ArticlePDF Available

Abstract

We construct a new indicator to capture media sentiment about the European Central Bank monetary policy and its relevant environment by analyzing 25,000 articles from five major international newspapers. Using named entity recognition and part-of-speech tagging, we propose a methodology to dissociate the dissemination of official communications of the central bank from the media comments. The resulting (daily) index correlates with some (monthly) standard measures of economic sentiment but reveals idiosyncratic information on monetary policy. Analyzing the determinants of our index, we find that both press conference and inter-meeting communications of the President significantly affect media sentiment. We then show that, controlling for a large range of factors, daily changes in media sentiment have predictive power for financial market inflation expectations.
Media sentiment on monetary policy: determinants and
relevance for inflation expectations
Matthieu Picault1, Julien Pinter2, and Thomas Renault3
1Univ. Orléans, LEO
2University of Minho
3Université Paris 1 Panthéon-Sorbonne
accepted at Journal of International Money and Finance
Abstract
We construct a new indicator to capture media sentiment about the European Central
Bank monetary policy and its relevant environment by analyzing 25,000 articles from five
major international newspapers. Using named entity recognition and part-of-speech tagging,
we propose a methodology to dissociate the dissemination of official communications of the
central bank from the media comments. The resulting (daily) index correlates with some
(monthly) standard measures of economic sentiment but reveals idiosyncratic information on
monetary policy. Analyzing the determinants of our index, we find that both press conference
and inter-meeting communications of the President significantly affect media sentiment. We
then show that, controlling for a large range of factors, daily changes in media sentiment have
predictive power for financial market inflation expectations.
Keywords: central bank communication, European Central Bank, textual analysis, inflation
expectations, media sentiment.
JEL Classification : E43, E52, G12
Corresponding author: julien.pinter@gmail.com.
The authors are particularly grateful to the participants of the IMF Research Department internal seminar, the
participants of the GDRE and AFSE conferences, and the participants of internal seminars at Université Paris 1
Panthéon-Sorbonne and Poitiers University. We are also particularily grateful to Roman Horvath, Pavel Gertler,
Paul Hubert, Bernd Hayo and Guntehr Capelle-Blancard for their useful remarks, and to Pavel Gertler for kindly
sharing his data with us. We also thank Paul Tetlock, Diego Garcia, and Hamza Benani for sharing their thoughts
at different stages of this paper.
CRediT author statement : Picault Matthieu: conceptualization, software, formal analysis, investigation (data
collection, coding), writing - review and editing, visualization (presentations), project administration; Pinter Julien:
conceptualization, formal analysis, investigation (data collection, econometrics), writing - original draft, writing
- review and editing, visualization (presentations), project administration; Thomas Renault: conceptualization,
software, formal analysis, writing - review and editing.
1
1 Introduction
The media coverage of central bank actions is of central importance for monetary policy ef-
fectiveness (Berger et al., 2011). As financial market participants rely heavily on media reports
to get information on central banks (Hayo and Neuenkirch, 2015), media coverage can facilitate
the transmission of the central bank policy while simultaneously enhancing the perception of its
actions. However, media are not just a mere relay of the information released by central banks.
Through their analysis of the central bank actions and communications, the media could also
influence the perception of the central bank policies and their relevant economic environment by
market participants and, consequently, have an impact on the economic outcome. While such a
media channel is acknowledged in research, such as Hayo and Neuenkirch (2015) or Blinder et al.
(2008), its existence has not been, to our knowledge, empirically documented.
In this paper, we address the question of whether central banks do affect media sentiment
through their actions and regular communications and whether media sentiment does in turn mat-
ter for financial markets’ inflation expectations. We thus adopt a framework consistent with the
"more realistic view" of the effect of central bank communication of Hayo and Neuenkirch (2015),
where it is noted that media coverage is a potential key channel between central bank communi-
cation and financial market participants’ perception (see Figure 1). Doing so, we depart from a
large strand of research focusing solely on the direct link between central banks’ communications
and financial variables.1
Media sentiment refers to the net degree of positivity (optimism) or negativity (pessimism) of
the media content, in line with the connected literature (Kearney and Liu, 2014, Tetlock, 2007). We
quantify it using advanced text-mining methods on a corpus of about 25,000 newspapers articles
related to the European Central Bank (ECB henceforth), published between 2006 and 2017. When
doing so, we do not consider each sentence of the media content. We use an innovative approach
to disentangle the media dissemination of the central bank’s communications from the media
comments on the monetary policy and its environment. In fact, newspaper articles following
a central bank communication (press conference, interviews, speeches, etc.) often contain several
direct quotes of the official communication. If those quotes are considered to capture the sentiment
of a given article, media sentiment will not only relate to the opinions of the journalists and experts
but will also reflect the tone of the official communications. To avoid this, we use text-mining
1Important contributions include, among others, Jansen and De Haan (2005), Ehrmann and Fratzscher (2007a),
Ehrmann and Fratzscher (2007b), Jansen and de Haan (2007), Rosa and Verga (2008), Schmeling and Wagner
(2015), Gertler and Horvath (2018), Hubert and Labondance (2020), Beaupain and Girard (2020), and Tillmann
(2020). See Blinder et al. (2008) for an influential survey on the topic.
2
Figure 1: "Standard view" versus "more realistic view" of the transmission of the central bank
actions and communications, adapted from Hayo and Neuenkirch (2015)
Standard view
Central bank actions
and communications Economic outcome
More realistic view
Central bank actions
and communications Economic outcome
Perception by
financial markets
Media coverage
Monetary policy
related news
Our Focus
Note: This Figure presents the "standard view" in which central bank actions and communications are assumed
to directly affect the economic outcome, and the "more realistic view" (adapted from Hayo and Neuenkirch
(2015)) in which the perception by financial markets is presented as a potential key element between the central
bank action / communication and the economic outcome.
methods to remove all sentences and quotes where the subject of the sentence is one of the ECB
Governing Council members or the ECB.
Once non-relevant sentences are filtered out, we use standard lexicon approaches (Loughran
and McDonald, 2011) to capture the sentiment of each article related to the ECB and construct a
daily media sentiment indicator. The resulting index intends to capture media sentiment on the
ECB monetary policy and its relevant environment. Empirical investigations suggest that it does
so relatively well. The monthly average of our index indeed correlates with some monthly standard
measures of economic sentiment, but it also reveals idiosyncratic information on monetary policy.
When, in a second stage, we investigate the determinants of our index, we find that monetary policy
decisions, press conferences, and inter-meeting communications of the Presidents all significantly
affect our media sentiment index, even when controlling for other factors. Consistent with some
central bankers’ comments, we find a straightforward linear effect: hawkish communications on
monetary policy decrease media sentiment, while dovish communications increase media sentiment.
3
Communications related to the economic outlook are also found to matter in some cases, but to
a lower extent than communications related to monetary policy inclination. The same applies for
communications of other Governing Council members related to monetary policy inclination. In a
last stage, we analyze the relevance of our index for the inflation expectations of financial markets.
We focus on the 10-years and the 5-years-to-5-years inflation swap, of well known relevance for
policy makers (Draghi, 2014).2We find that our media sentiment index has predictive power for
the next-day change in the 5-year 5-year and 10-year inflation swap, with more positive sentiment
increasing inflation expectations. This holds true even after controlling for a large set of macroeco-
nomic surprises, central bank decisions, the textual content of press conferences, as well as central
bank inter-meeting communications, where the latter are taken from a rich external hand-coded
dataset. The effect we find is of low magnitude, suggesting a negligible economic relevance.
The paper first contributes to the literature on monetary policy and the media by providing
new evidence showing the impact of various types of communication from central bankers on media
sentiment. In particular we complement the findings of Berger et al. (2011) and Böhm et al. (2012),
who analyzed the determinants of media favorableness toward, respectively, the monetary policy
decisions of the ECB and the Czech Central Bank. Although our focus is on media sentiment
rather than favorableness, we find, as in Berger et al. (2011), that only the surprising component
of an interest rate decision matters for media. Tobback et al. (2017) build an index of perceived
hawkisness / dovishness in the media for the days following the ECB press conference and find that
the tone of the ECB communication (hawkish versus dovish) is reflected in the media perception
of the ECB actions. We differ from Tobback et al. (2017) and from the two previously mentioned
works insofar as we do not focus solely on what media say a few days after the conference: our
index is a continuous measure. Consequently, we also contribute to the literature by showing that
inter-meeting communications, especially those related to monetary policy inclination, do affect
media.3
We also contribute to the literature on the impact of media sentiment on financial variables.
Tetlock (2007) and Garcia (2013), among others, apply textual analysis methods to newspaper
content and show that the resulting sentiment index helps forecast stock returns. Fraiberger et al.
(2021) use articles from Reuters to show that general news sentiment predicts daily returns in
both advanced and emerging markets and drives worldwide capital flows. Instead, the novelty of
2"The 5year/5year swap rate (...) is the metric that we usually use for defining medium term inflation"
3Another strand of the literature has focused on uncertainty in media. Husted et al. (2019) build an index of
monetary policy uncertainty from three US newspapers and show that shocks to monetary policy uncertainty raise
credit spreads and reduce output. Bennani (2018) measures media uncertainty related to the ECB’s monetary
policy decisions and finds it to be related to future monetary policy decisions.
4
this paper is that we focus on financial markets’ inflation expectations. This variable is of special
interest for central bankers and policymakers, as it shows whether markets are convinced that a
central bank will keep the inflation rate within its set target. We find that sentiment also plays a
role in that context though of limited economic relevance.
We add to the literature on inflation expectations, a key component of the response of long-
term nominal rates to news (Bauer, 2015), by providing evidence that media sentiment has an
impact them. Several studies noted that financial markets’ long-term4inflation expectations were
not responsive to monetary policy or macroeconomic news surprises when the central bank was
perceived as credible in achieving an explicit inflation target (see Garcia and Werner (2018) for a
comprehensive review). For example, Beechey et al. (2011) find that long-term inflation expecta-
tions were unresponsive to monetary policy or macroeconomic surprises in the Euro area before
the global financial crisis, while they were in the US. In contrast, Garcia and Werner (2018) find
that long-term financial markets’ expectations have become more responsive to macroeconomic
news surprises after 2013 in the Euro area, while Ambler and Rumler (2019) find that some recent
unconventional monetary policy announcements moved financial markets’ inflation expectations.5
Lastly, we contribute to the overall literature on monetary policy transmission by providing a
new index related to media sentiment on monetary policy, which can be used for further research.
Our Media Sentiment Indicator is freely available online.6
The rest of the paper is organized as follows. Section 2 details the construction of our index and
discusses its interpretation. Section 3 describes the methodology and the empirical results on the
determinants of media sentiment. Section 4 presents the methodology and the empirical results
on the relevance of media sentiment for inflation expectations. Section 5 concludes the paper.
4By "long-term" we refer in this paper to inflation expectations where the relevant horizon is 10 years. Some
papers we survey later also refer to such horizons as the "medium term".
5Other important contributions on the anchoring of financial markets’ expectations in the Euro area include,
among others, Bundick and Smith (2018) and Galati et al. (2011).
6http://www.cbcomindex.com/data.php or https://sites.google.com/site/julienpinter/data-1
5
2 Measuring media sentiment on monetary policy
2.1 Construction
We first extract from the Factiva database all articles containing the keywords “ECB” or “Eu-
ropean Central Bank” published by five major international newspapers (The Financial Times,
The Wall Street Journal, The New-York Times, Barron’s, and The Times) from January 2006
to December 2016. These newspapers are selected because they have a large audience of profes-
sional investors and they are known to include journalists’ views and analysis on the information
compared to other outlets, like newswire services.7
Within these, we consider only articles of which the main topic is the ECB by imposing a
minimum of three8mentions of either "European Central Bank" or the name of a member of the
Governing Council.9Our final database contains 24,931 articles.
To dissociate the mere dissemination of central bank official communication by the media from
the sentiment of the media, we remove from each article all quotes where the subject of the sentence
is a member of the ECB or the ECB itself. To do so, we use the spaCy package in Python to
analyze the semantic structure of each sentence. We use part-of-speech (POS) tagging to classify
each word based on its context within a sentence. POS allows us to "tag" each word, that is,
to classify it as a noun, verb, determinant, or any other grammatical category and identify its
syntactic dependency in a given sentence.10
Table 1 shows an example of sentences that our methodology allows us to detect (and remove)
in an article extract from The Wall Street Journal (available entirely in Appendix A.1).
Following previous research on the quantification of textual content, we use the Loughran and
McDonald (2011) dictionary to compute a sentiment score for each article.11 We adapt it to
7Doing so, our indicator may not capture the overall sentiment of the market and may be relatively biased. In
particlular, the sentiment from the English-speaking newspapers we collected may differ from the views conveyed
in French or German newspapers for example.
8We read and manually classified 600 articles to select the threshold value to obtain an adequate balance between
type 1 errors (selecting an irrelevant article) and type 2 errors (not selecting a relevant article).
9The list thus includes "ECB", "European Central Bank", and the main members of the Governing Coun-
cil: Draghi, Coeure, Mersch, Asmussen, Constancio, Gonzales, Lautenschlager, Praet, Bini-Smaghi, Stark, Issing,
Papademos, Gonzalez Paramo, Trichet, Tumpel-Gugerell, Noyer, Orphanides, Smets, Visco, Weidmann, Wellink,
Knot, Linde, Reinesch, Lane, Villeroy de Galhau, Honohan, Weber, Ordonez, Quaden, Coene, Nowotny, Provopou-
los, and Stournaras.
10We remove not only the sentences in which the subject is explicitly the ECB or a member of the Governing
Council but also sentences where the subject is a pronoun indirectly referring to the member of the ECB or the
ECB itself (as identified by the subject of the previous sentence(s) in that case).
11This approach is probably the most widely used in the connected literature, and it has been shown to capture
relevant information in a central banking context (e.g., Schmeling and Wagner, 2015 or Hubert and Labondance,
6
Table 1: Media sentiment computation: example based on an article (extract)
Text format Meaning
in blue Positive words from LM dictionary
in red negative words from LM dictionary
crossed out Exact re-transciptions of ECB members’ words removed with part-of-speech tagging
The Wall Street Journal, September 9, 2011 :
The ECB’s sudden shift opens it to renewed criticism that -as in July 2008 when it increased rates
just weeks before the collapse of Lehman Brothers- officials didn’t recognize early warning signs
on the economy and overestimated inflation risks, exacerbating the slowdown. The ECB made the
right call, Mr. Trichet said Thursday. "We think what we did was appropriate" he said, referring
to this year’s rate hikes. (...) "I would very much like to hear the congratulations for an institution
that has delivered price stability in Germany...which is better than has ever been achieved in this
country," he said, responding to a question about Mr. Gabriel’s criticism.
Full article in Appendix A.1
our monetary policy context: to avoid a straightforward bias in the sentiment measure, we discard
expressions that undoubtedly refer directly to ECB monetary policy measures while containing the
following words from the Loughran and McDonald (2011) dictionary: "negative rate", "negative
interest/deposit rate", and "quantitative easing".12 Then, we define Sentimentias the difference
between the number of positive and negative words in the article idivided by the total number of
words in the article:
Sentimenti=P(P ositiv e wordsi)P(N egative W ordsi)
N umber of W ordsi
(1)
By construction, each article has a sentiment score between -1 and +1. Examples of very
negative and very positive articles are given in Appendix A.1. In the article extract of Table 1, we
also show the words captured by the Loughran and McDonald (2011) dictionary (the methodology
is illustrated in Appendix A.1 on the full article). The 20 positive and negative words with the
highest number of occurrences are detailed in Appendix A.2. Finally, at a daily frequency, we
compute an aggregate media sentiment indicator equal to the average sentiment of all articles
published during the period, Sentimentt, so that:
Sentimentt= 100 ×Pnt
i=1 Sentimenti
nt
(2)
2020). However, it has the well-known drawback of ignoring qualitative differences between words.
12This happens to be particularly important for "negative". By applying our filter, we remove about one-third
of its mentions.
7
where ntis the number of all our relevant articles published during day tin all of our five major
newspapers.
Built as such, our index intends to capture the media sentiment regarding the ECB monetary
policy and its relevant economic environment. A more negative media sentiment indicates that
the ECB monetary policy and its underlying environment are viewed with more pessimism, while
a more positive media sentiment indicates more optimism. Figure 2 shows the Media Sentiment
Index (MSI) as well as the evolution of the number of articles during our sample period, with a
weekly frequency.13
2.2 Discussion and comparison with other indexes
As Figure 2 highlights, many highs and lows of the MSI are related to important monetary
policy events. For example, the second allotment of the ECB Targeted Longer-Term Refinancing
Operations (TLTROs) in February 2012 and the beginning of the ECB Quantitative Easing (QE)
in January 2015 are both associated with an unusually sharp increase in media optimism. No other
policy after 2013 appears to be associated with a higher media sentiment than the one associated
with the launch of QE. The index also soars when the media convey high optimism on the economic
outlook, indicating that our index captures well information on economic developments which are
relevant to the ECB. Conversely, the media sentiment decreased sharply during the Great Recession
and during the Eurozone crisis. It is interesting to observe that not all unconventional monetary
policies are associated with a sharp increase in media sentiment. For example, the first TLTRO
allotment, at the end of 2011, is not associated with an increase in media sentiment while the second
one is. One potential explanation is that the allotment figures were better than expected for the
second TLTRO, while they were more in line with expectations for the first TLTRO, according to
some newspaper articles. We also do not observe any sharp increase in media sentiment the week
of the implementation of the Security Market Program (second week of May 2010).
The MSI significantly differs from other well-known measures related to monetary policy per-
ception, such as the newspaper-based Monetary Policy Uncertainty (MPU) index of Husted et al.
(2019). The MSI has a correlation of only -0.05 with the MPU index for the ECB, considering
monthly aggregate data (given that the MPU index is only available at a monthly frequency).
13In addition, we also corrected for any potential newspaper-specific political bias by subtracting the mean senti-
ment for each newspaper (over the whole time period) from each sentiment score of the articles of the corresponding
newspaper. The resulting index maintained a correlation of 99.5% with the original one. We still used this version
in our empirical analyses, although the transformation does not affect the results. Note also that the values of the
MSI are mostly negative because the Loughran and McDonald (2011) dictionary, as the English dictionary, contains
more negative words than positive words (2,355 negative words for 354 positive ones).
8
Figure 2: Weekly aggregated sentiment and number of media articles
Notes: Figure A (top) presents the media sentiment indicator (weekly frequency for readability). Positive
(negative) events are shown in green (red). Figure B (bottom) shows the number of articles per week.
This suggests that the two indexes capture very different aspects of monetary policy perception.
The MSI has a higher correlation (0.49) with the widely used market-based measure of investors’
sentiment of Baker and Wurgler (2006),14 a correlation of 0.51 with the Economic Sentiment Indi-
cator (ESI) from Eurostat, and 0.61 with the investor economic confidence index for the Eurozone
developed by Sentix, all available only at a monthly frequency. This suggests that the "relevant
economic environment" information component of our index is relatively important at a monthly
14Investor sentiment, as the authors define it (belief about future cash flows and investment risks that is not
justified by the facts at hand), is not conceptually equivalent to the sentiment we measure. Media sentiment, i.e.,
the positivity or negativity of newspapers’ tone, can reflect both fundamentals and autonomous beliefs. When we
take their measure not orthogonal to macroeconomic fundamentals measures, we get a correlation of -0.53. This
may suggest that our index is correlated to investor sentiment as "autonomous beliefs on the future state of the
economy", assuming that the comparison with US data is relevant.
9
frequency, while about half of the variations of our index still cannot be attributed to the variations
in the sentiment on the economic environment relevant for monetary policy. To better understand
the idiosyncratic information of our index relative to economic sentiment indexes, in Appendix
B, we plotted the residual component of the MSI index after having regressed it (its monthly
average) on the Sentix investor confidence index. We observed that the three highest points after
2010 corresponded in turn to the month when the "whatever it takes" speech was pronounced, the
month when forward guidance was introduced, and the month when QE was introduced. Overall,
this further confirms that our index captures relevant and idiosyncratic information on monetary
policy perception.
We also compared our index with a commonly used index capturing the overall tone of the
speech of the President during the ECB press conferences, measured also by the Loughran and
McDonald (2011) dictionary (LM press conference tone henceforth). We mean-standardized the
index, as well as our index, taking only its values on average over the day of the conference and
two days after it to ensure a meaningful comparison. The correlation with the LM press conference
and our index for these days is quite large, about 46%. It suggests that our index captures relevant
information on monetary policy. Investigating which idiosyncratic information is present in our
index for these particular days, we looked at the events with the largest differences between the
value of our index and the value of the LM press conference tone. The largest difference arises
in the meeting of February 2010 during which Jean-Claude Trichet "hinted strongly" that the
ECB emergency measures to support financial markets would further unwind, according to The
Financial Times, in a context of fears related to sovereign debts. That day, an analyst summarized
in The Financial Times that "Mr Trichet’s firm attitude has fuelled fears of horror scenarios".
The second largest difference arose during the meeting of January 2009. At that meeting, the
ECB decreased its key interest rate, but several newspapers report a substantial "disappointment"
among market actors, related to fears over growth prospects. Many other large differences arise
in similar contexts, thereby tending to confirm that our index captures idiosynchratic information
on the media perception on monetary policy that are not available in standard indexes based on
the official ECB communications.
10
3 The determinants of the media sentiment
3.1 Methodology
In investigating the determinants of media sentiment, we are particularly interested in analyz-
ing whether the central bank can affect it. We conjecture that the central bank can influence media
sentiment through two channels. First, its monetary policy decisions per se may impact media
sentiment. In this case, media sentiment might primarily reflect the favorableness/unfavorableness
of central bank news coverage, as studied by Berger et al. (2011). Second, monetary policy com-
munications of the President and other members of the Governing Council could, by themselves,
influence media sentiment on a regular basis. Taking into account relevant control variables, the
model we intend to estimate is as follows:
MSIt=
J
X
i=1
αiMSIti+β ECB_decisionst+φEC B_commt+τControlst+t(3)
where MSItcorresponds to our Media Sentiment Index on the ECB monetary policy at day t,
EC B_decisionstis a vector of variables related to ECB monetary policy decisions, ECB_commt
is a vector of variables related to the communication of ECB Governing Council members, and
Controlstis a set of economic and financial variables related to the economic environment. α=
(α1, α2, ...),β,φ, and τare the associated vectors of coefficients, and Jcorresponds to the number
of lags for the dependent variable accounting for the persistence of sentiment.
The variables are summarized and described in Table 2. The variable selection regarding the
monetary policy decisions is inspired by Böhm et al. (2012) and Berger et al. (2011). We also
include a variable aimed at capturing the unconventional monetary policy measures adopted by
the ECB by building a dummy variable, taking the value of 1 for each unconventional monetary
policy announcement and 0 otherwise, as well as a variable related to negative rates adoption.15
The sample period for which we estimate the model is January 1, 2010 until May 31, 2016. It
is limited by data availability on Governing Council members’ communications, which we discuss
subsequently.
Regarding monetary policy communications, we study communications during both the press
conference and the inter-meeting period. The first refers to the traditional press conference of the
ECB President, previously given once a month until January 2015 (once every six weeks afterward),
during which the ECB President gives the Governing Council’s assessment of the economic and
15Arguably, negative rates adoption had an important signal effect, which we intend to quantitatively unveil with
the inclusion of this variable.
11
financial conditions and further develops its monetary policy decisions. The second refers to
the various communications (e.g., media interviews, testimonies, conference talks) of the ECB
Governing Council members, during which they usually express their views on the future stance of
the ECB monetary policy and on the economic outlook. For press conference communications, we
gather an exhaustive database containing all the textual content of the ECB press conferences and
build quantitative measures for the policy inclination (hawkish versus dovish) and for the economic
outlook content inclination (positive economic outlook versus negative) of each communication
using the central bank-specific dictionary of Picault and Renault (2017) developed specifically for
that purpose. Appendix C provides further details. For inter-meeting communications, we use
a daily database kindly provided by Gertler and Horvath (2018). The authors manually coded
each communication from each Governing Council member for the above-mentioned period.16 We
obtained these daily detailed data from 2010 to May 2016, which in turn sets our sample size.17
The daily data give us the average policy inclination and economic outlook content inclination for
the day, both for the President and all other members of the Governing Council. We also include
a dummy variable which takes a value of 1 on the day of the "whatever it takes" speech, which is
known for being a particularly influential communication.
Control variables are included to limit the likelihood of any omitted variable bias in the esti-
mated coefficients for our variables of interest. We include a set of inflation and macro-economic
news surprises for the Eurozone and for major European countries. All of the 17 variables included
for that purpose are defined as the difference between the announced value and the median expec-
tation, as surveyed by Bloomberg, and detailed in Appendix D. We include variables to capture
financial uncertainty (the level of the VSTOXX index) and fiscal stress (change in sovereign spread
of peripheral economies relative to Germany). We also consider the changes in the EUR/USD ex-
change rate and in the EuroStoxx 50 to control for changes in general financial market conditions.
Descriptive statistics of the main variables are provided in Appendix F.
In addition to the above stated variables, we add four lags of the dependent variable (J= 4 ).
We do so both to limit potential auto-correlation in the residuals and to inform on the persistence
of media sentiment.18 The estimates are performed considering a daily frequency.19 In our baseline
16They follow the standard of the related literature and use Reuters news to get the raw communications. See
Gertler and Horvath (2018) for more details.
17Note that all the results related to the variables available for 2006–2010 are unchanged when we consider the
whole period (2006–2016).
18The fourth first lags only appeared significant in most of our regressions. Without the lags, the Breusch-Godfrey
LM and the Durbin’s alternative tests for autocorrelation both reject the null hypothesis of no autocorrelation at
the 1% level, when we consider autocorrelation of order 1 or higher (up to an order of 4).
19The sentiment value on Mondays is set at the average value over Saturday-Sunday-Monday. For days without
12
Table 2: Variables description
Variables Description
ECB decisions
iAnnounced changes in the ECB key interest rate.
Positive interest rate surprise Difference between the actual interest rate decided during the press con-
ference and the market expected policy rate (Bloomberg survey median)
(0 if the difference should be negative).
Negative interest rate surprise Difference between the actual interest rate decided during the press con-
ference and the market expected policy rate (Bloomberg survey median)
(0 if the difference should be positive).
APP Dummy variable equal to 1 if an unconventional monetary policy is
announced. Detailed list in Appendix D.
Negative rate Dummy variable equal to 1 the day a negative rate on the deposit facility
is announced.
Whatever it takes Dummy variable equal to 1 the day of Draghi’s "whatever it takes"
speech at the Global Investment Conference (July 26, 2012).
ECB communications
President MP - Press Conferences Net monetary policy (MP) inclination of the ECB President speech
during the Press Conference (PC), measured using Picault and Renault
(2017).
President EC - Press Conferences Net economic outlook view (EC) inclination of the ECB President press
conference speech, measured using Picault and Renault (2017).
President MP - Inter Meeting Net monetary policy (MP) inclination of the ECB President communi-
cations in the inter-meeting period, manually coded, based on Gertler
and Horvath (2018).
President EC - Inter Meeting Net economic view (EC) inclination of the ECB President communi-
cations in the inter-meeting, manually coded, based on Gertler and
Horvath (2018).
Other members MP - Inter Meeting Net monetary policy (MP) inclination of the members of the Govern-
ing Council inter-meeting (IM) communications (excluding President),
manually coded, based on Gertler and Horvath (2018).
Other members EC - Inter Meeting Net economic view (EC) inclination of the members of the Govern-
ing Council inter-meeting (IM) communications (excluding President),
manually coded, based on Gertler and Horvath (2018).
Macro-economic surprises Difference between the announced macroeconomic data and its median
expectation (from Bloomberg’s surveys). Detailed list in Appendix D.
Financial variables
Financial uncertainty Level of the VSTOXX index
Fiscal stress Changes in the following: the average of the 10-years bond yield for
Portugal, Spain, Italy, and Greece minus the 10-years bond yield for
Germany.
EURUSD Change in the log of the average EUR / USD exchange rate (USDs per
EUR).
EUROSTOX Change in the log of the average EUROSTOXX 50 level.
13
model, we include the contemporaneous values for all our variables except for the variables related
to the inter-meeting communications, for which we consider the lag of the variables. First, this
allows us to take into account the potential delay of the treatment of the information of these
communications by journalists. These communications are indeed likely to be treated with less
importance than a regular press conference and thus be read and processed with less priority;
the influence they may have on journalists’ writings cannot be expected to always be immediately
apparent.20 Second, it allows us to avoid potential simultaneity issues. Governing Council members
arguably have some flexibility over their speeches, and they might also react to media sentiment
(or to the market developments of the day), with Governing Council members attempting to offset
a more negative media sentiment (or an adverse market development) by communicating more
positively.
We perform our baseline estimates using standard OLS21 with Newey-West standard errors to
account for heteroskedasticity and potential remaining auto-correlation in the residuals.22
3.2 Results
We perform our estimates by first including only the variables related to the ECB communi-
cation, disentangling between communications of the President versus those of the other members
of the Governing Council (Model 1). We then add variables related to the ECB monetary policy
decisions (Model 2) and the different controls for the economic and financial environment (Models
3 and 4).23 Table 3 presents the results, where column ishows the estimates of Model i.
The first observation from Table 3 is that media sentiment exhibits some persistence. Higher
media optimism on a given day increases media sentiment for the following four days. This is
any sentiment observation (which represents about 1.6%of our sample), we simply linearly interpolate the missing
value. The results are not dependent on these choices.
20Blinder et al. (2008) acknowledge a potential media reaction delay to central banker communications and
present it as a challenge to identify what represents the timing of a monetary policy "event". We will more deeply
investigate this aspect in what follows.
21Standard unit root tests lead us to infer that our data are stationary.
22In the final model (including lags for the dependent variable), the results from both the Breusch-Godfrey LM
and the Durbin’s alternative tests for autocorrelation do not allow us to conclude that autocorrelation is absent
in the residuals. We chose to deal with this concern by using Newey-West standard errors, and we will consider
alternative approaches in the robustness section. Newey-West tests are performed using fixed-b critical values
and the truncation parameter rule S = 1.3 T1/2(S = 52 in our case), as recommended in Lazarus et al. (2018).
Standard Newey-West tests lead to very similar conclusions. Regarding heteroscedasticity, the Breusch-Pagan test
for heteroscedasticity rejects the null hypothesis of homoscedasticity at the 5% level (but not at the 1%), while the
White test for heteroscedasticity cannot reject the null hypothesis of homoscedasticity at the 10% level.
23Only macro-economic surprises that appeared close to significance (p-value less than 0.20) were kept in the final
estimations displayed here, which leads us to consider four surprises out of the seventeen.
14
Table 3: Results - Media Sentiment determinants
[1] [2] [3] [4]
MSIt10.295*** 0.295*** 0.293*** 0.266***
(0.033) (0.033) (0.033) (0.031)
MSIt20.070*** 0.070*** 0.071*** 0.050*
(0.027) (0.027) (0.027) (0.026)
MSIt30.044 0.044 0.045 0.030
(0.031) (0.031) (0.032) (0.031)
MSIt40.087*** 0.087*** 0.086*** 0.062***
(0.024) (0.024) (0.024) (0.024)
President MP - Press Conferencest-0.445*** -0.483*** -0.479*** -0.401***
(0.130) (0.136) (0.137) (0.122)
President EC - Press Conferencest0.542* 0.603* 0.583* 0.554*
(0.320) (0.325) (0.326) (0.306)
Other members MP - Inter Meetingt1-0.074*** -0.072*** -0.077*** -0.066**
(0.027) (0.028) (0.027) (0.029)
Other members EC - Inter Meetingt10.023 0.022 0.025 0.012
(0.044) (0.044) (0.043) (0.045)
President MP - Inter Meetingt1-0.174*** -0.174*** -0.166*** -0.166***
(0.042) (0.042) (0.043) (0.041)
President EC - Inter Meetingt10.224** 0.224** 0.224** 0.217**
(0.088) (0.088) (0.088) (0.091)
i-0.405 -0.391 -0.736
(0.529) (0.528) (0.631)
Negative Interest Surprise 1.591*** 1.589*** 1.733***
(0.491) (0.490) (0.609)
Positive Interest Surprise 1.319** 1.265** 0.829
(0.529) (0.532) (0.706)
APP -0.126* -0.120 -0.151
(0.074) (0.075) (0.142)
Negative Rate 0.196*** 0.195*** 0.166***
(0.036) (0.036) (0.037)
Whatever it takes 0.373*** 0.368*** 0.163***
(0.042) (0.043) (0.054)
Macro-economic surprises No No Yes Yes
Financial uncertainty -0.012***
(0.002)
Fiscal stress -0.135**
(0.052)
EUROSTOX 3.352***
(0.920)
EURUSD 1.635
(1.612)
Constant -0.044** -0.044** -0.044** 0.241***
(0.018) (0.018) (0.018) (0.049)
Adjusted R20.152 0.150 0.152 0.188
Obs. 1630 1630 1630 1630
Note: Newey-West Standard Errors in parenthesis. ***, **, and * represent statistical significance at respectively
1%, 5%, and 10%. M SI is the Media Sentiment Index. All control variables are defined in Table 2.
15
important in that it implies that any factor that affects media sentiment on a given day can be
expected to affect media sentiment with a certain degree of persistence.
Regarding the ECB press conference communications, the tonality of President speeches during
the press conference significantly affects media sentiment. Both the economic outlook tone and
monetary policy inclination are found to have an impact on media sentiment, though the coefficient
associated to the former variable is statistically significant only at the 10% level. A more hawkish
tone during the press conference is significantly associated with a decrease in media sentiment,
while a more positive communication on the economic outlook during the press conference is
associated with an increase in media sentiment. The effect actually appears stronger when we add
controls for monetary policy decisions (column 2 with respect to column 1).
Regarding the inter-meeting communications, we find that communications of the President,
both on the economic outlook and the future stance of monetary policy, significantly affect media
sentiment. More dovish communications are associated with a decrease in media sentiment, and
positive communications on the economic outlook increase media sentiment. Communications of
other Governing Council members on the future monetary policy stance also appear to matter for
media sentiment. The effect consistently goes in the same direction as the effect of the communi-
cations of the President (but it appears to be about three times lower) in our last estimate (Model
4). Inter-meeting communications about the economic outlook of the members of the Governing
Council other than the President are not significantly linked to media sentiment on the next day in
our regressions. The fact that we find a lower effect for the communications of Governing Council
members other than the President on the monetary policy stance and no effect for their com-
munications on the economic outlook may indicate that communications from Governing Council
members other than the President are perceived as less important than communications from the
President. It could be also that these differences reflect a difference in media attention. These two
interpretations are naturally difficult to isolate from each other. They should also be taken with
caution, insofar as standard Wald tests reject the null hypothesis of equality of the coefficients only
when we consider those associated with the variables related to communications on the economic
outlook.24
Regarding the monetary policy decisions, we find that interest rate surprises affect media
sentiment. This is in line with the results of Berger et al. (2011) who found that the tone of news
reports is generally more negative when the policy decision surprises the financial market analysts.
We robustly detect this effect only for negative interest rate surprises; when the decided policy
24For the coefficients of the variables related to communications on the monetary policy stance, the p-value of
the standard Wald test is about 0.12.
16
rate is lower than the market expectation, media sentiment decreases. We do not find an effect
for positive interest rate surprises when we control for financial variables (Model 4), although this
could arguably be due to the imperfect nature of such controls in this case (which may also react
to the interest rate surprise). As in Berger et al. (2011), interest rates moves per se do not appear
to matter. The announcements of asset purchase programs also do not appear to systematically
affect media sentiment in our baseline estimates. However, we find that the adoption of a negative
rate in June 2014 and the "whatever it takes" speech of Mario Draghi are both associated with a
significantly important increase in media sentiment.
In terms of economic magnitude, further analysis reveals that events related to the press confer-
ence usually have more impact on media sentiment than standard inter-meeting communications.
For example, based on column (4), a one standard deviation increase in the tone of the press
conference in terms of hawkishness decreases media sentiment by about 0.13. This is about a
fourth of the MSI standard deviation on the analyzed period, and it could make the MSI move
from one decile to another. The effect is of comparable magnitude to the effect of a one standard
deviation decrease in the tone of the press conference regarding the economic outlook (0.09 de-
crease in MSI). A negative interest rate surprise of 25 basis points is associated with a decrease
in media sentiment of about 0.44. As a comparison, a one standard deviation increase in the VS-
TOXX leads "only" to a 0.08 decrease in media sentiment. Direct and contemporaneous effects of
events linked to the press conference on the MSI thus appear of particularly important magnitude
when compared to the effect that alternative broader variables have on the MSI. Standard moves
in the tone of inter-meeting communications impact media sentiment with less magnitude; a one
standard deviation increase in the tone of President communications related to monetary policy
inclination decreases media sentiment from about 0.03, an effect almost similar to a one standard
deviation decrease in the tone of President communications related to economic outlook (0.02 de-
crease in MSI). Exceptional communications can, however, have very important impacts on media
sentiment. Considering column (3), the "whatever it takes" speech contemporaneously moved the
MSI index from about 0.37, about five times more than a one standard deviation increase in the
VSTOXX.
While such an analysis sheds light on the direct and contemporaneous effect (or next-day effect
for inter-meeting communications) of our variables on media sentiment, in Appendix E, Table 7,
we analyzed and took into account the potential effect of ECB communications on media sentiment
up to 4 days after the communication by including more lags for our variables of interest in the
regression. The tone of the press conference in terms of monetary policy inclination appears to
significantly move media sentiment several days after the press conference. A much larger effect
17
of the President speech on media sentiment is unveiled through this analysis. On average, a one
standard deviation increase in the tone of the press conference in terms of hawkishness will have
decreased media sentiment by about 0.64 after five days,25 and thus it will have decreased media
sentiment by more than a standard deviation. Other newly included lags are usually not found to
affect media sentiment, suggesting that the ECB communications move media sentiment only at
the timing we initially set.26
3.3 Robustness and interpretation of the results
The most important result that arose in the previous section is that both press conference
and inter-meeting communications significantly affect media sentiment. This is especially true for
those related to the monetary policy inclination, with hawkish (dovish) communications decreasing
(increasing) media sentiment. In this section, we test for its robustness and further discuss its
interpretation.
To test the robustness of our results, we performed a set of alternative regressions, shown in
Table 8 of Appendix E. We first performed the same estimates with a standard Huber-White ma-
trix for the residuals. All the coefficients (including the ones related to monetary policy decisions)
either maintained the same significance level or saw it increase when we did so (column (1)). We
then performed the same estimates but included the lag of all the control variables instead of
their contemporaneous value.27 The coefficients for the variables related to inter-meeting commu-
nications as well as their associated statistical significance were barely affected (column (2)). We
then dealt with the potential remaining autocorrelation in the residuals with alternative methods.
25The figure is computed based on the estimates of column (4) of Table 7 and takes into account the persistence
of the MSI.
26An exception is the second lag of the variable capturing the tonality of the inter-meeting communications of the
President regarding the economic outlook, which is statistically significant at the 5% level, while being negative.
This suggests that the positive effect of these communications on sentiment reverses over the next day. A standard
Wald test cannot reject the hypothesis that this coefficient is equal to the opposite of the coefficient related to
the first lag of the same variable, at conventional statistical significance levels. In addition, the second lag of the
variable capturing the tonality of the inter-meeting communications of the Governing Council members other than
the President regarding the economic outlook appears statistically significant at the 5% level in all regressions
(columns (1) to (4)). This suggests that these also affect media sentiment, but with more delay than those of the
President, possibly owing to a difference in media attention. Regarding policy decisions, the "whatever it takes
speech" is also found to positively affect media sentiment the day after it arouse, and the asset purchase programs
are found to positively impact media sentiment the day after as well as three days after their announcements
(coefficient statistically significant at the 1% level). The MSI is also found to decrease three days after a positive
interest rate surprise, reinforcing the interpretation that interest rate surprises negatively affect media sentiment.
These latter results are all available on request.
27Compared to the previous specification including several lags for all variables, such a specification has the
advantage of being much more parsimonious.
18
First, we kept three lags for each variable related to ECB communications and ECB decisions,
adding only the lags of the control variables which appeared statistically significant (up to three
lags). Such a specification has the advantage of being more parsimonious than the ones previously
tested on Table 7, while autocorrelation tests cannot here reject the null hypothesis that there is
no autocorrelation in the residuals at a 10% significance level.28 Column (5) of Table 7 presents
the results with a Huber-White matrix for the residuals, and column (6) presents the results with a
standard covariance matrix. The previous key results are unaffected with the former specification,
while with the second specification the statistical significance of the coefficients of the variables
related to Governing Council communications and the variable capturing the tone of the President
on the economic outlook during the press conference now slightly overpasses the conventional 10%
level. In fact, the variance of most coefficients increases with the second specification, and some
variables for which the impact on media sentiment may seem self-evident to some observers (such
as the dummy for the "whatever it takes" speech or the fiscal stress variable) also lose statistical
significance at the 10% level. Second, we repeated the same estimates without the lags for the de-
pendent variable. While these were included to limit serial correlation in the residuals, remaining
serial correlation together with the presence of the lags could bias downward the other coeffi-
cients (Wilkins, 2018). We ran the baseline regression without the lagged dependent variables,
using respectively OLS with Newey-West standard errors, OLS with a Huber-White matrix for
the residuals, the Prais–Winsten estimator, and an ARMAX model estimated through maximum
likelihood; the conclusions remained broadly unchanged.29
To assess the robustness of our results to alternative sentiment measures, we built our index
with other dictionary approaches and then performed the same regressions. We first included only
the negative words when building the sentiment score. In a second specification, we selected the
positive and negative words used to build our sentiment measure in equation 1 from a generic
well-known dictionary, the Harvard-IV dictionary. In a third specification, we used the Vader
sentence-level classifier.30 Though the Vader approach and Harvard-IV dictionary are not built
specifically for a financial context, using them here allows us to have an idea of the extent to
28Durbin’s alternative tests for autocorrelation (Breusch-Godfrey LM tests for autocorrelation) of the order 1 till
5 all reject the null hypothesis at the 20% threshold (at the 18% threshold).
29In all specifications, the coefficients of the variables related to ECB communications that were statistically sig-
nificant at the 5% level remained significant at this level, with the exception of the communications of the Governing
Council members (other than President) related to monetary policy inclination with the ARMAX specification. The
coefficient of the variable related to the tone of the President on the economic outlook during the press conference
was statistically significant at the 5 or 10% level in all specifications but the Prais-Winsten specification. The results
are available on request.
30See Shapiro et al. (2020) for a general discussion on Vader and on the Harvard-IV dictionary.
19
which our results are dependent on the dictionary we used. Results are displayed in columns
(3), (4), and (5) of Table 8. Most key findings remain irrespective of the approach chosen. The
press conference and inter-meeting communications of the President related to monetary policy
inclination are still found to affect media sentiment, with coefficients statistically significant at
1% levels except with the Harvard IV-4 dictionnary. The same applies for the inter-meeting
communications of other Governing Council members, related to monetary policy inclination. The
effect of all communications related to the economic outlook cannot however be seen as robust
when using the Harvard-IV and Vader dictionary.31 In a final robustness check, we gave a different
weight to each newspaper article based on a proxy for an audience of professional investors.32 The
last column of Table 8 shows that the results are mostly unaffected by this change.
Focusing on the interpretation of our results, we then test for potential non-linearities in the
effect of communications on sentiment. One might expect that hawkish communications decrease
media sentiment only when a restrictive monetary policy seems unwarranted (e.g, when inflation
is below its target), and that they should be positively associated with sentiment only when such a
policy seems warranted. Evidence of such effects would affect the interpretation of our main result.
To test this hypothesis, we create a dummy variable equal to 1 when CPI inflation is above the ECB
2% inflation target (0 otherwise) and we interact each central bank communication variable with
this dummy variable. We add both the dummy variable and the interaction terms to equation
3. Detailed results are presented in Table 9 in Appendix E. The results do not strongly point
out the presence of non-linearity in the effect of communications; none of the interactive terms
are significant at the 5% level. This confirms the interpretation of the results of the previous
section: more hawkish communications make media content more pessimist, and more dovish
communications make media content more optimist, unconditional on inflation being above or
below the target.33 While such an interpretation may appear puzzling to some observers, it is
also consistent with some recent central banker comments. Powell (2018), for example, recognized
that "looser policy leading to more positive sentiment in markets and tighter policy depressing
sentiment" was a possible simple link between monetary policy and risk sentiment: our results
31The fact that the adjusted R2is lower for the estimates performed with these three alternative measures (relative
to our initial measure) can be seen as a factor justifying the relevance of our initial measure choice.
32A key issue here is that there is no standardized data on the worldwide audience for newspapers. We used data
from Erdos &Morgan’s Professional Investment Community Study of 2010. We computed the daily sentiment score
for each newspaper and then built a daily weighted average, using a weight of 0.59 for The Wall Street Journal,
0.35 for The Financial Times, 0.28 for Barron’s, 0.24 for The New York Times, 0.12 for The Times.
33Note that, when performing the same estimates with the available data for 2006-2016, we get the same results:
the tone of press conference communications related to monetary policy inclination is linearly and significantly
associated to media sentiment, with no evidence of non-linearity.
20
suggest that a similar direct link applies for media sentiment on monetary policy.
4 Media sentiment and inflation expectations
We now focus on the relevance of media sentiment for financial markets’ long-term inflation
expectations.
4.1 Empirical strategy
It is usually considered that any factor affecting financial market long-term inflation expecta-
tions does so because the central bank is not perceived either as completely willing or as completely
able to keep inflation in line with its target in the future.34 The fact that financial market ex-
pectations for Euro area inflation in five years for the next five years have been falling below the
ECB inflation target since 2014 (Draghi, 2014) can be interpreted as a sign that investors see the
ECB as less able to keep inflation in line with its 2 percent target. If media sentiment is, as one
may expect, correlated with investor perceptions, the MSI gives a unique opportunity to directly
test this hypothesis that long-term inflation expectations do respond to investor perceptions on
monetary policy and its relevant environment. Following this intuition, we test whether daily vari-
ations in the MSI affect financial markets’ long-term inflation expectations. We adopt a framework
close to Tetlock (2007) and Garcia (2013) who analyze the effect of investor sentiment proxied by
media content on the stock returns of the next day, but we focus our attention on financial market
inflation expectations.
We measure inflation expectations using daily data on inflation-linked forward swap rates in the
Euro area from Thomson Reuters (see e.g, Garcia and Werner, 2018, Ambler and Rumler, 2019, or
Beechey et al., 2011 for a similar measure). We focus our analysis on the 5-year forward inflation
compensation five years ahead (5-y to 5-y forward rate henceforth) and 10-year forward inflation-
linked swap rate (10-y forward rate henceforth), insofar as they represent inflation expectations for
the long-term and as such are of primary relevance to policymakers. The 5-y to 5-year forward rate
is often seen as the most relevant measure of financial market inflation expectations in the Euro
area and it receives a considerable attention from policymakers, central bankers (Draghi, 2014),
34For example, in Beechey et al. (2011), the fact that long-term inflation expectations respond to inflation surprises
in the US while they do not in the Eurozone before 2008 can be interpreted as evidence that the ECB is perceived
as more willing to keep inflation at a target than the FeD. Garcia and Werner (2018) note that the lack of evidence
that long-term inflation expectations responded to macroeconomic news before the global financial crisis for the
ECB was interpreted as a sign that the central bank was perceived as credible in achieving its inflation target (thus
willing or able to achieve it).
21
and market participants (Rennison, 2019).35 Inflation expectations daily changes in basis points
(fm
tfm
t1) for inflation expected in five years for the next five years (m= 5/5) or for the average
inflation expected for the next 10 years (m= 10/0) are assumed to be linked to media sentiment
through the following model:
fm
t+1 fm
t=c+
n
X
i=0
βi(fm
tifm
ti1) +
n
X
i=0
φiMSIti+
m
X
i=0
αiXtm+σdowt+t(4)
where cis a constant, Xta set of exogenous variables defined thereafter, dowtare days-of-the-
week dummy variables, tis the error term. We follow Natoli and Sigalotti (2018) to deal with
persistence and heteroskedasticity common with such financial variables and model the squared
variance σ2
tusing a GARCH(1,1) model given by:
σ2
t=γ0+γ12
t1+γ2σ2
t1(5)
Xt= (EC B_decisionst,EC B_commt,Macro_surprisest,F inancialst,N ews_Sent) is the
set of exogenous variables considered, where ECB_decisionst,ECB_commt,M acro_surprisest
are, respectively, the ECB monetary policy decisions, the ECB communications, and the macroe-
conomic surprises. In our context, it is particularly important to control for central bank commu-
nications, central bank decisions, and macroeconomic surprises. Without doing so, an apparent
effect of sentiment on inflation expectations could be a reflection of these factors. To better purge
the effect of economic news broadly defined, we also include two sentiment indexes of (domestic
and international) economic news through the vector News_Sentt: the index calculated by Bor-
toli et al. (2018) from the content of the French newspaper Le Monde, and the index of Shapiro
et al. (2020) focusing on the US. For a similar purpose, we will also consider the financial variables
previously defined F inancialst.
The coefficients in each regression are estimated by Maximum Likelihood with the same sample
period as in the previous sections (January 1, 2010 until May 31, 2016), taking only trading days
into consideration. We consider n= 5 lags for the dependent variable and the MSI. The model
is first estimated without the Xvector of controls, and these are then added with lags (m) chosen
with parsimony considerations so that they match any potential statistically significant lag of our
variable of interest.
4.2 Results
Results are displayed in Table 4.
35The five- and ten-year maturities are also considered to concentrate a significant amount of liquidity relative
22
Table 4: 5Y-5Y and 10Y Inflation Expectations and Media Sentiment
5Y-5Y 5Y-5Y 5Y-5Y 10Y 10Y 10Y
[1] [2] [3] [4] [5] [6]
fm
tfm
t10.094*** 0.089*** 0.081*** 0.152*** 0.145*** 0.145***
(0.026) (0.026) (0.027) (0.027) (0.028) (0.029)
fm
t1fm
t20.074*** 0.076*** 0.082*** 0.030 0.038 0.038
(0.026) (0.027) (0.027) (0.028) (0.030) (0.030)
fm
t2fm
t3-0.019 -0.031 -0.029 0.010 -0.001 0.001
(0.028) (0.028) (0.028) (0.029) (0.030) (0.030)
fm
t3fm
t4-0.015 -0.010 -0.015 0.013 0.014 0.007
(0.026) (0.026) (0.026) (0.026) (0.027) (0.027)
fm
t4fm
t5-0.031 -0.024 -0.025 -0.041 -0.034 -0.036
(0.026) (0.026) (0.026) (0.028) (0.028) (0.028)
MSIt10.202** 0.216*** 0.186** 0.209*** 0.234*** 0.231***
(0.083) (0.082) (0.087) (0.074) (0.075) (0.084)
MSIt2-0.225** -0.217** -0.217** -0.129 -0.133* -0.161**
(0.088) (0.089) (0.090) (0.079) (0.078) (0.079)
MSIt3-0.038 -0.023 -0.010 -0.109 -0.108 -0.080
(0.092) (0.094) (0.093) (0.082) (0.083) (0.085)
MSIt4-0.013 -0.005 -0.006 -0.005 0.005 0.008
(0.090) (0.093) (0.094) (0.071) (0.075) (0.077)
MSIt50.029 -0.011 0.003 0.040 -0.007 0.008
(0.082) (0.083) (0.085) (0.083) (0.082) (0.083)
Constant -0.058 0.145 0.211 -0.060 0.225* 0.119
(0.042) (0.136) (0.240) (0.043) (0.129) (0.235)
ARCH
ARCH(1) 0.058*** 0.056*** 0.058*** 0.098*** 0.107*** 0.094***
(0.009) (0.008) (0.009) (0.010) (0.011) (0.011)
GARCH(1) 0.941*** 0.944*** 0.942*** 0.898*** 0.891*** 0.904***
(0.008) (0.008) (0.008) (0.010) (0.010) (0.010)
Constant 0.013* 0.010 0.011 0.041*** 0.038*** 0.029***
(0.007) (0.006) (0.007) (0.010) (0.010) (0.009)
Day-of-the-week Yes Yes Yes Yes Yes Yes
ECB Press conference communications No Yes Yes No Yes Yes
ECB Inter-meeting communications No Yes Yes No Yes Yes
Monetary Policy Decisions No Yes Yes No Yes Yes
Macro. Surprises No Yes Yes No Yes Yes
News-sent No Yes Yes No Yes Yes
Financial Variables No No Yes No No Yes
Log likelihood -3244.551 -3221.753 -3209.509 -3176.578 -3149.496 -3138.629
Obs. 1572 1572 1572 1572 1572 1572
Note: Standard Errors in parenthesis. ***, **, and * represent statistical significance at respectively 1%, 5%,
and 10%. 5Y-5Y (10Y) is the 5-y to 5-y (10 years) forward rate from inflation-linked swaps. Day-of-the-week are
dummy variables for each day of the week, News-sent are the economic news indexes of Shapiro et al. (2020) and
Bortoli et al. (2018). M SI is the Media Sentiment Index. All other control variables are defined in Table 2.
to all other inflation swap instruments. See Garcia and Werner (2018) for a thorough discussion.
23
Columns (1) and (4) correspond to the estimates of equation 4 for our two inflation expectations
measures, without the vector of controls X. We find in both cases that the first lag of our sentiment
measure has a positive and significant effect on next day inflation expectations. An exogenous
increase in our measure of media sentiment on monetary policy at day twill thus increase the 5-y
to 5-y and the 10-y inflation expectations in t+ 1. We also observe in both cases that the positive
effect at day t+ 1 seems to at least partly reverse at day t+ 2: the second lag of the MSI is
negative for both measures and statistically significant at the 1% level for the 5-y to 5-y inflation
swap measure. We also find that changes in inflation expectations exhibit some persistence with a
statistically significant coefficient for the first lag for both measures, while the second lag appears
to be statistically significant for the 5-y to 5-y inflation swap. This is not surprising in itself, as
we are dealing with inflation swaps, which are over-the-counter instruments, where daily changes
are relatively small and arbitrage opportunities limited. In columns (2) and (5), we added the first
two lags for our control variables (m= 2) except financial variables. Doing so only slightly affects
the magnitude of our coefficient estimates but renders the second lag of the MSI now statistically
significant at the 10%confidence level when the dependent variable is the change in the 10-year
inflation expectations (column (5)). Lastly, in columns (3) and (6), we included financial variables
in the regressions to make sure our results are not driven by any other market relevant information
that would matter for next-days inflation expectations. The results remained unchanged.
4.3 Robustness and interpretation
In separate estimates36, we used a GARCH(2,2) to model the variance in equation 5 and
obtained similar results. We also tried to estimate equation 4 directly through OLS with Newey-
West or Hubert-White standard errors. While these showed coefficients for the MSI of slightly
higher magnitude, the key observation was not changed; the MSI was found to increase next day
inflation expectations while the effect reversed the day after. In all cases, Wald tests could not
reject the null hypothesis that the coefficient associated with the first lag of the MSI was equal to
the opposite of the one associated with the second lag of the M SI , thereby clearly suggesting the
presence of a reversal pattern. We also repeated the estimates with Jordà (2005)’s local projections
and reached the same conclusion. We did not find evidence of non-linearities, when considering a
different impact depending on whether inflation is below or above its target, or for moves of the
MSI at press conference days.
The effect we unveiled in this section is of low magnitude when compared to stocks or exchange
36All results are available upon request.
24
rates daily changes: a positive two-standard deviations increase in the MSI , which could make
the MSI move from 6 deciles, is found to increase the next-day 5-y to 5-y (10-y) forward rate
from less than a basis point, namely around 0.20 bps (0.24 bps). It has to be recalled, however,
that this concerns long-term inflation swaps, for which the daily changes are usually small, and
which, if anchored, should barely respond to news. In addition, we are analyzing an impact on the
next trading day. The magnitude we find is in line with previous findings on the effects of central
banks’ communications or macroeconomic surprises on inflation swaps. Beechey et al. (2011)
find that long-term inflation expectations (nine to ten years ahead inflation compensation) were
unresponsive to macroeconomic surprises in the Euro area before the global financial crisis, while
the effects found in Garcia and Werner (2018) for macroeconomic surprises after 2013 are typically
less than a basis point, with, for example, German and Spanish surprises in flash estimates leading
to a change in 5-y to 5-y inflation compensation of around 0.6 and 0.5 basis point in the day of
their release on average over their sample period. The observed response for the change in the 5-y
to 5-y forward rate is still about thrice its unconditional mean from our sample period (-0.082),
and is as such comparable in magnitude to the results obtained in Tetlock (2007).37 As in Tetlock
(2007) or Garcia (2013), the effect of a sentiment shock reverses, but quicker in our case. The
small effect we find suggests nonetheless that media sentiment on monetary policy is of limited
economic relevance to understand financial market’s inflation expectations.
5 Conclusion
Communication is a key part of monetary policy. While there is a clear consensus that com-
munications from central bankers impact financial markets, little is known about the role of media
coverage in this transmission. In this paper, we shed light on this particular point.
We built a new measure of the media sentiment on the ECB monetary policy and its relevant
environment, using advanced natural language processing methods with a database of about 25,000
newspaper articles.The resulting index is, to our knowledge, the first index measuring media sen-
timent on the ECB monetary policy and its relevant environment, on a continuous basis. A key
innovation we have introduced is the purging of each article from news re-transcription, to capture
journalist interpretation. We used a rich external dataset on inter-meeting communications to
analyze the determinants of our index. We found that media sentiment is affected by the content
37In Tetlock (2007), the effect of a two-standard deviations change in his sentiment index on the next day’s Dow
Jones returns is 16.2 basis points, that is, a bit more than thrice the unconditional mean of Dow Jones returns in
his sample period (5.4 basis points).
25
of the press conference and by inter-meeting interventions of the President as well as interventions
of other members of the Governing Council related to monetary policy inclination. We analyzed
the relevance of our index for inflation expectations and found that media sentiment on monetary
policy has predictive power for the daily changes in the 5-years to 5-years inflation swap rate as well
as for the 10-years inflation swap rate. The magnitude of the effect we found do not suggest that
media sentiment on monetary policy is economically important to understand financial markets’
inflation expectations.
Anecdotal evidence already suggests that media coverage matters for central bankers. For ex-
ample, in 2013, the ECB signed a contract to monitor print media, online media, broadcast, and
social media and to perform a media analysis/reputation tracking.38 The ECB itself announced on
its website that "the conclusions drawn from [this] analysis help shape the ECB’s communication
strategy." This paper provides an empirical analysis unveiling the effect of different kinds of com-
munications and decisions on media coverage. It stresses in particular the key importance of the
communications of the President, especially those during the press conference related to monetary
policy inclination. In addition, it provides empirical evidence suggesting that media sentiment also
matters in the context of central banking and that the economic importance of media sentiment
is negligible in this context. We find a reversal pattern for the effect of sentiment on next-days
inflation expectations, as observed in Tetlock (2007) or Garcia (2013) in another context.
Naturally, our analysis suffers from the traditional limits of text analysis methods, which could,
in turn, lead us to underestimate the magnitude of the media effect. Qualitative differences between
words, sentences and articles are in particular neglected. Furthermore, although we only selected
media known for their analysis (in contrast to newswire services), removed all re-transcriptions,
controlled for several news variables and several indexes of central bank communications, and
looked at the relationship with next-day inflation expectations rather than contemporaneous rela-
tions, we cannot firmly eliminate the possibility that the effect may be driven by news. In further
research, it would be interesting to assess with alternative methods the magnitude of the media
channel and to investigate the links between media sentiment on monetary policy and households
or business’ expectations.
38Germany-Frankfurt-on-Main: ECB - Provision of media monitoring and analysis services 2013/S 170-294141
26
References
Ambler, S. and Rumler, F. (2019). The effectiveness of unconventional monetary policy announce-
ments in the euro area: An event and econometric study. Journal of International Money and
Finance, 94(C):48–61.
Baker, M. and Wurgler, J. (2006). Investor sentiment and the cross-section of stock returns. The
Journal of Finance, 61(4):1645–1680.
Bauer, M. D. (2015). Inflation Expectations and the News. International Journal of Central
Banking, 11(2):1–40.
Beaupain, R. and Girard, A. (2020). The value of understanding central bank communication.
Economic Modelling, 85:154–165.
Beechey, M. J., Johannsen, B. K., and Levin, A. T. (2011). Are long-run inflation expectations
anchored more firmly in the Euro Area than in the United States? American Economic Journal:
Macroeconomics, 3(2):104–29.
Bennani, H. (2018). Media coverage and ECB policy-making: Evidence from an augmented Taylor
rule. Journal of Macroeconomics, 57:26 – 38.
Berger, H., Ehrmann, M., and Fratzscher, M. (2011). Monetary policy in the media. Journal of
Money, Credit and Banking, 43(4):689–709.
Blinder, A. S., Ehrmann, M., Fratzscher, M., De Haan, J., and Jansen, D.-J. (2008). Central bank
communication and monetary policy: A survey of theory and evidence. Journal of Economic
Literature, 46(4):910–45.
Böhm, J., Král, P., and Saxa, B. (2012). The Czech National Bank’s monetary policy in the media.
European Journal of Political Economy, 28(3):341–357.
Bortoli, C., Combes, S., and Renault, T. (2018). Nowcasting GDP growth by reading newspapers.
Economie et Statistique, 505(1):17–33.
Bundick, B. and Smith, A. L. (2018). Does Communicating a Numerical Inflation Target Anchor
Inflation Expectations? Evidence & Bond Market Implications. Research Working Paper RWP
18-1, Federal Reserve Bank of Kansas City.
Draghi, M. (2014). Speech at Jackson Hole Conference, August 22, 2014. Economic Policy
Symposium.
Ehrmann, M. and Fratzscher, M. (2007a). Communication by central bank committee members:
Different strategies, same effectiveness? Journal of Money, Credit and Banking, 39(2-3):509–541.
Ehrmann, M. and Fratzscher, M. (2007b). The timing of central bank communication. European
Journal of Political Economy, 23(1):124–145.
27
Fraiberger, S. P., Lee, D., Puy, D., and Ranciere, R. (2021). Media sentiment and international
asset prices. Journal of International Economics, 133:103526.
Galati, G., Poelhekke, S., and Zhou, C. (2011). Did the Crisis Affect Inflation Expectations?
International Journal of Central Banking, 7(1):167–207.
Garcia, A. J. and Werner, S. (2018). Inflation news and euro area inflation expectations. IMF
Working Papers, 18:1.
Garcia, D. (2013). Sentiment during recessions. The Journal of Finance, 68(3):1267–1300.
Gertler, P. and Horvath, R. (2018). Central bank communication and financial markets: New
high-frequency evidence. Journal of Financial Stability, 36(C):336–345.
Hayo, B. and Neuenkirch, M. (2015). Self-monitoring or reliance on media reporting: How do
financial market participants process central bank news? Journal of Banking & Finance, 59:27–
37.
Hubert, P. and Labondance, F. (2020). Central Bank Tone and the Dispersion of Views within
Monetary Policy Committees. Documents de Travail de l’OFCE 2020-02, Observatoire Francais
des Conjonctures Economiques (OFCE).
Husted, L., Rogers, J., and Sun, B. (2019). Monetary policy uncertainty. Journal of Monetary
Economics.
Jansen, D.-J. and De Haan, J. (2005). Talking heads: the effects of ECB statements on the
euro–dollar exchange rate. Journal of International Money and Finance, 24(2):343 – 361.
Jansen, D.-J. and de Haan, J. (2007). Were verbal efforts to support the euro effective? A high-
frequency analysis of ECB statements. European Journal of Political Economy, 23(1):245–259.
Jordà, Ò. (2005). Estimation and inference of impulse responses by local projections. American
Economic Review, 95(1):161–182.
Kearney, C. and Liu, S. (2014). Textual sentiment in finance: A survey of methods and models.
International Review of Financial Analysis, 33:171–185.
Lazarus, E., Lewis, D. J., Stock, J. H., and Watson, M. W. (2018). HAR Inference: Recommen-
dations for practice. Journal of Business & Economic Statistics, 36(4):541–559.
Loughran, T. and McDonald, B. (2011). When is a liability not a liability? Textual analysis,
dictionaries, and 10-Ks. The Journal of Finance, 66(1):35–65.
Moessner, R. (2018). Effects of asset purchases and financial stability measures on term premia in
the euro area. Applied Economics, 50(43):4617–4631.
Natoli, F. and Sigalotti, L. (2018). Tail co-movement in inflation expectations as an indicator of
anchoring. International Journal of Central Banking, 14:1.
28
Picault, M. and Renault, T. (2017). Words are not all created equal: A new measure of ECB
communication. Journal of International Money and Finance, 79:136 – 156.
Powell, J. H. (2018). Monetary Policy Influences on Global Financial Conditions and International
Capital Flows : a speech at "Challenges for Monetary Policy and the GFSN in an Evolving
Global Economy". Speech, Board of Governors of the Federal Reserve System.
Rennison, J. (2019). Why are central banks fixated on inflation expectations ? The Financial
Times, June 26 2019.
Rosa, C. and Verga, G. (2008). The impact of central bank announcements on asset prices in real
time. International Journal of Central Banking, 4(2):175 – 217.
Schmeling, M. and Wagner, C. (2015). Does central bank tone move asset prices? Working Paper.
Shapiro, A. H., Sudhof, M., and Wilson, D. J. (2020). Measuring news sentiment. Journal of
Econometrics.
Tetlock, P. C. (2007). Giving content to investor sentiment: The role of media in the stock market.
The Journal of Finance, 62(3):1139–1168.
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.
Tobback, E., Nardelli, S., and Martens, D. (2017). Between hawks and doves: measuring central
bank communication. ECB Working Papers, (No 2085).
Wilkins, A. S. (2018). To lag or not to lag?: Re-evaluating the use of lagged dependent variables
in regression analysis. Political Science Research and Methods, 6(2):393–411.
29
Appendix A - Measuring Media Sentiment
.1 Example of media sentiment measurement for two articles
We present two examples of texts from which the MSI is built, in which we indicate:
Positive words (from LM dictionary) in blue
Negative words (from LM dictionary) in red
Exact re-transciptions of ECB members’ words removed with part-of-speech tagging crossed
out
The first article below has a sentiment score which belongs to the last decile of the MSI
distribution. The sentiment of the original text is about -2,7 (in the 7th decile of our MSI measure)
while purged out from re-transcriptions, it amounts to about -3,6 (in the last decile).
The Wall Street Journal, September 9, 2011
FRANKFURT — The ECB opened the door to interest-rate cuts if needed to bolster a weakening economic
recovery —a dramatic U-turn from its decision to raise interest rates just two months ago. Economic risks have
"intensified" to the downside with "enormous" uncertainty, ECB President Jean-Claude Trichet told reporters
after the central bank held its main policy rate at 1.5%. He called the ECB’s reassessment of the economic
outlook "significant" and highlighted weakening global growth, declines in equity markets and strains in euro-zone
government bond markets as trouble spots. At the same time, Mr. Trichet defended the ECB’s two rate hikes earlier
this year, saying they were needed to keep inflation in check. He also launched an uncharacteristically passionate
defense of the ECB’s recent decision to buy Italian and Spanish bonds, saying the moves were needed to restore
smooth transmission of the ECB’s interest-rate decisions to financial markets and the economy. We stand ready
to do whatever is necessary,” Mr. Trichet said, adding that the ECB will “monitor very closely all developments.”
Until Thursday, the ECB had only said that it would closely monitor what it saw as upside risks to inflation. Those
risks are now balanced, Mr. Trichet said. At 2.5%, annual inflation is still above the ECB’s 2%target. But ECB
staff economists expect it to fall to 1.7%in 2012, according to revised projections released Thursday. "This is an
ECB that’s worried" about the economy, said Nick Matthews, economist at Royal Bank of Scotland. "It’s clear
that the ECB has an easing bias and may be cutting rates by the end of the year" Mr. Matthews said. The
euro zone expanded just 0.7%, at an annualized rate, in the second quarter as the region’s two largest economies,
Germany and France posted little or no growth. A collapse in business and consumer sentiment last month raised
a grimmer prospect: The euro zone may be headed toward another recession. The risks of renewed contraction in
the world’s advanced economies "has gone up" the Organization for Economic Cooperation and Development said
Thursday. The Paris-based organization of developed and large developing economies urged central banks with the
"scope" to do so to consider easing rates. But recent data suggest some of the gloom may be overblown in Europe.
German industrial production jumped 4%in July from June, though business sentiment slid that month. Consumer
spending in the euro zone rose in July. The economy is softening but isn’t falling back into recession yet, analysts
say. Euro-zone gross domestic product "is expected to increase very moderately in the second half of this year" Mr.
Trichet said. ECB staff cut their 2012 growth forecast to 1.3%from 1.7%. If the economy stabilizes even at weak
growth rates, then the ECB will likely keep interest rates steady, some economists said. "It may take contraction
for them to actually cut rates" said Howard Archer, economist at consultancy IHS Global Insight. The ECB raised
rates in April and again in July even as other major developed-country central banks such as the Federal Reserve,
the Bank of England and the Bank of Japan held rates at crisis-era lows closer to zero. The Bank of England voted
Thursday to keep rates at a record-low 0.5%, but didn’t announce any plans to resume its asset-purchase program,
known as quantitative easing. The ECB’s sudden shift opens it to renewed criticism that—as in July 2008 when
30
it increased rates just weeks before the collapse of Lehman Brothers—officials didn’t recognize early warning signs
on the economy and overestimated inflation risks, exacerbating the slowdown. The ECB made the right call, Mr.
Trichet said Thursday. "We think what we did was appropriate" he said, referring to this year’s rate hikes. "We
preserved a solid anchoring of inflation expectations." Mr. Trichet vigorously defended the ECB’s decision, made
one month ago, to reactivate its government bond purchase program and expand it to include Italian and Spanish
bonds. The ECB has bought more than 50 billion in government bonds since restarting the program, which had
been dormant for four months. That provoked a backlash in Germany, where government bond purchases by central
banks are seen as an inflationary taboo that puts central bankers in the realm of fiscal policy. The head of Germany’s
central bank, Jens Weidmann, voted against reactivating the purchases. German President Christian Wulff, whose
position is largely ceremonial, has called the ECB’s bond purchases ¨politically and legally questionable.¨
The head
of German’s center-left SPD party, Sigmar Gabriel, has also blasted the move. His voice rising uncharacteristically,
Mr. Trichet said the ECB has displayed an “impeccable, impeccable” adherence to price stability during its 12-year
history. “I would very much like to hear the congratulations for an institution that has delivered price stability in
Germany...which is better than has ever been achieved in this country” he said, responding to a question about Mr.
Gabriel’s criticism. Mr. Trichet repeated his insistence that euro-zone parliaments approve changes to Europe’s
440 billion rescue fund that would allow it to purchase bonds in financial markets, alleviating the ECB of that task.
The ECB President dismissed concerns that the euro zone suffers from a lack of liquidity in its banking system, and
rejected claims that the region’s banks are in urgent need of more capital. International Monetary Fund managing
director Christine Lagarde caused a stir last month when she said European banks needed urgent recapitalization.
“I will not dramatize the situation as has been done by some” Mr. Trichet said, without referring specifically to Ms.
Lagarde’s comments.
The second article below has a sentiment score which belongs to the first decile of the MSI
distribution (very positive). Its score is about 0.017.
The Financial Times, September 4, 2013
Watch out for some upbeat comments - maybe even a joke? - from Mario Draghi today. Plenty has gone
right recently for the ECB president - with the economic data, that is, not with the politics in his native Italy.
Wednesday’s purchasing managers’ indices showed the eurozone’s recovery becoming firmly established. Unlike
previous, German-led upturns the return to solid expansion territory is broad-based. Much of the improvement
has been in Italy and Spain. The PMIs are watched closely by financial markets, and the turnround explains why
European shares have outperformed US shares over the past month. Tensions over Syria have knocked Europe’s
equity rally recently. And the US Federal Reserve’s plan to scale back its asset purchases, or quantitative easing,
could become a big problem for the eurozone and the ECB if it results in tighter monetary conditions in Europe.
But the strength of equity rallies in countries such as Greece - still up almost 20 per cent year to date - highlight
the shift in investor sentiment towards Europe’s monetary union. The big exception? Italy’s shares are back down
to early January levels. That is hard to explain on economic grounds alone. True, Spain has made more progress in
improving competitiveness by reducing unit labour costs. But Italy’s PMIs are almost as strong as Spain’s. Both
have seen modest improvements in credit conditions recently amid clear signs that the eurozone ¨
fragmentation¨story
- which resulted in widely diverging borrowing costs across the region - has gone into reverse. The only explanation
is that markets remain spooked by Italian politics - especially threats that Silvio Berlusconi could bring down the
administration led by Enrico Letta. As long as Italy’s government looks wobbly, investors will fret about reform
prospects, and Mr Draghi will keep the country on his worry list.
31
.2 Top 20 word occurrences
Positive Negative
Word Occurrences Word Occurrences
strong 8664 crisis 17534
good 7206 cut 14255
despite 6941 bailout 8697
stability 6113 negative 8688
boost 5977 late 7888
better 5615 easing 7632
highest 4935 unemployment 6715
gains 4351 concerns 6574
positive 4278 fears 5585
able 4193 weak 5551
leading 3684 losses 5353
stronger 3640 warned 5337
best 3585 decline 4825
gained 3378 deficit 4821
greater 3082 problems 4798
strength 2397 lost 4492
great 2371 recession 4291
gain 2362 dropped 4214
benefit 1797 problem 4209
progress 1770 default 4143
Note: The table shows the 20 words from the Loughran and McDonald (2011) dictionary with the highest number
of occurrences in our database. Top positive (first column) and top negative (second column) words are shown.
32
Appendix B - Sentix investor confidence index versus MSI
We first plot the monthly Sentix investor confidence index for the Eurozone together with the
MSI average at the monthly frequency (Figure 3, upper part), and then the residual component of
our index after having regressed it on the Sentix investor confidence index (Figure 3, lower part).
Figure 3: MSI (monthly average) versus Sentix investor economic confidence for the Euro area
Notes: the top graph shows the media sentiment indicator aggregated at a monthly frequency versus the Sentix
investor economic confidence index for the Euro area. The bottom graph shows the residuals of the regression of
the media sentiment index on the Sentix investor confidence index, labelled "idiosynchratic component with
respect to Sentix index". The events corresponding to important highs and lows are commented.
The residual component shows the idiosyncratic information to our index with respect to the
Sentix investor confidence index. We observe that it contains relevant information on monetary
policy perception. Focusing on the post-crisis period, the largest peak arises during the "whatever
it takes" speech of Mario Draghi in the summer of 2012. Our MSI index soars the month of
Draghi speech while investor sentiment stays at a low level. After that period, it peaks again when
33
forward guidance is introduced in 2013 and when QE talks start in late 2014. We also see that
in 2009 the idiosyncratic component of our index remains significantly positive for some months,
possibly revealing that the confidence in the central bank solving the financial crisis was relatively
high (in the context of a falling investor economic sentiment). The lowest peak arises in January
2008, when markets become relatively pessimistic on the economic outlook, while Trichet remains
hawkish in his communication and the ECB does not cut rate in the January meeting, contrasting
with the accommodating policy decided by the FeD few days before.
34
Appendix C - Data on ECB communications
Press conference communications are extracted from the ECB website. We measure their quali-
tative content in terms of monetary policy inclination (hawkish versus dovish) and communication
vis-à-vis the economic outlook (positive versus negative). To do so, we apply the field-specific
dictionary developed in Picault and Renault (2017), which aims at capturing these two aspects
of the ECB member’s speeches, and is exactly built from the ECB members communications’
wording. In contrast to standard methods, it uses contiguous sequence of words and a specific
term-weighting approach to quantify the content of communication. This ECB-specific dictionary
has been shown to better capture the central bankers’ communication subtleties compared to the
traditional alternatives and is thus a natural candidate to extract the qualitative content of the
ECB communications. With the monetary policy inclination dictionary, we build a variable reflect-
ing the hawkish/dovish inclination of the President speeches during the press conference as well as
a variable reflecting its communication on the economic outlook during the press conference. The
classification of the ECB press conferences’ content is presented in Figure 4.
Figure 4: ECB Press Conference Introductory statement content
Note: The Figure presents, at a daily frequency, the monetary policy (lower value indicates a more
accommodating content) and economic outlook (a lower value indicates a more negative content) inclination of the
ECB Press conference Introductory Statement using Picault and Renault (2017) lexicon and methodology.
Regarding inter-meeting communications, we use data kindly provided by Gertler and Horvath
(2018), who manually classified ECB inter-meeting communications in function of their monetary
policy and economic outlook inclination. We obtain daily data representing the average inclina-
tion of the communications, differentiated between the communications from the President and
the communications from the other members of the Governing Council (other Executive Board
members and Heads of national central banks). We use them to construct two distinct daily time
35
series. The the number of communications and the resulting data are displayed in Figure 5 and 6
respectively.
Figure 5: Inter-meeting communications from ECB officials
Note: The Figure presents the daily number of inter-meeting communications from ECB officials during our
sample period using the Gertler and Horvath (2018) database.
Figure 6: ECB Communications tonality regarding Monetary Policy and Economic Outlook
Note: The Figure presents, at a daily frequency, the monetary policy inclination of the ECB communications (on
the left - a lower value indicates a more accommodating content) and the economic outlook view of the ECB
communications (on the right - a lower value indicates a more negative content). We present here the
accumulated values for these data for readability purpose, in all estimates the first difference of these was taken.
36
Appendix D - List of macroeconomic surprises and unconven-
tional monetary policies considered
Table 5: Macroeconomic surprises:
Country or area Variable
Euro Area
Inflation surprises, flash CPI
Business climate surprises
Industrial confidence surprises
PPI surprises
Production surprises
Unemployment surprises
France Inflation surprises, HCPI
Industrial production surprises
Germany
Inflation surprises, HCPI
Business climate surprises
Unemployment surprises
Italy
Inflation surprises, HCPI
Manufacture confidence surprises
Unemployment surprises
Spain
Inflation surprises, HCPI
Industrial output surprises
Unemployment surprises
Table 6: Unconventional monetary policies:
Press release date Unconventional monetary policy
7 May 2009 Covered bond purchase programme (first) and 1-year LTRO
10 May 2010 Securities market programme
7 August 2011 Securities market programme, new annoucement
6 October 2011 Covered bond purchase programme (second) and new LTRO
6 September 2012 Technical features of OMT
5 June 2014 Asset-backed securities purchase programme and TLTRO
4 September 2014 Asset-backed securities purchase programme and Covered bond pur-
chase programme
22 January 2015 Public sector purchase programme
9 March 2015 Public sector purchase programme, new announcement
10 March 2016 Public sector purchase programme, new announcement (increase in size)
Note: From Moessner (2018), press release date is verified on the ECB website.
37
Appendix E - Media Sentiment determinants, effect of com-
munications - Robustness and further tests
Table 7: Media Sentiment determinants, effect of communications - model with additional lags
[1] [2] [3] [4] [5] [6]
MSIt10.280*** 0.278*** 0.275*** 0.238*** 0.252*** 0.252***
(0.033) (0.034) (0.034) (0.032) (0.033) (0.025)
MSIt20.069*** 0.067*** 0.068*** 0.048** 0.048* 0.048*
(0.025) (0.026) (0.026) (0.024) (0.026) (0.026)
MSIt30.045 0.045 0.047 0.032 0.032 0.032
(0.033) (0.033) (0.033) (0.032) (0.027) (0.025)
MSIt40.086*** 0.088*** 0.087*** 0.065** 0.063** 0.063**
(0.024) (0.025) (0.025) (0.026) (0.028) (0.025)
President MP - Press Conferencest-0.479*** -0.523*** -0.507*** -0.441*** -0.438*** -0.438***
(0.132) (0.142) (0.145) (0.132) (0.121) (0.162)
President MP - Press Conferencest1-0.525*** -0.510*** -0.515*** -0.383*** -0.367*** -0.367**
(0.095) (0.116) (0.117) (0.121) (0.122) (0.163)
President MP - Press Conferencest2-0.409*** -0.465*** -0.471*** -0.345*** -0.364*** -0.364**
(0.084) (0.110) (0.110) (0.117) (0.126) (0.164)
President MP - Press Conferencest3-0.402*** -0.330** -0.331** -0.266* -0.270* -0.270
(0.140) (0.144) (0.144) (0.147) (0.149) (0.164)
President EC - Press Conferencest0.555* 0.627* 0.595* 0.590* 0.570** 0.570
(0.329) (0.341) (0.332) (0.322) (0.284) (0.417)
President EC - Press Conferencest10.198 0.204 0.213 0.183 0.196 0.196
(0.205) (0.242) (0.243) (0.235) (0.222) (0.417)
President EC - Press Conferencest20.072 0.061 0.075 0.050 0.065 0.065
(0.263) (0.299) (0.298) (0.301) (0.298) (0.421)
President EC - Press Conferencest30.539 0.419 0.417 0.512 0.449 0.449
(0.376) (0.423) (0.429) (0.420) (0.390) (0.422)
Other members MP - Inter Meetingt1-0.061** -0.060** -0.061** -0.052* -0.054* -0.054
(0.029) (0.029) (0.029) (0.028) (0.032) (0.034)
Other members MP - Inter Meetingt2-0.045 -0.043 -0.042 -0.036 -0.036 -0.036
(0.028) (0.028) (0.028) (0.027) (0.034) (0.034)
Other members MP - Inter Meetingt30.016 0.016 0.011 0.028 0.035 0.035
(0.039) (0.039) (0.039) (0.039) (0.036) (0.035)
Other members EC - Inter Meetingt10.028 0.020 0.023 0.019 0.013 0.013
(0.047) (0.047) (0.046) (0.046) (0.053) (0.050)
Other members EC - Inter Meetingt20.109** 0.105** 0.106** 0.083** 0.075* 0.075
(0.043) (0.044) (0.044) (0.041) (0.044) (0.050)
Other members EC - Inter Meetingt3-0.002 -0.007 -0.008 -0.044 -0.046 -0.046
(0.050) (0.051) (0.051) (0.047) (0.046) (0.050)
President MP - Inter Meetingt1-0.177*** -0.168*** -0.156*** -0.166*** -0.170*** -0.170**
(0.043) (0.046) (0.047) (0.045) (0.054) (0.075)
President MP - Inter Meetingt20.058 0.058 0.056 0.047 0.039 0.039
(0.063) (0.065) (0.068) (0.070) (0.069) (0.075)
President MP - Inter Meetingt30.010 0.011 0.004 -0.000 0.013 0.013
(0.072) (0.074) (0.074) (0.071) (0.079) (0.074)
President EC - Inter Meetingt10.228** 0.231** 0.230** 0.216** 0.227** 0.227*
(0.091) (0.092) (0.090) (0.099) (0.093) (0.132)
President EC - Inter Meetingt2-0.328*** -0.329** -0.327*** -0.337*** -0.340** -0.340**
(0.127) (0.128) (0.127) (0.124) (0.139) (0.132)
President EC - Inter Meetingt3-0.140 -0.141 -0.145 -0.168 -0.162 -0.162
(0.145) (0.146) (0.149) (0.148) (0.190) (0.132)
ECB decisions, t to t-3 No Yes Yes Yes YesnYesn
Macro-economic surprises, t to t-3 No No Yes Yes YesnYesn
Financial variables, t to t-3 No No No Yes YesnYesn
Constant -0.061*** -0.063*** -0.062*** 0.202*** 0.215*** 0.215***
(0.018) (0.019) (0.019) (0.052) (0.047) (0.050)
Adjusted R20.165 0.159 0.158 0.197 0.197 0.197
Obs. 1630 1630 1630 1630 1630 1630
Note: Newey-West Standard Errors in parenthesis. ***, **, and * represent statistical significance at respectively 1%, 5%, and 10%. n:
only the controls which appeared statistically significant at the 10 percent level in column (4) included in the specification. M SI is the
Media Sentiment Index. All control variables are defined in Table 2.
38
Table 8: Media Sentiment determinants, effect of communications - Robustness
[1] [2] [3] [4] [5] [6]
H-W lag nega sent harvard sent vader sent weighted sent
MSIa
t10.266*** 0.255*** 0.146*** 0.160*** 0.136*** 0.196***
(0.033) (0.033) (0.035) (0.029) (0.025) (0.030)
MSIa
t20.050* 0.056** 0.097*** 0.036 0.041* 0.069**
(0.026) (0.026) (0.025) (0.024) (0.022) (0.029)
MSIa
t30.030 0.027 0.035 0.034 0.039* 0.015
(0.027) (0.031) (0.027) (0.029) (0.023) (0.031)
MSIa
t40.062** 0.070*** 0.102*** 0.074*** 0.065*** 0.094***
(0.028) (0.024) (0.025) (0.028) (0.020) (0.024)
President MP - Press Conferencest-0.401*** -0.428*** -0.369*** -3.840* -6.587*** -0.405***
(0.120) (0.134) (0.100) (2.147) (1.531) (0.127)
President EC - Press Conferencest0.554** 0.558 0.702*** -0.998 3.667 0.652**
(0.278) (0.341) (0.272) (3.857) (3.352) (0.330)
Other members MP - Inter Meetingt1-0.066** -0.059** -0.044** -1.427** -1.065*** -0.062**
(0.032) (0.027) (0.022) (0.684) (0.376) (0.028)
Other members EO - Inter Meetingt10.012 0.009 0.000 0.073 -0.118 0.021
(0.052) (0.042) (0.044) (0.727) (0.366) (0.055)
President MP - Inter Meetingt1-0.166*** -0.172*** -0.139*** -1.695* -1.883*** -0.144***
(0.052) (0.041) (0.040) (0.889) (0.489) (0.045)
President EO - Inter Meetingt10.217** 0.203** 0.023 -0.799 0.694 0.085
(0.087) (0.091) (0.095) (1.883) (0.778) (0.080)
i-0.736 -0.548 -0.852 0.637 2.477 -1.192
(0.648) (0.541) (0.572) (11.491) (6.744) (0.731)
Negative Interest Surprise 1.733*** 1.682*** 1.455*** 28.494** 6.742 2.050***
(0.631) (0.563) (0.554) (11.886) (9.501) (0.695)
Positive Interest Surprise 0.829 1.088** 0.327 2.086 -3.038 -0.468
(0.749) (0.540) (0.460) (8.990) (5.672) (0.856)
APP -0.151 -0.049 0.025 0.798 -0.602 -0.123
(0.144) (0.072) (0.079) (1.653) (1.051) (0.168)
Negative Rate 0.166*** 0.157*** 0.104*** 12.091*** 6.692*** -0.029
(0.034) (0.039) (0.029) (0.604) (0.433) (0.040)
Whatever it takes 0.163*** 0.330*** 0.004 1.765 -0.451 -0.139**
(0.056) (0.037) (0.055) (1.136) (0.866) (0.058)
Macro-economic surprisestYes No Yes Yes Yes Yes
Financial variablestYes No Yes Yes Yes Yes
Macro-economic surprisest1No Yes No No No No
Financial variablest1No Yes No No No No
Constant 0.241*** 0.206*** 0.165*** 4.431*** 2.766*** 0.223***
(0.045) (0.051) (0.042) (1.089) (0.773) (0.060)
Adjusted R20.188 0.171 0.123 0.101 0.108 0.142
Obs. 1630 1630 1630 1630 1630 1630
Note: Estimates of the coefficients of equation 3, with alternative controls and alternative constructions of the MSI. In parenthesis,
Huber-White Standard Errors in column (1) and Newey-West Standard Errors in columns (2), (3), (4), (5) and (6). In columns (3),
(4), (5) and (6) we use alternative dependant variables so that M SI ais built with only the negative words considered in equation 1
(column 3), or with the positive and negative words from the Harvard-IV dictionary (column 4), or using Vader classifier (column 5) or
by weighting the newspapers using a proxy for their audience (column 6, see footnote 32 for details). ***, **, and * represent statistical
significance at respectively 1%, 5%, and 10%. All control variables are defined in Table 2.
39
Table 9: Media Sentiment determinants, effect of communications when inflation is above/below
its target
[1] [2] [3] [4]
MSIt10.290*** 0.290*** 0.268*** 0.265***
(0.032) (0.032) (0.031) (0.031)
MSIt20.064** 0.063** 0.048* 0.048*
(0.026) (0.026) (0.026) (0.026)
MSIt30.036 0.036 0.025 0.026
(0.033) (0.033) (0.031) (0.031)
MSIt40.079*** 0.079*** 0.060** 0.061**
(0.024) (0.024) (0.024) (0.024)
President MP - Press Conferencest-0.405*** -0.455*** -0.405*** -0.405***
(0.129) (0.139) (0.126) (0.126)
President EC - Press Conferencest0.973* 0.976* 1.052** 1.016**
(0.514) (0.503) (0.434) (0.441)
Other members MP - Inter Meetingt1-0.088*** -0.089*** -0.073** -0.078**
(0.033) (0.034) (0.033) (0.033)
Other members EO - Inter Meetingt10.046 0.046 0.063 0.068
(0.054) (0.054) (0.054) (0.053)
President MP - Inter Meetingt1-0.200*** -0.200*** -0.190*** -0.186***
(0.048) (0.048) (0.045) (0.046)
President EO - Inter Meetingt10.184* 0.184* 0.166 0.166
(0.102) (0.102) (0.104) (0.104)
T arget -0.088** -0.088** -0.046 -0.049
(0.039) (0.038) (0.035) (0.035)
President MP - Press ConferencestxT arget -0.081 -0.025 -0.051 -0.047
(0.321) (0.311) (0.285) (0.286)
President EO - Press ConferencestxT arget -0.733 -0.774 -0.847 -0.821
(0.675) (0.662) (0.579) (0.585)
Other members MP - Inter Meetingt1xT arget 0.095* 0.102* 0.058 0.062
(0.056) (0.055) (0.059) (0.057)
Other members EO - Inter Meetingt1xT arget -0.103 -0.106 -0.190* -0.193*
(0.093) (0.093) (0.100) (0.099)
President MP - Inter Meetingt1xTar get 0.142 0.141 0.086 0.097
(0.096) (0.096) (0.096) (0.092)
President EO - Inter Meetingt1xTar get 0.237* 0.234* 0.267* 0.267*
(0.142) (0.142) (0.145) (0.145)
i-0.136 -0.464 -0.455
(0.472) (0.581) (0.588)
Negative Interest Surprise 1.406*** 1.593*** 1.596***
(0.468) (0.525) (0.532)
Positive Interest Surprise 0.937 0.357 0.329
(0.609) (0.805) (0.809)
APP -0.116 -0.155 -0.151
(0.079) (0.146) (0.147)
Negative Rate 0.182*** 0.163*** 0.162***
(0.035) (0.036) (0.035)
Whatever it takes 0.507*** 0.240*** 0.240***
(0.054) (0.061) (0.063)
Financial variables No No Yes Yes
Macro-economic surprises No No No Yes
Constant -0.021 -0.021 -0.020 0.194***
(0.017) (0.017) (0.017) (0.044)
Adjusted R20.155 0.154 0.186 0.188
Obs. 1630 1630 1630 1630
Note: T arget is a dummy variable equal to 1 if inflation is above the 2% target, 0 otherwise. Newey-West Standard Errors in parenthesis.
***, **, and * represent statistical significance at respectively 1%, 5%, and 10%. MS I is the Media Sentiment Index. All control variables
are defined in Table 2.
40
Appendix F - Descriptive statistics
Table 10: Descriptive Statistics (2010-2016)
Mean S. D. Max. Min. Obs.
Media Sentiment
MSIgross -1.508 0.558 0.902 -3.945 1630
MSI -0.071 0.543 2.338 -2.562 1630
MSIN egativeW ords -0.065 0.452 1.574 -2.103 1630
MSIHarvardI V 4-0.319 9.250 32.008 -41.962 1630
MSIV ader -1.700 6.889 77.591 -63.855 1630
MSIW eighted -0.126 0.583 2.292 -3.139 1630
ECB Communications
President MP - Press Conferences -0.013 0.088 0.545 -0.761 1630
President EC - Press Conferences -0.003 0.035 0.296 -0.473 1630
President MP - Inter Meeting -0.015 0.167 1.000 -1.000 1630
President EC - Inter Meeting -0.001 0.093 1.000 -1.000 1630
Other members MP - Inter Meeting -0.034 0.376 1.000 -1.000 1630
Other members EC - Inter Meeting -0.003 0.249 1.000 -1.000 1630
Monetary Policy Decisions
i-0.001 0.017 0.250 -0.250 1630
Negative Interest Surprise -0.000 0.009 0.000 -0.250 1630
Positive Interest Surprise 0.000 0.006 0.250 0.000 1630
APP 0.005 0.070 1.000 0.000 1630
Negative Rate 0.001 0.025 1.000 0.000 1630
Whatever it takes 0.001 0.025 1.000 0.000 1630
Financial Variables
Fiscal stress 0.001 0.178 1.376 -3.608 1630
EURUSD -0.000 0.006 0.030 -0.024 1630
VSTOXX 23.810 6.789 53.547 12.713 1630
EUROSTOX 0.000 0.014 0.098 -0.063 1630
5Y-5Y Inflation Expectations (swap) -0.082 2.092 10.507 -8.741 1608
10Y Inflation Expectations (swap) -0.089 2.033 10.000 -10.500 1608
41
... Other topics might give 46 rise to niche or underground debates, e.g. health practice [9,10]: These topics might be 47 more uncommonly discussed across media channels and thus become more difficult to be 48 identified by monitoring social/news systems [10]. To what extent is NLP able to capture 49 different risks by using insurance-specific reports? ...
... For example, Figure 2 illustrates the benefits of 287 the TF-IDF in drawing out the pertinent themes which would be lost if we were merely 288 highlighting the most common words. In this case, in 2021 the TF-IDF correctly highlights 289 major themes (tracked also by previous works) like the COVID-19 pandemic (see [9,44]), 290 the increasing risk for cyber-attacks and online manipulation (see [45][46][47]) and growing 291 concern around inflation (see [48]). ...
Preprint
Full-text available
The ability to identify and rank risk is essential for efficient and effective supervision of financial service firms, such as banks and insurers. Risk ranking ensures limited resources are allocated where they are most needed. Today, automatic risk identification within insurance supervision primarily relies on quantitative metrics based on numerical data (e.g. returns). The purpose of this work is to assess whether Natural Language Processing (NLP) and cognitive networks can achieve similar automated risk ranking and identification by analysing textual data, i.e. NIDT=829 investor transcripts from Bloomberg. To this aim, this work explores and tunes 3 NLP techniques: (1) keyword extraction enhanced by cognitive network analysis; (2) valence/sentiment analysis; and (3) topic modelling. Results highlight that keyword analysis, enriched by term frequency-inverse document frequency scores and semantic framing through cognitive networks, could detect events of relevance for the insurance system like cyber-attacks or the COVID-19 pandemic. Cognitive networks were found to highlight events that related to specific financial transitions: The semantic frame of "climate" grew in size by +538% between 2018 and 2020 and outlined an increased awareness that agents and insurers expressed towards climate change. A lexicon-based sentiment analysis achieved a Pearson’s correlation of ρ=0.16 (p<0.001,N=829) between sentiment levels and daily share prices. Although relatively weak, this finding indicates that insurance jargon is insightful to support risk supervision. Topic modelling is considered less amenable to support supervision, because of a lack of results’ stability and an intrinsic difficulty to interpret risk patterns. We discuss how these automatic methods could complement existing supervisory tools in automated risk ranking.
Article
Full-text available
This paper examines whether the South Africa Reserve Bank's (SARB) Monetary Policy Committee (MPC) communications between 2010 and 2021 trigger causality in the subsequent news reports from the Mail and Guardian newspaper in South Africa. The study examines whether SARB's post MPC statements' readability is reciprocated in the subsequent Mail and Guardian newspaper articles. The study relies on the Flesch-Kincaid Grade Level score and Flesch Reading Ease Score methodology to unpack both SARB's MPC statements and the corresponding Mail and Guardian newspaper articles to create a computation analysis which is ultimately used to examine Granger causality. The study finds that although there might appear to be a co-movement between SARB's MPC statements' Flesch-Kincaid Grade Level score and Flesch Reading Ease Score to Mail and Guardian's newspaper articles' Flesch-Kincaid Grade Level score and Flesch Reading Ease Score, the Granger casualty test reveals that there is no causality between the two variables. This is explained by the Mail and Guardian articles directly quoting SARB's MPC statements in essence creating an 'artificial' co-movement.
Article
Full-text available
We analyze the degree of anchoring of inflation expectations in the euro area during the post-crisis period, with a focus on the time span from 2014 onwards when long-term beliefs have substantially drifted away from the policy target. Using a new estimation technique, we look at tail co-movements between short-and long-term distributions of inflation expectations, estimated from daily quotes of inflation derivatives. We find that, during 2014, average correlations between short-and long-term inflation expectations rose sharply; moreover, negative tail events impacting short-term beliefs have been increasingly channeled to long-term views, triggering both downward revisions in expectations and upward changes in uncertainty. Overall, our results signal a risk of downside de-anchoring of long-term inflation expectations. JEL Codes: C14, C58, E31, E44, G13.
Article
Full-text available
We develop a field-specific dictionary to measure the stance of the European Central Bank (ECB) monetary policy (dovish, neutral, hawkish) and the state of the Eurozone economy (positive, neutral, negative) through the content of ECB press conferences. In contrast with traditional textual analysis, we propose a novel approach using term-weighting and contiguous sequence of words (n-grams) to better capture the subtlety of central bank communication. We find that quantifying ECB communication using our field-specific weighted lexicon helps to explain future ECB monetary decisions when considering an augmented Taylor rule. Regarding European stock market volatility, we find that markets are more (less) volatile on the day following a conference with a negative (positive) tone about the euro area economic outlook. Our indicators significantly outperform a textual classification based on the Loughran–McDonald or Apel–Blix Grimaldi dictionaries and a media-based measure of economic policy uncertainty.
Article
This paper studies the nonlinear response of the term structure of interest rates to monetary policy shocks and presents a new stylized fact. We show that uncertainty about monetary policy changes the way the term structure responds to monetary policy. A policy tightening leads to a significantly smaller increase in long‐term bond yields if policy uncertainty is high at the time of the shock. We also look at the decomposition of bond yields into expectations about future policy and the term premium. The weaker response of yields is driven by the fall in term premia, which fall more strongly if uncertainty about policy is high. Conditional on a monetary policy shock, higher uncertainty about monetary policy tends to make securities with longer maturities relatively more attractive to investors. As a consequence, investors demand even lower term premia. These findings are robust to the measurement of monetary policy uncertainty, the definition of the monetary policy shock, and to changing the model specification.
Article
We construct a new measure of uncertainty about Federal Reserve policy actions and their consequences, a monetary policy uncertainty (MPU) index. We evaluate the information content of our index and document the usefulness of our index in bridging periods of conventional and unconventional policy making. We also estimate the aggregate effects of shocks to MPU on output, credit spreads, and other variables. Finally, we investigate the transmission channels of MPU, finding that heightened MPU leads to protracted declines in firm investment through both real options and financial frictions channels.
Article
The classic papers by Newey and West (1987) and Andrews (1991) spurred a large body of work on how to improve heteroscedasticity- and autocorrelation-robust (HAR) inference in time series regression. This literature finds that using a larger-than-usual truncation parameter to estimate the long-run variance, combined with Kiefer-Vogelsang (2002, 2005) fixed-b critical values, can substantially reduce size distortions, at only a modest cost in (size-adjusted) power. Empirical practice, however, has not kept up. This article therefore draws on the post-Newey West/Andrews literature to make concrete recommendations for HAR inference. We derive truncation parameter rules that choose a point on the size-power tradeoff to minimize a loss function. If Newey-West tests are used, we recommend the truncation parameter rule S = 1.3T1/2 and (nonstandard) fixed-b critical values. For tests of a single restriction, we find advantages to using the equal-weighted cosine (EWC) test, where the long run variance is estimated by projections onto Type II cosines, using ν = 0.4T2/3 cosine terms; for this test, fixed-b critical values are, conveniently, tν or F. We assess these rules using first an ARMA/GARCH Monte Carlo design, then a dynamic factor model design estimated using a 207 quarterly U.S. macroeconomic time series.
Article
We study the effects of the announcements of ECB asset purchases and of financial stability measures in the euro area in the wake of the global financial crisis and the euro area sovereign debt crisis on 10-year government bond term premia in 11 euro area countries. We find that the term premia of euro area countries with higher sovereign risk, as measured by sovereign CDS spreads, decreased more in response to the announcements of asset purchases and financial stability measures. Term premia of countries with lowest sovereign risk either increased as in Germany, or were not significantly affected or fell slightly, as in the Netherlands and Finland.
Article
This paper examines the financial market impact of intermeeting communication of the members of the European Central Bank's Governing Council (GC) using high frequency data between July 2008 and January 2014. Constructing a rich dataset of GC members’ public statements (speeches, conference discussions and media interviews) between monetary policy meetings allows us to investigate a detailed pattern of market responses to the ad-hoc communication of central bankers. Using least squares and quantile regressions, we document the impact of policymakers’ public statements on interest rates and the stock market with very little or no impact on exchange rates. In general, we find little evidence that the timing, sequencing or content of communication matters in immediate response. On the contrary, the results suggest that the market concentrates on the communication of key members of the committee.