoskedasticity, fat tails, and a limit at value zero when the
conditional variance goes to zero from the right. Importantly, the extreme concentration
of the prior mass at one and strong shrinkage toward this value standardises the
structural shocks’ conditional variances around value σ2n.t = 1. This assures appropriate
scaling in the estimation of the structural matrix. In our empirical model, the
normalisation holds true as the estimated variances fluctuate closely around one.
To draw from the posterior distributions we use a Gibbs sampler. We provide an R
package bsvarTVPs facilitating the reproduction of our results as well as new
applications.1

1
Access the bsvarTVPs package at https://github.com/bsvars/bsvarTVPs

14
 Table 1: Monetary policy shocks identification patterns

third row fourth row

yt πt Rt TSt mt spt yt πt Rt TSt mt spt
baseline ∗ ∗ ∗ 0 0 0 unrest. ∗ ∗ ∗ ∗ ∗ ∗
with TS ∗ ∗ ∗ ∗ 0 0 with 0 ∗ ∗ 0 ∗ ∗ ∗
with m ∗ ∗ ∗ 0 ∗ 0
only m 0 0 ∗ 0 ∗ 0

Note: ∗ indicates an unrestricted parameter and 0 an exclusion restriction on the contemporaneous

parameters in the third or fourth row of B.

3. Time-varying identification of US monetary policy shocks

Over the last sixty years, economic conditions, the arrival of new technologies,
understanding the role of communication, and new strategies of conducting monetary
policy in the US seem to affect the most efficient ways of extracting the structural shocks
from data. Despite these circumstances, empirical studies on the effects of the US
monetary policy shock assume that restrictions used to identify the structural shocks do
not change over time.
We challenge this view and study time-variation in US monetary policy shock
identification using data from January 1960 to September 2024 on log differenced
industrial production, yt , log consumer price index inflation, πt , federal funds rates, Rt ,
term spread, TSt , measured as the 10-year treasury constant maturity rate minus the
federal funds rate, log M2 money supply, mt , and the log S&P500 stock price index, spt in
a model with six lags. The Online Appendix gives details on the data.
We specify four (K3 = 4) exclusion restriction patterns for the third row to identify the
US monetary policy shock, shown in the left panel of Table 1.2 The first identification
pattern, that we call baseline, imposes simultaneity in short-term interest rates, prices and
output (Leeper et al., 1996). The second identification pattern, with TS, relatively to the
baseline, allows additional simultaneous movements of the term spread, TSt , in spirit of
Vázquez et al. (2013) and Diebold et al. (2006). We consider this identification in response

2
We found that an unrestricted third row does not provide economically interpretable monetary policy
shocks. On the contrary, the additional exclusion restrictions sharpen the identification and interpretations.
15
to the monetary policy shock identification strategies by Baumeister and Benati (2013),
Feldkircher and Huber (2018), and Liu, Mumtaz, Theodoridis, and Zanetti (2019) but
emphasise the spread’s role in determining the interest rates.
The third identification, with m, allows for simultaneity in interest rates and money,
on top of output and prices, inspired by Andrés et al. (2006) and Belongia and Ireland
(2015). This identification suggests synchronous movements in interest rates and market
liquidity levels. The fourth identification pattern, only m, for the monetary policy shock
highlights the simultaneity between short-term interest rates and a monetary aggregate,
in line with the identification used by Leeper and Roush (2003) and Sims and Zha (2006).
We specify two (K4 = 2) restriction patterns for the fourth row, shown in the right
panel of Table 1. Here, we either set no restrictions, labeled unrestricted, or impose a zero
on the simultaneous movement of term spread with interest rates, labeled with 0. This
zero restriction implies that monetary policy interventions can have a narrowing effect
on the term spread while the short term interest rate does not move. Such a restriction on
the contemporaneous relations is in spirit with the notion that short term interest rates
do not react to a term spread shock for four quarters imposed by Baumeister and Benati
(2013) and Feldkircher and Huber (2018). We impose a lower-triangular structure on the
remaining rows of the structural matrix.

3.1. Model evaluation

In order to verify whether data support TVI, we evaluate the forecasting performance
of models.3 More specifically, we extend a heteroskedastic SVAR in two dimensions:
by applying a different number of regimes M ∈ {1, 2, 3}, and by using models with the
multinomial prior over the identification patterns listed in Table 1 as well as models with
fixed identification patterns over MS regimes.
We use relative predictive log-scores, as defined by Geweke and Amisano (2010),
to evaluate density forecasts and relative root-mean-squared forecast errors for point

3
Forecasting performance measures are unbiased as they are invariant to local identification of the SVARs
up to the signs and ordering of structural equations, the signs of the SV log-volatility components, and MS
regime labeling (see Geweke, 2007).
16
 1.15
relative root−squared forecast error
0
relative predictive log score
−50

1.10
−100

1.05
−150

1.00
−200
−250

0.95
−300

2008 2010 2012 2014 2016 2018 2020 2022 2008 2010 2012 2014 2016 2018 2020 2022

forecast evaluation period forecast evaluation period

(a) Density forecast performance (b) Point forecast performance

Figure 1: Forecasting performance of selected models

Note: The legend for models includes: MS models (dark colour): M = 3 with TVI, M = 2 baseline,
M = 2 with TS, M = 2 with m, M = 2 only m, no MS (M = 1, bright colour): with TVI,
baseline, with TS, with m, only m. Measures are relative to the benchmark model with M = 2
and TVI.

forecasts. In both cases the benchmark model is the heteroskedastic SVAR with M = 2
MS regimes and the TVI selection mechanism proposed in this paper. Therefore, a
relative log-score higher than 0 or a relative root-mean-squared forecast error lower than
1 indicates that a model outperforms the benchmark. We compute the measures based on
190 one- and twelve-step-ahead forecasts with the forecast origins starting in December
2007 and ending in September 2023.
Figure 1 reports the one-step-ahead forecasting performance based on density (panel
a) and point forecasts (panel b). We find strong support for using TVI and MS. Overall,
the model with two regimes and TVI used as the benchmark preforms best according to
both measures, followed by the model with three regimes and TVI (solid dark line). The
latter outperforms the benchmark at the beginning of the sample and around 2020 based
on density forecasts and roughly until 2015 based on point forecasts. The model with TVI
but no MS (solid bright line) exhibits accurate point forecasting performance, reasonably

17
close to the benchmark, but performs poorly in density forecasting. Models with TVI
outperform the alternatives with no TVI (dashed or dotted lines). The twelve-steps-ahead
forecasting performance, discussed in the Online Appendix, provides a bit more nuanced
picture. Still, the benchmark model is not uniformly outperformed by any other model
according to both measures.
Based on these findings, we provide more insights using the benchmark model with
two MS regimes and TVI.

3.2. TVI probabilities

We find that data support TVI of US monetary policy shocks. To show this, first, we
estimate the posterior probability of the structural matrix featuring different identification
in the two regimes: Pr [κ3 (m = 1) , κ3 (m = 2) ∨ κ4 (m = 1) , κ4 (m = 2) | YT ]. Given our
model setup, the probability includes the intersection of changing the identification in
both rows, third and fourth. In the two estimated MS regimes, data favor different
identification schemes with strong posterior support of 0.87.
Moving on to the monetary policy shock identification, we find remarkably sharp
posterior probabilities of the TVI indicators reported in Figure 2. In the first regime
indicated by the light color in Figure 2, we find strong data support that the monetary
policy shock is identified by simultaneous movements of interest rates with output,
inflation, and the term spread, combined with no restrictions in the forth row on the
simultaneous movements of term spreads to all other variables. The posterior probability
of the TVI component indicators for the restriction pattern with TS and unrestricted are
numerically equal to one. This suggests that the monetary policy shock is best identified
with consideration to the whole bond yield curve signaling adjusted expectations on
future economic developments in the spirit of Diebold et al. (2006).
In the second regime visualized by the dark color in Figure 2, we find evidence that
data support a mixed identification of the monetary policy shock. The highest posterior
probability is attached to the restriction pattern allowing interest rates also to respond
simultaneously to changes in the monetary aggregate. The posterior value of the TVI

18
 1

1
regime 1 regime 1
component probability

component probability
regime 2 regime 2
0

0
baseline with TS with m only m unrestricted with 0

(a) third row (b) fourth row

Figure 2: Posterior probabilities of the TVI components in both regimes

component indicator for the third exclusion restriction pattern (with m) is 0.45. The
posterior probability of the remaining components is lower, with first and second both
having 0.27 posterior probability and the fourth (only m) 0.01.
Our findings are in line with Belongia and Ireland (2015) supporting that after 2008,
the monetary policy shock is best identified considering the high variability of the
monetary aggregates. While the Fed officially targeted money supply before 1982,
Belongia and Ireland (2015) argue that unconventional monetary policy can be seen as
an attempt to increase money growth. Similar to the support of data for money in the
identification pattern for the monetary policy shock in the second regime, the authors
show that excluding a monetary measure from the interest rate rule is rejected by the
data in the sample to 2007. Similarly, Lütkepohl and Woźniak (2020) find empirical
evidence for monetary policy shock identification allowing for simultaneity in interest
rates and money for data up to 2013. Neither of these papers considers identification
pattern including the term spread or time variation of the structural matrix.
The posterior probabilities of the TVI component indicator for the fourth row show a
dominant role of the unrestricted identification in the first regime, and equal support for
unrestricted and with 0 in the second one. This indicates that in the second regime, which
predominately covers the period of the zero-lower-bound environment, monetary policy
interventions partly have a narrowing effect on the term spread while the short term

19
interest rate does not move. However, since the second regime also covers periods
outside the zero lower bound period, we find only weak empirical support for the
restriction which is in spirit of the assumption imposed by Baumeister and Benati (2013)
and Feldkircher and Huber (2018) that short term interest rates do not immediately react
to a term spread shock.
Selecting two different identification patterns over time is remarkably robust with
respect to extending or limiting the set of possible exclusion restrictions in the search
(such as allowing for interest rates to move simultaneously with stock prices, including
patterns with zero restrictions on money and stock prices or zero restrictions on interest
rates, money and stock prices in the fourth row, excluding the variant only m, or allowing
only for simultaneous movements of interest rates, output, and inflation with term spread
or money in the third row). Moreover, models with changes in the prior set-ups on the
contemporaneous relations, or changes in the variables (using the consumer price index
and industrial production index instead of growth rates, or the interpolated GDP deflator
instead of CPI inflation) support time-variation in the identification of the monetary policy
shock.4 We report the robustness results in the Online Appendix.

3.3. Within-regime identification through heteroskedasticity

Our search for regime-specific identification patterns requires that the exclusion
restrictions are over-identifying and, thus, can be tested using statistical methods. We
achieve this by identifying the monetary policy shock through heteroskedasticity.
Figure 3 plots histograms of the standard deviation of the third shock ω3 (m) and of
the fourth shock ω4 (m) for both regimes, m ∈ {1, 2}. We find clear evidence for
heteroskedasticity within both regimes for the third and fourth equation visualized by
the bi-modality of the posterior distributions. The bi-modality results from the fact that

4
While we focus on reporting robustness to the changes in the prior specification most relevant
to the discussed feature, all our main findings regarding TVI (Section 3.2), identification through
heteroskedasticity (3.3), and Markov-switching probabilities (3.4) are robust to changing hyper-parameters
on the priors set on the contemporaneous relationships, autoregressive coefficients, SV or the
Markov-process, to different lag lengths (from 1 to 12), and using log first differences of money and
stock prices. Results are available upon request.
20
 3.0

regime 1 regime 1

3.0
regime 2 regime 2
2.0

2.0
Density

Density
p(ω3(m)) p(ω4(m))
1.0

1.0
0.0

0.0
−1.5 −1.0 −0.5 0.0 0.5 1.0 1.5 −1.5 −1.0 −0.5 0.0 0.5 1.0 1.5

ω3(m) ω4(m)

(a) ω3 (b) ω4

Figure 3: Marginal posterior and prior densities of the regime-specific standard deviations

ω3 (m) and ω4 (m), as well as hn.t , are identified up to their simultaneous change of sign,
which is implied by equation (9). This is taken into account when verifying
heteroskedasticity via evaluating ωn (m) = 0 following Chan (2018). Verifying that ω3 (m)
and ω4 (m) differ from zero implies identifying the underlying equations. In both cases,
the mass of the posterior distributions is away from zero. Therefore, the third and fourth
row are identified through heteroskedasticity and our selection of identifying
restrictions is valid in both regimes. We then label the third shock as the monetary
policy shock based on the exclusion restrictions imposed.
We find that the identification of the third and fourth row through heteroskedasticity
is robust to changing the scale hyper-parameter of the prior controlling for the
regime-dependent standard deviation parameter level to 0.5 or 2, see results in the
Online Appendix. Also, using alternative measurements for our variables, such as the
consumer price index, industrial production, or the interpolated GDP deflator, has no
impact on identification through heteroskedasticity.

3.4. Two Markov-switching regimes

Having established that the data support regime-specific identification of US monetary

policy shocks, we now take a closer look at the two estimated MS regimes. Figure 4 shows
the estimated regime probabilities for the second regime (a gray shaded area), together

21
 A B C D E F G H I J K LM N O P
component probability

1
0

1961 Q1 1971 Q1 1981 Q1 1991 Q1 2001 Q1 2011 Q1 2021 Q1

time
List of events:
A Dec 1982 peak of 1981-1982 recession
B Mar 2000 dot-com bubble burst
C Sep 2001 September 11 attacks
D Sep 2008 bankruptcy of Lehman Brothers on September 15, 2008
E Nov 2008 FOMC announces Quantitative Easing 1
F Jun 2009 NBER declares end of the US great recession
G Nov 2010 FOMC announces Quantitative Easing 2
H Aug 2011 FOMC announces to keep the federal funds rate at its effective zero lower bound “at least
through mid-2013”∗
I Sep 2011 FOMC announces Operation Twist
J Sept 2012 FOMC announces Quantitative Easing 3
K Dec 2012 FOMC announces to purchase 45B$ of longer-term Treasuries per month for the future,
and to keep the federal funds rate at its effective zero lower bound
Dec 2013 FOMC announces to start lowering the purchases of longer-term Treasuries and
L mortgage-backed securities (to $40B and $35B per month, respectively)
M Dec 2014 FOMC announces “it can be patient in beginning to normalize the stance of monetary
policy”
N Mar 2015 FOMC announces “an increase in the target range for the federal funds rate remains
unlikely at the April FOMC meeting”
O Mar 2020 start of COVID-19 pandemic. First unscheduled FOMC meeting on March 3 in response to
the outbreak of the pandemic announcing to “lower the target range for the federal funds
rate by 1/2 percentage point” and to “use its tools and act as appropriate to support the
economy.” On March 23, FOMC statement to “continue to purchase Treasury securities
and agency mortgage-backed securities in the amounts needed to support smooth market
functioning”.
P Mar 2022 First indications of changes in monetary policy with FOMC annoucing that one member
voted against keeping the federal funds rate at the same rate. In May and June the federal
funds rate was increased with the largest hikes since May 2000 and 1994, respectively.
∗ all quotes from https://www.federalreserve.gov/newsevents/pressreleases.htm.

Figure 4: Regime probabilities of the Markov process

with highlighted periods when the probability for the second regime is larger than 0.8
(dark-colored thick line), and specific events (black vertical lines topped by letters).
Regime probabilities are volatile and often not sharp. This is different from
“traditional” MS-VAR models where the MS process is driven by changes in
autoregressive parameters or volatility of the residuals. In our model, the MS
probabilities predominantly capture changes in the contemporaneous structural
relations. Importantly, the two regimes are characterized by two different identification
22
schemes as revealed by the estimated TVI probabilities. Hence, the estimated
regime-specific structural matrices differ sharply by imposing a zero on a specific
parameter versus estimating this parameter freely.
We interpret high regime probabilities as overwhelming evidence that monetary
shock identification is best captured by the TVI selected identification pattern during
this period. Especially, the second regime’s supported identification of a monetary
policy shock requires strong simultaneous movements in money and interest rates.
Hence, high probabilities for the second regime show up in times when data feature
such strong simultaneity, which we discuss further in Section 3.5. The identification
scheme selected in the second regime captures such data characteristics for the sake of
sharper identification of the shocks.
Regime one is predominant in the first part of the sample until 2000. It is present
in periods of rather normal economic developments (aka non-crisis times). After 2000
the second regime occurs more frequent and the periods are characterized by frequent
switches between the two regimes. The second regime clearly prevails during crisis
periods, in particular the financial and COVID-19 ones. In the aftermaths of these events,
regime 2 gains persistence. Additionally, it is present at earlier extra-ordinary times such
as the peak of the 1981–1982 crisis, the burst of the dot-com bubble in March 2000, or 9/11.
It also predominately covers periods of unconventional monetary policy actions such as
Quantitative Easing 1, Operation Twist, and several subsequent announcements of the
Fed, starting in August 2011, stating that they would keep the federal funds rate at its
effective zero-lower bound for a substantial period.
Notably, we find that the regime allocations are robust to various changes in the model
specification. The estimated regime probabilities, shown in Figure 5, are very similar to a
model where we impose the prior belief that regimes are less persistent or more persistent
by setting the prior expected regime duration of the Markov process to one month, called
prior MS 1, or 2 years doubling the expected prior regime duration, called prior MS 24,
respectively. Moreover, models using alternative measurements for our variables, such
as the consumer price index and industrial production index, model called CPI, IP, or
23
 prior MS 1
prior MS 24
component probability

CPI, IP
GDPDEF
0

1961 Q1 1971 Q1 1981 Q1 1991 Q1 2001 Q1 2011 Q1 2021 Q1

time

Figure 5: Regime probabilities of the Markov process across alternative models

the interpolated GDP deflator, model called GDPDEF, yield similar regime allocation
probabilities.

3.5. Three results illustrating regime differences

Next, we discuss differences across the regimes in more detail by looking at sample
moments of data, estimates of the contemporaneous relations in the interest rate equations,
and dynamic responses to US monetary policy shocks.

Regime-specific sample moments of data. Regime-specific sample moments reported in

Table 2 show that the federal funds rate is on average lower and the term spread higher
in the second regime while they both exhibit lower volatility then. Money is twice as
volatile in the second regime. Moreover, the covariance between interest rates and term
spreads is higher in the first than in the second regime. Differences are strong with
respect to the relation between money and interest rates. In the first regime the
covariance is positive, while in the second it is negative.
The sample regime-specific moments indicate that the second regime captures
monetary policy that is characterized by cutting the interest rates and narrowing the
bond yield curve supplemented by expanding the monetary basis. This supports our
TVI results on the dominate role of money in the second regime. It suggest that the
differences in the correlation of the variables picked up in our model by the
contemporaneous relations are a decisive feature to determine regime allocation and

24
 Table 2: Regime-specific sample moments of data

Regime 1 Regime 2
mean sd cov with Rt mean sd cov with Rt
Rt 5.64 3.62 2.54 2.73
TSt 0.84 1.69 −3.79 1.43 1.34 −2.15
mt 6.33 3.86 1.43 7.38 8.63 −2.03

Note: Table reports sample means, standard deviations, and the covariance with interest rates for variables
in both regimes. The regime-specific moments are given for the series in first differences for mt .

differences across the regimes.

Parameter estimates of the interest rate equation. The posterior means of the
contemporaneous coefficients in the interest rate equation, the third row of the structural
matrix with pattern with TS and with m, with the bounds of the 68% highest density
interval in parentheses,

mps
Regime 1 0.01 yt −0.07 πt +3.64 Rt +4.09 TSt = lags + ût
(−0.00; 0.01) (−0.09; −0.05) (3.45; 3.82) (3.92; 4.25)
mps
Regime 2 −0.03 yt −0.02 πt +14.00 Rt −0.45 mt = lags + ût
(−0.03; −0.02) (−0.05; 0.02) (11.32; 15.94) (−0.71; −0.02)

show notable differences across the regimes. In the first regime, interest rates and term
spreads are clearly contemporaneously related. The posterior mean of the
contemporaneous coefficient on term spreads is positive, with strong evidence that it is
different from zero. Interest rates move simultaneously with inflation, whereas the
coefficient on output is not different from zero.
In the second regime, money aggregate plays a dominant role, as evidenced by a
significant and relatively highest in absolute terms coefficient on this variable. This
regime also observes an increase in the importance of output relative to inflation. The
coefficient on the latter is not different from zero. This fact, together with the timing of the
second regime, is in line with the shift in the emphasis in the monetary policy after 2000
described by Belongia and Ireland (2016). For the remaining two identification pattern

25
 0.6

−0.2
0.5

change in pi [pp]
0.4

change in R [pp]
change in y [pp]

−0.4
0.0

0.2
−0.5

−0.6
0.0
−1.0

−0.8
−0.2
−1.5

−1.0
−0.4
1 2 3 4 5 1 2 3 4 5 1 2 3 4 5

horizon [years] horizon [years] horizon [years]

(a) response of y (b) response of π (c) response of R

2.0
change in TS [pp]

change in sp [%]
change in m [%]
0.0

5
1.5
1.0
−0.5

0
0.5
−1.0

−5
0.0

1 2 3 4 5 1 2 3 4 5 1 2 3 4 5

horizon [years] horizon [years] horizon [years]

(d) response of TS (e) response of m (f) response of sp

Figure 6: Impulse response to US monetary policy shocks

which get posterior weights in the second regime the posterior estimates

mps
Regime 2: −0.02 yt −0.01 πt +14.86 Rt = lags + ût
(−0.03; −0.01) (−0.04; 0.02) (12.24; 16.87)
mps
Regime 2 −0.02 yt −0.01 πt +14.83 Rt +0.02 TSt = lags + ût
(−0.03; −0.01) (−0.04; 0.02) (11.93; 16.58) (−0.45; 0.53)

support the emphasis on output while at the same time the coefficient on term spreads is
not different from zero.

Dynamic responses to US monetary policy shocks. We find notable differences in the dynamic
responses to the US monetary policy shock across the regimes. Based on its effects in the
first regime, monetary policy can be characterized as conventionally inflation-targeting,
whereas in the second one, as expansionary with specific macro-financial interactions.
Our analysis is based on Figure 6 reporting the median impulse responses to a monetary
policy shock decreasing interest rates by 25 basis points in the first (light color) and second
(dark color) regime together with the 68% posterior highest density sets. We discuss the
effects of an expansionary monetary policy shock.
In the first regime, an unexpected monetary loosening leads to a persistent level shift
26
in the federal funds rate. The response falls to the lowest point after two years at 80 basis
points. This shock is effective in increasing inflation with the highest impact of 0.40 pp
after seven months. The effect on output growth is positive in the long run while the
response is not different from zero at short horizons. It suggests that in the short run
the inflationary push overrules potential direct effects of interest rate decreases on output
growth in rather calm and economically stable periods. The system features the liquidity
effect with the money supply substantially and permanently increasing. In line with Ang,
Boivin, Dong, and Loo-Kung (2011), the lower interest rates decrease the slope of the
bond yield curve for nearly a year, with the trough response narrowing the term spread
by 102 basis points after one month. They also lead to a sharp increase in stock prices.
In the second regime, a negative monetary policy shock decreases the federal funds
rate permanently but slightly more moderately than in the first regime. In this regime,
the monetary policy is strongly expansionary in the long run as output growth increases
permanently and after a re-bouncing effect by 0.54 pp within the second year. At the same
time, inflation’s response is negligible. Therefore, this regime observes an increase in the
importance of output relative to inflation in the monetary policy reaction function that
together with the timing of the second regime, is in line with the shift in the emphasis in
the monetary policy after 2000 described by Belongia and Ireland (2016).
In these times of crises, the liquidity effect is positive. The slope of the bond yield
curve is only slightly affected as the term spread widens by 32 basis points after six
months, strengthening the conclusion by Tillmann (2020) on time variation in this effect.
The response cools off gradually but remains significant at all horizons. The stock prices
contract by more than 5% on impact. This result closely follows the change over time of
the impulse responses reported by Galı́ and Gambetti (2015) that the authors explained
by the existence of rational bubbles in the stock market.
These effects are in line with the characterisation of the monetary policy in the US as
unconventional in the periods aligning with the second regime. The main objectives of
these policies were to stimulate the economy in the aftermaths of the crises and provide
liquidity, all of which was implemented without substantially increasing inflation in line
27
with “missing inflation“ during the 2010s (see Bobeica and Jarocinski, 2019), a period
which is covered by the second regime.
Our results indicate time-variation in the monetary policy shock identification.
Time-varying parameters and identification play a crucial role in the interpretability of
the shocks. It is not obtained if any of these features are omitted from the model
formulation.

4. Conclusions

We propose a new Bayesian structural vector autoregression which allows for

time-varying identification. Our model facilitates a data-driven search of regime-specific
exclusion restriction patterns. It features Markov-switching regimes in the structural
matrix and a standard deviation parameter as well as structural shock variances
following a stochastic volatility process. These components allow us to study
time-variation in structural shock identification, time-varying impulse responses, and
provide an easy way to check identification through heteroskedasticity within each
regime as a validity condition of regime-specific identification.
In an empirical illustration, we show that data strongly support time-variation in US
monetary policy shock identification. We find evidence that during crisis and periods
of unconventional monetary policy actions monetary policy shock identification is best
captured by exclusion restrictions which allow money and interest rate to co-move. Before
2000 and during normal times, no restriction on the term spread is the data supported
identification scheme for monetary policy shocks.

References

Andrés, J., J. D. López-Salido, and J. Vallés (2006, April). Money in an Estimated Business Cycle Model of
the Euro Area. Economic Journal 116(511), 457–477.
Ang, A., J. Boivin, S. Dong, and R. Loo-Kung (2011). Monetary Policy Shifts and the Term Structure. The
Review of Economic Studies 78(2), 429–457.
Arias, J. E., J. F. Rubio-Ramı́rez, M. Shin, and D. F. Waggoner (2024). Inference Based on Time-Varying
SVARs Identified with Sign Restrictions . Working Papers 24-05, Federal Reserve Bank of Philadelphia.
28
Arias, J. E., J. F. Rubio-Ramı́rez, and D. F. Waggoner (2018). Inference based on structural vector
autoregressions identified with sign and zero restrictions: Theory and applications. Econometrica 86(2),
685–720.
Bacchiocchi, E., E. Castelnuovo, and L. Fanelli (2017). Gimme a Break: Identification and Estimation of the
Macroeconomic Effects of Monetary Policy Shocks in the US. Macroeconomic Dynamics 22, 1613–1651.
Baumeister, C. and L. Benati (2013). Unconventional Monetary Policy and the Great Recession: Estimating
the Macroeconomic Effects of a Spread Compression at the Zero Lower Bound. International Journal of
Central Banking 9(2), 165–212.
Baumeister, C. and J. D. Hamilton (2018). Inference in structural vector autoregressions when the identifying
assumptions are not fully believed: Re-evaluating the role of monetary policy in economic fluctuations.
Journal of Monetary Economics 100, 48–65.
Belongia, M. T. and P. N. Ireland (2015). Interest Rates and Money in the Measurement of Monetary Policy.
Journal of Business & Economic Statistics 33(2), 255–269.
Belongia, M. T. and P. N. Ireland (2016). The evolution of U.S. monetary policy: 2000–2007. Journal of
Economic Dynamics and Control 73, 78–93.
Bobeica, E. and M. Jarocinski (2019). Missing Disinflation and Missing Inflation: A VAR Perspective.
International Journal of Central Banking 15(1), 199–232.
Chan, J. C. C. (2018). Specification tests for time-varying parameter models with stochastic volatility.
Econometric Reviews 37(8), 807–823.
Diebold, F. X., G. D. Rudebusch, and S. Boraǧan Aruoba (2006). The macroeconomy and the yield curve: A
dynamic latent factor approach. Journal of Econometrics 131(1-2), 309–338.
Doan, T., R. B. Litterman, and C. A. Sims (1984). Forecasting and Conditional Projection Using Realistic
Prior Distributions. Econometric Reviews 3(June 2013), 37–41.
Feldkircher, M. and F. Huber (2018). Unconventional U.S. Monetary Policy: New Tools, Same Channels?
Journal of Risk and Financial Management 11(4), 1–31.
Galı́, J. and L. Gambetti (2015). The Effects of Monetary Policy on Stock Market Bubbles: Some Evidence.
American Economic Journal: Macroeconomics 7(1), 233–257.
Geweke, J. (1996). Variable selection and model comparison in regression. In Bayesian Statistics 5.
Geweke, J. (2007). Interpretation and inference in mixture models: Simple MCMC works. Computational
Statistics & Data Analysis 51(7), 3529–3550.
Geweke, J. and G. Amisano (2010). Comparing and evaluating bayesian predictive distributions of asset
returns. International Journal of Forecasting 26(2), 216–230.
Giacomini, R., T. Kitagawa, and A. Volpicella (2022). Uncertain identification. Quantitative Economics 13(1),
95–123.

29
Hamilton, J. D. (1989). A new approach to the economic analysis of nonstationary time series and the
business cycle. Econometrica: Journal of the econometric society 57(2), 357–384.
Kastner, G. and S. Frühwirth-Schnatter (2014). Ancillarity-sufficiency interweaving strategy (asis) for
boosting mcmc estimation of stochastic volatility models. Computational Statistics & Data Analysis 76, 408
– 423.
Kimura, T. and J. Nakajima (2016). Identifying conventional and unconventional monetary policy shocks:
a latent threshold approach. The B.E. Journal of Macroeconomics 16(1), 277–300.
Leeper, E. and J. E. Roush (2003). Putting ”M” back in monetary policy. Journal of Money, Credit and
Banking 35, 1217–1256.
Leeper, E. M., C. A. Sims, T. Zha, R. E. Hall, and B. S. Bernanke (1996). What Does Monetary Policy Do?
Brookings Papers on Economic Activity 1996(2), 1–78.
Liu, P., H. Mumtaz, K. Theodoridis, and F. Zanetti (2019). Changing Macroeconomic Dynamics at the Zero
Lower Bound. Journal of Business & Economic Statistics 37, 391–404.
Lütkepohl, H. and T. Schlaak (2022). Heteroscedastic Proxy Vector Autoregressions. Journal of Business &
Economic Statistics 40(3), 1268–1281.
Lütkepohl, H., F. Shang, L. Uzeda, and T. Woźniak (2024). Partial identification of heteroskedastic structural
vars: Theory and bayesian inference. Working paper, University of Melbourne.
Lütkepohl, H. and T. Woźniak (2020). Bayesian inference for structural vector autoregressions identified by
Markov-switching heteroskedasticity. Journal of Economic Dynamics and Control 113, 103862.
Pagliari, M. S. (2024). Does one (unconventional) size fit all? Effects of the ECB’s unconventional monetary
policies on the euro area economies. European Economic Review, 104817.
Primiceri, G. E. (2005). Time Varying Structural Vector Autoregressions and Monetary Policy. The Review of
Economic Studies 72(3), 821–852.
Rubio-Ramı́rez, J. F., D. F. Waggoner, and T. Zha (2010). Structural Vector Autoregressions: Theory of
Identification and Algorithms for Inference. Review of Economic Studies 77(2), 665–696.
Sims, C. A. (1980). Macroeconomics and Reality. Econometrica 48(1), 1–48.
Sims, C. A. and T. Zha (2006). Were There Regime Switches in U.S. Monetary Policy? American Economic
Review 96(1), 54–81.
Tillmann, P. (2020). Monetary Policy Uncertainty and the Response of the Yield Curve to Policy Shocks.
Journal of Money, Credit and Banking 52(4), 803–833.
Vázquez, J., R. Marı́a-Dolores, and J.-M. Londono (2013). On the informational role of term structure in the
US monetary policy rule. Journal of Economic Dynamics and Control 37(9), 1852–1871.
Waggoner, D. F. and T. Zha (2003). A Gibbs sampler for structural vector autoregressions. Journal of Economic
Dynamics and Control 28(2), 349–366.

30
Wu, J. C. and F. D. Xia (2016). Measuring the macroeconomic impact of monetary policy at the zero lower
bound. Journal of Money, Credit and Banking 48(2-3), 253–291.

31
 Online Supplement:
Time-Varying Identification
of Structural Vector Autoregressions
Annika Camehla,∗, Tomasz Woźniakb
a
Erasmus University Rotterdam, Burgemeester Oudlaan 50, 3062 PA Rotterdam, The Netherlands
b
University of Melbourne, 111 Barry St., 3053 Carlton, VIC, Australia

Appendix A. Gibbs sampler

For simplicity, we write the VAR model in compact form

yt = Axt + εt (A1)

 
where the N × (Np + d)-matrix A = A1 . . . Ap Ad collects autoregressive and constant
 ′
parameters and the (Np + d)-vector xt = yt−1 . . . yt−p dt contains lagged endogenous
′ ′ ′

variables and constant terms.

Sampling parameters in the structural matrix

To estimate the parameters in the structural matrix, we use a Gibbs sampler. The full
conditional posterior distribution of the unrestricted elements in regime m, equation n,
and TVI component k is generalised-normal, proportional to

 
1 −1
| det (Bm.k ) |Tm exp − bn.m.k Vn.m.k ΩB.nm V′n.m.k b′n.m.k (A2)
2

where Bm.k is the matrix from regime m and TVI component k, with scale matrix

−1
X
ΩB.nm = γ−1
B.n IN + (yt − Axt )(yt − Axt )′ (A3)
t:st =m

∗
Corresponding author. Email address: [email protected].
1
where Tm is the number of observations in regime m. Within each regime, we sample the
structural matrix and the TVI indicators jointly.

1. Sample parameters of the structural matrix: B row-by-row following the algorithm

by Waggoner and Zha (2003) by sampling K vectors bn.m.k from (A2) given κ(m) = k
for all k ∈ {1, . . . , K}.
2. Sample TVI indicators: by computing the conditional posterior probabilities of
each of these TVI components, pn.m.k , proportional to the product of the likelihood
function, L(θ | YT ), and the prior distributions:

pn.m.k ∝ L(θ | YT , κ(m) = k)p Bn.m.k | γB , κ(m) = k p (κ(m) = k) . (A4)

Sample the TVI indicator from the multinomial distribution using the probabilities
pn.m.k for k = 1, . . . , K, and return the draw from the joint posterior (bn.m.k , k).
3. Sample structural hyper-parameters: γB.n , sB.n , and sγB from their respective full
conditional posterior distributions:
 
 X X 
γB.n | Bn.m.k , k, sB.n ∼ IG2 sB.n + bn.m.kn b′n.m.kn , νB + rn.m.kn  (A5)
m m
 
sB.n | γB , sγB ∼ G (s−1
γB
+ (2γ B.n )−1 −1
) , νγB
+ 0.5ν B
(A6)
 N

 X 
sγB | sB ∼ IG2 ssB + 2 sB.n , νsB + 2NνγB  (A7)
n=1

If TVI is not implemented in a particular row, sample bn.m bypassing Step 2 of the
algorithm.

Sampling of the remaining parameters

Estimation of the remaining is implemented by the Gibbs sampler. The rows of the
transition probabilities matrix, P, and the initial state probabilities vector, π0 , are all
sampled from independent M-variate Dirichlet full conditional posterior distributions as
in Frühwirth-Schnatter (2006). The sampling procedure for the Markov process
2
realisations follows the forward-filtering backward-sampling algorithm proposed by
Chib (1996).

Sampling autoregressive parameters

We follow Chan, Koop, and Yu (2024) and sample the autoregressive parameters of the A
matrix efficiently row-by-row, where the nth row is denoted by [A]n· . Denote by An=0 the A
matrix with its nth row set to zeros. To derive the sampler for the row of the autoregressive
matrix, rewrite the structural form in an equivalent form as:

Bm.k yt − An=0 xt = [Bm.k ]·n ⊗ x′t [A]′n· + εt (A8)

 
Define an n × 1 vector znt = Bm.k yt − An=0 xt and an N × (Np + d) matrix Znt = [Bm.k ]·n ⊗ x′t
and rewrite equation (A8) as:

znt = Znt [A]′n· + εt , (A9)

where the shocks follow a normal distribution. This results in the multivariate normal
full conditional posterior distribution for the rows of A:

 ′

[A]′n· | YT , [A]n=0 , Bm.k , σ 21 , . . . σ 2T , γA.n ∼ NNp+d ΩA.n mA.n , ΩA.n (A10)
−1
X  −1
−1
ΩA.n = γ−1 Ω
A.n A + Z n′
t diag σ 2t Znt (A11)
t
X  −1
′ −1 ′
mA.n = γ−1
A.n ΩA mA.n + Zn′
t diag σ 2t znt (A12)
t

3
Finally, the hyper-parameters of the autoregressive shrinkage hierarchical prior are
sampled from their respective full conditional posterior distributions:

    ′ 
γA.n | [A]n· , sA.n ∼ IG2 sA.n + [A]n· − mA.n Ω−1 A.n [A] n· − mA.n , νA + Np + d (A13)
 
sA.n | γA.n , sγA ∼ G (s−1
γA
+ (2γA.n )−1 )−1 , νγA + 0.5νA (A14)
 N

 X 
sγA | sA ∼ IG2 ssA + 2 sA.n , νsA + 2NνγA  (A15)
n=1

Sampling stochastic volatility parameters

Gibbs sampler is facilitated using the auxiliary mixture sampler proposed by Omori,
Chib, Shephard, and Nakajima (2007).
Specify the nth structural shock as:

 
1
un.t = exp ωn (st )hn.t zn.t , (A16)
2

where zn.t is a standard normal innovation. Transform this equation by squaring and
taking the logarithm of both sides obtaining:

un.t = ωn (st )hn.t + e

e zn.t , (A17)

un.t = log u2n.t and e

where e zn.t = log z2n.t . The distribution of e
zn.t is log χ21 . This non-standard
distribution is approximated precisely by a mixture of ten normal distributions defined
by Omori et al. (2007). This mixture of normals is specified by qn.t = 1, . . . , 10 – the mixture
component indicator for the nth equation at time t, the normal component probability
πqn.t , mean µqn.t , and variance σ2qn.t . The latter three parameters are fixed and given in
Omori et al. (2007), while qn.t augments the parameter space and is estimated. Its prior
distribution is multinomial with probabilities πqn.t .
 ′
Define T × 1 vectors: hn = hn.1 . . . hn.T collecting the log-volatilities,
 ′  ′
qn = qn.1 . . . qn.T collecting the realizations of qn.t for all t, µqn = µqn.1 . . . µqn.T , and

4
  ′
σ 2qn = σ2qn.1 . . . σ2qn.T collecting the nth equation auxiliary mixture means and variances,
 ′  ′
e
Un = e un.1 . . . eun.T , and ωn = ωn (s1 ) . . . ωn (sT ) collecting the volatility of the
volatility parameters according to their current time assignment based on the sampled
realizations of the Markov process. Whenever a subscript on these vectors is extended
by the Markov process’ regime indicator m it means that the vector contains only the Tm
e n.m . Finally, define a T × T matrix Hρn with ones
observations for t such that st = m, e.g., U
on the main diagonal, value −ρn on the first subdiagonal, and zeros elsewhere, and a
Tm × Tm matrix Hρn .m with rows and columns selected according to the Markov state
allocations such that st = m. Sampling latent volatilities hn proceeds independently for
each n from the following T-variate normal distribution parameterized following Chan
−1
and Jeliazkov (2009) in terms of its precision matrix Vhn and location vector hn as:

 
hn | YT , sn , qn , B, A, ωn , ρn ∼ NT Vhn hn , Vhn (A18)
−1    
Vhn = diag ω2n diag σ −2 ′
qn + Hρn Hρn (A19)
  
hn = diag (ωn ) diag σ −2
qn u
e n − µ qn
(A20)

−1
The precision matrix, Vhn , is tridiagonal, which greatly leads to a simulation smoother
proposed by McCausland, Miller, and Pelletier (2011).
The regime-dependent volatility of the volatility parameters are sampled
independently from the following normal distribution:

ωn (m) | Ym , sn.m , sn.m , hn.m , σ2ωn ∼ N vωn.m ωn.m , vωn.m (A21)
−1
 
vωn.m = h′n.m diag σ −2
qn .m hn.m + σωn
−2
(A22)
  
ωn = h′n.m diag σ −2
qn .m e un.m − µqn .m (A23)

The autoregressive parameters of the SV equations are sampled independently from

5
the truncated normal distribution using the algorithm proposed by Robert (1995):

 −1  T   T−1 −1 

X T−1
  X  X  

ρn | YT , hn ∼ N  h2n.t   hn.t hn.t−1  , 
 h2n.t   I |ρn | < 1 . (A24)
 
t=0 t=1 t=0

The prior variances of parameters ωn (m), σ2ωn , are a posteriori independent and sampled
from the following generalized inverse Gaussian full conditional posterior distribution:
 
 1 X 2 
σ2ωn | YT , ωn (1), . . . , ωn (M) ∼ GIG MA − , ωn (m), 
2
(A25)
2 m S

The auxiliary mixture indicators qn.t are each sampled independently from
a multinomial distribution with the probabilities proportional to the product of the prior
probabilities πqn.t and the conditional likelihood function.
Finally, proceed to the ancillarity-sufficiency interweaving sampler proposed by
Kastner and Frühwirth-Schnatter (2014). Our implementation proceeds as follows:
Having sampled random vector hn and parameters ωn (m), compute the parameters of
the centered parameterization h̃n.t = ωn (m)hn.t and σ2υn = ω2n (m). Then, sample σ2υn .m from
the generalised inverse Gaussian full conditional posterior distribution:

 
Tm − 1 ′
σ2υn .m | Ym , h̃n.m , σ2ωn ∼ GIG − , h̃n.m H′ρn .m Hρn .m h̃n.m , σ−2
ωn (A26)
2

using the algorithm introduced by Hörmann and Leydold (2014). Resample ρn using
the full conditional posterior distribution from equation (A24) where the vector hn is
q
replaced by h̃n . Finally, compute ωn (m) = ± σ2υn .m and hn.t = ωn1(m) h̃n.t and return them as
the MCMC draws for these parameters.

Appendix B. Data

Rt and TSt are expressed in annual terms, as is πt computed as the logarithmic rates
of returns on the consumer price index expressed in percent. yt is taken in log-difference
multiplied by 100. mt , and spt are taken in log-levels multiplied by 100. Series Rt , yt , and
6
mt are downloaded from Board of Governors of the Federal Reserve System (2023a,b,c)
respectively, πt is provided by Organization for Economic Co-operation and Development
(2023), TSt by Federal Reserve Bank of St. Louis (2023), and spt by yahoo finance (2023).
Additionally, we used a series of shadow rates by Wu and Xia (2016) downloaded from
https://sites.google.com/view/jingcynthiawu/shadow-rates [accessed 15.08.2023].

Appendix C. Forecasting performance 12 periods ahead

1.2
0

relative root−squared forecast error

relative predictive log score

1.1
−100000

1.0
0.9
−200000

0.8
−300000

0.7

2008 2010 2012 2014 2016 2018 2020 2022 2008 2010 2012 2014 2016 2018 2020 2022

forecast evaluation period forecast evaluation period

(a) Density forecast performance (b) Point forecast performance

Figure Appendix C.1: Twelve-step-ahead forecasting performance of selected models

Note: The legend for models includes: MS models (dark colour): M = 3 with TVI, M = 2 baseline,
M = 2 with TS, M = 2 with m, M = 2 only m, no MS (M = 1, bright colour): with TVI,
baseline, with TS, with m, only m. Measures are relative to the benchmark model with M = 2
and TVI.

While the performance at at twelve horizons supports the benchmark model, the
picture is more nuanced. The model with two regimes and TVI outperforms all
alternatives except one based on both measures. The exception is the model with two
regimes and restrictions allowing for contemporaneous movement of interest rates with
output, inflation, and term spreads in the third row of B (dark dotted line). This model,
however, performs weak on short horizon forecasts. Log scores across models show that
7
MS clearly improves forecasting while due to the higher uncertainty accompanying
longer horizon forecasts the extra flexibility of TVI leads to small losses in forecasting
accuracy. The evaluation based on point forecasts, however, provides evidence that
models with TVI improve forecasting performance.

Appendix D. Robustness

Appendix D.1. Time-varying identification

We show robustness on the posterior probabilities of the TVI components in Table

D.1, with probabilities reported for our main model, labeled benchmark, and eight
alternative models in rows for the first (upper panel) and second regime (lower panel)
for TVI in the third row of B (left part) and fourth row (right part). Selecting two
different identification patterns over time is remarkably robust with respect to extending
or limiting the set of possible exclusion restrictions in the search. First, broadening the
set of exclusion restrictions in the third row leads to the same posterior probabilities of
the TVI mechanism for the monetary policy shock identification – shown for a model
which extends the restriction patterns of the benchmark model by allowing for interest
rates to move simultaneously with stock prices, called with sp. Expanding the restriction
patterns in the forth row does not change the conclusion on the support of zero
exclusion restriction on interest rates in the forth row, see the model called exp 4eq which
also includes patterns with zero restrictions on money and stock prices (“LT”) or zero
restrictions on interest rates, money and stock prices (“LT0”).
Second, excluding the variant only m, does not change the posterior probability of
the remaining components, see model ex only m. Limiting the set of possible exclusion
restrictions in the third row to simultaneous movements of interest rates, output, and
inflation with term spread or money supports the role of the term spread and money in
the respective regimes, model called with TS, m.
Moreover, models with changes in the prior set-ups on the contemporaneous relations
by cutting the prior standard deviation of B in half or doubling the prior variance, models

8
 Table D.1: Posterior probabilities of the TVI components across alternative models

Regime 1
third row fourth row
base with TS with m with sp only m UR with 0 LT LT0
benchmark 0.00 1.00 0.00 0.00 1.00 0.00
plus sp 0.00 1.00 0.00 0.00 0.00 1.00 0.00
exp 4eq 0.00 1.00 0.00 0.00 0.49 0.00 0.51 0.00
ex only m 0.00 1.00 0.00 1.00 0.00
with TS, m 1.00 0.00 1.00 0.00
sd(B) /2 0.00 1.00 0.00 0.00 1.00 0.00
Var(B) x2 0.00 1.00 0.00 0.00 1.00 0.00
CPI, IP 0.00 1.00 0.00 0.00 1.00 0.00
GDPDEF 0.00 1.00 0.00 0.00 1.00 0.00
Regime 2
benchmark 0.27 0.27 0.45 0.01 0.51 0.49
plus sp 0.16 0.15 0.38 0.25 0.05 0.58 0.42
exp 4eq 0.28 0.26 0.44 0.02 0.55 0.44 0.00 0.00
ex only m 0.25 0.25 0.50 0.54 0.46
with TS, m 0.42 0.58 0.55 0.45
sd(B) /2 0.26 0.24 0.47 0.03 0.54 0.46
Var(B) x2 0.27 0.27 0.45 0.01 0.51 0.49
CPI, IP 0.26 0.25 0.49 0.01 0.52 0.48
GDPDEF 0.29 0.28 0.43 0.00 0.51 0.49

Note: Table shows the posterior probabilities of κn (m) for the identification of the monetary policy shock
for regime one (upper panel) and regime two (lower panel) for nine different models.

called sd(B) /2 and Var(B) x2, respectively, support time-variation in the identification of
the monetary policy shock. Also, changes in the variables (using the consumer price
index and industrial production index instead of the growth rates, model called CPI, IP,
or the interpolated GDP deflator instead of CPI inflation, model called GDPDEF) yield
similar TVI probabilities.

Appendix D.2. Identification through heteroskedasticity

We find that the identification of the third and fourth row through heteroskedasticity
is robust to changing the scale hyper-parameter of the prior controlling for the
regime-dependent standard deviation parameter level to 0.5 or 2, see models called prior

9
 prior sv 0.5 prior sv 0.5
regime 1 regime 1
3.0

3.0
prior sv 2 prior sv 2
CPI, IP regime 2 CPI, IP regime 2
GDPDEF GDPDEF
2.0

2.0
Density

Density
1.0

1.0
0.0

0.0
−1.5 −1.0 −0.5 0.0 0.5 1.0 1.5 −1.5 −1.0 −0.5 0.0 0.5 1.0 1.5

ω3(m) ω4(m)

(a) ω3 (b) ω4

Figure Appendix D.1: Marginal posterior densities of the standard deviations across alternative models

sv 0.5 and prior sv 2 in Figure Appendix D.1. Also, using alternative measurements for
our variables, such as the consumer price index, industrial production, or the
interpolated GDP deflator, has no impact on identification through heteroskedasticity,
see models called CPI, IP and GDPDEF.

References

Board of Governors of the Federal Reserve System (2023a). Federal Funds Effective Rate [DFF]. Data.
Retrieved from FRED database [accessed August 15, 2023].
Board of Governors of the Federal Reserve System (2023b). Industrial Production: Total Index [INDPRO].
Data. Retrieved from FRED database [accessed August 15, 2023].
Board of Governors of the Federal Reserve System (2023c). M2 [M2SL]. Data. Retrieved from FRED
database [accessed August 15, 2023].
Chan, J. and I. Jeliazkov (2009). Efficient Simulation and Integrated Likelihood Estimation in State Space
Models. International Journal of Mathematical Modelling and Numerical Optimisation 1, 101–120.
Chan, J. C. C., G. Koop, and X. Yu (2024). Large order-invariant bayesian vars with stochastic volatility.
Journal of Business & Economic Statistics 42(2), 825–837.
Chib, S. (1996). Calculating Posterior Distributions and Modal Estimates in Markov Mixture Models. Journal
of Econometrics 75(1), 79–7.
Federal Reserve Bank of St. Louis (2023). 10-Year Treasury Constant Maturity Minus Federal Funds Rate
[T10YFFM]. Data. Retrieved from FRED database [accessed August 15, 2023].
Frühwirth-Schnatter, S. (2006). Finite Mixture and Markov Switching Models. New York: Springer
Science+Business Media, LLC.
10
Hörmann, W. and J. Leydold (2014). Generating generalized inverse gaussian random variates. Statistics
and Computing 24(4), 547–557.
Kastner, G. and S. Frühwirth-Schnatter (2014). Ancillarity-sufficiency interweaving strategy (asis) for
boosting mcmc estimation of stochastic volatility models. Computational Statistics & Data Analysis 76, 408
– 423.
McCausland, W. J., S. Miller, and D. Pelletier (2011). Simulation smoothing for state–space models: A
computational efficiency analysis. Computational Statistics & Data Analysis 55(1), 199–212.
Omori, Y., S. Chib, N. Shephard, and J. Nakajima (2007). Stochastic Volatility with Leverage: Fast and
Efficient Likelihood Inference. Journal of Econometrics 140(2), 425–449.
Organization for Economic Co-operation and Development (2023). Main Economic Indicators: Consumer
Price Index: All Items: Total for United States [CPALTT01USM661S]. Data. Retrieved from FRED
database [accessed August 15, 2023].
Robert, C. P. (1995). Simulation of truncated normal variables. Statistics and computing 5(2), 121–125.
Waggoner, D. F. and T. Zha (2003). A Gibbs sampler for structural vector autoregressions. Journal of Economic
Dynamics and Control 28(2), 349–366.
Wu, J. C. and F. D. Xia (2016). Measuring the macroeconomic impact of monetary policy at the zero lower
bound. Journal of Money, Credit and Banking 48(2-3), 253–291.
yahoo finance (2023). S&P 500 ˆGSPC. Data. [accessed August 15, 2023].

11