Gauging the Sentiment of Federal Open Market Committee Communications through the Eyes of the Financial Press

Finance and Economics Discussion Series Federal Reserve Board, Washington, D.C. ISSN 1936-2854 (Print) ISSN 2767-3898 (Online) Gauging the Sentiment of Federal Open Market Committee Communications through the Eyes of the Financial Press Shantanu Banerjee, Paul Cordova, Michiel De Pooter, and Olesya V. Grishchenko 2025-048 Please cite this paper as: Banerjee, Shantanu, Paul Cordova, Michiel De Pooter, and Olesya V. Grishchenko (2025). “Gauging the Sentiment of Federal Open Market Committee Communications through the Eyes of the Financial Press,” Finance and Economics Discussion Series 2025-048. Washington: Board of Governors of the Federal Reserve System, https://doi.org/10.17016/FEDS.2025.048. NOTE: Staff working papers in the Finance and Economics Discussion Series (FEDS) are preliminary materials circulated to stimulate discussion and critical comment. The analysis and conclusions set forth are those of the authors and do not indicate concurrence by other members of the research staff or the Board of Governors. References in publications to the Finance and Economics Discussion Series (other than acknowledgement) should be cleared with the author(s) to protect the tentative character of these papers.

Gauging the Sentiment of Federal Open Market Committee ∗ Communications through the Eyes of the Financial Press Shantanu Banerjee Paul Cordova Michiel De Pooter Olesya V. Grishchenko July 3, 2025 Abstract We apply natural language processing tools to news articles in the financial press to construct a sentiment index — an index of the perceived semantic orientation of monetary policy communications around scheduled Federal Open Market Committee (FOMC) meetings. To that end, we develop several dictionaries that capture various monetary policy tools: conventional monetary policy, asset purchases, and forward guidance. The surprises in the sentiment index around FOMC meetings announcements explain variation in major asset prices classes between May 1999 and November 2022. Sentiment index surprises are important for explaining the variation in asset prices beyond monetary policy surprises. JEL Classification: E00, E40, E58, G12 Keywords: Textualanalysis,semanticorientation,sentimentindex,FederalReserve,FOMC, hawkish, dovish, asset prices, policy expectations, conventional monetary policy, asset purchases, forward guidance, zero-lower-bound, COVID. ∗The views expressed in this paper are solely those of the authors and should not be interpreted as reflecting the views of the Board of Governors of the Federal Reserve System or of any other employee of the Federal Reserve System. Michiel De Pooter’s contribution to the paper was completed prior to the author joining Amazon. This publication and its contents are not related to Amazon and do not reflect the position of the company and its subsidiaries. We thank Stephen Brown, Mengqiao Du, Michael Erhmann, Stefan Greppmair, Klodiana Istrefi, Andreas Joseph, Don Kim, Ellen Meade, Brad Strum, Alexander Wagner, and Min Wei for helpful comments as well as participants of the seminars at the University of Rennes, University of Queensland, Victoria University at Wellington, Federal Reserve Board, Brock University, New Economic School, Workshop on Central Banking Communications at the Bank of Canada, the 2018 Summer Camp of the Research School of Finance of the Australian National University, the 2019 Annual Meeting of the Swiss Finance Society, the 2019 Biennial Conference of the Georgetown Center for Economic Research, the 2019 Swiss Finance Society Annual Meeting, the 2019 meetings of the European Finance Association, the 2022 Annual Meeting of the French Finance Association, the ECONDAT 2023 Spring Meeting ”Economics with non-traditional data and analytical tools”, the 6th Conference ”Nontraditional Data, Machine Learning, and Natural Language Processing for Macroeconomics”, the ECONDAT 2024 Fall Meeting at the Bank of Italy. Cait Walsh provided excellent research assistance. All errors are our own. Shantanu BanerjeeisattheFederalReserveBankofNewYork. PaulCordovaisatColumbiaUniversity. MichielDePooteris at Amazon.com. Olesya V. Grishchenko at the Board of Governors of the Federal Reserve System. Corresponding author: olesya.v.grishchenko@frb.gov; Board of Governors of the Federal Reserve System, Washington, DC 20551, phone 202-452-2981.

1 Introduction Over the past decades, central bank communications have changed significantly as central banks around the world have put more and more emphasis on transparency. For example, through the early 1990s, monetary policy decisions by the Federal Reserve’s Federal Open Market Committee (FOMC) were not announced to the public. As such, market participants had to infer whether the FOMC had changed its policy stance from the open market operations conducted by the Federal Reserve Bank of New York. In contrast, currently the FOMC uses a wide range of communication tools — including policy statements and press conferences following the conclusion of each FOMC meeting, the release of the Summary of Economic Projections (SEP) at every other meeting, as well as testimonies and speeches by the Fed Chair and FOMC members — to convey its views on U.S. economic developments and its monetary policy stance. Other central banks around the globe have similarly increased their transparency over time. Consequently, central bank communications are constantly in the spotlight and heavily scrutinized by various media outlets, in particular, by the financial press. As then-member of the Federal Reserve Board and FOMC Ben Bernanke emphasized in his remarks during the American Economic Association meeting (Bernanke, 2004) “... because the world is complex and ever changing, policy actions alone, without explanation, will never be enough to provide the public with the information it needs to predict policy actions. Words are also necessary.” Not surprisingly, the variation in the tone, or sentiment, of central banking communications has attracted a lot of attention among academic researchers and market participants alike. Our paper is related to the literature on central banking communications (see, e.g., Woodford, 2005; Blinder et al., 2008, for a comprehensive overview). More recently, the relevant literature particularly focused on the bag-of-words approach. The studies include, for example, Baker et al. (2016) who relate frequency of relevant words related to economic policy uncertainty, whereas Husted et al. (2020) similarly construct monetary policy uncertainty index, and Caldara and Iacoviello (2022) construct the geopolitical risk index. A specific part of this literature is related to measurement of semantic orientation of central bank communications. This literature falls into two broad categories: Researchers in this area use natural language processing (NLP) tools to study either the textual content of direct central banking communications (Romer and Romer, 2004; Meade et al., 2015; Acosta and Meade, 2015; Cannon, 2015; Hansen et al., 2018; Gardner et al., 2022; Acosta, 2023; Schmeling and Wagner, 2025) or public perceptions and interpretations of the central bank communications (Lucca and Trebbi, 2009; Carvalho et al., 2016; Schmanski et al., 2023).1 Our work falls into the second category. 1Schmanski et al. (2023) uses both direct communications and news articles, including Twitter, to measure monetary policy surprises related to sentiment. 1

Recent work on semantic orientation of central bank communications is part of a much larger and rapidly growing literature in economics and finance that focuses on extracting quantitative signals and measures from a range of sources containing written words, with applications to constructing uncertainty measures (e.g., Baker et al., 2016; Husted et al., 2020), predicting stock returns (Born et al., 2014; Azar and Lo, 2016; Heston and Sinha, 2020), predicting future firm earnings (Druz et al., 2020), measuring the value of economist forecasts (Sharpe et al., 2023), and — economic sentiment (Shapiro et al., 2022), among many others.2 Inthispaper, wecontributetothetextualanalysisliteraturethatfocusesspecificallyoncentral bank communications by constructing an index that measures semantic orientation of perceived FOMC communications around FOMC meetings. In particular, we construct a sentiment index (SI) from a corpus of relevant published articles collected from the premier news resource Dow Jones Factiva. The corpus of articles for our study consists of all relevant news articles from six major financial press outlets published in event windows of different lengths around FOMC meetingsbetweenMay1999andNovember2022. Therearetwobenefitsofassessingsentimentofpolicy communications via press articles. The first one is that it allows us to circumvent the complex task of assessing the sentiment of direct Fed communications, as official FOMC communications are notably complex and have become more complex over time (see, e.g., Hernandez-Murillo and Shell, 2014). The second one is that looking at the media articles before and after the FOMC meetings allows us to construct surprises in the sentiment and relate these surprises to changes in asset prices around FOMC communications — something that is not possible to do when looking only at the FOMC released statements. Our sample period starts from the May 1999 FOMC meeting and ends with the November 2022 FOMC meeting. Thus, it includes diverse periods of monetary policy communications: the GFC, following zero-lower-bound period, subsequent monetary policy normalization, the COVID- 19 pandemic and post-pandemic periods. As we discuss later, the presence of these periods in our sample also reveals the limitations of many existing methodologies used to assess the sentiment of FOMCcommunications, particularlythelackofwordsforcapturinghawkishanddovishsentiment associated with forward guidance and asset purchases. Toaddresseverincreasingcomplexityinmonetarypolicycommunications, ourapproachdiffers from existing studies in novel and significant ways. First, we collect the relevant stories from Dow Jones Factiva. Then we categorize the stories into four topic-keywords categories, based on general mentions of the monetary policy by the financial press (General), or specific mentions of monetary policy conduct in various regimes, whether they are related to conventional monetary policy (Policy Rates), or to unconventional policies such as asset purchases or forward guidance, 2A recent overview and introduction to the use of “text” as an input to economic research is given in Gentzkow et al. (2019). 2

naturally called (Asset Purchases) and (Forward Guidance). Second, we create modifier-keywords dictionaries related to each of the topic-keywords category. Modifier-keywords, in general, reflect and determine semantic orientation of topic-keywords (such as hawkish or dovish characterization of topic-keywords) based on the monetary policy actions and tools related to each topic. This distinction appears to us extremely important as different monetary policy regimes may (will) have different modifiers that characterize tightening or easing stance in the monetary policy. As oneofthemostobviousexamples, ahawkishcharacterizationofthemonetarypolicy(e.g., increase [rates]) during the conventional monetary policy regime when the Fed is using the fed funds rate as its main policy tool may become dovish characterization during the unconventional monetary policy regime (e.g., increase [purchases]) when asset purchases become one of the main tools, as during Large Scale Asset Purchase (LSAP) programs during the Global Financial Crisis (GFC). Third, we construct the sentiment index (SI) for each relevant story. This allows us to account for the length of a story and to weigh the frequency of modifier-keywords in each individual story, giving larger weights to stories that have a substantive discussion of the FOMC’s expected or realized policy stance compared to stories where the policy action is barely mentioned. We then aggregate the SIs of individual stories in the FOMC-specific SI in the window that starts on Monday of the week of the FOMC and ends at midnight on Wednesday after the FOMC.3 Fifth, we use a more elaborate set of rules to deal with the complexity of text and in particular how to handle verb tenses, and negations that can change the meaning of hawkish words to dovish and vice versa. In the second part of our paper we gauge the SI informational content. To do so, we use SI changes around the FOMC statement release time as a measure of sentiment surprises and in a simple linear regression framework, we examine whether surprises in SI help explain movements in a range of asset prices in arrow event windows around the releases of the FOMC statements.4 We find that they do. Our empirical results indicate that SI performs reasonably well in terms of explaining movements in prices of money market futures contracts, nominal and inflation-linked Treasury securities, equity indexes, and foreign currency pairs, all of which are measured in the narrow intraday event windows around FOMC meetings. Moreover, they do so even after we control for the presence of monetary policy surprises, based on (see, e.g., Bauer and Swanson, 2023). Our results are robust to controlling for the Global Financial Crisis (GFC) and for the zer-lower-bound (ZLB) period. We also analyze SI informational content separately during several subsample periods and reporttheseresultsinAppendix B.Thesesubsampleperiodsarethepre-GFCsample1999-2009,un- 3As a robustness exercise, we consider several other time frames around FOMC communications, over which we collect and score our stories. 4In recent years, the SEP have been released at the same time with the FOMC statement. See Appendix A for more details. 3

conventional monetary policy periods of 2009-2015, the COVID-19 pandemic, and post-pandemic sample 2020-2022. We find that most of the explanatory power of SI surprises is concentrated in the pre-GFC period, and we also find that SI surprises have explanatory power during the post-COVID period. Our index is relatively mute during the unconventional monetary policy period. This result may be attributed in part to the fact that policy communications became notably more transparent in our second sample period: In April 2011 the Federal Reserve then- Chairman Bernanke started conducting press conferences following the conclusions of the FOMC meetings after every other meeting, and Chairman Powell currently conducts press conferences after conclusion of every FOMC meeting. In addition to comparing our index’s ability to explain asset price movements in narrow windows around FOMC meetings, we also compare the performance of our SI to several alternative indexesdevelopedbyotherresearchers, mostlyatothercentralbanks. Wereconstructthemethodologies of Lucca and Trebbi (2009, developed at the Federal Reserve Bank of New York), Cannon (2015, developed at the Federal Reserve Bank of Kansas City), Carvalho et al. (2016, developed at the Federal Reserve Bank of San Francisco), Apel and Grimaldi (2012, developed at the Riksbank), and Nyman et al. (2018, developed at the Bank of England). We compare the various indexes and assess their informational content for explaining asset price movements. The rest of the paper is organized as follows. Section 2 describes the financial press stories and the financial market data sets. Section 3 describes our methodology and construction of our FOMC communications sentiment index, Section 4 discusses our index in terms of both its level and its surprises, Section 5 discusses our regression results with the index surprises and provides a comparison with other indexes, and Section 6 concludes. 2 Data 2.1 Factiva corpus of relevant media articles Our sample starts with the May 1999 FOMC meeting when the Committee first began issuing statements following each FOMC meeting, irrespective of whether the Committee had changed the stance of monetary policy or not.5 Our sample ends with the November 2022 FOMC meeting. In total, there are 188 scheduled FOMC meetings in our sample.6 5PriortoMay1999,theFOMConlyreleasedastatementwhenithaddecidedtochangethestanceofmonetary policy. The FOMC began to do so following the February 4, 1994 meeting. Before that, the FOMC had not issued any statements at all. 6The FOMC meeting calendar and FOMC statements can be found at https://www.federalreserve.gov/ monetarypolicy/fomccalendars.htm. During our sample period the FOMC has also held several unscheduled meetings. Wedidnotincludetheseinouranalysis,becauseourmethodologyrequirestohavenewsstoriespublished the before the FOMC statement release in order to construct the surprise in press sentiment around FOMC 4

We select corpus for our study from articles available in Dow Jones Factiva, a global news database that contains content from over 30,000 sources, including licensed publications, websites, and blogs. We focus on six outlets: the Associated Press, the Financial Times, the New York Times, Reuters News, the Wall Street Journal (both in print and online), and the Washington Post. The reason for this specific and, arguably, limited selection of sources is that these major outlets had consistently covered central banking communications since the start of our sample in 1999.7 The following Factiva filters are used to select a corpus for our study. First, for the paper to be included in our sample, it must contain one or more of the following keywords: FOMC, Federal Reserve, or Federal Open Market Committee. Second, we exclude stories that discuss non-financial industries (or industries not directly related to the financial sector), such as agriculture; automotive; basic materials and resources; energy; health care and life sciences; industrial goods; leisure, arts and hospitality; media and entertainment; real estate and construction; retail and wholesale; technology; telecommunication services; transportation and logistics; utilities; and consumer goods. Third, we exclude articles published in all regions other than the United States. We select Factiva stories in a five-day window around each FOMC statement release, from Sunday midnight to Friday midnight of the week when the FOMC meeting takes place. We then narrow down the event window to four alternative shorter windows over which we select stories for the SI construction. Figure 2 shows the timeline for each event window around the FOMC statement release on Wednesday at 2:00 p.m.8 The starting time for each event window is Sunday midnight. The set of stories that fall within thewindowfromSundaymidnighttothereleaseofthestatementconstitutesoursampleof“before stories” (indicated by the red section of the timelines in Figure 2). — stories written before the release and should contain expectations about the upcoming FOMC communications. The end time of considered event windows varies: midnight on Thursday, midnight on Wednesday, close-ofbusiness on Wednesday, and 30 minutes after the release on Wednesday. Naturally, these “after” windows have increasingly fewer stories in them (indicated by the blue section of the timelines in Figure 2).9 statement releases. 7Although our smaller set goes against the typical big data principle that more data tends to be better, we chose the smaller set of sources for their full-sample availability. In this way, we potentially reduce the presence of noise by excluding minor and less relevant news outlets. Furthermore, because we had to manually collect stories from Factiva, we leave testing of our index methodology on Factiva’s full set of sources for further research. 8We refer to the statement release day and time as Wednesday 2:00 p.m. for simplification in this Figure. However, the day and the timing of the releases of FOMC statements varied in our sample. We account for this details on our calculation of the sample and provide these details in Appendix A. 9We consider the stories in our sample that are published exactly at the time of the statement release (in the case of Figure 2, at 2:00 p.m.), as the ones published after the release as they were most likely written during the embargo period when certain members of the press already had access to the FOMC statement. 5

Our preferred event window is the middle one in the picture. It starts on Sunday midnight and ends on the FOMC meeting day midnight. This event window provides a reasonable balance between having stories that specifically describe the post-FOMC policy stance (and not affected by other events or shocks, such as macroeconomic releases that may sway the sentiment in wider windows shown in the top and second top of the Figure) and having a sufficient number of stories in the ”after” window to create an after-SI. Table 1 reports the number of stories for each of these event window specifications. The first line in each panel of this table reports the total number of stories downloaded from Factiva for six outlets and the number of stories for individual outlets in the corresponding event window. The second line reports the number of stories after we remove duplicate stories.10 The third line reports the number of stories in the corpus after we remove time non-stamped stories on the day of the FOMC, because they cannot be used in constructing surprises in the SI. The fourth and final line in each panel reports the number of remaining relevant stories — the stories that contain at least one of the words defined in the dictionary of entity words from table 2 and at least one of the topic keywords from table 3. In our preferred event window (from Sunday midnight to Wednesday midnight, reported in Panel C), our search criteria and filters result in a total of 8,878 stories released around the 188 FOMC meetings in our sample. On average, there are about 47 news stories per meeting from which we derive the FOMC-specific SI. Figure 3 shows granular information about the composition of our corpus. The stacked bars in Panel A shows the total number of stories per media outlet for each individual FOMC meeting. In general, we observe that the coverage of FOMC meeting-related news is quite dense. Press coverage of the FOMC communications increased dramatically during the financial crisis and again towards the end of 2015, when the FOMC began its most recent tightening cycle. Panel B provides information about the average number of stories per FOMC meeting per each individual mediaoutlet. Inoursample, Reuters Newshasthemostnumberofstoriespermeeting, onaverage, followed by the Associated Press, then the Wall Street Journal, the Financial Times, the New York Times, and lastly the Washington Post. Figure4showsthedistributionofthepublishedstoriesoverthe10-hourwindowfollowedbythe release of the FOMC statement for Reuters News, Associated Press, and the Wall Street Journal. The vertical grey bar indicates the timing of the press conference. Our post-meeting window captures a good amount of after stories, especially for Reuters News stories. As top panel shows, Reuters News mostly releases publications in the first half an hour after the statement release, while the peak of the stories published by Associated Press and the The Wall Street Journal occurs 10We define duplicates as stories with identical metadata, i.e., stories that have identical authors, titles, and an outlet. If the metadata is the same, we then check the cosine similarity of the frequency of words used in the story. If the cosine of the frequency vectors is greater than 0.95, then we flag the story as duplicate and remove it. 6

approximately between two to four hours of the statement release, as the middle and the bottom panels show.11 2.2 Financial market data We examine variation in several asset classes in relation to SI surprises and monetary policy (MP) surprises. For policy sensitive rates, we use implied rates on federal funds (FF) futures contracts one to six months ahead and Eurodollar (ED) futures contracts 1 to 4 quarters ahead. For nominal Treasury securities, we use yields on 3- and 6-month Treasury bills, as well as yields for 2-, 5-, 10-, and 30-year on-the-run nominal Treasury securities. For real rates, we use 5- and 10-year yields on Treasury Inflation-Protected Securities (TIPS). We also look at 5-, 10-year, and 5-year 5 years ahead inflation compensation, defined as the difference between nominal and TIPS yields of comparable maturities. For equity prices, we use Standard & Poor’s (S&P) 500 and NASDAQ Stock Market equity indexes. Finally, we also look at the responses of Bloomberg’s DXY dollar index, as well as the euro-dollar (EURO), and yen-dollar (YEN) currency pairs.12 We look at the intraday changes in these securities around the FOMC statement releases. All intraday data is from Bloomberg. We examined three intraday windows in measuring changes in asset prices: a tight event window, defined as the change in quotes 10 minutes before the release of the FOMC statement to 20 minutes afterwards; a wide event window, defined as the change in quotes 15 minutes before the release of the FOMC statement to 45 minutes afterwards; and a close-ofbusiness (COB) window, defined as the change in quotes 10 minutes before the FOMC statement release to close-of-business (COB), which we take as 3:45 p.m. Tight and wide windows follow Gurkaynak et al. (2005). The COB window also captures the market reaction to the post-meeting press conference in the second half of our sample. Our main results below are presented for the tight event window.13 3 Methodology In this section we discuss how we build our corpus of relevant articles from the financial press and, from that, construct our FOMC communications sentiment index. In Section 3.1 we discuss the construction of SI: Specifically, we outline a selection of entity 11We find that after the FOMC day and up to the end of the FOMC week, a significant number of additional stories that cover monetary policy communications, become available. However, these stories are outside of our preferred window. 12The DXY dollar index is defined as the spot U.S. dollar (USD) rate relative to a weighted average of six major currencies (euro, Japanese yen, British pound, Canadian dollar, Swedish krona, and the Swiss franc). 13Results that use wide and COB event windows are used as robustness checks. They are not reported but available upon request. 7

keywords that determine relevant sentences in the stories, a selection topic and modifier keywords, we describe how we determine the backward-looking and forward-looking sentences, and how we deal with negations. In Section 3.2 we discuss how we construct sentiment surprises. 3.1 Sentiment index construction We process each story in our sample separately as we find first relevant sentences in each story. For each story we first remove stop words and punctuation and then lemmatize the remaining words, as it is commonly done in the NLP literature. Specifically, we use the WordNetLemmatizer available in the Natural Language Toolkit (NLTK) module in Python.14 As the first step, we find sentences that contain entity words. If the sentence contains an entity word, we look for a topic-specific keyword. Next, if the sentence also contains a topic-specific keyword, we look whether the temporal orientation (backward- or forward-looking) is correct (according to whether the sentence is belongs to the story in the before- or the after-FOMC category) following the appropriate dictionary. If it is correct, we keep this relevant sentence, if not, we discard it. We then look for modifier keywords from the topic-specific dictionary and negotiations that change the semantic orientation of the modifier. Figure 1 presents the summary of these methodological step. Below, we describe these steps in detail: Entity keywords. Table 2 reports the entity keywords and their overall count in our sample. We apply entity words to search for relevant sentences. For our analysis we keep only those sentences in the stories that have at least one entity word in them. The entity words are supposed to ensure that the topic that we are picking is related to monetary policy communications specifically. Most common entity words are fed, federal reserve, and official found 21,465, 5,715, and 2,735 times, respectively, in our sample. Topic keywords. After the sentences that contain the entity keywords have been selected, we search for the topic-specific keywords in these sentences. We build a separate dictionary for a separate topic. There are a total of four dictionaries associated with four topics, specifically: General, Policy Rates, Asset Purchases, and Forward Guidance. Table 3 reports the topicspecific keywords in four dictionaries and their count in the sample. The topic keywords are usedasacontext-switchingtoolthatnextpointtoaspecificdictionaryformodifierkeywords (which we characterize as either hawkish or dovish words). As was discussed in Section 1, the verb increase can have a hawkish meaning in a policy-rate context, whereas it can have 14For a description of NLTK WordNetLemmatizer module, see https://www.nltk.org/_modules/nltk/stem/ wordnet.html and Miller (1995). 8

a dovish meaning in an asset- purchases context.15 As table 3 reports, the Policy Rates dictionary has 25,167 keywords followed by the General dictionary (16,147 keywords), then by the Asset Purchases dictionary (9,555 keywords), and, finally, by the Forward Guidance dictionary (4,229 keywords). Forward- and backward-looking rules. Once we determined to which topic a particular sentence belongs, we look at the temporal allocation of this sentence, namely, whether the sentence is forward- or backward looking. We proceed as follows. First, we divide all stories around each FOMC meeting into those that occur before the meeting — the before-stories — and those that occur after the meeting — the after-stories — by looking at the date and the time stamp of each story. Next, we apply a classification function to ensure that the before stories reflect only forward-looking information. However, we keep relevant sentences with both backward- and forward-looking directionality in the after stories to reflect the surprise about the past meeting as well as surprises about expected future policy actions (a so-called path surprise). Thus, the before stories should reflect only expectations about upcoming and future meetings at t, t+1, t+2, and so on. While, the after stories should reflect both information about the most recent meeting or policy action at time t as well as the path surprise about the meetings at t+1, t+2, and so on.16 To determine the directionality of a particular sentence, we have two dictionaries of words that are used to classify a sentence as either forward- or backward- looking. These dictionaries are listed in table 5. For forward-looking classification of sentences, we run two independent checks: • If a sentence has any word that is in our forward-looking dictionary, it is classified forward-looking. • If a sentence has a modal verb (will, could, might, may, should, can, and must) — not accompanied by another verb in past tense (e.g., could have) — it is classified as forward-looking. For backward-looking classification of sentences, we run three checks: • The sentence must not have been classified as forward-looking. • If the sentence has a word in our backward-looking dictionary, it will be classified as backward-looking. • If a sentence has a past-tense verb, it will be classified as backward-looking. 15This is in contrast to Gardner et al. (2022) who keep the same modifier-keywords dictionary for each topickeyword they study. 16Also here our methodology differs from Lucca and Trebbi (2009) who simply drop the past tense to remove any reference to the policy action at time t in all of their stories. 9

If a sentence is not classified as either forward- or backward-looking, it is dropped and not considered relevant. Modifier keywords and negations. Table 4 reports the counts of each of the hawkish and dovish words in the sample. They are grouped by the context. Conditional on the contextswitch, hawkish words increase the sentiment of a story (leaning towards more positive SI), while dovish words decrease the sentiment of a story (leaning towards more negative SI). For each modifier dictionary, we also report the total number of negations that would switch hawkish characterization into dovish characterization and vice verse. Words like not, and verbs than end with n’t make up most of the negation phrases that occur in our corpus. Our strategy for handling negations is simple: the sentence is searched up to three words before the modifier word, and if a negation is detected, the orientation of the modifier word is flipped (multiplied by -1). Similar to negations, the polarity of nearby words may flip the meaning of a modifier word but we did not find that polarity checks changed our results.17 Using a set of relevant sentences defined above in a story s that has W words in total, and H s S and D hawkish and dovish words, we construct a sentiment score SL for each story as: s S (cid:18) (cid:19) (cid:18) (cid:19) (cid:18) (cid:19) H D W +H s s s s SL ≡ log +1 −log +1 = log (1) s W W W +D s s s s with SL the sentiment level for story s and W the total number of words in the story.18 We have s s thus defined SL as the log-difference between the number of hawkish and dovish words relative s to the length of the story. Wethendefinetheoverallsentimentindexlevelformeetingtastheaverageofthestory-specific sentiment levels SL during the chosen time window: s 1 (cid:88) SI ≡ SL (2) τ s N τ s∈Sτ Inconstructingoursentimentindexthiswaywescoreeachstoryseparatelythereforeandweight stories equally. As a result, our approach is different from many studies in the literature that do not distinguish between stories and simply use a bag-of-sentences approach where sentences are 17For example, consider a hypothetical sentence “The Fed failed to raise rates.” The verb raise by itself is hawkish, but the preceding negative polarity word failed changes the meaning of the phrase to being dovish. In contrast, if one of the modifier keywords is preceded by a positive polarity word, the meaning of a that modifier will not be changed. As an example, consider the hypothetical sentence “The Fed successfully raised rates”. Here the adjective successfully has positive polarity. Therefore, it will be ignored. In our corpus, there were very few instances of modifier orientation changes based on polarity (via Loughran and McDonald (2011) dictionary). The index does not significantly change based on the presence or absence of a polarity check or changing polarity dictionaries. 18We add one in the log function to preclude situations where a story has only hawkish or only dovish words. 10

weighed equally in the overall corpus. Instead, for each story we weight the number of hawkish or dovish words by the total number of words in the story. We find such a weighting scheme more intuitive as it takes into account the context of the story in which hawkish and dovish words appear. For example, once would expect a single relevant hawkish or dovish sentence in an otherwise irrelevant story to be less informative about perceived FOMC sentiment than a story of the same length that contains multiple relevant sentences. 3.2 Sentiment index surprises construction Whereas we average over all the stories in a window for determining the sentiment level, for the surprises in the SI index we distinguish between the stories published before and after the FOMC statement release. In particular, we compute SI surprise as the difference between SI and before SI , the SI levels computed for the corpus stores published before the FOMC statement release after and afterwards in a specific window: 1 (cid:88) 1 (cid:88) ∆SI ≡ SI −SI = SL − SL , (3) τ after before s s N N after before s∈S s∈S after before with S , N and S , N referring to the set and the number of before and after before before after after stories, respectively, for a meeting τ. In computing surprises (3) we discard all stories from our sample that are released on the day of the FOMC statement without a time stamp. We do so because we cannot determine with certainty whether a particular article came out before or after the statement release. The third rows in each of the panels of table 1 shows the effect to the size of the our corpus of filtering out these time-non-stamped stories.19 4 Sentiment index and surprises Figure 5 shows the SI level in Panel A and SI surprises in Panel B, following the base-case event window from midnight on Sunday to midnight on the day of the statement release. For illustrative purposes, index levels in Panel A are scaled to be between -1 and 1, with zero indicating a neutral sentiment stance, and -1 the most dovish sentiment level across all FOMC meetings. Surprises in Panel B are standardized and expressed in standard deviation units. The grey bars represent NBER-identified recessions. 19Inourfive-daycorpusofno-duplicatestories,thereare11storiesoutof18,271storiesthatarenottimestamped on the same day as the FOMC statement release, which constitutes only about 0.06 percent of our sample. Note that the lack of an accurate release time is specific to the Financial Times and the New York Times. Stories from all other outlets tend to be properly time-stamped. 11

The level of our index appears to accurately captures the state of the business cycle and the stance of monetary policy during that cycle. Prior to recessions, the stance of monetary policy is perceived as hawkish as the FOMC is hiking rates. However, when a recession approaches, sentiment about the FOMC’s policy stance quickly turns sharply dovish and bottoms out during recessions as the FOMC aggressively cuts rates and then keeps rates low to spur an economic recovery. There are two interesting periods to highlight in particular. First, during the ZLB period approximately from 2010 to late 2014, the sentiment index hovers around neutral. This is somewhat surprising given that the FOMC actively used a range of accommodative policy tools during that period, in particular, asset purchases and forward guidance. It is likely that this period was also marked by an increasing number of central banking communications aimed to clarify the policies and increase transparencies. Second, at the end of the sample there is a sharp reversal in sentiment from being hawkish (positive sentiment) to becoming dovish (negative sentiment). That accurately coincides with the recent chances in the FOMC policy stance from tightening policy in 2018 to easing in 2019.20 The SI subsequently reversed its trend and became positive when the Fed engaged in the aggressive post-COVID tightening cycle with the aim to curb historically high inflation. Interestingly, our sentiment index appears to lead U.S. recessions. It is falling to negative territories shortly before the start of recessions. Sentiment indexsurprisessimilarlyappearintuitive. Goingintorecessions theFOMCsurprises markets on the dovish side. For example, in the middle and second half of 2018 and in the post-COVID period of 2022 we see more positive surprises, meaning that the financial press was surprised by the more-hawkish-than-expected stance of policy. This lines up well with financial market commentary around some of the FOMC meetings in the latter part of 2018. 5 Informational content of surprises in the SI We next explore the informational context of our SI level surprises for explaining movements in a broad range of asset prices around the FOMC meetings. Our empirical setup is motivated by the findings of Bernanke and Kuttner (2005) who document that equity prices significantly react to changes in monetary policy. To that end, in our analysis we first gauge the reaction in policy-sensitive derivatives, such as money market futures rates, nominal and inflation-linked Treasury yields, measures of TIPS-based breakeven inflation, equity prices, and several currency pairs relative to the dollar, to changes (surprises) in the SI. Finding notable significance, we then run a second round of regressions where we control for traditional financial-market-based measures of monetary policy surprises. 20The Fed cut policy rates for the first time in ten years since 2009 at the July 2019 FOMC meeting. 12

5.1 Empirical results Our benchmark regression is as follows. Let ∆y be the change in policy-sensitive rates, yields, t equity prices, or currency prices in the event window.21 Again, in the base-case the event window runs from midnight on Sunday to midnight on the day of the FOMC meeting. We then regress ∆y = α+β ∆SI +β MP +(cid:15) , (4) t 1 τ 2 t t where MP is the monetary policy shock used in Bauer and Swanson (2023), measured as the t orthogonalized monetary policy surprises in the first principal component of the one- to fourquarter ahead Eurodollar(ED) futures rates in the 30-minute window around the FOMC meeting announcements starting 10 minutes before each FOMC announcement and ending 20 minutes afterwards.22 The stochastic error term (cid:15) captures the effect of other factors that influence the t asset in question. The changes in asset prices are defined from intraday quotes in the same event window.23 We also include postGFC and ZLB dummy variables in our basic regressions, to account for different economic environments following the GFC and during ZLB periods: ∆y = α+β ∆SI +β MP +λ +(cid:15) , (5) t 1 τ 2 t Event t where λ = {λ ,λ }. λ dummy is defined as one from the December 2008 Event postGFC ZLB postGFC FOMC meeting to the November 2022 FOMC meeting and zero otherwise and λ dummy is ZLB defined as one from the December 2008 FOMC meeting to the October 2015 FOMC meeting and from the April 2020 FOMC meeting to the January 2022 FOMC meeting and zero otherwise. We run regressions (4) and (5) using SI and MP surprises. Sections 5.1.1 and 5.1.2 describe the results of tables 6 through 10. Panels A of these tables report regression results of various asset classes with respect to SI and MP surprises for the full sample period that correspond to regression (4) results. Panels B and C of these tables report regression (5) results, that account for either the post-GFC period or the ZLB bound, respectively. In addition, we report regressions (4) results for specific sub samples in Appendix B. 21Changes for rates are all recorded in basis points, while changes in equities and currencies are computed in percent. 22The orthogonalization is conducted with respect to changes in macroeconomic and financial variables around the FOMC window. The authors argue that such monetary policy surprises are a cleaner and a more informative measureofmonetarypolicysurpriseswheninformationthatispotentiallyrelatedtochangesinmacroeconomicand financial variables around the FOMC announcement is removed from these surprises. For a detailed description of the MP measure, see Bauer and Swanson (2023). t 23We also conducted robustness checks using the wider event window definition and obtained broadly similar results. Results are available upon request. 13

5.1.1 SI surprises and changes in policy-sensitive rates Table6reportsbaselineregressionresultsfrom(4)forthefederalfundsfutures(FF1,FF2,...,FF6) contracts one to six months ahead. The left-hand side variable of the regression (4) corresponds to changes in the federal funds futures contracts 10 minutes before to 20 minutes after the release of the FOMC statement. First, as panel A shows, the response of the fed funds futures contracts reflects the surprise in sentiment; the sentiment surprise coefficient β is positive and statistically 1 significant for fed funds futures contracts four to six months out (FF4 to FF6), despite a strongly significant coefficient for monetary policy surprise at all six horizons. In other words, positive (or, hawkish) sentiment surprises are associated with positive movements in future policy-sensitive rates and negative (or, dovish) sentiment surprises are associated with negative movements in future policy-sensitive rates. Second, the response of the fed funds futures contracts to a change in media sentiment following the FOMC announcement is statistically significant at the five-percent level for FF5 and FF6 and at the ten-percent level for FF4 contracts. The MP surprise drives out significance of the sentiment surprise for the very near term contracts, namely, FF1 through FF3, but not for contracts at longer horizons. Interestingly, the explanatory power of the sentiment and monetary policy surprises increases with horizon and notably higher at longer (FF4 through FF6) than ta the shorter (FF1 through FF3) horizons. Our results are robust to accounting for the GFC and ZLB dummies, which suggests that the results are not driven by these particular periods in our sample. Table 7 reports results for the Eurodollar futures ED1 through ED4, one to four quarters ahead. These results broadly follow results reported in table 6, but these results are even stronger. The SI surprise is strongly significant for contracts ED2 through ED4 at the 1 percent level of statistical significance. Further, the SI surprise is not driven away by the MP surprises. The R2 in these regressions is very high, from 67 to 76 percent, across horizons. Like in table 6, results in table 7 are robust when we account for the GFC or ZLB periods. Overall, the results in tables 6 and 7 are consistent as they both point out to the significance of information contained in SI beyond the very near-term (ED2 to ED4 contracts). This suggests that SI surprises have a complimentary information for the “path” part of changes in the policysensitive rates, while MP surprises are the main driver in changes in the ED1 contract. 5.1.2 SI surprises and changes in other asset classes prices Table 8 reports regression (4) results for nominal Treasury yields implied by the 2-, 5-, and 10year Treasury futures contracts. We find that SI surprises around FOMC announcements are important for predicting movements in nominal Treasury yields at all considered horizons, despite strong significance of MP shocks and regardless of the sample period (that is, whether we control 14

for the postGFC or ZLB periods, or not). The predictive content in both Si and MP diminishes with horizon as R2 falls from about 70 percent for the 2-year Treasury futures contract to about 33 percent for the 10-year contract. Table 9 reports regression results for one-the-run 5- and 10-year TIPS yields and 5- and 10-year inflation compensation as well as five-year fine-year ahead inflation compensation. SI surprises appear to explain movements in 5- and 10-year TIPS yields regardless of the sample period, but not in the inflation compensation series. Interestingly, MP surprises are strongly significant in explaining variation in TIPS yields but not in inflation compensation at any horizon. There are may be two reasons for that. First is a mechanical reason: both nominal yields and TIPS yields react to SI and MP surprises with similar regressions coefficients, especially for 5-year yields, so their effect on inflation compensation could be washed out because inflation compensation is defined as nominal minus TIPS yields of comparable maturities. Second reason is an economical one: insignificance of the MP and SI surprises for explaining changes in inflation compensation around the FOMC meeting announcements serves as an evidence of relatively well anchored inflation expectations, in general, in our sample. This result is not changed when include postGFC or ZLB dummies in the regressions, as shown in Panels B and C, respectively. Table 10 reports regression results related to changes in the S&P 500 and NASDAQ equity indexes as well as changes in the DXY, EURO, and YEN. First, we find that SI surprises matter little explaining movements in equity prices around the FOMC announcement windows and that MP surprises are very strong with a negative coefficient sign. This means that in our sample, on average, monetary policy acts strongly via a discount rate channel: A positive monetary policy surpriselowersequitypricesbecauseofhigherdiscountrates, andamore-restrictive-than-expected monetary policy. According to the results in Panels B and C, λ and λ are slightly postGFC ZLB positive post-GFC or ZLB periods, thus, in those subperiods, stock prices fall slightly less, than in the usual periods of monetary policy. These findings are directionally consistent with Bernanke and Kuttner (2005) who show that the unanticipated tightening (or easing) in monetary policy is associated with a decline (or increase) in broad stock indexes. Turning to regressions with currency indexes on the left-hand side, we find that SI surprises explain changes in DXY index with a positive sign, regardless of whether we control for the postGFC and/or ZLB periods or not. A positive surprise in sentiment increase the value of the dollar. We also find that positive sentiment surprise decreases the value of EURO index. The value of YEN index is relatively mute to SI surprises. In Appendix B, we take a micro view of SI relevance for explaining variation in asset prices and report regression results for specific six subsample of the full sample period from May 1999 to November 2022. 15

5.1.3 Comparison with other indexes InadditiontoconstructingSIthattakesintoaccountdifferentperiodsofunconventionalmonetary policy periods, we replicated five alternative methodologies developed by economists at various central banks and run regressions (4). We summarize these methodologies and regression results briefly in Appendix C. The indexes constructed using five available methodologies are shown in Figure C1.24 We find that our index is most highly correlated with the BoE index (correlation is around 0.8), then with the LT index (correlation is around 0.6), then with the KC Fed (correlation is around 0.5). The correlation of our index surprises with either Riksbank or SF Fed are very low. 6 Conclusion In this paper we constructed a sentiment index of monetary policy communications using texts in financial press articles published around the FOMC meetings from the May 1999 FOMC meeting to November 2022 FOMC meeting, a period that covered 188 scheduled meetings. We found that surprises in the sentiment index around the FOMC meetings have additional explanatory power for movements in asset prices, beyond a financial market-based measure of the monetary policy surprises. Our sentiment surprises also have some informational content for asset price movements in the post-COVID period. We have compared our results with other alternative indexes and find thatourindexhashigherexplanatorypowerformovementsinvariousfinancialassetpricesaround FOMC meetings, even when we control for financial market-based measures of monetary policy surprises. The reason for our sentiment index being more informative than existing sentiment indexesisthatweconstructedpolicies-specificdictionariesthattreatconventionalmonetarypolicy, quantitative easing, and forward guidance policies separately. Each of these dictionaries calls for a different set of modifiers of the monetary policy stance, either hawkish or dovish. This methodology was not considered before and should be considered when various sentiment indexes are constructed with the aim of obtaining additional information about monetary policy stance beyond financial market-based measures. 24By plotting the Riksbank index, we subtracted one from the Riksbank index to make it more comparable to other indexes. 16

References Acosta, Miguel, 2023, A new measure of central bank transparency and implications for the effectiveness of monetary policy, International Journal of Central Banking 19. Acosta, Miguel, and Ellen E. Meade, 2015, Hanging on every word: Semantic analysis of the FOMC’s postmeeting statement, FEDS Notes, September 30, 2015. Apel, Mikael, and Marianna Grimaldi, 2012, The information content of central bank minutes, Riksbank Research Paper Series No. 92. Azar, Pablo, and Andrew W. Lo, 2016, The wisdom of twitter crowds: Predicting stock market reactions to FOMC meetings via Twitter feeds, Journal of Portfolio Management 42, 123–134. Baker, Scott R., Nicholas Bloom, and Steven J. Davis, 2016, Measuring economic policy uncertainty, The Quarterly Journal of Economics 131, 1593–1636. Bauer, Michael D., and Eric T. Swanson, 2023, A reassessment of monetary policy surprises and high-frequency indentification, NBER Macroeconomics Annual 37, 87–155. Bernanke, Ben, 2004, Fedspeak, Remarks by Governor Ben Bernanke at the Meetings of the American Economic Association, San Diego, California. Bernanke, Ben S., and Kenneth N. Kuttner, 2005, What explains the stock market’s reaction to Federal Reserve policy?, Journal of Finance 60, 1221–1257. Blinder, Alan S., Michael Ehrmann, Marcel Fratzscher, Jakob De Haan, and David-Jan Jansen, 2008, Central bank communication and monetary policy: A survey of theory and evidence, Journal of Economic Literature 46, 910–45. Born, Benjamin, Michael Ehrmann, and Marcel Fratzscher, 2014, Central bank communication on financial stability, The Economic Journal 124, 701–734. Caldara, Dario, and Matteo Iacoviello, 2022, Measuring geopolitical risk, American Economic Review 112, 1194–1225. Cannon, San, 2015, Sentiment of the FOMC: Unscripted, Economic Review of the Federal Reserve Bank of Kansas City 100, 5–31. Carvalho, Carlos, Eric Hsu, and Fernanda Nechio, 2016, Measuring the effect of the zero lower bound on monetary policy, Federal Reserve Bank of San Francisco working paper 2016-06. Druz, Marina, Ivan Petzev, Alexander F. Wagner, and Richard J. Zeckhauser, 2020, When managers change their tone, analysts and investors change their tune, Financial Analysts Journal 76. Gardner,Ben,ChiaraScotti,andClaraVega,2022,Wordsspeakasloudlyasactions: Centralbank communication and the response of equity prices to macroeconomic announcements, Journal of Econometrics 231, 387–409, Special Issue: The Econometrics of Macroeconomic and Financial Data. Gentzkow, Matthew, Bryan T. Kelly, and Matt Taddy, 2019, Text as data, Journal of Economic Literature 57, 535–574. 17

Gurkaynak, Refet, Brian Sack, and Eric Swanson, 2005, Do actions speak louder than words? the response of asset prices to monetary policy actions and statements, International Journal of Central Banking 1, 55 – 93. Hansen, Stephen, Michael McMahon, and Andrea Prat, 2018, Transparency and deliberation within the FOMC: A computational linguistics approach, The Quarterly Journal of Economics 133, 801–288. Hernandez-Murillo, Rub´en, and Hannah G. Shell, 2014, The rising complexity of the FOMC statement, Federal Reserve Bank of St. Louis Economic Synopsis No. 2014-23. Heston, Steven L., and Nitish R. Sinha, 2020, News versus sentiment: Predicting stock returns from news stories, Financial Analysts Journal 73, 67–83. Husted, Lucas, John Rogers, and Bo Sun, 2020, Monetary policy uncertainty, Journal of Monetary Economics) 115, 20–36. Loughran, Tim, and Bill McDonald, 2011, When is a liability not a liability? Textual analysis, dictionaries, and 10-ks, Journal of Finance 66. Lucca, David, andFrancescoTrebbi, 2009, Measuringcentralbankcommunication: Anautomated approach with application to FOMC statements, NBER working paper 15367. Meade, Ellen E., Nicholas A. Burk, and Melanie Josselyn, 2015, The FOMC meeting minutes: An assessment of counting words and the diversity of views, FEDS Notes, May 26, 2015. Miller, George A., 1995, Wordnet: A lexical database for english, Communications of the ACM 38, 39–41. Nyman, Rickard, SujitKapadia, DavidTuckett, DavidGregory, PaulOrmerod, andRobertSmith, 2018, Newsandnarrativesinfinancialsystems: Exploitingbigdataforsystemicriskassessment, Bank of England staff working paper No. 704. Romer, Christina D., and David H. Romer, 2004, A new measure of monetary shocks: Derivation and implications, American Economic Review 94, 1055–1084. Schmanski, Bennett, Chiara Scotti, Clara Vega, and Hedi Benamar, 2023, Fed communication, news, Twitter, and echo chambers, https://ssrn.com/abstract=4458545. Schmeling,Maik,andChristianaWagner,2025,Doescentralbanktonemoveassetprices?,Journal of Financial and Quantitative Analysis 60, 36–67. Shapiro, Adam Hale, Moritz Sudhof, and Daniel J. Wilson, 2022, Measuring news sentiment, Journal of Econometrics 228, 221–243. Sharpe, Steven A., Nitish R. Sinha, and Christopher A. Hollrah, 2023, The power of narrative sentiment in economic forecasts, International Journal of Forecasting 39, 1097–1121. Woodford, Michael, 2005, Central bank communication and policy effectiveness, NBER working paper 11898. 18

Figure 1: Methodology Summary for Identifying and Determining the Semantic Direction of a Relevant Sentence Entity words (cid:63) Topic key words (cid:16)(cid:80) (cid:16) (cid:80) (cid:16) (cid:0)(cid:64) (cid:80) (cid:16) (cid:80) (cid:16) (cid:80) (cid:16) (cid:0) (cid:64) (cid:80) (cid:16) (cid:80) (cid:16) (cid:80) (cid:16) (cid:0) (cid:64) (cid:80) (cid:16) (cid:80) (cid:16) (cid:80) (cid:16) (cid:0) (cid:64) (cid:80) (cid:16) (cid:80) (cid:16) (cid:80) (cid:16)(cid:41)(cid:16) (cid:0)(cid:9)(cid:0) (cid:64)(cid:64)(cid:82) (cid:80)(cid:80)(cid:113) Policy Asset Forward General rates purchases guidance Temporal allocation of a sentence: backward or forward-looking sentence. If a sentence is temporally allocated, proceed to find topic-specific modifiers (cid:63) (cid:63) (cid:63) (cid:63) Topic-specific Topic-specific Topic-specific Topic-specific modifiers modifiers modifiers modifiers and negations and negations and negations and negations Notes: The figure presents the methodology summary for identifying and computing the semantic orientation of relevant sentences in financial media coverage of monetary policy communications around the Federal Open Market Committee (FOMC) statement releases after scheduled meetings, according to the description of Section 3. 19

Figure 2: Timing of the Sentiment Index Construction FOMC Statement release Sun midnight Fr midnight Before After FOMC Sun midnight Th midnight Before After FOMC Sun midnight Wed midnight Before After FOMC Sun midnight Wed COB Before After FOMC Sun midnight Wed FOMC + 30 minutes Before After Notes: This figure shows five alternative event windows around the release of the Federal Open Market Committee (FOMC) statement, for which the surprise in the sentiment index constructed. The timing of the start point for each of the five event windows is Sunday midnight before the statement release. The period between the release of the FOMC statement and the end of the event window differs. Going from top of the page to the bottom: For the event window shown on top, the end point corresponds to Friday midnight after the FOMC statement release; For the second event window – to Thursday midnight after the statement release; For the event window shown in the middle – to Wednesday midnight after the statement release; For thefourtheventwindow– toWednesday, 3:45p.m.; Fortheeventwindowshownatthebottom– to(usually Wednesday) 30 minutes after the statement release. The time before and after the statement release is shown in red and in blue, respectively. 20

Figure 3: Distribution of News Sources and Corpus Stories Panel A: Total number of Stories 160 Reuters News 150 The Wall Street Journal Associated Press 140 Financial Times The New York Times 130 The Washington Post 120 110 100 90 80 70 60 50 40 30 20 10 0 2000 2002 2004 2006 2008 2010 2012 2014 2016 2018 2020 2022 Panel B: Average number of stories per outlet per FOMC meeting 45 40 37 35 30 25 20 15 8 8 10 5 2 1 1 0 Reuters News The Wall Street Journal Associated Press Financial Times The New York Times The Washington Post Notes: The figure shows the media coverage of monetary policy communications around the Federal Open Market Committee (FOMC) statement releases after scheduled meetings. The top panel shows the total number of Factiva stories in our sample per media source and per FOMC meeting. The bottom panel reports the average number of stories per media source per FOMC meeting. The stories in our sample are collected from Factiva using the search criteria described in Section 2.1. The sample period consists of the 188 scheduled FOMC meetings between May 18, 1999, and November 2, 2022. Source: Dow Jones, Factiva; authors’ calculations. 21

Figure 4: Distribution of Post-meeting News Publications Reuters News Hours After FOMC seirotS fo rebmuN 120 Press Conference 80 COB 40 0 0 2 4 6 8 10 Associated Press Hours After FOMC seirotS fo rebmuN 30 Press Conference COB 20 10 0 0 2 4 6 8 10 The Wall Street Journal Hours After FOMC seirotS fo rebmuN 12 COB Press Conference 8 4 0 0 2 4 6 8 10 Notes: The figure shows the temporal distribution of news publications up to 10 hours following the Federal Open Market Committee (FOMC) statement release, at the fiveminute intervals. The distribution is shown for outlets that always have time-stamped stories in our sample: Reuters News, Associated Press, and the Wall Street Journal, in PanelsA,B,andC,respectively. Theverticaltripwire“COB”oneachchartcorresponds totheclose-of-businesstiming. Theverticalgray-shadedbaroneachchartrepresentsthe timing of the Chair’s press conference. The sample period consists of the 188 scheduled FOMC meetings between May 18, 1999 and November 2, 2022. Source: Dow Jones, Factiva; authors’ calculations. 22

Figure 5: Sentiment Index and Surprises Panel A: Sentiment Level 1.0 0.5 0.0 −0.5 −1.0 2000 2002 2004 2006 2008 2010 2012 2014 2016 2018 2020 2022 Date Panel B: Sentiment Surprise 2 1 0 −1 −2 −3 2000 2002 2004 2006 2008 2010 2012 2014 2016 2018 2020 2022 Date Notes: This figure shows sentiment index (SI) level and SI surprises around the Federal Open Market Committee (FOMC) communications, in Panels A and B, respectively. The SI and SI surprises are shown for the event window that extends from Sunday midnight before the FOMC statement release to the midnight oftheFOMCmeetingday. Thesampleperiodconsistsofthe188scheduledFOMC meetings between the May 1999 FOMC meeting and the November 2022 FOMC meeting. The frequency is FOMC. The gray-shaded bars indicate the National Bureau of Economic Research recessions. Source: Dow Jones, Factiva; authors’ calculations. 23

Table 1: Summary of Financial Press Articles Number of articles by outlet Total Reuters AP WSJ FT NYT WP Panel A: 5-day window, FOMC week All 19687 11124 2174 3459 1770 759 401 Net Of duplicates 18271 10745 2095 3043 1395 649 344 Net of time-non-stamped 18260 10735 2094 3043 1395 649 344 Net of irrelevant stories 15911 9026 1848 2843 1255 624 315 Panel B: 4-day window: Sun midnight to Thu FOMC week All 16409 9229 1952 2881 1354 653 340 Net Of duplicates 15067 8908 1887 2465 981 543 283 Net of time-non-stamped 15056 8898 1886 2465 981 543 283 Net of irrelevant stories 13019 7407 1668 2294 869 522 259 Panel C: 3-day window: Sun midnight to FOMC midnight All 12015 7270 1596 1926 778 275 170 Net Of duplicates 10738 6999 1535 1511 414 165 114 Net of time-non-stamped 10736 6997 1535 1511 414 165 114 Net of irrelevant stories 8878 5644 1309 1341 342 148 94 Panel D: 3-day window: Sun midnight to COB FOMC All 11568 6939 1523 1883 778 275 170 Net Of duplicates 10322 6694 1467 1468 414 165 114 Net of time-non-stamped 10311 6684 1466 1468 414 165 114 Net of irrelevant stories 6971 4508 809 1070 342 148 94 Panel E: 3-day window: Sun midnight to 30-min after FOMC All 5774 3328 695 1055 417 165 114 Net Of duplicates 5624 3204 672 1055 414 165 114 Net of time-non-stamped 5613 3194 671 1055 414 165 114 Net of irrelevant stories 4459 2446 519 910 342 148 94 Notes: ThistablepresentsstatisticsofourcorpusofFactivaarticlesforeachofthesixoutletsinoursample: the Associated Press (AP), the Financial Times (FT), the New York Times (NYT), Reuters (RT), the Wall Street Journal(WSJ)(bothinprintandonline),andtheWashington Post(WP).PanelsAthroughEreport the number of stories used to construct SIs with various asymmetric windows around the day of the release oftheFOMCstatementduringtheweekoftheFOMC.Thestartingpointofallthewindowsismidnighton Sunday. TheendpointofthewindowsareFriday(PanelA),Thursday(PanelB),midnightonthedayofthe FOMC (Panel C), close-of-business on the day of the FOMC (Panel D), and 30 minutes after the statement releaseonthedayoftheFOMC(PanelE).Eachpanelreportsthetotalnumberofarticlesdownloadedfrom Factiva and the number of stories after i) removing duplicates, ii) removing non-time-stamped stories, and iii) removing stories that have no sentences with entity or topic-keywords or correct temporal directionality. The sample period consists of the 188 scheduled FOMC meetings between the May 1999 FOMC meeting and the November 2022 FOMC meeting. Source: Dow Jones, Factiva; authors’ calculations. 24

Table 2: Dictionary of Entity Words Entity Words Count fed 21465 federal reserve 5715 official 2735 central bank 2726 fomc 2452 powell 1491 committee 1019 yellen 876 policy maker 711 greenspan 507 bernanke 266 chairman 254 policymaker 42 chair 21 federal open market committee 7 federal openmarket committee 7 chairwoman 3 Notes: Thistablepresentsthedictionaryofentity words that are used to classify the sentences, to which hawk and dove sentences are applied. Source: DowJones,Factiva;authors’calculations. 25

Table 3: Dictionaryies of Topic Keywords Key Words No. Key Words No. Key Words No. Panel A: General Dictionary Panel B: Policy Rates Dictionary statement 5195 rate 14100 policy 4386 interest rate 6910 monetary policy 1566 target 1696 credit 1012 federal funds rate 764 action 861 funds rate 487 stimulus 841 target range 436 announcement 712 borrowing cost 377 stance 454 target rate 245 reserve 403 benchmark rate 151 liquidity 269 rate regime 1 policy accommodation 123 monetary stimulus 101 policy stance 84 accommodation 48 policy tool 33 easy money 30 monetary accommodation 16 open market operations 13 Total 16147 Total 25167 Panel C: Asset Purchases Dictionary Panel D: Forward Guidance Dictionary treasury 2418 qe 67 pace 1402 bond 1356 treasury note 27 near zero 845 security 862 securities holdings 26 pledge 432 balance sheet 587 mortgage securities 24 gradual 348 bond purchases 503 holdings of securities 22 guidance 299 purchase 502 reinvestment 17 patient 284 mortgage 469 purchase programs 16 extended period 183 asset 460 maturing securities 13 forward guidance 99 portfolio 409 bond purchase program 11 measured pace 98 bond buying 343 asset purchase program 7 considerable period 87 bond buying program 265 treasury portfolio 7 considerable time 84 quantitative easing 246 buying program 5 exceptionally low 66 treasury securities 212 mortgage backed securities 3 forward policy guidance 2 asset purchases 207 mortgage portfolio 3 treasury bond 183 asset backed securities 1 government bond 102 asset purchasing program 1 government debt 95 asset purchasing programs 1 purchase program 84 security holdings 1 Total 9555 Total 4229 Notes: This table presents the dictionary of topic keywords in four specific categories (General, Policy Rates, Asset Purchases, and Forward Guidance) in Factiva-filtered stories. Respective counts of key words in our sample are provided in the columns next to them. Source: Dow Jones, Factiva; authors’ calculations. 26

Table 4: Dictionaries of Keyword Modifiers Hawkish No. Dovish No. Hawkish No. Dovish No. Panel A: General Dictionary Panel B: Policy Rates Dictionary raise 1329 cut 1052 raise 5606 cut 4067 increase 731 lower 175 increase 3177 lower 925 hike 678 buy 134 hike 2884 fall 491 tighten 205 ease 127 boost 591 drop 412 boost 109 drop 118 gain 418 reduce 392 gain 91 fall 98 tighten 349 ease 300 reduce 79 reduce 97 grow 322 contract 167 firm 66 continue 94 firm 270 decrease 34 grow 57 contract 65 expand 122 loose 27 sell 54 extend 50 advance 75 taper 54 increase 39 expand 52 expand 34 cut 51 grow 20 slow 40 loose 20 shrink 35 decrease 9 trim 27 advance 20 decrease 4 negation 19 negation 68 negation 100 negation 278 Panel C: Asset Purchases Dictionary Panel D: Forward Guidance Dictionary raise 611 buy 745 raise 530 drop 113 increase 370 cut 423 increase 507 cut 112 reduce 352 continue 331 hike 325 continue 49 hike 327 increase 178 reduce 53 fall 47 taper 240 reduce 140 slow 53 buy 38 sell 204 lower 135 gain 52 reduce 35 cut 192 extend 129 boost 50 lower 26 slow 167 expand 108 grow 44 increase 17 shrink 153 fall 107 tighten 36 ease 15 boost 88 grow 85 expand 30 expand 14 gain 84 ease 57 taper 19 contract 9 trim 75 drop 56 firm 15 loose 6 firm 63 contract 15 cut 13 grow 5 tighten 53 decrease 7 advance 12 extend 4 expand 40 loose 3 shrink 9 decrease 1 grow 37 trim 6 advance 12 sell 4 decrease 11 decrease 2 negation 26 negation 64 negation 4 negation 16 Notes: Panels A, B, C, and D of this table present the respective dictionaries and the number of (stemmed) keyword modifiers in the Factive corpus of financial media articles, in each of four topic categories: General, Policy Rates, Asset Purchases, and Forward Guidance, respectively. The sample period consists covers scheduled meetings between the May 18, 1999 and November 2, 2022 meetings, a total of 188 meetings. The bottom row in each panel reports the number of hawkish and dovish words modified by a negation in a relevant keyword dictionary. Source: Dow Jones, Factiva; authors’ calculations. 27

Table 5: Dictionary of Backward- and Forward-looking Words Forward-looking words Backward-looking words expects after likely appeared next seem unlikely last anticipates yesterday possibly been believe have projecting has going follow forward peak outlook record warn development if beat see sold look held think experience sometime remain affecting doing attempt already envision points support show view Notes: Thistablespresentsthedictionaryofwordsthatdetermine the forward- or backward-looking directionality of a sentence. Source: Dow Jones, Factiva; authors’ calculations. 28

Table 6: Federal Funds Futures Contracts FF1 FF2 FF3 FF4 FF5 FF6 Panel A: No Dummies SI surprise 0.003 0.003 0.005 0.007∗ 0.008∗∗ 0.010∗∗ (0.002) (0.003) (0.003) (0.003) (0.003) (0.003) MP surprise 0.177∗∗∗ 0.389∗∗∗ 0.449∗∗∗ 0.521∗∗∗ 0.606∗∗∗ 0.671∗∗∗ (0.050) (0.089) (0.074) (0.067) (0.061) (0.064) Constant −0.004∗ -0.004 −0.007∗∗∗ −0.007∗∗∗ −0.007∗∗∗ −0.010∗∗∗ (0.002) (0.002) (0.002) (0.002) (0.002) (0.002) R2 0.190 0.286 0.428 0.533 0.600 0.572 Panel B: PostGFC Dummy SI surprise 0.003 0.003 0.005 0.007∗ 0.007∗∗ 0.010∗∗ (0.002) (0.003) (0.003) (0.003) (0.003) (0.003) MP surprise 0.177∗∗∗ 0.388∗∗∗ 0.448∗∗∗ 0.521∗∗∗ 0.606∗∗∗ 0.672∗∗∗ (0.051) (0.089) (0.074) (0.067) (0.061) (0.065) PostGFC 0.003 -0.001 -0.002 0.001 0.001 0.002 (0.003) (0.005) (0.004) (0.004) (0.004) (0.005) Constant -0.005 -0.004 -0.006 −0.008∗ −0.008∗ −0.011∗∗ (0.003) (0.004) (0.004) (0.003) (0.003) (0.004) R2 0.194 0.286 0.429 0.533 0.600 0.572 Panel C: ZLB Dummy SI surprise 0.003 0.003 0.005 0.007∗ 0.008∗∗ 0.010∗∗ (0.002) (0.003) (0.003) (0.003) (0.003) (0.003) MP surprise 0.179∗∗∗ 0.389∗∗∗ 0.452∗∗∗ 0.525∗∗∗ 0.609∗∗∗ 0.676∗∗∗ (0.051) (0.089) (0.074) (0.067) (0.061) (0.065) ZLB 0.003 0.001 0.005 0.006 0.005 0.009∗ (0.003) (0.004) (0.004) (0.004) (0.004) (0.004) Constant -0.005 -0.005 −0.009∗∗ −0.009∗∗∗ −0.009∗∗∗ −0.013∗∗∗ (0.002) (0.003) (0.003) (0.003) (0.003) (0.003) R2 0.196 0.286 0.434 0.540 0.604 0.579 Notes: This table reports regression results of the changes in the federal funds futures (FF1, FF2, ..., FF6) contracts one to six months ahead around the scheduled Federal Open Market Committee (FOMC) announcements regressed on sentiment index (SI) surprises and monetary policy (MP) surprises. The changes in the federal funds futures contracts are computed as differences in quotes 10minutesbeforetheFOMCstatementreleaseto20minutesafterwards. SIsurprisesarecomputed asdifferencesintheaverageSIlevelofthestorieswithin-2-days-beforetheFOMCstatementrelease to within-10-hours-afterwards. The sample is from the May 1999 FOMC meeting to the November 2022 FOMC meeting, a total of 182 meetings. he Post-Global-Financial-Crisis (PostGFC) dummy is defined as one from the December 2008 FOMC meeting to the November 2022 FOMC meeting and zero otherwise. The zero-lower-bound (ZLB) dummy is defined as one from the December 2008 FOMC meeting to the October 2015 FOMC meeting and from the April 2020 FOMC meeting to the January 2022 FOMC meeting and zero otherwise. The ordinary least squares standard errors and t-statistics are provided in parentheses and the significance levels are as follows: ∗ corresponds to p−values <0.05, ∗∗ —p-values <0.01, and ∗∗∗ —p-values <0.001. Source: Bauer and Swanson (2023) (for MP surprises), authors’ calculations (for changes in federal funds futures contracts and SI surprises). 29

Table 7: Eurodollar Futures Contracts ED1 ED2 ED3 ED4 Panel A: No Dummies SI surprise 0.006 0.010∗∗∗ 0.013∗∗∗ 0.013∗∗∗ (0.003) (0.002) (0.002) (0.003) MP surprise 0.725∗∗∗ 0.901∗∗∗ 1.031∗∗∗ 1.091∗∗∗ (0.076) (0.050) (0.059) (0.079) Constant −0.007∗∗∗ −0.010∗∗∗ −0.012∗∗∗ −0.013∗∗∗ (0.002) (0.002) (0.002) (0.002) R2 0.667 0.829 0.820 0.756 Panel B: PostGFC Dummy SI surprise 0.006∗ 0.010∗∗∗ 0.012∗∗∗ 0.013∗∗∗ (0.003) (0.002) (0.002) (0.003) MP surprise 0.725∗∗∗ 0.902∗∗∗ 1.032∗∗∗ 1.092∗∗∗ (0.076) (0.050) (0.059) (0.079) PostGFC -0.001 0.001 0.003 0.004 (0.004) (0.003) (0.004) (0.005) Constant -0.006 −0.011∗∗∗ −0.014∗∗∗ −0.015∗∗∗ (0.003) (0.002) (0.003) (0.004) R2 0.667 0.830 0.821 0.757 Panel C: ZLB Dummy SI surprise 0.006 0.010∗∗∗ 0.013∗∗∗ 0.013∗∗∗ (0.003) (0.002) (0.002) (0.003) MP surprise 0.725∗∗∗ 0.904∗∗∗ 1.035∗∗∗ 1.096∗∗∗ (0.076) (0.050) (0.059) (0.078) ZLB 0.001 0.004 0.007 0.009∗ (0.004) (0.003) (0.004) (0.005) Constant −0.007∗∗ −0.011∗∗∗ −0.014∗∗∗ −0.016∗∗∗ (0.002) (0.002) (0.002) (0.003) R2 0.667 0.831 0.824 0.761 Notes: ThistablereportsregressionresultsofthechangesintheEurodollar(ED1,ED2, ED3, ED4) futures contracts (one to four quarters ahead) around the scheduled Federal Open Market Committee (FOMC) announcements regressed on sentiment index (SI) surprises and monetary policy (MP) surprises. The changes in the Eurodollar futures contracts are computed as differences in quotes 10 minutes before the FOMC statement releaseto20minutesafterwards. SIsurprisesarecomputedasdifferencesintheaverage SI level of the stories within-2-days-before the FOMC statement release to within-10hours-afterwards. The sample is from the May 1999 FOMC meeting to the November 2022 FOMC meeting, a total of 182 meetings. The Post-Global-Financial-Crisis (Post- GFC)dummyisdefinedasonefromtheDecember2008FOMCmeetingtotheNovember 2022FOMCmeetingandzerootherwise. Thezero-lower-bound(ZLB)dummyisdefined as one from the December 2008 FOMC meeting to the October 2015 FOMC meeting and from the April 2020 FOMC meeting to the January 2022 FOMC meeting and zero otherwise. The ordinary least squares standard errors and t-statistics are provided in parentheses and the significance levels are as follows: ∗ corresponds to p−values <0.05, ∗∗ —p-values <0.01, and ∗∗∗ —p-values <0.001. Source: BauerandSwanson(2023)(forMPsurprises),authors’calculations(forchanges in Eurodollar futures contracts and SI surprises). 30

Table 8: Treasury Futures Contracts TNOTE02 TNOTE05 TNOTE10 TBOND Panel A: No Dummies SI surprise 0.010∗∗ 0.009∗ 0.007∗ 0.004 (0.003) (0.003) (0.003) (0.003) MP surprise 0.798∗∗∗ 0.724∗∗∗ 0.477∗∗∗ 0.268∗∗∗ (0.063) (0.079) (0.063) (0.061) Constant −0.009∗∗∗ −0.007∗ -0.004 -0.001 (0.002) (0.003) (0.003) (0.003) R2 0.689 0.502 0.336 0.129 Panel B: PostGFC Dummy SI surprise 0.010∗∗ 0.009∗∗ 0.007∗ 0.004 (0.003) (0.003) (0.003) (0.003) MP surprise 0.798∗∗∗ 0.723∗∗∗ 0.475∗∗∗ 0.268∗∗∗ (0.063) (0.078) (0.062) (0.060) PostGFC -0.000 -0.005 -0.006 -0.003 (0.004) (0.005) (0.005) (0.005) Constant −0.009∗∗ -0.004 -0.001 0.001 (0.003) (0.004) (0.003) (0.003) R2 0.689 0.504 0.341 0.130 Panel C: ZLB Dummy SI surprise 0.011∗∗ 0.009∗ 0.007∗ 0.004 (0.003) (0.004) (0.003) (0.003) MP surprise 0.802∗∗∗ 0.727∗∗∗ 0.478∗∗∗ 0.269∗∗∗ (0.062) (0.078) (0.063) (0.060) ZLB 0.007 0.005 0.001 0.001 (0.004) (0.006) (0.006) (0.006) Constant −0.011∗∗∗ −0.009∗∗ −0.005∗ -0.002 (0.003) (0.003) (0.002) (0.002) R2 0.693 0.504 0.336 0.129 Notes: This table reports regression results of the changes in the 2-year to 30-year nominal Treasury notes and bond futures around the scheduled Federal Open Market Committee (FOMC) announcements regressed on the sentiment index (SI) surprise and monetary policy (MP) shocks. The changes in the Treasury futures are computed as differences in quotes 10 minutes before the FOMC statement release to 20 minutes afterwards. SI surprises are computed as differences in the average SI level of the stories within-2-days-before the FOMC statement release to within-10-hours-afterwards. The sample is from the May 1999 FOMC meeting to the November 2022 FOMC meeting, a total of 182 meetings. The Post-Global-Financial-Crisis (PostGFC) dummy is defined as one from the December 2008 FOMC meeting to the November 2022 FOMC meeting and zero otherwise. The zero-lower-bound (ZLB) dummy is defined as one from the December2008 FOMCmeeting totheOctober2015 FOMCmeetingand fromtheApril 2020 FOMC meeting to the January 2022 FOMC meeting and zero otherwise. The ordinary least squares standard errors and t-statistics are provided in parentheses and the significancelevelsareasfollows: ∗ correspondstop−values<0.05,∗∗ —p-values<0.01, and ∗∗∗ —p-values <0.001. Source: BauerandSwanson(2023)(forMPsurprises),authors’calculations(forchanges in Treasury futures contracts and SI surprises). 31

Table 9: TIPS and Inflation Compensation TIPS5 TIPS10 IC5 IC10 IC5Y5F Panel A: No Dummies SI surprise 0.008∗ 0.010∗∗ 0.013 -0.055 -0.126 (0.004) (0.004) (0.012) (0.056) (0.129) MP surprise 0.736∗∗∗ 0.540∗∗∗ 0.599 -1.648 -4.930 (0.101) (0.068) (0.476) (1.711) (5.027) Constant −0.009∗ -0.005 0.009 -0.041 -0.092 (0.004) (0.003) (0.009) (0.042) (0.093) Observations 143 158 143 158 143 R2 0.370 0.326 0.042 0.023 0.028 Panel B: PostGFC Dummy SI surprise 0.008∗ 0.010∗∗ 0.013 -0.052 -0.126 (0.004) (0.004) (0.012) (0.053) (0.129) MP surprise 0.736∗∗∗ 0.540∗∗∗ 0.597 -1.657 -4.920 (0.102) (0.067) (0.475) (1.720) (5.022) PostGFC -0.004 -0.005 0.019 -0.066 -0.167 (0.007) (0.006) (0.017) (0.068) (0.172) Constant -0.006 -0.002 -0.005 0.002 0.031 (0.005) (0.004) (0.007) (0.017) (0.062) Observations 143 158 143 158 143 R2 0.371 0.328 0.046 0.025 0.031 Panel C: ZLB Dummy SI surprise 0.009∗ 0.010∗∗ 0.012 -0.052 -0.111 (0.004) (0.004) (0.011) (0.053) (0.115) MP surprise 0.743∗∗∗ 0.542∗∗∗ 0.590 -1.607 -4.819 (0.099) (0.067) (0.467) (1.674) (4.927) ZLB 0.010 0.003 -0.012 0.067 0.160 (0.007) (0.007) (0.016) (0.066) (0.162) Constant −0.014∗∗ −0.006∗ 0.014 -0.069 -0.166 (0.004) (0.003) (0.016) (0.069) (0.168) Observations 143 158 143 158 143 R2 0.379 0.327 0.044 0.025 0.031 Notes: This table reports regression results of the changes in five- and ten-year yields on Treasury Inflation-Protected Securities (TIPS) and inflation compensation measures around the scheduled Federal Open Market Committee (FOMC) announcements regressed on sentiment index (SI) surprises and monetary policy (MP) surprises. The changes in TIPS yields inflation compensation are computed as differences in quotes 10 minutes before the FOMC statement release to 20 minutes afterwards. SI surprises are computed as differences in the average SI level of the stories within-2-days-before the FOMC statement release to within-10-hours-afterwards. The sample is from the May 1999 FOMC meeting to the November 2022 FOMC meeting, a total of 182 meetings. The Post-Global-Financial-Crisis (PostGFC) dummy is defined as one from the December 2008 FOMC meeting to the November 2022 FOMC meeting and zero otherwise. The zero-lower-bound (ZLB) dummy is defined as one from the December 2008 FOMC meetingtotheOctober2015FOMCmeetingandfromtheApril2020FOMCmeetingto the January 2022 FOMC meeting and zero otherwise. The ordinary least squares standard errors and t-statistics are provided in parentheses and the significance levels are as follows: ∗ corresponds to p−values < 0.05, ∗∗ —p-values < 0.01, and ∗∗∗ —p-values <0.001. Source: BauerandSwanson(2023)(forMPsurprises),authors’calculations(forchanges in the TIPS yields, inflation compensation measures, and SI surprises). 32

Table 10: Equities and Currencies Indexes S&P500 NASDAQ DXY EURO YEN Panel A: No Dummies SI surprise 0.038 0.013 0.081∗ −0.055∗ 0.044 (0.037) (0.038) (0.037) (0.025) (0.028) MP surprise −6.121∗∗∗ −5.112∗∗∗ 0.944 −4.652∗∗∗ 2.887∗∗∗ (0.893) (0.975) (0.767) (0.568) (0.508) Constant 0.013 0.030 −0.068∗ 0.044 -0.023 (0.035) (0.036) (0.030) (0.025) (0.025) Observations 182 161 181 182 182 R2 0.303 0.230 0.052 0.336 0.167 Panel B: PostGFC Dummy SI surprise 0.024 -0.003 0.081∗ −0.056∗ 0.043 (0.035) (0.036) (0.036) (0.026) (0.028) MP surprise −6.050∗∗∗ −5.090∗∗∗ 0.944 −4.649∗∗∗ 2.890∗∗∗ (0.889) (0.949) (0.763) (0.566) (0.506) PostGFC 0.249∗∗∗ 0.262∗∗∗ -0.004 0.010 0.011 (0.066) (0.070) (0.062) (0.051) (0.051) Constant −0.131∗ −0.141∗ -0.065 0.038 -0.030 (0.054) (0.059) (0.050) (0.034) (0.036) Observations 182 161 181 182 182 R2 0.354 0.290 0.052 0.336 0.167 Panel C: ZLB Dummy SI surprise 0.045 0.022 0.081∗ −0.056∗ 0.046 (0.037) (0.039) (0.038) (0.024) (0.028) MP surprise −5.999∗∗∗ −4.993∗∗∗ 0.944 −4.673∗∗∗ 2.919∗∗∗ (0.870) (0.939) (0.761) (0.569) (0.504) ZLB 0.209∗∗ 0.211∗∗ -0.001 -0.036 0.055 (0.071) (0.074) (0.065) (0.057) (0.055) Constant -0.063 -0.056 -0.067 0.057∗ -0.043 (0.042) (0.045) (0.038) (0.025) (0.027) Observations 182 161 181 182 182 R2 0.337 0.271 0.052 0.338 0.172 Notes: Thistablereportsregressionresultsofthechangesintheequityindexesandcurrenciesaround the scheduled Federal Open Market Committee (FOMC) announcements regressed on sentiment index (SI) surprises and monetary policy (MP) surprises. The changes in the equity indexes (S&P 500 and NASDAQ) and currencies indexes (DXY, Euro, and Yen) are computed as differences in quotes 10 minutes before the FOMC statement release to 20 minutes afterwards. SI surprises are computed as differences in the average SI level of the stories within-2-days-before the FOMC statement release to within-10-hours-afterwards. The changes in the S&P 500 index refer to the changes in the E-mini S&P 500 futures contract. The sample is from the May 1999 FOMC meeting to the November 2022 FOMC meeting, a total of 182 meetings. The Post-Global-Financial-Crisis (PostGFC)dummyisdefinedasonefromtheDecember2008FOMCmeetingtotheNovember2022 FOMC meeting and zero otherwise. The zero-lower-bound (ZLB) dummy is defined as one from the December 2008 FOMC meeting to the October 2015 FOMC meeting and from the April 2020 FOMCmeetingtotheJanuary2022FOMCmeetingandzerootherwise. Theordinaryleastsquares standarderrorsandt-statisticsareprovidedinparenthesesandthesignificancelevelsareasfollows: ∗ corresponds to p−values <0.05, ∗∗ —p-values <0.01, and ∗∗∗ —p-values <0.001. Source: Bauer and Swanson (2023) (for MP surprises and changes in the E-mini S&P 500 futures contract), authors’ calculations (for changes currencies indexes and SI surprises). 33

A Appendix This section provides historic details about the day and timing of the releases of the FOMC statements. The day of the FOMC statement release varied in the early part of our sample that we account for when we subsequently construct the sentiment surprises. In the early part of our sample FOMC meetings were mostly held on Tuesdays with periodic two-day meetings that concluded predominantly on Wednesdays but also occasionally on Thursdays. Starting from the April 2012 FOMCmeeting, theFOMChasonlyheldscheduledtwo-daymeetingsandthesetypicallyconclude on Wednesdays, although several have also concluded on Thursdays.25 The timing of the FOMC statement release also varied in early part of our sample. Prior to the March 2013 FOMC meeting, the statement was typically released at or around 2:15 p.m. Since the March 2013 FOMC meeting, the release time has been consistent at 2.00 p.m. In addition to the statement release, since October 2007 the FOMC has also released the Summary of Economic Projections (SEP) at the every other meeting (March, June, September, and December meetings). Finally, since the April 2011 FOMC meeting, the FOMC Chair has held press conferences after the meetings accompanied by the SEP. Since the March 2018 FOMC meeting, the FOMC Chair has held press conferences after the conclusion of each FOMC meeting. Currently, press conferences currently start at 2:30 p.m., 30 minutes after the statement is released and typically run for about an hour.26 We account for the variation in the date and time of the FOMC statement releases and we capture news stories in response to the suite of FOMC communications around each meeting. 25Among the total of 188 meetings in our sample, 57 concluded on a Tuesday, 125 on a Wednesday, and 6 on a Thursday. 26Until the December 2012 FOMC meeting, statements that were accompanied by the SEP have been released at 12:30 p.m., while statements of the ”non-SEP” meetings have been released at 2 p.m. 34

B Appendix B.1 Sentiment surprises regressions for policy sensitive rates Turning to subsample analysis, our most intriguing result is that most of the explanatory power in Panel A appears to be driven by the pre-ZLB period (Panel B), and that in the following subperiods the sentiment surprise is almost never significant (Panels C, D, E of Table 6) never significant with the exception of post-COVID period (Panel F). This result is robust to several alternative specifications of the index.27 We broadly attribute this robust result to the fact that monetary policy communications became increasingly more transparent with time: Since the April 2011 FOMC meeting, then-Chair of the Federal Reserve Ben S. Bernanke started holding press conferences that discussed and clarified policy actions, after every other FOMC meeting. Since the March 2018 FOMC meeting current Chair of the Federal Reserve Jerome H. Powell holds press conferences after each FOMC meeting. Table B1 reports baseline regression results from (4) for the federal funds futures contracts one to six months ahead, FF1, FF2, ..., FF6. The left-hand side variable of the regression (4) corresponds to changes in the federal funds futures contracts 20 minutes before to 10 minutes after the release of the FOMC statement. First, as panel A shows, the response of the fed funds futures contracts reflects the surprise in sentiment; the sentiment surprise coefficient β is positive 1 and statistically significant for fed funds futures contracts four to six months out (FF4 to FF6), despite a strongly significant coefficient for monetary policy surprise at all six horizons. In other words, positive (or, hawkish) sentiment surprises are associated with positive movements in future policy-sensitive rates and negative (or, dovish) sentiment surprises are associated with negative movements in future policy-sensitive rates. Second, the response of the fed funds futures contracts to a change in media sentiment following the FOMC announcement is statistically significant at the five-percent level for FF5 and FF6 and at the ten-percent level for FF4 contracts. The MP surprise drives out significance of the sentiment surprise for the very near term contracts, namely, FF1 through FF3, but not for contracts at longer horizons. Interestingly, the explanatory power of the sentiment and monetary policy surprises increases with horizon and notably higher at longer (FF4 through FF6) than ta the shorter (FF1 through FF3) horizons. Our results suggest that MP shocks do not fully explain the change in implied rates in fed funds futures in the tight window around the FOMC announcements and that other factors are at play as well. As such, MP surprises appear to capture the surprise in the policy rate itself, while the sentiment surprise has some informational content for the “path” surprise that reflects changes in rates further out in the horizon. B.2 Sentiment surprises regressions for other asset classes Table B3 reports the results of regressing Treasury bills and on-the-run nominal Treasury yields of maturities ranging between 3 months and 30 years on the SI surprises. Sentiment has some informational content for the on-the-run securities in the pre-ZLB period (Panel B). Table B4 reports regression results of the one-the-run 5- and 10-year TIPS yields and inflation compensation of corresponding maturities. The last column of this table also reports regression results of the 27A reports regression results for a tighter time frame of collection of stories, namely, the event window closes at the COB of the FOMC statement release day. 35

five-year, five-year-forward inflation compensation. Sentiment appears to explain movements in 10-year TIPS rates in the full sample (Panel A), and in 10-year TIPS rates and 10-year inflation compensation in the post-ZLB period (Panel D). This arguably makes sense as discussions related to inflation in policy communications are apparently dominant during this period, given the focus on inflation in the current macroeconomic environment. Table B5 reports regression results of the effect that SI surprises have on changes in the S&P 500 and NASDAQ equity indexes as well as on the changes in the currency indexes DXY, EURO, and YEN. The surprises in SI explain changes in DXY (with positive sign) and in EURO (with negativesign)inthefullsampleperiod(seePanelA).TheSIsurprisealsoappearstobeexplaining movements in the S&P500 prices in the pre-ZLB period (see Panel B). This means that a hawkish surprise is associated with the increases in equity prices. In contrast, a positive MP1 shock is associated with the declines in equity prices. This finding is interesting because it contrasts with thefindingsinTablesB1andB3, wheresentimentsurpriseandmonetarypolicyshockaffectpolicy rates and nominal Treasury yields in the same direction. Our findings regarding the MP regressor are directionally consistent with Bernanke and Kuttner (2005) who show that the unanticipated tightening (or easing) in monetary policy is associated with a decline (or increase) in broad stock indexes. 36

stcartnoC serutuF sdnuF laredeF :1B elbaT 6FF 5FF 4FF 3FF 2FF 1FF 6FF 5FF 4FF 3FF 2FF 1FF 8002 rebmeceD ot 9991 yaM :doireP BLZ-erP :B lenaP 2202 rebmevoN ot 9991 yaM :elpmaS lluF :A lenaP ∗∗∗710.0 ∗∗∗310.0 ∗∗∗310.0 ∗∗∗110.0 700.0 500.0 ∗∗010.0 ∗∗800.0 ∗700.0 500.0 300.0 300.0 esirprus IS )300.0( )300.0( )300.0( )300.0( )400.0( )300.0( )300.0( )300.0( )300.0( )300.0( )300.0( )200.0( ∗∗∗177.0 ∗∗∗707.0 ∗∗∗336.0 ∗∗∗575.0 ∗∗∗225.0 ∗∗∗142.0 ∗∗∗176.0 ∗∗∗606.0 ∗∗∗125.0 ∗∗∗944.0 ∗∗∗983.0 ∗∗∗771.0 esirprus PM )370.0( )470.0( )280.0( )780.0( )111.0( )760.0( )460.0( )160.0( )760.0( )470.0( )980.0( )050.0( ∗∗110.0− ∗900.0− ∗800.0− ∗700.0− 500.0− 600.0− ∗∗∗010.0− ∗∗∗700.0− ∗∗∗700.0− ∗∗∗700.0− 400.0− ∗400.0− tnatsnoC )400.0( )300.0( )300.0( )400.0( )400.0( )300.0( )200.0( )200.0( )200.0( )200.0( )200.0( )200.0( 87 87 87 87 87 87 281 281 281 281 281 281 snoitavresbO 607.0 417.0 966.0 095.0 024.0 862.0 275.0 006.0 335.0 824.0 682.0 091.0 2R 0202 yraunaJ ot 5102 rebmeceD :doirep BLZ-tsoP :D lenaP 5102 rebmevoN ot 9002 yraunaJ :doirep BLZ :C lenaP 100.0− 200.0− 300.0− 300.0− 600.0− 000.0 200.0 200.0 200.0 100.0 000.0− 100.0− esirprus IS )600.0( )400.0( )400.0( )400.0( )300.0( )100.0( )300.0( )200.0( )200.0( )100.0( )100.0( )100.0( ∗∗∗627.0 ∗∗∗506.0 ∗∗∗005.0 ∗∗223.0 081.0 850.0 ∗∗∗243.0 ∗∗∗942.0 ∗∗∗702.0 ∗∗∗381.0 ∗701.0 930.0 esirprus PM )111.0( )540.0( )050.0( )790.0( )301.0( )940.0( )550.0( )940.0( )840.0( )440.0( )250.0( )220.0( 800.0− 400.0− 400.0− 500.0− 200.0 100.0− ∗∗400.0− ∗300.0− 100.0− 200.0− 100.0− 100.0− tnatsnoC )500.0( )300.0( )300.0( )300.0( )400.0( )200.0( )100.0( )100.0( )100.0( )100.0( )100.0( )100.0( 43 43 43 43 43 43 55 55 55 55 55 55 snoitavresbO 826.0 367.0 176.0 093.0 622.0 170.0 415.0 424.0 343.0 033.0 511.0 370.0 2R 2202 rebmevoN ot 2202 yraunaJ :doirep DIVOC-tsoP :F lenaP 1202 rebmeceD ot 0202 yraurbeF :doirep DIVOC :E lenaP 490.0 440.0 ∗240.0 ∗030.0 020.0 400.0 ∗710.0 ∗610.0 ∗510.0 700.0 010.0 300.0 esirprus IS )730.0( )030.0( )410.0( )700.0( )210.0( )900.0( )600.0( )500.0( )400.0( )600.0( )010.0( )600.0( 483.0 904.0 712.0 280.0 810.0 900.0 ∗334.0 ∗∗783.0 ∗∗132.0 320.0 110.0 720.0− esirprus PM )493.0( )603.0( )831.0( )460.0( )821.0( )660.0( )801.0( )280.0( )640.0( )590.0( )911.0( )250.0( 131.0− 960.0− ∗170.0− ∗∗∗560.0− 140.0− 900.0− 100.0− 200.0− 200.0− 100.0− 000.0 100.0 tnatsnoC )050.0( )040.0( )910.0( )400.0( )610.0( )110.0( )200.0( )200.0( )200.0( )200.0( )300.0( )200.0( 7 7 7 7 7 7 8 8 8 8 8 8 snoitavresbO 474.0 824.0 416.0 026.0 081.0 230.0 427.0 277.0 547.0 383.0 923.0 012.0 2R eht dnuora daeha shtnom xis ot eno stcartnoc )6FF ,... ,2FF ,1FF( serutuf sdnuf laredef eht ni segnahc eht fo stluser noisserger stroper elbat sihT :setoN ehT .sesirprus )PM( ycilop yratenom dna sesirprus )IS( xedni tnemitnes eht no desserger stnemecnuonna )CMOF( eettimmoC tekraM nepO laredeF deludehcs .sdrawretfa setunim 02 ot esaeler tnemetats CMOF eht erofeb setunim 01 setouq ni secnereffid sa detupmoc era stcartnoc serutuf sdnuf laredef eht ni segnahc .sdrawretfa-sruoh-01-nihtiw ot esaeler tnemetats CMOF eht erofeb-syad-2-nihtiw seirots eht fo level IS egareva eht ni secnereffid sa detupmoc era sesirprus IS dradnats serauqs tsael yranidro ehT .sgniteem 281 fo latot a ,gniteem CMOF 2202 rebmevoN eht ot gniteem CMOF 9991 yaM eht morf si doirep elpmas ehT ∗∗∗ dna ,10.0 < seulav-p— ∗∗ ,50.0 < seulav−p ot sdnopserroc ∗ :swollof sa era slevel ecnacfiingis eht dna sesehtnerap ni dedivorp era scitsitats-t dna srorre .100.0< seulav-p— .)sesirprus IS dna stcartnoc serutuf sdnuf laredef ni segnahc rof( snoitaluclac ’srohtua ,)sesirprus PM rof( )3202( nosnawS dna reuaB :ecruoS 37

stcartnoC serutuF rallodoruE :2B elbaT 4DE 3DE 2DE 1DE 4DE 3DE 2DE 1DE 8002 rebmeceD ot 9991 yaM :doireP BLZ-erP :B lenaP 2202 rebmevoN ot 9991 yaM :elpmaS lluF :A lenaP ∗∗∗910.0 ∗∗∗810.0 ∗∗∗310.0 800.0 ∗∗∗310.0 ∗∗∗310.0 ∗∗∗010.0 600.0 esirprus IS )400.0( )300.0( )300.0( )400.0( )300.0( )200.0( )200.0( )300.0( ∗∗∗980.1 ∗∗∗140.1 ∗∗∗249.0 ∗∗∗797.0 ∗∗∗190.1 ∗∗∗130.1 ∗∗∗109.0 ∗∗∗527.0 esirprus PM )601.0( )970.0( )660.0( )501.0( )970.0( )950.0( )050.0( )670.0( ∗∗∗510.0− ∗∗∗410.0− ∗∗∗210.0− ∗800.0− ∗∗∗310.0− ∗∗∗210.0− ∗∗∗010.0− ∗∗∗700.0− tnatsnoC )400.0( )300.0( )300.0( )400.0( )200.0( )200.0( )200.0( )200.0( 87 87 87 87 281 281 281 281 snoitavresbO 387.0 358.0 868.0 186.0 657.0 028.0 928.0 766.0 2R 0202 yraunaJ ot 5102 rebmeceD :doirep BLZ-tsoP :D lenaP 5102 rebmevoN ot 9002 yraunaJ :doirep BLZ :C lenaP 400.0 300.0 100.0 000.0 700.0− 400.0− 100.0 400.0 esirprus IS )300.0( )200.0( )200.0( )100.0( )500.0( )500.0( )400.0( )300.0( ∗∗∗721.1 ∗∗∗610.1 ∗∗∗538.0 ∗∗∗206.0 ∗∗∗861.1 ∗∗∗079.0 ∗∗∗207.0 ∗∗∗393.0 esirprus PM )401.0( )680.0( )550.0( )430.0( )170.0( )970.0( )290.0( )270.0( ∗∗∗710.0− ∗∗∗610.0− ∗∗∗210.0− ∗∗∗800.0− ∗∗900.0− ∗∗∗800.0− ∗∗600.0− 300.0− tnatsnoC )300.0( )300.0( )200.0( )100.0( )300.0( )200.0( )200.0( )200.0( 43 43 43 43 55 55 55 55 snoitavresbO 508.0 848.0 298.0 298.0 937.0 457.0 666.0 094.0 2R 2202 rebmevoN ot 2202 yraunaJ :doirep DIVOC-tsoP :F lenaP 1202 rebmeceD ot 0202 yraurbeF :doirep DIVOC :E lenaP ∗∗∗910.0 ∗∗∗810.0 ∗∗∗310.0 800.0 ∗∗∗310.0 ∗∗∗310.0 ∗∗∗010.0 600.0 esirprus IS )400.0( )300.0( )300.0( )400.0( )300.0( )200.0( )200.0( )300.0( ∗∗∗980.1 ∗∗∗140.1 ∗∗∗249.0 ∗∗∗797.0 ∗∗∗190.1 ∗∗∗130.1 ∗∗∗109.0 ∗∗∗527.0 esirprus PM )601.0( )970.0( )660.0( )501.0( )970.0( )950.0( )050.0( )670.0( ∗∗∗510.0− ∗∗∗410.0− ∗∗∗210.0− ∗800.0− ∗∗∗310.0− ∗∗∗210.0− ∗∗∗010.0− ∗∗∗700.0− tnatsnoC )400.0( )300.0( )300.0( )400.0( )200.0( )200.0( )200.0( )200.0( 87 87 87 87 281 281 281 281 snoitavresbO 387.0 358.0 868.0 186.0 657.0 028.0 928.0 766.0 2R sretrauq ruof ot eno stcartnoc serutuf )4DE dna ,3DE ,2DE ,1DE( rallodoruE eht ni segnahc eht fo stluser noisserger stroper elbat sihT :setoN yratenom dna sesirprus )IS( xedni tnemitnes no desserger stnemecnuonna )CMOF( eettimmoC tekraM nepO laredeF deludehcs eht dnuora daeha tnemetats CMOF eht erofeb setunim 01 setouqni secnereffid sa detupmoc era stcartnoc serutuf rallodoruE eht ni segnahc ehT .sesirprus )PM( ycilop CMOF eht erofeb-syad-2-nihtiw seirots eht fo level IS egareva eht ni secnereffid sa detupmoc era sesirprus IS .sdrawretfa setunim 02 ot esaeler latot a ,gniteem CMOF 2202 rebmevoN eht ot gniteem CMOF 9991 yaM eht morf si elpmas ehT .sdrawretfa-sruoh-01-nihtiw ot esaeler tnemetats ∗ :swollof sa era slevel ecnacfiingis eht dna sesehtnerap ni dedivorp era scitsitats-t dna srorre dradnats serauqs tsael yranidro ehT .sgniteem 281 fo .100.0< seulav-p— ∗∗∗ dna ,10.0< seulav-p— ∗∗ ,50.0< seulav−p ot sdnopserroc .)sesirprus IS dna stcartnoc serutuf rallodoruE eht ni segnahc eht rof( snoitaluclac ’srohtua ,)sesirprus PM rof( )3202( nosnawS dna reuaB :ecruoS 38

stcartnoC serutuF yrusaerT :3B elbaT DNOBT 01ETONT 50ETONT 20ETONT DNOBT 01ETONT 50ETONT 20ETONT 8002 rebmeceD ot 9991 yaM :doireP BLZ-erP :B lenaP 2202 rebmevoN ot 9991 yaM :elpmaS lluF :A lenaP 300.0 ∗800.0 ∗∗310.0 ∗∗∗610.0 400.0 ∗700.0 ∗900.0 ∗∗010.0 esirprus IS )300.0( )400.0( )500.0( )500.0( )300.0( )300.0( )300.0( )300.0( ∗∗∗672.0 ∗∗∗534.0 ∗∗∗466.0 ∗∗∗267.0 ∗∗∗862.0 ∗∗∗774.0 ∗∗∗427.0 ∗∗∗897.0 esirprus PM )470.0( )180.0( )501.0( )380.0( )160.0( )360.0( )970.0( )360.0( 000.0 100.0− 400.0− ∗800.0− 100.0− 400.0− ∗700.0− ∗∗∗900.0− tnatsnoC )300.0( )300.0( )400.0( )300.0( )300.0( )300.0( )300.0( )200.0( 87 87 87 87 281 281 281 281 snoitavresbO 782.0 984.0 695.0 927.0 921.0 633.0 205.0 986.0 2R 0202 yraunaJ ot 5102 rebmeceD :doirep BLZ-tsoP :D lenaP 5102 rebmevoN ot 9002 yraunaJ :doirep BLZ :C lenaP 200.0 200.0 200.0 200.0 710.0 510.0 500.0 300.0− esirprus IS )200.0( )200.0( )300.0( )200.0( )120.0( )810.0( )210.0( )600.0( ∗822.0 ∗∗∗506.0 ∗∗∗839.0 ∗∗∗179.0 544.0 ∗∗∗038.0 ∗∗∗141.1 ∗∗∗729.0 esirprus PM )290.0( )301.0( )821.0( )990.0( )342.0( )881.0( )631.0( )860.0( ∗500.0− ∗∗∗110.0− ∗∗∗610.0− ∗∗∗610.0− 000.0− 400.0− 500.0− ∗700.0− tnatsnoC )200.0( )300.0( )300.0( )300.0( )600.0( )500.0( )600.0( )300.0( 43 43 43 43 55 55 55 55 snoitavresbO 402.0 375.0 276.0 457.0 880.0 342.0 614.0 316.0 2R 2202 rebmevoN ot 2202 yraunaJ :doirep DIVOC-tsoP :F lenaP 1202 rebmeceD ot 0202 yraurbeF :doirep DIVOC :E lenaP 520.0 710.0 220.0 020.0 900.0 420.0 430.0 220.0 esirprus IS )810.0( )610.0( )910.0( )720.0( )120.0( )320.0( )430.0( )410.0( 850.0 ∗173.0 ∗∗486.0 ∗∗849.0 110.0 981.0 882.0 204.0 esirprus PM )590.0( )901.0( )621.0( )531.0( )412.0( )641.0( )632.0( )061.0( 520.0− 320.0− 330.0− 620.0− 800.0 600.0 700.0 300.0 tnatsnoC )220.0( )910.0( )120.0( )230.0( )900.0( )700.0( )010.0( )500.0( 7 7 7 7 8 8 8 8 snoitavresbO 994.0 728.0 319.0 029.0 760.0 961.0 031.0 292.0 2R deludehcs eht dnuora serutuf dnob dna eton yrusaerT lanimon raey-03 ot raey-2 eht ni segnahc eht fo stluser noisserger stroper elbat sihT :setoN ehT .sesirprus )PM( ycilop yratenom dna sesirprus )IS( xedni tnemitnes no desserger stnemecnuonna )CMOF( eettimmoC tekraM nepO laredeF .sdrawretfa setunim 02 ot esaeler tnemetats CMOF eht erofeb setunim 01 setouq ni secnereffid sa detupmoc era serutuf yrusaerT eht ni segnahc -sruoh-01-nihtiw ot esaeler tnemetats CMOF eht erofeb-syad-2-nihtiw seirots eht fo level IS egareva eht ni secnereffid sa detupmoc era sesirprus IS tsael yranidro ehT .sgniteem 281 fo latot a ,gniteem CMOF 2202 rebmevoN eht ot gniteem CMOF 9991 yaM eht morf si elpmas lluf ehT .sdrawretfa ∗∗ ,50.0 < seulav−p ot sdnopserroc ∗ :swollof sa era slevel ecnacfiingis eht dna sesehtnerap ni dedivorp era scitsitats-t dna srorre dradnats serauqs .100.0< seulav-p— ∗∗∗ dna ,10.0< seulav-p— .)sesirprus IS dna stcartnoc serutuf yrusaerT ni segnahc rof( snoitaluclac ’srohtua ,)sesirprus PM rof( )3202( nosnawS dna reuaB :ecruoS 39

noitasnepmoC noitaflnI dna SPIT :4B elbaT F5Y5CI 01CI 5CI 01SPIT 5SPIT F5Y5CI 01CI 5CI 01SPIT 5SPIT 8002 rebmeceD ot 9991 yaM :doireP BLZ-erP :B lenaP 2202 rebmevoN ot 9991 yaM :elpmaS lluF :A lenaP ∗∗600.0 000.0 200.0− 800.0 010.0 621.0− 550.0− 310.0 ∗∗010.0 ∗800.0 esirprus IS )200.0( )200.0( )300.0( )400.0( )500.0( )921.0( )650.0( )210.0( )400.0( )400.0( ∗∗∗861.0 420.0 930.0 ∗∗∗774.0 ∗∗∗355.0 039.4− 846.1− 995.0 ∗∗∗045.0 ∗∗∗637.0 esirprus PM )540.0( )240.0( )280.0( )980.0( )201.0( )720.5( )117.1( )674.0( )860.0( )101.0( 100.0 000.0 000.0− 300.0− 700.0− 290.0− 140.0− 900.0 500.0− ∗900.0− tnatsnoC )300.0( )200.0( )300.0( )400.0( )500.0( )390.0( )240.0( )900.0( )300.0( )400.0( 93 45 93 45 93 341 851 341 851 341 snoitavresbO 293.0 310.0 940.0 355.0 516.0 820.0 320.0 240.0 623.0 073.0 2R 0202 yraunaJ ot 5102 rebmeceD :doirep BLZ-tsoP :D lenaP 5102 rebmevoN ot 9002 yraunaJ :doirep BLZ :C lenaP ∗300.0 ∗300.0 200.0 ∗600.0 400.0 800.0− 300.0− 300.0 220.0 210.0 esirprus IS )100.0( )100.0( )100.0( )200.0( )300.0( )210.0( )800.0( )500.0( )120.0( )210.0( 221.0 ∗∗761.0 ∗∗∗112.0 ∗∗∗786.0 ∗∗∗931.1 020.0− 540.0 011.0 ∗∗∗048.0 ∗∗∗432.1 esirprus PM )470.0( )850.0( )750.0( )111.0( )121.0( )441.0( )701.0( )151.0( )922.0( )922.0( 200.0− 300.0− ∗300.0− ∗∗∗310.0− ∗∗∗910.0− 300.0 200.0 200.0 100.0− 400.0− tnatsnoC )200.0( )200.0( )200.0( )300.0( )400.0( )400.0( )300.0( )300.0( )600.0( )600.0( 43 43 43 43 43 55 55 55 55 55 snoitavresbO 023.0 444.0 164.0 695.0 517.0 910.0 800.0 630.0 512.0 814.0 2R 2202 rebmevoN ot 2202 yraunaJ :doirep DIVOC-tsoP :F lenaP 1202 rebmeceD ot 0202 yraurbeF :doirep DIVOC :E lenaP 967.0 743.0 570.0− 220.0 520.0 120.0 620.0 230.0 840.0 170.0 esirprus IS )829.3( )687.1( )553.0( )920.0( )660.0( )620.0( )030.0( )530.0( )340.0( )160.0( 103.23− 304.41− 594.3 ∗855.0 137.0 312.0 802.0 302.0 333.0 635.0 esirprus PM )749.14( )330.91( )188.3( )421.0( )745.0( )762.0( )182.0( )892.0( )522.0( )663.0( 727.2− 232.1− 362.0 910.0− 150.0− 700.0− 600.0− 600.0− 100.0 200.0 tnatsnoC )853.5( )434.2( )094.0( )630.0( )280.0( )110.0( )210.0( )310.0( )210.0( )910.0( 7 7 7 7 7 8 8 8 8 8 snoitavresbO 521.0 121.0 461.0 038.0 892.0 021.0 841.0 861.0 502.0 791.0 2R noitasnepmocnoitaflnidna)SPIT(seitiruceSdetcetorP-noitaflnIyrusaerTraey-01dna-5nisegnahcehtfostlusernoissergerstroperelbatsihT :setoN yratenom dna sesirprus )IS( xedni tnemitnes no desserger stnemecnuonna )CMOF( eettimmoC tekraM nepO laredeF deludehcs eht dnuora serusaem esaeler tnemetats CMOF eht erofeb setunim 01 setouq ni secnereffid sa detupmoc era noitasnepmoc noitaflni ni segnahc ehT .sesirprus )PM( ycilop tnemetats CMOF eht erofeb-syad-2-nihtiw seirots eht fo level IS egareva eht ni secnereffid sa detupmoc era sesirprus IS .sdrawretfa setunim 02 ot fo latot a ,gniteem CMOF 2202 rebmevoN eht ot gniteem CMOF 9991 yaM eht morf si doirep elpmas ehT .sdrawretfa-sruoh-01-nihtiw ot esaeler ∗ :swollof sa era slevel ecnacfiingis eht dna sesehtnerap ni dedivorp era scitsitats-t dna srorre dradnats serauqs tsael yranidro ehT .sgniteem 281 .100.0< seulav-p— ∗∗∗ dna ,10.0< seulav-p— ∗∗ ,50.0< seulav−p ot sdnopserroc .)sesirprusISdna,serusaemnoitasnepmocnoitaflni,SPITnisegnahcrof(snoitaluclac’srohtua,)sesirprusPMrof()3202(nosnawSdnareuaB :ecruoS 40

sexednI seicnerruC dna seitiuqE :5B elbaT NEY ORUE YXD QADSAN 005P&S NEY ORUE YXD QADSAN 005P&S 8002 rebmeceD ot 9991 yaM :doireP BLZ-erP :B lenaP 2202 rebmevoN ot 9991 yaM :elpmaS lluF :A lenaP 750.0 ∗750.0− 090.0 860.0 ∗101.0 440.0 ∗550.0− ∗180.0 310.0 830.0 esirprus IS )330.0( )820.0( )150.0( )850.0( )940.0( )820.0( )520.0( )730.0( )830.0( )730.0( ∗∗546.1 ∗∗∗575.3− 923.0− ∗∗∗368.4− ∗∗∗271.6− ∗∗∗788.2 ∗∗∗256.4− 449.0 ∗∗∗211.5− ∗∗∗121.6− esirprus PM )075.0( )366.0( )179.0( )962.1( )971.1( )805.0( )865.0( )767.0( )579.0( )398.0( 920.0− 240.0 760.0− 011.0− 501.0− 320.0− 440.0 ∗860.0− 030.0 310.0 tnatsnoC )530.0( )530.0( )840.0( )750.0( )450.0( )520.0( )520.0( )030.0( )630.0( )530.0( 87 87 77 75 87 281 281 181 161 281 snoitavresbO 731.0 493.0 250.0 333.0 804.0 761.0 633.0 250.0 032.0 303.0 2R 0202 yraunaJ ot 5102 rebmeceD :doirep BLZ-tsoP :D lenaP 5102 rebmevoN ot 9002 yraunaJ :doirep BLZ :C lenaP 100.0− 100.0 000.0− 700.0− 030.0− 580.0 702.0− 471.0 151.0− 002.0− esirprus IS )620.0( )820.0( )720.0( )530.0( )820.0( )941.0( )131.0( )890.0( )001.0( )811.0( ∗∗∗102.7 ∗∗∗210.8− ∗∗∗086.7 ∗866.4− ∗∗891.4− ∗∗∗186.7 ∗∗∗577.01− 494.2 ∗∗∗483.6− ∗∗∗677.7− esirprus PM )560.1( )059.0( )598.0( )170.2( )405.1( )374.1( )619.1( )591.2( )256.1( )319.1( 660.0− ∗670.0 ∗570.0− ∗401.0 ∗∗511.0 710.0 100.0− 550.0− ∗511.0 901.0 tnatsnoC )530.0( )730.0( )430.0( )740.0( )040.0( )640.0( )940.0( )850.0( )250.0( )550.0( 43 43 43 43 43 55 55 55 55 55 snoitavresbO 865.0 026.0 346.0 522.0 572.0 862.0 654.0 770.0 142.0 972.0 2R 2202 rebmevoN ot 2202 yraunaJ :doirep DIVOC-tsoP :F lenaP 1202 rebmeceD ot 0202 yraurbeF :doirep DIVOC :E lenaP 360.0− ∗171.0 341.0− 815.0 554.0 452.0 244.0− 592.0 502.0 730.0 esirprus IS )911.0( )950.0( )460.0( )763.0( )322.0( )212.0( )882.0( )042.0( )175.0( )183.0( ∗∗940.4 ∗∗860.4− ∗∗040.4 ∗109.8− ∗080.7− 492.3 ∗767.4− ∗466.3 ∗004.31 259.5 esirprus PM )357.0( )375.0( )335.0( )206.2( )201.2( )383.1( )014.1( )763.1( )999.4( )499.3( 430.0− 590.0− 560.0 976.0− 626.0− 020.0 350.0 850.0− 411.0 480.0 tnatsnoC )541.0( )160.0( )560.0( )044.0( )092.0( )360.0( )380.0( )370.0( )771.0( )431.0( 7 7 7 7 7 8 8 8 8 8 snoitavresbO 598.0 719.0 429.0 117.0 537.0 081.0 462.0 571.0 223.0 191.0 2R tekraM nepO laredeF deludehcs eht dnuora sexedni seicnerruc dna sexedni ytiuqe eht ni segnahc eht fo stluser noisserger stroper elbat sihT :setoN ytiuqe eht ni segnahc ehT .sesirprus )PM( ycilop yratenom dna sesirprus )IS( xedni tnemitnes no desserger stnemecnuonna )CMOF( eettimmoC CMOF eht erofeb setunim 01 setouq ni secnereffid sa detupmoc era )neY dna ,oruE ,YXD( sexedni seicnerruc dna )QADSAN dna 005 P&S( sexedni eht erofeb-syad-2-nihtiw seirots eht fo level IS egareva eht ni secnereffid sa detupmoc era sesirprus IS .sdrawretfa setunim 02 ot esaeler tnemetats .tcartnoc serutuf 005 P&S inim-E eht ni segnahc ot refer xedni 005 P&S eht ni segnahC .sdrawretfa-sruoh-01-nihtiw ot esaeler tnemetats CMOF serauqs tsael yranidro ehT .sgniteem 281 fo latot a ,gniteem CMOF 2202 rebmevoN eht ot gniteem CMOF 9991 yaM eht morf si doirep elpmas ehT seulav-p— ∗∗ ,50.0< seulav−p ot sdnopserroc ∗ :swollof sa era slevel ecnacfiingis eht dna sesehtnerap ni dedivorp era scitsitats-t dna srorre dradnats .100.0< seulav-p— ∗∗∗ dna ,10.0< eht ni segnahc rof( snoitaluclac ’srohtua ,)tcartnoc serutuf 005 P&S inim-E eht ni segnahc dna sesirprus PM rof( )3202( nosnawS dna reuaB :ecruoS .)sesirprus IS dna ,sexedni seicnerruc ,xedni QADSAN 41

C Appendix In this appendix, we present a brief overview of five alternative methodologies developed by researchers in various central banks and report results of regression (4) based on SI construction following each of the five methodologies. C.1 Alternative methodologies Lucca-Trebbi index: developed by Lucca and Trebbi (2009). This methodology also starts with the Factiva database search and look for the articles in the three-day window around FOMC announcements (i.e., the day before, the day of, and the day after the FOMC) with words “Federal Reserve, Fed, FOMC”. The authors filter out stories on the day of the FOMC with no time stamps. Subsequently, in those articles, the authors use only relevant sentences, thatis, thesentencesthatincludewords“Rates, Policies, policies, statement, announcement, Fed, FOMC, Federal Reserve”. In those sentences, they look for hawkish words (hawkish, tighten, hike, raise, increase, boost) and dovish words (dovish, ease, cut, lower, decrease, loose). They subsequently compute a Factiva Semantic Orientation (FSO) score, similar to a computation in eqution (1). Direct negations in the methodology are handled by switching hawkish into dovish words and vice verse if these words are preceded by “not”. While not made explicit in the LT paper/appendices, we assume that LT include “n’t” contractions as well (e.g. didn’t). In addition, the authors remove all past tense verbs so they can only capture the future policy actions. Cannon index: Cannon (2015) looks at positive and negative words in the FOMC meeting transcripts. Atthetextprocessingleveltheauthoreliminatesstopwords, convertletterstolower case, remove numbers, and remove punctuation. The negation is accounted for by checking polarity of word directly preceding positive or negative (in our case, hawkish or dovish) words using two sentiment dictionaries, either Loughran and McDonald (2011) or Hu-Liu dictionary for product reviews. Finally, she defines the tone of a particular communication as: the ratio of the difference between positive and negative words to the sum of positive and negative words. Carvalho, Hsu, and Nechio index: Carvalho et al. (2016) use the Lucca and Trebbi (2009) approach but count in the articles only the number of the word “hawkish” or the word “dovish” and construct the words in the FSO similar to Lucca and Trebbi (2009) and our sentiment index (1). Apel and Grimaldi index: ApelandGrimaldi(2012)applytheirmethodologytotheRiskbank communications released in Swedish. The authors look at combinations of words to determine if a statement is dovish or hawkish. They use a prespecified list of seven nouns (i.e., inflation, cyclical position, growth, price, wages, oil price, development) and a set of either dovish adjectives (decreasing, slower, weaker, lower) or hawkish adjectives (increasing, faster, stronger, and higher) determining if the adjective-noun combination is hawkish or dovish. The adjectives used include all forms (stemmer) of a particular adjective (e.g., fast, faster, and fastest). Then they define the four indexes using the count of hawk and dove adjective-noun combinations: 42

• Net Hawkish Index = [(hawk/(hawk+dove)) - (dove/(hawk + dove))] + 1; • Net Dovish Index = [(dove/(hawk + dove)) - (hawk/(hawk+dove))] + 1; • Hawkish Index = (hawk/(hawk+dove)); • Dovish Index = (dove/(hawk + dove)). Nyman, Kapadia, Tuckett, Gregory, Ormerod, and Smith index: TheNymanetal.(2018) methodology counts the number of excitement and anxiety words in the financial market text-based data. The authors use Loughran and McDonald (2011) dictionary to determine the presence of negation. If a negative word occurs within three words of a hawk/dove word then the hawk/dove word is counted as the opposite. The set of negative words used is {no, not, none, neither, never, nobody}. The sentiment is defined as the difference between the number of hawkish and dovish words in a particular text scaled by the size of the text. C.2 Regression results Tables C1 through C5 show regression results that for the full sample period for five alternative methodologies described above for the SI construction.28 Table C1 presents results for the Lucca and Trebbi (2009) index for the full sample period. As the table shows, their index surprise is almost never significant except a few occurrences: the FF3 and FF4 contracts, 10-year TIPS yield and the EURO index (Panel C). Table C2 presents the results for the Carvalho et al. (2016) index. Thisindexhasmoresignificantloadings, forexample, itissignificantinexplainingchangesinTIPS yields, equity indexes, and currency indexes. Interestingly, this index appears to become more dovish as other indexes indexes become more hawkish during the economic upturn and at the start of the late 2015 Fed tightening cycle, according to Figure C1. Turning to Cannon (2015) index, surprises in this index appear to have little explanatory power for changes in asset prices except for a couple of fed funds futures contracts, FF4 and FF5. Apel and Grimaldi (2012) methodology yields a few mildly significant results, according to table C4, but overall, changes in asset prices are relatively mute to SI surprises constructed using this methodology.29 Finally, table C5 reports results based on the Nyman et al. (2018) methodology and has only two significant loadings: the FF1 contract and the DXY currency index. Note that the Nyman et al. (2018) methodology is based on extracting semantic content from financial-market text-based data. It is not necessarily directly related to BoE communications. We nevertheless apply the index to the corpus of our Factiva stories. Overall, none of these indexes appears to have results comparable and consistent across several asset classes and, most importantly, consistent across horizons. One explanation could be is that these methodologies are not designed to capture economic environment marked by unconventional monetary policies, when different sets of dictionaries are needed,as we argue in the paper. 28Results for the pre- and post-GFC periods for these alternative indexes are not presented in the paper but available upon request. 29For Riksbank index, we report results based on one out of their four variations of the index, namely, the Net Hawkish Index. Results for the other three variations are available upon request. 43

Figure C1: Sentiment Index: Alternative Methodologies Lucca-Trebbi #10-3 Bank of England 1 2 0.5 1 0 0 -0.5 -1 -1 -2 -1.5 -3 Jan-00 Jan-10 Jan-00 Jan-10 Kansas City Fed Riks Bank 0.5 0.3 0.2 0.1 0 0 -0.1 -0.5 -0.2 -0.3 Jan-00 Jan-10 Jan-00 Jan-10 San Francisco Fed 0.1 0 -0.1 -0.2 Jan-00 Jan-10 This figure shows the sentiment index (SI) of the Federal Open Market Committee (FOMC) monetary policy communications constructed following five alternative methodologiesofLuccaandTrebbi(2009)(LSI),Cannon(2015)(KansasCityFed), Nymanetal.(2018)(BankofEngland),ApelandGrimaldi(2012)(Riksbank),and Carvalhoetal.(2016)(SanFranciscoFed)methodologies. TheSIsarecomputedfor the period of 36 hours before the FOMC statement release to 36 hours afterwards. Thesampleperiodconsistsof146scheduledFOMCmeetings,betweentheMay1999 FOMC meeting to the June 2017 FOMC meeting. The frequency is FOMC. The blue-shaded bars indicate the National Bureau of Economic Research recessions. Source: Dow Jones, Factiva; authors’ calculations. 44

ygolodohtem )9002( ibberT dna accuL :1C elbaT serutuF dnuF laredeF :A lenaP 6FF 5FF 4FF 3FF 2FF 1FF 410.0 410.0 ∗710.0 ∗110.0 200.0 700.0 esirprus IS )52.1( )75.1( )24.2( )90.2( )16.0( )18.1( ∗∗∗177.0 ∗∗∗817.0 ∗∗∗527.0 ∗∗∗167.0 ∗∗∗998.0 ∗∗∗384.0 -rus PM esirp )59.8( )63.7( )91.8( )72.11( )15.02( )04.7( 500.0- 300.0- 300.0- 200.0- 100.0 100.0tnatsnoC )65.1-( )23.1-( )43.1-( )01.1-( )20.1( )17.1-( 641 641 641 641 641 641 snoitavresbO 9884.0 4935.0 0236.0 0847.0 2449.0 7597.0 2R sdleiY yrusaerT lanimoN nur-eht-nO :B lenaP Y03NURNO Y01NURNO Y5NURNO Y2NURNO M6NURNO M3NURNO 810.0 220.0 920.0 520.0 110.0 800.0 esirprus IS )88.1( )79.1( )09.1( )86.1( )04.1( )13.1( 740.0- 361.0 ∗∗∗033.0 ∗∗∗284.0 ∗∗∗635.0 ∗∗∗694.0 -rus PM esirp )37.0-( )88.1( )57.3( )35.5( )34.7( )57.5( 200.0- 400.0- 400.0- 600.0- 400.0- 300.0tnatsnoC )94.0-( )08.0-( )98.0-( )04.1-( )58.1-( )47.1-( 641 641 641 641 641 641 snoitavresbO 1110.0 6820.0 3380.0 3571.0 6194.0 7375.0 2R seicnerruC dna ,seitiuqE ,noitasnepmoC noitaflnI :C lenaP NEY ORUE YXD QADSAN 005PS F5Y5CI Y01CI Y5CI Y01SPIT Y5SPIT 251.0 ∗672.0- 141.0 090.0 811.0 000.0- 500.0- 800.0- ∗520.0 020.0 esirprus IS )76.1( )20.2-( )22.1( )94.0( )76.0( )20.0-( )01.1-( )50.1-( )99.1( )60.1( ∗156.1 ∗∗∗874.2- 067.0 819.3- ∗∗984.4- ∗∗∗842.0 ∗∗431.0 821.0 ∗∗323.0 ∗∗∗225.0 -rus PM esirp )14.2( )17.3-( )78.0( )58.1-( )66.2-( )25.3( )63.3( )13.1( )33.3( )15.5( 310.0- 720.0 ∗370.0- 220.0- 540.0- 400.0 200.0 200.0 300.0- 400.0tnatsnoC )83.0-( )87.0( )89.1-( )35.0-( )10.1-( )84.1( )14.1( )31.1( )36.0-( )37.0-( 641 641 641 521 641 701 221 701 221 701 snoitavresbO 2640.0 5590.0 4310.0 4390.0 2011.0 5590.0 6780.0 2460.0 5870.0 0121.0 2R -ecnuonnagniteem)CMOF(eettimmoCtekraMnepOlaredeFdeludehcsehtdnuorasegnahcecirptessafostlusernoissergerstroperelbatsihT :setoN seigolodohtem gniwollof detcurtsnoc era sesirprus PM dna IS .sesirprus PM dna sesirprus )IS( xedni tnemitnes eht ni segnahc eht no desserger stnem detupmoc era snoisserger eht ni segnahc secirp tessa dna sesirprus PM .ylevitcepser ,)5002( .la te kanyakruG dna )TL ,9002( ibberT dna accuL ni eht ni ecnereffid eht sa detupmoc era sesirprus IS .sdrawretfa setunim 02 ot esaeler tnemetats CMOF eht erofeb setunim 01 setouq ni secnereffid sa dradnatsserauqstsaelyranidroehT .sdrawretfa-sruoh-63-nihtiwotesaelertnemetatsCMOFehterofeb-sruoh-63-nihtiwseirotsehtfolevelISegareva ,10.0 < seulav-p— ∗∗ ,50.0 < seulav−p ot sdnopserroc ∗ :swollof sa era slevel ecnacfiingis eht dna sesehtnerap ni dedivorp era scitsitats-t dna srorre a ,gniteem CMOF 7102 enuJ eht ot gniteem CMOF 9991 yaM eht fo doirep elpmas eht rof detroper era stluser ehT .100.0 < seulav-p— ∗∗∗ dna fo rebmun eht nehw seires noitasnepmoc noitaflni dna )SPIT( seitiruceS detcetorP-noitaflnI yrusaerT fo noitpecxe eht htiw ,snoitavresbo 641 fo latot .doirep elpmas eht ni ylrae seitiruces SPIT eht fo ytilibaliavanu ot eud rewef si snoitavresbo .snoitaluclac ’srohtuA :ecruoS 45

ygolodohtem )6102( .la te ohlavraC :2C elbaT serutuF sdnuF laredeF :A lenaP 6FF 5FF 4FF 3FF 2FF 1FF ∗460.0 ∗450.0 530.0 ∗030.0 200.0 600.0 esirprus IS )51.2( )82.2( )27.1( )20.2( )03.0( )46.0( ∗∗∗577.0 ∗∗∗427.0 ∗∗∗537.0 ∗∗∗767.0 ∗∗∗009.0 ∗∗∗884.0 -rus PM esirp )97.8( )81.7( )60.8( )41.11( )62.02( )62.7( 400.0- 300.0- 200.0- 100.0- 100.0 100.0tnatsnoC )74.1-( )71.1-( )60.1-( )09.0-( )12.1( )34.1-( 641 641 641 641 641 641 snoitavresbO 4394.0 8045.0 1426.0 5547.0 1449.0 3097.0 2R sdleiY yrusaerT lanimoN nur-eht-nO :B lenaP Y03NURNO Y01NURNO Y5NURNO Y2NURNO M6NURNO M3NURNO 600.0 650.0 490.0 390.0 ∗740.0 220.0 esirprus IS )71.0( )94.1( )86.1( )59.1( )71.2( )84.1( 530.0- 471.0 ∗∗∗243.0 ∗∗∗294.0 ∗∗∗935.0 ∗∗∗005.0 -rus PM esirp )35.0-( )19.1( )45.3( )13.5( )12.7( )76.5( 100.0- 300.0- 300.0- 500.0- 300.0- 200.0tnatsnoC )82.0-( )36.0-( )76.0-( )22.1-( )18.1-( )46.1-( 641 641 641 641 641 641 snoitavresbO 2100.0 7320.0 3180.0 9771.0 9594.0 6175.0 2R seicnerruC dna ,seitiuqE ,noitasnepmoC noitaflnI :C lenaP NEY ORUE YXD QADSAN 005PS F5Y5CI Y01CI Y5CI Y01SPIT Y5SPIT 284.0 ∗∗081.1- 017.0 ∗670.1- ∗141.1- 210.0 320.0 130.0 ∗880.0 ∗141.0 esirprus IS )54.1( )59.2-( )66.1( )62.2-( )63.2-( )26.0( )04.1( )74.1( )62.2( )15.2( ∗917.1 ∗∗∗875.2- 308.0 358.3- ∗413.4- ∗∗∗742.0 ∗∗231.0 021.0 ∗∗333.0 ∗∗∗125.0 -rus PM esirp )34.2( )95.3-( )29.0( )97.1-( )55.2-( )25.3( )61.3( )51.1( )02.3( )69.4( 700.0- 810.0 960.0- 610.0- 630.0- 400.0 200.0 100.0 200.0- 300.0tnatsnoC )22.0-( )25.0( )19.1-( )04.0-( )68.0-( )35.1( )03.1( )39.0( )74.0-( )66.0-( 641 641 641 521 641 701 221 701 221 701 snoitavresbO 4440.0 9801.0 5020.0 5911.0 9821.0 8690.0 2390.0 3170.0 1180.0 9851.0 2R -ecnuonnagniteem)CMOF(eettimmoCtekraMnepOlaredeFdeludehcsehtdnuorasegnahcecirptessafostlusernoissergerstroperelbatsihT :setoN niseigolodohtemgniwollofdetcurtsnocerasesirprusPMdnaIS .sesirprusPMdnasesirprus)IS(xednitnemitnesehtnisegnahcehtnodessergerstnem secnereffid sa detupmoc era snoisserger eht ni segnahc secirp tessa dna sesirprus PM .ylevitcepser ,)5002( .la te kanyakruG dna )6102( .la te ohlavraC IS egareva eht ni ecnereffid eht sa detupmoc era sesirprus IS .sdrawretfa setunim 02 ot esaeler tnemetats CMOF eht erofeb setunim 01 setouq ni srorre dradnats serauqs tsael yranidro ehT .sdrawretfa-sruoh-63-nihtiw ot esaeler tnemetats CMOF eht erofeb-sruoh-63-nihtiw seirots eht fo level ∗∗∗ dna ,10.0< seulav-p— ∗∗ ,50.0< seulav−p ot sdnopserroc ∗ :swollof sa era slevel ecnacfiingis eht dna sesehtnerap ni dedivorp era scitsitats-t dna 641 fo latot a ,gniteem CMOF 7102 enuJ eht ot gniteem CMOF 9991 yaM eht fo doirep elpmas eht rof detroper era stluser ehT .100.0< seulav-p— snoitavresboforebmunehtnehwseiresnoitasnepmocnoitaflnidna)SPIT(seitiruceSdetcetorP-noitaflnIyrusaerTfonoitpecxeehthtiw ,snoitavresbo .doirep elpmas eht ni ylrae seitiruces SPIT eht fo ytilibaliavanu ot eud rewef si .snoitaluclac ’srohtuA :ecruoS 46

ygolodohtem )5102( nonnaC :3C elbaT serutuF sdnuF laredeF :A lenaP 6FF 5FF 4FF 3FF 2FF 1FF 720.0 ∗720.0 ∗420.0 710.0 300.0 ∗800.0 esirprus IS )38.1( )12.2( )32.2( )79.1( )77.0( )72.2( ∗∗∗577.0 ∗∗∗327.0 ∗∗∗337.0 ∗∗∗667.0 ∗∗∗009.0 ∗∗∗784.0 -rus PM esirp )83.9( )65.7( )44.8( )95.11( )15.02( )83.7( 500.0- 300.0- 300.0- 200.0- 100.0 100.0tnatsnoC )26.1-( )04.1-( )13.1-( )11.1-( )00.1( )46.1-( 641 641 641 641 641 641 snoitavresbO 6694.0 1845.0 4436.0 7057.0 4449.0 9497.0 2R sdleiY yrusaerT lanimoN nur-eht-nO :B lenaP Y03NURNO Y01NURNO Y5NURNO Y2NURNO M6NURNO M3NURNO 910.0 810.0 310.0 510.0 900.0 700.0 esirprus IS )33.1( )30.1( )95.0( )07.0( )58.0( )48.0( 930.0- ∗571.0 ∗∗∗743.0 ∗∗∗694.0 ∗∗∗145.0 ∗∗∗005.0 -rus PM esirp )95.0-( )99.1( )57.3( )36.5( )35.7( )37.5( 200.0- 300.0- 300.0- 500.0- 400.0- 200.0tnatsnoC )24.0-( )96.0-( )86.0-( )52.1-( )97.1-( )66.1-( 641 641 641 641 641 641 snoitavresbO 5800.0 5220.0 8860.0 1461.0 7784.0 6075.0 2R seicnerruC dna ,seitiuqE ,noitasnepmoC noitaflnI :C lenaP NEY ORUE YXD QADSAN 005PS F5Y5CI Y01CI Y5CI Y01SPIT Y5SPIT 980.0 081.0- 591.0 730.0 451.0 100.0 600.0- 900.0- 410.0 600.0esirprus IS )96.0( )31.1-( )61.1( )91.0( )67.0( )01.0( )80.1-( )30.1-( )77.0( )72.0-( ∗047.1 ∗∗∗736.2- 128.0 088.3- ∗∗734.4- ∗∗∗842.0 ∗∗331.0 721.0 ∗∗233.0 ∗∗∗735.0 -rus PM esirp )15.2( )88.3-( )69.0( )28.1-( )26.2-( )15.3( )82.3( )82.1( )13.3( )21.5( 800.0- 910.0 270.0- 020.0- 540.0- 400.0 200.0 200.0 200.0- 200.0tnatsnoC )42.0-( )55.0( )19.1-( )74.0-( )00.1-( )83.1( )14.1( )51.1( )15.0-( )64.0-( 641 641 641 521 641 701 221 701 221 701 snoitavresbO 4830.0 8570.0 5410.0 7190.0 3011.0 6590.0 0880.0 9260.0 9660.0 8211.0 2R stnemecnuonna )CMOF( eettimmoC tekraM nepO laredeF deludehcs eht dnuora segnahc ecirp tessa fo stluser noisserger stroper elbat sihT :setoN gniwollof detupmoc era sesirprus PM dna IS .sesirprus )PM( ycilop yratenom dna sesirprus )IS( xedni tnemitnes ni segnahc eht no desserger detupmoc era snoisserger eht ni segnahc secirp tessa dna sesirprus PM .ylevitcepser ,)5002( .la te kanyakruG dna )5102( nonnaC ni seigolodohtem eht ni ecnereffid eht sa detupmoc era sesirprus IS .sdrawretfa setunim 02 ot esaeler tnemetats CMOF eht erofeb setunim 01 setouq ni secnereffid sa dradnatsserauqstsaelyranidroehT .sdrawretfa-sruoh-63-nihtiwotesaelertnemetatsCMOFehterofeb-sruoh-63-nihtiwseirotsehtfolevelISegareva ,10.0 < seulav-p— ∗∗ ,50.0 < seulav−p ot sdnopserroc ∗ :swollof sa era slevel ecnacfiingis eht dna sesehtnerap ni dedivorp era scitsitats-t dna srorre a ,gniteem CMOF 7102 enuJ eht ot gniteem CMOF 9991 yaM eht fo doirep elpmas eht rof detroper era stluser ehT .100.0 < seulav-p— ∗∗∗ dna fo rebmun eht nehw seires noitasnepmoc noitaflni dna )SPIT( seitiruceS detcetorP-noitaflnI yrusaerT fo noitpecxe eht htiw ,snoitavresbo 641 fo latot .doirep elpmas eht ni ylrae seitiruces SPIT eht fo ytilibaliavanu ot eud rewef si snoitavresbo .snoitaluclac ’srohtuA :ecruoS 47

ygolodohtem )2102( idlamirG dna lepA :4C elbaT serutuF sdnuF laredeF :A lenaP 6FF 5FF 4FF 3FF 2FF 1FF ∗230.0 700.0 010.0 900.0 100.0- ∗510.0 esirprus IS )60.2( )15.0( )77.0( )39.0( )51.0-( )11.2( ∗∗∗477.0 ∗∗∗727.0 ∗∗∗637.0 ∗∗∗867.0 ∗∗∗109.0 ∗∗∗584.0 -rus PM esirp )86.8( )02.7( )99.7( )59.01( )24.02( )68.7( 400.0- 300.0- 200.0- 100.0- 100.0 100.0tnatsnoC )83.1-( )90.1-( )00.1-( )38.0-( )12.1( )53.1-( 641 641 641 641 641 641 snoitavresbO 1694.0 8235.0 6126.0 7347.0 1449.0 0108.0 2R sdleiY yrusaerT lanimoN nur-eht-nO :B lenaP Y03NURNO Y01NURNO Y5NURNO Y2NURNO M6NURNO M3NURNO ∗430.0 120.0 910.0 120.0 ∗720.0 ∗120.0 esirprus IS )52.2( )11.1( )28.0( )19.0( )74.2( )20.2( 140.0- 571.0 ∗∗∗643.0 ∗∗∗594.0 ∗∗∗835.0 ∗∗∗894.0 -rus PM esirp )56.0-( )29.1( )25.3( )71.5( )95.7( )40.6( 100.0- 300.0- 300.0- 400.0- 300.0- 200.0tnatsnoC )42.0-( )75.0-( )85.0-( )21.1-( )07.1-( )25.1-( 641 641 641 641 641 641 snoitavresbO 2610.0 1220.0 7960.0 2561.0 1305.0 2385.0 2R seicnerruC dna ,seitiuqE ,noitasnepmoC noitaflnI ,SPIT :C lenaP NEY ORUE YXD QADSAN 005PS F5Y5CI Y01CI Y5CI Y01SPIT Y5SPIT 901.0- 101.0- 762.0- 243.0 792.0 300.0- ∗910.0- 020.0- 800.0 230.0esirprus IS )67.0-( )35.0-( )43.1-( )03.1( )31.1( )72.0-( )14.2-( )86.1-( )44.0( )71.1-( ∗977.1 ∗∗∗456.2- 119.0 ∗669.3- ∗∗264.4- ∗∗∗742.0 ∗∗∗831.0 021.0 ∗∗233.0 ∗∗∗925.0 -rus PM esirp )45.2( )66.3-( )50.1( )20.2-( )28.2-( )45.3( )77.3( )51.1( )41.3( )43.5( 600.0- 410.0 760.0- 910.0- 930.0- 400.0 200.0 200.0 200.0- 300.0tnatsnoC )91.0-( )04.0( )68.1-( )84.0-( )19.0-( )55.1( )63.1( )20.1( )14.0-( )35.0-( 641 641 641 521 641 701 221 701 221 701 snoitavresbO 4830.0 4070.0 6610.0 9401.0 8411.0 8590.0 2511.0 9970.0 3460.0 7021.0 2R -ecnuonnagniteem)CMOF(eettimmoCtekraMnepOlaredeFdeludehcsehtdnuorasegnahcecirptessafostlusernoissergerstroperelbatsihT :setoN seigolodohtem gniwollof detcurtsnoc era sesirprus PM dna IS .sesirprus PM dna sesirprus )IS( xedni tnemitnes eht ni segnahc eht no desserger stnem sa detupmoc era snoisserger eht ni segnahc secirp tessa dna sesirprus PM .ylevitcepser ,)5002( .la te kanyakruG dna )2102( idlamirG dna lepA ni eht ni ecnereffid eht sa detupmoc era sesirprus IS .sdrawretfa setunim 02 ot esaeler tnemetats CMOF eht erofeb setunim 01 setouq ni secnereffid dradnatsserauqstsaelyranidroehT .sdrawretfa-sruoh-63-nihtiwotesaelertnemetatsCMOFehterofeb-sruoh-63-nihtiwseirotsehtfolevelISegareva ,10.0 < seulav-p— ∗∗ ,50.0 < seulav−p ot sdnopserroc ∗ :swollof sa era slevel ecnacfiingis eht dna sesehtnerap ni dedivorp era scitsitats-t dna srorre a ,gniteem CMOF 7102 enuJ eht ot gniteem CMOF 9991 yaM eht fo doirep elpmas eht rof detroper era stluser ehT .100.0 < seulav-p— ∗∗∗ dna fo rebmun eht nehw seires noitasnepmoc noitaflni dna )SPIT( seitiruceS detcetorP-noitaflnI yrusaerT fo noitpecxe eht htiw ,snoitavresbo 641 fo latot .doirep elpmas eht ni ylrae seitiruces SPIT eht fo ytilibaliavanu ot eud rewef si snoitavresbo .snoitaluclac ’srohtuA :ecruoS 48

ygolodohtem )8102( .la te namyN :5C elbaT serutuF sdnuF laredeF :A lenaP 6FF 5FF 4FF 3FF 2FF 1FF 376.7 928.6 215.5 941.4 062.0 ∗976.3 esirprus IS )91.1( )43.1( )12.1( )20.1( )31.0( )11.2( ∗∗∗967.0 ∗∗∗717.0 ∗∗∗927.0 ∗∗∗367.0 ∗∗∗009.0 ∗∗∗384.0 -rus PM esirp )33.9( )15.7( )14.8( )07.11( )08.02( )54.7( 500.0- 300.0- 300.0- 200.0- 100.0 100.0tnatsnoC )47.1-( )74.1-( )83.1-( )22.1-( )51.1( )37.1-( 641 641 641 641 641 641 snoitavresbO 9194.0 3045.0 0626.0 0647.0 1449.0 8797.0 2R sdleiY yrusaerT lanimoN nuR-eht-nO :B lenaP Y03NURNO Y01NURNO Y5NURNO Y2NURNO M6NURNO M3NURNO 158.3 302.7 034.9 547.9 694.3 971.3 esirprus IS )89.0( )73.1( )14.1( )25.1( )90.1( )90.1( 140.0- 761.0 ∗∗∗533.0 ∗∗∗484.0 ∗∗∗835.0 ∗∗∗794.0 -rus PM esirp )26.0-( )39.1( )77.3( )95.5( )54.7( )47.5( 100.0- 400.0- 400.0- 600.0- 400.0- 300.0tnatsnoC )83.0-( )47.0-( )28.0-( )93.1-( )88.1-( )66.1-( 641 641 641 641 641 641 snoitavresbO 4300.0 6320.0 4570.0 9171.0 4884.0 3275.0 2R seicnerruC dna ,seitiuqE ,noitasnepmoC noitaflnI :C lenaP NEY ORUE YXD QADSAN 005PS F5Y5CI Y01CI Y5CI Y01SPIT Y5SPIT 654.98 485.89- ∗774.321 277.15 404.25 243.0 661.0 854.0- 865.9 260.8 esirprus IS )77.1( )66.1-( )51.2( )95.0( )07.0( )90.0( )70.0( )21.0-( )33.1( )57.0( ∗816.1 ∗∗∗815.2- 666.0 329.3- ∗∗884.4- ∗∗∗742.0 ∗∗231.0 421.0 ∗∗∗523.0 ∗∗∗615.0 -rus PM esirp )24.2( )98.3-( )08.0( )58.1-( )66.2-( )04.3( )32.3( )52.1( )83.3( )82.5( 610.0- 620.0 ∗180.0- 420.0- 640.0- 400.0 200.0 200.0 300.0- 400.0tnatsnoC )84.0-( )47.0( )31.2-( )55.0-( )10.1-( )03.1( )32.1( )88.0( )16.0-( )96.0-( 641 641 641 521 641 701 221 701 221 701 snoitavresbO 9250.0 4580.0 4130.0 2490.0 1011.0 6590.0 9280.0 3250.0 5270.0 7711.0 2R -ecnuonnagniteem)CMOF(eettimmoCtekraMnepOlaredeFdeludehcsehtdnuorasegnahcecirptessafostlusernoissergerstroperelbatsihT :setoN seigolodohtem gniwollof detcurtsnoc era sesirprus PM dna IS .sesirprus PM dna sesirprus )IS( xedni tnemitnes eht ni segnahc eht no desserger stnem secnereffidsadetupmocerasnoissergerehtnisegnahcsecirptessadnasesirprusPM .ylevitcepser,)5002(.latekanyakruGdna)8102(.latenamyNni IS egareva eht ni ecnereffid eht sa detupmoc era sesirprus IS .sdrawretfa setunim 02 ot esaeler tnemetats CMOF eht erofeb setunim 01 setouq ni srorre dradnats serauqs tsael yranidro ehT .sdrawretfa-sruoh-63-nihtiw ot esaeler tnemetats CMOF eht erofeb-sruoh-63-nihtiw seirots eht fo level ∗∗∗ dna ,10.0< seulav-p— ∗∗ ,50.0< seulav−p ot sdnopserroc ∗ :swollof sa era slevel ecnacfiingis eht dna sesehtnerap ni dedivorp era scitsitats-t dna 641 fo latot a ,gniteem CMOF 7102 enuJ eht ot gniteem CMOF 9991 yaM eht fo doirep elpmas eht rof detroper era stluser ehT .100.0< seulav-p— snoitavresboforebmunehtnehwseiresnoitasnepmocnoitaflnidna)SPIT(seitiruceSdetcetorP-noitaflnIyrusaerTfonoitpecxeehthtiw ,snoitavresbo .doirep elpmas eht ni ylrae seitiruces SPIT eht fo ytilibaliavanu ot eud rewef si .snoitaluclac ’srohtuA :ecruoS 49