UBC Theses and Dissertations

UBC Theses Logo

UBC Theses and Dissertations

Essays in information economics Martineau, Charles 2017

Your browser doesn't seem to have a PDF viewer, please download the PDF to view this item.

Item Metadata

Download

Media
24-ubc_2017_september_martineau_charles.pdf [ 1.92MB ]
Metadata
JSON: 24-1.0348752.json
JSON-LD: 24-1.0348752-ld.json
RDF/XML (Pretty): 24-1.0348752-rdf.xml
RDF/JSON: 24-1.0348752-rdf.json
Turtle: 24-1.0348752-turtle.txt
N-Triples: 24-1.0348752-rdf-ntriples.txt
Original Record: 24-1.0348752-source.json
Full Text
24-1.0348752-fulltext.txt
Citation
24-1.0348752.ris

Full Text

Essays in Information EconomicsbyCharles MartineauB.Comm., Concordia University, 2009M.Sc., HEC Montre´al, 2011A THESIS SUBMITTED IN PARTIAL FULFILLMENT OFTHE REQUIREMENTS FOR THE DEGREE OFDOCTOR OF PHILOSOPHYinThe Faculty of Graduate and Postdoctoral Studies(Finance)THE UNIVERSITY OF BRITISH COLUMBIA(Vancouver)July 2017c© Charles Martineau 2017AbstractI present three essays on Information Economics. The first essay consists ofanalyzing high-frequency price dynamics around earnings announcementsfor the largest 1,500 U.S. stocks between 2011 and 2015. Price discoveryfollowing earnings surprises mostly occurs in the after-hours market, fol-lowing the earnings announcement, and is generally complete by 10 a.m.Eighty percent of the price response to earnings surprises in the after-hoursmarket occurs upon arrival of the first trades. Price reactions are largelyexplained by earnings surprises and not by order flow, consistent with thetheoretical view that news can incorporate prices instantly. In the secondessay, co-authored with Oliver Boguth and Vincent Gre´goire, we show thatin an effort to increase transparency, the Chair of the Federal Reserve nowholds a press conference following some, but not all, Federal Open MarketCommittee announcements. Press conferences are scheduled independentlyof economic conditions and communicate little information. Evidence fromfinancial markets demonstrates that investors lower their expectations ofimportant decisions on days without press conferences, and we show thatthey shift attention away from these announcements. Both channels preventeffective monetary policy, as the committee is averse to surprising marketsand aims to coordinate market expectations. Correspondingly, we show thatannouncements without press conferences convey less price-relevant informa-tion. In the third essay, co-authored with Adlai J. Fisher and Jinfei Sheng,we construct indices of media attention to macroeconomic risks includingemployment, growth, inflation and monetary policy. Attention rises aroundmacroeconomic announcements and following changes in fundamentals overquarterly, annual, and business cycle horizons. The effect is asymmetric,with bad news raising attention more than good news. Increases in aggre-gate trade volume and volatility coincide with rising attention, controllingfor announcements. Finally, changes in attention prior to the unemploymentannouncement predict both the announcement surprise and stock returns onthe announcement day. We conclude that media attention to macroeconomicfundamentals provides useful information beyond the dates and contents ofmacroeconomic announcements.iiLay SummaryPublic information releases from corporations and financial institutions havea significant impact on financial markets and stock prices. A long-standingissue in financial economics is to understand how fast the information getsincorporated into stock prices. This issue is often referred to the notion ofprice discovery. It is also important to understand how the information getsreleased to the public (e.g., newspaper articles, press conferences) influenceprice discovery. In addition, how recent technological development in finan-cial markets influence price discovery and how it impacts the social welfareof investors is an on-going debate. This thesis sheds light on these issuesand provides new empirical findings on price discovery following two publicinformation releases, that is, earnings and macroeconomic announcements.iiiPrefaceChapter 2 is based solely on my own work. Chapter 3 is a co-authoredproject with Assistant Professor Oliver Boguth of Arizona State Universityand Assistant Professor Vincent Gre´goire of the University of Melbourne. Iinitiated this project from previous research of mine. We contributed equallyto the writing and to the empirical analysis. Chapter 4 is a co-authoredproject with my adviser Professor Adlai J. Fisher and Ph.D. colleague JinfeiSheng. We contributed equally to the writing and to the empirical analysis.ivTable of ContentsAbstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iiLay Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iiiPreface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ivTable of Contents . . . . . . . . . . . . . . . . . . . . . . . . . . . . vList of Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . viiiList of Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ixAcknowledgments . . . . . . . . . . . . . . . . . . . . . . . . . . . . xii1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12 How is Earnings News Transmitted to Stock Prices? . . . . 22.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22.2 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82.2.1 Earnings Announcements Sample . . . . . . . . . . . . 82.2.2 NASDAQ Limit Order Book-Level Data . . . . . . . . 92.2.3 Displayed and Hidden Liquidity . . . . . . . . . . . . . 102.2.4 Summary Statistics . . . . . . . . . . . . . . . . . . . . 112.3 Price Discovery of Earnings Surprises: When is it Complete? 162.3.1 Are there Daily Post-Earnings Announcement drifts? . 162.3.2 Are there Intraday Post-Earnings Announcement Drifts? 212.3.3 The Response of After-Hours Returns to Earnings Sur-prises . . . . . . . . . . . . . . . . . . . . . . . . . . . 242.3.4 The Dynamics of Price Discovery following EarningsAnnouncements at the Opening of Markets . . . . . . 31vTABLE OF CONTENTS2.4 Price Discovery following Earnings Surprises in the After-Hours Market . . . . . . . . . . . . . . . . . . . . . . . . . . . 382.4.1 Market Activity in the After Hours around EarningsAnnouncements . . . . . . . . . . . . . . . . . . . . . . 382.4.2 The Dynamics of Price Discovery in the After-HoursMarket . . . . . . . . . . . . . . . . . . . . . . . . . . 412.4.3 How is Earnings News Transmitted to Stock Prices? . 462.5 The Impact of Earnings Surprises on Volatility, Liquidity, andTrade Volume . . . . . . . . . . . . . . . . . . . . . . . . . . . 532.6 Hidden Liquidity around Earnings Announcements . . . . . . 612.7 Conclusion to Chapter 2 . . . . . . . . . . . . . . . . . . . . . 653 Shaping Expectations and Coordinating Attention. . . . . . 673.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 673.2 The Federal Open Market Committee . . . . . . . . . . . . . 733.3 Financial Markets around FOMC Announcements . . . . . . 763.3.1 Press Conferences and Market Expectations . . . . . . 773.3.2 Resolution of Uncertainty at FOMC Announcements . 883.3.3 The Pre-FOMC Announcement Drift . . . . . . . . . . 933.4 Investor Attention to FOMC Announcements . . . . . . . . . 943.4.1 Institutional Investor Attention . . . . . . . . . . . . . 943.4.2 Retail Investor Attention . . . . . . . . . . . . . . . . 1023.4.3 Google Search Volume . . . . . . . . . . . . . . . . . . 1033.4.4 The Information Content of Press Conferences . . . . 1043.4.5 International Evidence . . . . . . . . . . . . . . . . . . 1063.5 Shaping Expectations and Coordinating Attention . . . . . . 1073.6 Conclusion to Chapter 3 . . . . . . . . . . . . . . . . . . . . . 1114 Media Attention, Macroeconomic Fundamentals. . . . . . . 1134.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1134.2 Macroeconomic Attention Indices . . . . . . . . . . . . . . . . 1184.2.1 Construction of the Attention Indices . . . . . . . . . 1224.2.2 Empirical Properties of the Attention Indices . . . . . 1234.3 Attention and Macroeconomic Fundamentals . . . . . . . . . 1324.3.1 Macroeconomic Announcements . . . . . . . . . . . . 1334.3.2 Macroeconomic Fundamentals . . . . . . . . . . . . . . 1354.4 Attention and Stock Market Activity . . . . . . . . . . . . . . 1404.5 Using Attention for Forecasting . . . . . . . . . . . . . . . . . 1424.5.1 Unemployment Announcements . . . . . . . . . . . . . 1424.5.2 FOMC Announcements . . . . . . . . . . . . . . . . . 149viTABLE OF CONTENTS4.6 Conclusion to Chapter 4 . . . . . . . . . . . . . . . . . . . . . 1515 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1535.1 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . 154Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155Appendices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .A Appendix to Chapter 2 . . . . . . . . . . . . . . . . . . . . . . 168A.1 Trading Hours on NASDAQ . . . . . . . . . . . . . . . . . . . 168A.2 Post-Earnings Announcement Drifts since 1984 . . . . . . . . 169A.3 Institutional Details about Hidden Orders on NASDAQ ITCH 171A.4 High-frequency Trading Activites in the After-Hours Market . 173A.5 Additional Results on the Impact of Earnings Surprises. . . . . 175B Appendix to Chapter 3 . . . . . . . . . . . . . . . . . . . . . . 177C Appendix to Chapter 4 . . . . . . . . . . . . . . . . . . . . . . 180C.1 Sample of news articles mentioning macroeconomic funda-mentals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 180C.1.1 Additional Figures and Results . . . . . . . . . . . . . 182viiList of Tables2.1 Descriptive Statistics . . . . . . . . . . . . . . . . . . . . . . . 122.2 Cumulative Daily Abnormal Returns following Earnings An-nouncements . . . . . . . . . . . . . . . . . . . . . . . . . . . 192.3 OLS Regression: Cumulative Abnormal Returns on EarningsSurprises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 212.4 Logit Regression: Determinants to After-Hours Trading fol-lowing Earnings News . . . . . . . . . . . . . . . . . . . . . . 262.5 OLS Regression: After-Hours Returns on Earnings Surprises . 282.6 Price Discovery following Earnings Surprises at the Openingof Markets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 352.7 OLS Regression: Stock Returns on Earnings Surprises andOrder Imbalance . . . . . . . . . . . . . . . . . . . . . . . . . 512.8 OLS Regression: Realized Spreads on Displayed and HiddenLimit Orders . . . . . . . . . . . . . . . . . . . . . . . . . . . 643.1 FOMC Announcement Calendar . . . . . . . . . . . . . . . . 753.2 FOMC Announcement Returns . . . . . . . . . . . . . . . . . 793.3 FOMC Announcement Returns: Regressions . . . . . . . . . . 823.4 Probability of Interest Rate Changes before FOMC Announce-ments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 863.5 Returns of VIX at FOMC Announcements . . . . . . . . . . . 913.6 Attention before FOMC Announcements . . . . . . . . . . . . 1003.7 Realized Volatility during Press Conferences . . . . . . . . . . 1053.8 Attention before Announcements in Canada and New Zealand 1083.9 Regressions with Time Trends . . . . . . . . . . . . . . . . . . 1104.1 Newspapers Search Words . . . . . . . . . . . . . . . . . . . . 1194.2 Macroeconomic Attention and Macroeconomic Fundamentals 1214.3 Descriptive Statistics . . . . . . . . . . . . . . . . . . . . . . . 1244.4 Persistence of Macroeconomic Attention . . . . . . . . . . . . 1304.5 Macroeconomic Attention and Macroeconomic Fundamentals 137viiiLIST OF TABLES4.6 Media Attention and Aggregate Trade Volume . . . . . . . . 1414.7 Media Attention and Implied Volatility . . . . . . . . . . . . . 1434.8 Unemployment Surprise Forecasts on Employment SituationAnnouncement Days . . . . . . . . . . . . . . . . . . . . . . . 1454.9 S&P Return Forecast on Employment Situation Announce-ment Days . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1484.10 Forecasts on FOMC Announcements . . . . . . . . . . . . . . 150A.1 High-Frequency Trading Activities during Regular and After-Market Hours . . . . . . . . . . . . . . . . . . . . . . . . . . . 174C.1 Descriptive Statistics and Correlation . . . . . . . . . . . . . 185C.2 Panel A: Descriptive Statistics for Monthly Unadjusted MAI 185C.3 Persistence of Macroeconomic Attention . . . . . . . . . . . . 187C.4 Media Attention and Macroeconomic Fundamentals . . . . . 188C.5 Media Attention and Aggregate Trade Volume . . . . . . . . 190C.6 Media Attention and Implied Volatility . . . . . . . . . . . . . 191C.7 Unemployment Surprise Forecasts . . . . . . . . . . . . . . . 192C.8 S&P Return Forecast on Employment Situation Announce-ment Days . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 196ixList of Figures2.1 Abnormal Daily Returns around Earnings Announcements . . 172.2 Cumulative Abnormal Intraday Returns around Earnings An-nouncements . . . . . . . . . . . . . . . . . . . . . . . . . . . 232.3 The Response of Stock Returns to Earnings Surprises at theOpening of Markets . . . . . . . . . . . . . . . . . . . . . . . 322.4 An Example of Price Response to Earnings Announcementsat High Frequency . . . . . . . . . . . . . . . . . . . . . . . . 392.5 Statistics on After-Hours Trading following Earnings Announce-ments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 402.6 Cumulative Returns following Earnings Announcements inthe After-Hours Market . . . . . . . . . . . . . . . . . . . . . 432.7 The Response of Stock Returns to Earnings Surprises in theAfter-Hours Market . . . . . . . . . . . . . . . . . . . . . . . 452.8 Order Imbalance following Earnings Announcements in theAfter-Hours Market . . . . . . . . . . . . . . . . . . . . . . . 482.9 Explanatory Power of Earnings Surprises and Order Imbal-ance to Stock Returns in the After-Hours Market . . . . . . . 492.10 Average Volatility, Quoted Spread, and Turnover prior toEarnings Announcements . . . . . . . . . . . . . . . . . . . . 572.11 The Response of Abnormal Volatility, Abnormal Quoted Spread,and Abnormal Turnover to Earnings Surprises around Earn-ings Announcements . . . . . . . . . . . . . . . . . . . . . . . 583.1 Cumulative E-mini Return around FOMC Announcements . 803.2 Term Structure of the Probability of Target Rate Changes . . 873.3 Term Structure of the Probability of Target Rate Changes:Regression . . . . . . . . . . . . . . . . . . . . . . . . . . . . 893.4 Cumulative VIX Return around FOMC Announcements . . . 963.5 Cumulative VIX Return around FOMC Announcements (2006-2011) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 973.6 FOMC Pre-Announcement Drift and Press Conferences . . . 98xLIST OF FIGURES3.7 Attention Level Before FOMC Announcements . . . . . . . . 994.1 Attention to Unemployment . . . . . . . . . . . . . . . . . . . 1174.2 Macro Attention and Macroeconomic Fundamentals . . . . . 1274.3 Autocorrelation in Macroeconomic Attention . . . . . . . . . 1324.4 Macroeconomic Attention around Macroeconomic Announce-ments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1344.5 Attention to Unemployment around Employment SituationAnnouncements . . . . . . . . . . . . . . . . . . . . . . . . . . 147A.1 Regular and After-Hours Trading for the NASDAQ Stock Ex-change . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168A.2 Historical Cumulative Abnormal Daily Returns around Earn-ings Announcements . . . . . . . . . . . . . . . . . . . . . . . 170A.3 The Response of Abnormal Volatility, Abnormal Quoted Spread,and Abnormal Turnover to Earnings Surprises around Earn-ings Announcements . . . . . . . . . . . . . . . . . . . . . . . 175C.1 Media Attention and Macroeconomic Fundamentals . . . . . 182xiAcknowledgmentsMy survival in the Ph.D. program and the fact that I will have a job inacademia would not have been possible without the support from manypeople. In particular, I owe a large part of my survival and appreciationfor great quality research to main my adviser Adlai Fisher. Adlai allowedme to join the Ph.D. program at the University of British Columbia andhelped me tremendously to understand empirical finance. Adlai taught mehow to appreciate data and, most importantly, how to let the data “speak”- i.e., let the data tell you things that you do not expect without imposingany constraints and how to analyze data from many angles. It took me awhile to live by this approach but I have to say that this research methodfits perfectly with me. Ali Lazrak played an important role in forcing meto explain my empirical finding to a theoretician and dedicated many hoursletting me explain what I was working on. These “sessions” really helpedme see the bigger picture of my work. Murray Carlson pushed me to look atthe finer details of my work and meticulously made me redo important partsof my work to make sure I covered all possible ground. Murray was alwaysmy go-to adviser for meticulous details. Beside their guidance, I thank theseindividuals for countless time dedicated to me and their encouragement.I would also like to thank the Finance department at the Sauder School,of Business. There are many great quality researchers in this school andfollowing my experience on the job market, I now understand why this de-partment, despite being small in comparison, stands out from many schools.I would like to thank two junior members of the finance department, MarkusBaldauf, and William Gornall. Both Markus and Will provided great train-ing for preparing the job market. Moreover, Markus taught me a lot ofinvaluable institutional details in the world of microstructure. I have alsoreceived valuable feedback from Professor Dale Griffin of the Marketing de-partment and Russell Lundholm of the Accounting department. In theEconomic department, I want to thank three superb theoretical econome-tricians: Hiroyuki (Hiro) Kasahara, Vadim Marmer, and Kevin Song. Theyare the ones who made me love my trip to the economic department. Hiro-xiiAcknowledgmentssan also played an important role guiding me with many tips for researchand the job market. Finally, it is important to highlight the great admin-istrative support that all Ph.D. students receive from Sally Bei and ElaineCho. They really simplify our lives.My peer group of Ph.D. students has been an important source of inspi-ration and for hard work. We helped and learned a lot from each other. Mypeer group is a strong class of researchers and I am happy that I will see allof them in future conferences. I have to highlight the great teamwork I havedeveloped with Jinfei Sheng. Jinfei is full of energy and has a sharp mindfor research and his a great co-author.I also thank my co-authors Oliver Boguth and Vincent Gre´goire (or Oliand Vinny). I learned a lot working with these two. Vincent really pushedmy understanding and comprehension of computer science to another level.I also thank all the support staff at Compute Canada and Westgrid for theaccess to supercomputing facilities.Beyond UBC, I thank the seminar participants at the HEC Montre´al,University of Virginia, Temple University, the University of Colorado atBoulder, Nanyang Technology University, University of Melbourne, Univer-sity of Toronto, Rice University, and McGill University. Many individualsgave helpful and constructive comments and I have already incorporatedsome of them into this thesis.I also have to thank many organizations for financial support that I havereceived during the Ph.D. They are: the NASDAQ OMX Educational Foun-dation, Montreal Exchange (Bourse de Montre´al), Bank of Montreal CapitalGroup, Canadian Securities Institute Research Foundation, and the SocialSciences and Humanities Research Council of Canada.Among the many personal acknowledgments that I owe, the first mustbe my wife Natsumi Fuwa who had to go through a lot during the Ph.D.She provided me with much support and happiness. I look forward to thenew stage of our lives together following the Ph.D. I must not forget tothank my parents who, despite not understanding what I do in life, alwaysencourage me to do what I love and most importantly to work hard and nottake anything for granted.Finally, this dissertation is dedicated to my sons Alex and Sasha whoxiiiAcknowledgmentsboth appeared in my life during the Ph.D. I owe a lot to my wife to be sucha dedicated mother.xivChapter 1IntroductionThis thesis is a collection of three essays at the intersection of Informa-tion Economics. Although the topics are diverse, they share the commonobjective of studying the interplay between asset prices and public newsevents at the corporate and institutional level. In the first essay, I exam-ine the speed at which unexpected news content of earnings announcementincorporate stock prices, i.e., price discovery, and the role of trade volumeto price discovery during the after-hours market. To investigate this ques-tion, I use a unique dataset from the nasdaq stock exchange that containprices and signed trade volume data for U.S. stocks between 2011 and 2015.In the second essay, I investigate the impact of the Federal Open MarketCommittee (fomc) announcements on financial markets and on investor at-tention to monetary policy when the announcement is accompanied withand without a press conference by the chairperson of the Federal Reserve.To study this question, I look at the response of asset prices and changes todifferent investor attention proxies before and after fomc announcements.I then compare the response of asset prices and investor attention for fomcannouncement with and without press conferences. The third essay docu-ments a novel channel to study the impact of the macroeconomy on assetprices through investor attention to macroeconomic risks. To measure in-vestor attention, I use daily newspaper article counts mentioning particularmacroeconomic risks from the Wall Street Journal and New York Times andstudy the relationship between investor attention and asset prices.Because each essay investigates a different topic in the field of Informa-tion Economics, chapters were designed to be self-contained. I thus leave amore exhaustive discussion of the research question and contribution to theintroduction specific to each chapter.1Chapter 2How is Earnings NewsTransmitted to Stock Prices?2.1 IntroductionA fundamental objective in financial economics is to understand how in-formation is transmitted to asset prices. Fama, Fisher, Jensen, and Roll(1969) present early evidence of how stock prices adjust to firm-level newsat a monthly frequency. More recent research shows how asset prices respondover short horizons to systematic news such as macroeconomic announce-ments (e.g., Andersen, Bollerslev, Diebold, and Vega, 2003a; Hu, Pan, andWang, 2015a).1 High-frequency price formation of individual stock pricesaround firm-level news announcements is less understood.In this paper, I examine price discovery following earnings announce-ments for the largest 1,500 U.S. stocks between 2011 and 2015. This topicis difficult to study at high frequency because a large proportion of earn-ings announcements, which are the most important type of firm-level news,occurs outside of regular trading hours (9:30 a.m. to 4 p.m. est). By incor-porating the after-hours market into my analysis of price formation, I amable to address several important questions.21Other related work on price formation following macroeconomic news includes Jones,Lamont, and Lumsdaine (1998a); Fleming and Remolona (1999a); Balduzzi, Elton, andGreen (2001a); Green (2004); Andersen, Bollerslev, Diebold, and Vega (2007a); Evansand Lyons (2008); Brogaard, Hendershott, and Riordan (2014) and Chordia, Green, andKottimukkalur (2016).2Patell and Wolfson (1984) and Woodruff and Senchack (1988) were the first to doc-ument intraday prices responses to earnings surprises. More recently, Jiang, Likitapiwat,and McInish (2012) show for a sample of S&P 500 stocks that an important share of pricevariation occurs in the after-hours market. Santosh (2014) study the impulse response pathof stock returns in business- and calendar-time units following earnings surprises in theafter-hours market and over the course of five trading days. Li (2016) implements a trad-ing strategy to take advantage of price drifts in the after-hours market following earningsannouncements. I study price discovery at high frequency using a similar methodology asAndersen, Bollerslev, Diebold, and Vega (2003a) and focus on when the impact of earningssurprises on the conditional mean changes in stock returns dissipates.22.1. IntroductionI first ask how quickly earnings surprises are incorporated into stockprices. Formally, I test for horizons at which earnings surprises have ex-planatory power. I show that, for my sample, price changes are affected byearnings surprises until 10 a.m. on the first session of regular trading follow-ing the earnings announcement. After 10 a.m., I find no evidence of post-earnings announcement drifts at any frequency, including the daily horizon.This result contrasts with literature that documents slow price formationfollowing earnings surprises.3 It is, however, well-known that slow price for-mation following earnings announcements is more pronounced in small andilliquid stocks (see e.g., Hou and Moskowitz, 2005; Chordia, Goyal, Sadka,Sadka, and Shivakumar, 2009).4To examine how quickly earnings surprises are incorporated into stockprices at high frequency, I utilize real-time quotations, transaction prices,and signed order flow from a limit order book exchange. I begin this anal-ysis at the 9:30 a.m. opening of markets by comparing two sets of stocks:stocks with and without after-hours trading following earnings announce-ments. Indeed, for 38 percent of my sample of earnings announcements, Ido not observe trades following earnings announcements in the after-hoursmarket. I document that stocks that are small and have low analyst andmedia coverage, low institutional ownership, and wider bid-ask spreads havea higher probability of no after-hours trading following earnings announce-ments. These stocks are predicted to have slower price discovery becauseof poor information quality (see Brennan, Jegadeesh, and Swaminathan,1993; Zhang, 2006). Controlling for the probability of having no after-hourstrading, I find that the after-hours close-to-open returns for stocks withafter-hours trading respond to earnings surprises by 40 percent more thanstocks with no after-hours trading. Using a similar methodology as Ander-sen, Bollerslev, Diebold, and Vega (2003a, 2007a), I show that stocks withno after-hours trading have significant price discovery that lasts 30 minutesfollowing the opening of markets. On the other hand, stocks with after-hours trading have no significant price discovery at the opening of markets,3Early papers documenting slow price formation to earnings news are Ball and Brown(1968) and Bernard and Thomas (1989). More recent evidence includes Doyle, Lundholm,and Soliman (2006), Hirshleifer, Lim, and Teoh (2009), and DellaVigna and Pollet (2009).4Boguth, Carlson, Fisher, and Simutin (2016) provide evidence of fast price formationof systematic news in large stocks. Bai, Philippon, and Savov (2016) show that marketshave become more efficient over time and this may explain why I observe no slow priceformation following earnings surprises at the daily frequency. In Section A.2 of the Ap-pendix, I show how the post-earnings announcements drift has changed since 1984 forsame sample selection criteria.32.1. Introductionwhich implies that all price discovery occurs in the after-hours market.5I then characterize the high-frequency dynamics of price discovery in theafter-hours market. I find that more than 80 percent of the total responseof stock returns to earnings surprises in the after-hours market occurs uponthe arrival of the first trades. I show that the initial price adjustments toearnings surprises occur as “jumps” followed by a price drift in the samedirection as the earnings surprise but the impact of earnings surprise dissi-pates in the after-hours market. Because earnings announcements lead toimportant price change in the after-hours market, this explains in part therecent findings of Bollerslev, Li, and Todorov (2016) regarding the higherrisk premium attached to estimated market betas using overnight close-to-open returns.6It is important to note that my results complements those of Santosh(2014).7 Santosh uses earnings surprises as instruments in structural equa-tions to estimate cumulative impulse response functions over five tradingdays following earnings announcements to test the invariance hypothesis ofKyle and Obizhaeva (2016). In its investigation, the author finds a cu-mulative impulse response that reflects 71 percent of the earnings news atthe opening of markets and close to 90 percent for stocks with high after-hours trading. It is comforting that I find similar results using anothermethodology commonly used in the literature of price discovery followingmacroeconomic news (e.g., Andersen, Bollerslev, Diebold, and Vega, 2003a,2007a). This methodology allows me to explicitly show when earnings sur-prises have no explanatory power to explain stock returns. Moreover, themethodology also allows me to investigate whether prices adjust more toearnings surprises or to order flow at the time of the announcement, whichconsist of the second objective of this paper.Santosh (2014) argues that price discovery following earnings announce-ments occurs through the arrival of order flow consistent with classical mi-crostructure models that suggest that transactions do affect prices becausethey convey information that is not common knowledge (e.g., Glosten andMilgrom, 1985; Kyle, 1985). Orders may be necessary to move prices fol-5These results do not imply that price discovery occurs in the after-hours marketbecause of actual trading. Liquidity providers can provide liquidity following earningsannouncements at prices that reflects instantly the news and trading can occur eventhough prices already reflect the new information (see Beaver, 1968).6Earnings announcements can increase stocks’ market betas because earnings an-nouncements generate systematic news (Patton and Verardo, 2012).7Santosh (2014) uses taq data and with a larger sample of stocks that spans the timeperiod of 2006 to 2011.42.1. Introductionlowing public announcements when liquidity providers (who are responsiblefor adjusting prices) have more limited information processing abilities thansome other traders (Kim and Verrecchia, 1994). On the other hand, theoryof public information associates the arrival of public news with instanta-neous price adjustment (e.g., Milgrom and Stokey, 1982; French and Roll,1986). In my data, I have signed order flow that allows me to investigatewhether prices adjust more to the actual news as predicted in French andRoll (1986) or to incoming order flow as in Kyle (1985) and Glosten andMilgrom (1985) and argued by Santosh (2014).I follow Evans and Lyons (2002) and document the explanatory power(R2) of earnings surprises and the net order imbalance (i.e., the differencebetween the total number of market-initiated buys and sells) to explainstock returns in the after-hours market following earnings announcements.8I find that the initial response of stock prices to earnings surprises occursdirectly. The R2 associated with the arrival of news explains ten percentof stock returns whereas net order imbalance explains only two percent.The explanatory power of earnings surprises on subsequent price changes is,however, short-lived and small, while the explanatory power of order imbal-ance remains sizable for the entire duration of the after-hours market. Pastresearch in foreign exchange markets largely attributes price adjustmentsaround macroeconomic news to order flow (see Evans and Lyons, 2008), butin the case of earnings announcements I find that the news itself largelyexplains the initial price adjustment. This implies that liquidity providersare capable at processing public information and incorporating news intoprices without relying on order flow.9The third objective of this paper is to examine how the magnitude ofearnings surprises impacts high-frequency abnormal stock price volatilities,abnormal bid-ask spreads, and abnormal trade volumes. Several empiricalpapers linked changes in price volatilities to price discovery following thearrival of news (see e.g., Ederington and Lee, 1993; Jones, Lamont, andLumsdaine, 1998a; Evans and Lyons, 2008). It is also important to extendthe analysis to trade volume and bid-ask spreads. Microstructure theorysuggests that changes in trade volume and bid-ask spreads are related to8Evans and Lyons (2002) examine the impact of order imbalance and nominal interestrate (public information) on daily foreign exchange prices. I refer the reader to Evansand Lyons (2002) and the working paper version Evans and Lyons (1999) for a simplestructural model motivating the empirical approach used in this paper.9Chordia, Green, and Kottimukkalur (2016), Brogaard, Hendershott, and Riordan(2015), and Baldauf and Mollner (2016) also provide evidence that liquidity providersplay a large role in price discovery.52.1. Introductionprice volatility and also reflect the arrival of information. I focus the analysisduring regular market hours prior to and after earnings announcements.I find significantly wider abnormal bid-ask spreads, lower abnormal stockprice volatility, and lower abnormal trade volume at high-frequency on trad-ing days prior to large earnings surprises. These results suggest that marketsanticipate the magnitude of earnings surprises and further suggests that thelarge earnings forecast errors in some stocks are explained, in part, by poorinformation quality (e.g., Kasznik and Lev, 1995; Lang and Lundholm, 1996)surrounding these stocks and, in turn, implies higher information asymme-try.10 Theory predicts that when information asymmetry is higher, tradingvolume may decrease before announcements because discretionary liquiditytraders postpone trading after the announcement is made (e.g., Admati andPfleiderer, 1988).I then examine the response of price volatility, bid-ask spreads, andtrade volume to earnings surprises following earnings announcements.11 Ifind that large earnings surprises lead to an increase in abnormal volatility,abnormal quoted spreads, and abnormal trade volumes at the opening ofmarkets following earnings announcements. As for the duration, the impactof earnings surprises on volatilities, spreads, and trade volumes graduallydecays over the course of regular trading hours following the opening ofmarkets, even though earnings surprises have no more impact on the ad-justments on the conditional mean of prices. The dynamics portrayed bythe abnormal volatility and trade volume are consistent with the theoreti-cal findings of Banerjee and Kremer (2010). The authors argues that tradevolume and volatility increases in the level of disagreement among investorson the interpretation of a public signal (i.e., agree to disagree) followed by agradual decay with the possibility of no adjustment in the conditional meanof prices.The last objective of this paper is to shed light on liquidity provisionaround earnings announcements. Liquidity provision is an important role ofstock markets and matters to price discovery (O’Hara, 2003). I find that ap-proximately 40 percent of incoming trade volume is executed against hidden10These results are similar to the “calm-before-storm” effect documented in Jones, La-mont, and Lumsdaine (1998a) and Akbas (2016).11An important literature documents the dynamics in trade volumes, volatilities, andspreads following earnings announcements at the daily (e.g., Beaver, 1968; Morse, 1981;Atiase and Bamber, 1994; Kandel and Pearson, 1995; Bamber, Barron, and Stober, 1997)and intraday horizon (Lee, Mucklow, and Ready, 1993). But, to my knowledge, this isthe first paper that documents the intraday dynamics conditioning on the magnitude ofthe earnings surprises.62.1. Introductionorders in the after-hours market following earnings announcements versus 12percent in regular market hours.12 This finding is significant because the ac-ceptance of hidden orders in financial markets is not unanimous among SECregulators and some suggest that hidden orders may deter the effectivenessof price discovery (see Shapiro, 2010). A liquidity provider may prefer hid-den orders because it helps uninformed traders to mitigate the option valueof limit orders that are expected to remain standing in the limit order bookfor a long period and, in turn, mitigate the risk of adverse selection (Harris,1996).13 On the other hand, Moinas (2011), Boulatov and George (2013),and Bloomfield, O’Hara, and Saar (2015) argue that informed traders mayprefer hidden orders.To understand whether hidden orders are beneficial to liquidity providers,I investigate the profitability of hidden orders versus displayed limit ordersfollowing earnings announcements in the after-hours market. If liquidityproviders earn higher profits with hidden orders than with displayed ordersthis would suggest that abolishing hidden orders could deter liquidity pro-vision and in turn deter price discovery following earnings announcements.I find that liquidity providers achieve profits (measured by realized spread)with displayed orders that are not statistically different from zero. But,liquidity providers that opt for hidden orders achieve significant positiveprofits. This finding suggests that abolishing hidden orders may deter theeffectiveness of price discovery following earnings announcements becauseliquidity traders may be less inclined to provide liquidity without the use ofhidden orders.The remainder of this paper is organized as follows. Section I describesthe data sources. In Section II, results on price discovery following earningssurprises, for both daily and intraday horizons, are presented. Price discov-ery in the after-hours market and the role of order flow to price discovery arepresented in Section III. The results of the impact of earnings surprises onvolatilities, bid-ask spreads, and trade volumes around earnings announce-ments are presented in Section IV. The profitability of hidden and displayedorders following earnings announcements is presented in Section V. Finally,12Hidden limit orders, like displayed limit orders, have price priority but always lose ontime priority against displayed limit orders. About 25 percent of incoming trade volumeis executed against hidden orders in the after-hours market when there are no earningsannouncements.13For example, a liquidity provider who is not fast enough to cancel their limit orderat the arrival of new information faces a higher risk of being “sniped” by a trader thatprocesses new information faster with a displayed order than with a hidden order. Bessem-binder, Panayides, and Venkataraman (2009) provide empirical support for the argumentof Harris (1996).72.2. DataSection VI concludes.2.2 Data2.2.1 Earnings Announcements SampleThe time coverage of this study is from January 1, 2011 to December 31,2015. I first select from the Center for Research in Security Prices (crsp)database stocks with nyse, nasdaq, or amex as their primary listing withshare code 10 or 11. Each stock must have Compustat data, precisely totalassets and market capitalization at the end of December of the previouscalendar year. I use these accounting metrics to later match each stock toone of the Fama-French 25 size and book-to-market portfolios. I then rankthe stocks by their market capitalization at the end of June of each year andselect the largest 1,500 stocks starting from 2010. I limit my sample to thelargest 1,500 stocks to minimize the computational constraint involved inprocessing the limit order book data, which I describe in the next section.I identify quarterly earnings announcements for the chosen sample stocksusing the announcement dates and times recorded in the Thomson ReutersI/B/E/S database. Because I/B/E/S timestamps are not always accurate(see Li, 2016; Santosh, 2014), I use the timestamps of the actual earningsnews in RavenPack to improve the accuracy. I match 87 percent of the earn-ings announcements from I/B/E/S with the earnings news in RavenPack.14For the missing 13 percent, I use the timestamps in I/B/E/S.When estimating the impact of earnings announcements on daily stockprices, announcements recorded as occurring at or after 4 p.m. on a givendate are relabeled for the purpose of this empirical analysis to have thefollowing trading day’s date, to reflect the fact that reactions to such an-nouncements are impounded in the stock’s price only on the following trad-ing day. This means that “day 0” in the event window is the trading day onwhich the reaction of investors to the earnings announcements trading on aU.S. exchange gets to impact the announcing firm’s stock price at the dailyhorizon.For each earnings announcement, I calculate the earnings surprise, de-fined as the scaled difference between actual and expected earnings:Si,t =epsi,t − êpsi,tPi,t−5, (2.1)14RavenPack is an intraday newswire provider. In the Internet Appendix of this paperI explain how to process RavenPack data and how to merge them with crsp.82.2. Datawhere epsi,t is the earnings per share of company i announced on day t, andêpsi,t is the forcasted earnings per share, calculated as the median consensusanalyst forecast. I scale the surprise using the stock price five trading daysbefore the announcement. I define the consensus analyst forecast as themedian of all analyst forecasts issued over the 90 days before the earningsannouncement date. If an analyst revises their forecasts during this interval,I use only their most recent forecasts. If a scheduled earnings announcementhas no earnings forecast, the earnings announcement observation is removedfrom the sample. I further winsorize earnings surprises at the 1st and 99thpercentile.In this paper, I focus only on after-hours earnings announcements (be-tween 4 p.m. and 9:30 a.m.), which represent 97 percent of the earningsannouncements in my sample. The final sample is composed of 25,552 earn-ings announcements with an average of 1,440 firms per year and a totalof 1,900 different firms between January 1, 2011 and December 31, 2015.15The earnings announcements are distributed as follows: 51.6 percent of theearnings announcements occur between 4 p.m. and 8 p.m., 47.1 percentoccur between 4 a.m. and 9:30 a.m., and 1.3 percent occur between 8 p.m.and 4 a.m.2.2.2 NASDAQ Limit Order Book-Level DataThroughout the paper I use high-frequency stock prices and trade volumedata from quotes and transactions from nasdaq’s TotalView-itch (here-after, nasdaq itch) limit order book, versions 4.1 and 5.0.16 nasdaq itchcontains a series of messages that describe orders added to, removed from,and executed on nasdaq for nasdaq-, nyse-, nyse Amex-, nyse Arca, andbats-listed securities. I construct a message-by-message limit-order book,where the book is updated whenever there is a new message that entersthe nasdaq exchange.17 nasdaq itch data differ from the commonly usedTrades and Quotes (taq) data provided by the nyse. Holden and Jacobsen(2014) document that taq can suffer from liquidity measurement problemsand errors in trade-quote matching due to insufficient timestamp granular-15On any given year, the sample of stocks represents approximately 90 percent of thetotal U.S. stock market capitalization traded on nyse, nasdaq, or amex with share code10 or 11.16See NASDAQ (2016a,b) for the official documentation on the data.17A Python code, developed in partnership with Vincent Gre´goire that constructs thelimit order book for nasdaq itch data version 4.1 and 5.0 will be made available on theMarket Empirical Analysis Toolbox for Python website http://www.meatpy.com. Thecode is adapted for multiprocessing.92.2. Dataity. On the other hand, itch data are publicly available at no cost and donot suffer from liquidity measurement problems and errors in trade-quotematching. But, processing these data and constructing the limit order bookare computationally costly. All trades in nasdaq itch are signed, excepttrades against hidden (i.e., non-displayed) limit orders starting from July14, 2014. I describe hidden orders in subsequent sections.18 Trades are notsigned in taq; the researcher must infer if a trade is a buy or a sell us-ing trade classification algorithms.19 When the empirical analysis requiressigned trades, the sample period starts on January 1, 2011 and ends on July13, 2014. Moreover, I observe every initiated trade that arrives in nasdaqitch, including the nasdaq portion of the Reg nms Intermarket SweepOrder and odd-lot orders.20After constructing the limit order book, I have for each stock an event-time midquote (the bid-ask mid point) timestamped to the nanosecond (abillionth of a second) from 9:30 a.m. to 4 p.m. I then aggregate the midquoteat a lower frequency (e.g., one- or five-minute intervals) using the last ob-servations at each interval. I also have for each stock the transaction data(price and quantity) and whether the trade was a market-initiated buy ormarket-initiated sell order from 4 a.m. to 8 p.m. After-hours trading onnasdaq is from 4 p.m. to 8 p.m. and resumes from 4 a.m. to 9:30 a.m.21I also observe crossing prices. Crossing prices are the price set at theopening and closing auctions (where the supply and demand curves meet atthe opening and closing auction). In addition, I process the spy ExchangeTraded-Fund (etf) that tracks the S&P 500 broad market index. I use thespy etf as a proxy for the intraday market return.2.2.3 Displayed and Hidden LiquidityBeing able to distinguish between hidden and displayed limit orders is impor-tant. When a trader wishes to provide liquidity with a limit order, she hasthe choice to display or hide the limit order. Hidden limit orders maintainprice priority but lose time priority to displayed orders at the same price.Therefore, displaying an order increases the chance of faster execution. Har-18See Section A.3 in the Appendix for more institutional details surrounding hiddenorders in nasdaq itch.19These trade classification algorithms are not flawless (see Chakrabarty, Pascual, andShkilko, 2015). Because liquidity is largely hidden in the after-hours market, it imposesimportant constraints on the effectiveness of trade classification algorithms.20Odd-lot orders are trades with less than 100 shares, can represent up to 60 percent ofthe total transactions (O’Hara, Yao, and Ye, 2014), and are not reported in taq.21See Figure A.1 for a graphical presentation of the trading hours on nasdaq.102.2. Dataris (1996) argues that hidden orders are effective for uninformed traders whowish to mitigate the option value of limit orders that are expected to re-main standing on the book for a long period and, in turn, mitigate therisk of adverse selection. On the other hand, Moinas (2011), Boulatov andGeorge (2013), and Bloomfield, O’Hara, and Saar (2015) argue that in-formed traders may prefer hidden orders. In Section 2.6, I document theimplication of hidden orders to price discovery following earnings announce-ments.2.2.4 Summary StatisticsTable 2.1 Panel A shows the sample stocks’ market capitalization at the endof June and analyst coverage breakdown by year and Panel B shows thecharacteristics of earnings announcements.112.2. DataTable 2.1: Descriptive StatisticsThis table reports descriptive statistics for the sample stocks, earnings an-nouncements, and trading activity used in this study. Panel A reports thedescriptive statistics on stock’s market capitalization (mcap) in million $at the end of June and analyst coverage. Panel B reports the descriptivestatistics for the earnings announcements. The after-hours announcementreturns are calculated between 4 p.m. prior to earnings announcements to9:30 a.m. on the following trading day. Panel C reports the descriptivestatistics for the trading activity on the nasdaq itch TotalView limit orderbook. Hidden corresponds to trades executed against hidden orders (i.e.,non-visible limit orders). Panel D reports the trading statistics by tradesize. ea corresponds to earnings announcements, ah corresponds to after-hours, and P25, P50, and P75 stand for the 25th, 50th, and 75th percentile.The sample period is January 1, 2011 to December 31, 2015.Panel A: Descriptive statistics on firm size (in million $) and analystcoverage2011 2012 2013 2014 2015MCAP min 794 721 901 1100 1199MCAP median 2924 2576 3147 4103 4115MCAP max 400885 547363 401730 556574 715600Number analysts P25 5 5 4 4 4Number analysts P50 9 9 8 8 8Number analysts P75 14 14 14 14 14122.2. DataPanel B: Descriptive statistics on earnings announcements2011 2012 2013 2014 2015Number of EA 5155 5015 5136 5142 5104% of earnings on Mond. 10 10 9 9 10% of earnings on Tues. 23 21 23 21 22% of earnings on Wed. 26 27 27 27 26% of earnings on Thurs. 33 34 33 34 33% of earnings on Frid. 5 6 6 6 6% of EA with AH trading 71 64 61 55 57Earnings surprisesMean 0.0008 0.0007 0.0006 0.0005 0.0005St. dev. 0.0039 0.0039 0.0036 0.0033 0.0035P25 -0.0001 -0.0002 -0.0002 -0.0002 -0.0002P50 0.0005 0.0005 0.0004 0.0004 0.0004P75 0.0018 0.0017 0.0015 0.0013 0.0014AH returns around EAMean 0.0000 0.0000 0.0007 0.0002 -0.0006St. dev. 0.0507 0.0533 0.0596 0.0527 0.0570P25 -0.0193 -0.0194 -0.0188 -0.0215 -0.0209P50 0.0005 0.0011 0.0021 0.0019 0.0011P75 0.0233 0.0222 0.0236 0.0264 0.0244132.2. DataPanel C: Descriptive statistics on trading activityMarket Hours After Hours After Hours (EA)P25 P50 P75 P25 P50 P75 P25 P50 P75Number of trades 669 1592 3679 1 3 8 4 16 104% hidden trade 8 11 16 12 25 50 19 29 40% hidden trade volume 8 12 18 8 25 60 21 41 59Panel D: Descriptive statistics on trading sizeNumber of shares per trade against displayed orders (%)< 100 100-500 500-1,000 > 1,000Market hours 32 66 1 1After hours 33 56 7 5After hours (EA) 30 60 6 4Number of shares per trade against hidden orders (%)< 100 100-500 500-1,000 > 1,000Market hours 27 71 2 1After hours 27 63 7 4After hours (EA) 22 61 8 8142.2. DataTrade size, in $ per trade, against displayed orders (%)< 1,000 1,000-5,000 5,000-50,000 > 50,000Market hours 17 55 28 0After hours 16 43 39 2After hours (EA) 12 46 40 2Trade size, in $ per trade, against hidden orders (%)< 1,000 1,000-5,000 5,000-50,000 > 50,000Market hours 15 52 32 1After hours 15 39 43 3After hours (EA) 11 36 47 6An important aspect of the data is worth mentioning. Despite firmsmaking earnings announcements in the after-hours market, I do not observetrades between the time of the announcement and the opening of marketsat 9:30 a.m. for approximately 38 percent of the earnings announcements.I show in the following section that a lack of after-hours trading followingearnings announcements indicates, in part, poor information quality sur-rounding these stocks, which results in slower price discovery.Panel C of Table 2.1 shows the percentiles for the number of trades andthe fraction of trades against hidden orders during regular market hours,in the after-hours market, and in the after-hours market when there is anearnings announcement across the sample of stocks. I observe that thelevel of trading activity increases in the after-hours market when there is anearnings announcement. Yet, the median number of trades in the after-hoursmarket, when there is an earnings announcement, is only 15. Note that themedian number of initiated trades and trade volume against hidden ordersis higher when there is an earnings announcement.22 Panel D presents thestatistics on the percentage of orders, by the number of shares per trade andby trade size (in dollars), that are executed against displayed and hiddenorders. Trades against hidden orders have a larger trade size than displayedorders and more so in the after-hours market. Large trade size indicates ahigher likelihood of the presence of institutional traders than retail tradersin the after-hours market.22Chakrabarty and Shaw (2008) also find more trades initiated against hidden orderson earnings announcement days.152.3. Price Discovery of Earnings Surprises: When is it Complete?2.3 Price Discovery of Earnings Surprises: Whenis it Complete?I now examine price discovery of earnings surprises at the daily horizon andat high frequency during regular market hours following earnings announce-ments.2.3.1 Are there Daily Post-Earnings Announcement drifts?To examine price formation at the daily horizon, I calculate for each stockin my sample the cumulative abnormal daily return starting five days beforeand ending 61 days after the earnings announcement. Following the sameprocedure as Hirshleifer, Lim, and Teoh (2009), I calculate the abnormaldaily return to account for return premia associated with size and book-to-market. I deduct from stock returns the return on the size and book-to-market benchmark portfolios obtained from Ken French’s website.23 Stocksare matched to one of 25 portfolios at the end of June of every year basedon their market capitalization at the end of June and their book-to-marketratio, calculated as the book equity of the last fiscal year end in the priorcalendar year divided by the market value of equity at the end of Decemberof the previous year.I plot in Figure 2.1 the average buy-and-hold abnormal returns (bhar)within each earnings surprises quintile and their corresponding 95 percentconfidence intervals around earnings announcements.23Data source: http://mba.tuck.dartmouth.edu/pages/faculty/ken.french/index.html162.3. Price Discovery of Earnings Surprises: When is it Complete?Figure 2.1: Abnormal Daily Returns around Earnings AnnouncementsThe figure shows the buy-and-hold cumulative abnormal returns (bhar)around earnings announcement announced on day 0 for each earnings sur-prise quintile sorts. I define the bhar for stock i from day τ to T (τ < T )as:BHAR[τ, T ]i =T∏k=τ(1 +Ri,k)−T∏k=τ(1 +Rp,k),where Ri,k is the return of the stock i and Rp,k is the return on the sizeand book-to-market matching Fama-French portfolio on day k. The figurerepresents the bhar [-5, T ] from five-days before the announcement (τ =−5) to day T following the announcement where T varies from T = −4 toT = 61 trading days. The shaded areas are pointwise 95% confidence bandsaround the average abnormal returns. The vertical line corresponds to theearnings announcement day. The sample consists of earnings announcementsfrom the largest 1,500 U.S. stocks between 2011 and 2015.−5 0 10 20 30 40 50 60Days since announcement−0.04−0.020.000.020.04CumulativeabnormalreturnTop quintileQuintile 4Quintile 3Quintile 2Bottom quintile172.3. Price Discovery of Earnings Surprises: When is it Complete?The first striking result is how “flat” the bhar are following earningsannouncements at day 0. Earnings surprises appear to be incorporated intothe first trading day. I report in Table 2.2 Panel A the tabulated format ofthe abnormal returns (ar) and the bhar over different trading day horizonsfollowing earnings announcements. The t-statistics are reported in bracketswhere the null is the ar and car are equal to zero. Panel B of Table 2.2shows the difference in ar and bhar between each quintile and quintile 3.Panel C shows the average ar and bhar for the top and bottom earningssurprises decile and the difference between both deciles. Table 2.2 shows noevidence of slow price formation at the daily horizon.182.3. Price Discovery of Earnings Surprises: When is it Complete?Table 2.2: Cumulative Daily Abnormal Returns following Earnings An-nouncementsPanel A of this table reports the abnormal returns (ar) and the buy-and-hold abnormal returns (bhar) at different horizons following earnings an-nouncements for each earnings surprises quintile. Panel B shows the differ-ence in the ar and the bhar between each earnings surprises quintile andquintile 3. Panel C shows the ar and bhar for the top and bottom earn-ings surprises deciles. The t-statistics where the null is zero are reported insquare brackets. The ar and bhar are calculated as follows:AR[τ ]i,q = Ri,τ −Rp,τ ,BHAR[τ, T ]i,q =T∏k=τ(1 +Ri,k)−T∏k=τ(1 +Rp,k),where Rik is the return of the stock i and Rpk is the return on the sizeand book-to-market matching portfolio on day k. The announcement dateof quarter q’s earnings occurs on day 0. The sample consists of earningsannouncements from the largest 1,500 U.S. firms between 2011 and 2015.Panel A: ar and car by earnings surprises quintileAR[0] AR[1] BHAR[2,5] BHAR[6,30] BHAR[31,61] BHAR[2,61]Top 0.03 0.002 0.001 -0.002 0.0 -0.001[31.2] [4.23] [0.94] [-1.33] [-0.07] [-0.63]Quintile 4 0.015 0.0 -0.001 0.003 0.0 0.002[17.99] [1.27] [-1.19] [2.42] [-0.13] [1.05]Quintile 3 0.004 -0.001 -0.001 0.004 0.0 0.003[5.19] [-2.0] [-2.22] [3.92] [-0.38] [1.53]Quintile 2 -0.012 -0.001 0.0 0.001 -0.002 -0.001[-15.53] [-3.0] [0.69] [0.67] [-1.79] [-0.74]Bottom -0.033 -0.001 0.001 0.001 -0.002 0.0[-31.7] [-3.34] [1.19] [1.01] [-1.06] [0.09]192.3. Price Discovery of Earnings Surprises: When is it Complete?Panel B: Difference in ar and car between each quintile and quintile 3AR[0] AR[1] BHAR[2,5] BHAR[6,30] BHAR[31,61] BHAR[2,61]Top-Q3 0.026 0.002 0.002 -0.006 0.0 -0.004[15.15] [3.25] [1.55] [-2.51] [0.13] [-1.03]Q4-Q3 0.011 0.001 0.001 -0.001 0.0 -0.001[7.13] [1.62] [0.63] [-0.69] [0.11] [-0.17]Q2-Q3 -0.016 0.0 0.001 -0.003 -0.002 -0.004[-10.45] [-0.5] [1.51] [-1.63] [-0.76] [-1.1]Bottom-Q3 -0.037 -0.001 0.002 -0.003 -0.001 -0.002[-20.62] [-0.99] [1.7] [-1.2] [-0.44] [-0.61]Panel C: ar and car for top and bottom earnings surprises decilesAR[0] AR[1] BHAR[2,5] BHAR[6,30] BHAR[31,61] BHAR[2,61]Top 0.034 0.002 0.001 -0.004 0.00 -0.003Bottom -0.037 -0.001 0.00 0.001 -0.003 -0.001Top-Bottom 0.071 0.003 0.001 -0.005 0.003 -0.001[21.719] [2.796] [0.482] [-1.249] [0.642] [-0.182]I report in Table 2.3 the estimated coefficients of a cross-sectional re-gression of ar and bhar on stock i’s respective earnings surprise Si,t.202.3. Price Discovery of Earnings Surprises: When is it Complete?Table 2.3: OLS Regression: Cumulative Abnormal Returns on EarningsSurprisesThis table reports the results of an ols regression of abnormal returns (ar)and cumulative abnormal returns (car) following earnings announcementsat different horizons on earnings surprises (Si,t). Standard errors are clus-tered by date and are reported in parentheses. Asterisks denote statisticalsignificance at the 5-percent level. The sample period is January 1, 2011 toDecember 31, 2015.AR[0] AR[1] CAR[2,5] CAR[6,30] CAR[31,61] CAR[2,61]Si,t 4.965* 0.293* 0.088 -0.259 0.225 0.157(0.165) (0.059) (0.094) (0.220) (0.251) (0.345)Intercept -0.002* -0.000 -0.000 0.002* -0.001 0.000(0.000) (0.000) (0.000) (0.001) (0.001) (0.001)Obs. 25552 25552 25548 25380 24088 24088Adj-R2 0.08 0.00 0.00 0.00 0.00 -0.00As expected, earnings surprises positively impact abnormal returns onthe earnings announcement day (AR[0]). An earnings surprise of 0.002,which is approximately the inter-quartile range between the 25th and 75thpercentile of earnings surprises, increases AR[0] by one percent. Also, earn-ings surprises positively and significantly impact AR[1] returns. Yet, theireconomic magnitudes are small, at about six basis points for an earnings sur-prise of 0.002 with a zero percent R2. More importantly, earnings surpriseshave no explanatory power on the car at any horizon.In Section A.2 of the Appendix, I show how the post-earnings announce-ments drift has changed since 1984 for the largest 1,500 U.S. stocks. It isobvious that markets have become more efficient at incorporating earningssurprises and only recently do we observe no strong evidence of post-earningsannouncement drift at the daily horizon.2.3.2 Are there Intraday Post-Earnings AnnouncementDrifts?I now investigate at high frequency the stock return response to earningsannouncements. In Figure 2.2, I plot the average cumulative abnormal logreturns at a five-minute frequency for each earnings surprises quintile start-ing on the trading day before the earnings announcement until the closing of212.3. Price Discovery of Earnings Surprises: When is it Complete?markets on the following trading day. The cumulative abnormal log returnis the difference between the cumulative log return of the announcing firm’sstock and the cumulative market log return proxied by the spy etf. At thisstage, I ignore the returns in the after-hours trading session. The overnight(close-to-open) return is calculated using the closing price at 4 p.m. and themidquote (mid-point between the best bid and best ask price) at 9:45 a.m.on the following trading day. I use midquotes starting at 9:45 a.m. becausefor a small number of observations I find that midquote prices in the orderbook between 9:30 a.m. and 9:45 a.m., are “noisy” (i.e., the midquote is farfrom the previous transaction price).222.3. Price Discovery of Earnings Surprises: When is it Complete?Figure 2.2: Cumulative Abnormal Intraday Returns around Earnings An-nouncementsThis figure shows the stocks’ cumulative abnormal five-minute log returnsbeginning at 9:45 a.m. on the trading day preceding an after-hours earn-ings announcement until 4 p.m. the following trading day. The cumulativeabnormal returns are calculated as the cumulative log returns for stock iminus the cumulative log returns of spy etf, a proxy for market returns.The overnight (close-to-open) return is calculated using prices at 4 p.m.preceding the earnings announcements and prices at 9:45 a.m. the follow-ing trading day. Each line represents a different quintile sort for earningssurprises. The shaded areas are pointwise 95% confidence bands around theaverage cumulative abnormal log returns. The sample period is January 1,2011 to December 31, 2015.9:45 16:00 9:45 16:00−0.04−0.03−0.02−0.010.000.010.020.030.04CumulativeabnormalreturnTop quintileQuintile 4Quintile 3Quintile 2Bottom quintileFrom Figure 2.2, we see a similar picture to Figure 2.1 where there is aclear demarcation between the earnings surprises quintiles. Moreover, thecar are also close to “flat” after the opening of markets. This suggests that232.3. Price Discovery of Earnings Surprises: When is it Complete?most, if not all, price discovery occurs in the after-hours market.2.3.3 The Response of After-Hours Returns to EarningsSurprisesIn this section, I quantify the impact of earnings surprises on after-hoursreturns calculated using prices at the closing (4 p.m.) and the opening ofmarkets (9:30 a.m.) on the trading day following the earnings announce-ment. More importantly, I examine whether a stock that has trading inthe after-hours market following earnings announcements influences the re-sponse of after-hours returns to earnings surprises. As previously shown, Ido not observe after-hours trading following earnings announcements on thenasdaq itch limit order book for 38 percent of earnings announcements inmy sample.24 A stock may not have after-hours trading following earningsannouncements due to factors such as stock visibility, information qualitysurrounding the stock, limited investor attention to the news, or that thenews is too complicated to process for liquidity providers to feel confidentto provide liquidity.The dominant economic factors that explain why a stock is more likely tohave after-hours trading following earnings announcements is an interestingtopic meriting further understanding, but is beyond the scope of this paper.Nonetheless, important literature documents slow price formation for stockswith poor information quality.25 I examine whether common proxies of in-formation quality surrounding a stock influence the likelihood of observinga trade in the after-hours market following earnings announcements. I re-port in Table 2.4 the estimated coefficients and marginal effects from a logitregression where the dependent variable is equal to one if the stock has noafter-hours trading following earnings announcements and zero otherwise.The independent variables are firm size, analyst and media coverage, in-stitutional ownership, and average bid-ask spreads. Firm size is based onthe logarithm market capitalization on the day prior to the earnings an-nouncement. Analyst coverage is the number of analyst forecasts prior toearnings announcements, and media coverage is the log of the total num-24It is possible that I may not observe a trade for a particular stock in the nasdaq itchlimit order book but a trade may have actually occurred on another exchange (i.e., darkpools, nyse limit order book). Yet, as I will show, stocks with no after-hours trading onnasdaq itch have slower price discovery. Therefore, this implies that price discovery didnot occur on another exchange.25See e.g., Brennan, Jegadeesh, and Swaminathan (1993); Hong, Lim, and Stein (2000);Hou and Moskowitz (2005); Zhang (2006), and Boguth, Carlson, Fisher, and Simutin(2016).242.3. Price Discovery of Earnings Surprises: When is it Complete?ber of articles in RavenPack with a relevance score of 90 or more in the21 trading days prior to earnings announcements. Institutional ownershipis the percentage of shares outstanding held by institutions from ThomsonReuters 13-F filings. The bid-ask spread is calculated using the averageof the one-second quoted spread measure (i.e., bid-ask spread divided bythe midquote) during regular trading hours in the 40 trading days prior toearnings announcements.2626I provide more details on the calculation of bid-ask spreads in Section 2.5.252.3. Price Discovery of Earnings Surprises: When is it Complete?Table 2.4: Logit Regression: Determinants to After-Hours Trading followingEarnings NewsThis table reports the results of a logit regression, where the dependent vari-able is equal to one if stock i has no trade in the after-hours market followingits earnings announcement and zero otherwise. The independent variablesare the stocks’ log market capitalization (Mcapi,t), the number of analystforecasts (Analystsi,t), the log of total number of newswire articles in Raven-Pack in the 21 trading days prior to earnings announcements (Mediai,t), thefraction of shares outstanding held by institutions (Institutioni,t), and theaverage quoted spread during regular trading hours in the 40 trading daysprior to earnings announcements (Spreadsi,t). The marginal effects are eval-uated at the mean. Asterisks denote statistical significance at the 5-percentlevel. The sample period is January 1, 2011 to December 31, 2015.Estimated coefficients Marginal effects (dy/dx)Mcapi,t -0.197* -0.045*(0.016) (0.004)Analystsi,t -0.054* -0.012*(0.003) (0.001)Mediai,t -0.068* -0.016*(0.015) (0.003)Institutioni,t -0.965* -0.221*(0.080) (0.018)Spreadsi,t 261.884* 59.922*(18.884) (4.255)Intercept 4.881*(0.384)Obs. 25133Pseudo-R2 0.09As expected, Table 2.4 shows that all of the coefficients for the indepen-dent variables are statistically significant with the correct predicted signs.This result emphasizes that stocks with no after-hours trading can be ex-plained, in part, by low information quality surrounding these stocks.I next use the predicted values from the logit regression to investigatewhether after-hours returns for stocks with a higher likelihood of after-hours262.3. Price Discovery of Earnings Surprises: When is it Complete?trading activity are more responsive to earnings surprises. To investigatethis possibility, I estimate the following regression:rahi,t = α+β1Si,t +Si,t ·β2ProbNoTradei,t +β3ProbNoTradei,t + i,t, (2.2)where time t denotes the after-hours time interval that starts at 4 p.m. priorto an earnings announcement and ends at 9:30 a.m. on the next tradingday. rahi,t denotes the log abnormal after-hours return and Si,t the earningssurprise for stock i. The abnormal after-hours return is calculated usingthe closing and opening prices from the auction if available; otherwise, Iuse the midquote from the limit order book.27 I then subtract the after-hours market return using the spy etf. ProbNoTradei,t corresponds tothe predicted values of having no trades in the after-hours market from thepreviously estimated logit regression.28I report the results in the first three columns of Table 2.5.27I exclude observations with after-hours returns in the top and bottom 1/1,000th ofthe distribution.28ProbNoTradei,t is a generated regressor. The error terms from the logit regressionand the regression specified in 2.2 are essentially uncorrelated (0.01); thus, adjustment forthe generated regressors is minimal.272.3. Price Discovery of Earnings Surprises: When is it Complete?Table 2.5: OLS Regression: After-Hours Returns on Earnings SurprisesThis table reports the regression results of the after-hours abnormal logreturn on earnings surprises. The after-hours abnormal returns are calcu-lated using the closing price at 4 p.m. prior to earnings announcements andopening price at 9:30 a.m. on the following trading day minus the marketreturn proxied by the spy etf over the same interval. Si,t is the earningssurprise. ProbNoTradei,t is the predicted probability of having no tradesin the after-hours market following earnings announcements based on thelogit regression reported in Table 2.4. NoTradei,t is a dummy variable equalto one if there is no trade in the after-hours market following the earningsannouncement and zero otherwise. BMOi,t is a dummy variable equal toone if the earnings announcement occurs before the market opens (12:00a.m. to 9:30 a.m.) and zero otherwise. Anni,t is the number of earnings an-nouncements in the after-hours market. Fridayt is a dummy variable equalto one if the earnings announcement occurs on a Friday and zero otherwise.Mediai,t is the stocks’ media coverage based on the log of the total num-ber of newswire articles in RavenPack following the earnings announcementuntil the opening of markets. Standard errors are clustered by date andare reported in parentheses. Asterisks denote statistical significance at the5-percent level. The sample period is January 1, 2011 to December 31, 2015.282.3. Price Discovery of Earnings Surprises: When is it Complete?(1) (2) (3) (4) (5) (6)Si,t 3.850* 4.868* 4.463* 4.812* 5.475* 4.512*(0.108) (0.251) (0.132) (0.252) (0.367) (0.465)Si,t × ProbNoTradei,t -2.369* -0.873 -1.199* -0.767(0.556) (0.600) (0.610) (0.619)Si,t ×NoTradei,t -2.016* -1.929* -1.823* -1.751*(0.181) (0.198) (0.204) (0.207)Si,t ×BMOi,t -0.838* -0.865*(0.209) (0.210)Si,t ×Anni,t -0.001 -0.001(0.003) (0.003)Si,t × Fridayi,t -0.215 -0.184(0.384) (0.381)Si,t ×Mediai,t 0.337*(0.100)NoTradei,t 0.005* 0.003 0.003 0.003(0.002) (0.002) (0.002) (0.002)ProbNoTradei,t 0.003* 0.002* 0.002* 0.002*(0.001) (0.001) (0.001) (0.001)BMOi,t 0.000 -0.000(0.000) (0.000)HighAnni,t 0.002* 0.002*(0.001) (0.001)Fridayi,t 0.003* 0.003*(0.001) (0.001)Mediai,t -0.001(0.000)Intercept -0.006* -0.008* -0.006* -0.007* -0.009* -0.007*(0.002) (0.002) (0.002) (0.002) (0.002) (0.002)Obs. 25133 25133 25133 25133 25133 25133Adj-R2 0.08 0.08 0.09 0.09 0.09 0.09Year-Month FE Y Y Y Y Y YColumns (1) and (2) show a positive and significant relationship betweenearnings surprises and after-hours returns. In Column (1), for an increase inearnings surprises (Si,t) of 0.002, the after-hours return increases by 77 basispoints. In Column (2), I find that the after-hours return of stocks with a 100percent probability of no after-hours trading following an after-hours earn-ings announcement respond 49 percent less to earnings surprises than stockswith a zero percent probability of no after-hours trading. Next, I replaceProbNoTradei,t with NoTradei,t, which corresponds to a dummy variableequal to one if I observe no actual after-hours trading followings earningsannouncements and zero otherwise. The results in Column (3) show thatthe impact of NoTradei,t on after-hours returns is quantitatively similarto ProbNoTradei,t. In Column (4), I combine both the actual realizationand the probability of having no trades in the after-hours market. Theresults in Column (4) show that, controlling for the probability of havingno after-hours trading, the after-hours returns for stocks with after-hourstrading respond to earnings surprises 40 percent more than stocks with no292.3. Price Discovery of Earnings Surprises: When is it Complete?after-hours trading. In Column (5), I report the results from the previousregression by including additional control variables related to investor atten-tion. I include an interaction variable Si,t ×BMOi,t, where BMOi,t equalsone if the announcement occurs before the market opens (between 12:00 a.m.and 9:30 a.m.). Intuition suggests that firms that announce earnings beforethe market opens give investors less time to process the news than earningsannounced the night before. I further interact the earnings surprise with adummy variable, Fridayt, which equals one if the earnings announcementoccurs on a Friday, and an additional interaction term, Annt, which corre-sponds to the total number of earnings announcements in the after-hoursmarket on date t. Hirshleifer, Lim, and Teoh (2009) and DellaVigna andPollet (2009) respectively show that when firms announce earnings on Fri-days or on days with a high number of earnings announcements, investorsare more likely to be inattentive and the price reaction to earnings surprisesis weaker and subject to more persistent price drifts. I report the resultsin Column (5). I find no statistical significance at the five percent levelfor the interaction between the earnings surprises and Fridayt and Annt.29But, the interaction term Si,t × BMOi,t is significant and negative, whichindicates potential additional price discovery at the opening of markets forstocks with earnings announcements that occur before market opens.Another factor likely to influence the response of after-hours returns toearnings surprises is media coverage. Peress (2008) finds that stocks with lessmedia coverage have longer post-earnings announcement drifts. To proxy formedia coverage, I count the total number of articles appearing in the intra-day newswire database RavenPack between the time of the announcementand the opening of markets. I interact the earnings surprise with Mediai,t,which is the log of the total number of articles about stock i. I report theresults in Column (6). The interaction term is positive and statisticallysignificant at the five percent level.Overall, the results show that stocks’ after-hours returns around earningsannouncements are less responsive to earnings surprises if there is no after-hours trading following the announcement. We should, therefore, expectadditional and significant price discovery for these stocks at the opening ofmarkets. Moreover, stocks with low media coverage and stocks with earningsannouncements that occur before the market opens are also expected to haveadditional price discovery.29Chakrabarty, Moulton, and Wang (2015) show that, with the advent of high-frequencytradings, the impact of limited attention on cumulative abnormal returns after earningsannouncements is diminished.302.3. Price Discovery of Earnings Surprises: When is it Complete?2.3.4 The Dynamics of Price Discovery following EarningsAnnouncements at the Opening of MarketsIn this section, I investigate whether any price discovery remains followingearnings surprises at the time the market opens at 9:30 a.m. The empiricalapproach is inspired from Andersen, Bollerslev, Diebold, and Vega (2003a,2007a).I first construct a panel dataset for each stock i that contains the five-minute log return ri,τ starting at 9:30 a.m. and ending at 10:30 a.m. (9:35a.m. is the first five-minute observation) following earnings announcementsusing the first transaction price starting at 9:30 a.m., the earnings surpriseSi,t, announced in the previous after-hours trading session prior to the open-ing of markets, the after-hours return rahi,t , and the five-minute market returnrmτ using the spy etf. I use transaction prices to calculate the returns. Notethat τ corresponds to a five-minute interval, for a total of twelve five-minuteintervals between 9:30 a.m. and 10:30 a.m. I estimate the following cross-sectional ordinary least squares (ols) regression:ri,τ = α+ βτSi,t + γτrahi,t + δrmτ + i,τ . (2.3)I control for after-hours return rahit because it may influence how the mar-kets respond to earnings surprises at opening. Because the model containsso many variables, it would prove counterproductive to report all of the pa-rameters estimates. The coefficients of interest are the estimated βˆτ andare plotted in Figure 2.3 with their corresponding 95 percent confidenceintervals. The standard errors are calculated using the Driscoll-Kraay ex-tension of the Newey-West hac estimator (Driscoll and Kraay, 1998). TheDiscoll-Kraay method is a generalized method of moments technique forlarge cross-sectional and time dimensions panel datasets. The coefficient es-timates are identical to ols estimates but the standard errors are robust toheteroskedasticity and to general forms of spatial and temporal dependence.312.3. Price Discovery of Earnings Surprises: When is it Complete?Figure 2.3: The Response of Stock Returns to Earnings Surprises at theOpening of MarketsThis figure shows the estimated response coefficients βˆτ from the stock returnconditional mean regression (2.3):ri,τ = α+ βτSi,t + γτrahi,t + δrmτ + i,τ .τ corresponds to a five-minute interval between 9:30 a.m. and 10:30 a.m.Earnings announcements are announced in the after-hours market preced-ing the opening of markets at 9:30 a.m. The shaded areas are pointwise95% confidence bands around the estimated coefficients. The standard er-rors are calculated using the Driscoll and Kraay (1998) method. Panel Ashows the estimated coefficients for the full sample of earnings announce-ments and Panel B and Panel C respectively show the results for stocks withno after-hours trading and with after-hours trading following earnings an-nouncements. The sample period is January 1, 2011 to December 31, 2015.322.3. Price Discovery of Earnings Surprises: When is it Complete?−0.20.00.20.40.60.81.0ResponsePanel A: Full sample−0.20.00.20.40.60.81.0ResponsePanel B: No after-hours trading9:359:409:459:509:5510:0010:0510:1010:1510:2010:2510:30−0.20.00.20.40.60.81.0ResponsePanel C: With after-hours tradingIn Figure 2.3, Panel A shows the estimated coefficients βˆτ for the fullsample of earnings announcements. Also, Panel B and Panel C respectivelyshow the estimated coefficients for stocks with and without after-hours trad-ing. I previously documented that no after-hours trading is the strongestfactor influencing the response of stocks’ after-hours (close-to-open) returnsto earnings surprises. Prices of these stocks are less responsive to earningssurprises and therefore we should expect these stocks to have additional andsignificant price discovery at the opening of markets.Panel A shows a moderate impact of earnings surprise on stock returns (acoefficient of 0.4) at the opening of markets followed by a slow decay endingaround 10 a.m. For stocks with no after-hours trading, the general pattern332.3. Price Discovery of Earnings Surprises: When is it Complete?is one of a quick mean adjustment, characterized by a jump at the openingof markets followed by a slow decay. An increase in the earnings surpriseof 0.002 increases returns by 17 basis points and a total cumulative impactof 30 basis points by 10 a.m. In Panel C, we see that stocks with after-hours trading have on average small, if any, remaining price discovery whenmarkets open. For stocks in Panel C, we must then explore price discoveryin the challenging context of after-hours trading, which I undertake in thefollowing section.In Table 2.6 Panel A, I report in a tabulated format the estimated co-efficients βˆτ between 9:30 a.m. and 10 a.m. of Figure 2.3. I also reportthe estimated coefficients for different sub-samples based on high (top quar-tile) and low (bottom quartile) predictability of having after-hours tradingfollowing earnings announcements, announcement time (i.e., earnings an-nouncements before market opens or after market closes), and for high (topquartile) and low (bottom quartile) media coverage based on the total num-ber of articles in RavenPack between the time of the announcement and theopening of markets. I also report the sum of the estimated coefficients forboth βˆτ and γˆτ between 9:30 and 10 a.m. After-hours returns may containinformation about the news not captured by earnings surprises.342.3.PriceDiscoveryofEarningsSurprises:WhenisitComplete?Table 2.6: Price Discovery following Earnings Surprises at the Opening of MarketsPanel A of this table reports the estimated response coefficients βˆτ and γˆτ from thestock return conditional mean regression (2.4):ri,τ = α+ βτSi,t + γτrahi,t + δrmτ + i,τ .rahi,t is the after-hours return and rm is the market return proxied by the spy etf. After-hours (ah) returns are calculated using the stock price at 4 p.m. prior to earningsannouncements and the stock price at 9:30 a.m. following earnings announcements.After-hours trading refers to stocks with one or more trades following the earningsannouncement in the after-hours market. The probability of after-hours trading corre-sponds to the stocks’ predicted probability of having after-hours trading based on a logitregression reported in Table 2.4. After market closes refers to earnings announcementsbetween 4 p.m. and 11:59 p.m. and before market opens to earnings announcementsbetween 12:00 a.m. and 9:30 a.m. Media coverage corresponds to the total number ofnewswire articles in RavenPack between the earnings announcement time and 9:30 a.m.Low and high respectively correspond to the to the bottom and top quartile. Standarderrors are clustered by date and reported in parentheses. Asterisks denote statisticalsignificance at the 5-percent level. Panel B shows the R2 from two univariate regres-sions: (1) stock returns on earnings surprises Si,t and (2) stock returns on after-hoursreturns rahi,t using stock returns calculated from 9:30 a.m. to 10 a.m. and from 10 a.mto 4 p.m. The sample period is January 1, 2011 to December 31, 2015.352.3.PriceDiscoveryofEarningsSurprises:WhenisitComplete?Panel A: The response of stock returns to earnings surprises and after-hours returns at opening of marketsβτ∑τ βτ∑τ γτ9:30-9:35 9:35-9:40 9:40-9:45 9:45-9:50 9:50-9:55 9:55-10:00 9:30-10:00 9:30-10:00Full sample 0.284* 0.126* 0.145* 0.078* 0.034 0.042* 0.723* 0.098*Actual AH tradingNo AH Trading 0.791* 0.255* 0.175* 0.154* 0.058* -0.008 1.452* 0.283*With AH trading 0.063 0.073 0.136* 0.048 0.024 0.060* 0.412* 0.051*Probability of AH tradingLow 0.480* 0.248* 0.163* 0.065 0.023 0.018 1.076* 0.151*High 0.010 -0.054 0.092 0.018 -0.045 0.056 0.068 0.041*Announcement timeAfter market closes 0.307* 0.159* 0.173* 0.126* 0.012 0.044* 0.814* 0.081*Before market opens 0.225* 0.086* 0.125* 0.038 0.055* 0.041 0.604* 0.126*Media coverageLow 0.365* 0.142 0.174* 0.111* 0.071* 0.024 0.899* 0.109*High 0.046 0.074 0.064 -0.003 -0.010 0.063 0.257 0.072*362.3. Price Discovery of Earnings Surprises: When is it Complete?Panel B: Explanatory power (R2) of earnings surprises and after-hoursreturns to stock returns9:30-10:00 10:00-4:00R2Surprise R2AH Return R2Surprise R2AH ReturnFull sample 0.01 0.03 0.00 0.00Actual AH tradingNo AH trading 0.05 0.11 0.00 0.00With AH trading 0.01 0.01 0.00 0.00Probability of AH tradingLow 0.03 0.05 0.00 0.00High 0.00 0.01 0.00 0.00Announcement timeAfter market closes 0.01 0.02 0.00 0.00Before market opens 0.02 0.04 0.00 0.00Media coverageLow 0.02 0.03 0.00 0.00High 0.00 0.02 0.00 0.00I find that stocks with a high predictability of having after-hours tradinghave no significant price discovery at the opening of markets. This suggeststhat stocks with high information quality affect the speed of price discovery.Similarly, stocks with high media coverage have no significant price discoverybut I find the opposite for stocks with low media coverage. I find littledifference in price discovery for stocks with earnings announcements thatoccur before the market opens or after the market closes. Yet, the impactof after-hours returns is greater for stocks that announce before the marketopens.In Panel B, I show the explanatory power (R2) of a univariate regressionof stock returns on earnings surprises and stock returns on after-hours re-turns between 9:30 to 10 a.m. and from 10 a.m. and 4 p.m. I choose a cutoffof 10 a.m. because this is where price discovery following earnings surprisesis generally complete in Figure 2.3. Consistent with the results of PanelA, earnings surprises for stocks with no after-hours trading have the highestexplanatory power to explain stock returns (R2 of five percent) between 9:30a.m. and 10 a.m. Also, the after-hours return has a high explanatory power372.4. Price Discovery following Earnings Surprises in the After-Hours Market(R2 of eleven percent), for stocks with no after-hours trading. Stocks with ahigh probability of after-hours trading have an R2 of zero percent for earn-ings surprises and one percent for after-hours returns. After 10 a.m., I findthat all R2 are equal to zero for the full sample and across subgroups, whichsuggests that price discovery following earnings surprises and after-hoursreturns is generally complete by 10 a.m.2.4 Price Discovery following Earnings Surprisesin the After-Hours Market2.4.1 Market Activity in the After Hours around EarningsAnnouncementsBefore I examine price discovery in the after-hours market, it is worthwhileto highlight the differences in market activity across stocks in the after-hoursfollowing earnings announcements. I show in Figure 2.4 an example of stockprice and trade volume (in hundreds of shares) reactions around an earningsannouncement scheduled at 4:30 p.m. on October 18, 2011 for a large liquidfirm, Apple Inc. (aapl) at a one-minute frequency between 3:30 and 5:30p.m.30The figure shows little trading volume in the limit order book after themarket closes at 4 p.m. At the time of the announcement (4:30 p.m.),the stock price drops following a negative earnings surprise and high tradevolume occurs.In Figure 2.5 Panel A, I show the distribution of total trades (log scale)between the time of the earnings announcement and the opening of marketsat 9:30 a.m. for my sample of stocks with after-hours trading followingearnings announcements.30I calculate the stock price as the volume-weighted transaction price.382.4. Price Discovery following Earnings Surprises in the After-Hours MarketFigure 2.4: An Example of Price Response to Earnings Announcements atHigh FrequencyThis figure shows the stock price and trade volume (in hundreds of shares) ata frequency of one minute between 3:30 p.m. and 5:30 p.m. for the companyApple (ticker: aapl) around the earning announcement made at 4:30 p.m.on October 18, 2011. The black dots are the volume-weighted transactionprices. The positive blue bars are the initiated market buy orders and thenegative red bars are the initiated market sell orders.390395400405410415420425430Stockprice15:3015:4015:5016:0016:1016:2016:3016:4016:5017:0017:1017:20−1500−1000−50005001000Volume392.4. Price Discovery following Earnings Surprises in the After-Hours MarketFigure 2.5: Statistics on After-Hours Trading following Earnings Announce-mentsThis figure shows in Panel A the distribution of the total number of trades(in log scale) between the time of the earnings announcement and the open-ing of markets at 9:30 a.m. for all earnings announcements with after-hourstrading. Panel B shows the distribution of the trading time (in hours) be-tween the first trade following the earnings announcement and the actualearnings announcement. P25 and P75 stand for the 25th and 75th per-centiles, respectively.Panel A: Distribution of total trades in the after-hours market followingearnings announcements0 2 4 6 8 10 12Log total trades0500100015002000FrequencyMean: 3.05P25: 1.098Median: 2.70P75: 4.634Panel B: Lapse time distribution between the first trade and earningsannouncements0 2 4 6 8 10Time (in hours)0100020003000400050006000700080009000FrequencyMean: 1.28P25: 0.013Median: 0.31P75: 1.556402.4. Price Discovery following Earnings Surprises in the After-Hours MarketNote that the mean is 3.05 and the median is 2.70 (a total of 21 and 15trades), suggesting that there are indeed only a few trades for more thanhalf of the sample. But, for some earnings announcements, the total numberof trades is in the thousands. In Panel B, I show the lapse of time (in hours)between the first trade and the earnings announcement. The mean and themedian are 1.28 and 0.31 hours, respectively. For 25 percent of the sample,the first trade occurs within 47 seconds.31Another question of interest is who is participating in the after-hoursmarket. The nasdaq itch data do not contain trader identification foreach order entry in the limit order book. Barclay and Hendershott (2004)show that adverse selection risk is higher in the after-hours market, whichsuggests that traders who participate in the after-hours market are morelikely to be informed and sophisticated. As shown in Table 2.1 Panel Dand Panel E, trade size both in shares and in dollars is greater in the after-hours market than during regular market hours, consistent with the ideathat large trade size is more likely to come from institutional traders thanretail traders.322.4.2 The Dynamics of Price Discovery in the After-HoursMarketIn this section, I examine price discovery in the after-hours market. Becauseno liquidity providers have the obligation to provide liquidity in the after-hours, prices are not continuous. For example, we may observe availableliquidity only the bid side of the book and nothing on the ask side. Dur-ing market hours, each stock has a designated market maker that providesliquidity on both sides of the book. Moreover, a large share of liquidityis hidden. Therefore, working in calendar time using midquotes to calcu-late returns is not feasible. To overcome this challenge, for each stock withafter-hours trading I calculate returns over ten intervals denoted k usingthe arrival of trades to define an interval. For instance, if a firm has tentrades, each trade arrival represents a trade bin. If a firm has five tradesthen it has only five trade arrival bins k. If a firm has more than ten tradesthen I divide the number of total trades in the after-hours by ten (a fraction31Even large firms can have a delay between the announcement and the first tradebecause of trading halts imposed by the exchange.32In Section A.4 of the Appendix, I use another dataset to investigate whether high-frequency trading is predominant in the after-hours market. Compared to regular tradinghours, I find that high-frequency traders are less present in the after-hours market.412.4. Price Discovery following Earnings Surprises in the After-Hours Marketof total trades) and a trade bin k contains a fraction of the total trades.33Essentially, I use business-time units rather than calendar-time units to cal-culate stock returns. The return over a trade arrival bin is the sum of thelog returns using transaction prices. I use the last trade prior to the earn-ings announcement to calculate returns for the first trade bin. I choose thearrival of trades and not trading volume to construct trade bins because theliterature has shown that the arrival of trades has a larger impact on stockprice volatility than trade volume (see Jones, Kaul, and Lipson, 1994).Figure 2.6 shows the average cumulative return following earnings sur-prises at the announcement in business time in the after-hours market. Tradebin k = 1 is the first trade bin following the announcement.33For example, if a firm has 15 trades, this represent 1.5 trades per bin. The first binwill contain the first trade following the announcement, the second bin contains the secondand third trade, the third bin contains the fourth trade, and so on.422.4. Price Discovery following Earnings Surprises in the After-Hours MarketFigure 2.6: Cumulative Returns following Earnings Announcements in theAfter-Hours MarketThis figure shows the stocks’ cumulative returns following earnings an-nouncements in the after-hours market. The x-axis corresponds to tradebins. The definition of a trade bin is described in the main text. Each linerepresents a different quintile sort for earnings surprises. The shaded areasare pointwise 95% confidence bands around the average returns. Panel Ashows the cumulative returns for all stocks with after-hours trading follow-ing earnings announcements (ea). Panel B shows the cumulative returnsfor stocks with more than 20 trades in the after-hours market following ea.Panel C shows the cumulative returns for stocks with less than or equal to20 trades following ea. Panel D zooms in on the first trade bin of Panel Band shows cumulative returns over ten trade bins following ea. The dashedvertical line is the arrival of the first trade bin following the earnings an-nouncement. The sample period is January 1, 2011 to December 31, 2015.−0.06−0.04−0.020.000.020.04CumulativereturnPanel A: Full sampleTop quintile Quintile 4 Quintile 3 Quintile 2 Bottom quintilePanel B: High trade announcements0 2 4 6 8 10Trade bins since announcement−0.06−0.04−0.020.000.020.04CumulativereturnPanel C: Low trade announcements0 2 4 6 8 10Trade bins since announcementPanel D: Zoom in on the first bin for high trade announcements432.4. Price Discovery following Earnings Surprises in the After-Hours MarketPanel A shows the cumulative return for the full sample of firms withafter-hours trading. The figure shows a clear demarcation between the dif-ferent earnings surprises quintiles at the first trade bin. I then split thesample of firms into high trade announcements (more than 20 trades follow-ing the announcement) and low trade announcements (less than or equalto 20 trades) and plot their cumulative returns in Panels B and C respec-tively.34 Panel C shows longer price drift than in Panel B and the initialprice adjustment to earnings surprises is also more moderate. In Panel D,I “zoom in” on the first trade bin of Panel B. I take all trades in the firsttrade bin for firms with more than 20 total trades in the after-hours andonce more construct ten trade bins. We now also observe price drifts forlarge firms at higher frequency.I now quantify the impact of earnings surprises on stock returns on eachtrade bin by estimating the following model:ri,k = α+ βkSi,t + i,k, (2.4)where k defines a trade bin. Similar to Figure 2.6, I show in Figure 2.7 theestimated βˆk for the full sample in Panel A, for the high trade announce-ments in Panel B, for the low trade announcements in Panel C, and zoom-inon the first trade bin (k = 1) for high trade announcements in Panel D.34The mean number of trades in the after-hours is 20, and 48 percent of firms have morethan 20 trades.442.4. Price Discovery following Earnings Surprises in the After-Hours MarketFigure 2.7: The Response of Stock Returns to Earnings Surprises in theAfter-Hours MarketThis figure shows the estimated response coefficients βˆk of the conditionalmean regression (2.4):ri,k = α+ βkSi,t + i,k,where k corresponds to trade arrival bins. The definition of a trade bin isdescribed in the main text. The shaded areas are pointwise 95% confidencebands around the estimated coefficients. The standard errors are calcu-lated using the Driscoll and Kraay (1998) method. Panel A shows the stockprice response coefficients βˆk for all stocks with after-hours trading follow-ing earnings announcements (ea). Panel B shows the stock price responsecoefficients for stocks with more than 20 trades in the after-hours marketfollowing ea. Panel C shows the stock price response coefficients for stockswith less than or equal to 20 trades following ea. Panel D zooms in on thefirst trade bin of Panel B and shows the stock price response coefficientsover ten trade bins following ea. The sample period is January 1, 2011 toDecember 31, 2015.012345ResponsePanel A: Full sample Panel B: High trade announcements1 2 3 4 5 6 7 8 9 10Trade bin since announcement012345ResponsePanel C: Low trade announcements1 2 3 4 5 6 7 8 9 10Trade bins since announcementPanel D: Zoom in on the first bin for high trade announcements452.4. Price Discovery following Earnings Surprises in the After-Hours MarketPanel A shows that price discovery occurs over the first three trade bins.The impact of earnings surprises on returns is one of a “jump” followed bya quick decay in the remaining response of returns to earnings surprises.With an earnings surprise of 0.002, the initial jump amounts to an increasein return of 75 basis points. The initial jump represents approximately 83percent of the total price response to earnings surprises in the after-hoursmarket. The median completion time of the first trade bin in calendar time-units is 18 minutes. Panel B and Panel C show almost no difference inthe speed of price discovery between high and low trade announcements. Areason why the speed of price discovery appears similar is because speed ismeasured in business-time units (e.g., arrival of trades) rather than calendar-time units, consistent with the microstructure invariance hypothesis of Kyleand Obizhaeva (2016) and with the findings of Santosh (2014). But, thespeed of price discovery in calendar time is not similar between groups.Assuming that price discovery completes by the end of the third trade bin,I find that the median and mean time to completion of price discoveryof earnings surprises for high (low) trade count firms is, respectively, 0.61(1.31) and 1.84 (2.86) hours. Lastly, Panel D shows that, within the firsttrade arrival bin for stocks with a high trade count following announcements,we do indeed observe “slow” price discovery. Overall, the results show thata large share of price discovery for stocks with after-hours trading occursaround the arrival of the first trades.2.4.3 How is Earnings News Transmitted to Stock Prices?The previous results show that stock prices respond to earnings surprisesalmost immediately at the time of the first trades. What is not clear, how-ever, is whether earnings surprises impact prices directly, indirectly throughincoming trades (order flow), or both. French and Roll (1986) and Flemingand Remolona (1999a) argue that publicly available news may be incorpo-rated in prices instantaneously, even without trading.In the absence of news, it is generally assumed that asset prices primarilyadjust through incoming market order flow, specifically net order imbalance.This is consistent with classic theories of intermediation (e.g., Kyle, 1985;Glosten and Milgrom, 1985). Net order imbalance is the difference betweenbuyer-initiated and seller-initiated market orders - it is a measure of netbuying pressure. Net order imbalance conveys information that liquidityproviders need to aggregate into prices. If news impacts prices throughorder flow, then net order flow should largely explain price changes followingearnings announcements and not earnings surprises.462.4. Price Discovery following Earnings Surprises in the After-Hours MarketTo test whether earnings surprises (news) or order flow explain pricechanges following earnings announcements, I use the same methodology asEvans and Lyons (2002). These authors estimate a structural model wherechanges in daily foreign exchange rates are determined by public informationand aggregate order imbalances. Formally, the change in log price followingthe arrival of news in Evans and Lyons (2002) can be stated as∆Pt = St +OIt, (2.5)where St is the surprise, OIt is the order imbalance, and ∆Pt is the changein log price following the news over interval t. Evans and Lyons (2002,2008) show that order imbalance, and not public macroeconomic news (e.g.,changes in interest rate), is the main determinant of daily exchange ratesand argue that foreign exchange dealers have limited ability to interpret thenews. The model of Evans and Lyons (2002) is adaptable at high frequencyand one can show whether stock prices respond primarily to news or to orderflow following earnings announcements.Similar to Evans and Lyons (2002), I study the explanatory power (R2) ofnet order imbalance and earnings surprises to explain the response of stockreturns following earnings announcements in the after-hours market overeach trade arrival bin defined in the previous section. If the explanatorypower of earnings surprises is greater than order imbalance, then pricesrespond primarily to news and not order flow.I define market-initiated net order imbalance (OI) in trade bin k as:OIk =Bk − SkBk + Sk, (2.6)where Bk and Sk respectively correspond to trade buys and sells in sharesunits in trade bin k.35 Because I observe only trades that occurs on nasdaq,an important assumption is that at any moment in time, the OI is thesame across all other trading venues. Li (2016) shows that nasdaq hasthe highest fraction of trades following earnings announcements during theafter hours with 44% followed by nyse with 38%. I show in Figure 2.8the average order imbalance across all trade bins for each earnings surprisesquintile. The figure shows that negative earnings surprises lead to moreselling pressure and vice versa for positive news. Also, note that the bottomearnings surprises quintile leads to greater net order imbalance (in absoluteterms) than the highest earnings surprises quintile.35I find quantitatively the same result in the paper using the number of buy and selltrades instead of using trade buys and sells in shares units.472.4. Price Discovery following Earnings Surprises in the After-Hours MarketFigure 2.8: Order Imbalance following Earnings Announcements in theAfter-Hours MarketThis figure shows the average net order imbalance at each trade bin acrossdifferent earnings surprises quintiles following earnings announcements forstocks with after-hours trading. The definition of a trade bin is described inthe main text. Trade bin one corresponds to the first trade bin following theearnings announcement. The earnings surprises quintiles are sorted fromthe lowest (1) to the highest (5). The order imbalance is calculated as thedifference between market-initiated buy and sell orders (in shares units)divided by the total market-initiated buys and sells orders. The sampleperiod is January 1, 2011 to July 13, 2014.Trade bins1 2 3 4 5 6 7 8 9 10Earnings surprises quintile12345Orderimbalance-0.20-0.17-0.13-0.10-0.07-0.030.000.030.070.10In Figure 2.9, I show the R2 for two distinct sets of univariate regressionsof stock returns on earnings surprises (Si,t) and order imbalance (OIk) ateach trade arrival bin k following earnings announcements.36The figure shows that earnings surprises explain more than ten percentof the initial stock price reaction to the arrival of news whereas order imbal-ance explains slightly less than two percent. After the first trade arrival bin,earnings surprises have almost no explanatory power. On the other hand,the explanatory power of order imbalance is approximately three percent.36Note that the sample period ends on July 13, 2014. As previously noted, nasdaqitch does not include signed trades against hidden orders from July 14, 2014.482.4. Price Discovery following Earnings Surprises in the After-Hours MarketFigure 2.9: Explanatory Power of Earnings Surprises and Order Imbalanceto Stock Returns in the After-Hours MarketThis figure shows the R2 from a univariate regression of stock returns onearnings surprises (solid blue line) and stock returns on incoming net orderimbalance (dotted red line) at each trade bin k following earnings announce-ments in the after-hours market. Net order imbalance is the difference be-tween market-initiated buy and sell orders (in shares units) divided by thetotal market-initiated buy and sell orders. The x-axis units are the tradebins. The definition of a trade bin is described in the main text. The sampleperiod is January 1, 2011 to July 13, 2014.1 2 3 4 5 6 7 8 9 10Trade bins since announcement0.000.020.040.060.080.10R2Earnings Surprise (Si,t)Order Imbalance (OIi,k)Because the largest share of price discovery following earnings announce-ments occurs at the first trade bin (approximately 80 percent) and earningssurprises explain ten percent of the initial price adjustment, we can concludethat price discovery in the after-hours market largely occurs directly fromthe arrival of news. Yet, order flow remains sizable for the remaining of theafter-hours.I report in Table 2.7 Panel A the results of regressions of stock returns492.4. Price Discovery following Earnings Surprises in the After-Hours Marketbetween the earnings announcements and the opening of markets on earningssurprises and order imbalance. I also include as independent variables thelog of the total number of trades (Trdi,t), analyst dispersion (Dispi,t), andinteraction terms Si,t × Trdi,k, OIi,t × Trdi,t, OIi,t × Si,t, OIi,t × Dispi,t.Order imbalance may play a larger role if there is more trade, when the Si,tis negative, as depicted in Figure 2.8, or when analyst dispersion prior toearnings announcement is high (see Pasquariello and Vega, 2007). Analystdispersion is calculated as:Dispi,t =√Vt−1[epsi,t]|Et−1[epsi,t]| , (2.7)where Vt−1[epsi,t] is the variance of all the forecasts of earnings that ana-lysts issue for company i within an interval of ninety days before the an-nouncement. I calculate the dispersion only for companies with at least fouranalysts estimate prior to the earnings announcements.502.4. Price Discovery following Earnings Surprises in the After-Hours MarketTable 2.7: OLS Regression: Stock Returns on Earnings Surprises and OrderImbalanceThis table reports coefficients from regressions of the log stock returns fol-lowing earnings announcements in the after-hours market on earnings sur-prises (Si,t) order imbalance (OIi,k), log total number trades (Trdi,k), andanalyst dispersion (Dispi,t). The definition of a trade bin is described inthe main text. The order imbalance is calculated as the difference betweenmarket-initiated buy and sell orders (in shares units) divided by the totalmarket-initiated buy and sell orders. Panel A shows the results for all stockswith after-hours trading over the entire after-hours period following earningsannouncements. Panel B shows the results in the first trade bin (k = 1) andover all remaining trade bins (k > 1). Panel C shows the results for stockswith more than 20 trades following earnings announcements and zooms inon the first trade bin and reconstructs a new set of ten trade bins. Thestandard errors are clustered by date and reported in parenthesis. Asterisksdenote statistical significance at the 5-percent level. The sample period isfrom January 1, 2011 to July, 14, 2014.Panel A: After hours(1) (2) (3) (4) (5) (6) (7)Si,t 4.431* 2.518* 2.429* 2.445* 2.744*(0.147) (0.254) (0.249) (0.249) (0.279)OIi,t 0.010* 0.002* 0.002* 0.002* 0.002*(0.001) (0.001) (0.001) (0.001) (0.001)OIi,t × Trdi,t 0.005* 0.004* 0.004* 0.004*(0.001) (0.001) (0.001) (0.001)Si,t × Trdi,t 0.552* 0.540* 0.537* 0.459*(0.088) (0.087) (0.087) (0.092)Si,t ×OIi,t 0.129 0.055(0.177) (0.197)Dispi,t ×OIi,t 0.000(0.003)Trdi,t -0.001* -0.002* -0.002* -0.002* -0.002*(0.000) (0.000) (0.000) (0.000) (0.000)Dispi,t 0.001(0.003)Intercept -0.005* -0.001* 0.002* 0.000 0.001 0.001 0.000(0.001) (0.001) (0.001) (0.001) (0.001) (0.001) (0.001)Obs. 11255 11255 11255 11255 11255 11255 9555Adj-R2 0.10 0.01 0.03 0.11 0.12 0.12 0.12512.4. Price Discovery following Earnings Surprises in the After-Hours MarketPanel B: After hours - per trade binTrade bin k = 1 Trade bins k > 1(1) (2) (3) (4) (5) (6) (7) (8) (9) (10)Si,t 3.761* 2.992* 2.922* 3.106* 0.737* 0.302 0.226 0.264(0.127) (0.140) (0.138) (0.158) (0.096) (0.196) (0.193) (0.221)Si,t × Trdi,k 0.500* 0.481* 0.382* 0.119* 0.116 0.103(0.083) (0.083) (0.087) (0.060) (0.060) (0.067)OIi,k 0.005* 0.004* 0.004* 0.002* 0.002* 0.001(0.000) (0.000) (0.001) (0.001) (0.001) (0.001)OIi,k × Trdi,k 0.003* 0.002* 0.003* 0.003* 0.003* 0.003*(0.001) (0.001) (0.001) (0.000) (0.000) (0.000)Si,t ×OIi,k -0.102 0.099(0.143) (0.152)Dispi,t ×OIi,k 0.002 0.001(0.002) (0.003)Trdi,k -0.002* -0.001* -0.002* -0.002* -0.000 -0.000 -0.000 -0.000*(0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000)Dispi,t 0.002 -0.001(0.002) (0.002)Intercept -0.004* -0.002* 0.001 -0.001* -0.002* -0.001* -0.000 0.000 -0.000 0.000(0.000) (0.000) (0.001) (0.000) (0.001) (0.000) (0.001) (0.001) (0.001) (0.001)Obs. 11255 11255 11255 11255 9555 10040 10040 10040 10040 8570Adj-R2 0.11 0.12 0.02 0.13 0.13 0.01 0.01 0.03 0.03 0.03Panel C: After hours - zoom in on the first trade binTrade bin k = 1 Trade bins k > 1(1) (2) (3) (4) (5) (6) (7) (8) (9) (10)Si,t 3.448* 1.825* 1.997* 2.176* 1.278* 0.861* 0.825* 0.787*(0.162) (0.551) (0.552) (0.599) (0.145) (0.190) (0.186) (0.208)Si,t × Trdi,k 0.310* 0.246* 0.220 0.157 0.140 0.104(0.108) (0.109) (0.116) (0.088) (0.087) (0.097)OIi,k -0.010* -0.009* -0.009* 0.002* 0.002* 0.001(0.002) (0.002) (0.002) (0.001) (0.001) (0.001)OIi,k × Trdi,k 0.004* 0.003* 0.003* 0.003* 0.002* 0.002*(0.001) (0.000) (0.000) (0.000) (0.000) (0.000)Si,t ×OIi,k 0.114 -0.360*(0.201) (0.174)Dispi,t ×OIi,k -0.003 0.007*(0.003) (0.003)Trdi,k -0.002* -0.001* -0.002* -0.002* -0.000 -0.000 -0.000 -0.000(0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000)Dispi,t 0.004 -0.002(0.003) (0.002)Intercept -0.005* 0.004 0.006* 0.004 0.003 -0.003* -0.002* -0.001 -0.002* -0.001(0.001) (0.002) (0.002) (0.002) (0.002) (0.000) (0.001) (0.001) (0.001) (0.001)Obs. 5480 5480 5480 5480 4832 5480 5480 5480 5480 4832Adj-R2 0.11 0.11 0.04 0.14 0.14 0.03 0.03 0.02 0.05 0.05522.5. The Impact of Earnings Surprises on Volatility, Liquidity, and Trade VolumeComparing the R2 in columns (1) and (2) shows that returns are largelyexplained by the news and not order flow. Column (5) shows the resultswith the interaction term Trdi,t and including order flow improves the R2by one percent. Columns (6) and (7) show that OIi,t×Si,t and Dispi,t×OIi,tis not different from zero and does not improve the explanatory power.Table 2.7 Panel B reports the results of regressions of stock returns inthe first trade bin and over all remaining trade bins on earnings surprisesand order imbalance. In the first trade bin k, the results show that theearnings surprises largely explain returns and not order flow. I repeat thesame analysis in Panel C but zoom in on the first trade bin for a sub-sampleof stocks with more than 20 trades following earnings announcements andI reconstruct a new set of ten trade bins. R2 results show that returns aredriven largely by earnings surprises with an R2 of 11% but including orderflow and its interaction improve the R2 to 14%. It seems that when thereis little trading, order flow carries more information, yet earnings surprisesmatter more. If I extend the analysis during regular market hours for stockswith no after-hours trading, I find that order imbalance does not have anyexplanatory power to explain stock returns between 9:30 and 10 a.m.The overall results suggest that prices respond directly to public infor-mation. This indicates that liquidity providers are sophisticated at process-ing news and largely responsible for price adjustment in response to newsthrough limit order quote updates. This result supports the recent findingsof Brogaard, Hendershott, and Riordan (2015) and Chordia, Green, andKottimukkalur (2016), who show that price discovery largely comes fromquote adjustments.2.5 The Impact of Earnings Surprises onVolatility, Liquidity, and Trade VolumeFor a more comprehensive understanding of price formation following earn-ings surprises, one must go beyond the study of the impact of surprises onconditional mean changes in prices. For instance, volatility in prices is equiv-alent to information flow in a large class of models (e.g., Ross, 1989). Severalempirical papers (see e.g., Ederington and Lee, 1993; Jones, Lamont, andLumsdaine, 1998a; Andersen, Bollerslev, Diebold, and Vega, 2003a) studythe response of abnormal volatility in bond and foreign exchange prices fol-lowing macroeconomic news and associate the response to price discovery.3737Beaver (1968) argues that price changes in response to earnings news reflect changes inexpectations of the market as a whole while an increase in trade volume reflects changes in532.5. The Impact of Earnings Surprises on Volatility, Liquidity, and Trade VolumeIn this section, I examine how the magnitude of earnings surprises impact athigh frequency the dynamics of abnormal stock price volatilities, abnormaltrade volumes, and abnormal bid-ask spreads on three days around earningsannouncements during regular market hours. Microstructure theory sug-gests that changes in trade volume and bid-ask spreads are related to pricevolatility and also reflect the arrival of information.How is the magnitude in earnings surprises expected to impact volatility,trade volume, and bid-ask spreads? Stocks with large earnings surprises (i.e.large forecast error) is explained, in part, to poor information quality (e.g.,Kasznik and Lev, 1995; Lang and Lundholm, 1996) surrounding these stocks.Consequently, stocks with poor information quality force investors to acquirediverse information to better interpret the news. The poorer the informationquality surrounding the stock, the more diverse is information about theexpectation of the news among investors. Kim and Verrecchia (1991, 1994)argue that trade volume following earnings announcements increases in thelevel of asymmetry among investors prior to the announcement. Moreover,at the announcement, large surprises may also lead to larger dispersion inthe interpretation of the news among investors. Theory predicts that tradevolume also increases in the level of disagreement in the interpretation ofthe news (Kandel and Pearson, 1995; Banerjee and Kremer, 2010). Kimand Verrecchia (1994) further advance that higher information asymmetryat the announcement increases trading opportunities for informed traders,which leads to an increase in bid-ask spreads. When trade volume increases,volatility also increases (Kim and Verrecchia, 1994; Banerjee and Kremer,2010).I do not limit my analysis solely following earnings announcements butalso on trading days prior to announcements. Doing so provides an indi-cation of whether markets anticipate the magnitude of earnings surprisessimilar to the “calm-before-storm” effect before anticipated news as docu-mented in Jones, Lamont, and Lumsdaine (1998a) and Akbas (2016).To measure abnormal intraday volatility, I estimate the following modelfor each stock i separately:rτ = α+ ρrτ−1 + γrmτ + βτSt · 1{τ∈t} + τ , (2.8)where τ corresponds to a five-minute interval between 9:30 a.m. and 4 p.m.,the expectations of individual investors. Earnings news may be neutral and not change theexpectations of the market as a whole but greatly alter the expectations of individuals. Inthis case, we would observe no price change but there would be shifts in portfolio positionsreflected in trade volume and price volatility.542.5. The Impact of Earnings Surprises on Volatility, Liquidity, and Trade Volumerτ is the log five-minute returns using midquotes, rmτ is the market returnproxied by the spy etf, and St is the earnings surprise release on date t inthe after-hours market. The indicator variable 1{τ∈t} takes the value one ifthe five-minute interval τ belongs to the earnings announcement day t. Idefine the idiosyncratic volatility for stock i as |ˆτ |. There are in total 78five-minute intervals in a trading day t. I pool all 40 trading days prior toan earnings announcement and the day of the announcement to estimateEquation (2.8) for each stock i separately.Following the estimation of Equation 2.8, I sum the estimated |ˆτ | at each30-minute interval, for a total of 13 |ˆτ˜ |, which corresponds to a 30-minuteintraday volatility estimate for interval τ˜ on date t.I measure liquidity using the quoted bid-ask spread measure. For eachstock i, I have the best bid and ask prices at every second interval s duringregular market hours. I define the one-second quoted spread asQSi,s,t =Aski,s,t −Bidi,s,tPi,s,t, (2.9)where Pi,s,t is the midquote, (Aski,s,t+Bidi,s,t)/2, at the one second intervals on date t. I then average the QSi,s,t over a 30-minute interval to get atime-weighted quoted spread measure denoted QSi,τ˜ ,t.I calculate trade volume using the measure of turnover. Denote Vi,τ˜ ,t asthe total number of shares traded in a 30-minute interval τ˜ for stock i ondate t. I define trade turnover asTurni,τ˜ ,t =Vi,τ˜ ,tOuti,t, (2.10)where Outi,t is the current shares outstanding. I further scale Turni,τ˜ ,t byits standard deviation in the trading window (-40, -11) preceding an earningsannouncement for that year. I scale by the standard deviation to control forchanges in normal, non-announcement period turnover across time.In Figure 2.10 I show the average intraday volatility, quoted spreads, andturnover 40 to 11 trading days prior to earnings announcements per earningssurprises quintile. Even if we exclude two weeks (in trading days) prior tothe earnings announcement, we observe that stocks with upcoming largesurprises have higher volatility and quoted spreads and lower turnover. Ifwe compare stocks with large surprises (top or bottom quintiles) and stockswith no surprises (quintile 2) at 12 p.m., volatility is greater for stocks withlarge surprises by 23%. Quoted spreads are wider by 25% and turnover is7% lower for stocks with large surprises than for stocks with no surprises.552.5. The Impact of Earnings Surprises on Volatility, Liquidity, and Trade VolumeEmpirical evidence from the accounting literature suggests that stocks withupcoming large forecast errors are stocks with poor information quality, e.g.,less analyst coverage and less information disclosure coming from the firm(see e.g., Kasznik and Lev, 1995; Lang and Lundholm, 1996). Stocks withpoor information quality imply higher information asymmetry that leads towider bid-ask spread (Chae, 2005) and to higher information uncertaintythat leads to higher stock price volatility (Zhang, 2006).To estimate the impact of earnings surprises on abnormal volatility, Iestimate the following model:|ˆi,τ˜ |−|¯i,τ˜ |= a+ bτ˜ |Si,t|+cσd(t)√13+ ei,τ˜ , (2.11)where |ˆi,τ˜ |-|¯i,τ˜ | is the volatility for stock i for interval τ˜ minus the averagevolatility in the 40 to 11 trading days prior to earnings announcementsfor the same interval τ˜ . σd(t) is the daily volatility of the market, whichis the one-day-ahead volatility forecast for day d(t) from a simple dailyconditionally Gaussian garch (1, 1) using the broad stock market indexfrom Kenneth French’s website. I estimate Equation (2.11) on three tradingdays around the earnings announcement. In total, I estimate 39 bˆτ˜ (13 pertrading day).In Figure 2.11, Panel A, I plot the estimated bˆτ˜ .562.5. The Impact of Earnings Surprises on Volatility, Liquidity, and Trade VolumeFigure 2.10: Average Volatility, Quoted Spread, and Turnover prior to Earn-ings AnnouncementsThis figure shows the average 30-minute volatility, quoted spread, andturnover in the 40 to 11 trading days prior to earnings announcementsduring regular market hours for each absolute earnings surprises quintile.Volatility is the sum of the five-minute absolute value of the residuals inEquation (2.8) estimated for each stock i seperately:rτ = α+ ρrτ−1 + γrmτ + βτSt · 1{τ∈t} + τ ,over a 30-minute interval. Quoted spread is the average of the time-weightedone-second quoted spread defined as bid-ask spread divided by the midquotein a 30-minute interval. Turnover is the sum of total shares traded in a 30-minute interval divided by the number of shares outstanding and scaled bythe standard deviation of that year. The sample period is January 1, 2011to December 31, 2015.0.0040.0060.0080.0100.0120.014Panel A: VolatilityTop quintileQuintile 4Quintile 3Quintile 2Bottom quintile0.00050.00100.00150.00200.00250.0030Panel B: Quoted spread10:00 11:00 12:00 13:00 14:00 15:00 16:001.01.11.21.31.41.51.61.71.8Panel C: Turnover572.5. The Impact of Earnings Surprises on Volatility, Liquidity, and Trade VolumeFigure 2.11: The Response of Abnormal Volatility, Abnormal QuotedSpread, and Abnormal Turnover to Earnings Surprises around Earnings An-nouncementsThis figure shows the estimated coefficient responses of abnormal volatil-ity, abnormal quoted spread, and abnormal turnover to absolute earningssurprises around earnings announcements at each 30-minute interval dur-ing regular trading hours. The regression specifications are described in themain text. The left pane shows the day before the earnings announcement(ea), the middle pane is the ea day, and the right pane is the day afterthe ea. The ea occurs in the after-hours market (between 4 p.m. and 9:30a.m.) indicated by the straight dashed vertical lines. Volatility is the sum ofthe five-minute absolute value of the residuals in Equation (2.8) estimatedfor each stock i seperately:rτ = α+ ρrτ−1 + γrmτ + βτSt · 1{τ∈t} + τ ,over a 30-minute interval. Quoted spread is the average of the time-weightedone-second quoted spread defined as bid-ask spread divided by the midquotein a 30-minute interval. Turnover is the sum of total shares traded in a 30-minute interval divided by the number of shares outstanding and scaled bythe standard deviation of that year. The shaded areas are pointwise 95%confidence bands around the estimated coefficients. The standard errors arecalculated using the Driscoll and Kraay (1998) method.Panel A: Abnormal volatility response to earnings surprises10:0011:0012:0013:0014:0015:0016:00−0.50.00.51.01.52.02.53.0ResponseDay before EA10:0011:0012:0013:0014:0015:0016:00EA Day10:0011:0012:0013:0014:0015:0016:00Day after EA582.5. The Impact of Earnings Surprises on Volatility, Liquidity, and Trade VolumePanel B: Abnormal quoted spread response to earnings surprises10:0011:0012:0013:0014:0015:0016:00−0.10−0.050.000.050.100.15ResponseDay before EA10:0011:0012:0013:0014:0015:0016:00EA Day10:0011:0012:0013:0014:0015:0016:00Day after EAPanel C: Abnormal turnover response to earnings surprises10:0011:0012:0013:0014:0015:0016:00−2000200400600ResponseDay before EA10:0011:0012:0013:0014:0015:0016:00EA Day10:0011:0012:0013:0014:0015:0016:00Day after EAThe vertical dashed lines correspond to the after-hours trading sessionwith the earnings announcement. On the day before earnings announce-ments, stocks with an absolute earnings surprise of 0.003 (approximatelythe inter-quartile range in absolute earnings surprises) lead to a 0.075 per-cent decrease in abnormal volatility at the opening of markets until 2 p.m.This magnitude represents an approximate 15 percent decrease in volatilityaround 1 p.m. relative to the average volatility in the benchmark window592.5. The Impact of Earnings Surprises on Volatility, Liquidity, and Trade Volume(-40, -11). On the day of the announcement, for the same magnitude ofabsolute earnings surprises, abnormal volatility jumps by 0.9 percent at theopening of markets followed by a gradual decay. This increase in volatil-ity represents an approximate 82 percent increase in stock price volatilityat the opening of markets relative to the benchmark window. On the fol-lowing trading day, the estimated bˆτ˜ are in general negative. This suggeststhat stocks with higher volatility prior to earnings announcements have theirvolatilities move closer to the group of stocks with smaller earnings surprisesprior to earnings announcements.I next examine the impact of earnings surprises on bid-ask spreads. I es-timate Equation (2.11) with QSi,τ˜ −QSi,τ˜ as the dependent variable, whereQSi,τ˜ is the average quoted spread 40 to 11 trading days prior to earningsannouncements. I plot in Panel B the estimated bˆτ˜ . I find that liquidityproviders widen spreads in anticipation of large earnings surprises of approx-imately three percent at the opening of markets. The economic magnitudeis small but as shown in Figure 2.10, stocks with large upcoming surprisesalready have wider bid-ask spreads many days before the announcement.On the day of the announcement, quoted spreads widen by 12 percent atthe opening of markets relative to the benchmark window and the impactof earnings surprises on quoted spreads gradually decays. I show in FigureA.3 the comparison in the dynamics for stocks with and without after-hourstrading. The change in dynamics for quoted spreads is largely driven bystocks with no after-hours trading.Finally, I examine the impact of earnings surprises on trade volume. Iestimate Equation (2.11) with Turni,τ˜ −Turni,τ˜ as the dependent variable,where Turni,τ˜ is the average turnover 40 to 11 trading days prior to earningsannouncements. I also control for turnover in the spy etf to proxy formarket trade volume rather than market volatility. I plot in Panel C theestimated bˆτ˜ . The impact of earnings surprises on the day prior to earningsannouncements is economically large. At the opening of markets, for anabsolute earnings surprise of 0.003, turnover is lower by 52 percent relativeto the average turnover in the benchmark window (-41, -11). On the day ofthe announcement, turnover increases by 158 percent relative to the averageturnover in the benchmark window. The impact of earnings surprises onturnover gradually decays on the day of the announcement.Overall, the dynamics in volatility, bid-ask spread, and turnover lead-ing to earnings announcements indicate that markets anticipate the magni-tude of earnings surprises. The response of volatility, bid-ask spreads, andturnover to absolute earnings surprises on the earnings announcement day is602.6. Hidden Liquidity around Earnings Announcementsmore gradual than the impact of earnings surprises on the conditional meanadjustment of prices. The model of Banerjee and Kremer (2010) provides in-sights to this finding. In their model, the level of trade volume and volatilitygradually decays following a jump because of disagreement among investorson the interpretation of public information. The decay reflects convergencein beliefs among investors on the valuation of the asset. As beliefs convergesvolume and volatility decreases. On the other hand, asset prices reflect theaverage valuation among investors and the average may not change whilebeliefs on the valuation among investors still differ. Yet an interesting ques-tion remain. Why is the impact of earnings news volatilities, volumes, andspreads longer-lived than its impact on prices?2.6 Hidden Liquidity around EarningsAnnouncementsThe last objective of this paper is to shed light on an interesting fact aboutliquidity following earnings announcements in the after-hours market. Ifind that 41 percent of the trade volume involves hidden orders followingearnings announcements in the after-hours market versus only 12 percentduring regular market hours and 25 percent during after hours when thereis no earnings announcements. But, the acceptance of hidden orders by theSEC is still an on-going debate because hidden orders make markets lesstransparent (Shapiro, 2010).What is the rational for liquidity providers to choose hidden liquidity?Harris (1996) and Bessembinder, Panayides, and Venkataraman (2009) ar-gue that hidden orders are effective for mitigating adverse selection. On theother hand, Bloomfield, O’Hara, and Saar (2015) show in a lab experimentthat informed traders may prefer hidden orders so as to not reveal how muchthey are willing to buy or sell and earn higher profits. Recent theoreticalworks suggest that hidden orders lead to deeper limit order books (Moinas,2011), intensify competition among informed traders, and improve marketefficiency (Boulatov and George, 2013). Assuming that liquidity providersthat opt for hidden orders are indeed informed traders on the true fun-damental price of the stock following earnings announcements, abolishinghidden orders may deter the willingness of informed liquidity providers toparticipate and consequently deteriorate the speed of price discovery.3838When a trade occurs against a hidden order, market participants do learn that a tradegot executed against a hidden order. For example, on nasdaq, market participants seethe message order P when a trade gets executed against a hidden order. But, starting612.6. Hidden Liquidity around Earnings AnnouncementsI now investigate the profitability of hidden orders versus displayed limitorders from the perspective of liquidity providers following earnings an-nouncements. If, on average, the profitability associated with hidden ordersis not any different from displayed orders, then abolishing hidden orders maynot impact the price discovery process following earnings announcements.On the other hand, if hidden orders are associated with higher profitability,then abolishing hidden orders may deter the willingness of traders to provideliquidity and, in turn, deter price discovery.To measure the profitability of liquidity providers, I calculate for eachobserved trade j across all stocks with after-hours trading following an earn-ings announcement the realized spread measure, rsi,j , defined asrsi,j ={mi,j−pi,tmi,t−1 ∗ 100, if trade j was a passive buypi,t−mi,jmi,t−1 ∗ 100, if trade j was a passive sell,(2.12)where mi,t is the crossing price at the opening of markets if there was anauction or the midquote in the order book at 9:30 a.m if there was not.mi,t−1 is the closing crossing price prior to the announcement if there wasan auction or the midquote in the order book at 4 p.m. if there was not.39 I also winsorized the realized spreads at the 1st and 99th percentiles. Icalculate the realized spread for displayed and hidden orders separately.To examine the profitability of liquidity provision, I estimate the follow-ing ols regression:rsoi,k,t = β1Displayedi,k,t + β2Hiddeni,k,t + i,k,t. (2.13)rsoi,k,t corresponds to the average realized spread across all orders of typeo for stock i on earnings announcements of date t in trade bin k.40 Ordertype o is either displayed or hidden orders. Hiddeni,k,t is a dummy variableequal to one if the order type o represents hidden orders and zero otherwise.from July 14, 2014 market participant cannot infer from the message order P whether thetrade was an initiated market buy or sell order.39In the microstructure literature, calculation of the realized spread involves use of amidquote taken a few seconds or minutes after the trade but, as previously argued, onecannot use midquotes in the after-hours market. Choosing the opening price is thereforenot common but remains the best choice for a wide cross-sectional analysis of realizedspread in the after-hours market.40An alternative regression is a cross-section regression across all trades at differenttrade arrival bins. The inconvenience of this regression is that it gives more weight toearnings announcement events with a large number of trades.622.6. Hidden Liquidity around Earnings AnnouncementsSimilarly, Displayedi,k,t is a dummy variable equal to one if order type orepresents a displayed orders and zero otherwise.Table 2.8 Panel A shows the estimated coefficients estimate at differenttrade bins for earnings announcements with more than 20 trades and lessthan or equal to 20 trades.632.6. Hidden Liquidity around Earnings AnnouncementsTable 2.8: OLS Regression: Realized Spreads on Displayed and HiddenLimit OrdersThis table reports coefficients from regressions of realized spreads on adummy variable Hiddeni,k,t equal to one if the realized spread is for hiddenorders and zero otherwise, and a dummy variable Displayedi,k,t equal toone if the realized spread is for displayed orders and zero otherwise. Therealized spread is the average realized spread for each order type (hiddenor displayed) by earnings announcement dates and at each trade bin k foreach stock. The definition of a trade bin is described in the main text. Theregression is estimated for the first trade bin, for the second to the fifthtrade bins, and for the sixth to the tenth trade bins. High (low) trade an-nouncements correspond to earnings announcements with more than (lessor equal to) 20 trades in the after-hours market. The standard errors areclustered by date and reported in parentheses. Asterisks denote statisticalsignificance at the 5-percent level. The sample period is January 1, 2011 toJuly 13, 2014.High trade announcements Low trade announcementsk = 1 2 ≤ k < 5 k ≥ 5 k = 1 2 ≤ k < 5 k ≥ 5Hiddeni,k,t 0.23* 0.16* 0.08* 0.18* 0.24* 0.19*(0.04) (0.02) (0.02) (0.09) (0.06) (0.05)Displayedi,k,t -0.07 -0.06* -0.01 0.06 0.01 -0.01(0.04) (0.02) (0.02) (0.06) (0.04) (0.04)Obs. 13100 39826 80327 6262 13083 14578% displayed orders 66 66 66 74 71 70% hidden orders 34 34 34 26 29 30The results show that realized spreads for displayed orders are not sta-tistically different from zero at the five percent level, except in the secondcolumn for high trade firms where displayed orders earn a negative profit.On the other hand, realized spreads for hidden orders are all statisticallydifferent from zero at the five percent level and much larger than displayedorders. On average, the profit for a hidden order on a $50 stock is about 7.5cents for high trade announcements and 10 cents for low trade announce-ments across all trade bins.The positive profitability associated with hidden orders can be explained,in part, by the fact that adverse selection risk for displayed orders is high642.7. Conclusion to Chapter 2and hidden orders effectively mitigate this risk or that liquidity providersare at an informational advantage on future price drift following the news.Only future research with actual data on hidden order placement can ad-vance our knowledge as to why hidden orders are profitable. But, this resultis important to policy makers that wish to abolish hidden orders to increasemarket transparency; it may harm price discovery following earnings an-nouncements because some liquidity providers may only want to providehidden liquidity.2.7 Conclusion to Chapter 2This paper investigates how earnings surprises are incorporated into stockprices for the largest 1,500 U.S. stocks between 2011 and 2015. This occursdue to a two-stage adjustment process. First, prices adjust sharply and di-rectly to earnings surprises upon arrival of the first trades and more than80 percent of the share of after-hours price discovery occurring preciselyat this moment. Earnings surprises and not order flow largely explain thisinitial price adjustment. Second, after the initial adjustment, order flowimbalances explain the remaining price adjustment in the after-hours mar-ket. I find significant price discovery remaining at the opening of marketsfor stocks with no after-hours trading following earnings announcements.Around 10 a.m. following the opening of markets, earnings surprises haveno explanatory power to explain stock returns.I also find low abnormal volatility, low abnormal trade volume, and highabnormal quoted spread on the day prior to earnings announcements withlarge earnings surprises. This implies that markets anticipate the magnitudeof earnings surprises. The positive impact of large earnings surprises onthe adjustment process of price volatility, quoted spread, and trade volumefollowing earnings announcements is more gradual and persistent than theimpact of earnings surprises on prices.Last, I show that hidden orders are widely used following earnings an-nouncements and are more profitable than displayed orders for liquidityproviders. Hidden liquidity decreases market transparency but may, in fact,improve market efficiency following the arrival of news because liquidityproviders may be more inclined to supply liquidity with the use of hiddenorders.The findings of this paper shed light on existing theories on the roleof order flow and liquidity provision on price discovery but also proposenew avenues for future theoretical work. For instance, why is there an652.7. Conclusion to Chapter 2after-hours market? What are the economic determinants that explainswhy some investors trade in the after-hours market? Clearly, there is someheterogeneity among market participants, with some choosing to sit outof the active period of price formation when corporate announcements aremade outside of regular trading hours, and some staying or becoming active.66Chapter 3Shaping Expectations andCoordinating Attention: TheConsequence of FOMC PressConferences3.1 IntroductionThe Federal Open Market Committee (fomc), the monetary policy-makingbody of the U.S. Federal Reserve System (Fed), meets regularly to discussthe state of the economy and set monetary policy. Because asset prices reactstrongly to news about macroeconomic conditions, great care is given notjust to the decisions made, but also to how they are communicated to finan-cial markets after the meetings.41 While it was left to market participants toinfer decisions from the Fed’s open market operations prior to 1994, policydecisions are now announced in a press statement. In an effort to “provideadditional transparency and accountability” (Bernanke, 2011), since April2011 the fomc publishes economic projection materials and the Chair ofthe Board of Governors holds a press conference (pc) following half of theannouncements.Importantly, the decision to hold a pc does not depend onmacroeconomic conditions, as the schedule for both announcements and pcsis released at least six months in advance.In this paper, we study the economic consequences of having press con-41 A large literature documenting the response of asset prices to macroeconomic newsfor various asset classes includes Jones, Lamont, and Lumsdaine (1998b); Fleming andRemolona (1999b); Balduzzi, Elton, and Green (2001b); Andersen, Bollerslev, Diebold,and Vega (2003b, 2007b); Green (2004); ?, and Hu, Pan, and Wang (2015b). Cook andHahn (1989); Kuttner (2001); Bernanke and Kuttner (2005); Ozdagli and Weber (2015),and Bjornland and Leitemo (2009) focus their analysis on fomc announcements. Morebroadly, Savor and Wilson (2013a) and Lucca and Moench (2015a) find that investorsdemand a large premium for macroeconomic risks, and Savor and Wilson (2014) showthat this premium has important implications for asset pricing.673.1. Introductionferences following only some meetings. It is conceivable that the committeedefers important decisions for meetings when it has the opportunity to pro-vide explanations and context in a pc. The introduction of press conferencescould therefore lead to two classes of fomc announcements, with importantannouncements on days with pcs and lesser ones on days without. Such aseparation would reduce the frequency at which news about the economyand monetary policy is released to financial markets, and seriously questionwhether pcs increase transparency.In its official position, the Fed insists that all meetings and announce-ments, irrespective of press conferences, are equally important. For example,when asked if it is good “that the market expects big news to come when youhave a press conference and no news to come when you don’t have one,”Chairwoman Yellen replied that she “would really like to strongly discouragethe expectation that policy moves can only occur when there’s a scheduledpress conference” (Yellen, 2014). In a similar exchange nine months later,Chairwoman Yellen insists that “every meeting is a live meeting where theCommittee can make a decision to move to change our target for the federalfunds rate” (Yellen, 2015b).However, there is also reason to believe that pcs influence the timing ofimportant policy decisions. For example, in June 2015 Chairwoman Yellensuggested a first interest rate raise in “September [2015] or December [2015]or March [2016]” (Yellen, 2015a), three fomc meetings with scheduled pressconferences. The committee would also meet in July 2015, October 2015,and January 2016, each without press conference following the announce-ment of their decisions. When looking at actual policy decisions, we docu-ment that only two out of eight important monetary policy announcementsduring our sample period were made on days without pcs, which comprisenearly half of all announcement days. Moreover, just after our sample ends,the first interest rate increase following the financial crisis was announcedin December 2015, a day with pc.Of course, it is difficult to objectively quantify the gravity of the Fed’sdecisions, and the small number of important policy changes prohibits adetailed statistical analysis. We therefore instead focus our analysis on thebeliefs and behavior of market participants and rely on financial marketsto gauge the expectations of significant monetary policy decisions. Usingevidence spanning multiple asset classes, we document striking differencesin both markets’ expectations of and reactions to fomc announcements withand without pcs. We first show that average returns of the S&P 500 in the30 minutes after the fomc announcement are large and positive on dayswith pcs, averaging 0.29%. This estimate is statistically significant and683.1. Introductionrobust to outliers and bootstrapped small-sample statistics. In contrast,announcement returns are on average negative on days without pcs. Thedifference in announcement returns between pc and non-pc days is largeand significant at 0.57%, and remains robust to controlling for inflation andchanges to the unemployment rate, the two variables the fomc is mandatedto manage, as well as growth of gross domestic product (gdp) and pastmarket returns.We argue that this ex-post reaction to fomc announcements can beused to proxy for the ex-ante market expectation of the Fed’s decisions.The reasoning relies on the observation that throughout our sample similarinformation was revealed at both types of announcements. In particular, theFederal funds target rate, one of the main drivers of equity prices in fomcannouncements, remained unchanged at 0 to 0.25%. Since 2011, the fomchas therefore repeatedly surprised markets positively, with the magnitudeof the surprise directly proportional to ex-ante expectations of target rateincreases.42 The large market returns following announcements with pcsthen correspond to large ex-ante market expectations of rate increases.Two aspects about our analysis are important to emphasize. First, weanalyze announcement returns conditional on press conferences taking place,but the returns we study do not include information revealed during thepress conferences. Second, these findings are about the market reaction tofomc announcements. They are not returns in anticipation of announce-ments, as in Lucca and Moench (2015a), nor do they necessarily presentprofitable trading opportunities. Interestingly, in our more recent samplewe confirm a pre-fomc announcement return of similar magnitude as Luccaand Moench (2015a), but only on days with pcs. In contrast, if there is nopress conference, average market returns leading up to the announcementare zero.Stock price reactions to fomc announcements are only an indirect mea-sure of ex-ante expectations of changes to monetary policy. To overcomethis limitation, we directly measure expectations of target rate changes im-plied by Federal Fund Futures. On days with pcs, the probability of a ratechange is on average 2.8 percentage points, or a staggering 76%, higher thanon days without. The differential market assessments about probabilities of42Target rate announcements are of first-order importance for equity prices (Kuttner,2001). For example, Bernanke and Kuttner (2005) and Ozdagli and Weber (2015) estimatethat a surprise decrease in the Federal funds rate of 0.25% increases stock prices by 1%,whereas the analysis in Bjornland and Leitemo (2009) suggests an even bigger impact.Gu¨rkaynak, Sack, and Swanson (2005) confirm that rate announcements are important,but argue that the future path of policy also plays a role.693.1. Introductioninterest rate changes are not limited to the nearest fomc announcement;rather, they persist for at least three years into the future. This confirmsthat markets expect more important decisions on days with press confer-ences.We next investigate the effects of press conferences for monetary policy.In particular, we ask if, consistent with market expectations, the Fed makesmore important announcements on days with pcs. To answer this question,we use the option-implied volatility of the S&P 500, as measured by the vixindex, to proxy for uncertainty associated with monetary policy. Consistentwith findings in Beber and Brandt (2009), Savor and Wilson (2013a), andAmengual and Xiu (2015), the vix drops sharply by 2% on average at fomcannouncements, suggesting that the Fed provides valuable information toreduce uncertainty about the economy or monetary policy. Investigating theimpact of press conferences, we find that all of this decline comes on dayswhen a pc is scheduled, where the vix drops by over 4%. In contrast, on dayswithout pcs, the vix remains virtually unchanged after the announcement,and monetary policy uncertainty is not reduced.Taken together, our findings suggest that expectations of relevant changesto monetary policy are lower on fomc announcement days without pcs, andthat the fomc reveals less price-relevant information to markets on thosedays. In other words, the introduction of pcs separated fomc announce-ments into important and lesser ones.A possible concern regarding these conclusions is that really the upcom-ing release of the economic projection materials (epms), and not the sched-uled pcs, are responsible for the heightened market expectations. Althoughthe individual effects are difficult to disentangle as both events always occuron the same days, we employ a change in the timing of the release of theepms to show that they generally contain little information and are thereforeunable to explain our findings. Crucially, even if the specific channel thatseparates fomc announcements was the release of epms, our main findingsand conclusion would not be affected. The question would then become whyepms are not released at all meetings.What economic channels link press conferences with market expectationsand monetary policy decisions? Of course, if the Fed intended to makeimportant monetary policy announcements only on days with pcs, and ifmarkets understand this despite the Fed’s denial, we would expect to observeboth more important announcements and high market expectations on dayswith pcs. However, we argue that it is also possible that lowered marketexpectations on days without pcs impose constraints on the Fed through tworelated, but distinct, channels. First, if markets do not expect significant703.1. Introductionpolicy decisions, any announcement of such would therefore be a surprise.However, the Fed is frequently believed to be averse to surprising markets.43Market expectations can therefore become self-fulfilling, and this tensionalso increases the Fed’s incentives for the kind of informal communicationstudied in Cieslak, Morse, and Vissing-Jorgensen (2015a).Second, if investors now believe some meetings to be more importantthan others, it would be natural that they allocate more of their limitedattention to these meetings.44 But it has long been recognized that in-vestor attention and market expectations are critical to the transmission ofmonetary policy (Stein, 1989; Blinder, Goodhart, Hildebrand, Lipton, andWyplosz, 2001), and that therefore “monetary policy is more effective if itis more effective in coordinating market expectations” (Amato, Morris, andShin, 2002, p.496).45 Clearly, if investors pay less attention to its communi-cation, the Fed cannot effectively coordinate market expectations and mightfind it optimal to delay important announcements.We confirm that pcs indeed influence investors’ allocation of attention tofomc announcements. In particular, we show that media coverage of andinterest in the fomc is significantly higher prior to announcements withpcs than those without. The effect is large and holds both for measurestypically associated with attention of institutional and retail investors. Tocapture attention of institutional investors, we follow Ben-Rephael, Da, andIsraelsen (2016) and use Bloomberg articles and intraday newswires. Simplyallowing conditional means to vary between pc and non-pc events explains43For example, Stein and Sunderam (2015) model a central bank that is averse to bond-market volatility. See also Cieslak, Morse, and Vissing-Jorgensen (2015a) for a detaileddiscussion. In the press, a survey by the Wall Street Journal “underscores just how muchwork it would take for the Fed to create expectations of a rate increase at a meetingwithout a news conference” (Zumbrun, 2015).44Press conferences can therefore serve as a coordination device when investors have lim-ited capacity for processing information (Sims, 2003). Duffie and Sun (1990), Abel, Eberly,and Panageas (2007, 2013), and Huang and Liu (2007) show that investors optimally re-main inattentive to some information if they face information acquisition costs. Similarly,in Kacperczyk, van Nieuwerburgh, and Veldkamp (2016), investors allocate scarce atten-tion between different kinds of information and optimally focus on information about moreuncertain outcomes, i.e., information that has the largest impact on prices. In both typesof models, with indistinguishable fomc announcements investors will pay equal attentionto each. However, pcs designate some events to be more important than others, and theycoordinate investors to pay more attention to fomc announcements with pcs.45Highlighting this importance further, Blinder (1998, p.70) states: “central banks gen-erally control only the overnight interest rate, an interest rate that is relevant to virtuallyno economically interesting transactions. Monetary policy has important macroeconomiceffects only to the extent that it moves financial market prices that really matter – likelong-term interest rates, stock market values and exchange rates.”713.1. Introductionup to 39% of the variation in these media attention measures. Building onFang and Peress (2009a) and Da, Engelberg, and Gao (2011a), our proxiesfor retail investor attention are lower-frequency measures based on articlesin the print editions of major newspapers, and a broader attention measurebased on Google search volume in the week prior to fomc announcements.Just like institutional investors, retail investors also focus their attentionaround fomc announcement days with press conferences. We find a similarpattern for the Bank of Canada and the Reserve Bank of New Zealand, thetwo other central banks that follow communication policies similar to theone of the fomc.The elevated investor attention leading up to fomc announcements withpress conferences could reflect an increased interest to the announcement,or novel attention paid to the actual press conference. To answer this ques-tion, we show that pcs convey little new information to markets. Whilethe realized volatility of equities during the pc is elevated, it is not signif-icantly higher than during the corresponding time interval following fomcannouncements without pc. Further, virtually no changes in option impliedvolatility indicate that pcs do not reduce uncertainty.46 Given this evidence,it seems unlikely that press conferences themselves command the additionalattention; rather, markets pay more attention because they expect moreimportant fomc announcements.To answer whether the separation into important and lesser fomc an-nouncements might have been the Fed’s intention, or an unintended con-sequence of having press conferences, we show that most of our findingssignificantly strengthen throughout our sample. While the increasing role ofpcs on market expectations and investor attention could reflect slow learn-ing of investors about a possible new regime, we also find that the amount ofinformation released at fomc announcements with pcs increases over time.This slow trend suggests that the Fed did not initially choose to designatefomc meetings with pcs as more important than those without, but insteadis reacting to changes in market expectations and investor attention.Press conferences were introduced with the goal to increase transparency.Our analysis raises strong doubts about whether this goal is achieved. Aswe show, pcs convey little new information to markets. At the same time,our evidence suggests that the reduced information revealed at non-pc an-nouncements decreases transparency at these intermediate times. Taken46These tests measure information content only by the reaction of equity and optionmarkets. Information that does not immediately affect market prices, either because it isnot price relevant or takes longer to process, could of course still be revealed during pressconferences.723.2. The Federal Open Market Committeetogether, overall transparency probably decreased as a result of having pcsafter only some fomc announcements.The implications of this new fomc communication policy are difficultto gauge. While transparency is frequently viewed as positive, it is lessclear whether increased transparency really results in lower price volatilityor in prices that better reflect fundamental values. See, for example, Amato,Morris, and Shin (2002) and Banerjee, Davis, and Gondhi (2015).Taken to the extreme, our evidence raises the question why the fomcmeets and makes policy announcements on days without scheduled pressconferences. If the objective of the fomc is to increase transparency whilesimultaneously limiting market surprises and maintaining flexibility of ac-tion, it should consider following the practice of holding press conferencesafter every meeting, as adopted by the European Central Bank, the Bankof Japan, Sweden’s Riksbank and Norway’s Norges Bank.3.2 The Federal Open Market CommitteeThe fomc is the monetary policy-making body of the U.S. Federal ReserveSystem. It oversees the nation’s open market operations, i.e., purchases andsales of U.S. Treasury and Federal Agency Securities, which affect the costand availability of money and credit in the economy, under the statutorydual mandate of maximum employment and stable prices. The fomc iscomposed of the seven members of the Board of Governors and five of thetwelve Reserve Bank presidents. While the president of the Federal ReserveBank of New York serves on a continuous basis, the presidents of the otherReserve Banks serve one-year terms on a rotating basis. By law, the fomcmust meet at least four times a year. Since 1981, however, eight regularlyscheduled meetings have been held each year at intervals of five to eightweeks. Members may also be called on to participate in special meetingsif circumstances require consultation or consideration of an action betweenthese regular meetings. Prior to 1994, changes to the Federal funds rate werenot announced and market participants had to infer them by observing thesize and type of open market operations. In 1994, the fomc began announc-ing their policy decisions in a press statement, with the announcement datesand times released to the public in June of the previous year.Since April 2011, the Chair of the Board of Governors holds a press con-ference following half of the fomc announcements. Importantly, just likethe announcements themselves, pcs are scheduled at least six months in ad-vance, and the decision to hold a pc therefore does not depend on economic733.2. The Federal Open Market Committeeor market conditions.47 Press conferences last on average close to one hourand consist of an opening statement by the Chair of the Board of Gover-nors followed by a question and answer session with financial journalists.Between April 2011 and January 2013, fomc announcements with pcs werescheduled for 12:30 p.m., followed by pcs beginning at 2:15 p.m. Announce-ments without pcs were scheduled for 2:15 p.m. Since March 2013, fomcannouncements always occur at 2:00 p.m., and press conferences begin at2:30 p.m..Table 3.1 provides an overview of the fomc announcements, their sched-uled times, and the starting time of the associated press conferences. Thetable also reports the actual announcement times, obtained from ThomsonReuters Tick History (trth) as supplied by the Securities Industry ResearchCentre of Asia-Pacific (sirca). In total, our sample is comprised of 37 an-nouncements, 19 with and 18 without press conferences. After some initialirregularities, since June 2012 press conferences now follow every other fomcannouncement.47The schedule for a year is released in June of the previous year. The new communi-cation policy was first announced on March 24, 2011, five weeks before the first meetingwith a press conference.743.2. The Federal Open Market CommitteeTable 3.1: FOMC Announcement CalendarThis table shows the scheduled (Sched.) and actual (Act.) time of fomcannouncements and the scheduled time for press conferences (pcs) betweenApril 2011 and October 2015.Source: http://www.federalreserve.gov/monetarypolicy/fomccalendars.htmand trth.Date Sche. Act. pc Date Sched. Act. pc04/27/2011 12:30 12:32 14:15 09/18/2013 14:00 14:00 14:3006/22/2011 12:30 12:27 14:15 10/30/2013 14:00 14:0008/09/2011 14:15 14:18 12/18/2013 14:00 14:00 14:3009/21/2011 14:15 14:23 01/29/2014 14:00 14:0011/02/2011 12:30 12:32 14:15 03/19/2014 14:00 14:00 14:3012/13/2011 14:15 14:12 04/30/2014 14:00 14:0001/25/2012 12:30 12:27 14:15 06/18/2014 14:00 14:00 14:3003/13/2012 14:15 14:15 07/30/2014 14:00 14:0004/25/2012 12:30 12:32 14:15 09/17/2014 14:00 14:00 14:3006/20/2012 12:30 12:32 14:15 10/29/2014 14:00 14:0008/01/2012 14:15 14:13 12/17/2014 14:00 14:00 14:3009/13/2012 12:30 12:31 14:15 01/28/2015 14:00 14:0010/24/2012 14:15 14:15 03/18/2015 14:00 14:00 14:3012/12/2012 12:30 12:30 14:15 04/29/2015 14:00 14:0001/30/2013 14:15 14:15 06/17/2015 14:00 14:00 14:3003/20/2013 14:00 14:00 14:30 07/29/2015 14:00 14:0005/01/2013 14:00 14:00 09/17/2015 14:00 14:00 14:3006/19/2013 14:00 14:00 14:30 10/28/2015 14:00 14:0007/31/2013 14:00 14:00753.3. Financial Markets around FOMC AnnouncementsOne challenge that arises in studying fomc press conferences is thatthe number of events is quite limited. We address this issue in three ways.First, we provide bootstrapped standard errors and p-values for all our sta-tistical tests. Second, we analyze the effect of outliers on the distributionof announcement returns. Third, we provide a test that uses informationfrom futures market to estimate the effect of pcs going forward, effectivelyextending our sample by three years.Throughout our sample, the Federal funds target range remained con-stant at 0 to 0.25%. Nevertheless, our sample contains some changes inmonetary policy by means of quantitative easing (qe) to help revive theU.S. economy following the financial crisis. We now list some of the keyfomc announcements since 2011.June 22, 2011 (pc): the Fed announces the end of qe2.September 21, 2011 (no pc): the Fed announces Operation Twist,which consists of purchasing $400 billion of Treasuries with long ma-turities and selling an equal amount with shorter-term maturities.June 20, 2012 (pc): the Fed announces that it will continue OperationTwist.September 13, 2012 (pc): the Fed announces qe3.December 12, 2012 (pc): the Fed announces the expansion of qe3.June 19, 2013 (pc): During the pc, Chairman Bernanke suggests agradual moderation of the pace of bond purchases could begin in themonths to come.48September 18, 2013 (pc): the Fed decides to hold off on “tapering”.October 29, 2014 (no pc): the Fed announces the halt of bond pur-chases.3.3 Financial Markets around FOMCAnnouncementsIn this section, we investigate whether the schedule of press conferences af-fects financial markets and has any consequences for the Fed and monetary48Equity and fixed income markets reacted strongly to this information. Interestingly,on May 22, 2013, one month before this press conference, Chairman Bernanke made astatement using similar language in a testimony to Congress.763.3. Financial Markets around FOMC Announcementspolicy. Rather than attempting to measure the gravity of monetary policydecisions, we use evidence from equity and derivative markets to show thatthe introduction of pcs has significantly affected market behavior aroundfomc announcements. First, pcs influence the perceived importance offomc announcements as only events with pcs are associated with large ex-pectations of important monetary policy decisions. Second, on days withoutpcs, fomc announcements do not resolve uncertainty about monetary pol-icy, suggesting that the news is viewed as less momentous. Lastly, we showthat the pre-fomc announcement drift, a robustly positive stock marketreturn prior to fomc announcements documented by Lucca and Moench(2015a), prevails in our sample, but is limited to announcement days withpress conferences.3.3.1 Press Conferences and Market ExpectationsWe use two measures of stock market expectations of changes in monetarypolicy. First, we argue that ex-post stock market announcement returnsproxy for ex-ante expectations if the total information content in announce-ments, expected and unexpected, is constant throughout the sample. Sec-ond, we obtain a more direct measure of true ex-ante implied probabilitiesof target rate changes from Federal Funds Futures.Stock Market Announcement ReturnsWe begin our analysis by showing that stock market reactions to fomcannouncements differ across days with and without press conferences. Ifmarkets are efficient, these returns measure the unexpected component ofthe announcements. We argue that, specific to our sample, these surprisescan also be used to proxy for the expected part of the announcements.Our identification relies on the observation that there is little variation inthe total information content, expected and unexpected, of announcementsin our sample. In particular, the Federal funds target rate, the single mostclosely watched number associated with fomc announcements, has remainedat its lower bound of 0 to 0.25%. Any decisions regarding this rate cantherefore be thought of as binary: rates can either remain unchanged orincrease.4949In practice, unconventional monetary policies such as large-scale asset purchases canbe used to effectively overcome the zero lower bound (Swanson, 2015). On the other end,target rates could increase by more than 0.25%.773.3. Financial Markets around FOMC AnnouncementsSince unexpected rate increases typically lead to a drop in equity prices(Kuttner, 2001; Bernanke and Kuttner, 2005), in this scenario prices shouldrise when the Fed announces that rates remain low. The magnitude of therise, however, depends on the markets’ ex-ante expectations that rates wouldincrease. For example, if markets are certain that rates will not change, anannouncement of no increase should not affect prices. If on the other handmarkets have a large expectation of a rate increase, any announcement ofconstant rates should be considered a large positive surprise, and stock pricesshould therefore increase significantly.We focus on the liquid and arguably mostly efficient equity market, inparticular the shortest maturity E-mini S&P 500 Futures (e-mini), obtainedfrom trth. We define the e-mini price as the midpoint of the best outstand-ing bid and ask quotes, and convert this time-series of prices into one-secondmidquote returns. We further restrict our sample to regular equity marketstrading hours, i.e., 9:30 a.m. to 4:00 p.m. est.Figure 3.1 plots the average cumulative e-mini return around fomc an-nouncements, starting 2.5 hours before and ending 1.5 hours after the an-nouncement. The time interval is chosen to avoid potential effects fromovernight returns. As shown in Table 3.1, prior to 2013 announcementswith pcs were made no earlier than 12:27 p.m., or 2 hours and 57 minutesafter market open, and between August 2011 and January 2013 announce-ments without pcs were made no later than 2:23 p.m., or 1 hour and 37minutes before market close. Returns are normalized to zero at the time ofthe announcement.Panel A groups all fomc announcements from April 2011 to October2015. Consistent with the conjecture that fomc announcements through-out our sample contained good news for equity markets, there is a smallreturn of around 0.10% in the hour after the announcement. The 95% con-fidence interval, plotted in gray, suggests that this effect is not statisticallysignificant.A striking pattern emerges in Panel B, where we separate fomc an-nouncements into ones with and without press conferences. When thereis a pc (blue solid line), prices increase by an economically large and sta-tistically significant 0.40% after the announcement. In contrast, fomc an-nouncements without pcs (red dashed line) are accompanied by a drop inprices of about 0.20% during a volatile period following the announcement.In Table 3.2, we formally test the main insights from Figure 3.1. Thetable provides estimates of moments and associated statistical tests of an-nouncement returns, which we define as the cumulative e-mini return inthe 31-minute event window starting one minute before the announcement.783.3. Financial Markets around FOMC AnnouncementsTable 3.2: FOMC Announcement ReturnsThis table reports selected moments and percentiles of log returns of theshortest maturity S&P 500 E-mini Futures, in %, over the 31-minute inter-vals starting one minute before fomc announcements. Values are reportedfor the whole sample, as well as samples split into days with and withoutpress conferences (pcs) and their difference. Asymptotic and bootstrappedstandard errors are presented in parenthesis and square brackets, respec-tively, and bootstrapped p-values in italics. N denotes the number of obser-vations. Panel A is based on the whole sample, while Panel B repeats theanalysis on trimmed samples that omit the smallest and largest observation.The sample period is April 2011 to October 2015.Panel A: Full Sample Panel B: Trimmed SampleAll pc No pc Diff. All pc No pc Diff.Mean 0.015 0.292 -0.277 0.569 0.051 0.295 -0.181 0.476Std. Error (asympt.) (0.10) (0.11) (0.15) (0.19) (0.07) (0.09) (0.08) (0.12)Std. Error (bootstr.) [0.10] [0.11] [0.15] [0.18] [0.07] [0.09] [0.08] [0.12]p-value (bootstr.) 0.92 0.01 0.08 0.00 0.48 0.00 0.03 0.00Std. Deviation 0.623 0.478 0.636 0.432 0.371 0.327Minimum -2.450 -0.711 -2.450 -0.959 -0.267 -0.95925th Percentile -0.182 0.081 -0.369 -0.169 0.090 -0.350Median 0.080 0.274 -0.104 0.080 0.274 -0.10475th Percentile 0.293 0.502 0.076 0.283 0.453 0.069Maximum 1.238 1.238 0.356 1.052 1.052 0.221Proportion <0 0.405 0.211 0.611 0.400 0.176 0.625N 37 19 18 35 17 16We begin our announcement window one minute prior to the event to en-sure that our findings are not affected by either information leakage beforethe announcement or possible data errors with regard to the exact fomcannouncement time. The choice of end time follows Ozdagli and Weber(2015), and further ensures that announcement returns are not affected byinformation released during the press conferences.793.3. Financial Markets around FOMC AnnouncementsFigure 3.1: Cumulative E-mini Return around FOMC AnnouncementsThis figure shows the average cumulative log return, in %, of the shortestmaturity S&P 500 e-mini futures around fomc announcements. Returnsare normalized to zero at the time of the announcement. Panel A showsresults for the whole sample, while Panel B separates announcements intothose with press conferences (blue solid line) and those without (red dashedline). The shaded areas are pointwise 95% confidence bands around theaverage returns. The sample period is April 2011 to October 2015.-2.5hrs 0Announcement1.5hrs−0.3−0.2−0.10.00.10.20.30.4CumulativeE-minireturnPanel A: All announcements-2.5hrs 0Announcement1.5hrs−0.8−0.6−0.4−0.20.00.20.40.60.8CumulativeE-minireturnPanel B: Announcements with and without press conferencesWith press conferencesWithout press conferences803.3. Financial Markets around FOMC AnnouncementsThe full sample results in Panel A show an average announcement returnof 0.02%. On days with pcs, this figure rises to 0.29%, while it is -0.28% ondays without. Based on the asymptotic distribution, the mean return forall announcements is insignificant. Announcement returns on days with pcsare both significantly positive and significantly larger than those on dayswithout. Returns on days with pcs range from -0.71% to 1.24%, with only4 out of 19 observations (21%) negative.Our evidence is based on a rather small sample containing only 19 (18)observations for pc (non-pc) events. We address concerns about the sampledistribution of the test statistic and the effect of possible outliers in twoways. First, we also provide bootstrapped standard errors (in brackets)and p-values (in italics). All bootstrapped results are based on 1,000,000samples. The bootstrapped standard errors closely resemble the asymptoticones, and the p-values confirm the previous findings.Second, to investigate the potential impact of outliers, Panel B repeatsthe analysis on a trimmed sample that excludes both the largest and small-est announcement return observations in each group. Point estimates for themeans are, with one exception, little affected. Only on non-pc days, aver-age returns rise from -0.27% to -0.18%, the minimum increases from -2.45%to -0.96%, and the standard deviation declines from 0.64% to 0.33%. Thisimplies that the sample was affected by one very large negative observation.Crucially, even in the trimmed sample, the statistical inference remains un-changed. Announcement returns on days with pcs are significantly positive,and larger than those on days without.We test whether the announcement return differences between pc andnon-pc days can be explained by different economic environments in Table3.3. The first two specifications regress announcement returns on indicatorvariables for pc and non-pc days. These two tests confirm the results fromTable 3.2 under the additional assumptions ordinary least square regressionsimpose on the error distribution. Just allowing for differences in averagesbetween pc and non-pc days explains 19% of the variation in announcementreturns.In the third specification, we add monthly log changes in seasonally ad-justed consumer price index (∆CPI) and unemployment (∆UE) to controlfor the economic environment. These variables are the most natural can-didates to influence expected monetary policy, as they correspond to thefomc’s target measures under its dual mandate. Data are obtained fromthe U.S. Bureau of Labor Statistics, and we always use the most recentlyannounced data. We complement these with gdp growth (∆GDP ) from theU.S. Bureau of Economic Analysis. In the fourth specification, we further813.3. Financial Markets around FOMC AnnouncementsTable 3.3: FOMC Announcement Returns: RegressionsThis table reports coefficients from regressions of fomc announcement re-turns on a press-conference indicator PC, equal to one if a meeting is fol-lowed by a press conference and zero otherwise, non-PC = 1 − PC, andcontrol variables. Announcement returns are the log returns of the shortestmaturity S&P 500 E-mini Futures, in %, over the 31-minute intervals start-ing one minute before fomc announcements. ∆CPI, ∆UE, and ∆GDP arelog changes in, respectively, the consumer price index, the unemploymentrate, and the gross domestic product. RS&P is the S&P 500 log return overthe 21-day interval ending 3 days before the announcement. Asymptoticheteroscedasticity robust and bootstrapped standard errors are presented inparenthesis and square brackets, respectively, and bootstrapped p-values initalics. Adjusted R2 and the number of observations N are also reported.The sample period is April 2011 to October 2015.Announcement Returns(1) (2) (3) (4)Intercept -0.277 -0.292 -0.391(0.15) (0.23) (0.24)[0.13] [0.19] [0.18]0.01 0.10 0.02PC 0.292 0.569 0.589 0.592(0.11) (0.18) (0.17) (0.16)[0.12] [0.18] [0.18] [0.16]0.03 0.00 0.00 0.00non-PC -0.277(0.15)[0.13]0.01∆CPI -0.185 -0.235(0.36) (0.33)[0.37] [0.34]0.57 0.46∆UE -0.814 -1.546(0.57) (0.77)[0.78] [0.77]0.28 0.05∆GDP -0.019 -0.025(0.05) (0.05)[0.06] [0.06]0.75 0.66RS&P 0.058(0.05)[0.02]0.01Adjusted R2 0.192 0.192 0.146 0.253N 37 37 37 37823.3. Financial Markets around FOMC Announcementscontrol for the cumulative log return of the S&P 500 Total Return Indexover the 21 trading days ending three days before the event, RS&P, fromtrth. The specific window is chosen to avoid overlap with both the currentand the previous fomc meetings.Of the control variables, changes in the unemployment rate are signifi-cantly negatively and the prior 21-day S&P 500 returns significantly posi-tively related to announcement returns. The signs are consistent with ourinterpretation of the dependent variable. Following improvements in thestate of the economy, such as a decrease in the unemployment rate or risingstock prices, markets increase their expectation of a tightening in monetarypolicy. Announcements to keep policy unchanged therefore result in largepositive surprises. Importantly, none of the control variables have any im-pact on the coefficient on the pc indicator. The marginal impact of pcson announcement returns is very stable across specifications, ranging from0.57% to 0.59%, and highly statistically significant.Ex-Ante Implied Probabilities of Target Rate ChangesAnnouncement returns are ex-post measures that might be affected by thecontent of the announcement, and might be a noisy measure of ex-ante ex-pectations if the total information content of announcements varies through-out the sample. We now validate these results using a pure ex-ante measurefrom derivative markets that directly captures the expected gravity of fomcannouncements.We measure the ex-ante probabilities of target rate changes using Fed-eral Funds Futures (ff), for which we obtain settlement prices from trth.These contracts are listed for the first 36 calendar months and derive theirprice from the realized Federal funds overnight rate. Specifically, the set-tlement price is 100 minus the average daily transaction-volume-weightedFederal funds overnight rate for the delivery month. Futures prices thusreflect market expectations of the average daily Federal funds effective rate(FFER), which is published by the Federal Reserve Bank of New York eachday.To extract probabilities of rate movements from ff prices, we follow themethodology used by the CME Group.50 The expected target rate change50Alternatively, it is possible to use these contracts to estimate the announcement sur-prise following Kuttner (2001). However, to obtain surprises and therefore expectations,Kuttner’s approach requires the use of ff prices from the end of the announcement day.This is not suitable for our purposes, as the end-of-day prices contain information re-vealed during the press conference. For more details on the construction of probabilities833.3. Financial Markets around FOMC Announcementsin month m is computed asE(∆rm) = F̂FERm − F̂FERm−1, (3.1)where F̂FERm is the futures-implied FFER at the end of month m. It isimportant to note that these expected target rate changes can be negativeeven though the Federal funds target rate is at its zero lower bound. This isbecause rates are targeted to stay within an interval, in our case 0 to 0.25%,rather than at a specific number, whereas the ff settlement price is basedon realized market rates.To convert expected rate changes to probabilities, we assume that targetrates can only change by 0.25% at any given meeting and computeP (l) = |E(∆rm)| /0.25, (3.2)P (↑) = max [E(∆rm), 0] /0.25. (3.3)The calculation of F̂FERm depends on whether there is another fomcmeeting scheduled in month m+ 1. If there is, we estimateF̂FERm−1 = 100− FFm−1 (3.4)F̂FERm =1N −M [N(100− FFm)−M(100− FFm−1)] (3.5)where FFm is the price of the future expiring in month m, N is the numberof calendar days in month m, and M is the calendar day of the FOMCmeeting minus 1. If there is no meeting scheduled in the following month,we instead estimateF̂FERm−1 =1M[N(100− FFm)− (N −M)(100− FFm+1)] (3.6)F̂FERm = 100− FFm+1. (3.7)To test whether press conferences affect the probability of rate changes,we obtain for each fomcmeeting the ff implied probability computed on theprevious day. We then regress meeting-to-meeting changes in the ff impliedprobability onto changes in an indicator variable for press conferences andcontrol variables.Our findings are summarized in Table 3.4. In the first three columns,the dependent variable is based on the probability of changes in interestof rate movements, see http://www.cmegroup.com/trading/interest-rates/countdown-to-fomc.html.843.3. Financial Markets around FOMC Announcementsrates. The first specification only contains an intercept and changes in thepc indicator variable, ∆PC, which can take one of three values: one if theannouncement has a pc while the previous did not, minus one for the oppo-site case, and zero if both the current and prior announcement were followedby or not followed by pcs. It shows that, on average, the probability of ratechanges is 2.8 percentage points higher on days with press conferences thanon those without. The estimate is statistically significant and economicallylarge compared to the sample average of the probability of rate changes of5.1%. The estimate thus suggests that meetings with press conferences areassociated with a (5.1+2.8/2)/(5.1−2.8/2)−1 = 76% increased probabilityof a change in target rates relative to those without.When adding control variables in specifications (2) and (3), the coeffi-cient on ∆PC is unaffected, and press conferences remain associated witha higher probability of rate changes. Of the control variables, only the pastS&P 500 return is significant. The negative coefficient suggests that changesin market prices reflect the altered probabilities of interest rate changes.Since the target rate has been at its zero lower bound throughout oursample, we also perform the tests on the narrower probability of targetrate increases. The results, shown in columns (4)-(6) of Table 3.4, confirmthe previous findings. On days with press conferences, the probability ofa rate increase is 3.3 percentage points higher than on days without pressconferences. Relative to the unconditional average of a rate increase of 3.3%in our sample, this corresponds to a three-fold increase in probability onpress conference days relative to non-pc days.Federal Funds Futures are listed for the next 36 calendar months, pro-viding a rich source of information regarding long term expectations of ratechanges. To investigate the effects of press conferences on the term structureof market expectations, we first compute the probability of a rate changefor each FOMC meeting from 2011 to 2016 using ff settlement prices at theend of the first trading day of each calendar year. Results are presented inFigure 3.2, where full circles identify meetings with pcs while hollow dotsidentify those without.In the plot for 2011, the probabilities of rate changes are smoothly in-creasing over the next eight fomc announcements. The plot is based ondata from January 3, 2011, and the introduction of press conferences hadnot yet been announced. Therefore, not surprisingly, press conferences donot affect the probabilities. In the following years, we see a clear separationbetween meetings with pcs and those without. Probabilities of interest ratechanges are consistently higher for meetings associated with pcs.Next, we formally test the main insights from Figure 3.2. For this test, we853.3. Financial Markets around FOMC AnnouncementsTable 3.4: Probability of Interest Rate Changes before FOMC Announce-mentsThis table reports coefficients from regressions of meeting-to-meetingchanges in the probability of interest rate changes, in %, on changes ∆PCof an indicator variable equal to one if a meeting is followed by a press con-ference and zero otherwise, and control variables. Probabilities of changes,∆P (l), or increases, ∆P (↑), in Federal funds rates are derived from FederalFunds Futures as measured one day prior to each fomc meeting. ∆CPI,∆UE, and ∆GDP are log changes in, respectively, the consumer price in-dex, the unemployment rate, and the gross domestic product. RS&P is theS&P 500 log return over the 21-day interval ending 3 days before the an-nouncement. Asymptotic heteroscedasticity robust and bootstrapped stan-dard errors are presented in parenthesis and square brackets, respectively,and bootstrapped p-values in italics. Adjusted R2 and the number of ob-servations N are also reported. The sample period is April 2011 to October2015. Detailed information on the construction of implied probability mea-sures is provided in the text.∆P (l) ∆P (↑)(1) (2) (3) (4) (5) (6)Intercept 0.078 0.817 2.439 0.090 0.337 1.427(1.15) (1.48) (1.52) (1.04) (1.16) (1.12)[1.16] [2.26] [2.01] [1.04] [2.02] [1.92]0.95 0.72 0.22 0.94 0.87 0.45∆PC 2.825 2.795 2.926 3.262 3.204 3.292(1.25) (1.27) (1.09) (1.10) (1.11) (1.04)[1.21] [1.21] [1.04] [1.08] [1.08] [1.00]0.02 0.02 0.00 0.00 0.00 0.00∆CPI -2.209 -1.553 -0.436 0.004(4.48) (4.24) (3.89) (3.81)[5.01] [4.33] [4.48] [4.15]0.66 0.73 0.92 0.99∆UE 3.211 14.911 6.180 14.039(10.37) (7.68) (9.00) (6.81)[10.26] [9.48] [9.17] [9.08]0.75 0.12 0.50 0.12∆GDP -0.098 -0.017 0.159 0.213(0.82) (0.66) (0.77) (0.67)[0.86] [0.74] [0.77] [0.71]0.91 0.98 0.84 0.77RS&P -0.949 -0.638(0.29) (0.28)[0.27] [0.26]0.00 0.02Adjusted R2 0.106 0.029 0.250 0.177 0.110 0.211N 36 36 36 36 36 36 863.3. Financial Markets around FOMC AnnouncementsFigure 3.2: Term Structure of the Probability of Target Rate ChangesThis figure shows the implied probability of an interest rate change at eachof the eight annual fomc meetings. Implied probabilities are computedfrom settlement prices of Federal Fund Futures on the first trading day ofeach calendar year. Full circles identify meetings followed by press confer-ences while hollow dots identify those without. Detailed information on theconstruction of probability measures is provided in the text.1-20113-20114-20116-20118-20119-201111-201112-201105101520Probability(%)20111-20123-20124-20126-20128-20129-201210-201212-20120123456720121-20133-20135-20136-20137-20139-201310-201312-20130123456789Probability(%)20131-20143-20144-20146-20147-20149-201410-201412-20142345678920141-20153-20154-20156-20157-20159-201510-201512-20150102030405060Probability(%)20151-20163-20164-20166-20167-20169-201611-201612-201601020304050602016look at meetings after June 2012, when the regular pattern of quarterly pcswas announced.51 Using settlement prices from after the announcement on51The regular pattern allows investors to forecast dates of future press conferences.While the calendar of fomc meetings is released in June of the previous year, the approx-imate dates are generally predictable from past meetings. For this test, we assume thatparticipants knew the true meeting dates going forward, using the actual fomc calendarup to 2017. We supplement this with the following expected meetings dates for 2018:January 31, March 14 (pc), May 2, June 13 (pc), August 1, September 19 (pc), October31 and December 12 (pc).873.3. Financial Markets around FOMC Announcementseach fomc meeting date, we infer the probability of an interest rate changefor the following 22 meetings. For each observation date, we then computethe change in the probability of an interest rate change ∆P (l) between eachconsecutive future meeting pair along with the associated ∆PC indicator.This gives us of a total of 567 observations: 21 meeting pairs for each of the27 observation dates. We then run the following panel regression:∆P (l)t,i = α+21∑i=1β∆PCi∆PCt,i + εt,i (3.8)where t represents the observation date and i represents the ith pair ofconsecutive future meetings. Regression results are presented in Figure 3.3.Blue squares indicate coefficient estimates β∆PCi while errors bars indicatethe 95% confidence interval from standard errors clustered by observationdate and meeting pair. All 21 coefficient estimates are positive, and all buttwo are statistically significant at the 5% level. This suggests that marketsexpect more important decisions on days with press conferences not only forthe upcoming fomc meeting, but for at least three years into the future.Overall, the prices of Federal Fund Futures, combined with the reac-tions of equity markets to fomc announcements, paint a clear picture thatmarkets expect big changes in monetary policy only following fomc meet-ings with press conferences, and view the remaining announcements as lessimportant.3.3.2 Resolution of Uncertainty at FOMC AnnouncementsHaving established that markets view fomc announcements on days withpress conferences as more important than those without, we now ask if theFed reveals more information on these days. To quantify the amount ofinformation revealed, we follow Beber and Brandt (2009), Savor and Wilson(2013a), and Amengual and Xiu (2015) and use the option implied volatilityindex, vix, as proxy for uncertainty associated with monetary policy. Withthe arrival of new information, we generally expect uncertainty to decrease.52But volatility would change little if announcements merely confirm whatmarkets already expected, or if announcements provide little price-relevant52Beber and Brandt (2009) and Savor and Wilson (2013a) show a general link betweenresolution of macroeconomic uncertainty and changes in the vix index. Amengual andXiu (2015) are specifically interested in large downward jumps in the vix, and argue thatin addition to resolving uncertainty, the Fed usually intervenes in hard times, effectivelyproviding a put option to markets. For our purpose, the distinction between both inter-pretations is secondary.883.3. Financial Markets around FOMC AnnouncementsFigure 3.3: Term Structure of the Probability of Target Rate Changes: Re-gressionThis figure shows estimates from the panel regression:∆P (l)t,i = α+21∑i=1β∆PCi∆PCt,i + εt,i,where t represents the observation date and i represents the ith pair ofconsecutive future meetings. Observation dates are fomc meetings datesfrom July 2012 to October 2015. Using settlement prices from fomc an-nouncement days, we infer the probability of an interest rate change forthe following 22 meetings. For each observation date, we then compute thechange in the probability of an interest rate change ∆P (l), in %, betweeneach consecutive future meeting pair. Blue squares indicate coefficient es-timates β∆PCi while errors bars indicate the 95% confidence interval fromstandard errors clustered by observation date and meeting pair. Detailedinformation on the construction of probability measures is provided in thetext.1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21Meeting pair i−202468101214RegressionCoefficientβ∆PCiAdj. R2: 0.246Intercept: 1.319***N. Obs.: 567information. If on the other hand uncertainty in markets was large, and the893.3. Financial Markets around FOMC Announcementsannouncement resolves this uncertainty, we expect large declines in the vix.Figure 3.4 shows cumulative changes in the vix around the fomc an-nouncement, starting 2.5 hours prior and ending 1.5 hours after. The intra-day vix data is provided by trth. Across all fomc announcements (PanelA), the vix exhibits the expected pattern. There is little time-series vari-ation prior to the announcement, but the vix drops sharply by about 2%when the new information arrives. The release of the Fed’s monetary policydecisions clearly reduces uncertainty.A striking contrast emerges in Panel B, which separates fomc announce-ments into ones that are followed by a pc (blue solid line) and ones that arenot (red dashed line). While announcements with pcs see an average dropof over 4% in the volatility index, uncertainty remains virtually unaffectedby fomc announcements without pcs.Table 3.5 formally tests this finding. We first regress log changes invix from one minute prior to 30 minutes after the announcement on in-dicator variables for pc and non-pc days. Regression (1) shows that thevix decreases by a statistically and economically highly significant 4.3% ondays with pcs, and remains unchanged on days without. Including con-trol variables further increases the economic magnitude and the statisticalsignificance of the impact of press conferences.The large decrease in option-implied volatility suggests that a significantamount of uncertainty in equity markets is resolved at the time of the an-nouncement on days with pcs. In contrast, when there is no pc, uncertaintydoes not change around fomc announcements. In turn, this implies thatfomc announcements communicate price-relevant information only on pcdays, and markets correctly expect no relevant monetary policy changes ondays without pcs.A potential confounding effect stems from the publication of economicprojection materials (epms), which contain the economic projections of Fed-eral Reserve Board members and the Federal Reserve Bank presidents aboutgrowth, unemployment, inflation, and future policy. Prior to 2013, these ma-terials were not released until the beginning of the press conferences, andtherefore after the time window of our analysis ends. Since 2013, however,they are made public simultaneously with the fomc announcement. Whilethe relevance of these materials is often debated in the media, they nonethe-less represent additional information that can potentially contribute to thereduction in uncertainty. We address this issue in three ways.First, we introduce an indicator variable 1EPM equal to one for thetime period in which epms are released concurrently with the fomc an-nouncements (2013-2015), and zero otherwise. Regressions (5)-(7) extend903.3. Financial Markets around FOMC AnnouncementsTable 3.5: Returns of VIX at FOMC AnnouncementsThis table reports coefficients from regressions of returns in vix aroundfomc announcements on a press conference indicator PC, equal to one if ameeting is followed by a press conference and zero otherwise, non-PC = 1−PC, and control variables. vix announcement returns are the log changes inthe vix, in %, over the 31-minute intervals starting one minute before fomcannouncements. ∆CPI, ∆UE, and ∆GDP are log changes in, respectively,the consumer price index, the unemployment rate, and the gross domesticproduct. RS&P is the S&P 500 log return over the 21-day interval ending3 days before the announcement. 1EPM is an indicator variable equal toone for events between 2013 and 2015, and zero otherwise. Asymptoticheteroscedasticity robust and bootstrapped standard errors are presented inparenthesis and square brackets, respectively, and bootstrapped p-values initalics. Adjusted R2 and the number of observations N are also reported.The sample period is April 2011 to October 2015.∆V IX(1) (2) (3) (4) (5) (6) (7)Intercept 0.264 0.353 0.702 1.592 1.709 1.619(0.83) (1.36) (1.35) (1.48) (2.06) (1.53)[0.95] [1.38] [1.38] [1.54] [1.73] [1.68]0.78 0.80 0.61 0.30 0.32 0.34PC -4.338 -4.601 -4.855 -4.867 -3.957 -4.195 -3.099(1.02) (1.31) (1.27) (1.26) (1.82) (1.94) (1.55)[0.92] [1.32] [1.25] [1.23] [2.04] [1.96] [2.01]0.00 0.00 0.00 0.00 0.05 0.03 0.12PC x 1EPM -1.415 -1.379 -3.139(2.50) (2.51) (2.36)[2.58] [2.48] [2.63]0.58 0.58 0.23non-PC 0.264(0.83)[0.95]0.78∆CPI 3.141 3.317 2.036 2.193(2.37) (2.37) (2.55) (2.53)[2.62] [2.57] [2.53] [2.45]0.23 0.20 0.42 0.37∆UE 10.064 12.641 9.824 13.325(4.41) (5.04) (5.01) (6.22)[5.53] [5.84] [5.24] [5.50]0.07 0.03 0.06 0.02∆GDP 0.228 0.251 0.198 0.183(0.43) (0.43) (0.43) (0.43)[0.46] [0.45] [0.44] [0.43]0.62 0.58 0.65 0.66RS&P -0.204 -0.280(0.24) (0.21)[0.17] [0.17]0.23 0.101EPM -1.993 -1.770 -0.764(1.75) (1.95) (1.40)[1.89] [1.82] [1.87]0.29 0.33 0.69Adjusted R2 0.225 0.225 0.247 0.252 0.274 0.281 0.305N 37 37 37 37 37 37 37913.3. Financial Markets around FOMC Announcementsregressions (2)-(4) by interacting PC with 1EPM . The interaction term isnot significant in any of the specifications, suggesting that our results arenot driven by this simultaneous release of economic projection materials.Second, we confirm in untabulated results that the changes in vix aroundthe release of epms in 2011 and 2012, which occured at the beginning of thepc, do not suggest that epms reduce uncertainty. The mean log changes invix from one minute prior to 30 minutes after the release of the epms isnot statistically significantly different from zero, with only two out of eightobservations negative.Third, we look at the effect of summary of economic projections (seps)in the period prior to our sample. Quarterly seps, which were subsumed bythe more detailed epms in 2011, were first introduced following the October2007 meeting and released simultaneously with the meeting minutes. FromOctober 2007 to March 2011 there are 28 fomc meetings, 14 of them withseps. For those meetings, we regress in untabulated results the daily changein vix on the day of the release of fomc minutes on an seps dummy whichis equal to one if seps were released at the same time or zero otherwise.The coefficient on the seps dummy is both positive and insignificant, whichfurther suggests that seps do not reduce uncertainty.Taken together, these tests suggest that the information contained ineconomic projection materials does not cause our results, since their releasedoes not affect uncertainty. Even if that was the case, the conclusions andimplications of this paper would remain, but the specific channel would beunclear. Nonetheless, our evidence suggests that it is the mere presence of apc, and possibly the scheduled release of epms, that drives our results andnot the potential information they communicate to the market.Taken together, these tests cast doubt on the value of economic projec-tion materials as a source of information for markets. In particular, sincethe release of the epms does not affect uncertainty, these materials can notbe responsible for our findings. Importantly, even if it was the case thatepms are responsible for the patterns in market expectations we attributeto press conferences, the conclusions and implications of our findings wouldremain unchanged. Only the specific channel that coordinates expectationswould be different. Nonetheless, our evidence suggests that it is the merepresence of a pc, and possibly the scheduled release of epms, that drives ourresults and not the potential information they communicate to the market.The argument that important monetary policy decisions should reduceuncertainty in markets is general and, in contrast to the evidence usingstock market announcement returns, does not require that total informationrevealed at the announcement is constant in the sample. This allows us to923.3. Financial Markets around FOMC Announcementsinvestigate if the segregation of fomc announcements is a new effect causedby press conferences, or if historically some announcements have alwaysimplicitly carried a higher weight. Since most press conferences (15 outof 19) are scheduled following the second fomc meeting in each calendarquarter, we test if fomc announcements at quarter ends have always had alarger impact on uncertainty.Figure 3.5 shows changes in the vix around fomc announcements fromJanuary 2006 to March 2011, separately for the first (dashed red line) andsecond (solid blue line) announcements in each calendar quarter. In short,there is no difference. Therefore, there is no evidence to suggest that thetiming of press conferences simply reflects a previously existing pattern.Instead, the separation into important and less important fomc announce-ments seems to be caused by the advent of press conferences.3.3.3 The Pre-FOMC Announcement DriftIn this section, we revisit the pre-fomc announcement drift of Lucca andMoench (2015a, “lm”) in the recent sample and its relation to press confer-ences. If important fomc monetary policy announcements are associatedwith high anticipatory returns, we expect the magnitude of the pre-fomcannouncement drift to be related to the importance of the announcements.lm find that in the period from September 1994 to March 2011, the S&P500 index has on average increased by 49 basis points in the 24 hours priorto fomc announcements. This return has proven difficult to explain, andone might wonder if this anomaly was specific to the chosen sample or if itis a robust effect. Our sample begins in April 2011, and therefore does notoverlap with the original study.Figure 3.6 shows average cumulative e-mini log returns for the 2-dayswindow ending on fomc announcement days. Panel A shows average re-turns in the April 2011 to October 2015 sample (brown solid line) and forreference also shows the findings from November 1997 to March 2011 (blackdotted line). While this latter sample slightly differs from the one in lm dueto data availability, the cumulative returns are very similar. Panel B sepa-rates announcements from the April 2011 to October 2015 sample into thosewith press conference (blue solid line) and those without (red dashed line).Vertical dashed lines indicate the three announcement times throughout thesample period: 12:30 p.m. (pc 2011-2012), 2:00 p.m. (all announcementsfrom March 2013), and 2:15 p.m. (non-pc until January 2013).Interestingly, we observe a statistically significant pre-fomc announce-ment return, but only for meetings that are followed by press conferences.933.4. Investor Attention to FOMC AnnouncementsThe cumulative return from opening on the day prior until the announce-ment is approximately 50 bp, nearly identical to the one in the earlier sam-ple. While market prices in the lm sample smoothly increase from about24 hours prior to the fomc announcement until the actual announcement,we observe the returns occurring in two waves. Prices first increase on themorning of the previous day, and then jump further overnight. While study-ing the causes of these excess returns is beyond of the scope of this paper,a potential explanation is that the shift in timing is due to investors tryingto front-run the return documented by lm.In a stark contrast, average returns prior to fomc announcement with-out pc are flat. While it is hard to explain those pre-fomc announcementsreturns, it is nonetheless striking that an important pattern previously as-sociated with fomc meetings is now only observable around those with pcs.3.4 Investor Attention to FOMC AnnouncementsIn this section, we study investor attention before fomc meetings. If in-vestors have limited resources and information is costly to acquire or process,investors will optimally choose to focus their attention on information withlarger impact on prices (see, e.g., Sims, 2003; Abel, Eberly, and Panageas,2013; Huang and Liu, 2007; Kacperczyk, van Nieuwerburgh, and Veldkamp,2016). If investors truly expect big changes in monetary policy only follow-ing fomc meetings with press conferences, they should therefore be moreattentive to those meetings.Using different proxies for the attention of institutional and retail in-vestors, in particular media coverage spanning multiple media outlets andfrequencies and Google search volume, we show that interest in the fomcis higher prior to announcements with press conferences. We argue that wemeasure additional attention to the fomc announcements rather than atten-tion to the press conferences themselves, as pcs reveal little new informationto markets and therefore do not command attention. In an out-of-sampletest, we show that similar results obtain in Canada and New Zealand, thetwo other countries where central banks follow a comparable communicationpolicy.3.4.1 Institutional Investor AttentionWe begin our analysis with a proxy for institutional investors’ attentionbased on articles published on the Bloomberg (bb) terminal platform (Ben-Rephael, Da, and Israelsen, 2016). To construct a news intensity measure,943.4. Investor Attention to FOMC Announcementswe first obtain the daily number of articles related to the U.S. Federal Re-serve and then average these over the three business days prior to eachannouncement.53Historical levels of bb are presented in Panel A of Figure 3.7, wherefull circles identify meetings with pcs while hollow dots identify those with-out. After early 2012, we see a clear separation between meetings with pcsand those without. Those with pcs draw more attention, and this is alsoapparent in the other measures of attention that we describe later.Our findings for bb are summarized in Panel A of Table 3.6, wherethe dependent variable is the meeting-to-meeting log change in bb newsintensity. Performing our test on changes rather than levels avoids concernsthat variables might be non-stationary in-sample. The first specificationwithin each group contains only an intercept and changes in the pc indicatorvariable, ∆PC. It shows that, on average, Bloomberg coverage increases by27% on days with pcs relative to days without. This estimate is not onlyeconomically meaningful, but also statistically significant based on both theasymptotic and the bootstrapped distributions. The indicator variable aloneexplains 26% of the total variation in bb news intensity.53The total article count is retrieved from the Bloomberg Terminal using the searchword “Federal Reserve”. While also based on bb, our attention measure differs from theone used by Ben-Rephael, Da, and Israelsen (2016) because their proxy is only availablefor individual equities and not institutions such as the Federal Reserve.953.4. Investor Attention to FOMC AnnouncementsFigure 3.4: Cumulative VIX Return around FOMC AnnouncementsThis figure shows the average cumulative log return, in %, of vix aroundfomc announcements. Returns are normalized to zero at the time of theannouncement. Panel A shows results for the whole sample, while Panel Bseparates announcements into those with press conference (blue solid line)and those without (red dashed line). The shaded areas are pointwise 95%confidence bands around the average returns. The sample period is April2011 to October 2015.-2.5hrs 0Announcement1.5hrs−5−4−3−2−10123CumulativeVIXreturnPanel A: All announcements-2.5hrs 0Announcement1.5hrs−8−6−4−20246CumulativeVIXreturnPanel B: Announcements with and without press conferencesWith press conferencesWithout press conferences963.4. Investor Attention to FOMC AnnouncementsFigure 3.5: Cumulative VIX Return around FOMC Announcements (2006-2011)This figure shows the average cumulative log return, in %, of vix aroundfomc announcements. vix returns are normalized to zero at the announce-ment. Events are separated into the first (red dashed line) and second (bluesolid line) announcements in each calendar quarter. The shaded areas arepointwise 95% confidence bands around the average returns. The samplecontains 42 events from January 2006 to March 2011. -2.5hrs 0Announcement1.5hrs5432101Cumulative VIX return1st announcement in quarter2nd announcement in quarter973.4. Investor Attention to FOMC AnnouncementsFigure 3.6: FOMC Pre-Announcement Drift and Press ConferencesThis figure shows the average cumulative log return, in %, of the shortestmaturity S&P 500 e-mini futures in the 2-day window ending on fomcannouncement days. Panel A shows the results for the November 1997 toMarch 2011 (black dotted line) and the April 2011 to October 2015 (brownsolid line) samples, while Panel B separates announcements in the lattersample into those with press conference (blue solid line) and those without(red dashed line). The shaded areas are pointwise 95% confidence bandsaround the average returns. Vertical dashed lines indicate the three fomcannouncement scheduled times throughout the sample period.09:3010:3011:3012:3013:3014:3015:3009:3010:3011:3012:3013:3014:3015:30−0.50.00.51.0CumulativeE-minireturnPanel A: All announcementsNov. 1997 to March 2011April 2011 to Oct. 201509:3010:3011:3012:3013:3014:3015:3009:3010:3011:3012:3013:3014:3015:30−0.50.00.51.0CumulativeE-minireturnPanel B: Announcements with and without press conferencesWith press conferencesWithout press conferences983.4. Investor Attention to FOMC AnnouncementsFigure 3.7: Attention Level Before FOMC AnnouncementsThis figure shows the level of attention measures prior to each fomc an-nouncement. Full circles identify meetings followed by press conferenceswhile hollow dots identify those without. The sample period is April 2011to October 2015. Detailed information on the construction of attentionmeasures is provided in the text.2012 2013 2014 2015100150200250300350400BBPanel A: Bloomberg2012 2013 2014 201550100150200250300INWPanel B: Dow Jones Intraday Newswires2012 2013 2014 2015123456WSJPanel C: Wall Street Journal2012 2013 2014 20150.51.01.52.0NYTPanel D: New York Times2012 2013 2014 20155101520253035SVIPanel E: Google Search Volume Index993.4.InvestorAttentiontoFOMCAnnouncementsTable 3.6: Attention before FOMC AnnouncementsThis table reports coefficients from regressions of meeting-to-meeting logchanges in measures of attention, in %, on changes ∆PC of an indicatorvariable equal to one if a meeting is followed by a press conference andzero otherwise, and control variables. The measures of media attention arebased on articles published on the Bloomberg terminal platform (bb), theDow Jones intraday newswires (inw), or printed in the Wall Street Jour-nal (wsj), and the New York Times (nyt). The weekly Search VolumeIndex (svi) is obtained from Google Trends for searches for “fomc” andrelated terms. ∆CPI, ∆UE, and ∆GDP are log changes in, respectively,the consumer price index, the unemployment rate, and the gross domesticproduct. RS&P is the S&P 500 log return, over the 21-day interval ending3 days before the announcement. Asymptotic heteroscedasticity robust andbootstrapped standard errors are presented in parenthesis and square brack-ets, respectively, and bootstrapped p-values in italics. Adjusted R2 and thenumber of observations N are also reported. The sample period is April2011 to October 2015. Detailed information on the construction of mediaattention measures is provided in the text.1003.4.InvestorAttentiontoFOMCAnnouncementsPanel A: ∆BB Panel B: ∆INW Panel C: ∆WSJ Panel D: ∆NYT Panel E: ∆SVI(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) (13) (14) (15)Intercept -0.259 14.531 23.398 2.026 13.269 21.609 -0.184 8.570 20.412 2.288 -0.349 7.977 -1.135 -12.321 -9.917(6.74) (8.97) (7.08) (9.66) (11.47) (10.69) (8.42) (17.22) (14.46) (8.76) (17.92) (17.04) (7.41) (11.20) (11.23)[6.74] [12.43] [11.07] [9.60] [18.01] [17.52] [8.46] [16.27] [14.40] [8.70] [16.67] [16.06] [7.38] [13.85] [14.13]0.97 0.24 0.03 0.82 0.46 0.22 0.97 0.62 0.15 0.79 0.98 0.62 0.88 0.37 0.49∆PC 26.633 26.759 27.474 50.286 50.716 51.389 21.150 21.636 22.592 18.510 19.366 20.038 21.047 22.006 22.200(7.04) (7.00) (5.85) (10.03) (9.37) (8.73) (8.97) (8.93) (7.45) (9.02) (8.55) (8.20) (7.50) (7.03) (6.84)[7.03] [6.65] [5.76] [10.02] [9.64] [9.12] [8.83] [8.71] [7.50] [9.08] [8.92] [8.36] [7.71] [7.41] [7.36]0.00 0.00 0.00 0.00 0.00 0.00 0.02 0.01 0.00 0.04 0.03 0.02 0.01 0.00 0.00∆CPI -57.685 -54.101 -69.728 -66.356 -32.783 -27.995 -13.330 -9.964 3.464 4.436(23.39) (22.66) (35.48) (33.21) (27.92) (26.13) (26.49) (23.52) (29.24) (30.02)[27.59] [23.94] [39.92] [37.80] [36.12] [31.12] [36.94] [34.65] [30.72] [30.52]0.04 0.02 0.08 0.08 0.37 0.37 0.73 0.77 0.93 0.90∆UE -13.524 50.422 -44.163 15.984 -52.632 32.769 -90.319 -30.275 -100.061 -82.726(46.39) (49.53) (74.73) (84.39) (59.40) (55.92) (73.80) (75.60) (41.56) (45.39)[56.53] [52.33] [81.85] [82.77] [73.97] [68.10] [75.77] [75.94] [62.91] [66.74]0.81 0.33 0.59 0.85 0.48 0.63 0.24 0.69 0.11 0.21∆GDP -4.385 -3.942 -3.287 -2.870 -4.490 -3.898 -1.696 -1.280 1.035 1.156(4.04) (3.86) (6.21) (6.41) (6.08) (5.32) (6.31) (5.88) (4.74) (4.85)[4.73] [4.10] [6.85] [6.49] [6.19] [5.34] [6.34] [5.94] [5.28] [5.24]0.35 0.33 0.63 0.66 0.47 0.47 0.79 0.83 0.85 0.83RS&P -5.188 -4.880 -6.928 -4.871 -1.406(1.48) (2.19) (1.61) (1.41) (2.20)[1.50] [2.38] [1.95] [2.18] [1.92]0.00 0.04 0.00 0.03 0.45Adjusted R2 0.264 0.282 0.443 0.394 0.390 0.435 0.112 0.060 0.280 0.077 0.029 0.119 0.147 0.141 0.126N 36 36 36 36 36 36 36 36 36 36 36 36 36 36 361013.4. Investor Attention to FOMC AnnouncementsIn the second and third specifications, we add control variables. Of these,only changes in CPI and S&P 500 returns are significant. The negative co-efficient on returns suggests that interest in the Fed is higher after bad stockmarket realizations, consistent with well documented investor behavior, forexample under prospect theory (Kahneman and Tversky, 1979). Impor-tantly, none of the control variables affects the coefficient of interest. ∆PCremains economically large and statistically significant in all specifications.We next move to a high-frequency measure of institutional investor at-tention that is based on intraday newswires (inw) in the hours before fomcannouncements. From RavenPack’s global macroeconomic news database,we collect a comprehensive sample of news stories from the Dow Jones NewsWire. We keep only intraday news that are classified as full-article, and aretimestamped in the 24-hour window ending 1 minute before fomc announce-ments. To capture the predominance of the entities mentioned, RavenPackassigns to each news a relevance score between 0 and 100. We select newsarticles with a minimum relevance score of 90 for either the Federal Reserveor the Federal Open Market Committee.Our findings are summarized in Panel B of Table 3.6. As with Bloombergnews, the coefficient estimate on ∆PC is positive, highly significant, andunaffected by control variables. We find that on days with pcs the numberof articles related to the fomc in the intraday Dow Jones newswires increasesby 50%. Nearly 40% of the variation in the number of intraday newswirearticles on fomc announcement days can be attributed to pcs taking place.3.4.2 Retail Investor AttentionWe next turn to proxies for retail investors’ attention, which are based onlow-frequency printed news in the Wall Street Journal (wsj) and the NewYork Times (nyt) around fomc announcements. To measure daily news in-tensity, we follow Fisher, Martineau, and Sheng (2016) and divide the num-ber of articles related to the fomc or monetary policy by the total numberof articles published in the morning editions of each newspaper.54 We thenaverage daily intensity over windows that start three business days beforethe announcement and end with the morning edition on the announcementday. Fisher, Martineau, and Sheng (2016) provide a detailed overview of theconstruction of macroeconomic media attention indices and their statisticalproperties.54In particular, we search factiva for the following key words: ((federal reserve orfederal open market committee or fomc) and (interest rate or monetary or inflation oreconomy or economic or unemployment)).1023.4. Investor Attention to FOMC AnnouncementsOur findings are summarized in Panels C (wsj) and D (nyt) of Table3.6. As with our measures of institutional investor attention, the coefficientestimate on ∆PC is positive, highly significant, and unaffected by controlvariables. We find that on days with pcs, media attention in the wsj in-creases by 21%. The regression R2 suggests that 11% of the variation inthe number of printed news articles prior to fomc announcement days canbe attributed to pcs taking place. A similar picture emerges for our mediaattention measure for the nyt.3.4.3 Google Search VolumeWe conclude the attention analysis with the search volume index (svi) fromGoogle Trends, which measures the frequency of searches in Google for givenkeywords. Data obtained from Google Trends have previously been used tostudy the effects of investor attention (Da, Engelberg, and Gao, 2011a) andto obtain broad sentiment measures (Da, Engelberg, and Gao, 2015). Inparticular, the weekly svi is calculated by dividing the number of searchesfor specific keywords (“fomc” and related terms), by the total number ofsearches in a geographic area (“global”), and rescaling the resulting seriesso that the maximum is 100.In contrast to our previous measures, the search volume index proxiesfor the overall level of interest among Google’s users. Google is often usedas a universal shortcut to websites. We posit that the svi is a proxy forweb traffic to the fomc and other related websites, and therefore quantifiesinvestor attention. We use the svi in the last full week prior to each fomcmeeting, and again analyze meeting-to-meeting log changes.55The findings in Panel E of Table 3.6 mirror the previous ones. Searchvolume for “fomc” is 21% higher prior to announcements with pcs thanbefore those without. Overall, there is strong evidence that both mediaand investors attention has shifted since the introduction of fomc pressconferences. Rather than equally spreading their attention over all eightfomc announcements per year before pcs were introduced in 2011, investorsnow put more emphasis on the four announcements that are accompaniedby press conferences.55Given that the svi is based on calendar weeks, concerns might arise if some fomcannouncements are later in the week than others. In our sample, the vast majority ofannouncements fall on a Wednesday, only three on a Tuesday and two on a Thursday.Since the two Thursday announcements are followed by pcs while the three Tuesdayannouncements are not, a possible bias would work against our findings.1033.4. Investor Attention to FOMC Announcements3.4.4 The Information Content of Press ConferencesIncreased attention prior to fomc announcements with press conferences byitself does not need to be surprising. If press conferences themselves com-municate information that investors pay attention to, our findings would notrepresent increased attention to fomc announcements, but rather incremen-tal attention to pcs. We now demonstrate that press conferences reveal littlenew information to equity markets, and therefore do not command the extraattention.In efficient markets, the release of price relevant new information in-duces prices to move instantly. Consequently, in a large class of models,information flow is equivalent to volatility (e.g., Ross, 1989). We estimatehigh-frequency measures of realized volatility during pcs to proxy for theinformation revealed. While realized market volatility is generally high dur-ing pcs, this is due to the preceding fomc announcements. In particular,we show that realized volatility is not significantly higher during actual pcsthan during the same time frame following fomc announcements withoutpcs. Similarly, the vix index is largely unchanged during pcs, suggestingthat possible information revealed during pcs does not reduce monetarypolicy uncertainty.We define realized volatility during pcs as the square root of meansquared one-minute e-mini log returns, expressed in percent per year, inthe 60-minute window starting with the press conference. Panel A of Table3.7 shows that realized volatility during pcs is around 15.9%. Crucially,the point estimate for average realized volatility is not larger than the onefor the control sample, estimated during the times when pcs would takeplace on non-pc days. To the contrary, volatility in the control sample isslightly larger at 17.2%. The bootstrapped standard errors closely resemblethe asymptotic ones, and the p-values confirm the findings. Overall, there isno indication that volatility during actual press conferences might be higherthan at the same time on days without pcs.Basing conclusions of this test on the entire sample induces a possiblebias. In 2011 and 2012, fomc announcements with pc were held earlier inthe day than those without (see Table 3.1), and pcs started 1.75 hours afterannouncements. This implies that the time window for hypothetical pcs inthe control group comprises of the first trading hour of the next trading day,and volatility is known to vary throughout the day.Panel B shows the results for the reduced sample from March 2013 toOctober 2015 that is unaffected by differences in announcement times. Therealized volatility during pcs is 17.6%, higher than the 13.6% in the control1043.4. Investor Attention to FOMC AnnouncementsTable 3.7: Realized Volatility during Press ConferencesThis table reports the realized volatility (rv) of the shortest maturity S&P500 E-mini Futures returns during fomc press conferences (pcs). rv isdefined as the annualized mean of squared one-minute midquote log returns,in %, during the 60-minute interval starting at the press conference. Theratio of this rv relative to the announcement rv, estimated between 1 minuteprior and 30 minutes after the announcement, is also reported. On dayswithout pcs, rv is estimated during the corresponding event-time in whichpcs would take place. Asymptotic and bootstrapped standard errors arepresented in parenthesis and square brackets, respectively, and bootstrappedp-values in italics. N denotes the number of observations. The sample periodis April 2011 to October 2015.rv rv relative toannouncement rvpc No pc Difference pc No pc DifferencePanel A: Full SampleMean 15.92 17.18 -1.26 0.64 0.63 0.01Std. Error (asympt.) (3.09) (0.07)Std. Error (bootstr.) [3.00] [0.06]p-value (bootstr.) 0.65 0.82N 19 18 19 18Panel B: March 2013 to October 2015Mean 17.61 13.55 4.06 0.62 0.56 0.06Std. Error (asympt.) (2.85) (0.08)Std. Error (bootstr.) [2.72] [0.08]p-value (bootstr.) 0.14 0.43N 11 11 11 111053.4. Investor Attention to FOMC Announcementssample. The difference has a p-value of 0.14, suggesting that pcs do notconvey important information. This conclusion remains when we controlfor the information content of fomc announcements. Volatility estimatesduring both actual pcs and in the control group are nearly identical relativeto those estimated between one minute before and 30 minutes after fomcannouncements. Since realized volatility spikes immediately after announce-ments and then declines slowly, our evidence indicates that high volatilityduring pcs is driven by news revealed at the fomc announcement, and notthe press conferences themselves.56Overall, the evidence based on realized volatility does not suggest thatimportant price-relevant information is revealed during press conferences.We also confirm in untabulated results that the average change in vix fromthe beginning to the end of the pc is zero, corroborating our conclusion thatpress conferences provide little additional information to markets.3.4.5 International EvidenceWe now look at evidence from other countries as out-of-sample evidence forour findings. Most central banks hold press conferences following each oftheir regular meetings, for example the European Central Bank, the Bankof Japan, Sweden’s Riksbank and Norway’s Norges Bank. We are aware ofonly two central banks that follow a pattern similar to the one adopted bythe fomc: the Reserve Bank of New Zealand and the Bank of Canada.57Since March 1999, the Reserve Bank of New Zealand holds eight regularannual meetings, and every other meeting is followed by a press conference.Our sample ends in August 2016 and contains 141 meetings, 70 of whichhad pcs. The Bank of Canada follows the same pattern, but only startedpcs in January 2013. Until July 2016, there were 29 meetings, 15 of whichwere followed by a pc.Since not all our previous tests are applicable to an international setting,we repeat only the analysis using Bloomberg news intensity and Googlesearch volume in these two countries.58 We first obtain historical Bloomberg56Persistence in realized volatility following macroeconomic news is well documented inAndersen, Bollerslev, Diebold, and Vega (2003b), and investigating its causes is beyondthe scope of our paper.57Two additional central banks hold pcs only after only some announcements. TheBank of England’s Monetary Policy Committee holds monthly meetings, and issues aquarterly Inflation Report that is followed by a pc. However, until August 2015, theinflation report was released about one week after the monetary policy announcement.The Swiss National Bank hold quarterly meetings and semi-annual pcs.58The Wall Street Journal, the New York Times, and the intraday newswires are US-1063.5. Shaping Expectations and Coordinating Attentionnews intensity for announcements of both central banks considered, and theGoogle svi based on searches in the respective home country from GoogleTrends.59Our findings are summarized in Table 3.8. Our main attention results areconfirmed for both central banks considered. On days with pc, the attentionin Bloomberg news coverage increases by 31% in Canada and 20% in NewZealand. Similarly, Google search intensity increases by 24% in Canada and13% in New Zealand. These findings suggest that the shift in attentioninduced by post-announcement press conferences is not unique to the fomcbut present for all central banks that have adopted similar communicationpatterns.3.5 Shaping Expectations and CoordinatingAttentionIn this section, we investigate the economic mechanism underlying our find-ings. In particular, we want to understand why market expectations andinvestor attention are higher on days with pcs, and why the Fed makesdecisions that reduce monetary policy uncertainty only on these days.It is conceivable that the Fed instituted press conferences with the inten-tion to defer important decisions for meetings when it has the opportunity toprovide explanations and context in a pc. This is a natural and convincingargument, even though it was not mentioned when a possible introductionof press conferences was originally discussed, and Chairwoman Yellen main-tains that all fomc meetings are equally important.60Alternatively, it is also possible that the shifts in market expectationsand investor attention are unintended consequences of the press conferences.based media with sparse international coverage. In addition, Brusa, Savor, and Wilson(2016) show that, while fomc decisions impact international stock markets, those marketsdo not react significantly to decisions of their domestic central bank. Consequently, we donot expect foreign financial markets to react as the U.S. market does.59We adjust the timezone settings in Bloomberg prior to obtaining news count for theReserve Bank of New Zealand to avoid capturing news published after the announcement.Google Trends provides weekly data for a maximum time range of 5 years, and we thereforefollow our original sample and obtain weekly Google Trends data for 2011 to 2015.60Press conferences were first mentioned during a fomc conference call on October15, 2010. The discussion revolved about what other central banks do, about providing “alittle more clarity”, and that it “dovetails with some of the concerns about interpretations”(Bernanke, 2010). It is also acknowledged that “communicating what we are doing will bechallenging”, that pcs “would probably become obligatory on a regular basis”, and thatit “would be quite a commitment”.1073.5. Shaping Expectations and Coordinating AttentionTable 3.8: Attention before Announcements in Canada and New ZealandThis table reports coefficients from regressions of meeting-to-meeting logchanges in Bloomberg news count (BB) and the Google Search Volume In-dex (svi), in %, on changes ∆PC of an indicator variable equal to one ifa meeting is followed by a press conference and zero otherwise, for interestrate announcements of the Bank of Canada and the Reserve Bank of NewZealand. Asymptotic heteroscedasticity robust and bootstrapped standarderrors are presented in parenthesis and square brackets, respectively, andbootstrapped p-values in italics. Adjusted R2 and the number of observa-tions N are also reported. The sample period is January 2013 to July 2016for Canada, March 1999 to August 2016 for Bloomberg on New Zealandand January 2011 to December 2015 for Google on New Zealand. Detailedinformation on Bloomberg news and the svi is provided in the text.Canada New Zealand∆BB ∆SVI ∆BB ∆SVI(1) (2) (3) (4)Intercept -1.968 0.797 -1.398 0.241(10.36) (6.90) (3.03) (5.34)[10.35] [6.90] [3.03] [5.31]0.85 0.93 0.65 0.96∆PC 31.166 23.532 20.190 13.111(10.36) (6.90) (3.09) (5.34)[10.37] [6.91] [3.06] [5.31]0.00 0.00 0.00 0.01Adjusted R2 0.215 0.266 0.231 0.112N 28 28 140 391083.5. Shaping Expectations and Coordinating AttentionAt least two channels might be responsible. First, markets might falsely in-terpret the Fed’s intention and assign a small probability that the Fed wantsto time important decisions to be made on days with pcs. Investors conse-quently lower their expectations of monetary policy action on days withoutpcs, and shift their attention accordingly. The decreased expectations inturn imply that any action from the Fed would be a surprise to markets,which the Fed does not like. It effectively limits the range of actions the Fedcan take on non-pc days. This constraint of the Fed naturally feeds back tomarket expectations and investor attention.Second, with information acquisition costs, a small difference in the per-ceived importance of fomc meetings due to the introduction of pcs can besufficient to shift attention away from announcements without pcs. Thisattention shift in turn can influence the Fed because investor attention iscritical to the transmission of monetary policy (Stein, 1989; Blinder, Good-hart, Hildebrand, Lipton, and Wyplosz, 2001), and “monetary policy is moreeffective if it is more effective in coordinating market expectations” (Amato,Morris, and Shin, 2002, p.496). In this scenario, small initial changes in in-vestor attention lead the Fed to slightly shift their policy decisions, whichfeeds back to market expectations and investor attention.The data can help us understand whether the Fed intended to focus mon-etary policy to days with pcs, or whether this shift in focus was dictated bymarkets. In the first case, we should observe an immediate drop in the im-portance of monetary policy decisions on days without pcs. Depending onwhether investors understand this or learn from past policy announcements,we expect to see changes in expectations and attention either instantly or de-veloping over time. In the second case, while we cannot distinguish betweenthe exact mechanism of how unintended consequences might arise, we wouldexpect to see the magnitude of all of our findings to increase over time ratherthan observing an immediate impact. Of course, the two interpretations arenot mutually exclusive.To answer this question, we test whether press conferences had an im-mediate impact, or if their effects appeared gradually. Table 3.9 revisits ourmain regressions, but also interacts our variables of interest, PC or ∆PC,with a time trend variable T , which is set to 0 for the first meeting andincreases by 1/8 for each subsequent meeting. Since there are 8 meetingsper year, T increases by one for every year.The first three columns show results for our proxies of market expecta-tions. Focusing on the interaction term, we see that it is not significantlydifferent from zero for e-mini announcement returns, but positive for thetwo ex-ante measures of Fed Fund Futures implied probabilities of interest1093.5. Shaping Expectations and Coordinating AttentionTable 3.9: Regressions with Time TrendsThis table reports coefficients from regressions of log returns in e-mini andvix around fomc announcements on a press conference indicator PC, equalto one if a meeting is followed by a press conference and zero otherwise,and from regressions of meeting-to-meeting changes in the probability ofinterest rate changes and log attention measures, in %, on changes ∆PC inPC, on interaction with a time trend T and control variables. T is 0 for thefirst meeting and increases by 1/8 for each meeting (by one for every year).∆CPI, ∆UE, and ∆GDP are log changes in, respectively, the consumerprice index, the unemployment rate, and the gross domestic product. RS&Pis the S&P 500 log return, over the 21-day interval ending 3 days beforethe announcement. Asymptotic heteroscedasticity robust and bootstrappedstandard errors are presented in parenthesis and square brackets, respec-tively, and bootstrapped p-values in italics. Adjusted R2 and the number ofobservations N are also reported. The sample period is April 2011 to Octo-ber 2015. Detailed information on the construction of dependent variablesis provided in the text.Ret ∆P (l) ∆P (↑) ∆V IX ∆BB ∆INW ∆WSJ ∆NYT ∆SVI(1) (2) (3) (4) (5) (6) (7) (8) (9)Intercept -0.750 -1.140 -2.694 1.157 13.328 31.170 -3.223 6.981 -30.052(0.32) (2.31) (2.34) (1.74) (14.87) (21.70) (17.56) (21.85) (19.67)[0.25] [2.51] [2.35] [1.79] [15.11] [26.08] [17.93] [23.41] [19.71]0.00 0.66 0.25 0.52 0.38 0.23 0.85 0.76 0.12T 0.140 0.333 0.551 0.276 -0.534 -3.888 1.408 -3.065 3.040(0.09) (0.67) (0.73) (0.73) (4.18) (6.47) (4.54) (6.26) (6.11)[0.09] [0.69] [0.64] [0.65] [4.12] [7.12] [4.89] [6.38] [5.38]0.12 0.63 0.39 0.67 0.90 0.58 0.77 0.63 0.57PC (∆PC) 0.570 -6.258 -5.991 0.117 -9.689 52.850 -44.150 -6.816 -20.489(0.27) (2.67) (2.26) (2.03) (20.38) (31.29) (18.57) (21.84) (21.09)[0.34] [2.50] [2.34] [2.45] [15.07] [26.01] [17.88] [23.33] [19.66]0.09 0.01 0.01 0.96 0.52 0.04 0.01 0.77 0.30PC (∆PC)× T 0.022 3.690 3.730 -2.284 14.933 -0.587 26.819 10.790 17.154(0.11) (1.01) (0.98) (1.01) (6.76) (10.57) (6.36) (8.01) (8.70)[0.13] [0.94] [0.88] [0.97] [5.68] [9.80] [6.73] [8.79] [7.40]0.88 0.00 0.00 0.02 0.01 0.95 0.00 0.22 0.02∆CPI 0.058 2.202 4.124 0.333 -41.747 -72.745 -2.235 -5.083 24.144(0.37) (3.43) (2.62) (2.57) (24.71) (36.90) (24.57) (26.63) (26.28)[0.34] [3.90] [3.65] [2.52] [23.49] [40.49] [27.84] [36.28] [30.60]0.88 0.58 0.26 0.89 0.07 0.07 0.95 0.89 0.43∆UE -1.509 2.031 1.048 15.707 -1.938 17.543 -60.963 -68.453 -142.403(0.76) (6.97) (6.77) (6.26) (40.23) (84.55) (48.54) (78.20) (51.43)[0.74] [8.59] [8.04] [5.43] [51.69] [89.28] [61.31] [80.04] [67.33]0.05 0.81 0.90 0.00 0.97 0.84 0.32 0.39 0.04∆GDP -0.027 0.276 0.495 0.112 -2.643 -2.678 -1.712 -0.175 2.421(0.05) (0.57) (0.61) (0.40) (3.79) (6.68) (4.36) (5.58) (4.64)[0.06] [0.63] [0.59] [0.41] [3.78] [6.52] [4.48] [5.84] [4.93]0.63 0.66 0.40 0.78 0.48 0.68 0.71 0.97 0.62RS&P 0.060 -0.247 0.072 -0.391 -2.344 -4.985 -1.825 -2.812 1.854(0.04) (0.30) (0.24) (0.19) (1.78) (2.80) (2.10) (2.12) (2.47)[0.02] [0.29] [0.27] [0.17] [1.74] [3.01] [2.07] [2.70] [2.28]0.01 0.40 0.80 0.02 0.18 0.10 0.38 0.30 0.42Adjusted R2 0.317 0.437 0.435 0.358 0.503 0.400 0.466 0.106 0.185N 37 36 36 37 36 36 36 36 361103.6. Conclusion to Chapter 3rate changes. This suggests that the effect of pcs on announcement returnsis approximately constant throughout our sample period, but expectationsmeasured from Fed Funds Futures increased over time.Column (4) shows the results for the changes in the vix at announce-ments. The coefficient on the pc indicator of 0.117 is insignificant, suggestingthat resolution of macroeconomic uncertainty was unrelated to press confer-ences early in the sample. The interaction term, in contrast, is significantlynegative at -2.284. This implies that, with each passing year in our sample,the difference in resolution of monetary policy uncertainty between dayswith and without pcs increases by over two percentage points of the vixindex.Lastly, columns (5)-(9) show that the effect of pcs on investor attentionalso becomes more pronounced over time. In particular, the interactionterm is significantly positive for three of our five proxies, including articlespublished on the Bloomberg terminal and in the Wall Street Journal andthe Google search volume index. This is consistent with pcs acting as adevice for coordinating attention.Overall, our results for market expectations and investor attention clearlysupport the hypothesis that markets are slowly adjusting to the new com-munication policy. The increasing importance of press conferences on theamount of information released at fomc announcements, as measured bychanges to monetary policy uncertainty, is especially interesting. It suggeststhat the Fed did not initially choose to designate fomc meetings with pcsas more important than those without, but is adjusting to their new policyand reacting to changes in market expectations and investor attention.3.6 Conclusion to Chapter 3In an effort to increase transparency, the Chair of the Board of Governorsnow holds a press conference following half of the scheduled fomc announce-ments. While press conferences do not add significant information relativeto the preceding announcement, we document that this information practicehas unintended consequences: it curtails the range of actions the Fed cantake and counteracts the declared transparency goal.Holding press conferences after some, but not all, fomc meetings skewsexpectations of important monetary policy decisions towards announcementdays with press conferences. This is turn coordinates media and investorattention towards those meetings. Since managing market expectations iscentral to monetary policy, it is optimal for the Fed to focus their policy1113.6. Conclusion to Chapter 3efforts on times when markets pay close attention.As a result, the Fed, generally believed to be averse to surprising markets,now faces two obstacles to make important monetary policy decisions atmeetings without press conferences: markets do not expect big decisions,and investors pay less attention. This constrains the possible monetarypolicy decisions. Naturally, these constraints diminish information flow andreduce transparency.Taken to the extreme, our evidence raises the question why the fomcmeets and makes policy announcements when there are no press confer-ences. Resolving the constraints on actions and the associated reducedtransparency requires that markets perceive all fomc announcements equal.While this could be achieved by removing press conferences completely, inorder to maintain their goal of increased transparency, the Fed should in-stead consider holding press conferences after every meeting, as many othercentral banks do.112Chapter 4Media Attention,MacroeconomicFundamentals, and the StockMarket4.1 IntroductionClassical theories of asset pricing, based on exogenous information flowsand efficient market pricing (e.g., Merton, 1973), provide no explicit role forinvestor attention. A growing literature establishes however that investorattention, to both firm-level and aggregate news, plays an important rolein financial markets. For example, Da, Engelberg, and Gao (2011b) showthat investor attention to individual stocks positively predicts subsequentshort-run returns for those stocks.61 Andrei and Hasler (2014) develop the-oretical and empirical links between attention to the aggregate stock mar-ket and conditional moments of the aggregate stock market. Kacperczyk,Van Nieuwerburgh, and Veldkamp (2016) study interactions between firm-level and aggregate attention.If attention in general is important to understanding financial markets,then what other types of attention, beyond firm-level and aggregate atten-tion, might be worth studying? In this paper we propose new measuresof attention, derived from news media coverage, to separate categories ofmacroeconomic fundamentals such as unemployment, output growth, infla-tion, and oil prices.We focus on macroeconomic fundamentals for several reasons. First,the finance literature has long sought to connect asset prices to underly-ing macroeconomic factors (Chen, Roll, and Ross, 1986). Second, currentevidence establishes that scheduled macroeconomic announcements have61For further evidence regarding attention to individual stocks, see Huberman and Regev(2001); Barber and Odean (2008); DellaVigna and Pollet (2009).1134.1. Introductionstrong impacts on asset prices (Andersen, Bollerslev, Diebold, and Vega,2003a, 2007a; Savor and Wilson, 2013b), and we anticipate that such an-nouncements should also impact attention. Third, while the asset pricingliterature often tends towards stock-market based factors in describing thecross-section of returns (e.g., Fama and French, 1993), casual observationof news media coverage suggests that attention to systematic risks is morefrequently framed in terms of macroeconomic factors such as unemploymentand inflation as opposed to stock-market based factors like size and value.Finally, an interesting aspect of attention to macroeconomic fundamentals isthat we can relate the dynamics of attention to the dynamics of the underly-ing macroeconomic fundamentals. This allows us to answer questions suchas what types of changes in unemployment or output growth or inflationresult in increases or decreases in attention to these fundamentals.Our measures of attention are based on media coverage of different typesof fundamental news. The categories of macroeconomic fundamentals are:unemployment, output growth, inflation, credit ratings, the housing market,interest rates, monetary policy, oil, and the U.S. dollar. We create lists ofsearch words that capture attention to each of these fundamentals. Forexample, to capture attention to U.S. output growth, we use the followingset of words: gross domestic product, gdp, gross national product, andgnp. We count the number of articles in the Wall Street Journal (wsj) andNew York Times (nyt) starting in 1980 for nyt and 1984 for wsj until2015 that include any of these search terms. Scaling by the total number ofarticles published gives us a measure of relative attention to each categoryof macroeconomic fundamental.Our indices most directly measure media attention, but the media clearlyhas strong incentives to cover issues of interest to their readers, and priorliterature often uses media attention as a proxy for investor attention (e.g.,Barber and Odean, 2008; Yuan, 2015). A separate line of research, which wedo not contribute to, investigates the causal role of media attention (e.g.,Tetlock, 2007, 2010; Peress, 2014). We view media coverage as a usefulproxy for investor attention because of the long time series it permits. Ourindices permit daily estimates of attention beginning in 1980. More directmeasures of investor attention, such as Google search (e.g., Da, Engelberg,and Gao, 2011b) have other advantages but provide shorter time series.Henceforth, we do not distinguish between media and investor attention,although this could be an interesting topic for future research. Althoughnot the focus of our research, we do provide separate measures of attentionfor the nyt and wsj, which suggests heterogeneity in attention across thedifferent readerships of these outlets.1144.1. IntroductionOur macroeconomic attention indices (“mai”) show interesting empiricalproperties. We first address comovement in attention, and show that theindices are not driven by a single factor. They are imperfectly correlated,and over time attention shifts across inflation, employment, monetary policy,and the other fundamentals. If these shifts in attention reflect changes ininvestor concerns, then only in very special cases could efforts to price assetsreduce to a single factor representation of risk.We next address the duration of cycles in attention. For the macroeco-nomic fundamentals we consider, the attention indices are stationary, butpersistent. The conservative Bayesian Information Criterion suggests atmost four lags in a monthly autoregression framework. However, when weaggregate the attention indices over different window lengths, similar to themidas framework of Ghysels, Santa-Clara, and Valkanov (2006), we find thatmost of the series show evidence of cycles at multiple frequencies, rangingfrom one day to as long as one year. These aspects of attention are con-sistent with fractal behavior over a range of frequencies, producing a slowdecay in autocorrelations over a range of lags that is often associated withlong-memory. These patterns in attention are properties also observed in ag-gregate stock market volume and volatility in prior literature (see Andersen,Bollerslev, Diebold, and Ebens, 2001; Bollerslev and Mikkelsen, 1996).We next seek to relate attention to movements in economic fundamen-tals. We associate each of the attention indices with a related macroeco-nomic variable, and, where possible, at least one scheduled announcement.As expected, high frequency variations in attention do relate to schedulednews announcements, and we document which announcements have the mostimpact on attention. Lower frequency movements in attention relate tomovements in economic fundamentals. We decompose each of the economicseries (e.g., unemployment, inflation) into simple moving averages over dif-ferent window sizes. Attention relates to variations and squared variations inshorter-horizon simple moving averages of fundamentals relative to longer-horizon moving averages. All significant squared terms on variations arepositive, consistent with the idea that changes in fundamentals lead to in-creased attention. The directional effect of signed changes in fundamentalson attention is generally also consistent with intuition. For example, in-creases in unemployment increase attention, and decreases in house pricesincrease attention. These findings are consistent with Andrei and Hasler(2016) where the authors investigate whether asymmetry in attention is ra-tional and find that investors pay more attention to news the further awaythe predictive variable is from its long-term average.In some cases the relation between attention and fundamentals is very1154.1. Introductionstrong. For example, over 50% of the variation in our unemployment atten-tion index is explained by unemployment fundamentals, and the comove-ment is strong enough to be apparent in a simple plot (see Figure 4.1). Wealso document differences between the wsj and nyt in the strength of therelation between their attention indices and fundamentals.We further show that news media attention to macroeconomic funda-mentals relates to measures of daily stock market activity. Controlling formacroeconomic announcements, increases in attention correlate with higheraggregate volume and higher aggregate volatility.We then investigate how media attention to unemployment might act asa leading indicator to predict the surprise in the announced unemploymentrate, -i.e. the difference between the actual and expected unemploymentrate. Increasing media attention to unemployment leading to up to theemployment announcement predicts the surprise in the unemployment rateand the S&P 500 stock return on announcement day.Finally, we examine how media attention to monetary policy can predictstock returns, changes in vix, and changes in Fed fund rates on fomc an-nouncement days. We find that an increase in attention to monetary on dayspreceding fomc announcements predicts positive stock returns, a decreasein vix, and a decrease in Fed fun rates on fomc announcement days.This paper relates to at least three literatures. The first is research onthe links between attention and financial markets. Theoretical studies builton rational inattention framework highlights the importance of attentionallocation to asset prices (e.g., Sims, 2003; Peng and Xiong, 2006; Kacper-czyk, Van Nieuwerburgh, and Veldkamp, 2016). Andrei and Hasler (2014)establish the links between attention to aggregate stock market volatilityand risk premium and Andrei and Hasler (2016) show that attention istime-varying. Also, recent studies create direct measures of stock-specificinvestor attention using search frequency in Google and find that investorattention predicts stock prices (Da, Engelberg, and Gao, 2011b; Da, Gu-run, and Warachka, 2014). We extend this literature by creating measuresof attention to macroeconomic fundamentals and examining their links tofundamentals as well as the stock market.Second, we contribute to the literature relating macroeconomic news toasset prices. Andersen, Bollerslev, Diebold, and Vega (2003a, 2007a) showthat macroeconomic announcements have an impact on financial assets athigh-frequency. Boyd, Hu, and Jagannathan (2005) find that unemploymentannouncements impact stock prices condition on business cycle. Gilbert(2011) documents that macro announcements revisions have strong relationwith the stock market index. Recent studies find that Federal Open Market1164.1. IntroductionFigure 4.1: Attention to UnemploymentThis figure shows the monthly unemployment attention indices for the Wall Street Journal(mai-wu) and the New York Times (mai-nu) and the monthly unemployment rate. Theblue line is the attention index (mai) and the red dotted line is the unemployment rate.The units are in percentage. The gray vertical bars are nber recessions.0.00.51.01.52.02.53.0UnemploymentMAIMAI-NU (NYT)1980 1984 1988 1992 1996 2000 2004 2008 20120123456UnemploymentMAIMAI-WU (WSJ)Unemployment MAI Unemployment Rate34567891011UnemploymentRate(%)34567891011UnemploymentRate(%)1174.2. Macroeconomic Attention IndicesCommittee (fomc) announcements have significant impact on market riskpremium (Savor and Wilson, 2013b; Cieslak, Morse, and Vissing-Jorgensen,2015b). Media coverage of macroeconomic risks can also be used as a con-ditioning variable in testing asset pricing models (Matthies and Liu, 2015).We show that high-frequency movements in media attention to macro funda-mentals are linked to macroeconomic announcements, while lower-frequencyfluctuations are linked to the fundamentals itself. Further, we show thatchanges in media attention predict both surprises and stock returns on un-employment announcement days.Finally, our paper relates to the literature on text search methods. Ex-amples include Antweiler and Frank (2004), Tetlock (2007), Fang and Peress(2009b). In particular, Baker, Bloom, and Davis (2015) measure economicpolicy uncertainty using, in part, newspaper articles mentioning policy un-certainty. The authors show that economic policy uncertainty (epu) indexaffects both aggregate and firm-level activities. Our research differs by fo-cusing on attention to macroeconomic risks.4.2 Macroeconomic Attention IndicesWe create indices of news-media attention to the following macroeconomicrisks: output growth, inflation, employment, interest rates, monetary policy,housing, credit conditions, oil, and the U.S. dollar. For each fundamental,we create a list of related words and phrases, shown in Table 4.1. We aimfor the lists to be objectively reasonable.1184.2. Macroeconomic Attention IndicesTable 4.1: Newspapers Search WordsThis table presents the search words used to select the articles related tonine specific macroeconomic fundamentals in the Wall Street Journal (wsj)and New York Times (nyt). The nine macroeconomic fundamentals arecredit ratings, Gross Domestic Product (gdp), housing market, inflation,interest rate, monetary, oil, U.S. dollar, and unemployment.Category Newspapers search wordsCredit Rating (credit rating) or (bond rating)gdp gross domestic product or gdp or gnp or gross national productHousing Market (housing market) or (house sale) or (new home start) or(home construction) or (residential construction) or (housing sale)or (home price)Inflation inflation and (economy or economic or Federal Reserve)Interest Rate interest rate and (economic or economy or federal reserve)Monetary (federal reserve or federal open market committee or fomc)and (interest rate or monetary or inflationor economy or economic or unemployment)Oil oilU.S. Dollar U.S. dollar or U.S. exchange rate or U.S. currencyUnemployment (unemployment or population out of work)and (economy or economic)1194.2. Macroeconomic Attention IndicesWe search articles in the Wall Street Journal (wsj) and New York Times(nyt). These publications cover general news, economic news, and financialnews, and have been used in numerous prior studies. We use two differentpublications to provide a sense of the robustness, and also to illuminate dif-ferences in attention across outlets with different audiences. wsj is generallyregarded as having a tighter focus on the economy and financial markets aswell as a more conservative editorial slant, while nyt provides broader cov-erage of general news and has a more politically liberal reputation.62 Forthe nyt, the sample period is from June 1, 1980 to April 30, 2015. For thewsj, the sample period is from January 1, 1984 to April 30, 2015. Duringthese sample periods broad digital coverage of the publications is available.We consider only the newspaper print editions. Table 4.2 presents mai andreports the data sources for associated fundamentals to each mai.62The differences in media slant and its economic impact are well-documented in theliterature (see e.g., DellaVigna and Kaplan (2007); Gentzkow and Shapiro (2010)).1204.2.MacroeconomicAttentionIndicesTable 4.2: Macroeconomic Attention and Macroeconomic FundamentalsThis table presents the macroeconomic attention indices (mai) for credit ratings, gross domestic product(gdp), housing market, inflation, interest rate, monetary, oil, us dollar, and unemployment and its relatedmacroeconomic fundamentals and announcements. The table also reports the data sources for the fun-damentals. The announcement dates are from Bloomberg except for the historical gdp announcements(pre-1997) that are from the U.S. Bureau of Economic Analysis.MAI Fundamental Macroeconomic AnnouncementFundamental Source of Fundamental Name of Announcement FrequencyCredit Rating Corp. Relative Spread∗ Moody’s Corporate Bond Yieldgdp QtQ real gdp log growth rate Federal Reserve of St-Louis Gross Domestic Product (gdp) QuarterlyHousing Nominal Home Price Index Robert Shiller’s website∗∗ Case-Shiller Home Price MonthlyInflation log growth in cpi Bureau of Labor Statistics Consumer Price Index (cpi) MonthlyInterest Federal Fund Rate Federal Reserve of St-Louis Federal Open Market Committee 8 per yearMonetary Federal Fund Rate Federal Reserve of St-Louis Federal Open Market Committee 8 per yearOil Crude Oil Spot Price Energy Information Admin.Unemployment† Unemployment rate Bureau of Labor Statistics Employment Situation Monthlyusd Trade Weighted USD Index Federal Reserve of St-Louis∗ The relative spread is the difference between baa and aaa in corporate bond yields divided by aaa.∗∗ us home prices 1890 to present, http://www.econ.yale.edu/ shiller/data.htm.† Unemployment rates are from the initial release.1214.2. Macroeconomic Attention Indices4.2.1 Construction of the Attention IndicesEach day in the sample period, we count the number of articles in each pub-lication that satisfy the search criteria for each macro fundamental. Thisprovides a daily count Np,f,t, where p indexes the publication (wsj or nyt)of articles showing some form of attention to each fundamental f . We nor-malize these counts by dividing by the average number of articles per dayNˆp,t for publication p during the calendar month including observation t.The“unadjusted” macroeconomic attention index for each individual pub-lication p is:MAI-pUf,t =Np,f,tNˆp,t. (4.1)The unadjusted attention indices measure the percentage of articles on agiven day that have content related to the macroeconomic fundamental ofinterest.We define related measures that are demeaned, or alternatively de-meaned and standardized. Let µp,f and σp,f denote respectively the time-series means and standard deviations of the daily unadjusted attention in-dices MAI-pUf,t. The demeaned measures are denotedMAI-pDf,t = MAI-pUf,t − µp,f ,and the standardized measures are denotedMAI-pf,t = MAI-pDf,t/σp,f .We also define two composite indexes of attention. The first compositeindex, denoted mai-c1, is an average of the demeaned nyt and wsj indicesin time periods when both are available, and the nyt index only in the1980-1983 period:MAI-C1ft ={(MAI-WDft + MAI-NDft)/2 from Jan. 1, 1984 to Apr. 30, 2015,MAI-NDft from June 1, 1980 to Dec. 31, 1983.(4.2)Demeaning the individual publication indices before averaging ensures thatwe will not induce a level effect driven simply by the change in compositionthat occurs in 1984 when the WSJ data becomes available.The second composite index, denoted mai-c2, is an average of the stan-dardized nyt and wsj indices when both are available:MAI-C2ft ={(MAI-Wft + MAI-Nft)/2 from Jan. 1, 1984 to Apr. 30, 2015,MAI-Nft from June 1, 1980 to Dec. 31, 1983.(4.3)1224.2. Macroeconomic Attention IndicesStandardizing ensures that both publications contribute equally to the vari-ation of mai-c2. While the weighting of the two composite indices is differ-ent, neither is superior in any sense. The publication with more variationin its own attention index will be weighted more heavily in mai-c1 relativeto mai-c2. If one believes that greater variation in attention over time re-flects more information, then the weighting of mai-c1 may be preferred tomai-c2.All of the indices build on simple counts of the number of articles relatedto a macroeconomic fundamental, as a proportion of all articles. Manyelaborations of this approach are possible, for example weighting articles bytheir number of words, or attempting to measure the intensity of relevancerather than a simple binary coding. We take a basic approach for simplicity,and expect other measurement methods to be explored in future research.We emphasize that the indices measure attention only, and do not attempt todistinguish other possible article attributes such as positive versus negativesentiment.4.2.2 Empirical Properties of the Attention IndicesTable 4.3, Panel A provides summary statistics for the unadjusted dailyattention indices for both nyt and wsj. For the wsj, the index averagesrange from a low of about 0.5% of articles for credit to a high of over 2%for inflation and oil. nyt coverage of macroeconomic fundamentals is uni-formly lower as a proportion of all coverage. The nyt index means have alowest value of 0.08% for U.S. dollar coverage, and the highest index meansare inflation (0.90%), unemployment (0.81%), and oil (0.76%). Consistentwith the higher mean attention levels in the wsj, the standard deviation ofattention is also uniformly higher for the wsj than the nyt. This impliesthat the weight of the wsj in the composite indices mai-c1 will be higherthan in the composite indices mai-c2.1234.2.MacroeconomicAttentionIndicesTable 4.3: Descriptive StatisticsThis table presents the descriptive statistics for the macroeconomic attention indices (mai). PanelA shows the daily unadjusted media attention indices (mai) for the Wall Street Journal (mai-wuf,t) and New York Times (mai-nu f,t), the Economic Policy Uncertainty (epu) index, the impliedvolatility (vxo), and the three-month detrended log S&P 500 trade volume. Columns Mon toSun are the daily averages for each mai. Panels B shows the correlation between the demeanedmacroeconomic attention composite indices (mai-c1), epu, vxo, and the 60-day detrended S&P500 trade volume at the daily frequency. Obs. stands for the number of observations, and St.dev. stands for the standard deviation.Panel A: Daily unadjusted MAI descriptive statistics (1980-2015)Obs. Mean St. Dev. Min Max Mon Tues Wed Thur Frid Sat SunWall Street JournalCredit Rating 11443 0.46 0.89 0.00 9.67 0.50 0.58 0.73 0.57 0.62 0.22 0.00GDP 11443 1.41 1.54 0.00 12.91 2.09 1.65 1.82 1.77 1.94 0.62 0.00Housing 11443 0.71 1.46 0.00 17.18 0.62 0.68 1.40 0.84 0.99 0.42 0.00Inflation 11443 2.24 2.06 0.00 15.71 3.28 2.47 3.01 2.86 3.15 0.87 0.00Interest 11443 0.95 1.23 0.00 13.54 1.21 1.02 1.40 1.31 1.30 0.40 0.00Monetary 11443 1.91 1.95 0.00 18.62 2.60 2.11 2.61 2.63 2.50 0.90 0.00Oil 11443 2.34 2.57 0.00 19.47 2.82 2.98 3.37 3.05 3.16 0.97 0.00Unemp. 11443 1.44 1.64 0.00 14.07 2.00 1.48 2.09 1.59 2.18 0.73 0.00USD 11443 0.78 1.08 0.00 9.60 0.97 1.07 1.07 1.03 1.08 0.24 0.00New York TimesCredit Rating 12752 0.20 0.43 0.00 10.06 0.11 0.21 0.24 0.23 0.20 0.17 0.23GDP 12752 0.51 0.58 0.00 5.65 0.37 0.43 0.46 0.49 0.53 0.43 0.88Housing 12752 0.29 0.57 0.00 7.23 0.11 0.18 0.28 0.28 0.28 0.20 0.68Inflation 12752 0.90 0.91 0.00 12.26 0.66 0.70 0.93 0.89 0.94 0.82 1.37Interest 12752 0.26 0.38 0.00 3.12 0.19 0.21 0.27 0.28 0.26 0.24 0.34Monetary 12752 0.92 0.77 0.00 8.68 0.60 0.78 0.98 1.04 1.06 0.95 1.05Oil 12752 0.76 0.84 0.00 8.94 0.50 0.73 0.80 0.84 0.81 0.70 0.91Unemp. 12752 0.81 0.90 0.00 10.53 0.58 0.55 0.70 0.67 0.92 0.78 1.48USD 12752 0.08 0.20 0.00 3.34 0.01 0.08 0.07 0.08 0.08 0.07 0.18Other VariablesEPU 11077 102.61 70.29 3.38 719.07 111.25 102.56 96.44 90.01 93.26 90.70 134.02VXO 7386 20.73 9.06 8.51 150.19 20.80 20.67 20.68 20.79 20.74Volume 8798 20.17 1.48 16.52 23.16 20.09 20.19 20.20 20.19 20.17 20.20 20.161244.2.MacroeconomicAttentionIndicesPanel B: Daily MAI-C1 correlation (1980-2015)Credit Rating GDP Housing Inflation Interest Monetary Oil Unemp. USD EPU VXO VolumeCredit Rating 1.00 0.16 0.16 -0.02 0.13 0.17 0.14 0.15 0.15 0.13 0.20 0.29GDP 0.16 1.00 0.15 0.21 0.16 0.23 0.12 0.33 0.10 0.10 0.08 0.25Housing 0.16 0.15 1.00 0.08 0.24 0.26 0.13 0.16 0.06 0.04 0.02 0.38Inflation -0.02 0.21 0.08 1.00 0.34 0.45 0.31 0.22 0.18 0.02 0.02 -0.24Interest 0.13 0.16 0.24 0.34 1.00 0.57 0.33 0.14 0.29 0.08 0.16 0.14Monetary 0.17 0.23 0.26 0.45 0.57 1.00 0.29 0.27 0.24 0.16 0.17 0.20Oil 0.14 0.12 0.13 0.31 0.33 0.29 1.00 0.02 0.37 0.03 0.08 0.02Unemp. 0.15 0.33 0.16 0.22 0.14 0.27 0.02 1.00 -0.02 0.21 0.17 0.16USD 0.15 0.10 0.06 0.18 0.29 0.24 0.37 -0.02 1.00 0.02 0.23 0.03EPU 0.13 0.10 0.04 0.02 0.08 0.16 0.03 0.21 0.02 1.00 0.28 0.07VXO 0.20 0.08 0.02 0.02 0.16 0.17 0.08 0.17 0.23 0.28 1.00 0.10Volume 0.29 0.25 0.38 -0.24 0.14 0.20 0.02 0.16 0.03 0.07 0.10 1.001254.2. Macroeconomic Attention IndicesTable 4.3, Panel A also provides index means by day of the week. TheSaturday edition of wsj generally has less coverage of macro fundamentalsthan other days of the week. For nyt, the Saturday edition appears to haveroughly similar content to other days, while the large Sunday edition offersmore coverage than other days. While the effects of weekend news coverageare interesting and potentially important, for simplicity in the remainder ofour analysis we discard all non-trading days (weekends and holidays). Toaccount for potential day-of-the weak seasonalities in news coverage, all ofour empirical results use day-of-the-week dummy variables.Figure 4.2 plots the attention indices. For reference, each attention in-dex is associated with a series of macroeconomic fundamentals that seemsrelevant.63 For example, the output growth attention index is plotted onthe same axes with the log quarter-to-quarter growth in real gdp. The fulllist of attention indices versus the associated macroeconomic fundamentalsplotted in Figure 4.2 is given in Table 4.2.63This approach follows Carroll (2003), who plots a monthly news count index of in-flation from the New York Times and the Washington Post against cpi, from 1981 to2001.1264.2. Macroeconomic Attention IndicesFigure 4.2: Macro Attention and Macroeconomic FundamentalsThis figure shows the monthly macroeconomic attention indices (mai) for the Wall StreetJournal (mai-wu) and the New York Times (mai-nu) against related monthly macroeco-nomic fundamentals described in Table 4.2. The blue line represents a macroeconomicattention index (left y-axis) and the red dotted line (right y-axis) the mai related macroe-conomic fundamental (see Table 4.2). The units are in percentage. The gray vertical barsare nber recessions.1980 1984 1988 1992 1996 2000 2004 2008 20120.00.51.01.52.02.53.0CreditRatingMAIMAI-NU (NYT)1980 1984 1988 1992 1996 2000 2004 2008 20120.00.51.01.52.02.53.03.54.0MAI-WU (WSJ)Credit Rating MAI Corporate Relative Spread010203040506070010203040506070CorporateRelativeSpread1980 1984 1988 1992 1996 2000 2004 2008 20120.00.20.40.60.81.01.21.4GDPMAIMAI-NU (NYT)1980 1984 1988 1992 1996 2000 2004 2008 20120.51.01.52.02.53.03.5MAI-WU (WSJ)GDP MAI Real GDP Quarterly Growth Rate−3−2−10123−3−2−10123RealGDPQuarterlyGrowthRate(%)1980 1984 1988 1992 1996 2000 2004 2008 20120.00.20.40.60.81.01.21.41.61.8HousingMAIMAI-NU (NYT)1980 1984 1988 1992 1996 2000 2004 2008 201201234567MAI-WU (WSJ)Housing MAI Log Nominal Home Price Return−1.5−1.0−0.50.00.51.01.5−1.5−1.0−0.50.00.51.01.5LogHomePriceReturn1274.2. Macroeconomic Attention Indices1980 1984 1988 1992 1996 2000 2004 2008 20120.00.51.01.52.02.53.0InflationMAIMAI-NU (NYT)1980 1984 1988 1992 1996 2000 2004 2008 20121234567MAI-WU (WSJ)Inflation MAI Change in CPI−1.5−1.0−0.50.00.51.01.5−1.5−1.0−0.50.00.51.01.5ChangeinCPI(%)1980 1984 1988 1992 1996 2000 2004 2008 20120.00.20.40.60.81.0InterestMAIMAI-NU (NYT)1980 1984 1988 1992 1996 2000 2004 2008 20120.00.51.01.52.02.53.03.54.0MAI-WU (WSJ)Interest MAI Fed Funds Rate0510152005101520FedFundsRate(%)1980 1984 1988 1992 1996 2000 2004 2008 20120.00.51.01.52.02.53.03.54.04.5OilMAIMAI-NU (NYT)1980 1984 1988 1992 1996 2000 2004 2008 20120246810MAI-WU (WSJ)Oil MAI Oil Log Price2.02.53.03.54.04.55.02.02.53.03.54.04.55.0OilLogPrice1284.2. Macroeconomic Attention IndicesWe emphasize several properties of the attention indices. First, the in-dices do not appear to be driven by a single factor. They are imperfectlycorrelated, and over time attention shifts across different fundamentals. Sec-ond, attention is highly persistent. All series show fluctuations that last overperiods at least as long as several years, including both gradual trends andsharp changes. Third, the indices also show cycles at a range of higher fre-quencies, including short bursts of attention. Finally, attention seems to beat least loosely related to underlying fundamentals. This is seen most clearlyin the plot for employment, where broad patterns in attention seem to matchclosely with the level of the unemployment rate. We now investigate eachof these aspects of the plots using statistical analyses.Table 4.3 shows daily (Panel B) and monthly (Panel C) correlationsamong the composite attention indices mai-c1, as well as correlations withother series of interest: implied volatility (vxo) from the Chicago BoardOptions Exchange (cboe)64, economic policy uncertainty (epu) from Baker,Bloom, and Davis (2015)65, detrended S&P 500 trade volume (Volume) fromthe Center for Research in Security Prices (crsp), and lagged values ofthe vxo and Volume. The results confirm the imperfect correlation of theattention indices. In daily data, the highest inter-mai correlations mai arebetween monetary and inflation (0.45), monetary and interest rates (0.57),oil and inflation (0.31), us dollar and oil (0.37), and inflation and interestrates (0.34). Not all correlations are positive. For example, in monthly datathe mai for gdp and inflation are negatively correlated (-0.14) and creditrating and inflation (-0.18). We also are interested in correlations betweenthe attention indices and other variables. In the monthly data, the highestcorrelations with epu are unemployment (0.35), credit rating (0.28), andmonetary (0.15). The highest correlations with vxo are us dollar (0.33),credit rating (0.32), and unemployment (0.32).To address stationarity, we estimate ar (p) models for each attentionindex from monthly data. Following Campbell and Yogo (2006), we usethe lag length that minimized the Bayesian information criteria (bic). Theminimum bic for all of our mai occurs at four lags or less. Table 4.4 showsthese ar estimates, controlling for monthly fixed-effects. The table alsoreports Dickey-Fuller p-values for the null hypothesis that each series has aunit root. The df statistics reject the presence of unit roots except for theus dollar mai.6664Data source: https://www.cboe.com/micro/vix/historical.aspx.65The data is available at http://www.policyuncertainty.com/.66The us dollar mai-c2 rejects the unit root with a p-value of 0.09.1294.2. Macroeconomic Attention IndicesTable 4.4: Persistence of Macroeconomic AttentionPanel A of this table presents ar (p) models of the monthly de-meaned macroeconomic attention composite indices (mai-c1), controllingfor monthly time-fixed effects. df (p-value) are the p-values for the Dickey-Fuller (df) statistics that test the null of a unit root in each time series.Panel B reports the estimates from an ols regression of the daily demeanedmacroeconomic attention composite indices (mai-c1) on various moving av-erage lags of itself. L1 corresponds to the lag of itself and L5, L21, L62,L250, and L1000 are the moving average for 5, 21, 62, 250, and 1000 dayspreceding the observed values at time t. We control for day-of-week fixedeffects. The standard errors are reported in parenthesis and are calculatedusing Newey-West standard errors (10 lags). Obs. stands for the number ofobservations. *, **, and *** denote the statistical significance at the 10%,5%, 1% levels, respectively.Panel A: Monthly MAI-C1 AR(4) coefficients and DF statisticsCredit Rating GDP Housing Inflation Interest Monetary Oil Unemp. USDconst 0.01 0.03 -0.02 0.09** 0.02 0.07 0.14* 0.01 -0.02(0.03) (0.03) (0.04) (0.04) (0.03) (0.05) (0.08) (0.04) (0.03)AR(1) 0.70*** 0.25*** 0.47*** 0.51*** 0.58*** 0.50*** 0.71*** 0.62*** 0.69***(0.08) (0.04) (0.10) (0.05) (0.05) (0.04) (0.05) (0.06) (0.06)AR(2) -0.02 0.29*** 0.10 0.21*** 0.17** 0.13** 0.17*** 0.17*** 0.06(0.10) (0.04) (0.08) (0.04) (0.07) (0.05) (0.04) (0.05) (0.06)AR(3) -0.01 0.30*** 0.29*** 0.05 -0.00 0.15** 0.02 0.11** 0.01(0.07) (0.05) (0.10) (0.05) (0.06) (0.07) (0.08) (0.05) (0.05)AR(4) 0.15** 0.08 0.01 0.10** 0.10** 0.04 0.01 0.01 0.18***(0.07) (0.05) (0.06) (0.05) (0.05) (0.05) (0.04) (0.04) (0.04)DF (p-value) 0.00 0.02 0.04 0.00 0.01 0.00 0.00 0.00 0.13Adj-R2 0.58 0.70 0.63 0.67 0.62 0.54 0.79 0.78 0.82Obs. 415 415 415 415 415 415 415 415 415Panel B: Daily MAI-C1 regressions on lagged attentionCredit Rating GDP Housing Inflation Interest Monetary Oil Unemployment U.S. Dollarconst -0.09*** 0.08** -0.21*** 0.09** -0.04 -0.11** -0.21*** 0.04 -0.08***(0.02) (0.04) (0.03) (0.05) (0.03) (0.04) (0.05) (0.04) (0.02)L1 0.07*** 0.05*** 0.06** 0.03** 0.12*** 0.17*** 0.06*** 0.00 -0.01(0.02) (0.01) (0.03) (0.01) (0.02) (0.02) (0.02) (0.02) (0.02)L5 0.28*** 0.11*** 0.56*** 0.13*** 0.16*** 0.19*** 0.38*** 0.23*** 0.18***(0.05) (0.03) (0.06) (0.03) (0.03) (0.03) (0.05) (0.04) (0.04)L21 0.44*** -0.01 0.05 0.30*** 0.24*** 0.23*** 0.36*** 0.22*** 0.51***(0.07) (0.07) (0.09) (0.06) (0.07) (0.05) (0.05) (0.07) (0.07)L62 0.02 0.41*** 0.12** 0.34*** 0.18** 0.12* 0.13*** 0.30*** 0.13*(0.07) (0.10) (0.06) (0.07) (0.09) (0.07) (0.05) (0.08) (0.08)L250 0.12* 0.43*** 0.20** 0.09 0.25*** 0.23*** 0.03 0.26*** 0.19***(0.06) (0.10) (0.08) (0.06) (0.07) (0.08) (0.03) (0.07) (0.06)L1000 0.02 -0.04 -0.01 0.03 -0.01 0.01 0.02 -0.09*** -0.04(0.06) (0.06) (0.05) (0.05) (0.04) (0.05) (0.02) (0.03) (0.03)Obs. 8109 8109 8109 8109 8109 8109 8109 8109 8109Adj-R2 0.29 0.15 0.43 0.17 0.23 0.26 0.54 0.32 0.411304.2. Macroeconomic Attention IndicesTo further explore time-series dependence, Figure 4.3 shows autocorre-lation plots of each composite series mai-c1 for lag lengths from 1 to 250trading days. We plot the autocorrelations for residuals after controllingfor day-of-the-week dummies and month-of-the-year dummies. The plotsshow very slow decay in this range of frequencies, and the autocorrelationsare significantly larger than zero at 250 lags for all series. Several of theautocorrelation plots show apparent cycles in dependence. For example,gdp shows strong increases in correlations at each monthly interval. Otherseries (housing, us dollar) have increases in autocorrelations at weekly in-tervals. These cycles are consistent with the importance of periodic newsannouncements.To account for potential long-memory dependence as well as multiplecycles in news variation, we use regressions that aggregate the attentionindices over different horizons similarly to midas regression (see Ghysels,Santa-Clara, and Valkanov, 2006). Specifically, we construct simple movingaverages of the attention indices over window sizes of 1 day, 5 days, 21days (monthly), 62 days (quarterly), and 250 days (annual), and 1000 days(business cycle).Panel B of Table 4.4 shows results of regressing each attention index onlagged simple moving averages of its own history, for the full set of differ-ent window sizes. All of the series show persistence at multiple frequencies,with the majority having significant positive persistence in daily, weekly,monthly, quarterly, and annual-length moving averages in the multiple re-gression framework.One exception is credit rating attention, which does not show significant per-sistence beyond monthly horizons. A separate monthly cycle is not presentin gdp attention, although it does show significant persistence at all othercycle lengths between daily and annual. This result seems intuitive giventhe quarterly reporting cycle for gdp growth. These results are consistentwith slow, approximately hyperbolic decay in the persistence of attentionto each of the fundamental factors. The presence of multiple frequencies inattention to financial news are also broadly consistent with the motivationand theoretical framework in Calvet and Fisher (2007), who hypothesizefractal patterns in news about the fundamentals impacting asset prices. Wenext determine whether the fluctuations of the individual attention indicescan be related to macroeconomic fundamentals.1314.3. Attention and Macroeconomic FundamentalsFigure 4.3: Autocorrelation in Macroeconomic AttentionThis figure shows the autocorrelations (ρk) for residuals after controlling for day-of-the-week dummies and month-of-the-year dummies for each of the composite macroeconomicattention index mai-c1 for k lags ranging from 1 to 250 trading days. The dashed linerepresents the 95% critical value for the test ρk ≤ 0, where we use the “large-lag” standarderrors of Anderson (1976). These standard errors account for the observed autocorrelationsfor lags less than k.0.00.10.20.30.40.50.60.7Credit Rating GDP Housing0.00.10.20.30.40.50.60.7Inflation Interest Monetary0 50 100 150 200 250Lags0.00.10.20.30.40.50.60.7Oil0 50 100 150 200 250LagsUnemployment0 50 100 150 200 250LagsUS Dollar4.3 Attention and Macroeconomic FundamentalsIntuition suggests that high frequency fluctuations in attention could bedriven by economic announcements, while lower frequency variations might1324.3. Attention and Macroeconomic Fundamentalsbe related to movements in economic fundamentals. We test these ideas.4.3.1 Macroeconomic AnnouncementsPrior literature has established links between economic announcements andreturns and volatility for the foreign exchange and stock market (Ander-sen, Bollerslev, Diebold, and Vega, 2003a, 2007a). We now investigatethe relationship between macroeconomic announcements and attention tomacroeconomic fundamentals. Attention could be limited to simply report-ing on announcements. Alternatively, attention might be high in advance ofannouncements as news media strive to anticipate the content of announce-ments, or to put the potential outcomes of an announcement into a broadercontext for the benefit of their readers.Cross-sectionally, our analysis can tell us which types of announcementshave the largest impacts on macroeconomic attention. If the media play animportant role in the transmission of economic news, then understanding theallocation of media resources to covering different types of announcementsshould be informative about which announcement matters most to readers.The economic announcements we consider are: consumer price index(cpi), employment situation, and Federal Open Market Committee (fomc)announcements. The announcement dates span the entire sample lengthof our indices. The cpi, and employment situation announcement dates arefrom the Bureau of Labor Statistics and fomc announcement dates are fromthe Federal Reserve Board. Macroeconomic attention can be influenced bymultiple announcements, hence we study the most intuitive links betweenthe macroeconomic attention indices and macroeconomic announcements asshown in Table 4.2. The specification we use is:MAI-C1df,t = α+δ=4∑δ=−4βδAnnj,t+δ + t (4.4)where MAI-C1df,t is the composite index mai-c1 detrended by its own 60-day simple moving average. The variables Annj,t+δ are equal to 1 if there isan announcement on day-t+ δ, 0 otherwise, and we let δ take integer valuesfrom -4 to 4. Since the model specification contains many variables we showthe regression coefficients, βδ and their 95 percent confidence intervals inFigure 4.4. In the first row, attention to inflation increases leading up tothe cpi announcement, and the index is at its highest one day after theannouncement. cpi announcements also raise attention more moderately inthe monetary and oil attention indices.1334.3. Attention and Macroeconomic FundamentalsFigure 4.4: Macroeconomic Attention around Macroeconomic Announce-mentsThis figure shows the lag and forward estimated coefficients βδ from an ols regression ofdetrended macroeconomic attention indices mai-c1 on announcement dummies as specifiedin Equation (4.4). The shaded area corresponds to the 95% confidence interval around theestimated coefficients. The x-axis is the number days since the announcement. The firstrow shows attention around the consumer price index (cpi) announcements, the secondrow the Employment situation announcements, and the third row the Federal Open MarketCommittee (fomc) announcements for different mai-c1. The vertical line represents theday of the announcement.−4 −3 −2 −1 0 1 2 3 4−0.20.00.20.40.6ResponseCPI Ann. on Inflation MAI−4 −3 −2 −1 0 1 2 3 4CPI Ann. on Monetary MAI−4 −3 −2 −1 0 1 2 3 4CPI Ann. on Oil MAI−4 −3 −2 −1 0 1 2 3 4−0.4−0.20.00.20.40.60.8ResponseEmpl. Ann. on Inflation MAI−4 −3 −2 −1 0 1 2 3 4Empl. Ann. on Monetary MAI−4 −3 −2 −1 0 1 2 3 4Empl. Ann. on Unemployment MAI−4 −3 −2 −1 0 1 2 3 4Days relative to announcement−0.50.00.51.01.52.0ResponseFOMC Ann. on Monetary MAI−4 −3 −2 −1 0 1 2 3 4Days relative to announcementPre-1994 on FOMC Monetary MAI−4 −3 −2 −1 0 1 2 3 4Days relative to announcementPost-1994 FOMC on Monetary MAIFor unemployment announcements (second row), macroeconomic atten-tion increases two days in advance of the announcement, spikes on the an-1344.3. Attention and Macroeconomic Fundamentalsnouncement day, and remains high for two days after the announcement.Unemployment announcements do not impact other mai, such as inflationand monetary.fomc announcements (the third row) have moderate impacts on the at-tention index associated with monetary policy in the full sample. However,a subsample analysis shows that the effects are indistinguishable prior to1994, when policy actions were not publicly announced. After 1994 whenthe fomc started public announcements of the policy action, the pattern inattention becomes more pronounced. Boguth, Carlson, Fisher, and Simutin(2016) use our monetary policy attention index and show that times wheninvestors expect important decisions from the Federal Open Market Com-mittee, attention is high prior to committee meeting.4.3.2 Macroeconomic FundamentalsBeyond the link between economic announcements and daily spikes in at-tention, what accounts for the lower-frequency fluctuations in the attentionindices? Figure 4.1 and 4.2 suggests attention dynamics could reflect chang-ing economic conditions.Prior literature has attempted to establish links between macroeconomicvariables and financial market variables such as volatility (Schwert, 1989).We expect that macroeconomic attention connects economic news with fi-nancial markets, serving an intermediary function. A benefit of measuringmacroeconomic attention is that we can measure not just aggregate interestin financial and economic news, we can also tell what writers are talkingabout. Hence the low frequency variations in our different mai should pickup changing patterns in concerns for different macroeconomic fundamentals.To study how variations in macroeconomic fundamentals impact macroe-conomic attention, we decompose the macro variables into detrended movingaverages over different window sizes. That is, given a particular macroeco-nomic fundamental Ft (e.g., unemployment rate, change in log cpi, changein log house price index), we can decompose the fundamental into a set ofdetrended moving averages:Ft ≡ (Ft−F t,t−2) + (F t,t−2−F t,t−11) + (F t,t−11−F t,t−47) +F t,t−47, (4.5)where F t,t−k is the simple moving average of the fundamental from t− k tot. The components on the right hand side of the equation, each in paren-theses, are detrended moving averages over window sizes that are expandingapproximately geometrically. These could be capable of capturing the low-frequency patterns in autocorrelations documented for the attention indices1354.3. Attention and Macroeconomic Fundamentalsin Table 4.4. We regress the monthly attention indices on these detrendedmoving averages and their squared values:(4.6)MAIf,t = α+ β1(Ft − Ft,t−2) + β2(Ft − Ft,t−2)2 +β3(Ft,t−2 − Ft,t−11) + β4(Ft,t−2 − Ft,t−11)2 +β5(Ft,t−11 − Ft,t−47) + β6(Ft,t−11 − Ft,t−47)2 + t.Table 4.5 reports results for regression (4.6) for the nyt (Panel A) andwsj (Panel B) indices. The results show generally that attention respondsto changes in macro fundamentals. Adjusted R2 range from 0 to over 50%,with most of the regressions having at least one significant coefficient onfundamentals.1364.3.AttentionandMacroeconomicFundamentalsTable 4.5: Macroeconomic Attention and Macroeconomic FundamentalsThis table presents the results of an ols regression of monthly macroeconomic attention indices (mai) on differentmacroeconomic fundamentals. Panel A and Panel B report the results for the New York Times macroeconomicattention indices (mai-nu) and the Wall Street Journal (mai-wu) respectively. The general regression is specifiedin equation 4.6. F corresponds to the associated fundamental to each mai as described in Table 4.2 and Ft isthe moving average over t days of the respective macroeconomic fundamental. We control for monthly fixedeffects. The standard errors are reported in parenthesis and are calculated using Newey-West standard errors(10 lags). Obs. stands for the number of observations. *, **, *** denote the statistic significance at the 10%,5%, 1% levels, respectively.Panel A: MAI-NU (New York Times)MAI: Credit Rating GDP Housing Inflation Interest Monetary Oil Unemployment US DollarF: Credit Rating Spreads GDP Growth Home Price Ret ∆ CPI Fed Fund Fed Fund Oil Price Ret Unemp. Rate USD Index RetFt − Ft,t−2 0.022 -0.221* -0.171** -0.020 -0.022 -0.003 0.034 0.000(0.014) (0.122) (0.068) (0.018) (0.035) (0.004) (0.155) (0.001)Ft,t−2 − Ft,t−11 -0.001 0.059* -0.317*** -0.533*** 0.004 -0.010 0.005 0.063 -0.001(0.004) (0.031) (0.110) (0.163) (0.013) (0.034) (0.009) (0.091) (0.004)Ft,t−11 − Ft,t−47 -0.011 0.154 -0.013 0.641 -0.019*** -0.041* 0.044* 0.140*** -0.020(0.012) (0.100) (0.107) (0.758) (0.006) (0.021) (0.024) (0.048) (0.012)(Ft − Ft,t−2)2 0.000 0.538*** -0.476*** 0.030*** 0.059*** 0.002*** 0.632 0.000(0.001) (0.117) (0.170) (0.007) (0.017) (0.001) (0.737) (0.001)(Ft,t−2 − Ft,t−11)2 -0.000 0.055 0.242*** -0.260 0.014** 0.048*** 0.003*** 0.229** -0.004*(0.000) (0.039) (0.086) (0.177) (0.006) (0.014) (0.001) (0.104) (0.002)(Ft,t−11 − Ft,t−47)2 0.001 0.190 0.413** 6.503*** 0.007*** -0.005 -0.007 0.066*** -0.016(0.001) (0.150) (0.202) (2.207) (0.002) (0.008) (0.006) (0.025) (0.012)const 0.189*** 0.416*** 0.004 0.644*** 0.187*** 0.819*** 0.488*** 0.559*** 0.068***(0.038) (0.057) (0.043) (0.078) (0.026) (0.067) (0.083) (0.065) (0.018)Obs. 419 125 419 419 419 419 376 419 419Adj-R2 0.05 0.06 0.35 0.15 0.16 0.09 0.28 0.51 -0.001374.3.AttentionandMacroeconomicFundamentalsPanel B: MAI-WU (Wall Street Journal)MAI: Credit Rating GDP Housing Inflation Interest Monetary Oil Unemployment US DollarF: Credit Rating Spreads GDP Growth Home Price Ret ∆ CPI Fed Fund Fed Fund Oil Price Ret Unemp. Rate USD Index RetFt − Ft,t−2 0.053** -0.272 -0.259 -0.280 -0.488 -0.016 -0.193 0.007(0.023) (0.302) (0.185) (0.242) (0.361) (0.011) (0.268) (0.013)Ft,t−2 − Ft,t−11 0.024** 0.176 -0.680*** 0.704 0.161 0.198 0.016 0.141 -0.022(0.012) (0.120) (0.256) (0.444) (0.163) (0.241) (0.020) (0.247) (0.042)Ft,t−11 − Ft,t−47 0.022 0.294 -0.268 4.609*** 0.132 0.129 0.172* 0.241** -0.362***(0.023) (0.293) (0.318) (1.321) (0.090) (0.117) (0.099) (0.103) (0.136)(Ft − Ft,t−2)2 -0.002 0.486 -0.274 0.571 0.162 0.006*** 3.176** 0.016**(0.003) (0.479) (0.358) (0.640) (0.826) (0.001) (1.413) (0.008)(Ft,t−2 − Ft,t−11)2 0.001 0.315** 0.672*** 1.139** 0.362*** 0.343* 0.007*** 0.202 0.055**(0.001) (0.147) (0.236) (0.455) (0.123) (0.177) (0.001) (0.183) (0.022)(Ft,t−11 − Ft,t−47)2 0.001 0.399 2.393*** 12.976** 0.075** 0.070 -0.003 0.082* 0.295**(0.001) (0.454) (0.458) (6.190) (0.038) (0.065) (0.019) (0.043) (0.148)const 0.558*** 1.740*** 0.142 3.015*** 1.032*** 2.364*** 2.728*** 1.866*** 0.829***(0.084) (0.121) (0.106) (0.105) (0.110) (0.183) (0.359) (0.133) (0.159)Obs. 376 125 376 376 376 376 376 376 376Adj-R2 0.11 0.06 0.47 0.19 0.13 0.03 0.08 0.33 0.141384.3. Attention and Macroeconomic FundamentalsTo help synthesize the results, we first focus on aspects that are similaracross Panels A and B, or across attention in both the nyt and wsj. Con-firming the idea that change raises attention, many of the coefficients onsquared changes in fundamentals are significant and positive in both pan-els. For the nyt, of the fifteen significant coefficients on squared changesin fundamentals, thirteen are positive. For the wsj, all fifteen of the fifteensquared changes on fundamentals are positive. These results are consis-tent with theories where changes in fundamentals raise attention, such as inAndrei and Hasler (2014, 2016).A second intuitive idea is that for a given magnitude of the absolutechange, attention will be higher when the change is in a direction that isassociated with “bad” versus “good” times. Focusing on the significantcoefficients on signed changes in fundamentals, many of the series show con-sistent results across the nyt and wsj in the intuitive direction suggestingthat bad news raises attention: Attention to credit rises when relative creditspreads rise; attention to housing rises when house prices fall; attention tounemployment rises when unemployment increases.We also see interesting differences across the wsj and nyt attentionindices. In general, the R2 for the wsj attention index regressions onfundamentals are higher than for the nyt. One notable exception is un-employment. More than 50% of the variation of the nyt attention indexis explained by movements in the unemployment rate, consistent with thevery strong comovement apparent in Figure 1, compared to the lower R2 of33% for explaining wsj attention to unemployment. Why do unemploymentfundamentals have less explanatory power for wsj attention than for nyt at-tention? Examining the plots in Figure 1, the nyt has shown a consistentlypositive relation between unemployment and attention to unemployment.For the wsj, in the 1980’s and 1990’s attention moved almost inverselywith the unemployment level. Starting in the 2000’s and certainly by thefinancial crisis, wsj coverage of unemployment began to comove positivelywith changes in unemployment, similar to the nyt. This is consistent withthe idea that the readership and editorial policy of the nyt have been moreconsistently focused on unemployment than the wsj over time; however, fol-lowing the financial crisis, the wsj became more attentive to unemploymentin a manner similar to nyt.67Consistent with this idea of different focuses and audiences between thenyt and wsj, we also see a difference in how inflation impacts attention.67Another contributing factor could be the retirement of conservative editor RobertBartley, who retired from the wsj in 2000 after serving for thirty years.1394.4. Attention and Stock Market ActivityAn increase in inflation tends to raise attention to inflation at the wsj, butreduces attention at the nyt. This is again consistent with the idea that thewsj tends to be more politically conservative and associated with monetaristviews on inflation than the nyt, which tends towards more Keynesian viewson the economy.4.4 Attention and Stock Market ActivityBeber, Brandt, and Kavajecz (2011) conjecture that market participants arecontinually digesting news about the macroeconomy, which impacts theirpreferences, expectations, and risk tolerances. As a result, macroeconomicnews induce them to trade. The authors show that market trade volume seg-mented by economic sectors contain important macroeconomic informationand in turn predict important macroeconomic announcements.We study the link between daily macroeconomic attention and stockmarket activity. Let V lmdt be the logarithm of the daily aggregate tradevolume of S&P 500 firms, detrended by its own 60-day moving average,following Tetlock (2007). We run the regression:(4.7)V lmdt = αf + βfMAI5−20,f,t + γfAnnt + δfAnnt ·MAI5−20,f,t + f,t,where MAI5−20,tt is the difference between the five-day and twenty-day mov-ing average of mai-c1 to macro fundamental f . Annj,t is equal to 1 if thereis an announcement on day-t, zero otherwise.68Table 4.6 shows that for all mai, rising attention is associated with anincrease in market volume. When we include macro announcements in theregressions, many of the announcements have significant impacts on volume,but the inclusion of these variables does not alter inferences about the im-portance of attention. Interaction terms do not have a consistent sign, anddo not alter inference about the effects of attention or announcements ontrading volume.68To simplify the analysis, we do not differentiate between all gdp announcements(advance, preliminary, and final).1404.4. Attention and Stock Market ActivityTable 4.6: Media Attention and Aggregate Trade VolumeThis table presents the results of an ols regression of the daily detrendedS&P 500 trade volume on the difference between the 5-day and 20-day mov-ing average mai-c1 and a dummy (Ann) equal to one if there is a relatedannouncement specified in Table 4.2, zero otherwise. We detrend the logtrade volume using the moving average of the log trade volume of the past60 trading days. For all model specifications, we control for day-of-weekfixed effects. The standard errors are reported in parenthesis and are cal-culated using Newey-West standard errors (250 lags). Obs. stands for thenumber of observations. *, **, *** denote the statistical significance at the10%, 5%, 1% levels, respectively.MAI: Inflation Monetary InterestAnn: CPI FOMC FOMCMAI5−20 0.052*** 0.051*** 0.056*** 0.066*** 0.065*** 0.066*** 0.058*** 0.057*** 0.058***(0.009) (0.009) (0.009) (0.008) (0.008) (0.008) (0.013) (0.013) (0.013)Ann 0.034*** 0.043*** 0.026*** 0.027*** 0.030*** 0.031***(0.007) (0.007) (0.009) (0.010) (0.009) (0.009)MAI5−20×Ann -0.104*** -0.011 -0.043(0.024) (0.035) (0.039)const 0.002 0.000 0.001 0.002 0.002 0.002 0.002 0.002 0.002(0.006) (0.006) (0.006) (0.006) (0.006) (0.006) (0.006) (0.006) (0.006)Obs. 8787 8787 8787 8787 8787 8787 8787 8787 8787Adj-R2 0.06 0.06 0.06 0.07 0.07 0.07 0.05 0.05 0.05MAI: GDP Unemployment Credit Rating Oil USDAnn: GDP Report EmploymentMAI5−20 0.027*** 0.027*** 0.026*** 0.030*** 0.029*** 0.030*** 0.068*** 0.026*** 0.075***(0.010) (0.010) (0.010) (0.010) (0.010) (0.010) (0.018) (0.010) (0.019)Ann 0.005 0.003 0.013 0.018(0.008) (0.008) (0.011) (0.013)MAI5−20×Ann 0.035 -0.031(0.036) (0.034)const 0.002 0.002 0.002 0.002 -0.000 -0.000 0.002 0.013** 0.028***(0.006) (0.006) (0.006) (0.006) (0.007) (0.007) (0.006) (0.007) (0.006)Obs. 8787 8787 8787 8787 8787 8787 8787 7368 8321Adj-R2 0.05 0.05 0.05 0.05 0.05 0.05 0.05 0.05 0.061414.5. Using Attention for ForecastingAnother way to look at the impact of macroeconomic attention on stockmarket activity is to investigate the relationship between macroeconomic at-tention and implied volatility, measured by the vxo index, which is availablebeginning in 1986. We implement the following regression for each attentionindex:V XOt = αf + βfMAI20−250,f,t + γfAnnt + δfAnnt ·MAIf,20−250,t + f,t(4.8)Table 4.7 shows that increases in macroeconomic attention on interest rates,gdp, unemployment, credit ratings and usd positively relate to increases inimplied volatility. The R2 are highest for unemployment (13%) and gdp(7%). Results are similar if we detrend vxo using a 250-day moving average.Thus, controlling for macroeconomic announcements, increases in attentionis associated with an increase in both aggregate volume and volatility.Overall the results of this section provide strong evidence that increasesin attention to macro fundamentals is positively correlated with the aggre-gate stock market activities.4.5 Using Attention for ForecastingGiven the links between media attention and macroeconomic fundamen-tals, it is natural to consider whether media attention might help to predictfundamentals on macroeconomic announcements. We are particularly in-terested to understand the link between the mai to unemployment and theemployment situation announcements and the mai to monetary policy andfomc announcements. Our decision to focus on unemployment is partlymotivated by the plots in Figure 4.1 which suggest that the unemploymentattention indices might act as a leading indicator, and partly motivated byfindings in prior literature that the unemployment report is important forstock market returns (Boyd et al., 2005). We also ask whether attention tomonetary policy can forecast the stock returns, change in implied volatil-ity, and the Fed fund rate on fomc announcements. Lucca and Moench(2015b) show that a significant fraction of the risk premium is earned onfomc announcements. Savor and Wilson (2013b) further show that impliedvolatility significantly decrease on fomc announcements.4.5.1 Unemployment AnnouncementsWe construct measures of “surprises” in the monthly employment report intwo ways. First, we consider a simple random walk model of unemploy-1424.5. Using Attention for ForecastingTable 4.7: Media Attention and Implied VolatilityThis table presents the results of an ols regression of the daily impliedvolatility proxied by vxo regressed on the difference between the 20-day and250-day moving average mai-c1 and a dummy (Ann) equal to one if there isa related announcement specified in Table 4.2, zero otherwise. For all modelspecifications, we control for day-of-week fixed effects. The standard errorsare reported in parenthesis and are calculated using Newey-West standarderrors (250 lags). Obs. stands for the number of observations. *, **, ***denote the statistical significance at the 10%, 5%, 1% levels, respectively.MAI: Inflation Monetary InterestAnn: CPI FOMC FOMCMAI20−250 -2.730 -2.729 -2.750 3.443** 3.442** 3.448** 4.709* 4.708* 4.727*(3.362) (3.362) (3.335) (1.600) (1.599) (1.601) (2.606) (2.606) (2.606)Ann 0.259 0.266 -0.205 -0.207 -0.244 -0.246(0.182) (0.184) (0.224) (0.225) (0.237) (0.240)MAI20−250×Ann 0.438 -0.213 -0.591(0.764) (0.569) (1.112)const 20.720*** 20.703*** 20.703*** 20.722*** 20.722*** 20.722*** 20.732*** 20.733*** 20.733***(1.231) (1.227) (1.226) (1.249) (1.249) (1.249) (1.257) (1.257) (1.258)Obs. 7386 7386 7386 7386 7386 7386 7386 7386 7386Adj-R2 0.01 0.01 0.01 0.02 0.02 0.02 0.01 0.01 0.01MAI: GDP Unemployment Credit Rating Oil USDAnn: GDP Report EmploymentMAI20−250 11.370** 11.377** 11.398** 11.079*** 11.080*** 11.103*** 7.603*** 0.511 6.786**(4.613) (4.614) (4.600) (4.075) (4.074) (4.079) (2.898) (1.148) (2.654)Ann 0.286 0.279 0.207 0.206(0.200) (0.199) (0.153) (0.156)MAI20−250×Ann -0.420 -0.475(1.168) (0.761)const 20.650*** 20.628*** 20.628*** 20.645*** 20.598*** 20.598*** 20.765*** 20.762*** 20.805***(1.139) (1.135) (1.135) (1.087) (1.088) (1.088) (1.218) (1.252) (1.245)Obs. 7386 7386 7386 7386 7386 7386 7361 7361 7005Adj-R2 0.07 0.07 0.07 0.13 0.13 0.13 0.05 0.00 0.02ment, under which the prediction for the following month’s unemploymentrate is the prior month’s unemployment rate, and the surprise is definedas the change in unemployment. Second, we use the regression model ofBoyd, Hu, and Jagannathan (2005) to generate the unemployment forecasts,which we call the Boyd, Hu, and Jagannathan (2005) surprise. The authors’forecasting model uses information from related macroeconomic variables,including industrial production, T-bill rate, corporate bond yield spreads,and past unemployment rate. The surprise is defined as the difference be-tween the announced unemployment rate and the unemployment forecast.1434.5. Using Attention for ForecastingThe date of reference for the actual unemployment rate is the release date ofthe employment situation announcement made by the U.S. Bureau of LaborStatistics.For predictor variables, we carry out separate analyses using detrendedlevels of the composite indices mai-c1. Specifically, to capture very short runmovements, we use the difference between the 5-day simple moving averageand the 20-day simple moving average of the attention indices (mai 5−20).To capture a range of other movements, we similarly calculate 5-, 20-, and60-day moving averages detrended by the 252-day moving average (i.e., mai5−252, mai 20−252, mai 60−252). Following Boyd et al. (2005), we also interacteach of the predictor variables with nber recession dummies. Since the nberdummies are not known in advance, regressions using these interactionsare not predictive. Boyd et al. (2005) hypothesize that “bad news” forunemployment means different things in expansions and contractions, andthe interaction variables allow us to see whether the predictive ability ofattention, if it exists, concentrates in contractions.To investigate the link between unemployment surprises and our atten-tion index to unemployment, we estimate the following regression:Surpt = c+MAIt−1 +MAIt−1 ·NBER+ +et, (4.9)where Rett is the daily return of S&P 500 index, MAIt−1 is the detrendedmai-c1 for unemployment, NBER is an indicator variable for NBER reces-sion, and Surpt is unemployment announcement surprise.69Table 4.8 shows that the detrended unemployment attention variablesare significantly related to surprises in the unemployment report, and thatthe interaction variables are often important. Under the random walk model,attention indices positively predict future surprises in unemployment, andvariables are significant when interacted with the nber recession dummies.Hence, increases in macroeconomic attention to unemployment positivelypredict future changes in unemployment, and this relationship is strongduring recessions. Changes in macroeconomic attention retain the abilityto explain future changes in employment relative to the Boyd et al. (2005)regression model.69When the Employment Situation announcement occurs on Good Friday (U.S. holiday)we use the stock return on the following trading if the market is close.1444.5. Using Attention for ForecastingTable 4.8: Unemployment Surprise Forecasts on Employment Situation An-nouncement DaysThis table presents the results of an ols regression of the unemploymentsurprise regressed on the one-day lag detrended demeaned daily compositemai-c1 for unemployment at different frequencies and an interaction termbetween mai-c1 and an nber dummy. For example, MAI5−20 is the dif-ference between the five-day and twenty-day moving average of mai-c1 forunemployment. The nber dummy equals one if the unemployment surpriseoccurs during a nber recession, zero otherwise. The surprise is calculatedas the difference between the actual unemployment for month t reported inmonth t + 1 and the random-walk (i.e. the previous month unemploymentrate) in Panel A and the forecasted unemployment rate as in Boyd, Hu,and Jagannathan (2005) in Panel B. The standard errors are reported inparenthesis and are calculated using the White’s heteroskedasticity robuststandard errors. Obs. stands for the number of observations. *, **, ***denote the statistical significance at the 10%, 5%, 1% levels, respectively.Panel A: Random-WalkMAI: MAI5−20 MAI5−250 MAI20−250 MAI60−250MAI 0.040 0.020 0.074*** 0.042** 0.142*** 0.090** 0.216*** 0.110**(0.027) (0.026) (0.019) (0.019) (0.033) (0.035) (0.045) (0.052)MAI×NBER 0.298** 0.194*** 0.183** 0.375***(0.138) (0.051) (0.080) (0.083)const -0.010 -0.010 -0.012 -0.017* -0.002 -0.009 -0.001 -0.012(0.010) (0.010) (0.009) (0.009) (0.009) (0.010) (0.009) (0.009)Obs. 418 418 407 407 407 407 407 407Adj-R2 0.00 0.02 0.04 0.08 0.06 0.07 0.07 0.11Panel B: Boyd et al. (2005) SurpriseMAI: MAI5−20 MAI5−250 MAI20−250 MAI60−250MAI 0.024 0.017 0.046*** 0.036** 0.089*** 0.078*** 0.129*** 0.092**(0.023) (0.023) (0.016) (0.017) (0.024) (0.029) (0.034) (0.043)MAI×NBER 0.106 0.065 0.040 0.134**(0.095) (0.043) (0.054) (0.064)const -0.018** -0.018** -0.020*** -0.021*** -0.013* -0.015* -0.013* -0.017**(0.008) (0.008) (0.008) (0.008) (0.007) (0.008) (0.007) (0.008)Obs. 418 418 407 407 407 407 407 407Adj-R2 0.00 0.00 0.02 0.03 0.03 0.03 0.04 0.051454.5. Using Attention for ForecastingFigure 4.5 shows graphically how attention changes before and after un-employment surprises. There are four panels, corresponding to all combina-tions of the main two unemployment surprises, and the two unemploymentattention indices. For each unemployment surprise, we separate the datainto three equal-sized bins of small, medium, and large surprises. We thenplot in event time the average attention over a period one year prior to thesurprise, out to one year subsequent to the surprise.The results show similar patterns. When the unemployment surprise isparticularly low, on average attention to unemployment in the media hasbeen declining over the past year, and continues to decline over the followingyear. Conversely, when the unemployment surprise is large and positive, onaverage attention has been increasing over the prior year, and continuesto increase over the following year. When the unemployment surprise isin the middle tercile, on average attention is approximately flat over theprior and following years, and at a lower level than for large positive ornegative surprises. These findings are consistent with the regression results,and confirm that attention moves both before and after changes in reportedfundamentals.It is natural to think that if changing attention to unemployment predictsunemployment announcement surprises, then it may also predict market re-turns on the day of the employment announcement. This topic relates toprior research by Boyd et al. (2005), who show that unemployment surprisesgenerally relate positively to market returns on the announcement date, butthe relationship turns negative during nber recessions. In Table 4.9, werevisit their results using the two different measures of unemployment sur-prise defined previously, and adding measures of macroeconomic attentionas explanatory variables. We specify:Rett = c+MAIt−1 +MAIt−1 ·NBER+Surpt+Surpt ·NBER+et. (4.10)where Rett is the daily return of S&P 500 index.The first column of Table 4.9 shows results with only the variables usedby Boyd et al. (2005). The coefficient estimates are consistent with theirresults: unemployment surprises positively relate to market returns, butthe relationship turns negative in recessions. Both the surprise and theinteraction term are significant at the 5% and 10% level.The remaining columns of Table 4.9 consider as explanatory variables,separately and with the Boyd et al. (2005) surprise as controls, measures ofchanges in attention to unemployment. The short-horizon trend in attention1464.5. Using Attention for ForecastingFigure 4.5: Attention to Unemployment around Employment SituationAnnouncementsThis figure shows the daily 60-day moving average of the unemployment attention indexfor the Wall Street Journal (mai-wu) and the New York Times (mai-nu) around theemployment situation announcements. The window is 250 trading days before and aftereach announcement. We separate the random-walk and the Boyd, Hu, and Jagannathan(2005) surprises into terciles. The mai around low surprises is in blue (solid line), mediumsurprises is in red (dotted line), and high surprises is in black (dashed line).−200 −100 0 100 200Days relative to announcement1.71.81.92.02.1MAI-WU (WSJ)−200 −100 0 100 200Days relative to announcement0.550.600.650.700.750.80MAI-NU (NYT)Random-Walk Surprises−200 −100 0 100 200Days relative to announcement1.61.71.81.92.02.12.2MAI-WU (WSJ)−200 −100 0 100 200Days relative to announcement0.500.550.600.650.700.750.80MAI-NU (NYT)Low surprises Medium surprises High surprisesBoyd et al. (2005) Surprises1474.5. Using Attention for ForecastingTable 4.9: S&P Return Forecast on Employment Situation AnnouncementDaysThis table presents the results of an ols regression of the daily S&P 500 logreturn on the employment situation announcement date regressed on theBoyd, Hu, and Jagannathan (2005) surprise (SurpBoyd) of the unemploy-ment announcement, the surprise interacted with an nber dummy, the one-day lag detrended unemployment attention index composite index mai-c1,and the detrended unemployment attention index interacted with an nberdummy. For example, MAI5−20,t is the difference between the five-day andtwenty-day moving average of mai-c1 for unemployment. The nber dummyequal one if the unemployment surprise occurs during a nber recession, zerootherwise. We show the results for two different detrended frequencies forthe unemployment attention index. The standard errors are reported inparenthesis and are calculated using the White’s heteroskedasticity robuststandard errors. Obs. stands for the number of observations. *, **, ***denote the statistical significance at the 10%, 5%, 1% levels, respectively.MAI: MAI5−20 MAI20−250MAI 0.361** 0.319** 0.295* 0.278 -0.059 -0.106(0.159) (0.160) (0.161) (0.212) (0.223) (0.221)MAI×NBER 0.617 0.800 1.177** 1.442***(0.787) (0.721) (0.514) (0.511)SurpBoyd 0.620* 0.572 0.725**(0.354) (0.352) (0.366)SurpBoyd×NBER -2.022* -2.282* -3.184**(1.229) (1.278) (1.323)const 0.052 -0.015 -0.015 0.011 0.032 -0.015 0.009(0.057) (0.061) (0.061) (0.062) (0.058) (0.060) (0.060)Obs. 419 418 418 418 407 407 407Adj-R2 0.01 0.01 0.01 0.02 0.00 0.02 0.04(5-day minus 20-day moving average) is positive and significant at the 5%level in all specifications, and remains significant with the Boyd et al. (2005)variables as controls. The medium-horizon attention trend (20-day minus250-day moving average), positively relates to the market return, but isnot significant independently. However, interacted with the nber recession1484.5. Using Attention for Forecastingdummy, the coefficients are uniformly positive and significant. The sign isopposite to the coefficient on the surprise itself interacted with the nberrecession dummy.It is important to distinguish between the trend in attention, which re-flects anticipation, and the surprise itself, which reflects a realization. Con-sistent with the results of Boyd et al. (2005), during a recession a higherrealization of unemployment on the announcement date leads to lower mar-ket returns. We add to this that rising attention before the announcementdate tends to be associated with higher market returns on the announcementdate, as uncertainty is resolved.4.5.2 FOMC AnnouncementsWe now investigate whether our attention indices to monetary policy canpredict stock returns, changes in implied volatility, and changes in Fed fundrates on fomc announcements. We focus specifically on the period post1994 when fomc decisions are publicly announced. We use a similar olsregression framework as in Equation (4.9) but using the S&P 500 returns,changes in implied volatility proxied by vxo, or changes in Fed fund ratesas dependent variables. Changes in Fed fund rate consist of a random-walksurprise measure.Table 4.10, Panel A shows that, controlling for the interaction betweennber dummies and mai, our attention index to monetary policy predictspositive stock returns on fomc announcements. The short-horizon trend inattention (5-day minus 20-day moving average) is positive and significant atthe 5% level. Similar results hold for long-horizon trend in attention (60-dayminus 250-day moving average). We next investigate whether attention tomonetary can predict changes in vix on fomc announcement days.1494.5. Using Attention for ForecastingTable 4.10: Forecasts on FOMC AnnouncementsThis table presents the results of an ols regression of the daily S&P 500log returns (in percent) and changes in the implied volatility (∆vxo) on afomc dummy, a detrended monetary macro attention composite index (mai-c1), and the mai-c1 interacted with the fomc dummy. The fomc dummyequal one on fomc days. We show the results for two different detrendedfrequencies for the monetary attention index. For example, MAI5−20 is thedifference between the five-day and 20-day moving average of mai-c1. Thestandard errors are reported in parenthesis and are calculated using theNewey-West standard errors (six lags). *, **, *** denote the statisticalsignificance at the 10%, 5%, 1% levels, respectively. The sample period isJune 1, 1980 to April 30, 2015.Panel A: S&P 500 returns (1994-2015)MAI: MAI5−20 MAI5−250 MAI20−250 MAI60−250MAI 0.362 0.591** 0.115 0.287* -0.096 0.062 0.432 0.546**(0.265) (0.281) (0.154) (0.168) (0.204) (0.212) (0.368) (0.260)MAI×NBER -0.675 -0.489* -0.666 -0.561(0.480) (0.277) (0.546) (1.539)const 0.329*** 0.337*** 0.330*** 0.344*** 0.323*** 0.332*** 0.323*** 0.327***(0.089) (0.090) (0.091) (0.092) (0.091) (0.091) (0.089) (0.086)Obs. 171 171 171 171 171 171 171 171Adj-R2 0.01 0.02 -0.00 0.01 -0.00 -0.00 0.01 0.00Panel B: Changes in implied volatility (1994-2015)MAI: MAI5−20 MAI5−250 MAI20−250 MAI60−250MAI -4.423** -6.079*** -2.124* -3.152** -0.574 -1.043 -3.272* -3.660*(1.849) (2.236) (1.104) (1.369) (1.431) (1.747) (1.922) (1.899)MAI×NBER 4.887* 2.926* 1.975 1.901(2.876) (1.676) (2.767) (6.309)const -2.088*** -2.149*** -2.128*** -2.214*** -2.066*** -2.093*** -2.031*** -2.043***(0.579) (0.581) (0.599) (0.611) (0.608) (0.614) (0.585) (0.584)Obs. 171 171 171 171 171 171 171 171Adj-R2 0.05 0.06 0.02 0.03 -0.01 -0.01 0.01 0.00Panel C: Changes in Fed fund rates (1994-2008)MAI: MAI5−20 MAI5−250 MAI20−250 MAI60−250MAI -0.105 0.007 -0.127* -0.007 -0.218** -0.017 -0.326*** -0.096(0.151) (0.089) (0.072) (0.059) (0.109) (0.087) (0.124) (0.120)MAI×NBER -0.215 -0.251*** -0.615*** -0.731**(0.247) (0.084) (0.179) (0.288)const -0.028 -0.026 -0.025 -0.020 -0.025 -0.017 -0.012 -0.012(0.032) (0.032) (0.030) (0.030) (0.030) (0.029) (0.028) (0.028)Obs. 104 104 104 104 104 104 104 104Adj-R2 0.01 0.02 0.06 0.12 0.07 0.21 0.07 0.151504.6. Conclusion to Chapter 4Panel B shows that an increase in the short-horizon trend in attentionpredicts a decrease in vix, which suggests that an increase attention predictsa decrease in uncertainty on fomc announcement days. The coefficienton the interaction term MAI×NBER is positive and significant, indicatingthat an increase in attention predicts greater resolution of uncertainty onfomc announcements during expansion than during recessions. Finally,we examine the relationship between changes in the Fed fund rates andattention on fomc announcements.Panel C shows that our attention measure to monetary predicts nega-tive changes in Fed fund rates, meaning that attention increases before theFed announces a cut in Fed fund rates. This is consistent with the factthat during 1994-2008, most of the Fed’s decision on interest rate was tolower rather than increase the Fed fund rate.70 More importantly, the rela-tionship is stronger during recessions. This is consistent with Kacperczyk,Van Nieuwerburgh, and Veldkamp (2016), who show that investors pay moreattention to macroeconomic risks during recessions.4.6 Conclusion to Chapter 4We build indices of investor attention to macroeconomic fundamentals usingnews articles from wsj and nyt. Attention indices rises around macroeco-nomic announcements and following changes in fundamentals over quar-terly, annual, and business cycle horizons. The effect of announcements andchanges in fundamentals on indices is asymmetric, with bad news raisingattention more than good news. Attention indices have important implica-tions to financial markets, and we show that aggregate trade volume andvolatility coincide with rising attention, controlling for announcements. Wefurther show that attention predicts surprises as well as stock returns onunemployment and fomc announcement days.Our paper adds to the growing literature documenting the importanceof investor attention in financial markets (e.g. Andrei and Hasler, 2014; Da,Engelberg, and Gao, 2011b). Future work could go in many directions. Wefind evidence of time-varying attention to different macroeconomic funda-mentals in the news media. In the spirit of the Merton (1980) IntertemporalCapital Asset Pricing Model, such attention dynamics could be related totime-variation in the risks or risk premia associated with different types70We focus on the 1994-2008 period because the Fed reached the so-called ’zero lowerbound’ and did not change the Fed fund rate after 2008. The most recent rate change wasthe increase in December 2015.1514.6. Conclusion to Chapter 4of macroeconomic fundamentals. Another possible extension is to combineboth investors’ sentiment and attention to macroeconomic fundamentals andrelate to stock market returns.152Chapter 5ConclusionThis thesis is a collection of three essays on Information Economics. I fo-cus specfically on the impact of public news on financial markets and assetprices. The first essay, Chapter 2, examines the speed of price discovery fol-lowing earnings announcements and the role of order flow to price discovery.Past research usually assumes that liquidity providers are not sophisticatedenough to process public information and, in turn, rely on the incomingorder flow from sophisticated traders to adjust prices. Contrary to pastresearch, I find that earnings surprises are the main determinant that ex-plains price changes following earnings announcements and not order flow.Yet, important questions remain to be answered. For example, despite fastprice discovery following earnings surprises, why is the impact of earningssurprises on stock volatility and trade volume remains abnormally high forseveral hours or for even more than one trading day?In Chapter 3, I analyze the impact of Federal Open Market Committee(fomc) announcement press conferences on financial markets and investorattention to monetary policy. In an effort to increase transparency, the Chairof the Board of Governors now holds a press conference following half of thescheduled fomc announcements. I find that holding press conferences aftersome, but not all, fomc meetings skew expectations of important mone-tary policy decisions towards announcement days with press conferences. Inturn, the introduction of press conferences coordinates media and investorattention towards those meetings. This may pose a problem for the FederalReserve, which is generally believed to be averse to surprising markets. Ifthe Federal Reserve must announce an important decision on days with nopress conference, it risks surprising markets because investors did not expectany important news.In Chapter 4, I build indices of investor attention to macroeconomicfundamentals using news articles from the Wall Street Journal and the NewYork Times. I document the dynamics in attention, its fluctuation over time,and its relationship to macroeconomic fundamentals. Investor attention in-dices have important implications for financial markets, and we show thataggregate trade volume and volatility coincide with rising attention, con-1535.1. Future Worktrolling for announcements. I further show that attention predicts surprisesas well as stock returns on unemployment and fomc announcement days.More importantly, understanding investor attention to macro risk throughmedia attention to macroeconomic fundamentals provides useful informationbeyond the dates and contents of macroeconomic announcements.5.1 Future WorkI plan extend each chapter to new research projects. In the first essay, I onlyexplore the evolution of price discovery since the 1980s for the largest 1,500U.S. stocks. I plan to extend the analysis to the complete cross-section ofU.S. stocks and document the evolution of the post-earnings announcementdrifts over time. Despite not being a simple task, I would like to pin downthe main factor explaining the near disappearance of the post-earnings an-nouncement drifts at the daily horizon. Also, I would like to understandhow faster price discovery following earnings announcements influence assetpricing factors in the cross-section of stocks.In the second chapter, I am currently extending the analysis to price dis-covery following fomc announcements in the equity market using changesin eurodollar futures as a measure of fomc surprises. I find that surprisesare larger on fomc announcement days when there is a press conference.Moreover, fomc announcement surprises incorporate equity prices withinminutes. But, using a non-parametric approach to examine price formationfollowing fomc announcements relative to future indicative prices, I findthat prices following announcements remain noisy and that it takes sev-eral hours before price formation is complete. Despite large trade volumefollowing fomc announcements, prices are not efficient.With the third chapter of my thesis, there is a lot of potential researchavenue to explore with the newly constructed dataset on investor attentionto macroeconomic risks. For example, is a macroeconomic risk factor pricedin the cross-section of stocks conditioned only when investors pay attentionto that particular risk?154BibliographyAbel, Andrew B., Janice C. Eberly, and Stavros Panageas, 2007, Optimalinattention to the stock market, American Economic Review 97, 244–249., 2013, Optimal inattention to the stock market with informationcosts and transactions costs, Econometrica 81, 1455–1481.Admati, Anat R., and Paul Pfleiderer, 1988, A theory of intraday patterns:Volume and price variability, Review of Financial Studies 1, 3–40.Akbas, Ferhat, 2016, The calm before the storm, The Journal of Finance71, 225–266.Amato, Jeffery D., Stephen Morris, and Hyun Song Shin, 2002, Communica-tion and monetary policy, Oxford Review of Economic Policy 18, 495–503.Amengual, Dante, and Dacheng Xiu, 2015, Resolution of policy uncertaintyand sudden declines in volatility, Working Paper, University of Chicago,Chicago, IL.Andersen, Torben G., Tim Bollerslev, Francis X. Diebold, and Heiko Ebens,2001, The distribution of realized stock return volatility, Journal of Fi-nancial Economics 61, 43–76.Andersen, Torben G., Tim Bollerslev, Francis X. Diebold, and Clara Vega,2003a, Micro effects of macro announcements: Real-time price discoveryin foreign exchange, American Economic Review 93, 38–62., 2003b, Micro effects of macro announcements: Real-time pricediscovery in foreign exchange, American Economic Review 93, 38–62., 2007a, Real-time price discovery in global stock, bond and foreignexchange markets, Journal of International Economics 73, 251–277., 2007b, Real-time price discovery in global stock, bond and foreignexchange markets, Journal of international Economics 73, 251–277.155BibliographyAnderson, Oliver Duncan, 1976, Time series analysis and forecasting: theBox-Jenkins approach (Butterworths London).Andrei, Daniel, and Michael Hasler, 2014, Investor attention and stock mar-ket volatility, Review of Financial Studies 28, 33–72., 2016, Dynamic attention behavior under return predictability,Working paper.Antweiler, Werner, and Murray Z. Frank, 2004, Is all that talk just noise?the information content of internet stock message boards, Journal of Fi-nance 59, 1259–1294.Atiase, Rowland K., and Linda Smith Bamber, 1994, Trading volume re-actions to annual accounting earnings announcements: The incrementalrole of predisclosure information asymmetry, Journal of Accounting andEconomics 17, 309–329.Bai, Jennie, Thomas Philippon, and Alexi Savov, 2016, Have financial mar-kets become more informative?, Journal of Financial Economics, forth-coming.Baker, Scott R., Nicholas Bloom, and Steven J. Davis, 2015, Measuringeconomic policy uncertainty, Working Paper.Baldauf, Markus, and Joshua Mollner, 2016, Fast traders make a quick buck:The role of speed in liquidity provision, Working Paper.Balduzzi, Pierluigi, Edwin J. Elton, and T. Clifton Green, 2001a, Economicnews and bond prices: Evidence from the us treasury market, Journal ofFinancial and Quantitative Analysis 36, 523–543., 2001b, Economic news and bond prices: Evidence from the ustreasury market, Journal of financial and Quantitative analysis 36, 523–543.Ball, Ray, and Philip Brown, 1968, An empirical evaluation of accountingincome numbers, Journal of Accounting Research 6, 159–178.Bamber, Linda S., Orie E. Barron, and Thomas L. Stober, 1997, Tradingvolume and different aspects of disagreement coincident with earningsannouncements, Accounting Review 72, 575–597.156BibliographyBanerjee, Snehal, Jesse Davis, and Naveen Gondhi, 2015, When trans-parency improves, must prices reflect fundamentals better?, Working Pa-per, Northwestern University, Evanston, IL.Banerjee, Snehal, and Ilan Kremer, 2010, Disagreement and learning: Dy-namic patterns of trade, The Journal of Finance 65, 1269–1302.Barber, Brad M., and Terrance Odean, 2008, All that glitters: The effect ofattention and news on the buying behavior of individual and institutionalinvestors, Review of Financial Studies 21, 785–818.Barclay, Michael J., and Terrence Hendershott, 2004, Liquidity externalitiesand adverse selection: Evidence from trading after hours, The Journal ofFinance 59, 681–710.Beaver, William H., 1968, The information content of annual earnings an-nouncements, Journal of Accounting Research 6, 67–92.Beber, Alessandro, and Michael W. Brandt, 2009, Resolving macroeconomicuncertainty in stock and bond markets, Review of Finance 13, 1–45.Beber, Alessandro, Micheal W. Brandt, and Kenneth A. Kavajecz, 2011,What does equity sector orderflow tell us about the economy?, Review ofFinancial Studies 24, 3688–3730.Ben-Rephael, Azi, Zhi Da, and Ryan D. Israelsen, 2016, It depends on whereyou search: A comparison of institutional and retail attention, WorkingPaper, Indiana University, IN.Bernanke, Ben S., 2010, Conference call of the Fed-eral Open Market Committee on October 15, 2010,http://www.federalreserve.gov/monetarypolicy/files/FOMC20101015confcall.pdf,accessed March. 14, 2016., 2011, Transcript of ChairmanBernanke’s press conference, April 27, 2011,http://www.federalreserve.gov/mediacenter/files/FOMCpresconf20110427.pdf,accessed Nov. 18, 2015., and Kenneth N. Kuttner, 2005, What explains the stock market’sreaction to Federal Reserve policy?, Journal of Finance 60, 1221–1257.Bernard, Victor L., and Jacob K. Thomas, 1989, Post-earnings-announcement drift: Delayed price response or risk premium?, Journalof Accounting Research 27, 1–36.157BibliographyBessembinder, Hendrik, Marios Panayides, and Kumar Venkataraman, 2009,Hidden liquidity: An analysis of order exposure strategies in electronicstock markets, Journal of Financial Economics 94, 361–383.Bjornland, Hilde C., and Kai Leitemo, 2009, Identifying the interdependencebetween US monetary policy and the stock market, Journal of MonetaryEconomics 56, 275–282.Blinder, Alan, 1998, Central Banking in Theory and Practice (MIT Press:Cambridge, MA)., Charles Goodhart, Philipp Hildebrand, David Lipton, and CharlesWyplosz, 2001, How do central banks talk?, Geneva Reports on the WorldEconomy 3 London, Centre for Economic Policy Research.Bloomfield, Robert, Maureen O’Hara, and Gideon Saar, 2015, Hidden liq-uidity: Some new light on dark trading, The Journal of Finance 70, 2227–2274.Boguth, Oliver, Murray Carlson, Adlai J. Fisher, and Mikhail Simutin, 2016,Horizon effects in average returns: The role of slow information diffusion,Review of Financial Studies pp. 2241–2281.Bollerslev, Tim, Sophia Zhengzi Li, and Viktor Todorov, 2016, Roughingup beta: Continuous versus discontinuous betas and the cross section ofexpected stock returns, Journal of Financial Economics 120, 464–490.Bollerslev, Tim, and Hans Ole Mikkelsen, 1996, Modeling and pricing longmemory in stock market volatility, Journal of Econometrics 73, 151–184.Boulatov, Alex, and Thomas J. George, 2013, Hidden and displayed liq-uidity in securities markets with informed liquidity providers, Review ofFinancial Studies 26, 2096–2137.Boyd, John H., Jian Hu, and Ravi Jagannathan, 2005, The stock market’sreaction to unemployment news: Why bad news is usually good for stocks,Journal of Finance 60, 649–672.Brennan, M. J., N. Jegadeesh, and B. Swaminathan, 1993, Investment anal-ysis and the adjustment of stock prices to common information, Reviewof Financial Studies 6, 799–824.Brogaard, Jonathan, Terrence Hendershott, and Ryan Riordan, 2014, Highfrequency trading and price discovery, Review of Financial Studies 27,2267–2306.158Bibliography, 2015, Price discovery without trading: Evidence from limit orders,Working Paper.Brusa, Francesca, Pavel G. Savor, and Mungo I. Wilson, 2016, One centralbank to rule them all, Working Paper, University of Oxford, UK.Calvet, Laurent E., and Adlai J. Fisher, 2007, Multifrequency news andstock returns, Journal of Financial Economics 86, 178–212.Campbell, John Y., and Motohiro Yogo, 2006, Efficient tests of stock returnpredictability, Journal of Financial Economics 81, 27–60.Carroll, Christopher D., 2003, Macroeconomic expectations of householdsand professional forecasters, the Quarterly Journal of Economics pp. 269–298.Chae, Joon, 2005, Trading volume, information asymmetry, and timing in-formation, The Journal of Finance 60, 413–442.Chakrabarty, Bidisha, Pamela Moulton, and Xu Wang, 2015, Attention ef-fects in a high-frequency world, Working Paper.Chakrabarty, Bidisha, Roberto Pascual, and Andriy Shkilko, 2015, Evaluat-ing trade classification algorithms: Bulk volume classification versus thetick rule and the Lee-Ready algorithm, Journal of Financial Markets 25,52–79.Chakrabarty, Bidisha, and Kenneth W. Shaw, 2008, Hidden liquidity: Orderexposure strategies around earnings announcements, Journal of BusinessFinance & Accounting 35, 1220–1244.Chen, Nai-Fu, Richard Roll, and Stephen A. Ross, 1986, Economic forcesand the stock market, Journal of business 59, 383–403.Chordia, Tarun, Amit Goyal, Gil Sadka, Ronnie Sadka, and LakshmananShivakumar, 2009, Liquidity and the post-earnings-announcement drift,Financial Analysts Journal 65, 18–32.Chordia, Tarun, T. Clifton Green, and Badrinath Kottimukkalur, 2016, Dohigh frequency traders need to be regulated? Evidence from trading onmacroeconomic announcements, Working Paper.Cieslak, Anna, Adair Morse, and Annette Vissing-Jorgensen, 2015a, Stockreturns over the FOMC cycle, Working Paper, University of California atBerkeley, Berkeley, CA.159Bibliography, 2015b, Stock returns over the fomc cycle, Working Paper.Cook, Timothy, and Thomas Hahn, 1989, The effect of changes in thefederal-funds rate target on market interest-rates in the 1970s, Journalof Monetary Economics 24, 331–351.Da, Zhi, Joseph Engelberg, and Pengjie Gao, 2011a, In search of attention,Journal of Finance 66, 1461–1499., 2011b, In search of attention, Journal of Finance 66, 1461–1499., 2015, The sum of all FEARS: Investor sentiment and asset prices,Review of Financial Studies 28, 1–32.Da, Zhi, Umit G. Gurun, and Mitch Warachka, 2014, Frog in the pan:Continuous information and momentum, Review of Financial Studies 27,2171–2218.DellaVigna, Stefano, and Ethan Kaplan, 2007, The fox news effect: Mediabias and voting, The Quarterly Journal of Economics 122, 1187–1234.DellaVigna, Stefano, and Joshua M. Pollet, 2009, Investor inattention andFriday earnings announcements, The Journal of Finance 64, 709–749.Doyle, Jeffrey T., Russell J. Lundholm, and Mark T. Soliman, 2006, The ex-treme future stock returns following I/B/E/S earnings surprises, Journalof Accounting Research 44, 849–887.Driscoll, John C., and Aart C. Kraay, 1998, Consistent covariance matrixestimation with spatially dependent panel data, Review of Economics andStatistics 80, 549–560.Duffie, Darrell, and Tong-sheng Sun, 1990, Transactions costs and portfoliochoice in a discrete-continuous-time setting, Journal of Economic Dynam-ics and Control 14, 35–51.Ederington, Louis H, and Jae Ha Lee, 1993, How markets process informa-tion: News releases and volatility, The Journal of Finance 48, 1161–1191.Evans, Martin D., and Richard K. Lyons, 1999, Order flow and exchangerate dynamics, NBER working paper., 2002, Order flow and exchange rate dynamics, Journal of PoliticalEconomy 110, 170–180.160Bibliography, 2008, How is macro news transmitted to exchange rates?, Journalof Financial Economics 88, 26–50.Fama, Eugene F., Lawrence Fisher, Michael C. Jensen, and Richard Roll,1969, The adjustment of stock prices to new information, InternationalEconomic Review 10, 1–21.Fama, Eugene F., and Kenneth R. French, 1993, Common risk factors inthe returns on stocks and bonds, Journal of financial economics 33, 3–56.Fang, Lily, and Joel Peress, 2009a, Media coverage and the cross-section ofstock returns, Journal of Finance 64, 2023–2052.Fang, Lily H., and Joel Peress, 2009b, Media coverage and the cross-sectionof stock returns, Journal of Finance 64, 2023–2052.Fisher, Adlai J., Charles Martineau, and Jinfei Sheng, 2016, Media atten-tion, macroeconomic fundamentals, and the stock market, Working Paper,University of British Columbia, Vancouver, Canada.Fleming, Michael J., and Eli M. Remolona, 1999a, Price formation andliquidity in the u.s. treasury market: The response to public information,The Journal of Finance 54, 1901–1915., 1999b, Price formation and liquidity in the US treasury market:The response to public information, Journal of Finance 54, 1901–1915.French, Kenneth R., and Richard Roll, 1986, Stock return variances: Thearrival of information and the reaction of traders, Journal of FinancialEconomics 17, 5–26.Gentzkow, Matthew, and Jesse M Shapiro, 2010, What drives media slant?evidence from us daily newspapers, Econometrica 78, 35–71.Ghysels, Eric, Pedro Santa-Clara, and Rossen Valkanov, 2006, Predictingvolatility: getting the most out of return data sampled at different fre-quencies, Journal of Econometrics 131, 59–95.Gilbert, T., 2011, Information aggregation around macroeconomic an-nouncements: revisions matter, Journal of Financial Economics 101, 114–131.Glosten, Lawrence R., and Paul R. Milgrom, 1985, Bid, ask and transac-tion prices in a specialist market with heterogeneously informed traders,Journal of financial economics 14, 71–100.161BibliographyGreen, T. Clifton, 2004, Economic news and the impact of trading on bondprices, The Journal of Finance 59, 1201–1233.Gu¨rkaynak, Refet S., Brian Sack, and Eric T. Swanson, 2005, Do actionsspeak louder than words? The response of asset prices to monetary policyactions and statements, International Journal of Central Banking 1, 55–93.Harris, Larry, 1996, Does a minimum price variation encourage order expo-sure?, Unpublished working paper.Hirshleifer, David, Sonya S. Lim, and Siew H. Teoh, 2009, Driven to dis-traction: Extraneous events and underreaction to earnings news, Journalof Finance 64, 2289–2325.Holden, Craig W., and Stacey Jacobsen, 2014, Liquidity measurement prob-lems in fast, competitive markets: Expensive and cheap solutions, Journalof Finance 69, 1747–1785.Hong, Harrison, Terence Lim, and Jeremy C. Stein, 2000, Bad news travelsslowly: Size, analyst coverage, and the profitability of momentum strate-gies, The Journal of Finance 55, 265–295.Hou, K., and T. J. Moskowitz, 2005, Market frictions, price delay, and thecross-section of expected returns, Review of Financial Studies 18, 981–1020.Hu, Grace Xing, Jun Pan, and Jiang Wang, 2015a, Early peek advantage?,Working Paper., 2015b, Early peek advantage?, Working paper.Huang, Lixin, and Hong Liu, 2007, Rational inattention and portfolio selec-tion, Journal of Finance 62, 1999–2040.Huberman, Gur, and Tomer Regev, 2001, Contagious speculation and acure for cancer: A nonevent that made stock prices soar, The Journal ofFinance 56, 387–396.Jiang, Christine X., Tanakorn Likitapiwat, and Thomas H. McInish, 2012,Information content of earnings announcements: Evidence from after-hours trading, Journal of Financial and Quantitative Analysis 47, 1303–1330.162BibliographyJones, Charles M., Gautam Kaul, and Marc L. Lipson, 1994, Transactions,volume, and volatility, Review of Financial Studies 7, 631–651.Jones, Charles M., Owen Lamont, and Robin L. Lumsdaine, 1998a, Macroe-conomic news and bond market volatility, Journal of Financial Economics47, 315–337., 1998b, Macroeconomic news and bond market volatility, Journalof Financial Economics 47, 315–337.Kacperczyk, Marcin, Stijn Van Nieuwerburgh, and Laura Veldkamp, 2016,A rational theory of mutual funds’ attention allocation, Econometrica 84,571–626.Kacperczyk, Marcin, Stijn G. van Nieuwerburgh, and Laura Veldkamp,2016, A rational theory of mutual funds attention allocation, Economet-rica 84, 571–626.Kahneman, Daniel, and Amos Tversky, 1979, Prospect theory - analysis ofdecision under risk, Econometrica 47, 263–291.Kandel, Eugene, and Neil D. Pearson, 1995, Differential interpretation ofpublic signals and trade in speculative markets, Journal of Political Econ-omy pp. 831–872.Kasznik, Ron, and Baruch Lev, 1995, To warn or not to warn: Managementdisclosures in the face of an earnings surprise, Accounting Review pp.113–134.Kim, O., and R.E. Verrecchia, 1991, Market reaction to anticipated an-nouncements, Journal of Financial Economics 30, 273–309.Kim, Oliver, and Robert E. Verrecchia, 1994, Market liquidity and volumearound earnings announcements, Journal of Accounting and Economics17, 41–67.Kuttner, Kenneth N., 2001, Monetary policy surprises and interest rates:evidence from the Fed funds futures market, Journal of Monetary Eco-nomics 47, 523–544.Kyle, Albert S., 1985, Continuous auctions and insider trading, Economet-rica 53, 1315–1335., and Anna A. Obizhaeva, 2016, Market microstructure invariance:Empirical hypotheses, Econometrica 84, 13451404.163BibliographyLang, Mark H., and Russell J. Lundholm, 1996, Corporate disclosure policyand analyst behavior, Accounting Review 71, 467–492.Lee, Charles M.C., Belinda Mucklow, and Mark J. Ready, 1993, Spreads,depths, and the impact of earnings information: An intraday analysis,Review of Financial Studies 6, 345–374.Li, Jiasun, 2016, Slow price adjustment to public news in after-hours trading,Journal of Trading 11, 16–31.Lucca, David O., and Emanuel Moench, 2015a, The pre-FOMC announce-ment drift, Journal of Finance 70, 329–371.Lucca, David Ol, and Emanuel Moench, 2015b, The pre-fomc announcementdrift, The Journal of Finance 70, 329–371.Matthies, Ben, and Yukun Liu, 2015, News-based macroeconomic asset pric-ing, Working paper.Merton, Robert C., 1973, Theory of rational option pricing, The Bell Journalof economics and management science pp. 141–183., 1980, On estimating the expected return on the market: An ex-ploratory investigation, Journal of Financial Economics 8, 323–361.Milgrom, Paul, and Nancy Stokey, 1982, Information, trade and commonknowledge, Journal of Economic Theory 26, 17–27.Moinas, Sophie, 2011, Hidden limit orders and liquidity in order drivenmarkets, Working Paper.Morse, Dale, 1981, Price and trading volume reaction surrounding earningsannouncements: A closer examination, Journal of Accounting Researchpp. 374–383.NASDAQ, 2016a, NASDAQ TotalView-itch 4.1, http://www.nasdaqtrader.com/content/technicalsupport/specifications/dataproducts/nqtv-itch-v4_1.pdf., 2016b, NASDAQ TotalView-itch 5.0, http://www.nasdaqtrader.com/content/technicalsupport/specifications/dataproducts/NQTVITCHspecification.pdf.O’Hara, Maureen, 2003, Presidential address: Liquidity and price discovery,The Journal of Finance 58, 1335–1354.164Bibliography, Chen Yao, and Mao Ye, 2014, What’s not there: Odd lots andmarket data, Journal of Finance 69, 2199–2236.Ozdagli, Ali, and Michael Weber, 2015, Monetary policy through productionnetworks: Evidence from the stock market, Working Paper, University ofChicago, Chicago, IL.Pasquariello, Paolo, and Clara Vega, 2007, Informed and strategic order flowin the bond markets, Review of Financial Studies 20, 1975–2019.Patell, James M., and Mark A. Wolfson, 1984, The intraday speed of adjust-ment of stock prices to earnings and dividend announcements, Journal ofFinancial Economics 13, 223–252.Patton, A.J., and M. Verardo, 2012, Does beta move with news? Firm-specific information flows and learning about profitability, Review of Fi-nancial Studies 25, 2789–2839.Peng, Lin, and Wei Xiong, 2006, Investor attention, overconfidence andcategory learning, Journal of Financial Economics 80, 563–602.Peress, Joel, 2008, Media coverage and investors attention to earnings an-nouncements, Working paper.Peress, J., 2014, The media and the diffusion of information in financialmarkets: evidence from newspaper strikes, Journal of Finance 69, 2007–2043.Ross, Stephen A., 1989, Information and volatility: The no-arbitrage mar-tingale approach to timing and resolution irrelevancy, The Journal of Fi-nance 44, 1–17.Santosh, Shrihari, 2014, Public information and price discovery: Trade timevs. clock time, Working Paper.Savor, Pavel, and Mungo Wilson, 2013a, How much do investors care aboutmacroeconomic risk? Evidence from scheduled economic announcements,Journal of Financial and Quantitative Analysis 48, 343–375.Savor, P., and M. Wilson, 2013b, How much do investors care about macroe-conomic risk? evidence from scheduled economic announcements, Journalof Financial and Quantitative Analysis 48, 343–375.Savor, Pavel, and Mungo Wilson, 2014, Asset pricing: A tale of two days,Journal of Financial Economics 113, 171–201.165BibliographySchwert, G. William, 1989, Why does stock market volatility change overtime?, Journal of Finance 44, 1115–1153.Shapiro, Mary L., 2010, Speech by SEC chair-man: “Strengthening our Equity Market Structure”,https://www.sec.gov/news/speech/2010/spch090710mls.htm.Sims, Christopher A., 2003, Implications of rational inattention, Journal ofMonetary Economics 50, 665–690.Stein, Jeremy C., 1989, Cheap talk and the Fed - a theory of imprecisepolicy announcements, American Economic Review 79, 32–42., and Adi Sunderam, 2015, Gradualism in monetary policy: a time-consistency problem?, Working Paper, Harvard University, Cambridge,MA.Swanson, Eric T., 2015, Measuring the effects of unconventional monetarypolicy on asset prices, Working Paper, University of California at Irvine,Irvine, CA.Tetlock, Paul C., 2007, Giving content to investor sentiment: The role ofmedia in the stock market, Journal of Finance 62, 1139–1168., 2010, Does public financial news resolve asymmetric information?,Review of Financial Studies 23, 3520–3557.Weller, Brian M., 2016, Efficient prices at any cost: Does algorithmic tradingdeter information acquisition?, Working paper.Woodruff, Catherine S., and Andrew J. Senchack, 1988, Intradaily price-volume adjustments of NYSE stocks to unexpected earnings, The Journalof Finance 43, 467–491.Yellen, Janet L., 2014, Transcript of ChairYellen’s press conference, December 17, 2014,http://www.federalreserve.gov/mediacenter/files/fomcpresconf20141217.pdf,accessed Nov. 18, 2015., 2015a, Transcript of Chair Yellen’s press conference, June 17, 2015,http://www.federalreserve.gov/mediacenter/files/FOMCpresconf20150617.pdf,accessed Nov. 18, 2015.166, 2015b, Transcript of Chair Yellen’spress conference, September 17, 2015,http://www.federalreserve.gov/mediacenter/files/fomcpresconf20150917.pdf,accessed Nov. 18, 2015.Yuan, Yu, 2015, Market-wide attention, trading, and stock returns, Journalof Financial Economics 116, 548–564.Zhang, X, 2006, Information uncertainty and stock returns, The Journal ofFinance 61, 105–137.Zumbrun, Josh, 2015, WSJ survey: most economists don’tsee all FOMC meetings as ’live’, The Wall Street Journal,http://www.wsj.com/articles/wsj-survey-most-economists-dont-see-all-fomc-meetings-as-live-1444312805, accessed Nov. 18, 2015.167Appendix AAppendix to Chapter 2A.1 Trading Hours on NASDAQFigure A.1: Regular and After-Hours Trading for the NASDAQ Stock Ex-changeThis figure shows the regular trading hours (9:30 a.m. to 4 p.m) and theafter-hours trading sessions (4 p.m. to 8 p.m. and from 4 a.m. to 9.30 a.m.)on the nasdaq stock exchange.No tradingRegular trading hours9:30 16:00 20:00 4:00 9:30 16:00Trading day t Trading day t+1Regular trading hoursAfter hours After hours168A.2. Post-Earnings Announcement Drifts since 1984A.2 Post-Earnings Announcement Drifts since1984I plot in Figure A.2 the average cumulative abnormal returns (car) withineach earnings surprises quintile and their corresponding 95 percent confi-dence intervals around earnings announcements for the largest 1,500 U.S.stocks for different time periods between 1984 to 2010. In total there areclose to 114, 200 earnings announcements. Because I do not have the actualtimestamp of each earnings announcement, the sample contains earningsannouncements that were announced during both regular and after-markethours. Therefore, the day ”0” contains both abnormal returns of the date ofthe announcement and the following trading day. I further exclude observa-tions with returns in the top and bottom 5/1,000th of the distributions. But,I find that excluding outliers only have an impact on the bottom earningsquintile for the period of 2006 to 2010.169A.2. Post-Earnings Announcement Drifts since 1984Figure A.2: Historical Cumulative Abnormal Daily Returns around Earn-ings AnnouncementsThis figure shows the stocks’ cumulative abnormal returns (car) from fivetrading days preceding to 61 trading days following earnings announcementsfor each earnings surprises quintile. The car are calculated as follows:CAR[−5, 61]i,q =61∏k=−5(1 +Ri,k)−61∏k=−5(1 +Rp,k),where Ri,k is the return of the stock i and Rp,k is the return on the size andbook-to-market matching Fama-French portfolio on day k for quarter q’searnings. Each line represents a different quintile sort for earnings surprises.The shaded areas are pointwise 95% confidence bands around the averageabnormal returns. The vertical line corresponds to the earnings announce-ment day. The sample consists of earnings announcements from the largest1,500 U.S. stocks between 1984 and 2010.−0.04−0.020.000.020.04CumulativereturnPanel A: 1984-1994 Panel B: 1995-2000−5 0 10 20 30 40 50 60Days since announcement−0.04−0.020.000.020.040.06CumulativereturnPanel C: 2001-2005−5 0 10 20 30 40 50 60Days since announcementPanel D: 2006-2010Top quintile Quintile 4 Quintile 3 Quintile 2 Bottom quintile170A.3. Institutional Details about Hidden Orders on NASDAQ ITCHA.3 Institutional Details about Hidden Orderson NASDAQ ITCHThis note contains details about hidden order observations in nasdaq TotalView-itch.In nasdaq TotalView-itch, we do not observe submitted hidden ordersby liquidity providers. Prior to October 6, 2010, trades against a hiddenorder would display both the Order Reference Number associated with thehidden order and a Buy/Sell Indicator, which indicated whether the initiatedtrade was a buy or sell (see appendix in NASDAQ, 2016a). But, sinceOctober 6, 2010, all trades against hidden orders display a “0” as an OrderReference Number and, since July 14, 2014, all trades against hidden ordersdisplay “B” as a Buy/Sell Indicator.These changes impose challenges to empiricists who wish to understandthe drivers to the use of hidden orders versus displayed orders and the impactof hidden orders on stock prices, trade volume, etc. Just for example, in thispaper when I study the impact of market-initiated trade imbalance (i.e.,order flow imbalance) on stock returns, I must end my sample on July 13,2014 because I do not have the Buy/Sell Indicators on trades against hiddenorders from July 14, 2014 onward.Why did nasdaq do these changes? Some traders claim that providingthe Order Reference Number and the Buy/Sell Indicator help high-frequencytraders figure out market directions.71 For example, Order Reference Num-ber linked to a trade is cumulative. This means that every time a tradeexecutes against a fraction of the total shares from the same hidden order,the same Order Reference Number is attached to that trade. This allowsnasdaq itch subscribers to determine how many shares the hidden buyeror seller is willing to trade.The objective of using hidden orders is not to provide other traders theability to infer their strategies and potentially private information. Aftersome pressure from the investor community, nasdaq decided not to displayOrder Reference Number and the Buy/Sell Indicator in nasdaq itch. But,empiricists who want to understand the greater details of the functioningof financial markets now have less detailed data to work with. nasdaqitch was the only data source on hidden order activities on the U.S. stockexchanges.71See the 2010 white paper “Exchanges and Data Feeds: Data Theft on Wall Street”by Sal Arnuk and Joeph Saluzzi of Themis Trading at http://blog.themistrading.com/wp-content/uploads/2010/05/THEMIS-Data-Theft-On-Wall-Street-05-11-10.pdf171A.3. Institutional Details about Hidden Orders on NASDAQ ITCHnasdaq does provide at a monthly fee of $2,000 data on the market’sfull liquidity, including reserve and hidden interest. The data are calledModel View and provide a minute-by-minute summary of total displayedand hidden interest at each price point. The data are not available “live”and are reported with a two-week lag. Also, the minute-by-minute data areavailable only from 8 a.m. to 4 p.m. As shown in this paper, hidden ordersare heavily used in the after-hours market.172A.4. High-frequency Trading Activites in the After-Hours MarketA.4 High-frequency Trading Activites in theAfter-Hours MarketHigh-frequency traders (hft) now represent a large share of market trad-ing but are they also present in the after-hours market? To provide someinsights on this question, I use a dataset that contains a sample of 120 nas-daq-listed stocks that identify the liquidity taker and maker (provider) foreach trade as a high-frequency trader or non-high-frequency trader. Thedata identify 26 proprietary high-frequency trading firms. Though the timeseries of these data does not span the time series of the nasdaq itch dataused in this study, it provides interesting insights. This is the first datasetthat contains high-frequency traders identification for us stocks (see Bro-gaard, Hendershott, and Riordan (2014) for more details on this dataset).Table A.1 in the Appendix shows the fractions of hft that supply liquid-ity (makers) and take liquidity (takers), and the fraction of total trades forwhich the liquidity taker or maker is an hft. The data show that hft ac-tivities decrease in the after-hours with earnings announcements by morethan half for large firms, from 67 to 22 percent of total shares traded andfrom 73 to 30 percent for the total number of trades. For small firms, thetotal activity remains around 30 percent. These numbers suggest the pres-ence of more institutional traders in the after-hours market than hft. Buthigh-frequency trading can still play a role around earnings announcements.Weller (2016) shows that algorithmic trading deters information acquisitionprior to earnings announcements.173A.4. High-frequency Trading Activites in the After-Hours MarketTable A.1: High-Frequency Trading Activities during Regular and After-Market HoursThis table reports the average fraction of trades, both in shares and totaltrades, with high-frequency trading activities during regular market hours(9:30 a.m. to 4 p.m.) and in the after-hours market (4 p.m. to 9:30 a.m.)with and without earnings announcements (ea). Makers stands for liquid-ity making for trades executed against limit orders submitted by a high-frequency trader. Takers stands for liquidity taking for trades initiated bya high-frequency trader. Total stands for total high-frequency trading ac-tivities with either both or one side of the trade involving a high-frequencytrader. The numbers are in percentages. The sample consists of 120 nas-daq-listed stocks. Sample firms are separated into size-tercile groups. Thesample period is from January 1, 2008 to December 31, 2009.Shares Trades Shares TradesTrading Period Firm Size Makers Takers Makers Takers Total TotalMarket hours Small 20 12 21 12 30 31Medium 35 18 39 21 48 53Large 42 40 46 45 67 73After hours Small 13 12 13 13 23 23Medium 16 16 17 17 30 31Large 17 16 21 18 30 33After hours (EA) Small 17 20 17 23 34 36Medium 11 13 13 14 22 24Large 13 11 17 17 22 30174A.5. Additional Results on the Impact of Earnings Surprises. . .A.5 Additional Results on the Impact ofEarnings Surprises on Trade Volume,Volatility, and Bid-Ask SpreadsFigure A.3: The Response of Abnormal Volatility, Abnormal QuotedSpread, and Abnormal Turnover to Earnings Surprises around Earnings An-nouncementsThis figure shows the estimated coefficient responses of abnormal volatil-ity, abnormal quoted spread, and abnormal turnover to absolute earningssurprises around earnings announcements at each 30-minute interval dur-ing regular trading hours. The regression specifications are described inthe main text. The left pane shows the day before the earnings announce-ment (ea), the middle pane is the ea day, and the right pane is the dayafter the ea. The ea occurs in the after-hours market (between 4 p.m.and 9:30 a.m.) indicated by the straight dashed vertical lines. The circleblue line represents stocks with after-hours trading and the square red linerepresents stocks with no after-hours trading activity following earnings an-nouncements. Volatility is the sum of the five-minute absolute value of theresiduals in Equation (2.8):rτ = α+ ρrτ−1 + γrmτ + βτSt · 1{τ∈t} + τ ,over a 30-minute interval. Quoted spread is the average of the time-weightedone-second quoted spread defined as bid-ask spread divided by the midquotein a 30-minute interval. Turnover is the sum of total shares traded in a 30-minute interval divided by the number of shares outstanding and scaled bythe standard deviation of that year. The shaded areas are pointwise 95%confidence bands around the estimated coefficients. The standard errors arecalculated using the Driscoll and Kraay (1998) method.175A.5. Additional Results on the Impact of Earnings Surprises. . .Panel A: Abnormal volatility response to earnings surprises10:0011:0012:0013:0014:0015:0016:00−0.50.00.51.01.52.02.53.03.5ResponseDay before EA10:0011:0012:0013:0014:0015:0016:00EA Day10:0011:0012:0013:0014:0015:0016:00Day after EAWith after-hour tradingWithout after-hour tradingPanel B: Abnormal quoted spread response to earnings surprises10:0011:0012:0013:0014:0015:0016:00−0.10−0.050.000.050.100.150.200.250.30ResponseDay before EA10:0011:0012:0013:0014:0015:0016:00EA Day10:0011:0012:0013:0014:0015:0016:00Day after EAWith after-hour tradingWithout after-hour tradingPanel C: Abnormal turnover response to earnings surprises10:0011:0012:0013:0014:0015:0016:00−400−2000200400600800ResponseDay before EA10:0011:0012:0013:0014:0015:0016:00EA Day10:0011:0012:0013:0014:0015:0016:00Day after EAWith after-hour tradingWithout after-hour trading176Appendix BAppendix to Chapter 3FOMC Transcripts ExcerptsWhile the fomc minutes are typically released three weeks after each meet-ing, actual transcripts of meetings are made public only after 5 years. Atthe time of the writing of this paper, only the transcripts meetings up to andincluding 2011 are available. In this appendix, we summarize and presentexcerpts from relevant discussions pertaining to the creation of fomc pressconferences.72The idea of holding regular pcs after fomc announcements was firstdiscussed in a conference call on October 15, 2010. The general opinionwas favorable, with a notable word of caution from Ms. Yellen: “A pressconference does have some appeal, but it would probably become obligatoryon a regular basis and would be quite a commitment for the Chairman toundertake.” Only Ms. Duke (member of the Board of Governors of theFederal Reserve System) strongly opposed the idea, but would later speakin favor of it at the March 2011 meeting.pcs were further briefly discussed at the November 2010 meeting, withthe idea to be investigated further by the communication subcommitteeheaded by Governor Yellen. Transcripts form the subcommittee meetingsare non publicly available.Ms. Yellen reported the recommendation of the subcommittee to intro-duce regular pcs by the chairman at the March 2011 meeting. The fomcultimately decided to announce pcs two weeks later, with the first one tobe held following the April meeting. “In light of those considerations, thesubcommittee recommends that the Chairman conduct quarterly press confer-ences in the afternoon after the conclusion of each two-day FOMC meeting.”Note that prior to 2012, there were one-day and two-day meetings. Since2012, all fomc meetings take two-days. One of the motivation other thanincreased transparency was that the fomc appeared to be lagging othercountries on that aspect. In the words of Chairman Bernanke, “I think the72All relevant transcripts can be found athttps://www.federalreserve.gov/monetarypolicy/fomchistorical2011.htm.177Appendix B. Appendix to Chapter 3difference between the Fed and other central banks has become quite striking–every other central bank does have this method for communication.”Some members raised concerns regarding the possibility that quarterlypcs would differentiate meetings. For example, Mr. Kocherlakota felt that“it’s distinguishing the meetings in an unusual way. It’s not like we onlymake important decisions at two-day meetings that require a lot of clarifi-cation. So if we are going to go down this path, I actually would suggestthinking about doing it every time.” Ms. Yellen’s response was that “Thedistinguishing feature of the two-day meetings is the economic projectionsand the ability that that would give the Chairman to explain our overallframework and put decisions into the context of them.” Mr. Lacker won-dered what impact pcs would have on their decisions, “whether there wouldbe some hesitance to take actions in between press conference meetings, andI am not quite sure what the answer to that is, but I think it is worth consid-ering.” In the end, Mr. Lacker sided in favor of pcs: “I’d strongly supportthis press conference, and I think there are going to be some subtleties aboutit that are going to emerge in practice. I think we’re going to have to resistthe urge to wait to do things at just these quarterly meetings. I think whenwe want to do something, we’re going to have to have the courage to goahead and do it.” In the end, there was strong support for holding pcs.There are at least three occasions at subsequent meeting were the timingof pcs explicitly entered discussions about some decision. First, at the April2011 meeting which would be followed by the first pc, Mr. Lockart statedthat “I think it is possible with good communication to limit the announce-ment effect on the announcement of ceasing reinvestments, and I think wemay be able to limit an announcement effect even with the initiation of smallasset sales, but this will require skillful communication, and it seems to methat the timing would best coincide with the Chairman’s press conferencesso that he can explain that a rise in the fed funds rate is not necessarilyimminent.”Second, at the June 2011 meeting (pc), Mr. Lockart stated while dis-cussing the idea of changing the wording of the press release that “I thinktoday’s press conference affords the Chairman the opportunity, if you wish orif you get the question, to convey the Committee’s sense of the risk context.”Finally, at the September 2011 meeting (non-pc), Ms. Pianalto sug-gested delaying action until the following meeting because of the associatedpc: “I prefer to continue to reinvest maturing agency debt and MBS intoTreasuries. We told the public that we wanted to return our portfolio to aTreasury-only portfolio. If we decide that this is an appropriate way to go,178Appendix B. Appendix to Chapter 3I would rather wait to do this at our November meeting because that is ameeting where you will have a press conference. It will give you an opportu-nity to talk about the change in our reinvestment strategy.” Ultimately, thecommittee did not wait and adopted the measure at the September meeting,announcing Operation Twist.179Appendix CAppendix to Chapter 4C.1 Sample of news articles mentioningmacroeconomic fundamentalsWe present in this appendix samples of news articles from the Wall StreetJournal (wsj) and New York Time (nyt) that are selected to build ourmedia attention indices to macroeconomic fundamentals.Inflation1) Jonathan Fuerbringer, “Do Deficit Impede Recovery? New Analysis”,New York Times, January 21, 1983.“These levels give rise to the persistent fear of renewed inflation with theFederal Reserve being forced, in an effort to keep the economy going, toease its tight hold on the money supply and push down interest rates sothat the deficit is easier to finance and the recovery will not be tripped up.”Unemployment1) Ken Gilpin, “Jobs Data Push Bonds Up Sharply”, New York Times, July3, 1992.“Stunning weakness in labor statistics for June and the Federal ReserveBoard’s equally striking response to the data caused an eruption in thecredit markets yesterday. Prices of fixed-income securities rose sharply andinterest rates fell.”2) Jonathan Fuerbringer, “Greenspan Speaks: Recession’s Over,” NewYork Times, March 10, 2002.“The recovery, he told Congress, ’is already well under way.’ His commentsfollowed economic data showing a turnaround in manufacturing and a surgein the service sector. Then, on Friday, the Labor Department said theunemployment rate had slipped and that the number of lost jobs had shrunkto just 50,000. All this was uplifting for stocks and bad for bonds.”3) Kate Davidson, “Strong Jobs Report Clears Fed for Liftoff on Rates”Wall Street Journal, December 4, 2015.“The U.S. economy delivered another month of sturdy job growth in Novem-ber, clearing a path for the Federal Reserve to end later this month an180C.1. Sample of news articles mentioning macroeconomic fundamentalsextraordinary seven-year run of near-zero interest rates.”Monetary policy1) Greg Ip, Nicholas Kulish and Jacob M. Schlesinger, “New Model: ThisEconomic Slump Is Shaping Up to Be A Different Downturn,” Wall StreetJournal, January 5, 2001.“One reason is that investors may respond quickly to a cut in Fed interestrates – as they did with Wednesday’s huge rally in response to the surprisereduction of half a percentage point in short-term rates. That instantlyeased some of the pain that had spread through the economy. The stockmarket has become the most important transmission mechanism of mone-tary policy,’ says Jan Hatzius, senior economist at Goldman Sachs. Andthat’s one reason, adds Brad DeLong, an economist at the University ofCalifornia at Berkeley, that Fed moves have a bigger effect now.”2) Michael Derby, “Yield Curve, Fresh Data Are Unsettling Factors—Back From Holiday Break, Investors Will Get a Look at FOMC’s Dec. 12Mintues,” Wall Street Journal, January 3, 2006.“Not only will the market digest reports on manufacturing and employmentdata, but the publication of the minutes from the Federal Open MarketCommittee’s Dec. 13 meeting today also could help settle the debate overwhether a yield-curve inversion makes sense. . . The Fed’s role has becomemore important to the market after central bankers rejiggered their policystatement at their last gathering to suggest at least one more rise in thefederal-funds rate, bringing it to 4.50% from 4.25%, is likely.”181C.1. Sample of news articles mentioning macroeconomic fundamentalsC.1.1 Additional Figures and ResultsFigure C.1: Media Attention and Macroeconomic FundamentalsThis figure shows the monthly media attention indices for the Wall Street Journal (mai-wu), the New York Times (mai-nu), the demeaned composite index (mai-c1), and thedemeaned and standardized composite index (mai-c2) against related macroeconomic fun-damentals described in Table 4.2. The blue line represents a particular media attentionindex (mai) (y-axis) and the red dotted line (secondary-y axis) is the related macroe-conomic fundamental. The units are in percentage. The gray vertical bars are nberrecessions. See Table 4.21980 1984 1988 1992 1996 2000 2004 2008 20120.00.51.01.52.02.53.03.54.0CreditRatingMAIMAI-WU (WSJ)1980 1984 1988 1992 1996 2000 2004 2008 20120.00.51.01.52.02.53.0MAI-NU (NYT)1980 1984 1988 1992 1996 2000 2004 2008 2012−0.50.00.51.01.52.0MAI-C1 (WSJ+NYT)1980 1984 1988 1992 1996 2000 2004 2008 2012−1.0−0.50.00.51.01.52.02.53.03.5MAI-C2 (WSJ+NYT)Credit Rating MAI Corporate Relative Spread010203040506070010203040506070010203040506070010203040506070CorporateRelativeSpread1980 1984 1988 1992 1996 2000 2004 2008 20120.51.01.52.02.53.03.5GDPMAIMAI-WU (WSJ)1980 1984 1988 1992 1996 2000 2004 2008 20120.00.20.40.60.81.01.21.4MAI-NU (NYT)1980 1984 1988 1992 1996 2000 2004 2008 2012−0.6−0.4−0.20.00.20.40.60.81.0MAI-C1 (WSJ+NYT)1980 1984 1988 1992 1996 2000 2004 2008 2012−1.5−1.0−0.50.00.51.01.52.02.53.0MAI-C2 (WSJ+NYT)GDP MAI Real GDP Quarterly Growth Rate−3−2−10123−3−2−10123−3−2−10123−3−2−10123RealGDPQuarterlyGrowthRate(%)1980 1984 1988 1992 1996 2000 2004 2008 20121234567InflationMAIMAI-WU (WSJ)1980 1984 1988 1992 1996 2000 2004 2008 20120.00.51.01.52.02.53.0MAI-NU (NYT)1980 1984 1988 1992 1996 2000 2004 2008 2012−1.0−0.50.00.51.01.52.02.5MAI-C1 (WSJ+NYT)1980 1984 1988 1992 1996 2000 2004 2008 2012−1.0−0.50.00.51.01.52.02.5MAI-C2 (WSJ+NYT)Inflation MAI Change in CPI−1.5−1.0−0.50.00.51.01.5−1.5−1.0−0.50.00.51.01.5−1.5−1.0−0.50.00.51.01.5−1.5−1.0−0.50.00.51.01.5ChangeinCPI(%)182C.1. Sample of news articles mentioning macroeconomic fundamentals1980 1984 1988 1992 1996 2000 2004 2008 20120.00.51.01.52.02.53.03.54.0InterestMAIMAI-WU (WSJ)1980 1984 1988 1992 1996 2000 2004 2008 20120.00.20.40.60.81.0MAI-NU (NYT)1980 1984 1988 1992 1996 2000 2004 2008 2012−1.0−0.50.00.51.01.5MAI-C1 (WSJ+NYT)1980 1984 1988 1992 1996 2000 2004 2008 2012−1.0−0.50.00.51.01.52.0MAI-C2 (WSJ+NYT)Interest MAI Fed Funds Rate05101520051015200510152005101520FedFundsRate(%)01234567MonetaryMAI1981 1985 1989 1993 1997 2001 2005 2009 2013051015200.51.01.52.02.53.03.54.04.55.0MonetaryMAIMAI-WU (WSJ)0.00.51.01.52.02.51981 1985 1989 1993 1997 2001 2005 2009 2013051015200.51.01.52.02.53.03.54.04.55.0MAI-NU (NYT)−1.5−1.0−0.50.00.51.01.52.02.51981 1985 1989 1993 1997 2001 2005 2009 2013051015200.51.01.52.02.53.03.54.04.55.0MAI-C1 (WSJ+NYT)−1.0−0.50.00.51.01.52.01981 1985 1989 1993 1997 2001 2005 2009 2013051015200.51.01.52.02.53.03.54.04.55.0FedFundsRate(first)BalanceSheet(second)MAI-C2 (WSJ+NYT)Monetary MAI Fed Fund Rate Fed Assets1980 1984 1988 1992 1996 2000 2004 2008 201201234567HousingMAIMAI-WU (WSJ)1980 1984 1988 1992 1996 2000 2004 2008 20120.00.20.40.60.81.01.21.41.61.8MAI-NU (NYT)1980 1984 1988 1992 1996 2000 2004 2008 2012−1.0−0.50.00.51.01.52.02.53.03.5MAI-C1 (WSJ+NYT)1980 1984 1988 1992 1996 2000 2004 2008 2012−1.0−0.50.00.51.01.52.02.53.0MAI-C2 (WSJ+NYT)Housing MAI Log Nominal Home Price Return−1.5−1.0−0.50.00.51.01.5−1.5−1.0−0.50.00.51.01.5−1.5−1.0−0.50.00.51.01.5−1.5−1.0−0.50.00.51.01.5LogHomePriceReturn183C.1. Sample of news articles mentioning macroeconomic fundamentals1980 1984 1988 1992 1996 2000 2004 2008 20120246810OilMAIMAI-WU (WSJ)1980 1984 1988 1992 1996 2000 2004 2008 20120.00.51.01.52.02.53.03.54.04.5MAI-NU (NYT)1980 1984 1988 1992 1996 2000 2004 2008 2012−2−1012345MAI-C1 (WSJ+NYT)1980 1984 1988 1992 1996 2000 2004 2008 2012−1.0−0.50.00.51.01.52.02.53.03.5MAI-C2 (WSJ+NYT)Oil MAI Oil Log Price2.02.53.03.54.04.55.02.02.53.03.54.04.55.02.02.53.03.54.04.55.02.02.53.03.54.04.55.0OilLogPrice1980 1984 1988 1992 1996 2000 2004 2008 20120123456UnemploymentMAIMAI-WU (WSJ)1980 1984 1988 1992 1996 2000 2004 2008 20120.00.51.01.52.02.53.0MAI-NU (NYT)1980 1984 1988 1992 1996 2000 2004 2008 2012−1.0−0.50.00.51.01.52.02.5MAI-C1 (WSJ+NYT)1980 1984 1988 1992 1996 2000 2004 2008 2012−1.0−0.50.00.51.01.52.02.53.0MAI-C2 (WSJ+NYT)Unemployment MAI Unemployment Rate34567891011345678910113456789101134567891011UnemploymentRate(%)1980 1984 1988 1992 1996 2000 2004 2008 20120.00.51.01.52.02.53.03.5USDMAIMAI-WU (WSJ)1980 1984 1988 1992 1996 2000 2004 2008 20120.000.050.100.150.200.250.300.350.400.45MAI-NU (NYT)1980 1984 1988 1992 1996 2000 2004 2008 2012−1.0−0.50.00.51.01.5MAI-C1 (WSJ+NYT)1980 1984 1988 1992 1996 2000 2004 2008 2012−1.0−0.50.00.51.01.5MAI-C2 (WSJ+NYT)USD MAI USD Log Price Index4.24.34.44.54.64.74.84.95.04.24.34.44.54.64.74.84.95.04.24.34.44.54.64.74.84.95.04.24.34.44.54.64.74.84.95.0USDLogPriceIndex184C.1.SampleofnewsarticlesmentioningmacroeconomicfundamentalsTable C.1: Descriptive Statistics and CorrelationThis table presents the descriptive statistics for the monthly unadjusted media attention indices(mai) for the Wall Street Journal (mai-wu) and New York Times (mai-nu), the Economic PolicyUncertainty (epu) index, the implied volatility (vxo), and the three-month detrended log S&P500 trade volume. Columns Jan to Dec are the monthly averages for each mai. Panels B showsthe correlation between the demeaned macroeconomic attention composite indices (mai-c1), epu,vxo, and the 60-day detrended S&P 500 trade volume at the monthly frequency.Table C.2: Panel A: Descriptive Statistics for Monthly Unadjusted MAIObs. Mean St. Dev. Min Max Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov DecWall Street JournalCredit Rating 376 0.60 0.56 0.00 3.87 0.59 0.61 0.60 0.51 0.52 0.65 0.61 0.68 0.56 0.58 0.62 0.65GDP 376 1.86 0.61 0.73 4.10 1.93 1.92 1.79 1.77 1.70 1.78 1.83 2.03 1.83 1.85 1.95 1.90Housing 376 0.90 1.01 0.00 6.47 1.00 0.87 0.86 0.86 0.93 0.96 0.94 0.96 0.88 0.92 0.83 0.83Inflation 376 2.96 0.82 1.43 6.85 3.15 3.08 2.93 2.81 3.00 3.05 2.79 3.00 2.98 2.81 2.87 3.01Interest 376 1.24 0.69 0.13 3.91 1.34 1.12 1.25 1.13 1.18 1.31 1.20 1.39 1.22 1.24 1.26 1.31Monetary 376 2.49 1.06 0.42 6.26 2.66 2.45 2.49 2.24 2.36 2.56 2.36 2.61 2.63 2.41 2.47 2.60Oil 376 3.07 1.94 0.61 9.37 3.13 2.87 3.13 3.09 3.08 2.99 2.89 3.15 3.13 3.20 3.03 3.22Unemp. 376 1.87 0.80 0.57 5.38 2.03 1.91 1.74 1.68 1.68 1.78 1.85 1.90 1.98 1.86 1.99 2.03USD 376 1.04 0.79 0.00 3.45 1.21 0.99 1.01 0.99 0.97 0.89 1.08 1.05 1.08 1.07 1.12 1.05New York TimesCredit Rating 419 0.20 0.23 0.00 2.91 0.23 0.19 0.17 0.17 0.17 0.18 0.20 0.21 0.19 0.21 0.23 0.22GDP 419 0.46 0.23 0.11 1.55 0.51 0.45 0.42 0.46 0.40 0.43 0.45 0.43 0.46 0.46 0.48 0.50Housing 419 0.23 0.28 0.00 1.62 0.28 0.27 0.21 0.18 0.18 0.17 0.23 0.28 0.25 0.26 0.20 0.22Inflation 419 0.82 0.48 0.03 2.70 0.97 0.85 0.81 0.74 0.82 0.87 0.83 0.81 0.82 0.78 0.74 0.82Interest 419 0.24 0.14 0.00 0.94 0.24 0.23 0.25 0.21 0.24 0.23 0.26 0.27 0.24 0.24 0.21 0.24Monetary 419 0.89 0.36 0.12 2.27 1.02 0.96 0.91 0.77 0.81 0.88 0.90 0.94 0.94 0.85 0.82 0.89Oil 419 0.74 0.58 0.00 4.46 0.82 0.75 0.78 0.72 0.68 0.71 0.72 0.78 0.73 0.75 0.63 0.77Unemp. 419 0.68 0.45 0.04 2.68 0.81 0.71 0.61 0.55 0.61 0.61 0.70 0.66 0.72 0.76 0.76 0.71USD 419 0.06 0.09 0.00 0.42 0.06 0.07 0.07 0.06 0.08 0.06 0.05 0.08 0.05 0.07 0.06 0.06Other VariablesEPU 360 101.33 41.96 37.27 271.83 127.67 106.13 94.75 82.98 86.87 89.70 94.48 95.44 107.89 112.99 111.94 105.12VXO 352 20.77 8.36 9.54 61.41 21.04 20.54 20.50 19.40 19.21 18.82 19.84 20.91 22.67 23.88 21.91 20.63Volume 419 0.01 0.09 -0.35 0.31 0.12 -0.04 0.05 0.02 -0.03 0.02 0.05 -0.03 0.00 0.07 -0.08 -0.04185C.1.SampleofnewsarticlesmentioningmacroeconomicfundamentalsPanel B: Monthly MAI-C1 correlation (1980-2015)Credit Rating GDP Housing Inflation Interest Monetary Oil Unemp. USD EPU VXO VolumeCredit Rating 1.00 0.48 0.30 -0.18 0.32 0.40 0.22 0.31 0.30 0.28 0.32 -0.01GDP 0.48 1.00 0.36 -0.14 0.20 0.40 0.07 0.64 0.10 0.13 0.18 -0.08Housing 0.30 0.36 1.00 0.03 0.45 0.48 0.16 0.20 0.06 -0.07 0.05 0.06Inflation -0.18 -0.14 0.03 1.00 0.35 0.36 0.43 -0.05 0.23 -0.01 0.03 0.06Interest 0.32 0.20 0.45 0.35 1.00 0.77 0.59 0.04 0.56 0.04 0.23 0.03Monetary 0.40 0.40 0.48 0.36 0.77 1.00 0.45 0.28 0.42 0.15 0.27 0.04Oil 0.22 0.07 0.16 0.43 0.59 0.45 1.00 -0.11 0.59 0.07 0.08 0.05Unemp. 0.31 0.64 0.20 -0.05 0.04 0.28 -0.11 1.00 -0.17 0.35 0.32 -0.05USD 0.30 0.10 0.06 0.23 0.56 0.42 0.59 -0.17 1.00 0.07 0.33 0.03EPU 0.28 0.13 -0.07 -0.01 0.04 0.15 0.07 0.35 0.07 1.00 0.44 0.05VXO 0.32 0.18 0.05 0.03 0.23 0.27 0.08 0.32 0.33 0.44 1.00 0.06Volume -0.01 -0.08 0.06 0.06 0.03 0.04 0.05 -0.05 0.03 0.05 0.06 1.00186C.1. Sample of news articles mentioning macroeconomic fundamentalsTable C.3: Persistence of Macroeconomic AttentionPanel A of this table presents ar (p) models of the monthly demeaned andstandardized media attention composite indices (mai-c2), controlling formonthly time-fixed effects. df (p-value) are the p-values for the Dickey-Fuller (df) statistics that test the null of a unit root in each time series.Panel B reports the estimates from an ols regression of the daily demeanedand standardized media attention composite indices (mai-c2) on variousmoving average lags of itself. L1 corresponds to the lag of itself and L5,L21, L62, L250, and L1000 are the moving average for 5, 21, 62, 250, and1000 days preceding the observed values at time t. We control for day-of-week fixed effects. The standard errors are reported in parenthesis and arecalculated using Newey-West standard errors (10 lags). Obs. stands for thenumber of observations. *, **, and *** denote the statistical significance atthe 10%, 5%, 1% levels, respectively.Panel A: Monthly MAI-C2 AR(4) Coefficients and DF statisticsCredit Rating GDP Housing Inflation Interest Monetary Oil Unemp. USDconst 0.02 0.05 -0.01 0.08** 0.03 0.03 0.11** -0.01 -0.04(0.05) (0.04) (0.05) (0.03) (0.04) (0.04) (0.05) (0.04) (0.03)AR(1) 0.66*** 0.26*** 0.60*** 0.49*** 0.53*** 0.47*** 0.66*** 0.67*** 0.54***(0.07) (0.06) (0.10) (0.05) (0.05) (0.04) (0.05) (0.06) (0.06)AR(2) 0.01 0.28*** 0.09 0.25*** 0.15** 0.15*** 0.18*** 0.13** 0.19***(0.07) (0.04) (0.08) (0.05) (0.07) (0.05) (0.05) (0.06) (0.05)AR(3) 0.05 0.31*** 0.14 0.08 -0.03 0.08* 0.08 0.10* 0.13**(0.05) (0.06) (0.09) (0.05) (0.05) (0.04) (0.10) (0.06) (0.05)AR(4) 0.09 0.06 0.03 0.09** 0.17*** 0.06 -0.02 0.01 0.07(0.05) (0.05) (0.08) (0.04) (0.04) (0.04) (0.06) (0.05) (0.06)DF (p-value) 0.00 0.05 0.02 0.00 0.00 0.00 0.00 0.00 0.09Adj-R2 0.55 0.66 0.64 0.76 0.52 0.44 0.75 0.78 0.77Obs. 415 415 415 415 415 415 415 415 415Panel B: Daily MAI-C2 Frequency RegressionsCredit Rating GDP Housing Inflation Interest Monetary Oil Unemployment U.S. Dollarconst -0.15*** 0.00 -0.21*** -0.02 -0.10*** -0.20*** -0.18*** -0.03 -0.22***(0.03) (0.03) (0.02) (0.03) (0.03) (0.03) (0.03) (0.03) (0.02)L1 0.08*** 0.07*** 0.04* 0.06*** 0.13*** 0.19*** 0.11*** 0.04** 0.01(0.02) (0.01) (0.02) (0.01) (0.02) (0.02) (0.03) (0.02) (0.01)L5 0.28*** 0.12*** 0.46*** 0.13*** 0.15*** 0.18*** 0.39*** 0.22*** 0.16***(0.06) (0.03) (0.07) (0.03) (0.03) (0.03) (0.04) (0.03) (0.03)L21 0.40*** 0.06 0.23*** 0.26*** 0.27*** 0.23*** 0.30*** 0.25*** 0.39***(0.09) (0.07) (0.08) (0.06) (0.06) (0.05) (0.06) (0.06) (0.06)L62 0.06 0.34*** 0.06 0.36*** 0.15* 0.13* 0.13** 0.26*** 0.29***(0.06) (0.10) (0.07) (0.07) (0.08) (0.07) (0.05) (0.08) (0.07)L250 0.08 0.41*** 0.17** 0.08 0.25*** 0.20*** 0.01 0.23*** 0.14**(0.06) (0.11) (0.08) (0.06) (0.07) (0.07) (0.03) (0.06) (0.05)L1000 0.02 -0.05 0.01 0.05 -0.01 0.00 0.03 -0.08*** -0.03(0.05) (0.06) (0.06) (0.04) (0.04) (0.05) (0.02) (0.03) (0.03)Obs. 8109 8109 8109 8109 8109 8109 8109 8109 8109Adj-R2 0.28 0.18 0.42 0.20 0.18 0.25 0.52 0.36 0.34187C.1.SampleofnewsarticlesmentioningmacroeconomicfundamentalsTable C.4: Media Attention and Macroeconomic FundamentalsThis table presents the results of an ols regression of monthly macroeconomic media attention indices (mai) on different macroeconomicfundamentals. Panels A and Panel B report the results for the demeaned composite index (mai-c1) and the demeaned and standardizedcomposite index (mai-c2), respectively. The general regression is specified in equation 4.6. F corresponds to the associated fundamentalto each mai as described in Table 4.2 and Ft is the moving average over t days of the respective fundamental. We control for monthlyfixed effects. The standard errors are reported in parenthesis and are calculated using Newey-West standard errors (5 lags). Obs.stands for the number of observations. *, **, *** denote the statistic significance at the 10%, 5%, 1% levels, respectively.Panel A: MAI-C1 (Demeaned)MAI: Credit Rating GDP Housing Inflation Interest Monetary Oil Unemployment US DollarF: Credit Rating Spreads GDP Growth Home Price Ret ∆ CPI Fed Fund Fed Fund Oil Price Ret Unemp. Rate USD Index RetFt − Ft,t−3 0.034** -0.250 -0.234** -0.042 -0.031 -0.009 -0.013 0.004(0.015) (0.176) (0.104) (0.040) (0.057) (0.006) (0.175) (0.006)Ft,t−3 − Ft,t−12 0.011 0.117 -0.462*** -0.085 -0.005 -0.015 0.010 0.164 -0.007(0.007) (0.072) (0.160) (0.234) (0.033) (0.049) (0.013) (0.125) (0.019)Ft,t−12 − Ft,t−48 0.003 0.224 -0.097 2.268*** 0.010 -0.000 0.108** 0.171*** -0.186***(0.015) (0.184) (0.180) (0.648) (0.028) (0.041) (0.054) (0.062) (0.063)(Ft − Ft,t−3)2 -0.001 0.517* -0.407* 0.007 0.018 0.004*** 1.022 0.007**(0.002) (0.269) (0.218) (0.023) (0.025) (0.001) (0.782) (0.004)(Ft,t−3 − Ft,t−12)2 0.000 0.185** 0.451*** 0.288 0.015 0.040** 0.005*** 0.232** 0.023**(0.000) (0.084) (0.141) (0.234) (0.015) (0.020) (0.001) (0.104) (0.010)(Ft,t−12 − Ft,t−48)2 0.001 0.295 1.418*** 9.858*** 0.007 0.001 -0.005 0.075*** 0.141**(0.001) (0.296) (0.329) (1.605) (0.007) (0.013) (0.011) (0.026) (0.067)const -0.031 -0.076 -0.472*** -0.062 -0.006 0.010 -0.300 -0.061 -0.099(0.041) (0.076) (0.054) (0.077) (0.068) (0.093) (0.183) (0.078) (0.070)Obs. 419 125 419 419 419 419 376 419 419Adj-R2 0.10 0.08 0.49 0.19 0.01 0.01 0.15 0.50 0.14188C.1.SampleofnewsarticlesmentioningmacroeconomicfundamentalsPanel B: MAI-C2 (Demeaned and Standardized)MAI: Credit Rating GDP Housing Inflation Interest Monetary Oil Unemployment US DollarF: Credit Rating Spreads GDP Growth Home Price Ret ∆ CPI Fed Fund Fed Fund Oil Price Ret Unemp. Rate USD Index RetFt − Ft,t−3 0.049** -0.312* -0.216** -0.054 -0.016 -0.005 -0.024 0.004(0.023) (0.177) (0.086) (0.049) (0.044) (0.004) (0.171) (0.006)Ft,t−3 − Ft,t−12 0.010 0.300* -0.501*** -0.378* -0.001 -0.017 0.006 0.184 -0.007(0.009) (0.171) (0.164) (0.203) (0.032) (0.032) (0.008) (0.113) (0.018)Ft,t−12 − Ft,t−48 -0.006 0.636 -0.045 1.729** -0.008 -0.017 0.060** 0.166*** -0.225***(0.023) (0.463) (0.180) (0.704) (0.024) (0.025) (0.027) (0.053) (0.069)(Ft − Ft,t−3)2 -0.001 0.697*** -0.456** 0.053** 0.039** 0.003*** 0.949 0.007*(0.003) (0.225) (0.189) (0.022) (0.015) (0.000) (0.801) (0.004)(Ft,t−3 − Ft,t−12)2 0.000 0.414** 0.450*** -0.028 0.032* 0.050*** 0.003*** 0.236** 0.009(0.001) (0.191) (0.135) (0.183) (0.017) (0.016) (0.001) (0.119) (0.012)(Ft,t−12 − Ft,t−48)2 0.002 0.819 1.172*** 9.650*** 0.015** -0.000 -0.005 0.070*** 0.081(0.001) (0.751) (0.344) (1.955) (0.006) (0.008) (0.006) (0.026) (0.074)const -0.045 -0.194 -0.451*** -0.109 -0.064 -0.027 -0.219*** -0.067 -0.091(0.059) (0.205) (0.056) (0.072) (0.064) (0.068) (0.084) (0.070) (0.080)Obs. 419 125 419 419 419 419 376 419 419Adj-R2 0.08 0.08 0.47 0.22 0.12 0.07 0.25 0.54 0.09189C.1. Sample of news articles mentioning macroeconomic fundamentalsTable C.5: Media Attention and Aggregate Trade VolumeThis table presents the results of an ols regression of the daily detrendedS&P 500 trade volume on the difference between the 5-day and 20-day mov-ing average mai-c2 and a dummy (Ann) equal to one if there is a relatedannouncement specified in Table 4.2, zero otherwise. We detrend the logtrade volume using the moving average of the log trade volume of the past60 trading days. For all model specifications, we control for day-of-weekfixed effects. The standard errors are reported in parenthesis and are cal-culated using Newey-West standard errors (250 lags). Obs. stands for thenumber of observations. *, **, *** denote the statistical significance at the10%, 5%, 1% levels, respectively.MAI: Inflation Monetary InterestAnn: CPI FOMC FOMCMAI5−20 0.059*** 0.058*** 0.063*** 0.086*** 0.085*** 0.086*** 0.049*** 0.048*** 0.049***(0.013) (0.013) (0.014) (0.011) (0.011) (0.011) (0.011) (0.011) (0.011)Ann 0.035*** 0.042*** 0.027*** 0.027*** 0.030*** 0.031***(0.007) (0.007) (0.009) (0.010) (0.009) (0.009)MAI5−20×Ann -0.114*** -0.011 -0.033(0.032) (0.038) (0.032)const 0.003 0.000 0.001 0.002 0.002 0.002 0.003 0.002 0.002(0.006) (0.006) (0.006) (0.006) (0.006) (0.006) (0.006) (0.006) (0.006)Obs. 8787 8787 8787 8787 8787 8787 8787 8787 8787Adj-R2 0.06 0.06 0.06 0.07 0.07 0.07 0.05 0.05 0.05MAI: GDP Unemployment Credit Rating Oil USDAnn: GDP Report EmploymentMAI5−20 0.019* 0.019* 0.017 0.034*** 0.033*** 0.034*** 0.043*** 0.043** 0.027*(0.011) (0.011) (0.011) (0.012) (0.012) (0.012) (0.013) (0.017) (0.014)Ann 0.005 0.003 0.014 0.017(0.008) (0.008) (0.011) (0.012)MAI5−20×Ann 0.058 -0.031(0.041) (0.039)const 0.002 0.002 0.002 0.003 -0.001 -0.000 0.002 0.013** 0.028***(0.006) (0.006) (0.006) (0.006) (0.007) (0.007) (0.006) (0.007) (0.006)Obs. 8787 8787 8787 8787 8787 8787 8787 7368 8321Adj-R2 0.05 0.05 0.05 0.05 0.05 0.05 0.05 0.05 0.06190C.1. Sample of news articles mentioning macroeconomic fundamentalsTable C.6: Media Attention and Implied VolatilityThis table presents the results of an ols regression of the daily impliedvolatility proxied by vxo regressed on the difference between the 20-day and250-day moving average mai-c2 and a dummy (Ann) equal to one if there isa related announcement specified in Table 4.2, zero otherwise. For all modelspecifications, we control for day-of-week fixed effects. The standard errorsare reported in parenthesis and are calculated using Newey-West standarderrors (250 lags). Obs. stands for the number of observations. *, **, ***denote the statistical significance at the 10%, 5%, 1% levels, respectively.MAI: Inflation Monetary InterestAnn: CPI FOMC FOMCMAI20−250 -2.427 -2.425 -2.466 5.647** 5.646** 5.668** 5.671** 5.670** 5.698**(4.705) (4.706) (4.667) (2.415) (2.415) (2.416) (2.558) (2.558) (2.562)Ann 0.265 0.277 -0.178 -0.187 -0.196 -0.204(0.185) (0.189) (0.221) (0.224) (0.222) (0.229)MAI20−250×Ann 0.881 -0.750 -0.846(1.157) (0.732) (1.053)const 20.728*** 20.711*** 20.711*** 20.719*** 20.720*** 20.720*** 20.724*** 20.724*** 20.724***(1.240) (1.236) (1.236) (1.245) (1.245) (1.245) (1.253) (1.253) (1.253)Obs. 7386 7386 7386 7386 7386 7386 7386 7386 7386Adj-R2 0.00 0.00 0.00 0.03 0.03 0.03 0.03 0.03 0.03MAI: GDP Unemployment Credit Rating Oil USDAnn: GDP Report EmploymentMAI20−250 12.939*** 12.946*** 12.995*** 14.035*** 14.037*** 14.075*** 5.462*** 1.148 4.202**(5.008) (5.009) (4.994) (4.866) (4.866) (4.879) (1.719) (1.781) (1.921)Ann 0.297 0.284 0.222 0.221(0.199) (0.202) (0.155) (0.159)MAI20−250×Ann -0.973 -0.781(1.097) (0.996)const 20.632*** 20.609*** 20.609*** 20.633*** 20.583*** 20.582*** 20.766*** 20.763*** 20.777***(1.124) (1.120) (1.120) (1.066) (1.067) (1.066) (1.216) (1.252) (1.250)Obs. 7386 7386 7386 7386 7386 7386 7361 7361 7005Adj-R2 0.08 0.08 0.08 0.15 0.15 0.15 0.05 0.00 0.01191C.1. Sample of news articles mentioning macroeconomic fundamentalsTable C.7: Unemployment Surprise ForecastsThis table presents the results of an ols regression of the unemploymentsurprise regressed on various detrended daily media attention indices at dif-ferent frequencies and an interaction term between the detrended mediaattention indices and an nber dummy. The nber dummy is equal to one ifthe unemployment surprise occurs during a nber recession, zero otherwise.Panel A shows the result for mai-wu, mai-nu in Panel B, and mai-c2 inPanel C. We use three different unemployment surprises. Each surprise iscalculated as the difference between the actual unemployment for month treported in month t + 1 and (1) the random-walk (i.e. the previous monthunemployment rate), (2) the forecasted unemployment rate as in Boyd, Hu,and Jagannathan (2005), or (3) the median of the forecasted unemploy-ment rate by economists surveyed by Bloomberg. The standard errors arereported in parenthesis and are calculated using the White’s heteroskedas-ticity robust standard errors. Obs. stands for the number of observations.*, **, *** denote the statistical significance at the 10%, 5%, 1% levels,respectively.Panel A: MAI-WU (Wall Street Journal)Random-WalkMAI: MAI5−20 MAI5−250 MAI20−250 MAI60−250MAI 0.030* 0.015 0.035*** 0.013 0.054** 0.006 0.096** 0.002(0.016) (0.016) (0.013) (0.012) (0.026) (0.025) (0.037) (0.037)MAI×NBER 0.200*** 0.128*** 0.174*** 0.319***(0.066) (0.029) (0.053) (0.051)const -0.013 -0.013 -0.011 -0.014 -0.004 -0.011 -0.003 -0.014(0.010) (0.010) (0.009) (0.009) (0.009) (0.009) (0.009) (0.009)Obs. 375 375 364 364 364 364 364 364Adj-R2 0.01 0.04 0.02 0.07 0.02 0.05 0.03 0.09Boyd et al. (2005) SurpriseMAI: MAI5−20 MAI5−250 MAI20−250 MAI60−250MAI 0.019 0.014 0.024** 0.016 0.044** 0.025 0.068*** 0.034(0.013) (0.013) (0.011) (0.011) (0.018) (0.020) (0.025) (0.027)MAI×NBER 0.057 0.047* 0.068* 0.117***(0.057) (0.028) (0.039) (0.045)const -0.020*** -0.020*** -0.019** -0.020*** -0.014* -0.017** -0.014* -0.018**(0.008) (0.008) (0.008) (0.008) (0.007) (0.008) (0.007) (0.008)Obs. 375 375 364 364 364 364 364 364Adj-R2 0.00 0.00 0.02 0.02 0.02 0.02 0.02 0.03192C.1. Sample of news articles mentioning macroeconomic fundamentalsPanel A: Continued.Bloomberg SurpriseMAI: MAI5−20 MAI5−250 MAI20−250 MAI60−250MAI 0.033** 0.021 0.019* 0.009 0.005 -0.014 0.013 -0.028(0.015) (0.015) (0.011) (0.012) (0.020) (0.025) (0.029) (0.037)MAI×NBER 0.138*** 0.049** 0.059 0.118**(0.046) (0.022) (0.040) (0.051)const -0.039*** -0.039*** -0.035*** -0.037*** -0.031*** -0.035*** -0.031*** -0.037***(0.010) (0.010) (0.010) (0.010) (0.010) (0.011) (0.010) (0.011)Obs. 217 217 217 217 217 217 217 217Adj-R2 0.02 0.05 0.01 0.02 -0.00 -0.00 -0.00 0.01193C.1. Sample of news articles mentioning macroeconomic fundamentalsPanel B: MAI-NU (New York Times MAI)Random-WalkMAI: MAI5−20 MAI5−250 MAI20−250 MAI60−250MAI 0.000 0.001 0.079*** 0.051** 0.186*** 0.131*** 0.294*** 0.178***(0.037) (0.036) (0.026) (0.026) (0.039) (0.040) (0.057) (0.062)MAI×NBER -0.005 0.210** 0.224** 0.503***(0.181) (0.104) (0.112) (0.141)const -0.006 -0.006 -0.008 -0.013 -0.002 -0.009 -0.003 -0.013(0.010) (0.009) (0.009) (0.009) (0.009) (0.009) (0.009) (0.009)Obs. 418 418 407 407 407 407 407 407Adj-R2 -0.00 -0.00 0.03 0.05 0.06 0.08 0.08 0.12Boyd et al. (2005) SurpriseMAI: MAI5−20 MAI5−250 MAI20−250 MAI60−250MAI -0.001 -0.002 0.041* 0.034 0.095*** 0.090** 0.164*** 0.125**(0.032) (0.034) (0.021) (0.023) (0.031) (0.035) (0.048) (0.058)MAI×NBER 0.005 0.052 0.021 0.170*(0.111) (0.057) (0.077) (0.101)const -0.015** -0.015** -0.017** -0.018** -0.014* -0.015* -0.014* -0.018**(0.008) (0.008) (0.008) (0.008) (0.007) (0.008) (0.007) (0.008)Obs. 418 418 407 407 407 407 407 407Adj-R2 -0.00 -0.00 0.01 0.01 0.02 0.02 0.04 0.04Bloomberg SurpriseMAI: MAI5−20 MAI5−250 MAI20−250 MAI60−250MAI -0.001 0.010 0.019 0.014 0.048 0.025 0.015 -0.069(0.038) (0.040) (0.029) (0.032) (0.045) (0.058) (0.065) (0.080)MAI×NBER -0.150 0.032 0.069 0.270**(0.118) (0.070) (0.091) (0.130)const -0.031*** -0.031*** -0.032*** -0.033*** -0.031*** -0.033*** -0.031*** -0.037***(0.010) (0.010) (0.010) (0.010) (0.010) (0.011) (0.010) (0.010)Obs. 217 217 217 217 217 217 217 217Adj-R2 -0.00 -0.00 -0.00 -0.01 0.00 -0.00 -0.00 0.01194C.1. Sample of news articles mentioning macroeconomic fundamentalsPanel C: MAI-C2 (Demeaned and Standardized MAI)Random-WalkMAI: MAI5−20 MAI5−250 MAI20−250 MAI60−250MAI 0.036 0.017 0.083*** 0.051** 0.158*** 0.110*** 0.234*** 0.136***(0.032) (0.031) (0.021) (0.021) (0.034) (0.034) (0.046) (0.051)MAI×NBER 0.228 0.211*** 0.180* 0.382***(0.170) (0.066) (0.093) (0.103)const -0.009 -0.008 -0.011 -0.017* -0.002 -0.009 -0.002 -0.012(0.010) (0.010) (0.009) (0.009) (0.009) (0.009) (0.009) (0.009)Obs. 418 418 407 407 407 407 407 407Adj-R2 0.00 0.01 0.04 0.08 0.07 0.08 0.09 0.12Boyd et al. (2005) SurpriseMAI: MAI5−20 MAI5−250 MAI20−250 MAI60−250MAI 0.021 0.013 0.049*** 0.038** 0.092*** 0.084*** 0.135*** 0.099**(0.028) (0.029) (0.018) (0.019) (0.025) (0.030) (0.038) (0.048)MAI×NBER 0.096 0.070 0.031 0.142**(0.104) (0.048) (0.057) (0.071)const -0.017** -0.017** -0.019** -0.021*** -0.013* -0.015* -0.013* -0.017**(0.008) (0.008) (0.008) (0.008) (0.007) (0.008) (0.007) (0.008)Obs. 418 418 407 407 407 407 407 407Adj-R2 -0.00 -0.00 0.02 0.03 0.03 0.03 0.04 0.05Bloomberg SurpriseMAI: MAI5−20 MAI5−250 MAI20−250 MAI60−250MAI 0.049 0.036 0.031 0.017 0.027 -0.002 0.018 -0.058(0.033) (0.034) (0.022) (0.025) (0.035) (0.047) (0.050) (0.065)MAI×NBER 0.335** 0.072 0.079 0.212**(0.168) (0.047) (0.072) (0.093)const -0.036*** -0.038*** -0.034*** -0.036*** -0.031*** -0.034*** -0.031*** -0.038***(0.011) (0.011) (0.010) (0.010) (0.010) (0.011) (0.010) (0.011)Obs. 217 217 217 217 217 217 217 217Adj-R2 0.01 0.02 0.01 0.01 -0.00 -0.00 -0.00 0.01195C.1. Sample of news articles mentioning macroeconomic fundamentalsTable C.8: S&P Return Forecast on Employment Situation AnnouncementDaysThis table presents the results of an ols regression of the daily S&P 500 logreturn on the employment situation announcement date regressed on the un-employment surprise as in Boyd, Hu, and Jagannathan (2005), the surpriseinteracted with an nber dummy, the daily detrended unemployment mediaattention index composite index mai-c2, and the detrended unemploymentmedia attention index interacted with an nber dummy. The nber dummyis equal to one if the unemployment surprise occurs during a nber recession,zero otherwise. We show the results for two different detrended frequenciesfor the unemployment media attention index. The standard errors are re-ported in parenthesis and are calculated using the White’s heteroskedasticityrobust standard errors. Obs. stands for the number of observations. *, **,*** denote the statistical significance at the 10%, 5%, 1% levels, respectively.MAI: MAI5−20 MAI20−250MAI 0.395** 0.372** 0.350** 0.282 -0.053 -0.105(0.172) (0.174) (0.175) (0.194) (0.193) (0.192)MAI·NBER 0.288 0.443 1.256** 1.502***(0.756) (0.724) (0.488) (0.483)SurpBoyd 0.615* 0.585* 0.724**(0.354) (0.351) (0.368)SurpBoyd×NBER -1.938* -2.174* -3.070**(1.133) (1.273) (1.283)const 0.047 -0.009 -0.009 0.017 0.031 -0.017 0.007(0.057) (0.061) (0.061) (0.062) (0.058) (0.059) (0.059)Obs. 423 418 418 418 407 407 407Adj-R2 0.01 0.01 0.01 0.01 0.00 0.02 0.04196

Cite

Citation Scheme:

        

Citations by CSL (citeproc-js)

Usage Statistics

Share

Embed

Customize your widget with the following options, then copy and paste the code below into the HTML of your page to embed this item in your website.
                        
                            <div id="ubcOpenCollectionsWidgetDisplay">
                            <script id="ubcOpenCollectionsWidget"
                            src="{[{embed.src}]}"
                            data-item="{[{embed.item}]}"
                            data-collection="{[{embed.collection}]}"
                            data-metadata="{[{embed.showMetadata}]}"
                            data-width="{[{embed.width}]}"
                            async >
                            </script>
                            </div>
                        
                    
IIIF logo Our image viewer uses the IIIF 2.0 standard. To load this item in other compatible viewers, use this url:
http://iiif.library.ubc.ca/presentation/dsp.24.1-0348752/manifest

Comment

Related Items