Journal of Monetary Economics 144 (2024) 103548

Contents lists available at ScienceDirect

Journal of Monetary Economics

journal homepage: www.elsevier.com/locate/jme

Estimating the Fed’s unconventional policy shocks✩

Marek Jarociński
European Central Bank DG-Research, Frankfurt am Main, 60640, Germany

ARTICLE INFO ABSTRACT

JEL classification: Financial market responses to Fed monetary policy announcements are often very small, but
E52 sometimes very large and the mix of news contained in these announcements varies over time. I
E58 exploit these features of the data to estimate different types of Fed policy shocks. The resulting
E44
shocks can be naturally labeled as standard monetary policy, Odyssean forward guidance, large
Keywords: scale asset purchases and Delphic forward guidance. They affect risk-free interest rates, stock
High-frequency identification prices and the dollar on impact and have delayed but pronounced effects on corporate bond
Non-Gaussianity
spreads and breakeven inflation rates.
Fat tails
Forward guidance
Asset purchases

1. Introduction

Modern central banks deploy a variety of policies in their efforts to steer the economy and measuring the effects of these different
policies is of paramount importance for monetary economics. Since Kuttner (2001), many papers identify monetary policy shocks
from the changes of financial asset prices in a narrow time window around Federal Open Market Committee (FOMC) announcements.
Prior to the announcement, asset prices reflect the consensus view on the state of the economy and the Fed’s expected response to
it. Afterwards, asset prices incorporate also any unexpected news conveyed in the announcement. These news could be about the
current fed funds rate or its future path, asset purchases, the Fed’s view on the state of the economy, etc. They represent different
structural shocks that may affect the economy differently, so it is crucial to disentangle their effects.
This paper estimates the structural shocks that underlie the financial market reactions to FOMC announcements. While the nature
of the shocks is not specified ex ante, ex post the estimated shocks can be naturally labeled as the current fed funds rate policy,
an ‘‘Odyssean’’ forward guidance (a commitment to a future course of policy rates), a large scale asset purchase and a ‘‘Delphic’’
forward guidance (a statement about the future course of policy rates understood as a forecast of the appropriate stance of the policy
rather than a commitment, see Campbell et al., 2012).
To estimate the structural shocks I exploit a striking, yet hitherto neglected feature of the data. Namely, the high-frequency
reactions of financial variables, such as interest rates and stock prices, to FOMC announcements are usually very small, but sometimes
very large, i.e. they have very fat tails, or excess kurtosis. This feature implies that these data may contain information about the
nature of the underlying structural shocks. Given the importance of the Fed policies, it is vital to exploit this available information as
well as possible. Previous literature on high-frequency reactions of financial variables to central bank announcements has ignored

✩ The opinions in this paper are those of the author and do not necessarily reflect the views of the European Central Bank. I thank anonymous referees, Jesús
Fernández-Villaverde, Michael Johannes, Peter Karadi, Burçin Kısacıkoğlu, Ulrich Müller, Giorgio Primiceri and numerous seminar and conference participants
for useful comments.
E-mail address: [email protected].
URL: https://marekjarocinski.github.io.

https://doi.org/10.1016/j.jmoneco.2024.01.001
Received 21 June 2023; Received in revised form 5 January 2024; Accepted 7 January 2024
Available online 9 January 2024
0304-3932/© 2024 Elsevier B.V. All rights reserved.
M. Jarociński Journal of Monetary Economics 144 (2024) 103548

it, treating the shocks explicitly or implicitly as Gaussian. This paper is, to my knowledge, the first attempt to tap this valuable
source of information.
Intuitively, fat-tailed shocks can be identified from the data because they tend to produce informative case studies. When we see
a significant market reaction to an FOMC announcement, and the underlying shocks are independent and fat-tailed, there is a high
chance that only a subset of the shocks is driving this reaction, while the others are very small. This greatly facilitates detecting
the unique patterns in the data corresponding to individual shocks. As a result, the shocks are identifiable from the data via the
likelihood function alone, even in the absence of economic identifying restrictions. One contribution of this paper is to provide
the intuition of the identification based on fat tails using a simple supply and demand example. The example also illustrates how
identification weakens, and possibly vanishes, when the independence assumption is relaxed, which is potentially relevant in the
case of disentangling Fed policies.
It is a separate question why the Fed shocks’ reflections in financial data are fat tailed. One reason could be that the Fed generally
avoids surprising financial markets, until occasionally it is forced to do it big time. Another reason could be that investors process
information imperfectly and focus only on the most salient dimensions (Van Nieuwerburgh and Veldkamp, 2010).
My baseline model expresses the surprises (i.e., the high-frequency reactions to FOMC announcements) in the near-term fed funds
futures, 2- and 10-year Treasury yield and the S&P500 stock index as linear combinations of four Student-t distributed shocks. It
turns out that these four shocks are very precisely estimated and ex post have natural economic interpretations. The first shock raises
the near-term fed funds futures, with a diminishing effect on longer maturities, and depresses the stock prices. It can be naturally
labeled as the standard monetary policy shock. The second shock increases the 2-year Treasury yield the most and depresses the
stock prices. It can be naturally labeled as the (Odyssean) forward guidance shock. The third shock increases the 10-year Treasury
yield the most and plays a large role in some of the most important asset purchase announcements. It can be naturally labeled
as the asset purchase shock. The fourth shock has a similar impact on the yield curve as the Odyssean forward guidance shock,
but triggers an increase, rather than a decrease, in the stock prices. Therefore, this shock matches the concept of Delphic forward
guidance introduced by Campbell et al. (2012). I find very similar four shocks when repeating the estimation on the principal
components of a larger dataset and under a variety of other modifications of the baseline model.
The findings of this paper are relevant for the ongoing research on the effectiveness of non-standard monetary policies. I track
the effects of the estimated shocks using daily local projections. The shocks gradually propagate through the financial system and
after a few days get reflected in the corporate bond spreads and breakeven inflation rates. Also the Delphic forward guidance shocks
have significant and persistent effects on financial variables and contribute to the historical narrative of Fed policies. One of the
largest Delphic shocks occurs in August 2011, when the Fed stated that exceptionally low interest rates will be warranted at least
through mid-2013, triggering pessimism about the economy.
It is important that the results are robust to relaxing the assumption that the structural shocks are independent. If different Fed
shocks tend to be large simultaneously (e.g. if they have some common stochastic volatility), the identification from fat-tails gets
diluted and eventually vanishes (e.g. Montiel Olea et al., 2022). However, even accounting for this possibility, I still find enough
independent variation to yield tight identification and virtually the same estimated shocks.
Previous research has used a variety of approaches and assumptions to decompose the high-frequency financial market reactions
into economically interpretable components (see Gürkaynak et al., 2005; Cieślak and Schrimpf, 2019; Jarociński and Karadi, 2020;
Swanson, 2021; Inoue and Rossi, 2021; Miranda-Agrippino and Ricco, 2021; Bauer and Swanson, 2023; Lewis, 2023, and others).
Most of these papers ignore the non-Gaussianity in the data and construct the shocks with the a priori assumed features. They
often use identifying restrictions familiar from the Structural VAR literature. For example, Gürkaynak et al. (2005) separate the
target factor (standard monetary policy) from the path factor (forward guidance) imposing a zero restriction on the response of short
term rates to forward guidance. Swanson (2021) imposes a narrative restriction that the asset purchase shock is small prior to the
Zero Lower Bound (ZLB) period. Jarociński and Karadi (2020) separate monetary policy (a summary of standard and non-standard
policies) from information (Delphic) shocks using sign restrictions. It is very interesting that, although I do not impose any of these
restrictions, the shocks I estimate satisfy them (sometimes up to a numerical approximation). They are also highly correlated with
their counterparts in these papers. Thus, my approach provides a statistical validation of the assumptions imposed in these papers.
That said, I refine these papers’ interpretations of the data by distinguishing four main shocks, while they identify at most three.
Identification through non-Gaussianity, such as the excess kurtosis exploited here, has been known since the 1990s but
economic applications have started to appear only recently. This source of identification underlies the Independent Components
Analysis (ICA) (Comon, 1994; Hyvärinen et al., 2001), which is widely used in signal processing, telecommunications and medical
imaging. Bonhomme and Robin (2009) use ICA to identify factor loadings. Methodologically closest paper to the present one
is Lanne et al. (2017) who identify structural VARs (SVARs) with Student-t shocks. Gouriéroux et al. (2017) extend the inference
on SVARs to pseudo-maximum likelihood. Gouriéroux et al. (2020) show that also the Structural Vector Autoregression Moving
Average (SVARMA) model is identified under shock non-Gaussianity. Fiorentini and Sentana (2020) study the effects of distributional
misspecification and identify an SVAR of volatility indices. Drautzburg and Wright (2021) use non-Gaussianity to strengthen the
identification in sign-restricted in VARs. Braun (2023) applies identification through non-Gaussianity to the oil market, Anttonen
et al. (2023) use it to study fiscal multipliers. Davis and Ng (2022) provide econometric theory for VARs with disaster-type shocks
and apply it to economic uncertainty and Covid shocks.
Several recent papers exploit the non-Gaussianity of financial variables for the identification of macroeconomic SVARs that
include a monetary policy shock. Maxand (2020) estimates an SVAR for the US that includes the fed funds rate and stock prices,
and finds a non-Gaussian monetary policy shock. Lanne and Luoto (2020) and Anttonen et al. (2021) study the effects of the US
monetary policy shock in the SVAR of Uhlig (2005) generalized to allow for non-Gaussian shocks and find that the monetary policy

2
M. Jarociński Journal of Monetary Economics 144 (2024) 103548

shock has very fat tails. Lanne and Luoto (2021) estimate an illustrative 3-variable non-Gaussian SVAR including the fed funds rate,
with a highly non-Gaussian monetary policy shock. Andrade et al. (2023) document that many proxies of monetary policy shocks
studied in the earlier literature exhibit fat tails and use this feature to sharpen the identification in their non-Gaussian SVARs. These
papers work with financial variables at the monthly frequency and identify a single monetary policy shock. The present paper is
different in that it works with the high-frequency reactions of financial variables to FOMC announcements and uses non-Gaussianity
to distinguish between different types of Fed policies.
There are analogies between identification by non-Gaussianity and identification by heteroskedasticity (Rigobon, 2003). Both
approaches are examples of a statistical identification exploiting that the shocks arrive ‘‘irregularly’’. For some recent applications of
identification by heteroskedasticity see e.g. Lewis (2023, 2022, 2021), Brunnermeier et al. (2021), Miescu (2021). In particular, Lewis
(2023) also identifies the effects of the Fed policies from high-frequency financial data, and, remarkably, finds similar four
dimensions of monetary policy. This is notable, because his approach is very different from the present paper. Lewis exploits the
intraday time variation of the asset price volatility on the days of FOMC announcements. On each of these days he fits a separate time
series model and performs a separate identification. By contrast, here each FOMC announcement contributes only one observation
and I rely on contrasting financial market reactions across the announcements.
The rest of the paper is structured as follows. Section 2 presents the data, highlighting their excess kurtosis. Section 3 lays out the
baseline econometric model and explains the identification with a simple example. Section 4 reports the estimation results for the
baseline model. Section 5 summarizes the lessons from alternative estimations reported in detail in the Online Appendix. Section 6
tracks the longer term effects of the shocks using daily local projections. Section 7 concludes.

2. Data

The data on high-frequency financial market reactions to FOMC announcements come from the widely-used dataset of Gürkaynak
et al. (2005) (Gürkaynak et al., 2005 from now on) updated by Gürkaynak et al. (2022). This dataset contains the changes of financial
variables in a 30-minute window around FOMC announcements (from 10 min before to 20 min after the announcement). The sample
studied here contains 241 FOMC announcements from 5 July 1991 to 19 June 2019.
In the baseline analysis I consider a vector of four variables (later I also extract factors from a larger set of variables). I refer
to the variables using their well-known (Gürkaynak et al., 2005) database identifiers. MP1, or the first fed funds future adjusted
for the number of the remaining days of the month (see Gürkaynak et al., 2005 for details) is the expected fed funds rate after the
FOMC announcement. ONRUN2 and ONRUN10 are the 2- and 10-year Treasury yields. Finally, SP500 is the Standard and Poors
500 blue chip stock index.
The choice of MP1, ONRUN2 and ONRUN10 follows Swanson (2021), who finds that these three variables approximately span
the target, path and LSAP factors that he constructs. I add the SP500 in order to capture the effects beyond the yield curve.
The responses of the four baseline variables to FOMC announcements are very non-Gaussian. Fig. 1 reports, for each variable, the
histogram, a Gaussian density and a Student-t density each fitted into the data by maximum likelihood. We can clearly see that the
Gaussian densities, plotted in red, fit the histograms poorly. First, the Gaussian distributions predict too few near-zero observations.
Second, the observed 4-, 6- and even 8-standard deviation outliers are unlikely under the Gaussian distribution. The fitted Student-t
densities, which agree with the histograms quite well, have very low shape parameters (𝑣 = 0.6, 1.7, 2.4, 2.3, respectively) implying
very large departures from Gaussianity.
If we think of these variables as being linear combinations of Fed policy shocks, their non-Gaussianity implies that at least one,
and possibly more of these shocks are non-Gaussian.

3. The econometric model and identification

Throughout this paper I assume that market responses to FOMC announcements are generated by the following simple model
driven by potentially fat-tailed shocks:

𝑦𝑡 = 𝐶 ′ 𝑢𝑡 , 𝑢𝑡 ∼ i.i.d.𝑝(𝑢𝑡 ). (1)

𝑦𝑡 = (𝑦1,𝑡 , … , 𝑦𝑁,𝑡 )′ is a vector of 𝑁 variables observed at time 𝑡. 𝑢𝑡 = (𝑢1,𝑡 , … , 𝑢𝑁,𝑡 )′ is a vector of unobserved, structural
(i.e. uncorrelated) shocks coming from a density 𝑝(𝑢𝑡 ) which may exhibit fat tails. 𝐶 is an 𝑁 × 𝑁 matrix whose 𝑖, 𝑗-th element
𝐶(𝑖, 𝑗) contains the effect of shock 𝑖 on variable 𝑗.1

3.1. The intuition behind the identification

The purpose of this section is to provide a simple illustration how structural relationships get revealed in the data in the presence
of fat tails (excess kurtosis) and a sufficient degree of independence. For formal proofs that non-Gaussianity (of a more general form)
of all but one shocks implies identification see e.g. Lanne et al. (2017), Proposition 2, or the discussion in Sims (2021).2

1 Eq. (1) is a special case of a Structural VAR with no lags of 𝑦 . The lack of any lags in Eq. (1) is consistent with the standard assumption in the high-frequency
𝑡
monetary policy surprise literature that those surprises should not be predictable with any information that predates the surprise.
2 Departures from Gaussianity other than fat tails also yield identification and do not necessarily require independence. See Bonhomme and Robin (2009),

Anttonen et al. (2023) for conditions under which skewed shocks are identifiable and Mesters and Zwiernik (2022) for other minimal conditions for identifiability.

3
M. Jarociński Journal of Monetary Economics 144 (2024) 103548

Fig. 1. The empirical distributions of the baseline variables.

Note. Each plot contains the histogram of the data, the Gaussian density and the Student-t density each fitted into the data by maximum likelihood. The
histograms are scaled so that they integrate to 1.

For a simple illustration, consider a market for good A. Market prices 𝑃 and quantities 𝑄 are determined by demand and supply,
each subject to shocks. 𝛥𝑃 and 𝛥𝑄 are the innovations in 𝑃 and 𝑄 in response to shocks. Can we identify the slopes of the demand
and supply curves from the data on 𝛥𝑃 and 𝛥𝑄?
Consider two structural models. In Model 1 the demand schedule is flat and the supply schedule is steep, while in Model 2 it is
the reverse. Models 1 and 2 satisfy Eq. (1) with coefficients 𝐶1 and 𝐶2 respectively,
( ) ( 𝑠) ( ) ( )
𝛥𝑄 ′ 𝑢 ′ 0.94 0.33 ′ 0.14 0.99
= 𝐶𝑖∈{1,2} , with 𝐶 = , 𝐶 = , (2)
𝛥𝑃 𝑢𝑑 1 −0.14 0.99 2 −0.94 0.33
where 𝑢𝑠 is a supply shock and 𝑢𝑑 a demand shock. In Model 1 a unit supply shock 𝑢𝑠 increases the quantity supplied by 0.94 while
the market price falls by 0.14, revealing a flat demand curve with the slope of −0.14∕0.94 ≈ −0.15. The slope of the supply curve is
0.99∕0.33 = 3. In Model 2 the slopes are −6.7 and 0.33 respectively. Panels A and B of Fig. 2 plot these demand and supply curves.
When the shocks 𝑢𝑠 and 𝑢𝑑 are Gaussian, we cannot identify the slopes from the data on 𝛥𝑃 and 𝛥𝑄. The second row of Fig. 2
presents the combinations of 𝛥𝑃 and 𝛥𝑄 obtained from Model 1 in panel C and from Model 2 in panel D, when the shocks 𝑢𝑠 and
𝑢𝑑 are drawn from independent Gaussian distributions with mean 0 and variance 1. In this example 𝐶1 × 𝐶1′ = 𝐶2 × 𝐶2′ = ( 0.2 1 0.2 ).
1
Consequently, in both cases (𝛥𝑃 , 𝛥𝑄) are Gaussian with the same first two moments, (0, 0) and ( 0.2 1 0.2 ), so the samples look the
1
same.
However, when shocks 𝑢𝑠 and 𝑢𝑑 are independent Student-t, the situation changes. Now Models 1 and 2 produce systematically
different combinations of 𝛥𝑃 and 𝛥𝑄. This is illustrated in the third row of Fig. 2. The samples in the third row are generated from
(2) but this time shocks 𝑢𝑠 and 𝑢𝑑 are drawn from independent Student-t distributions with mean 0 and shape parameter 𝑣 = 1.5. For
comparability with the previous example, the drawn shocks are re-scaled to ensure that their sample variance is 1. Hence, (𝛥𝑃 , 𝛥𝑄)
continue to have the same first two sample moments, (0, 0) and ( 0.2 1 0.2 ). Nevertheless, the samples in panels E and F look very
1
differently from each other and even an observer lacking any statistical training will have no problem matching each sample with
the correct structural model.
What helps here is the high kurtosis of the Student-t distribution, i.e. the fact that the shocks are often tiny, but sometimes large.
For an outlying observation, chances are that only one of the shocks was large, while the other was tiny. Hence, these observations
cluster around the demand and supply schedules, revealing their slopes. Obviously, if we can identify the structural model visually,
we can also do it numerically by evaluating the likelihood function (Online Appendix A presents an example).

4
M. Jarociński Journal of Monetary Economics 144 (2024) 103548

Fig. 2. Stylized example: demand and supply of good A.

Note. Each scatter plot has 1000 observations. The samples in the left column are generated from (2) with coefficients 𝐶1 , and the samples in the right column
are generated with coefficients 𝐶2 . In panels C and D the shocks 𝑢𝑑 and 𝑢𝑠 are independent Gaussian with mean 0 and variance 1. In panels E and F the shocks
𝑢𝑑 and 𝑢𝑠 are independent Student-t with mean 0, scale parameter 1 and shape parameter 𝑣 = 1.5. In panels G and H the shocks 𝑢𝑑 and 𝑢𝑠 are drawn from the
Partially Dependent Multivariate t distribution defined in Online Appendix G, with mean 0 and parameters 𝑣0 = 0.85 and 𝑣̄ = 0.65. In panels I and J the shocks
are drawn from the Multivariate Student-t with mean 0, scale parameter identity matrix and shape parameter 𝑣 = 1.5. Before feeding to the model, the shocks
are re-scaled so that their sample standard deviations equal 1. In all scatter plots, 𝛥𝑃 and 𝛥𝑄 have sample mean zero, sample variance 1 and sample correlation
0.2.

5
M. Jarociński Journal of Monetary Economics 144 (2024) 103548

Fig. 2. (continued).

Independence of the shocks helps but is not necessary for identification. In panels G and H the shocks are no longer independent.
In particular, they continue to be marginally Student-t and orthogonal, but when one of them is large in absolute value, the other
one is also more likely to be large in absolute value. This blurs the picture compared with the case of independence, because there
are more cases when both shocks are large, producing outliers that lay far away from either of the curves. Nevertheless, we can
still distinguish the models. Only in the limiting case of extreme dependence, shown in panels I and J, identification vanishes (the
shocks here come from a 2-dimensional Multivariate Student-t distribution). In the presence of less extreme dependence the models
continue to be identifiable.

3.2. Estimation

In the rest of the paper model (1) is taken to the data. For a sample of 𝑇 observations the model can be written as
𝑌 = 𝑈 𝐶, (3)
𝑦′𝑡
where 𝑌 is the 𝑇 × 𝑁 matrix with in row 𝑡 and 𝑈 is the corresponding 𝑇 × 𝑁 matrix of structural shocks.
It is convenient to reparameterize the model in terms of 𝑊 = 𝐶 −1 , so that we can write 𝑌 𝑊 = 𝑈 .

3.2.1. The baseline model

∏
The baseline model assumes that the shocks in period 𝑡 are independent Student-t, i.e. 𝑝(𝑢𝑡 ) = 𝑁 𝑛=1 𝑝(𝑢𝑛,𝑡 ) and shock 𝑢𝑛,𝑡 comes
from a Student-t density with 𝑣𝑛 degrees of freedom, denoted  (𝑣𝑛 ), with the probability density function
𝑣𝑛 +1
𝑝(𝑢𝑛,𝑡 ) = 𝑐(𝑣𝑛 ) (1 + 𝑢2𝑛,𝑡 ∕𝑣𝑛 )− 2 , (4)
( )−1
−1∕2 𝑣
where 𝑐(𝑣𝑛 ) = 𝑣𝑛 𝐵 21 , 2𝑛 is the integrating constant and 𝐵(⋅, ⋅) is the beta function. As 𝑣𝑛 → ∞ the Student-t distribution
converges to the Gaussian distribution, so this model accommodates the case where some or all shocks are Gaussian. Let 𝑣 denote
the length-𝑁 vector 𝑣 = (𝑣1 , … , 𝑣𝑁 ).

6
M. Jarociński Journal of Monetary Economics 144 (2024) 103548

The log-likelihood of sample 𝑌 is

∑𝑇 ∑ 𝑁
𝑣𝑛 + 1 ∑𝑁
log 𝑝(𝑌 |𝑊 , 𝑣) = 𝑇 log | det 𝑊 | − log(1 + 𝑢2𝑛,𝑡 ∕𝑣𝑛 ) + 𝑇 log 𝑐(𝑣𝑛 ), (5)
𝑡=1 𝑛=1
2 𝑛=1

where 𝑢𝑛,𝑡 = 𝑦′𝑡 𝑤𝑛 , with 𝑤𝑛 the 𝑛th column of 𝑊 .

I start by maximizing the likelihood (5) to obtain point estimates 𝑊̂ and 𝑣̂ = (𝑣̂ 1 , … , 𝑣̂ 𝑁 ), the structural shocks 𝑈̂ = 𝑌 𝑊̂ and the
impact matrix 𝐶̂ = 𝑊̂ −1 . Online Appendix B provides the analytical expression for the gradient of the likelihood which improves
the speed and precision of the maximization.
To cross-check these results and to assess the estimation uncertainty I estimate the model with the Bayesian approach using a
noninformative prior for 𝑊 (the prior for 𝑣 needs to be proper for technical reasons, see Bauwens and Lubrano 1998, but I use a
conservative prior that mildly works against identification). I use Bayesian simulation (Metropolis–Hastings algorithm) to study the
shape of the likelihood function more systematically and to make the inference about nonlinear functions of 𝑊 , such as 𝐶, as precise
as possible, instead of relying on asymptotic approximations. Online Appendix B provides the details on the Bayesian estimation.

3.2.2. Relaxing independence: the PDMT model

In a robustness check I relax the independence assumption and assume that the shocks come from the new Partially Dependent
Multivariate t-distribution (PDMT) defined in Online Appendix G. In this joint distribution each shock 𝑢𝑡𝑛 is marginally Student-t
and uncorrelated with the other shocks, but they are not independent because their large absolute values tend to occur at the same
time. The extent to which this happens is governed by the estimated parameters of the distribution.

3.3. The problem of shock permutation

One peculiarity of model (1) is that it is only identified up to permutation and signs of the shocks: permuting the shocks
(𝑁! possibilities) and flipping their signs (2𝑁 possibilities) does not change the value of the likelihood function. Therefore, if the
maximization of the likelihood detects a local mode 𝑊̂ , 𝑣,̂ we know that there are also 𝑁! × 2𝑁 − 1 other local modes with the same
value of the likelihood function. A Bayesian posterior simulator may visit the neighborhoods of different modes, and for a given
draw of 𝑊 one cannot be sure to what signs and ordering of the shocks it corresponds.
This paper proposes a practical approach for normalizing the draws, i.e. mapping them into the same signs and ordering of
the shocks. It takes one local mode 𝑊̂ as the reference point. Then, for each draw it finds shock signs and ordering that has the
highest probability under the Gaussian approximation of the likelihood function around 𝑊̂ . Online Appendix C provides the details and
an illustration.
In the literature the dominant solution of the permutation and sign switching problem is to use a non-exchangeable prior on the
columns of 𝑊 (e.g. Brunnermeier et al., 2021). The advantage of this paper’s approach is that it does not require an informative
prior on 𝑊 . Its disadvantage is that it relies on the local Gaussian approximation of the shape of the likelihood and might not
perform well when this approximation is poor.

4. Estimation results for the baseline model

I define 𝑦𝑡 =(MP1, ONRUN2, ONRUN10, SP500) and estimate model (1) with independent Student-t shocks (4) by maximum
likelihood, obtaining the point estimates 𝑊̂ , 𝑣,
̂ 𝑈̂ = 𝑌 𝑊̂ , 𝐶̂ = 𝑊̂ −1 . Next, I simulate the Bayesian posterior under a flat prior for 𝑊
and store 2000 approximately independent draws. Either approach suggests that the model is well identified, with four very fat-tailed
shocks and modes of 𝑊 corresponding to alternative signs and permutations well separated by regions of very low likelihood.

4.1. The impact effects of the four baseline shocks

̂ together with the 95% Bayesian posterior probability ranges.

Fig. 3 reports the impacts of the shocks implied by the estimated 𝐶,
To facilitate interpretation and comparisons with the literature, Fig. 3 reports the effects of one-sample-standard-deviation shocks.
That is, each row in this figure represents a row of 𝐶 scaled by the sample standard deviation of the corresponding column of 𝑈̂ .3
Furthermore, the responses of interest rates are arranged into a yield curve, with the maturities on the 𝑥-axis.
The shocks reported in Fig. 3 are tightly estimated and have intuitive economic interpretations.
𝑢1 looks like a standard contractionary monetary policy shock. The fed funds rate increases by 7.5 basis points and other interest
rates follow, with a weaker effect for longer maturities. The 2-year Treasury yield increases by almost 4 basis points and the 10-year
Treasury yield by about 1.6 basis points. The SP500 index drops by 26 basis points.
𝑢2 looks like the effect of forward guidance. The fed funds rate does not change in the near term, but the 2-year yield increases
by more than 4 basis points and the 10-year yield by 3.4 basis points in the half-hour window around the FOMC announcement.
This shock is very contractionary and the SP500 drops by 43 basis points.

3 Parameter 𝐶 gives the responses of 𝑌 to the shocks 𝑈 that have different scales determined by the properties of  (𝑣𝑛 ). The population standard deviations
√ 𝑣
of the shocks are 𝑛
𝑣 −2
for 𝑣𝑛 > 2 and infinite or undefined otherwise, but one can always compute the sample standard deviation of 𝑈̂ .
𝑛

7
M. Jarociński Journal of Monetary Economics 144 (2024) 103548

Fig. 3. Responses of the variables to standardized shocks, 95% bands.

𝑢3 mostly affects the 10-year yield, while having little effect on anything else, except the 2-year rate, which falls a little. However,
I show later that in the second half of the sample this shock has a significant negative impact on the stock market and a positive
impact on the 2-year rate. Furthermore, its large realizations coincide with important announcements of asset purchase policies,
which justifies calling it an LSAP shock.4
Finally, 𝑢4 moves the yield curve similarly as the forward guidance shock 𝑢2 , only is about two-thirds of the size. However,
by contrast to 𝑢2 , this shock is accompanied by an increase in the SP500 index by 35 basis points, which can be rationalized by
the presence of the Fed information effect. In particular, this shock perfectly matches the notion of the Delphic forward guidance
of Campbell et al. (2012).
The effects of 𝑢1 and 𝑢2 on MP1 and Treasury yields are very similar to the effects of the target factor and path factor of Gürkaynak
et al. (2005) and Swanson (2021) (see e.g. Swanson’s Table 3). This is in spite of the fact that I do not impose any of their identifying
restrictions. Furthermore, the estimation uncertainty is very small. We can conclude that the maximum likelihood estimation that
exploits the kurtosis of the data validates these earlier studies and their assumptions.
Another important lesson is that Fed information effects matter, as witnessed by the nontrivial role of 𝑢4 , and they manifest
themselves as the Delphic forward guidance. The theoretical models of Melosi (2017) and Nakamura and Steinsson (2018) focus on
the information effects that accompany current fed fund rate changes, but this paper’s agnostic estimation picks up the information
effects in the forward guidance.

4.2. The estimated shocks: the history

Fig. 4 reports the history of the shocks over time. To facilitate interpretation, in this figure the shocks are rescaled so that a
unit 𝑢1 shock raises the MP1 by 1 basis point, a unit 𝑢2 and 𝑢4 raises the ONRUN2 by 1 basis point, and a unit 𝑢3 shock raises the
ONRUN10 by 1 basis point.5 The top panel of Fig. 4 shows the pre-ZLB period 1991–2008 and the bottom panel the remaining
period 2009–2019. Vertical bars highlight many of the same events as Gürkaynak et al. (2005) and Swanson (2021). (For reference,
Online Appendix J provides the responses of the variables 𝑦𝑡 to each of these events.)
The history of the standard monetary policy shock 𝑢1 agrees with the accepted accounts. This is not surprising because 𝑢1 is
essentially equal to MP1 (the rank correlation of 𝑢1 with MP1 is 0.99). It is also highly correlated with the Gürkaynak et al. (2005)

4 Swanson (2021) also finds that his LSAP shock has an insignificant effect on the stock prices in the full sample.
5 To achieve that, 𝑢̂ 1 is multiplied by its effect on MP1, given by 𝑐̂1,1 , 𝑢̂ 2 and 𝑢̂ 4 are multiplied by their effects on ONRUN2, given by 𝑐̂2,2 and 𝑐̂4,2 , respectively,
and 𝑢̂ 3 is multiplied by its effect on ONRUN10, given by 𝑐̂3,3 .

8
M. Jarociński Journal of Monetary Economics 144 (2024) 103548

Fig. 4. The estimated shocks over time. Note. The shocks are re-scaled as explained in footnote 5. IM denotes an ‘‘inter-meeting’’ announcement.

target factor/(Swanson, 2021) fed funds rate shock (rank correlation 0.76, linear correlation 0.95, see Table 1). In the 1991–2008
plot we can see that, as is frequently noted, the largest realizations of standard policy shocks occur at inter-meeting announcements
(labeled ‘‘IM’’ in the plot). Unsurprisingly, in the Zero Lower Bound period the standard monetary policy shocks are negligible.
The Student-t model interprets some of the forward guidance episodes as Odyssean, 𝑢2 and some as Delphic, 𝑢4 , or the mix
of both. Table 1 reports that the forward guidance shock of Swanson (2021) is highly positively correlated with both 𝑢2 and 𝑢4
(rank correlations of 0.74 and 0.48 respectively). The 1991–2008 plot in Fig. 4 highlights the dates of the ten forward guidance
episodes discussed in Gürkaynak et al. (2005) (their Table 4, ‘‘Ten Largest Observations of the Path Factor’’). They are labeled
with the key word of the FOMC statement or a one-word description of its message. The Odyssean forward guidance, 𝑢2 dominates
the announcements marked ‘overshooting’ (December 1994, markets expect future tightening after Blinder’s recent comments of
‘overshooting’), ‘unsettled’ (October 1998), ‘tightening’ (May and October 1999) and ‘drop considerable’ (January 2004, dropping
of the commitment to a ‘considerable period’ of the same policy). The Delphic forward guidance 𝑢4 dominates the episodes labeled
‘Jan3,2001’ and ‘weakness’ (August 2002). The remaining highlighted announcements (‘first easing’, ‘unwelcome’ and ‘considerable’)
are mixtures of both types of forward guidance.
The announcement on January 3, 2001 triggers the largest Delphic shock in the sample. This is a large inter-meeting rate cut
that, as discussed in Gürkaynak et al. (2005), caused financial markets to mark down the probability or a recession and as a result
expect higher rates down the road. The Gürkaynak et al. (2005) methodology picks it up as a combination of a target factor easing
and a path factor (forward guidance) tightening. In this paper’s methodology the forward guidance is of the Delphic kind and
therefore reinforces the stock market gains rather than dampening them, which helps match the extremely large, 400bp increase in

9
M. Jarociński Journal of Monetary Economics 144 (2024) 103548

Table 1
Pairwise rank and linear correlations with baseline shocks 𝑢1 , 𝑢2 , 𝑢3 and 𝑢4 .
Obs. 𝑢1 𝑢2 𝑢3 𝑢4
Changing the sample
Drop January 3, 2001 240 𝑢1 : 1.00 𝑢2 : 0.92 𝑢3 : 0.95 𝑢4 : 0.99
(1.00) (0.89) (0.98) (1.00)
Drop QE1 (March 18, 2009) 240 𝑢1 : 0.999 𝑢2 : 0.998 𝑢3 : 0.998 𝑢4 : 0.996
(1.000) (0.999) (0.998) (0.999)
Sample 1991–2004 120 𝑢1 : 0.94 𝑢2 : 0.88 𝑢3 : 0.92 𝑢4 : 0.97
(0.99) (0.95) (0.94) (0.99)
Sample 2005–2019 121 𝑢1 : 1.00 𝑢2 : 0.82 𝑢3 : 0.88 𝑢4 : 0.96
(1.00) (0.77) (0.97) (0.98)
Other papers
Swanson (2021) 241 ff: 0.79 fg: 0.75 lsap: −0.66 fg: 0.47
(0.95) (0.81) (−0.84) (0.48)
Jarociński and Karadi (2020) FF4 221 MP: 0.43 MP: 0.63 MP: −0.05 CBI: 0.59
(0.69) (0.45) (−0.04) (0.66)
Jarociński and Karadi (2020) 1stPC 241 MP: 0.39 MP: 0.71 MP: −0.08 CBI: 0.77
(0.67) (0.56) (−0.02) (0.81)

Note. Rank (Spearman’s) correlations on top, regular font; linear (Pearson’s) correlations below, in brackets, italics. ‘ff’, ‘fg’ and
‘lsap’ stand for fed funds, forward guidance and large scale asset purchase shocks. ‘MP’ and ‘CBI’ stand for monetary policy and
central bank information shocks.

the S&P500. Since this 𝑢4 shock is so large and its market interpretation tricky to interpret,6 I test the robustness of the results to
dropping the January 3, 2001 observation from the sample and re-estimating the model. The results without this observation are
very similar. The rank correlation of the two estimates of 𝑢4 on the remaining dates is 0.99 (Table 1).
In the announcement labeled ‘weakness’ on August 13, 2002 the FOMC stated that the balance of risks has shifted towards
economic weakness. This stimulated both pessimism, reflected in stock market losses, and expectations of lower rates in the future.
Therefore, while the announcement did not promise a rate cut explicitly, it worked as a Delphic forward guidance.
In the 2009–2019 plot in Fig. 4 the largest Delphic shock is the ‘mid-2013’ announcement, issued on August 9, 2011, in which
the FOMC stated that the ‘‘economic conditions ... are likely to warrant exceptionally low levels for the federal funds rate at least
through mid-2013’’. It is intuitive that such a wording of the forward guidance is prone to trigger a Delphic interpretation (e.g. Del
Negro et al., 2023 discuss the Delphic nature of this announcement). By contrast, the forward guidance episodes from December
2014 to March 2016 are either Odyssean, 𝑢2 or mixes of Delphic and Odyssean.
Interestingly, the ‘dovish’ announcement on September 17, 2015, which is a major forward guidance shock in Swanson (2021),
does not show up as such here. On that day markets priced in some probability that the Fed would raise the rates for the first time
since 2008. The Fed did not change the rates and the MP1 dropped by 6.4 basis points upon the announcement. This is interpreted
here as a standard fed funds rate shock 𝑢1 of −6.4 basis points, accompanied by a mix of small Odyssean and Delphic forward
guidance shocks of −1.5 basis points each. However, there are few other so clear discrepancies between the two approaches.
The largest by far LSAP shock 𝑢3 accompanies the announcement of the expansion of the QE1 program (March 18, 2009). I
check the robustness of the results to omitting this observation, but all the lessons remain almost unchanged (see the second line of
Table 1). As in Swanson’s analysis, this shock is accompanied by a large expansionary Odyssean forward guidance shock. Another
sizable expansionary LSAP shock happens at the announcement of the ‘Operation Twist’ (September 21, 2011). Finally, there is first
a contractionary and then an expansionary LSAP shock during the ‘‘taper tantrum’’ episode, the first on June 19, 2013 (‘taper’)
the second on September 18, 2013 (‘no taper’). Also consistently with Swanson’s findings, there are no expansionary LSAP shocks
during the announcements of QE2 and QE3 programs.
Table 1 shows the correlations between 𝑢1 , 𝑢2 , 𝑢3 , 𝑢4 and the related shocks identified with very different techniques by Swanson
(2021) and Jarociński and Karadi (2020). The standard policy shock 𝑢1 is highly correlated with Swanson’s Fed Funds rate shock,
the two forward guidance shocks 𝑢2 and 𝑢4 are both highly correlated with Swanson’s single forward guidance shock (he does not
distinguish between Delphic and Odyssean forward guidance) and the LSAP shock 𝑢3 is highly correlated with Swanson’s LSAP shock
(I scales this shock with the opposite sign). Jarociński and Karadi (2020) have a single catch-all monetary policy shock which is
highly correlated both with 𝑢1 and with 𝑢2 (but does not capture asset purchases 𝑢3 ). The Delphic shock 𝑢4 is highly correlated with
the central bank information (CBI) shock of Jarociński and Karadi (2020), which also picks up the positive correlation between
interest rate surprises and stock price surprises. For the baseline CBI shock, which uses the fourth fed funds future (FF4) as the
summary of the interest rate surprises, the rank correlation is 0.59. For the CBI shock based on the first principal component of
futures with maturities up to 1 year as the summary of interest rate surprises, (reported by Jarociński and Karadi, 2020 in the
Appendix), the rank correlation is even higher, 0.77.

6 The Fed stated that ‘‘there is little evidence to suggest that longer-term advances in technology and associated gains in productivity are abating’’ and cited

recent signals of economic weakness as the reason for the cut, but market commentary focused mainly on the latter.

10
M. Jarociński Journal of Monetary Economics 144 (2024) 103548

4.3. Further estimation results

Inspecting matrix 𝑊 (reported in Online Appendix D) we can see that 𝑢1 is essentially equal to the MP1, while the other
shocks are non-trivial linear combinations of all variables. As can be inferred already from Fig. 4, all the shocks are very fat
tailed. The maximum likelihood estimates of the degree-of-freedom parameters are 𝑣̂ 1 = 0.6 (0.07), 𝑣̂ 2 = 2.14 (0.4), 𝑣̂ 3 = 2.13 (0.4)
and 𝑣̂ 4 = 2.0 (0.4) (standard deviations in parentheses). Online Appendix D reports their posterior densities, which peak around
the maximum likelihood estimates. When 𝑣 ≤ 2 the Student-t density does not have a finite variance. The finding that the Fed
shocks are well characterized by a density without a finite variance echoes a similar finding in Anttonen et al. (2023) obtained with
different data and with a more general shock density. These authors ascribe their finding to the presence of unmodeled conditional
heteroskedasticity and the same reasoning applies here. Especially the first shock has very low volatility during the Zero Lower Bound
period and much higher volatility otherwise. Explicit modeling of conditional heteroskedasticity might tighten the identification
further but this is left for future research.
Online Appendix D reports also additional estimation results. The marginal posterior densities of the elements of 𝑊 and 𝐶 are
very tight and approximately Gaussian. The posterior modes corresponding to different shock permutations are separated by very
low probability regions and the posterior simulator never jumped between the permutations. The shocks do not have significant
rank correlations. Large shocks tend to occur together (as can be seen also in Fig. 4), suggesting that it may be useful to relax the
independence assumption, but, as shown in the appendix, this has very little effect on the results.

5. Other exercises

This section summarizes other empirical results reported in detail in the Online Appendix.
First, the results are not sensitive to mis-estimation or mistaken restrictions on 𝑣, as long as it does not exceed the value of about
12. Online Appendix E studies this sensitivity. Instead of estimating 𝑣, it imposes values on the grid from 0.5 to 30, and estimates
only 𝐶. It shows that the results are very similar for values 𝑣 < 12.
When the model is estimated on subsamples, the LSAP shock 𝑢3 is only detectable in the later part of the sample, which is
intuitive because the LSAP policies were only introduced in the wake of the Great Financial Crisis. The remaining shocks are present
throughout the sample and have reasonably similar effects. The shocks estimated in subsamples are quite highly correlated with the
respective shocks estimated on the full sample (see Table 1). Detailed sub-sample results are provided in Online Appendix F.
The variant of the model that assumes the PDMT distribution of the shocks does detect some dependence between the shocks.
However, the shocks still exhibit enough independent variation to yield basically the same 𝐶, with only marginally larger posterior
uncertainty. To push the model further I force the degree of dependence to be even higher by imposing a priori restrictions. These
restrictions widen the uncertainty bands more visibly, but the key features of the shocks are still distinguishable. Moreover, when I
evaluate these restrictions formally using Bayes factors they are very strongly rejected. More details of these exercises are provided
in Online Appendix G. The key lesson from this analysis is that the shocks need not be fully independent to achieve meaningful
identification. In the present empirical application the baseline results are robust to relaxing the assumption of independence.
Very similar four shocks are obtained when instead of the four variables included in the baseline model, the model is estimated
on the first four principal components extracted from a larger set of GSS variables that includes more interest rates at various
maturities. Online Appendix H reports the details. It also explores models with three or five shocks and with other information sets.
Models with three shocks that include the SP500 surprise detect shocks very similar to 𝑢1 , 𝑢2 , 𝑢4 . Models with three shocks that only
include interest rates recover 𝑢1 , 𝑢3 and a composite forward guidance shock (an amalgam of 𝑢2 and 𝑢4 ). The resulting three shocks
closely resemble the shocks that Swanson (2021) estimates from the same information set but with very different methods. This
result serves as another statistical validation of his exercise. Models with more than four shocks feature similar baseline four shocks
and either yield additional Delphic shocks or a new shock that mainly affects the exchange rate.

6. Longer term effects: daily local projections

To study the effects of the four baseline shocks beyond the first thirty minutes after the FOMC announcement I estimate local
projections:

𝑥𝑡+ℎ − 𝑥𝑡−1 = 𝛼 + 𝛽ℎ𝑖 𝑢𝑖,𝑡 + 𝑒𝑡 , (6)

where 𝑥𝑡 is a daily financial variable and 𝑡 is an FOMC announcement day. I consider horizons ℎ = 1, 3, 5, 10, 15, 20, 25 business
days. 𝑢𝑖,𝑡 , 𝑖 = 1, 2, 3, 4 are one sample standard deviation shocks from the baseline model. The shocks are included in the regressions
one-by-one. The quantity of interest is 𝛽ℎ𝑖 , the effect of a one sample standard deviation shock. Eq. (6) is estimated with OLS and
with heteroskedasticity-robust errors, and a bootstrap procedure accounts for the uncertainty in the estimation of the shocks.7
Fig. 5 reports local projection results where variable 𝑥 is, respectively, the federal funds effective rate, 2-year and 10-year treasury
yields, and the S&P500 stock price index (i.e. the daily analogues of the high-frequency variables entering the baseline model) as
well as corporate bond spreads, breakeven inflation rates and exchange rate. Online Appendix I reports data sources and robustness
to alternative variables.

7 For each draw 𝑊 𝑚 , 𝑚 = 1, … , 2000 from the posterior (i) compute 𝑈 𝑚 = 𝑌 𝑊 𝑚 and standardize it; (ii) estimate regression (6) by OLS and store the point

estimate of 𝛽ℎ𝑖 and its Eicker–Huber–White (EHW) variance. The reported plots represent the means of the 2000 point estimates and the total variances computed
from the 2000 EHW variances and point estimates using the law of total variance. Uncertainty about 𝑈 increases these bands by on average 2.6% (and at most
14%) of the width of the bands conditional on the maximum likelihood estimates of the shocks.

11
M. Jarociński Journal of Monetary Economics 144 (2024) 103548

Fig. 5. The effects of the shocks on daily financial variables, local projection estimates
Note. The figure reports the effect of a 1-sample standard deviation shock based on local projections (6). The left-hand-side variables are in percent. Solid
line: OLS estimate conditional on the maximum likelihood estimate of the shocks. Darker shade: 1 standard deviation bands (68% probability). Lighter shade:
1.645 standard deviation bands (90% probability). The standard deviations are heteroskedasticity robust and account for the uncertainty in the estimation of
the shocks, as explained in footnote 7.

12
M. Jarociński Journal of Monetary Economics 144 (2024) 103548

Four main lessons follow from these local projection results. First, the effects of the shocks on interest rates and stock prices in
the first 30 min given by the matrix 𝐶 are not just temporary blips. Most of them persist in the following days and weeks, some of
them die out gradually and some, most notably for 𝑢4 , get amplified over time. 𝑢1 has a persistent effect on the short and medium
(2-year) interest rates and the stock prices. 𝑢2 has a persistent effect on the 10-year rate and stock prices, but its effect on the 2-year
rate dies out within about 2 weeks, suggesting that it may be a weak instrument for the 2-year rate in lower frequency data. 𝑢3 has
a persistent effect on the 10-year rate and not on much else, echoing its high-frequency effects (the positive effect on stock prices
is just a temporary blip). The Delphic shock 𝑢4 triggers a dynamics that is different from all other shocks: its initial effects are not
significant for the first four days, but gradually get amplified and both interest rates and stock prices are higher after one month
than at the end of the FOMC day.
The second lesson from Fig. 5 is that the shocks gradually propagate through the financial system and eventually have pronounced
effects on the financial conditions in the economy. In particular, corporate bond spreads gradually increase after 𝑢1 , 𝑢2 , 𝑢3 reflecting
the tightening of monetary policy, and decline after 𝑢4 , consistently with its Delphic character. By the end of the month the responses
are large and statistically significant. Standard 𝑢1 and 𝑢2 shocks increase the high yield spread by about 20 bp, 𝑢3 increases it by about
10 bp and 𝑢4 reduces it by 20 bp. While Fig. 5 plots the high yield bond spread, Online Appendix I shows that for other bond ratings
the effects go in the same direction and that the spreads account for much of the responses of corporate bond yields, suggesting the
importance of financial frictions in the transmission of monetary policy. It also reports similar responses of the broader National
Financial Conditions Index.
The third lesson is that breakeven inflation rates gradually decline after contractionary monetary policy shocks 𝑢1 , 𝑢2 , 𝑢3 and
gradually increase after the Delphic shock 𝑢4 . This might reflect intuitive reactions of inflation expectations to the shocks. These
results could be also due to liquidity and/or inflation risk premia increasing after 𝑢1 , 𝑢2 , 𝑢3 and declining after 𝑢4 . Fig. 5 presents the
results for the 5-year breakeven rates. Online Appendix I reports on the term structure of breakeven rates, showing that the effects
are more backloaded for 𝑢1 and more frontloaded for 𝑢2 and 𝑢3 , with 𝑢4 falling in-between.
The fourth lesson is that contractionary monetary policy shocks 𝑢1 , 𝑢2 , 𝑢3 strengthen the dollar exchange rate, while the Delphic
shock 𝑢4 does not affect it much. Standard policy and forward guidance shocks 𝑢1 and 𝑢2 strengthen the dollar vs the euro statistically
significantly by 20–25 bp. The effect of the asset purchase shock 𝑢3 is quantitatively similar but estimated with a large uncertainty.
The effect of the Delphic shock 𝑢4 on the dollar is 10 bp on impact but is completely reversed after one day. This shock’s little
impact on the exchange rate is consistent with the role of the dollar as a barometer of financial market risk-taking capacity (Avdjiev
et al., 2019). A positive Delphic shock increases the financial markets’ appetite for risk and this pushes the dollar down, working
against the standard uncovered interest rate parity effect of higher US interest rates. Online Appendix I shows that the findings on
the euro exchange rate mostly extend to other currencies, especially those of the advanced economies.

7. Conclusions

This paper exploits the high kurtosis of financial market responses to pin down four main dimensions of FOMC announcements,
which can be naturally labeled as: standard monetary policy, Odyssean forward guidance, LSAP and Delphic forward guidance.
These shocks have plausible effects on financial markets and provide intuitive interpretations of the FOMC announcements in the
sample. The paper explains the intuition behind the fat tails-based identification and shows that it requires only a sufficient degree
of independence, rather than full independence. It tracks the delayed effects of the estimated shocks on financial markets and finds
that they eventually have pronounced effects on breakeven inflation rates and financial conditions, especially on the corporate bond
spreads.

Data availability

Data will be made available on request.

Appendix A. Supplementary data

Supplementary material related to this article can be found online at https://doi.org/10.1016/j.jmoneco.2024.01.001.

References

Andrade, P., Ferroni, F., Melosi, L., 2023. Identification Using Higher Order Moments Restrictions. Working Paper Series WP 2023-28, Federal Reserve Bank of
Chicago, http://dx.doi.org/10.21033/wp-2023-28.
Anttonen, J., Lanne, M., Luoto, J., 2021. Statistically Identified SVAR Model With Potentially Skewed and Fat-Tailed Errors. Technical Report, University of
Helsinki, http://dx.doi.org/10.2139/ssrn.3925575, (Posted 20 September 2021; last revised 8 March 2023) Available at SSRN: http://ssrn.com/abstract=
3925575.
Anttonen, J., Lanne, M., Luoto, J., 2023. Bayesian Inference on Fully and Partially Identified Structural Vector Autoregressions. Technical Report, University of
Helsinki, http://dx.doi.org/10.2139/ssrn.4358059, (February 14, 2023) Available at SSRN: http://ssrn.com/abstract=4358059.
Avdjiev, S., Du, W., Koch, C., Shin, H.S., 2019. The dollar, bank leverage, and deviations from covered interest parity. Am. Econ. Rev.: Insights 1 (2), 193–208.
http://dx.doi.org/10.1257/aeri.20180322.
Bauer, M.D., Swanson, E.T., 2023. A reassessment of monetary policy surprises and high-frequency identification. NBER Macroecon. Annu. 37 (1), 87–155.
http://dx.doi.org/10.1086/723574.

13
M. Jarociński Journal of Monetary Economics 144 (2024) 103548

Bauwens, L., Lubrano, M., 1998. Bayesian inference on GARCH models using the Gibbs sampler. Econom. J. 1 (1), 23–46. http://dx.doi.org/10.1111/1368-
423X.11003.
Bonhomme, S., Robin, J.M., 2009. Consistent noisy independent component analysis. J. Econometrics 149 (1), 12–25. http://dx.doi.org/10.1016/j.jeconom.2008.
12.019.
Braun, R., 2023. The importance of supply and demand for oil prices: Evidence from non-Gaussianity. Quant. Econ. 14 (4), 1163–1198. http://dx.doi.org/10.
3982/QE2091.
Brunnermeier, M., Palia, D., Sastry, K.A., Sims, C.A., 2021. Feedbacks: Financial markets and economic activity. Amer. Econ. Rev. 111 (6), 1845–1879.
http://dx.doi.org/10.1257/aer.20180733.
Campbell, J.R., Evans, C.L., Fisher, J.D.M., Justiniano, A., 2012. Macroeconomic effects of federal reserve forward guidance. Brook. Pap. Econ. Act. 1–80, URL
https://EconPapers.repec.org/RePEc:bin:bpeajo:v:43:y:2012:i:2012-01:p:1-80.
Cieślak, A., Schrimpf, A., 2019. Non-monetary news in central bank communication. J. Int. Econ. 118, 293–315. http://dx.doi.org/10.1016/j.jinteco.2019.01.012.
Comon, P., 1994. Independent component analysis, A new concept? Signal Process. 36 (3), 287–314. http://dx.doi.org/10.1016/0165-1684(94)90029-9.
Davis, R., Ng, S., 2022. Time Series Estimation of the Dynamic Effects of Disaster-Type Shock. Technical Report, http://dx.doi.org/10.48550/arXiv.2107.06663,
ArXiV, arXiv:2107.06663v2.
Del Negro, M., Giannoni, M.P., Patterson, C., 2023. The forward guidance puzzle. J. Political Econ. Macroecon. 1 (1), 43–79. http://dx.doi.org/10.1086/724214.
Drautzburg, T., Wright, J.H., 2021. Refining Set-Identification in VARs through Independence. NBER Working Papers 29316, National Bureau of Economic
Research, Inc, http://dx.doi.org/10.3386/w29316.
Fiorentini, G., Sentana, E., 2020. Discrete Mixtures of Normals Pseudo Maximum Likelihood Estimators of Structural Vector Autoregressions. Working Papers
wp2020-2023, CEMFI, URL https://ideas.repec.org/p/cmf/wpaper/wp2020_2023.html.
Gouriéroux, C., Monfort, A., Renne, J.P., 2017. Statistical inference for independent component analysis: Application to structural VAR models. J. Econometrics
196 (1), 111–126. http://dx.doi.org/10.1016/j.jeconom.2016.09.007, URL http://www.sciencedirect.com/science/article/pii/S0304407616301749.
Gouriéroux, C., Monfort, A., Renne, J.P., 2020. Identification and estimation in non-fundamental structural VARMA models. Rev. Econom. Stud. 87 (4), 1915–1953.
http://dx.doi.org/10.1093/restud/rdz028, arXiv:https://academic.oup.com/restud/article-pdf/87/4/1915/33461576/rdz028.pdf.
Gürkaynak, R.S., Karasoy-Can, H.G., Lee, S.S., 2022. Stock market’s assessment of monetary policy transmission: The cash flow effect. J. Finance 77 (4), 2375–2421.
http://dx.doi.org/10.1111/jofi.13163, URL http://refet.bilkent.edu.tr/research.html.
Gürkaynak, R.S., Sack, B., Swanson, E., 2005. Do actions speak louder than words? The response of asset prices to monetary policy actions and statements. Int.
J. Central Bank. 1 (1), 55–93, URL https://ideas.repec.org/a/ijc/ijcjou/y2005q2a2.html.
Hyvärinen, A., Karhunen, J., Oja, E., 2001. Independent Component Analysis. Wiley.
Inoue, A., Rossi, B., 2021. A new approach to measuring economic policy shocks, with an application to conventional and unconventional monetary policy.
Quant. Econ. 12 (4), 1085–1138. http://dx.doi.org/10.3982/QE1225.
Jarociński, M., Karadi, P., 2020. Deconstructing monetary policy surprises - the role of information shocks. Am. Econ. J.: Macroecon. 12 (2), 1–43. http:
//dx.doi.org/10.1257/mac.20180090.
Kuttner, K.N., 2001. Monetary policy surprises and interest rates: Evidence from the fed funds futures market. J. Monetary Econ. 47 (3), 523–544. http:
//dx.doi.org/10.1016/S0304-3932(01)00055-1.
Lanne, M., Luoto, J., 2020. Identification of economic shocks by inequality constraints in Bayesian structural vector autoregression. Oxf. Bull. Econ. Stat. 82 (2),
425–452. http://dx.doi.org/10.1111/obes.12338.
Lanne, M., Luoto, J., 2021. GMM estimation of non-Gaussian structural vector autoregression. J. Bus. Econom. Statist. 39 (1), 69–81. http://dx.doi.org/10.1080/
07350015.2019.1629940.
Lanne, M., Meitz, M., Saikkonen, P., 2017. Identification and estimation of non-Gaussian structural vector autoregressions. J. Econometrics 196 (2), 288–304.
http://dx.doi.org/10.1016/j.jeconom.2016.06.002.
Lewis, D.J., 2021. Identifying shocks via time-varying volatility. Rev. Econom. Stud. 88 (6), 3086–3124. http://dx.doi.org/10.1093/restud/rdab009.
Lewis, D.J., 2022. Robust inference in models identified via heteroskedasticity. Rev. Econ. Stat. 104 (3), 510–524. http://dx.doi.org/10.1162/rest_a_00963.
Lewis, D.J., 2023. Announcement-specific decompositions of unconventional monetary policy shocks and their effects. Rev. Econ. Stat. 1–46. http://dx.doi.org/
10.1162/rest_a_01315.
Maxand, S., 2020. Identification of independent structural shocks in the presence of multiple Gaussian components. Econom. Stat. 16, 55–68. http://dx.doi.org/
10.1016/j.ecosta.2018.10.005.
Melosi, L., 2017. Signalling effects of monetary policy. Rev. Econom. Stud. 84 (2), 853–884. http://dx.doi.org/10.1093/restud/rdw050.
Mesters, G., Zwiernik, P., 2022. Non-Independent Component Analysis. Technical Report, Universitat Pompeu Fabra, URL https://www.geertmesters.com/.
Miescu, M., 2021. Forward Guidance and Conventional Monetary Policy Shocks: Identification and Assessment. mimeo, Lancaster University.
Miranda-Agrippino, S., Ricco, G., 2021. The transmission of monetary policy shocks. Am. Econ. J.: Macroecon. 13 (3), 74–107. http://dx.doi.org/10.1257/mac.
20180124.
Montiel Olea, J.L., Plagborg-Møller, M., Qian, E., 2022. SVAR identification from higher moments: Has the simultaneous causality problem been solved? AEA
Pap. Proc. 112, 481–485. http://dx.doi.org/10.1257/pandp.20221047.
Nakamura, E., Steinsson, J., 2018. High-frequency identification of monetary non-neutrality: The information effect. Q. J. Econ. 133 (3), 1283–1330. http:
//dx.doi.org/10.1093/qje/qjy004.
Rigobon, R., 2003. Identification through heteroskedasticity. Rev. Econ. Stat. 85 (4), 777–792. http://dx.doi.org/10.1162/003465303772815727.
Sims, C.A., 2021. SVAR Identification through Heteroskedasticity with Misspecified Regimes. Technical Report, Princeton University, URL http://sims.princeton.
edu/yftp/bpss/IDHmsspcfdRgms.pdf.
Swanson, E.T., 2021. Measuring the effects of Federal Reserve forward guidance and asset purchases on financial markets. J. Monetary Econ. 118, 32–53.
http://dx.doi.org/10.1016/j.jmoneco.2020.09.003.
Uhlig, H., 2005. What are the effects of monetary policy on output? Results from an agnostic identification procedure. J. Monetary Econ. 52 (2), 381–419.
http://dx.doi.org/10.1016/j.jmoneco.2004.05.007.
Van Nieuwerburgh, S., Veldkamp, L., 2010. Information acquisition and under-diversification. Rev. Econom. Stud. 77 (2), 779–805. http://dx.doi.org/10.1111/
j.1467-937X.2009.00583.x.

14