@prefix vivo: . @prefix edm: . @prefix ns0: . @prefix dcterms: . @prefix skos: . vivo:departmentOrSchool "Business, Sauder School of"@en, "Finance, Division of"@en ; edm:dataProvider "DSpace"@en ; ns0:degreeCampus "UBCV"@en ; dcterms:creator "Sheng, Jinfei"@en ; dcterms:issued "2018-05-16T18:18:13Z"@en, "2018"@en ; vivo:relatedDegree "Doctor of Philosophy - PhD"@en ; ns0:degreeGrantor "University of British Columbia"@en ; dcterms:description "The relationship between information and stock returns is one of the most fundamental questions in finance and economics. This thesis aims to enhance our understanding of this relationship by using novel datasets and methods. In particular, I present a collection of three essays on the impacts of information on stock returns. I study both traditional sources of information, such as earnings announcements, macroeconomic news, and media coverage, as well as non-traditional sources of information, such as online employee reviews. The first essay “Asset Pricing in the Information Age: Employee Expectations and Stock Returns” studies the investment value of employees' information in financial markets, using a novel dataset of nearly one million employee reviews. This essay shows that employee expectations of their employers' business prospects predict future returns. I find that employee reviews are related to firms' fundamentals because they predict cash flow news. This essay highlights the importance of online information about firms' fundamentals, which is beyond traditional information sources such as analyst forecasts. Investors often face multiple types of news at the same time. Thus, the interaction between different types of news is crucial for understanding how information is incorporated into stock prices. My second essay “Macro News, Micro News, and Stock Prices” investigates interactions between macro-announcements and the processing of earnings news. Existing theories suggest that macro-news should crowd out attention to firm-level news, implying less efficient pricing. However, I find the opposite: on macro-news days price efficiency of earnings announcements is better when macroeconomic announcements are released on the same day. The news text is a new source of data which allows us to measure intangible but important variable, such as investor attention. In my third essay, “Media Attention, Macroeconomic Fundamentals, and the Stock Markets” (co-authored with Adlai Fisher and Charles Martineau), we construct indices of media attention to macroeconomic risks including employment, growth, and monetary policy. We study the properties of these attention indices and link them to stock markets. We conclude that media attention to macroeconomic fundamentals provides market-relevant information beyond the contents and dates of macroeconomic announcements."@en ; edm:aggregatedCHO "https://circle.library.ubc.ca/rest/handle/2429/65931?expand=metadata"@en ; skos:note "Essays on Information and Stock ReturnsbyJinfei ShengBA., Nankai University, 2007MA., Nankai University, 2009M.S., Texas A&M University, 2011A THESIS SUBMITTED IN PARTIAL FULFILLMENT OFTHE REQUIREMENTS FOR THE DEGREE OFDOCTOR OF PHILOSOPHYinThe Faculty of Graduate and Postdoctoral Studies(Business Administration)THE UNIVERSITY OF BRITISH COLUMBIA(Vancouver)May 2018© Jinfei Sheng 2018The following individuals certify that they have read, and recommend to the Faculty of Graduate and Post-doctoral Studies for acceptance, the dissertation entitled:Essays on Information and Stock Returnssubmitted by Jinfei Sheng in partial fulfillment of the requirements forthe degree of Doctor of Philosophyin Business AdministrationExamining Committee:Adlai Fisher, Business AdministrationCo-supervisorMurray Carlson, Business AdministrationCo-supervisorKai Li, Business AdministrationSupervisory Committee MemberWerner Antweiler, Business AdministrationUniversity ExaminerGiovanni Gallipoli, EconomicsUniversity ExaminerAdditional Supervisory Committee Members:Jan Bena, Business AdministrationSupervisory Committee MemberCarolin Pflueger, Business AdministrationSupervisory Committee MemberJesse Perla, EconomicsSupervisory Committee MemberiiAbstractThe relationship between information and stock returns is one of the most fundamental questions in financeand economics. This thesis aims to enhance our understanding of this relationship by using novel datasetsand methods. In particular, I present a collection of three essays on the impacts of information on stockreturns. I study both traditional sources of information, such as earnings announcements, macroeconomicnews, and media coverage, as well as non-traditional sources of information, such as online employee re-views. The first essay “Asset Pricing in the Information Age: Employee Expectations and Stock Returns”studies the investment value of employees’ information in financial markets, using a novel dataset of nearlyone million employee reviews. This essay shows that employee expectations of their employers’ businessprospects predict future returns. I find that employee reviews are related to firms’ fundamentals because theypredict cash flow news. This essay highlights the importance of online information about firms’ fundamen-tals, which is beyond traditional information sources such as analyst forecasts. Investors often face multipletypes of news at the same time. Thus, the interaction between different types of news is crucial for under-standing how information is incorporated into stock prices. My second essay “Macro News, Micro News,and Stock Prices” investigates interactions between macro-announcements and the processing of earningsnews. Existing theories suggest that macro-news should crowd out attention to firm-level news, implyingless efficient pricing. However, I find the opposite: on macro-news days price efficiency of earnings an-nouncements is better when macroeconomic announcements are released on the same day. The news text isa new source of data which allows us to measure intangible but important variable, such as investor attention.In my third essay, “Media Attention, Macroeconomic Fundamentals, and the Stock Markets” (co-authoredwith Adlai Fisher and Charles Martineau), we construct indices of media attention to macroeconomic risksincluding employment, growth, and monetary policy. We study the properties of these attention indices andlink them to stock markets. We conclude that media attention to macroeconomic fundamentals providesmarket-relevant information beyond the contents and dates of macroeconomic announcements.iiiLay SummaryThis essay presents a collection of three essays on the relationship between information and stock returns,which is one of the most fundamental questions in financial economics. The first essay “Asset Pricingin the Information Age: Employee Expectations and Stock Returns” studies the investment value of em-ployees’ information in financial markets, using a novel dataset of nearly one million employee reviews.My second essay “Macro News, Micro News, and Stock Prices” investigates interactions between macro-announcements and the processing of earnings news and shows novel evidence of the complementaritybetween two types of information. In the third essay, “Media Attention, Macroeconomic Fundamentals, andthe Stock Markets” (co-authored with Adlai Fisher and Charles Martineau), we construct indices of mediaattention to macroeconomic risks including employment, growth, and monetary policy and study their im-pact on stock markets. In summary, this thesis helps us to better understand the importance of informationin stock markets.ivPrefaceChapters 2 and 3 are based solely on my own work. Chapter 4 is a co-authored project with my adviserProfessor Adlai J. Fisher and former Ph.D. colleague Charles Martineau. We contributed equally to thewriting and to the empirical analysis.vTable of ContentsAbstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iiiLay Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ivPreface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vTable of Contents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . viList of Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . viiiList of Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xAcknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xiDedication . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xii1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12 Asset Pricing in the Information Age: Employee Expectations and Stock Returns. . . . . . . . 22.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22.2 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62.3 Employee outlook: is it valuable to the stock market? . . . . . . . . . . . . . . . . . . . . 92.4 Information processing: alpha decay and trading . . . . . . . . . . . . . . . . . . . . . . . 152.5 Information hierarchies: Employee ranks and the value of employee outlooks . . . . . . . . 172.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 213 Macro News, Micro News, and Stock Prices. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 383.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 383.2 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 423.3 The effect of macro news on the processing of earnings news . . . . . . . . . . . . . . . . . 433.4 Investor attention . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 483.5 Alternative explanations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 493.6 Robustness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 523.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54vi4 Media Attention, Macroeconomic Fundamentals, and the Stock Market . . . . . . . . . . . . 714.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 714.2 Macroeconomic Attention Indices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 744.3 Attention and Macroeconomic Fundamentals . . . . . . . . . . . . . . . . . . . . . . . . . 784.4 Attention and Stock Market Activity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 804.5 Using Attention for Forecasting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 814.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 865 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108AppendicesA Appendix to Chapter 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117B Appendix to Chapter 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128C Appendix to Chapter 4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132viiList of Tables2.1 Distribution of employee reviews . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 252.2 Summary statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 262.3 Employee expectations and firm characteristics . . . . . . . . . . . . . . . . . . . . . . . . 272.4 Employee expectations and stock returns . . . . . . . . . . . . . . . . . . . . . . . . . . . . 282.5 Employee expectations and stock returns: subsample . . . . . . . . . . . . . . . . . . . . . 292.6 Fama-MacBeth regressions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 302.7 Employee expectations and firms’ earnings . . . . . . . . . . . . . . . . . . . . . . . . . . 312.8 Employee expectations and profitability . . . . . . . . . . . . . . . . . . . . . . . . . . . . 322.9 Employee expectations and stock returns: robustness . . . . . . . . . . . . . . . . . . . . . 332.10 Employee expectations and institutional trading . . . . . . . . . . . . . . . . . . . . . . . . 342.11 Employee outlook and insider trading . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 352.12 Information hierarchies within firms—evidence from returns . . . . . . . . . . . . . . . . . 362.13 Information hierarchies within firms—evidence from textual analysis . . . . . . . . . . . . . 373.1 Summary statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 563.2 The macro-news effect–top groups . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 573.3 The macro-news effect–all firms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 583.4 Drift over different horizons . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 593.5 Lead and lag effects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 603.6 Heterogeneity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 613.7 Many macroeconomic announcements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 623.8 Trading strategy on drift portfolios . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 633.9 Investor attention . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 643.10 Volume reaction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 653.11 Changes in risk . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 663.12 Trading frictions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 673.13 Information spillover from macro news . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 683.14 Strategic timing of earning announcements . . . . . . . . . . . . . . . . . . . . . . . . . . . . 693.15 Robustness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 704.1 Newspapers Search Words . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 944.2 Macroeconomic Attention and Macroeconomic Fundamentals . . . . . . . . . . . . . . . . 954.3 Descriptive Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96viii4.4 Descriptive Statistics (cont.) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 974.5 Persistence of Macroeconomic Attention . . . . . . . . . . . . . . . . . . . . . . . . . . . . 984.6 Macroeconomic Attention and Macroeconomic Fundamentals . . . . . . . . . . . . . . . . 994.7 Media Attention and Aggregate Trade Volume . . . . . . . . . . . . . . . . . . . . . . . . . 1014.8 Media Attention and Implied Volatility . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1024.9 Forecasting Unemployment Surprises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1034.10 Forecasting Returns and Changes in VIX on Employment Situation Announcements . . . . 1044.11 Forecasting Returns and Changes in VIX on FOMC Announcements . . . . . . . . . . . . . 105A.1 Mean values of positive outlook and star ratings in each month . . . . . . . . . . . . . . . . 124A.2 Employee outlook and other components in reviews . . . . . . . . . . . . . . . . . . . . . 125A.3 Correlations and PCA analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126A.4 Machine learning methods and their performance . . . . . . . . . . . . . . . . . . . . . . . 127A.5 Top five words in each topic from LDA . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127B.1 Characteristics of Macroeconomic Announcements . . . . . . . . . . . . . . . . . . . . . . . . 130B.2 Volume reaction with an alternative measure . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130B.3 Macro announcements and market risk premium . . . . . . . . . . . . . . . . . . . . . . . . . . 131C.1 Descriptive Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137C.2 Persistence of Macroeconomic Attention . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138C.3 Macroeconomic Attention and Macroeconomic Fundamentals . . . . . . . . . . . . . . . . 139C.4 Macroeconomic Attention and Aggregate Trade Volume . . . . . . . . . . . . . . . . . . . 140C.5 Media Attention and Implied Volatility . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141C.6 Forecasting Unemployment Surprises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142C.6 Forecasting Unemployment Surprises (cont.) . . . . . . . . . . . . . . . . . . . . . . . . . 143C.6 Forecasting Unemployment Surprises (cont.) . . . . . . . . . . . . . . . . . . . . . . . . . 144C.7 Changes in VIX before and on Announcement Days . . . . . . . . . . . . . . . . . . . . . . 145C.8 Forecasting Returns and Changes in VIX on Employment Situation Announcements . . . . 146C.9 Forecasting Returns and Changes in VIX on FOMC Announcements . . . . . . . . . . . . . 147ixList of Figures2.1 Top words in reviews with positive and negative outlook . . . . . . . . . . . . . . . . . . . 222.2 Cumulative returns . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 232.3 Return predictability over different horizons . . . . . . . . . . . . . . . . . . . . . . . . . . 243.1 Performance of drift at different horizons . . . . . . . . . . . . . . . . . . . . . . . . . . . 554.1 Attention to Unemployment and the Unemployment Rate . . . . . . . . . . . . . . . . . . 884.2 Macro Attention and Macroeconomic Fundamentals . . . . . . . . . . . . . . . . . . . . . . 894.3 Macroeconomic Attention and Macroeconomic Fundamentals (cont.) . . . . . . . . . . . . 904.4 Autocorrelation in Macroeconomic Attention . . . . . . . . . . . . . . . . . . . . . . . . . 914.5 Macroeconomic Attention around Macroeconomic Announcements . . . . . . . . . . . . . 924.6 Attention to Unemployment around Employment Situation Announcements . . . . . . . . . 93A.1 Review Form . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120A.2 Salaries information . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121A.3 Histogram of reviews by star ratings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122A.4 The distribution of reviews over different months . . . . . . . . . . . . . . . . . . . . . . . 123C.1 Media Attention and Macroeconomic Fundamentals . . . . . . . . . . . . . . . . . . . . . . 132C.1 Media Attention and Macroeconomic Fundamentals (cont.) . . . . . . . . . . . . . . . . . . 133C.1 Media Attention and Macroeconomic Fundamentals (cont.) . . . . . . . . . . . . . . . . . . 134C.2 Macroeconomic Attention around Macroeconomic Announcements . . . . . . . . . . . . . 135C.3 Detrended VIX before Employment Situation and FOMC announcements . . . . . . . . . . 136xAcknowledgementsMy journey of six-year study in Vancouver at UBC has been wonderful and extremely rewarding. I havebeen lucky to meet many people who have taught, helped, and encouraged me throughout my doctoral study.I would like to use this opportunity to thank them.I am grateful to my main supervisor and co-author Adlai Fisher who has been always encouraging andhelpful. I learned so much by interacting and working with him. Most importantly, Adlai taught me thata good empiricist should have a theoretical mind. Murray Carlson is my co-supervisor and has guided meto become a curious and rigorous researcher. Kai Li read through all my papers and gave me detailed andextremely helpful suggestions. I also learned a lot from her about how to become an efficient researcher. Ithank Jan Bena for numerous extensive discussions about my papers since my first year. He taught me tothink very hard about the economics behind empirical findings. Carolin Pflueger motivated me to do goodresearch and her taste of research has affected on the type of research I am working on and I will pursue inthe future. Jesse Perla always saw the big picture of my papers and provided insightful comments.I would like also to acknowledge the support of my co-authors Jack Favilukis at UBC and Mike Simutinat the University of Toronto, to both of whom I am extremely thankful for their help for my papers and thejob market process. I also thank Werner Antweiler, Paul Beaudry, Ruth Freedman, Lorenzo Garlappi, RonGiammarino, Will Gornall, Hernan Ortiz-Molina, and Elena Simintzi for their generous help in my researchand teaching. I thank Canadian Securities Institute Research Foundation for the financial support for thisthesis.One of my best experiences at UBC is the interaction with my Ph.D. colleagues. Charles Martineau wasa great person to talk to and an excellent co-author. Terry Zhang and I disagree on many things, but we forman amazing team to work on exciting projects. Guangli Lu is my life-time friend since our undergraduatestudy at Nankai University, and his support is very important for me. I also thank Alex Corhay, KairongXiao, and Ting Xu for their help during my days at UBC.Finally, I would like to send my special thank you to my family. Thank you to my grandmother AdaWu who brought me up. She passed away in 2012 winter when I was in Vancouver. I owe her a farewell.I thank my parents, Jinde and Xuefeng, for their endless support and love despite me being far away fromhome. Thank you to my loving wife Jinzhi and my son Liangkai (Liam), who brought so much happinessto my life. Without the constant support of my family, I could not have completed my thesis.xiDedicationTo my family.xiiChapter 1IntroductionHow does information affect stock prices? The answer to this question is crucial for understanding thefunctioning of stock markets. In the information age, ever-increasing access to information has made theproblem of how investors process information more relevant. This thesis contains three essays on the rela-tionship between information and stock returns. The first essay studies the value of online employee reviewsto stock markets, using a novel dataset of nearly one million employee reviews. Employee beliefs about theiremployers’ business prospects predict future returns at one- to five-month horizons, delivering an annualizedabnormal return of 7% to 9%. Employee reviews are related to firms’ fundamentals because they predictcash flow news. In addition, the reviews predict future trading activity by hedge funds, suggesting somesophisticated investors exploit this information or its underlying sources. There are information hierarchieswithin firms in the sense that relatively high-level employees’ expectations are better in predicting futurestock returns. This essay highlights the importance of online information about firms’ fundamentals, whichis beyond traditional information sources such as analyst forecasts.In the second essay, I move to the interaction between different types of information. In particular,I investigate interactions between macro-announcements and the processing of earnings news. Existingtheories suggest that macro-news should crowd out attention to firm-level news, implying less efficientpricing. However, I find the opposite: on macro-news days price reactions to earnings news are 17% strongerand the post-earnings announcement drift is 71% weaker. To explain these results, I show that institutionalinvestor attention is higher on macro-news days. Hence, macro-news appears not to be a distraction fromfirm-level news, but instead serves to enhance overall attention to financial markets. I suggest extensions ofexisting theories that could be consistent with these findings.In the third essay, co-authored with Adlai Fisher and Charles Martineau, we construct indices of me-dia attention to macroeconomic risks including employment, growth, and monetary policy. Attention risesaround macroeconomic announcements and following changes in fundamentals over quarterly, annual, andbusiness cycle horizons. The effect is asymmetric, with bad news raising attention more than good news.Attention relates to the stock market in two ways. First, increases in aggregate trade volume and volatilitycoincide with rising attention, controlling for announcements. Second, changes in attention before unem-ployment and FOMC announcements predict announcement surprises, stock returns, and implied volatilitychanges on the announcement day. We conclude that media attention to macroeconomic fundamentals pro-vides market-relevant information beyond the contents and dates of macroeconomic announcements.Because each essay investigates a different topic, chapters were designed to be self-contained. I thusleave a more exhaustive discussion of methodology and contribution to literature to the introduction specificto each chapter.1Chapter 2Asset Pricing in the Information Age:Employee Expectations and Stock Returns2.1 IntroductionThe arrival of the information age has significantly changed the landscape of financial markets. Investors canaccess more information than ever. Not only has the amount of information that investors face increased, thecomposition of information producers has also changed dramatically. In the past, investors relied on profes-sionals, such as sell-side analysts, to acquire value-relevant information regarding the future performance offirms.1 The rise of the internet and social media has greatly facilitated the generation and dissemination ofinformation by non-professionals, including retail investors, consumers, and employees. Yet, little is knownabout the value of non-professional forecasts to the stock market. I address this question using a uniquedataset of nearly one million online employee reviews.Whether employee forecasts are valuable to the stock market is not clear ex ante. On the one hand, priorresearch finds that employees do not seem to have superior information about their firms’ future returns,in the context of investments in stocks of their own companies in 401(k) accounts (Benartzi, 2001; Cohen,2008). Also, employees are not experts in forecasting firms’ fundamentals and their information is limitedto their own occupations, and may be influenced by their own experiences. On the other hand, employees asinsiders may possess a wealth of information about their firms. Upper-level management often knows moreabout the performance of their firms than outsiders. Even non-executive employees can possess informationrelevant to a firm’s future prospects through their day-to-day jobs (Babenko and Sen, 2015). Informationfrom employees helps financial analysts make accurate forecasts (Malloy, 2005) and investment managersachieve good performance (Coval and Moskowitz, 1999).2 Thus, even if employees post only a subset oftheir information online, it may be valuable to the stock market. Moreover, despite potential bias and errorsin individual forecasts, the aggregation process may help to filter out noise, thereby resulting in high qualityinformation.To examine the role of non-expert forecasts, I extract employee reviews of companies from Glassdoor,1There were about 6000 sell-side analysts at the world’s 12 largest investment banks in 2016 (Financial Times, 2017, “Sell-sideresearch would be little missed”).2These studies find that local investment managers and analysts possess an information advantage, and this advantage translatesinto better performance. One major source of their information advantage is access to private information from top executives andemployees. For example, Coval and Moskowitz (2001) argue that “Investors located near a firm can visit the firm’s operations,talk to suppliers and employees, as well as assess the local market conditions in which the firm operates” (p. 839). Similarly,Malloy (2005) claims that “The ability of local analysts to make house calls rather than conference calls, during which time theycan meet CEOs face to face and survey the firm’s operations directly, provides them with an opportunity to obtain valuable privateinformation” (p. 721).2the largest job review website in the world. As of June 2017, Glassdoor had 45 million unique visitorsper month 3. Glassdoor is a website where current and former employees anonymously review companiesand their management. The reviews evaluate companies on certain criteria, including company reviews,compensation and benefits, and interview experiences. Although Glassdoor reviews are not directly relatedto investment opinions, they provide a variety of unique information.4 Most importantly, one component ofGlassdoor reviews focuses on the employee’s opinion about the firm’s business outlook (hereafter employeeoutlook).5 I measure employee forecasts using this outlook variable.Anecdotal evidence suggests that employee outlook contains value-relevant information to the stockmarket. For instance, Sears, the famous 130-year-old retailer, has been struggling since 2010. The fractionof reviews that had positive employee outlook dropped 3.3% in November 2013 compared to the previousthree months, with employee comments such as: “Not enough customers and not competitively pricedenough for sales,” “Online shopping from within store very frustrating - products weren’t available andcustomers were disappointed,” and “My store didn’t see a lot of customers (like many Sears stores thesedays).” The stock price of Sears dropped from $64 in November 2013 to $49 in December 2013 (i.e., -23%in returns), and never since recovered to the $64 level. The decline in positive outlook also coincided witha negative earnings surprise for the last quarter of 2013. Employees as a group can therefore hold valuableinformation about a firm’s fundamentals and stock prices.While anecdotal evidence is appealing, it is not necessarily the case that this pattern holds systemicallyacross all firms. Using review data from Glassdoor for S&P 1500 firms from 2012 to 2016, I provide acomprehensive study of employee outlook. Information embedded in employee outlook is novel and differ-ent from known information from financial statements and media coverage. I find evidence that employeeoutlook contains value-relevant information. Higher abnormal positive outlook predicts increased futurestock returns. A long-short portfolio trading strategy delivers an abnormal return of 0.61% to 0.78% permonth for a one-month holding period. The return predictability of employees’ information is robust tocontrolling for firm characteristics such as size, market beta, book-to-market, media coverage, profitability,and employee satisfaction (Edmans, 2011). The return predictability of employee outlook is strongest overone to four weeks, and declines gradually over five months. The abnormal returns do not reverse during a12-month holding period. This alpha decay pattern holds for weekly and daily results as well, suggestingthat this information is incorporated into stock prices over time.To understand the information contained in employee outlook, I investigate the relation between em-ployee outlook and firms’ future earnings. I find that employee outlook positively predicts subsequentearnings surprises and changes in profitability. The results are robust to controlling for other variables thataffect earnings, such as analyst coverage and past stock returns. The economic magnitude is significant.For instance, one standard deviation change of abnormal positive outlook is associated with an increase3Glassdoor, 2017, Fact and site statistics, https://www.glassdoor.com/press/facts/4Job seekers, investors, and corporations use information from Glassdoor for important decisions. For instance, a news articlein the Wall Street Journal reports that Zillow’s CEO Mr.Rascoff and leaders at other companies look at information from Glassdoorof acquisition candidates. Mr Rascoff says “we walked away from dozens of acquisitions that looked good on paper and madestrategic sense.”5The specific question Glassdoor asks is: “In the next six months, do you believe your company’s business will perform better,worse or remain the same?”3of 29% of the mean of earnings surprises. I further examine the extent to which the predictive power ofemployee outlook for future stock returns can be explained by future earnings. I find that the return pre-dictability declines 48% after controlling for future earnings surprises, suggesting that future earnings canaccount for about half of the return predictability. These findings suggest that employee outlook containsnovel information about firms’ fundamentals.Because employee outlook contains value-relevant information about stocks, an interesting question iswhether sophisticated investors exploit this information from Glassdoor. I examine this question by lookingat the trading activities of various types of institutional investors. Anecdotal evidence suggests that hedgefunds exploit Glassdoor data and, indeed, I find that abnormal positive outlook predicts the net purchasesby hedge funds. In contrast, abnormal positive outlook does not predict net purchases by mutual funds orother non-hedge-fund institutions. My results show that abnormal positive outlook also predicts a decreasein short selling quantity and costs. Although I cannot distinguish whether those sophisticated investors aretrading directly on the employee reviews, or similar information through other channels, my analysis showsthat employees in aggregate hold useful investment information, and acquire this information before thesophisticated investors trade on that information.Several recent studies point out the importance of looking at the heterogeneity in analyst forecasts (Chi-ang et al., 2016; Michaely et al., 2017). Motivated by these studies, I examine whether significant qualitydifferences exist in online forecasts by different levels of employees. Are high-level employees’ outlooksbetter than low-level employees’ outlooks? The answer to this question is not clear ex ante. While sometheories of organization hierarchies predict that high-level employees have better information (Garicano,2000), other theoretical evidence stresses the benefits of having a different structure where low-level em-ployees possess important information (Landier et al., 2009). To examine the heterogeneity in employeeoutlook, I classify employee job titles into three groups (high-, middle-, and low-level) within a firm basedon wages, and examine the return predictability of outlook by different levels of employees. I find that,on average, the return predictability of employee outlook increases with employee rank, suggesting thepresence of information hierarchies within firms.Interestingly, a firm’s information hierarchy depends on its complexity. I use three proxies to measurea firm’s complexity: organization hierarchy, size, and whether the firm is a conglomerate. For complexfirms, high-level employees’ outlooks tend to have better return predictability, and middle- and low-levelemployees’ outlooks do not predict future stock returns. However, for simple firms, middle- and low-levelemployees’ outlooks also predict future returns. The return predictability among high-, middle-, and low-level employees’ outlooks is statistically indistinguishable. These findings are consistent with theoreticalevidence and economic intuition that the information structure is flatter among simple firms because thecommunication cost among different levels of employees is smaller.Why are high-level employees’ outlooks better in predicting future returns? I address this question byexamining whether high-level employees’ reviews convey different information compared to those fromlow-level employees. Using machine learning to detect latent topics in review texts, I compare the topicsfrom reviews by different levels of employees. High-level employees’ reviews are often about businessgrowth, which is more related to a firm’s fundamentals. Low-level employees’ reviews focus on work4hours and personnel development, which are less related to a firm’s fundamentals. Although employees’opinions on work hours and personnel development may affect a firm’s valuation in the long run, they areslow-moving variables and thus less important in predicting short-run earnings and stock returns.This paper makes four main contributions to the finance and economic literature. First, this study isthe first to test the investment value of online non-professional forecasts of firms’ fundamentals. Whileseveral studies examine the impact of social media on stock returns (e.g., Tumarkin and Whitelaw, 2001;Antweiler and Frank, 2004; Chen et al., 2014),6 their results are based on investment opinions and stressthe importance of peer-effect among investors.7 Different from these studies, this paper focuses on onlineforecasts that are directly about firms’ fundamentals. This unique feature allows me to distinguish between afundamental story (where employee outlook contains relevant information about firms’ fundamentals) and abehavioral story (where investors just react to employee outlook for behavioral reasons and the outlook doesnot provide valuable information about firms’ fundamentals). Also, previous studies often use a relativelysmall sample of firms or have a small number of reviews for each firm. My study utilizes a comprehensivesample of S&P 1500 firms where an average firm has about 700 reviews. Because the power of onlinedata arises from the aggregation of a large amount of data, my large sample allows for reliable statisticalinferences. Thus, this paper not only extends earlier results to another category of social media, but alsosheds light on the underlying reasons why social media matters for stock markets.Second, by studying non-professional forecasts, this paper highlights a new type of information interme-diary that is different from traditional types such as sell-side analysts (e.g., Womack, 1996; Jegadeesh et al.,2004). Although financial analysts are experts in producing earnings forecasts, employees’ forecasts maycontain novel information for several reasons. First, employee outlook is based on first-hand informationfrom day-to-day jobs. This type of information, called “serendipitous information” by Subrahmanyam andTitman (1999), is diffuse but genuine. Second, social media is unique in the sense that it allows a largenumber of people to produce information at a very low cost. Although individual employee’s forecasts maynot be precise about a firm’s future prospects, employee forecasts in aggregate may contain informationthat is not fully incorporated into earnings forecasts. Third, prior studies find that financial analysts tend tobehave strategically and do not always reveal their true opinions due to conflicts of interest (e.g., Michaelyand Womack, 1999; Kadan et al., 2009). In contrast, employees are more likely to provide genuine opinions,because their forecasts are anonymous.8 Moreover, an individual employee has little incentive to intention-ally spread false information and mislead investors because a single forecast is not likely to exert significantinfluence on stock prices. This study contributes to this literature by showing that non-professional forecastscan provide value-relevant information that is beyond analyst forecasts, which is a novel phenomenon in theinformation age. Several papers use employee review data, but they have a different focus on corporateculture and employee satisfaction (Grennan, 2013; Green et al., 2018).Third, my findings shed new light on the debate over whether rank-and-file employees possess value-6In general, this paper fits into the literature on the impact of media on stock returns (e.g., Tetlock, 2007; Fang and Peress, 2009;Dougal et al., 2012; Huang, 2018; Fisher et al., 2017).7Campbell et al. (2017) show that 36% of investment platform users are financial professionals.8This does not mean that online anonymous reviews are unbiased. Online reviews often suffer from response bias. I findresponse bias is not severe in Glassdoor data due to their policy to encourage more balanced reviews (see Section 1.2).5relevant information. Prior studies report contradicting findings based on indirect evidence from employees’investment decisions (Benartzi, 2001; Cohen, 2008; Babenko and Sen, 2015). While inferring employees’information from their stock purchases is useful, it is not direct. Even if employees’ stock purchases do notpredict firms’ future returns, this does not necessarily mean that employees do not have valuable informationabout their firms. Employees might not exploit their information due to a lack of financial literacy orsignificant fixed costs of participation. Using a direct measure of employee information from their onlinereviews, I show that rank-and-file employees do possess valuable information among firms with simplestructure.Fourth, this paper shows the first direct empirical evidence of the existence and magnitude of informa-tion hierarchies within firms. While various theories of organization hierarchies study information structurewithin firms (Garicano, 2000), direct empirical evidence is limited. Empirical testing of information hier-archies is challenging because measuring employees’ information is extremely hard, and it is difficult toevaluate the value of information. I use employees’ online reviews as a proxy for their information and usethe stock market as a laboratory to evaluate the value of information. If one group of employees’ informationcan better predict future returns, then their information is better. This study complements the literature onorganization hierarchies by explicitly examining the information structure within firms.2.2 Data2.2.1 Employee review dataI retrieve employee reviews from Glassdoor, the largest website for employee reviews. Launched in June2008, Glassdoor provides information about companies that is posted by current and former employees,including company reviews, compensation and benefits, and interview experiences. Since then, hundreds ofthousands of users have posted over 33 million reviews and insights for approximately 700,000 companiesaround the world. The website is widely used and had 45 million global users visit in July 2017.Each company review contains numerous components: an overall rating as well as ratings on compensa-tion and benefits, work-life balance, company culture and values, and senior management, all of which aremeasured on a five-point scale. Each company review also reveals whether the reviewer would recommendthe company to a friend, and whether they approve of the CEO. Most importantly, it contains an assessmentof a firm’s 6-month business outlook (see Figure A.1 in the Appendix for the full questionnaire). Otherimportant information in a review includes reviewer’s job title and location. Reviewer’s job title is crucialto analyze information distribution within the firm. For each job title, I collect its average salary from thesame website (see Figure A.2 in the Appendix for an example). The outlook information only became avail-able since March 2012, and thus my main analysis focuses on the sample period of 2012-2016. I also usemachine learning to backfill the outlook variable to 2008 and the results are robust.A common concern about online review data is response bias. Glassdoor uses a “Give to get” policy asan incentive to encourage more neutral and balanced company ratings. Under this policy, users must sharetheir opinions of their own employer to access information on Glassdoor. In a report released in October2017, Glassdoor shows that the polarization bias on its website is less severe than other online review sites,6such as Amazon and Yelp.9 Nevertheless, I check whether there is serious bias in the data. Luca and Zervas(2016) find that suspect reviews tend to have a bimodal distribution of either very low or very high scores. Assuch, bimodally distributed employee reviews would indicate a serious response bias in the sample. FigureA.3 in the Appendix shows that star ratings are approximately normally distributed rather than bimodallydistributed. Although this evidence does not fully eliminate the possibility of response bias, it makes bias aless serious concern. In a robustness test, I also remove firms with less than 5 reviews per month during thesample period because these firms are more likely to be affected by extreme responses.2.2.2 Summary statisticsIn this paper, I focus on S&P 1500 firms and collect firm returns and volume data from CRSP, accountingvariables from Compustat, and analyst forecasts from I/B/E/S. After matching CRSP and Compustat toemployee reviews data, the final sample consists of 1422 firms with about one million reviews. I alsoremove stocks that have less than 20 reviews in total during the sample period.10 Table 2.1 presents thedistribution of reviews and firms over Fama-French 12 industries. The top three industries in terms of thenumber of employee reviews are wholesale, business equipment, and finance. In a robustness test, I excludefinancial firms and find similar results. On average, a firm has about 700 reviews from 2012 to 2016.To construct the monthly panel sample, I aggregate daily reviews to monthly frequency. For outlook,I calculate the fraction of reviews that have positive outlook (Positive outlook). Similarly, I calculate thefraction of reviews that state “recommend to a friend” (Recommend). For the star rating variables, I calculatethe monthly average of the overall rating (Overall), culture and values rating (Culture), work-life balancerating (WorkLife), senior management rating (Management), compensation and benefits (Compensation),and career opportunities (Career). To measure the new information conveyed by employee outlook (i.e.,surprises in employee outlook), I use average positive outlook over the prior three months as a benchmarkfor employees’ expectations of a firm’s future prospects. I thus measure the abnormal positive outlook(AbnOutlook) as the difference between the Positive outlook and its mean over the prior three months. Byusing this measure, I difference out time-invariant biases in employee outlook. In a robustness test, I use themeans over the prior six months as benchmarks to calculate AbnOutlook, and get similar results.Table 2.2 reports the summary statistics on these review variables. The mean value of Positive outlooksuggests that, on average, 41.7% of employees who post a review have a positive opinion of their firm’snear-term business prospects. The mean AbnOutlook is 0.35%, which is not statistically different fromzero. On average, about 56.01% of employees said that they would recommend their company to a friend.The average Overall rating is 3.18 (out of 5), indicating that the average employee posting a review hasa generally positive view about the firm. The five subcomponents have similar means, with Managementhaving a slightly lower value. Table 2.2 also reports the descriptive statistics of firm characteristics. Thesummary statistics of firm characteristics are comparable to the literature for the S&P 1500 sample.9See the Appendix for further discussion of “Give to get” policy and the Glassdoor report on this policy.10This is to address the concern that the results might be driven by firms with very few reviews, which are biased. These firmstend to be small firms, for which the return predictability is stronger. Thus, removing these firms is conservative and results in alower bound. The results are robust when including these firms.7Does employee outlook convey novel information or just reflect stale news from existing financial state-ments and media reports? To examine this question, I regress employee outlook on lagged positive mediacoverage and firm characteristics including size, book-to-market, market beta, institutional ownership, an-alyst coverage, profitability, and trading volume.11 The results, presented in Table 2.3 , show that none ofthese variables can predict abnormal positive outlook. This finding suggests that abnormal positive outlookis largely independent of the information contained in accounting statements, analyst coverage, and mediacoverage.2.2.3 Employee outlookI examine the nature of employee outlook by looking at other components of an employee review: starratings, employee characteristics, and texts. Specifically, I run the following regression:Outlookit = γ0+ γ1Xit + εit (2.1)where Outlookit is a dummy variable that is equal to one if outlook is positive, and zero otherwise; andXit includes three sets of variables: (i) employee characteristics: current vs. former, work in headquarterstates vs. non-headquarter states; (ii) star ratings as discussed in Section 2.2; and (iii) text characteristics:number of words in title, pros, cons, and advice to management sections, and number of words of the wholereview.12First, I examine what type of employees are more likely to have positive outlook. Table A.2 Panel A inthe Appendix presents the results. Current employees are more likely to have positive outlook than formerones (Column (1)), and employees who work in the headquarter state do not have a more optimistic outlookthan others (Column (2)). Second, I also look at the recommendation variable and the star ratings and findthat employees who state “recommend to a friend” or give high star ratings are more likely to have positiveoutlook (Columns (3)-(4)).Finally, I investigate the differences in review texts among reviews with positive and negative outlookalong two dimensions: length and meaning of text. A review text has four parts: review title, pros, cons,and advice to management. I count the number of words in each part. If business outlook indeed reflectsemployees’ opinions of the firm’s future prospects, it might be related to the length of discussion in pros,cons, and advice parts of the review. Table A.2 Panel B in the Appendix shows that employees who havea positive outlook tend to write a shorter title, say less in cons and advice to management sections, and saymore in pros section (Column (1)). In general, employees with a positive outlook tend to write a shorterreview than those who do not (Column (2)).I examine the relationship between outlook and meaning of review text by finding the top words inreviews. If employee outlook is indeed related to the business operation of the firm, employees who havepositive outlook may also tend to say something good about the firm in the text. I use the bag-of-wordsmethod to convert raw reviews into words, and remove stop words (i.e., meaningless words such as the,11RavenPack provides a sentiment score (out of 100) for each news article. A news article for a firm is positive if the score isabove 50.12I also use logit and probit regression methods to conduct this test and get similar results.8an, and). I then convert the words into numeric vectors. To account for the relative importance of a wordin each review, a term-frequency-inverse document frequency (tf-idf) method is used in the transformationprocess. After transformation, each review is represented as distributions of words with their weights.Finally, I run a logistic regression where the dependent variable is a dummy variable that is equal to one ifemployee outlook is positive, and zero otherwise, and the independent variables are distributions of words.The regression coefficients give the relative importance of each word.Figure 2.1 reports the results of such a test and shows the top 10 words associated with both positive andnegative outlook. The blue bars are words that are the most important for a review with a positive outlook.These words include great, awesome, love, growing, amazing, excellent, continue, fantastic, superb, andincredible, all of which are positive words. The red bars are words that are the most important for a reviewwith a negative outlook. These words include horrible, downhill, sinking, worst, unstable, uncertain, layoffs,declining, poor, and terrible, all of which are negative words. Overall, the list of top words from textualanalysis seems to be consistent with the opinion of outlook.2.3 Employee outlook: is it valuable to the stock market?There is no systemic evidence of the existence and magnitude of the value of employee outlook to the stockmarket. I examine this question by looking at whether employees’ positive outlook can predict a significantincrease in future stock returns, controlling other factors affecting stock returns. I use both portfolio sortingand panel regression approaches, followed by a battery of robustness tests. To understand the underlyinginformation sources of employee outlook, I test whether it is linked to future fundamental information.2.3.1 Employee outlook and stock return predictabilityI use a portfolio approach to examine the investment value of employee opinions about the prospects of firms.In each month from June 2012 through December 2016, I sort sample stocks into tercile portfolios basedon AbnOutlook. That is, stocks with a high (low) AbnOutlook are assigned into the top (bottom) portfolio,and the rest are assigned into the middle portfolio. I then track the performance of the three portfolios overthe following month. I employ two weighting methods across firms, equal weighting and value weighting.I also use a weighting method based on the number of reviews in each month in a robustness test.Figure 2.2 graphically presents the stock performance of high- and low-AbnOutlook firms. It showsthe cumulative return of the high-AbnOutlook and low-AbnOutlook portfolios that are formed at the endof June 2012, rebalanced at the end of each month, and held to December 2016. Over my sample period,the equal-weighted (value-weighted) high-AbnOutlook portfolio has cumulative returns that are 73% (45%)higher than the low-AbnOutlook portfolio.To formally test the investment value of employee outlook, I first examine whether it can predict excessreturns adjusted by risk-free rate. Table 2.4 Panel A shows that portfolio returns monotonically increase withpositive outlook value. For stocks with low abnormal positive outlook (Portfolio 1), their average excessreturn is 0.71% (equal weighting). On the other hand, the average excess return of stocks with high positiveoutlook (Portfolio 3) is 1.57%. The difference (0.86%) between the two portfolios is statistically significant9at the 1% level. The pattern is similar for value-weighted portfolios. This evidence suggests that firms withmore positive outlook tend to have higher future stock returns after adjusting the risk-free rate.While the evidence from excess returns is significant, it is possible that stocks in Portfolio 3 are differentfrom Portfolio 1 in systematic ways. For example, stocks in Portfolio 3 may have more exposure to riskfactors than stocks in Portfolio 1. To ensure the outperformance based on abnormal positive outlook doesnot result from risk, I use the Fama-French-Carhart four-factor model (Fama and French, 1993; Carhart,1997) to adjust returns. I compute a four-factor alpha by regressing monthly portfolio excess returns on themonthly returns from the risk factors:Rit = α+βMKT MKT RFt +βSMBSMBt +βHMLHMLt +βMOMMOMt + εit (2.2)where Rit is the excess return adjusted by the risk-free rate from portfolio i at time t. For the long-shortportfolio, Rit is the return difference between portfolio 1 and portfolio 3. MKT RFt , SMBt , HMLt , andMOMt are market, size, value, and momentum risk factor returns, respectively.13Table 2.4 Panel B reports the alphas and factor loadings from this regression for both equal-weighted andvalue-weighted portfolios. Stocks with high AbnOutlook outperform the four-factor benchmark by 0.51%to 0.56% per month. On the other hand, stocks with low AbnOutlook underperform by 0.05% to 0.26%per month. A long-short portfolio that buys stocks in the top tercile of positive outlook and sells stocks inthe bottom tercile outperforms the benchmark by 0.61% to 0.78% per month, or about 7% to 9% annually,which is significant at the 1% level. While the magnitude of alpha is sizable, it is not unusually large forinvestment strategies based on information that is not from traditional sources. It is comparable to abnormalreturns of long-short strategies based on managerial ownership (4%-10% in Lilienfeld-Toal and Ruenzi,2014), employee stock purchases (10% in Babenko and Sen, 2015), and firm complexity (11.8% in Cohenand Lou, 2012).To understand the sources of the predictability of future stock returns based on employee outlook, Iconduct various subsample analyses. First, prior literature has established that the information environmentof firms can affect the processing of information. Firms with a more opaque information environment willhave stronger return predictability because their information gets less attention, resulting in slower incorpo-ration of news. I use firm size and analyst coverage as proxies for the information environment. Informationabout small firms and firms with lower analyst coverage is more likely to be ignored by investors. If a worseinformation environment slows the incorporation of the information embedded in employee outlook, thepredictability should be concentrated among small firms and firms with low analyst coverage. I partition thesample into two subsamples based on their market capitalization.14 Small firms are firms with a market capbelow the median, and large firms are those with a market cap above the median. Similarly, I partition thesample based on the median of analyst coverage. Table 2.5 Panels A and B show that the return predictabilityis stronger among small firms, and comes mainly from firms with fewer analysts.Second, because the information is from employees, the relative importance of employees within firmsmay also affect the return predictability. Prior literature finds that the level of human capital or organization13These risk factor returns are downloaded from Kenneth French’s website: http://mba.tuck.dartmouth.edu/pages/faculty/ken.french/data_library.html.14I use the market capitalization from previous year to avoid forward-looking bias.10capital in a firm affects its stock returns (Eisfeldt and Papanikolaou, 2013; Boguth et al., 2016b). Moreover,for firms with more employees, the information quality from Glassdoor is better because these firms tendto have more reviews. I use labor intensity at the industry level to proxy for the importance of employeesin a firm. Employees’ outlook for firms with high labor intensity may have greater impact on their stockreturns, and the return predictability will be stronger for these firms. I partition firms into two groups basedon their labor intensity at the industry level. Firms in industry with labor intensity above the median are highlabor intensity firms, while the rest are low labor intensity firms. Table 2.5 Panel C shows that the returnpredictability of employee outlook is much greater among firms in higher labor intensity industries.Third, the information quality of employee outlook also depends on employees’ characteristics. Cur-rent employees and employees who work near the headquarters are more likely to have better informationabout their firms. Former employees do not work in the company anymore and therefore their informationmay be stale. Employees who work far away from headquarters may only have limited information abouttheir division. Thus, the outlook of current employees and employees who work near headquarters likelyhas stronger return predictability. To test this conjecture, I partition the sample based on employee statusand employee location. Table 2.5 Panel D shows that the return predictability is mainly driven by currentemployees. Table 2.5 Panel E shows that, although the return predictability of employees who work in theheadquarter state is greater than that of other employees, information from employees in a non-headquarterstate is still valuable in predicting future performance.2.3.2 Fama-MacBeth regressionsThe results so far show that employee outlook contains value-relevant information. Is this information new,or does it simply capture return predictability from contemporaneous accounting variables? To address thisconcern, I conduct Fama-MacBeth regressions by explicitly controlling for accounting variables and otherknown predictors in the cross-section of stock returns. Fama-MacBeth regressions are performed in twosteps (Fama and MacBeth, 1973). For the first step, in each month, I run the following cross-sectionalregression:R j,t+1 = η0+η1AbnOutlook j,t +η2BC j,t +θZ j,t + ε j,t (2.3)where Ri,t+1 is the excess return adjusted by the risk-free rate of stock j in month t + 1; AbnOutlook isthe abnormal positive outlook of stock j in month t; and BC is a dummy variable that is equal to one ifthe firm is on Fortune’s list of the “100 Best Companies to Work for” in that year, and zero otherwise.15Z j,t is a vector of firm characteristics for firm j in month t. I include the following firm characteristics ascontrols: exposure to market risk (Beta), market capitalization (Size), cumulative stock returns in the prior12 months, institutional ownership, analyst coverage, gross profitability of Novy-Marx (2013), and dollartrading volume. Media coverage, which is the number of news articles that mentioned the firm in the DowJones News Archives (from RavenPack), is also included as a control variable.15The list of the “100 Best Companies to Work for” from 1984 to 2012 can be downloaded from Alex Edmans’s website:http://alexedmans.com/data/. I obtain the list of companies from 2013 to 2016 from Fortune.11For the second step, I estimate the time-series averages of the cross-sectional regression coefficients.Table 2.6 reports the results of the second step. Column (1) shows the results of a univariate regression.The coefficient on employee outlook is positive and significant at the 1% level, suggesting that employeeoutlook can predict future stock returns. Edmans (2011) finds that firms on the “100 Best Companies toWork for” list tend to have higher future long-term returns. Column (2) reports the results of a univariateregression of BC on future returns. The coefficient is negative and insignificant, suggesting that employeesatisfaction cannot predict short-run returns. In Column (3), I regress both AbnOutlook and BC on futurereturns and find that the coefficient on employee outlook remains similar at both statistical and economiclevels. Column (4) reports the results of multivariate regressions with a battery of additional controls andshows that change in employee outlook continues to predict future stock return after adding these controls.The economic magnitude is significant. For instance, the coefficient estimate on AbnOutlook in Column (4)indicates that one standard deviation change of AbnOutlook is associated with an increase of about 0.84%in the one-month-ahead excess returns. Overall, these results suggest that the information from employeeoutlook is likely new information, which is different from other sources such as accounting statements, pastreturns, media coverage, and employee satisfaction.2.3.3 Employee outlook and earningsThe fact that employee outlook predicts future stock returns suggests that it may contain new informationabout a firm’s earnings. If this is the case, then employee outlook may predict a firm’s earnings news. Totest this conjecture, I use earnings surprises to capture new information in earnings and examine whetheremployee outlook can predict earnings surprises. Because earnings are released at a quarterly frequency,I calculate the employee outlook at a quarterly frequency as the fraction of reviews with positive outlookduring a quarter. AbnOutlook is then defined as the quarterly positive outlook minus its mean over the priorthree quarters. Following the literature, I run the following panel regression:SUEi,q = γ0+ γ1AbnOutlooki,q+ γ2SUEi,q−1+θZi,q−1+ εi,q (2.4)where SUE is earnings surprises for firm i in quarter q (as defined in Section 2); AbnOutlooki,q is theabnormal positive outlook for firm i in quarter q; and Zi,q−1 is a vector of firm characteristics of firm i inquarter q− 1, including lagged earnings surprises, size, market beta, book-to-market, profitability, stockreturns in the past 12 months, media coverage, institutional ownership, analyst coverage, and trade volume.I also add time and firm fixed effects in some specifications. Note that abnormal positive outlook is measuredbefore the release of quarterly earnings news, which typically occurs about 30 days after a quarter-end.Table 2.7 reports the regression results. Abnormal employee outlook positively predicts earnings sur-prises regardless of the specifications. The economic magnitude is significant. For instance, in Column (3)the coefficient on AbnOutlook indicates that one standard deviation change of AbnOutlook is associated withan increase of 0.02 % in SUE, which is about 29% of the mean of SUE.I also use cumulative abnormal returns (CAR) around the earnings announcement window to proxyfor earnings surprises. Specifically, I use the CAR[-1,1] where returns are measured by using a three-day12window centered on the announcement date and adjusted using the market model. I use the same regressionas in Equation (3.4), except replace the SUE with CAR[-1,1]. The regression results with CAR as the measureof earnings surprises are reported in Columns (4)-(6) of 2.7 . The coefficient on AbnOutlook is positive andsignificant regardless of specifications. The economic magnitude is also significant. One standard deviationchange of AbnOutlook is associated with an increase of 0.44% in CAR.Moreover, I investigate whether employee outlook predicts firms’ future performance. I use two mea-sures of performance: change in return on asset (∆ROA) and change in operating profitability (∆Profitability).I regress these two measures on lagged AbnOutlook with some controls and time and firm fixed effects. Ta-ble 2.8 reports the results. The coefficient on AbnOutlook is positive and significant for all specifications,suggesting employee outlook positively predicts a firm’s future performance for both measures.To further examine the extent to which the predictive power of employee outlook for future stock returnscan be explained by future earnings, I use Fama-MacBeth regression as in Equation (2.3) and add subsequentearnings surprises as an additional control variable. I find that the coefficient on AbnOutlook declines from0.93 to 0.48, suggesting the return predictability declines 48% after controlling for future earnings surprises.This finding indicates that firms’ future earnings can account for about half of the return predictability fromemployee outlook.Overall, the results regarding employee outlook and subsequent earnings surprises and firm performancesuggest information embedded in employee outlook contains novel information about firms’ fundamentals.Information regarding earnings, ROA, and profitability is released at a quarterly frequency, while informa-tion conveyed from employee outlook is analyzed at a monthly frequency. Employees who work in a firmwith strong earnings, ROA, and profitability may know that their firm is doing well and express their con-fidence about the firm’s prospect on Glassdoor, which predicts the subsequent release of earnings, ROA,and profitability. It also indicates that the average analyst does not fully incorporate the information fromemployee reviews in their forecasts.2.3.4 RobustnessI conduct a battery of tests to assess the robustness of the main results. First, these review components arenot independent of each other. As discussed in Section 2, employees who have a positive outlook of thecompany tend to give a high rating for Culture, WorkLife, Management, Compensation, and Career as well.This suggests that a set of latent variables may explain reviewers’ responses across various components ofthe reviews. My main focus is whether employee reviews contain information about firms’ future perfor-mance, which may be most directly captured by the outlook variable. But, it is possible that other variablesalso contain valuable information about the near-term prospects of the firm. To further understand this, Iconduct a Principal Component Analysis (PCA) and extract the top three components out of eight variables.Table A.3 Panel B in the Appendix shows that the first and most important component is mainly loadedby outlook variable. The other two components are mainly about work-life balance and compensation andbenefits. I construct an expected positive outlook variable, which is a linear combination of employee out-look, recommendation dummy, and star ratings based on the weights from PCA. The expected AbnOutlookis calculated as expected positive outlook minus its mean over the prior three months. Table 2.9 Panel A13shows that a long-short portfolio that buys stocks in the top tercile of expected AbnOutlook and sells stocksin the bottom tercile outperforms the Fama-French-Carhart four-factor benchmark by 0.30% to 0.63% permonth. Notice that the magnitude of alpha based on expected AbnOutlook is smaller than alpha based onactual AbnOutlook, suggesting the outlook variable contains more value-relevant information.Second, I use several alternative risk benchmarks to adjust the returns, including the Fama and French(1993) three-factor model, the Fama and French (2015) five-factor model, the Fama-French-Carhart six-factor model, the Hou et al. (2015) four-factor model, and the Fama-French-Carhart four-factor model aug-mented with a liquidity factor (Pástor and Stambaugh, 2003).16 Table 2.9 Panel A shows that the resultsare similar to those obtained from the main specification in terms of statistical significance and economicmagnitude.Third, I construct an abnormal positive outlook using the mean of positive outlook over the prior sixmonths, rather than three months, as a benchmark. The AbnOutlook is measured by positive outlook minusits mean over the prior six months. The results, reported in Table 2.9 Panel A, show that the four-factoralphas on the long-short portfolio continue to be positive and significant at the 1% level.Fourth, one may also be concerned that the sample period of 2012-2016 is too short to draw any reliableconclusion. In fact, it is normal to use a relatively short sample period due to the availability of online data.17Nevertheless, I use rating variables and text to backfill outlook values from June 2008 to February 2012 withmachine learning methods. Specifically, I use ratings variables and text of all reviews from March 2012 toDecember 2016 as training and test samples to find the relationship between these variables and outlook, andthen predict outlook based on the training samples. I use various machine learning methods: KNN, logisticregression, linear SVC, decision tree, random forest, gradient boosted regression trees, and deep learning.18Table A.4 in the Appendix presents the out-of-sample accuracy of these methods. While KNN and othermachine learning methods provide decent accuracy, deep learning outperforms all of them with an accuracyscore of 91% (see the Appendix for more details). Thus, I use the predicted outlook from a deep learningmethod in the end. With a reliable predicted outlook back to 2008, I then sort portfolios based on predictedAbnOutlook and calculate the Fama-French-Carhart four-factor alphas. Table 2.9 Panel B shows that firmswith high AbnOutlook outperform firms with low AbnOutlook by 0.39% to 0.53% per month.I also use several alternative samples to assess the robustness of the return predictability. While thesample of S&P 1500 firms is already large and accounts for more than 90% of the market capitalization ofthe US stock market, I also extend the sample to all Compustat-CRSP firms. I then use the final matchedsample of 2108 firms with employee reviews to form long-short portfolios and calculate the Fama-French-Carhart four-factor alphas. To address concerns on financial firms, I remove them from the sample andconstruct the portfolios with non-financial firms. To address the concern that the information is noisy andthe quality of data is not high for firms with very few reviews, I remove stocks with less than 5 reviews permonth and redo the tests as in the main specification. For a similar reason, I remove reviews from outsideof the US, and then conduct portfolio regression tests. The results for these alternative samples, reported16The liquidity factor returns data are from Lubos Pastor’s website: http://faculty.chicagobooth.edu/lubos.pastor/research.17For instance, Antweiler and Frank (2004) use Yahoo! message board data in 2000 to examine its influence in financial markets.Da et al. (2011) use Google trend data from 2004-2008 to measure investor attention and study their impact on stock returns.18See Friedman et al. (2001) for a detailed discussion on machine learning. See LeCun et al. (2015) for a review on deep learning.14in Table 2.9 Panel B, are similar to the main results in terms of both statistical significance and economicmagnitude.Finally, I use an alternative weighting scheme, weighting by the number of reviews. This weightingmethod lessens the possibility that the portfolios are dominated by firms with a relatively small number ofreviews that may contain more noise. Table 2.9 Panel C reports the results of review-weighted portfolioreturns using the same regression specification as in Equation (2.2). The results are similar to those ofequal-weighted portfolio returns in the main specification.2.4 Information processing: alpha decay and trading2.4.1 Return predictability over different holding horizonsTo understand the processing of information in employee outlook, I examine the performance of long-shortportfolios over one to 12 months after portfolio formation. Again, a long-short portfolio buys stocks in thetop tercile of abnormal positive outlook and sells stocks in the bottom tercile.Figure 2.3 Panel A reports the results of portfolios constructed over different holding periods. Thereis clearly a decay pattern of alpha over the holding horizon. Although the four-factor alpha of the long-short portfolios with a two-month horizon is statistically significant, the economic magnitude (0.20%) isjust one-third of the alpha of portfolios with a one-month horizon. The alphas are still significant up toa five-month holding horizon.19 These results indicate that the return predictability of employee outlookworks in the short run, which is consistent with evidence that managers are unable to forecast returns past100 days (Jenter et al., 2011). Together with results from Fama-MacBeth regression in Section 1.3, thesefinding suggest the underlying mechanism of the return predictability of employee outlook is likely differentfrom that of employee satisfaction, which predicts long-run returns (Edmans, 2011).To further understand the information processing and address the concern that most of the monthly alphamay come from the first week after portfolio formation, I conduct a similar exercise at a weekly frequency.I sort stocks into tercile portfolios based on AbnOutlook in each month. I then track the performance ofthe three portfolios over one to four weeks after portfolio formation. The results, reported in Figure 2.3Panel B, show a decay pattern in alphas over different holding weeks that is consistent with the notionthat information is incorporated into stock prices gradually and the return predictability of this informationweakens over time. But, the abnormal return (0.14%) of the long-short portfolios for the fourth week afterportfolio formation is still statistically and economically significant (t-statistic is 3.24).Moreover, I track how stock prices react to this information at a daily frequency, using a 30-day rollingwindow to aggregate employee outlook. Specifically, for each firm I calculate the fraction of positive outlookover the past 30 days, and determine abnormal positive outlook by comparing to its mean over the past 120days. Daily return on portfolio is noisy due to market microstructure issues. Instead, I use a regressionapproach where the dependent variable is abnormal return adjusted by the market model and the independentvariables are daily abnormal positive outlook based on a rolling window and controls (size, turnover, trading19Di Mascio et al. (2017)use institutional investors’ trading to infer their private information and document an alpha decay patternover 12 months.15volume, and past month returns). Figure 2.3 Panel C reports the coefficient on abnormal positive outlookfor up to 20 trading days. Again, there is a decay pattern in the impact of employee outlook on stock returnsand the impact is still significant on day 20.2.4.2 Institutional tradingHaving demonstrated the return predictability of employee outlook, one might wonder whether investorsexploit this information. Also, the alpha decay pattern indicates that the information from employee outlookis incorporated into stock prices over time, which suggests that investors may trade on this information. Iinvestigate this question by looking at trading activities by institutional investors.Anecdotal evidence suggests that some hedge funds trade on information from Glassdoor20, and I per-form a systematic test. Using a list of 1274 hedge funds from Agarwal et al. (2013), I collect their holdingsfrom Thomson Reuters’ 13f filings data. For each stock, I calculate net purchases by hedge funds as thedifference in hedge-fund ownership between this quarter and the last quarter. If hedge funds exploit thisinformation, they should trade in the same direction as the employee outlook signal. To test this conjecture,I regress net purchases on the lagged abnormal positive outlook and controls including stock returns, firmcharacteristics, and media coverage.Table 2.10 presents the results of this regression. The coefficient on abnormal positive outlook is positiveand significant, suggesting that hedge fund managers are trading in the same direction as abnormal positiveoutlook. In terms of economic magnitude, one standard deviation change of abnormal positive outlook isassociated with a 0.01% increase in hedge funds’ net purchases, which is about 4% of the mean of netpurchases. I also do similar tests for non-hedge funds and mutual funds. The results, reported in Columns(3)-(6), show that abnormal positive outlook does not predict net purchases by mutual funds or non-hedgefunds in general. This is consistent with the fact that hedge funds are more active in trading on onlineinformation.Notice that these findings do not exclude the possibility that sophisticated investors have their ownsources of private information about firms’ fundamentals. If their private information is correlated with em-ployee outlook from Glassdoor, their trading activities can reflect either their own information or employeeoutlook signal. One way to disentangle the underlying sources is to examine whether trading activities canpredict abnormal positive outlook. In untabulated results, I show that neither hedge funds’ nor short sellers’trading activities can predict abnormal positive outlook. Although this finding does not completely rule outthat sophisticated investors have their own information sources, it suggests that employee outlook is at leastone type of information that they exploit. Moreover, the fact that abnormal returns of trading strategy basedon employee outlook decline over the holding horizon indicates that investors trade on this information.2.4.3 Insider tradingIn this section, I test whether employee outlook is related to insider trading. Although the overlap betweenemployees who are required to report insider trading (top executives or directors) and employees who post20Financial Times, January 21, 2017, “Hedge funds and private equity tap Glassdoor for investment tips.”16reviews on Glassdoor is not big, they may share the same information source as the reviewers. For example,both non-executive employees and top executives know that their company is doing well this quarter. Beforethe earnings announcement, non-executive employees might post reviews with positive outlook, while topexecutives might buy more shares of their company. SEC requires high-level employees to report trans-actions involving their own company stocks under certain conditions.21 I use the reported insider tradingtransactions from SEC EDGAR to test whether the information reported on Glassdoor is associated withinsider trading. Specifically, I aggregate the sells (buys) of insider trading for each firm to the monthly leveland calculate the abnormal sells (buys), which is the difference between sells (buys) and their mean in theprior three months. To test the conjecture, I regress the one-month-ahead insider abnormal sells (buys) onabnormal positive outlook and control variables.Table 2.11 presents the results for the insider trading test. In Columns (1)-(3), where the dependentvariable is insider sells, the coefficient on abnormal positive outlook is negative and significant for univariateregression and multivariate regressions with various controls and fixed effects. This finding suggests thatinsiders reduce their sells if employee outlook is very positive. As I show in Section 1.3, positive outlookpredicts higher future returns, and it is rational for insiders to reduce their sales if their know stock priceswill increase in the following month. Thus, it seems that top insiders are sharing similar opinions of thefirm’s future prospects with employees who write reviews on Glassdoor. Interestingly, this is not the casefor insider buys. In Columns (4)-(6), I find that the coefficient on abnormal positive outlook is not significantfor inside buys. This has something to do with the fact that it is relatively easier to sell a stock than to buyone as an employee. For many companies, employees can only buy shares of their own company in themiddle or at the end of each month, but they can sell their shares at any time.Overall, the evidence regarding insider trading suggests that the online information of employee outlookis strongly linked to information possessed by top employees, which is revealed in their insider tradingactivities.2.5 Information hierarchies: Employee ranks and the value of employeeoutlooksThe results so far show that employee outlook contains value-relevant information to the stock market. Re-cent studies on analysts forecasts show that it is important to examine the heterogeneity in analyst forecastsrather than just look at forecast concensus (Chiang et al., 2016; Michaely et al., 2017). Thus, an interestingquestion is whether there are significant quality differences in online forecasts across different levels of em-ployees. Are high-level employees’ outlooks more reliable than low-level employees’ outlooks? Anothermotivation is related to the literature on theories of organization hierarchies within firms. If a firm has a“top-down” structure (Garicano, 2000), high-level employees may possess better information than low-levelemployees. However, if a firm has a “bottom-up” structure, low-level employees as a group may hold better21Every director, officer, or owner of more than 10% of a class of equity securities registered under Section 12 of the SecuritiesExchange Act of 1934 must file a statement of ownership regarding such security with the United States Securities and ExchangeCommission. This file is called Form 4; see SEC website (https://www.sec.gov/files/form4data%2C0.pdf) for more details.17information than high-level employees. Even for a firm with a “top-down” structure, low-level employeesmay hold a wealth of information that is useful to predict future performance. Thus, it is an empiricalquestion whether high-level employees have better information than low-level employees.22I define the difference of information quality across employee ranks as “information hierarchies.” Adirect empirical test for information hierarchies is challenging for two reasons. First, it is hard to measureemployees’ information. Second, it is difficult to evaluate whose information is better or whether theirinformation is different. In this section, I use employees’ online reviews as a proxy for their information,and use the return predictability as a tool to gauge the value of information. If high-level employees’outlooks are better than low-level employees, then one would expect that the return predictability of high-level employees’ information to be better than low-level employees’ information.2.5.1 Rank hierarchies within firmsAlthough the employee review data contains a job title for each review, it is hard to classify hierarchiesjust based on job titles. For example, is a senior engineer at Google a high-level employee or middle-levelone? In the literature of firm organization hierarchies, some studies use international data that come withhierarchy classifications (Caliendo et al., 2015). Although the hierarchy code is not available for US firms,prior studies often use occupation title to define hierarchies (Bloom et al., 2012). Because employee rankis positively associated with wage (Caliendo et al., 2015), I rank employees’ job titles based on the averagewage of each title within a firm.23 For each firm, I then assign all job titles into three layers: high, middle,and low. Although having a large number of layers is appealing and allows us to explore rich dynamics, arelatively small number of three is conservative. If I find information hierarchies within three layers, I wouldexpect to see an even larger spread of information quality difference when there are more than three layers.Thus, the estimated magnitude of information hierarchies from three layers provides a lower bound.2.5.2 Information hierarchies: return predictabilityTo show information hierarchies, I test whether high-level employees’ outlooks can better predict futurestock returns than middle-level employees’ outlooks, and whether middle-level employees’ outlooks canbetter predict stock returns than low-level employees’ outlooks. I first match the employee review data withwage based on the reviewer’s job title and then assign all reviews into three groups based on wage withineach firm. I then form portfolios based on reviews by a certain level of employees. Specifically, in eachmonth, I construct abnormal positive outlook based on outlooks by high-level employees, assign stocksinto three portfolios, and track the performance of each portfolio in the next month. I run a Fama-French-Carhart four-factor regression test as in Equation (2.2) for the long-short portfolio. I repeat the steps formiddle- and low-level employees. For simplicity, I only report results based on value-weighted returns. Theresults, reported in Table 2.12 , show that high-level employees’ outlooks have strong return predictability22Several papers study the empirical side of organization hierarchy (e.g., Caliendo et al., 2015), but direct testing from aninformation perspective is limited.23Liu et al. (2017)find that distributions of wages on Glassdoor are representative and comparable to that of the AmericanCommunity Survey.18(Column (1)). AbnOutlook of middle-level employees can predict future return, but only at the 10% level.AbnOutlook of low-level employees does not predict increase in future stock returns. This finding providesevidence of information hierarchies within firms in the sense that high-level employees’ outlooks have betterreturn predictability than that of middle- and low-level employees.In theories of firm organization hierarchies, the degree of information hierarchy depends on a firm’sorganization structure. Also, it is possible that information hierarchies may only be concentrated amongcertain types of firms. To investigate this conjecture, I partition the sample of firms into two groups based ontheir organization structure. Specifically, I label a firm as complex (simple) if its number of occupation titlesis greater (smaller) than the median. For both complex and simple firms, I then repeat the same exerciseas above to calculate the four-factor alphas based on high-, middle-, and low-level employees’ outlooks.The results are reported in Columns (2)-(3) of Table 2.12 . Interestingly, for complex firms, only high-levelemployees’ outlooks can predict returns (Column (2)) and the economic magnitude of the four-factor alpha(1.70 % per month) is twice the alpha in Table 2.4 . However, for simple firms, employees’ outlooks atall levels can predict returns (Column (3)). In terms of economic magnitude, the four-factor alpha basedon high-level employees’ outlooks is about 0.29% greater than that for low-level employee outlooks. Thisdifference is statistically insignificant (the p-value of the F-test is around 0.5). These results suggest thatinformation hierarchies are more pronounced among complex firms. Only if the number of job titles is largeenough is the information of high-level employees significantly better than that of middle- and low-levelemployees. For firms with relatively fewer job titles, employees at all levels possess valuable informationabout the firm and their outlooks can predict returns at almost the same magnitude.Another way to look at the heterogeneity of information hierarchies is by firm size. Large firms tendto have more complex hierarchies. Thus, it is possible that information hierarchies only exist among largefirms. To test this conjecture, I partition the sample of firms into two groups based on their size. For bothlarge and small firms, I then repeat the same exercise as above to calculate the four-factor alphas based onhigh-, middle-, and low-level employees’ outlooks. The results, presented in Columns (4)-(5) of Table 2.12,show that high-level employees’ outlooks have stronger return predictability than middle- and low-levelemployees’ outlooks among large firms. However, this is not the case for small firms where employees at alllevels have valuable information. In terms of economic magnitude, middle-level employees’ outlooks are asgood as high-level employees’ outlooks for predicting future returns.The heterogeneity of information hierarchies also depends on whether a firm is a conglomerate or astandalone firm. Conglomerates tend to have more complex organization structures than standalone firms.Following the literature (Cohen and Lou, 2012; Boguth et al., 2016a), I define a firm as a conglomerate if itoperates in two or more industries.24 For both conglomerates and standalone firms, I then repeat the sameexercise as above to calculate the four-factor alphas based on high-, middle-, and low-level employees’outlooks. The last two columns in Table 2.12 show the results. While high-level employees’ outlooksare better than middle- and low-level employees’ outlooks among conglomerates, this is not the case forstandalone firms where employees at all levels have valuable information. This finding is consistent with24Industries are defined by Fama-French 49-industry classifications. The results are robust to using alternative classificationssuch as 3-digit SIC or 4-digit SIC code.19Cohen and Lou (2012), who find that the complicated structure of conglomerates leads to slow incorporationof information. Overall, there is evidence that information hierarchies are prevalent among firms. Thispattern is more pronounced among firms with complex organizational forms, large firms, and conglomerates.2.5.3 Information hierarchies: textual analysisWhy do high-level employees’ outlooks have better return predictability? Do their reviews contain differentinformation compared to low-level employees’? I examine information hierarchies from another perspec-tive: review texts. One popular method in the literature is to use word count, where researchers count thenumber of words of interest in a text and then divide by the total number of words in that text. While thismethod has its merits, it requires researchers to have good prior knowledge about what type of words theyare looking for in the texts. Researchers without prior knowledge about what they should expect from textsrequire more advanced methods, such as machine learning and deep learning.25Similar to other textual analysis methods, machine learning methods involve a first step to remove use-less information (i.e., stop words) and then represent the text as data. The representation process convertstext data to vectors of numbers by detecting latent patterns in transformed data (Gentzkow et al., 2017). Iuse Latent Dirichlet Allocation (LDA), which is a popular method in the finance and economic literature todetect latent topics among employee reviews (e.g., Hansen et al., 2017). Specifically, I use LDA to analyzepotential topics in employees’ reviews for three samples of firms: all firms, complex firms, and simple firms.The definitions of complex and simple firms are the same as noted above. For each sample, I use LDA tofind ten topics in the review texts (see the Appendix for details). After transforming each review into adistribution of ten topics, I calculate the average weight of each topic among high-, middle-, and low-levelemployees (defined the same way as in Section 1.5). The two most important topics are reported.The outputs of machine learning are summarized in Table 2.13 . Column (1) shows different topicsacross employee levels for all firms. There is clearly a difference between topics of high-, middle-, and low-level employees. High-level employee reviews focus more on business growth and career opportunities,middle-level employees focus on salary and culture of the company, and low-level employees care mostabout personnel development and work hours. Although it is hard to compare the importance of thesetopics, it is clear that some topics, such as business growth, are more related to a firm’s fundamentals, whileother topics, such as work hours, are less directly linked to a firm’s fundamentals.Column (2) presents the results for complex firms. Again, high-level employees’ reviews are aboutbusiness growth and career opportunities, while low-level employees’ reviews focus on benefits and workhours. However, the results for simple firms, reported in Column (3), display a different pattern: differentlevels of employees tend to have more overlap in LDA topics, such as leadership. This finding suggeststhat information contained in high-, middle-, and low-level employees’ reviews for firms with relativelysimple organization structures tend to focus on similar topics. Overall, the textual analysis indicates a largedispersion among latent topics in reviews by different levels of employees. The dispersion of topics in textsis further evidence of information hierarchies within firms.25A fast growing body of literature uses machine learning methods (Antweiler and Frank, 2004; Li, 2010).202.6 ConclusionThe relation between information and stock prices is one of the most fundamental questions in finance.Prior reseach highlights the importance of looking at alternative information sources that is beyond publicdomain (Roll, 1988), such as private information (Koudijs, 2015). Increasing access to information hasmade the problem of how non-traditional information is incorporated into stock prices ever more relevant.I use a novel dataset of one million online employee reviews to understand the value and processing of anew type of information, online forecasts, in stock markets. I first show that employee outlook containsvalue-relevant information and predicts future stock returns. The information embedded in employee out-look is different from accounting information, past returns, media coverage, and employee satisfaction. Thereturn predictability of employee outlook decays over five months. Employee outlook predicts trading ac-tivities by hedge funds and short sellers, suggesting that sophisticated investors exploit this information orits underlying sources.These findings highlight the role of non-experts in forecasting firms’ fundamentals through online plat-forms, which is beyond traditional information intermediaries such as sell-side analysts. Employees as agroup can potentially provide both more timely and better quality information on firms’ performance thananalysts.The results in this paper also have important implications with respect to the efficient market hypothesis(Fama, 1970). The strong-form efficient market hypothesis, where stock prices should reflect all public andprivate information, is often violated (e.g., Schwert, 2003). Private information, and even some public infor-mation, is not fully incorporated into stock prices. This study shows that some new public information fromemployees is valuable to the stock market and is incorporated into stock prices within five month, whichindicates market inefficiency. My results also contribute to the debate on whether greater information pro-duction leads to the improvements in financial market efficiency in past decades (Bai et al., 2016; Farboodiet al., 2017).21Figure 2.1: Top words in reviews with positive and negative outlookThis figure presents the top 10 words in reviews with positive or negative outlook. Starting from the raw texts ofreview title, pros, cons, and advice to management, I remove stop words and transform the remaining words to numericvectors. To account for the relative importance of a word in each review, a term-frequency-inverse document frequency(tf-idf) method is used in the transformation process. After transformation, each review is represented as a distributionof words with their weights. I then run a logistic regression where the dependent variable is a dummy variable that isequal to one if employee outlook is positive, and zero otherwise, and the independent variables are weights of words.The regression coefficients give the relative importance of each word and are reported on the y-axis. The blue bars arefor positive outlook, and red bars are for negative outlook.22Figure 2.2: Cumulative returnsThis figure shows the cumulative portfolio returns over the sample period for a high AbnOutlook and low AbnOutlookportfolio. Portfolios are formed at the end of June 2012, rebalanced each month, and then held to December 2016. Thesolid red line is the long portfolio and the dashed blue line is the short portfolio. Panel A shows the equal-weightedresults and Panel B shows the value-weighted results.Panel A. Equal weighted returnsPanel B. Value weighted returns23Figure 2.3: Return predictability over different horizonsThis figure plots the impact of employee outlook on stock returns over different horizons. In Panel A, I sort stocksinto tercile portfolios based on abnormal positive outlook in each month. I then track the performance of the threeportfolios two to 12 months after portfolio formation. In Panel B, I sort stocks into tercile portfolios based on abnormalpositive outlook in each month. I then track the performance of the three portfolios over one to four weeks afterportfolio formation. Portfolio returns are calculated using a value weighting method. All alphas are calculated usingthe Fama-French-Carhart four-factor model. In Panel C, I use a regression approach where the dependent variable isabnormal return adjusted by the market model and the independent variables include daily abnormal positive outlookand controls (size, turnover, trading volume, and past 30-day returns). Abnormal positive outlook is calculated for a30-day rolling window with its mean over the past 120 days as a benchmark. The regression coefficient on abnormalpositive outlook is reported in the figure.Panel A. Monthly portfolio returns Panel B. Weekly portfolio returnsPanel C. Daily regression coefficient24Table 2.1: Distribution of employee reviewsThis table reports summary statistics for the sample of employee reviews for S&P 1500 firms from June 2012 toDecember 2016. It reports the number of reviews and firms for 12 Fama-French industries.Fama-French 12 industries # of reviews # of firmsConsumer NonDurables 29,537 96Consumer Durables 10,669 36Manufacturing 52,368 151Oil, Gas, and Coal Extraction and Products 10,424 47Chemicals and Allied Products 16,600 40Business Equipment 235,885 261Telephone and Television Transmission 46,356 37Utilities 7,712 46Wholesale, Retail, and Some Services 295,272 188Healthcare, Medical Equipment, and Drugs 39,227 117Finance 143,637 205Other 85,330 198Total 973,017 1,42225Table 2.2: Summary statisticsThis table reports summary statistics for the sample of employee reviews for S&P 1500 firms from June 2012 toDecember 2016. Positive outlook is the number of reviews with positive outlook divided by the total number reviewsin each month for each firm. Abnormal positive outlook is Positive outlook minus its mean over the prior three months.Recommend is the fraction of reviews that state “recommend to a friend” divided by the total number of reviews in eachmonth for each firm. Overall is the average overall star rating (on a scale of 1 to 5) in each month for each firm. Themonthly average of culture and values rating (Culture), work-life balance rating (WorkLife), senior management rating(Management), compensation and benefits (Compensation), and career opportunities (Career) are calculated in thesame way. Market cap is market capitalization of the firm, calculated as the number of shares outstanding multipliedby the stock price. Market beta is a firm’s exposure to market risk, calculated from a 3-year rolling regression ofmonthly excess stock returns of the firm on market returns. Stock returns 12 months is the cumulative stock returnsover the past 12 months. Dollar volume is trading volume multiplied by stock price. Profitability is gross profitability,calculated as the ratio of income before extraordinary items to book value of assets. Institutional ownership is thefraction of shares owned by institutional investors. Analyst coverage is the number of analysts that cover a firm.Mean StdDev 25th Median 75thEmployee reviewsPositive outlook (%) 41.70 32.98 10.71 40.00 63.16Abnormal positive outlook (%) 0.35 32.81 -16.67 0.00 16.67Recommend (%) 56.01 33.05 33.33 57.89 81.25Overall 3.18 0.89 2.67 3.22 3.77Culture 3.15 0.97 2.57 3.17 3.83WorkLife 3.18 0.90 2.67 3.18 3.80Management 2.78 0.93 2.17 2.83 3.33Compensation 3.25 0.84 2.79 3.27 3.88Career 2.96 0.86 2.50 3.00 3.50Firm characteristicsMarket cap ($ mil) 17297.06 42043.39 1476.84 4340.63 14087.20Market Beta 1.22 0.57 0.84 1.17 1.53Stock returns 12 months (%) 14.01 34.50 -5.14 11.97 29.78Dollar volume 2524.75 6522.59 247.45 910.87 2711.98Profitability 0.20 0.14 0.10 0.17 0.27Institutional ownership 0.74 0.17 0.65 0.74 0.84Analyst coverage 13.84 8.69 7.00 13.00 20.0026Table 2.3: Employee expectations and firm characteristicsThis table reports determinants regression of abnormal positive outlook and lagged firm characteristics from June2012 to December 2016. Abnormal positive outlook is the fraction of reviews with positive outlook in each monthfor each firm minus its mean over the prior three months. Size is the logarithm of market capitalization of the firm,calculated as the number of shares outstanding multiplied by the stock price. Book-to-market is a firm’s book valuedivided by its market value. Beta is a firm’s exposure to market risk, calculated from a rolling regression of monthlyexcess stock returns of the firm on market returns. Stock returns 12 months is the cumulative stock returns over thepast 12 months. Positive media coverage is the number of positive news articles that mentioned a firm in the DowJones News Archives, provided by RavenPack. Institutional ownership is the fraction of shares owned by institutionalinvestors. Analyst coverage is the number of analysts that cover a firm. Profitability is gross profitability, calculatedas the ratio of income before extraordinary items to book value of assets. Dollar volume is trading volume multipliedby stock price. Firm and time fixed effects are included. Numbers in parentheses are t-statistics that are adjusted forheteroscedasticity and clustered by industry and time. ***, **, and * indicate statistical significance at the 1%, 5%,and 10% levels, respectively.(1) (2) (3)Abnormal positive outlookSize 0.00 0.01(0.16) (1.08)Book-to-market -0.00 -0.01(-0.01) (-0.66)Beta -0.00 -0.01(-0.37) (-0.57)Stock returns 12 months -0.00 -0.00(-0.24) (-0.61)∆Positive media coverage -0.00 -0.00(-0.58) (-0.33)Instititional ownership -0.03 -0.04(-1.09) (-1.16)Analyst coverage -0.00 -0.00(-0.18) (-0.29)Profitability -0.07 -0.08(-1.54) (-1.64)Log(Dollar volume) -0.00 -0.01(-0.93) (-1.18)Observations 26,534 26,534 26,534Adjusted R-squared 0.012 0.012 0.012Firm, Time FEs Y Y Y27Table 2.4: Employee expectations and stock returnsThis table reports portfolio returns results. In each month from June 2012 to December 2016, I sort sample stocks intotercile portfolios based on AbnOutlook, which is the fraction of reviews with positive outlook minus its mean over theprior three months. I then track the performance of the three portfolios over the following month. The portfolio returnsare calculated by two weighting schemes: equal- and value-weighting. Portfolio returns are adjusted by risk-free rate.Long-short portfolio buys the top tercile portfolio and sells the bottom tercile portfolio. The regression results in PanelB are based on Equation (2.2). Numbers in parentheses are t-statistics calculated using Newey-West standard errorswith four lags. ***, **, and * indicate statistical significance at the 1%, 5%, and 10% levels, respectively.Panel A. Portfolio excess returnsEqual-weighting Value-weightingPortfolio 1 (Low positive outlook) 0.71 1.03Portfolio 2 0.99 1.10Portfolio 3 (High positive outlook) 1.57 1.72Long-short (High-Low) 0.86*** 0.69***Panel B. Portfolio alphas(1) (2) (3) (4) (5) (6)Equal-weighting Value-weightingVARIABLES Portfolio 1 Portfolio 3 Long-short Portfolio 1 Portfolio 3 Long-shortAlpha -0.26** 0.51*** 0.78*** -0.05 0.56*** 0.61***(-2.09) (3.23) (3.79) (-0.49) (2.93) (3.62)MktRF 1.03*** 1.02*** -0.00 0.95*** 1.00** 0.05(3.52) (3.25) (-0.03) (2.93) (2.23) (0.90)SMB 0.54*** 0.42** -0.11*** 0.03 -0.09** -0.12(2.93) (2.60) (-2.87) (0.75) (-2.57) (-1.64)HML 0.07** 0.10 0.03 -0.07 0.02 0.09(2.19) (1.63) (0.51) (-1.17) (0.40) (0.80)MOM -0.10*** -0.04 0.06 -0.15*** -0.11*** 0.05(-3.15) (-1.22) (1.56) (-3.78) (-2.98) (0.80)28Table 2.5: Employee expectations and stock returns: subsampleThis table reports the Fama-French-Carhart four-factor alphas on a monthly long-short portfolio based on abnormalpositive outlook for various subsamples. In Panel A, I partition the sample of stocks into a large firm group and asmall firm group based on the median of market cap in the past year. In Panel B, I partition the sample of stocksinto a high-analyst-coverage group and a low-analyst-coverage group based on the median of analyst coverage in thepast quarter. In Panel C, I partition the sample of stocks based on industry labor intensity, which is the number ofemployees divided by Property Plant and Equipment. Stocks that belong to an industry with labor intensity higherthan the median are assigned to the high labor intensity group, while the rest are assigned to the low labor intensitygroup. In Panel D, I partition the sample based on employee status: current vs. former employees. In Panel E, Ipartition the sample based on whether an employee works in the headquarter state or not. Within each group of stocksin each month, I form a long-short portfolio that buys the top tercile stocks and sells the bottom tercile stocks basedon abnormal employee outlook and calculate the four-factor alphas. Numbers in parentheses are t-statistics calculatedusing Newey-West standard errors with four lags. ***, **, and * indicate statistical significance at the 1%, 5%, and10% levels, respectively.Equal-weighting Value-weightingPanel A: Firm sizeSmall firms 1.03*** 0.68***(3.09) (3.53)Large firms 0.40** 0.32**(2.17) (2.61)Panel B: Analyst coverageLow analyst coverage 0.88*** 0.69***(2.90) (3.34)High analyst coverage 0.47** 0.20(2.11) (0.70)Panel C: Industry labor intensityLow labor intensity 0.61** 0.35*(2.37) (1.83)High labor intensity 0.99*** 0.66***(3.05) (3.51)Panel D. Employee statusCurrent employee 0.78*** 0.63***(3.69) (3.48)Former employee 0.28** 0.29(2.65) (1.38)Panel E. Employee LocationHeadquarter state 0.83*** 0.82***(3.01) (2.85)Non-headquarter state 0.48** 0.32**(2.42) (2.31)29Table 2.6: Fama-MacBeth regressionsThis table reports the coefficient estimates of Fama-MacBeth regressions of one-month-ahead excess stock returns on abnormalpositive outlook and other cross-sectional predictors of stock returns as in Equation (2.3). Abnormal positive outlook is the fractionof reviews with positive outlook in each month for each firm minus its mean over the prior three months. 100 Best companiesis a dummy variable that is equal to one if a firm is on Fortune’s list of the “100 Best Companies to Work for” in that year, andzero otherwise. Size is the logarithm of market capitalization of the firm, calculated as the number of shares outstanding multipliedby the stock price. Book-to-market is a firm’s book value divided by its market value. Beta is a firm’s exposure to market risk,calculated from a rolling regression of monthly excess stock returns of the firm on market returns. Stock returns 12 months is thecumulative stock returns over the prior 12 months. Positive media coverage is the number of positive news articles that mentionedthe firm in the Dow Jones News Archives in each month, provided by RavenPack. Institutional ownership is the fraction of sharesowned by institutional investors. Analyst coverage is the number of analysts covering the firm. Profitability is the gross profitability,calculated as the ratio of income before extraordinary items to book value of assets. Dollar volume is trading volume multipliedby stock price. Numbers in parentheses are t-statistics calculated using Newey-West standard errors with four lags. ***, **, and *indicate statistical significance at the 1%, 5%, and 10% levels, respectively.(1) (2) (3) (4)VARIABLES One-month-ahead excess returnsAbnormal positive outlook 0.93*** 0.93*** 0.84***(2.82) (2.83) (2.81)100 Best companies -0.14 -0.13 -0.11(-0.60) (-0.56) (-0.49)Size -0.25*(-1.84)Book-to-market 0.78**(2.43)Beta -0.32(-1.54)Stock returns 12 months 0.00(0.00)∆Positive media coverage 0.02***(2.95)Institutional ownership -0.96*(-1.69)Analyst coverage 0.00(0.25)Profitability 0.99(1.53)Log(Dollar volume) -0.07(-0.50)Observations 27,023 27,023 27,023 27,023R2 0.004 0.002 0.006 0.06430Table 2.7: Employee expectations and firms’ earningsThis table reports regressions of earnings surprises and profitability on employee outlook and controls as in Equation (2.3) in thepaper. Abnormal positive outlook is the fraction of reviews with positive outlook in each quarter for each firm minus its mean overthe past three quarters. Size is the logarithm of market capitalization of the firm, calculated as the number of shares outstandingmultiplied by the stock price. Book-to-market is a firm’s book value divided by its market value. Beta is a firm’s exposure tomarket risk, calculated from a rolling regression of monthly excess stock returns of the firm on market returns. Stock returns 12months is the cumulative stock returns over the prior 12 months. Positive media coverage is the number of positive news articlesthat mentioned the firm in the Dow Jones News Archives in each quarter, provided by RavenPack. Institutional ownership is thefraction of shares owned by institutional investors. Analyst coverage is the number of analysts that cover the firm. Profitability isgross profitability, calculated as the ratio of income before extraordinary items to book value of assets. Dollar volume is tradingvolume multiplied by stock price. SUE is standardized unexpected earnings, calculated as the difference between realized quarterlyearnings and consensus of earnings by analyst forecasts divided by stock price at the end of that quarter. Cumulative abnormalreturns (CAR) are returns measured using a three-day window centered on the announcement date and adjusted using the marketmodel. Numbers in parentheses are t-statistics that are adjusted for heteroscedasticity and clustered by industry and time. ***, **,and * indicate statistical significance at the 1%, 5%, and 10% levels, respectively.(1) (2) (3) (4) (5) (6)VARIABLES Earnings surprises (SUE) CARAbnormal positive outlook 0.037*** 0.032** 0.028** 0.437** 0.443** 0.473**(2.97) (2.46) (2.04) (2.11) (2.10) (2.16)SUE_lag 0.059*** 0.044*** 0.091 0.071(3.91) (2.78) (1.29) (0.86)Size -0.024** -0.025* -0.069 -0.063(-2.27) (-1.68) (-0.52) (-0.40)Book-to-market 0.066*** 0.077** -0.015 0.202(2.89) (2.21) (-0.07) (0.66)Beta 0.013 0.007 0.004 -0.136(1.20) (0.39) (0.03) (-0.62)Stock returns 12 months 0.011 0.011 0.689*** 0.574*(0.72) (0.57) (2.69) (1.66)∆Positive media coverage 0.001 0.002 0.048*** 0.062***(0.81) (1.16) (2.69) (2.71)Instititional ownership -0.038 -0.078 -0.018 0.519(-0.93) (-1.10) (-0.04) (0.97)Analyst coverage 0.001 0.002 0.004 0.015(1.10) (1.53) (0.41) (1.15)Profitability -0.067* -0.085 -1.163*** -2.488***(-1.69) (-1.28) (-3.27) (-3.82)Log(Dollar volume) 0.023** 0.022 0.030 0.037(2.02) (1.45) (0.23) (0.23)Time, Firm FEs N N Y N N YObservations 13,947 13,947 13,947 13,905 13,905 13,905Adjusted R2 0.000 0.009 0.146 0.000 0.001 0.10531Table 2.8: Employee expectations and profitabilityThis table reports regressions of earnings surprises and profitability on employee outlook and controls as in Equation (2.3) in thepaper. Abnormal positive outlook is the fraction of reviews with positive outlook in each quarter for each firm minus its mean overthe past three quarters. Size is the logarithm of market capitalization of the firm, calculated as the number of shares outstandingmultiplied by the stock price. Book-to-market is a firm’s book value divided by its market value. Beta is a firm’s exposure tomarket risk, calculated from a rolling regression of monthly excess stock returns of the firm on market returns. Stock returns 12months is the cumulative stock returns over the prior 12 months. Positive media coverage is the number of positive news articlesthat mentioned the firm in the Dow Jones News Archives in each quarter, provided by RavenPack. Institutional ownership is thefraction of shares owned by institutional investors. Analyst coverage is the number of analysts that cover the firm. Profitability isgross profitability, calculated as the ratio of income before extraordinary items to book value of assets. Dollar volume is tradingvolume multiplied by stock price. SUE is standardized unexpected earnings, calculated as the difference between realized quarterlyearnings and consensus of earnings by analyst forecasts divided by stock price at the end of that quarter. Cumulative abnormalreturns (CAR) are returns measured using a three-day window centered on the announcement date and adjusted using the marketmodel. Return on asset (ROA) is income before extraordinary items divided by total assets. Numbers in parentheses are t-statisticsthat are adjusted for heteroscedasticity and clustered by industry and time. ***, **, and * indicate statistical significance at the 1%,5%, and 10% levels, respectively.(1) (2) (3) (4) (5) (6)VARIABLES 4ROA 4ProfitabilityAbnormal positive outlook 0.083*** 0.076*** 0.060*** 0.066*** 0.075** 0.046*(3.02) (3.61) (3.81) (3.67) (2.51) (1.79)Size 0.291 1.743 7.061*** 26.778***(0.24) (1.06) (3.84) (4.94)Book-to-market -2.844*** -1.683*** 1.261 -2.444***(-3.86) (-3.08) (0.49) (-2.62)Beta -3.020*** -2.919** -2.360 -0.238(-3.20) (-2.03) (-1.57) (-0.08)Stock returns 12 months 0.072*** 0.056*** 0.033 -0.012(4.57) (3.60) (1.48) (-0.46)∆Positive media coverage -0.006 -0.009 0.016 0.021(-0.47) (-0.62) (0.80) (0.87)Instititional ownership 6.486* -0.949 -1.645 22.733***(1.89) (-0.24) (-0.30) (2.68)Analyst coverage -0.158 -0.199** 0.017 -0.781*(-1.33) (-2.04) (0.10) (-1.95)Log(Dollar volume) 1.530 0.968 -1.800 -2.462(1.48) (0.84) (-1.32) (-0.78)Time, Firm FEs N N Y N N YObservations 17,608 17,608 17,608 17,608 17,608 17,608Adjusted R2 0.015 0.130 0.207 0.011 0.719 0.80932Table 2.9: Employee expectations and stock returns: robustnessThis table reports robustness tests of portfolio regression. Panel A presents results with various alternative methods tocalculate alphas. The HXZ four-factor alpha is calculated during the sample period of June 2012 to December 2015due to the availability of the HXZ factor data. Panel B presents results with a variety of alternative samples and all al-phas are calculated using the Fama-French-Carhart four-factor model. Panel C presents results with review-weightedportfolio returns returns with the alpha calculated using the Fama-French-Carhart four-factor model. Numbers inparentheses are t-statistics calculated using Newey-West standard errors with four lags. ***, **, and * indicate statis-tical significance at the 1%, 5%, and 10% levels, respectively.Equal-weighting Value-weightingPanel A: Alternative methodsUsing estimated outlook by PCA 0.63*** 0.30***(3.14) (2.75)Fama-French three-factor model 0.88*** 0.64***(3.53) (3.27)Fama-French five-factor model 0.89*** 0.64***(3.69) (3.16)Fama-French-Carhart six-factor model 0.87*** 0.61***(3.43) (2.74)Hou-Xue-Zhang four-factor model 0.90*** 0.65***(3.09) (3.41)Fama-French-Carhart with the liquidity factor 0.85** 0.61***(2.28) (2.77)AbnOutlook based on the mean over the prior 6 months 0.94*** 0.81***(3.94) (2.96)Panel B: Alternative samplesExtended sample 2008-2016 0.53*** 0.39***(3.10) (3.39)Sample with all CRSP-Compustat firms 1.08*** 0.64***(3.45) (2.78)Sample with more than 5 reviews per month 0.96*** 0.83***(3.24) (2.89)Sample excluding financial firms 0.68*** 0.53***(3.59) (2.98)Sample excluding non-US reviews 0.71*** 0.55***(2.88) (3.47)Panel C: Alternative weighting methodReview-weighted portfolios 0.71***(2.76)33Table 2.10: Employee expectations and institutional tradingThis table reports the results of regressions of one-quarter-ahead institutional trading on employee outlook. Hedge fund netbuy isthe change in hedge fund ownership compared to the past quarter. Similar definitions are applied to Non-hedge fund netbuy andmutual fund netbuy. Abnormal positive outlook is the fraction of reviews with positive outlook in each quarter for each firm minusits mean over the past three quarters. Stock returns is the stock returns at the end of that quarter. Stock returns 12 months is thecumulative stock returns over the past 12 months. Size is the logarithm of market capitalization of the firm, calculated as the numberof shares outstanding multiplied by the stock price. Book-to-market is a firm’s book value divided by its market value. Beta is afirm’s exposure to market risk, calculated from a rolling regression of monthly excess stock returns of the firm on market returns.Positive media coverage is the number of positive news articles that mentioned the firm in the Dow Jones News Archives, providedby RavenPack. Analyst coverage is the number of analysts that cover a firm. Profitability is gross profitability, calculated as theratio of income before extraordinary items to book value of assets. Dollar volume is trading volume multiplied by stock price.Numbers in parentheses are t-statistics that are adjusted for heteroscedasticity and clustered by industry and time. ***, **, and *indicate statistical significance at the 1%, 5%, and 10% levels, respectively.(1) (2) (3) (4) (5) (6)Hedge fund Non-Hedge fund Mutual fundnetbuy netbuy netbuyAbnormal positive outlook 0.012*** 0.011*** 0.001 0.003 0.003 -0.007(4.38) (3.87) (0.49) (0.90) (0.59) (-1.37)Stock returns 0.215*** -0.126*** 0.345***(13.29) (-6.95) (11.27)Stock returns 12 months -0.010* -0.012** 0.029***(-1.83) (-2.22) (7.45)Size -0.756 0.216 -4.516***(-1.16) (0.34) (-6.59)Book-to-market -0.244 0.379* -0.243(-1.24) (1.67) (-1.09)Beta 0.477 -0.523 0.746(0.80) (-0.86) (1.46)∆Positive media coverage 0.140 -0.170 0.243(1.32) (-1.37) (1.24)Analyst coverage -0.007 0.005 0.104**(-0.16) (0.11) (2.19)Profitability 0.256 1.649 2.398(0.12) (0.72) (1.15)Log(Dollar volume) -0.000* 0.000* -0.000(-1.85) (1.65) (-0.71)Time, Firm FEs N Y N Y N YObservations 16,102 16,102 16,102 16,102 16,102 16,102Adjusted R2 0.059 0.138 0.058 0.105 0.064 0.15434Table 2.11: Employee outlook and insider tradingThis table reports the results of regressions of one-month-ahead insider trading on employee outlook. Abnormal positive outlookis the number of reviews with positive outlook divided by the total number of reviews in each month for each firm minus its meanover the past three months. Stock returns is the stock returns in that month. Stock returns 12 months is the cumulative stock returnsover the past 12 months. Size is the logarithm of market capitalization of the firm, calculated as the number of shares outstandingmultiplied by the stock price. Book-to-market is a firm’s book value divided by its market value. Beta is a firm’s exposure to marketrisk, calculated from a rolling regression of monthly excess stock returns of the firm on market returns. Positive media coverage isthe number of positive news articles that mentioned the firm in the Dow Jones News Archives, provided by RavenPack. Institutionalownership is the fraction of shares owned by institutional investors. Analyst coverage is the number of analysts that cover a firm.Profitability is gross profitability, calculated as the ratio of income before extraordinary items to book value of assets. Dollarvolume is trading volume multiplied by stock price. Numbers in parentheses are t-statistics that are adjusted for heteroscedasticityand clustered by industry and time. ***, **, and * indicate statistical significance at the 1%, 5%, and 10% levels, respectively.(1) (2) (3) (4) (5) (6)VARIABLES one-month-ahead insider sells one-month-ahead insider buysAbnormal positive outlook -0.025*** -0.024*** -0.029*** -0.007 -0.009 -0.014(-2.69) (-2.62) (-3.06) (-0.69) (-0.84) (-1.26)Stock returns -0.400*** -0.443*** 0.369*** 0.360***(-11.29) (-10.74) (7.89) (7.01)Stock returns 12 months 0.039*** 0.018** -0.048*** -0.032***(6.76) (1.98) (-7.32) (-3.14)Size 0.005*** 0.072*** -0.002* -0.050***(4.07) (8.29) (-1.72) (-4.68)Book-to-market 0.000 0.019** 0.001 -0.033**(0.12) (2.00) (0.21) (-2.13)Beta 0.003 0.006 0.007** 0.012(1.32) (0.78) (2.57) (1.42)Media coverage -0.001*** -0.001* -0.000 -0.000(-3.27) (-1.79) (-1.64) (-0.21)Analyst coverage -0.000** -0.001 0.001** 0.002**(-2.51) (-0.82) (2.55) (2.03)Institutional ownership -0.014 -0.078** -0.017 -0.040(-1.39) (-2.15) (-1.57) (-0.91)Profitability -0.019** -0.058* -0.002 0.059(-2.33) (-1.65) (-0.16) (0.90)Log(Dollar volume) -0.000 -0.000 0.000 -0.000(-0.25) (-0.69) (0.20) (-0.84)Time, Firm FEs N N Y N N YObservations 26,226 26,226 26,226 26,226 26,226 26,226Adjusted R2 0.000 0.004 0.139 0.000 0.003 0.11235Table 2.12: Information hierarchies within firms—evidence from returnsThis table reports information hierarchy within firms. I first match the employee reviews data with wage based onreviewer’s job title and then assign all reviews into three groups based on wage within each firm. I then form portfoliosbased on reviews by a given level of employees. Specifically, in each month, I construct portfolios based on abnormalpositive outlook among high-level employees, assign stocks into three portfolios, and track the performance of eachportfolio in the following month. I run a four-factor regression test as in Equation (2.2) for the long-short portfolio.I repeat the steps for middle- and low-level employees. For simplicity, I only report results based on value-weightedreturns. In Columns (2)-(3), I partition the sample of firms into two groups based on number of job titles. A firm witha number of job titles that is greater (less) than the median is labeled a complex (simple) firm. Within each group offirms, I then repeat the tests described above based on reviews by high-, middle-, and low-level employees. Similarly, Ipartition the sample of firms based on the median of firm size in Columns (4)-(5). In Columns (6)-(7), a conglomerateis defined as a firm that operates in two separate industries, while a standalone is a firm that only operates in oneindustry. Numbers in parentheses are t-statistics calculated using Newey-West standard errors with four lags. ***, **,and * indicate statistical significance at the 1%, 5%, and 10% levels, respectively.All firmsHierarchy Firm size OrganizationComplex Simple Large Small Conglomerates Standalone(1) (2) (3) (4) (5) (6) (7)High-level employees 0.84*** 1.70** 0.69*** 0.93*** 0.88** 0.85*** 0.84***(3.73) (2.21) (3.89) (2.89) (2.03) (2.87) (3.63)Middle-level employees 0.37* 0.42 0.56** 0.48 0.83** 0.35 0.70***(1.74) (0.50) (2.56) (1.64) (2.15) (1.25) (4.06)Low-level employees 0.33 0.35 0.38*** 0.33 0.68** 0.37 0.48***(1.48) (0.51) (2.69) (1.57) (2.15) (1.58) (2.80)36Table 2.13: Information hierarchies within firms—evidence from textual analysisThis table reports information hierarchies within firms by analyzing latent topics in employee reviews. Using LatentDirichlet Allocation (LDA), I first find ten topics in review texts and transform each review as a distribution of topicswith a weight for each topic. I then calculate the average weight of each topic among high-, middle-, and low-levelemployees and report the two most important topics. In Columns (2) and (3), I partition the sample of firms into twogroups based on number of job titles. A firm with a number of job titles that is greater (less) than the median is labeleda complex (simple) firm. Within each group of firms, I repeat the LDA analysis and find the two most important topics.(1) (2) (3)All firms Complex firms Simple firmsHigh-level employees business growth business growth business growthcareer opportunities career opportunities leadershipMiddle-level employees salary flexibility leadershipculture personnel development flexibilityLow-level employees personnel development work hours salarywork hours benefits leadership37Chapter 3Macro News, Micro News, and Stock Prices3.1 IntroductionThe link between information and asset prices is one of the most important questions in finance. Investorsoften face multiple sets of information at the same time. Thus, the interaction between different typesof information is crucial for understanding how information is incorporated into stock prices. Broadlyspeaking, there are two types of news: economy-wide (macro) and firm-specific (micro) news. In this paper,I study the interaction between macro news and the processing of firms’ earnings announcements, as wellas the implications for asset pricing.Existing theories suggest that macro news should impede the processing of earnings news. Rationalattention theories state that investors have limited attention (e.g., time and cognitive resources). If they payattention to macro news, then they have less attention for firm-level news. For instance, Peng and Xiong(2006)provide theoretical evidence that retail investors tend to process market- and sector-wide informationbefore processing firm-specific information because there are economies of scale in processing the market-and sector-wide information. In Kacperczyk et al. (2016), institutional investors (mutual fund managers) ra-tionally allocate more attention to aggregate shocks in recessions and idiosyncratic shocks in booms becausethe aggregate shocks are more uncertain during the recession. In sum, these theories suggest a substituterelationship between macro and firm-level news because the presence of macro news distracts investors,leading them to pay less attention to earnings news. Prior studies show that investor attention is importantfor the processing of earnings news.26 Thus, the substitute relationship predicts that macro news reduces theefficiency of price reactions to earnings news.My empirical tests, however, find the opposite: macro news improves the processing of firm-level news.In particular, I examine whether macro news affects the processing of earnings news by focusing on a set offour important macro announcements. I find that the immediate price reaction to a firm’s earnings surpriseis significantly stronger and the drift is significantly weaker when the macro news is released on the sameday. This result suggests that earnings information released on macro-news days is incorporated into stockprices faster, leading to more efficient stock valuation. The finding remains after controlling for existingfactors that affect investor reaction to earnings news, such as the number of earnings news, the day of theweek, and the level of market returns. The economic magnitude of this effect is significant. Firms with the26It is well-documented that investors react to earnings announcements slowly and earnings news is incorporated into stockprices only gradually, which is called “post-earnings announcements drift.” One important reason for the drift is that not all of theinvestors pay attention to earnings announcements when the information is released, leading to an under-reaction in the aggregateresponse of prices to news (Ball and Brown, 1968; Bernard and Thomas, 1989). Recent studies find that investors are distracted byother firms’ earnings announcements (Hirshleifer et al., 2009), and they pay less attention when the earnings news is released onFriday (DellaVigna and Pollet, 2009).38largest earnings surprises on macro-news days experience a 17% higher immediate price reaction and a 71%lower post-earnings announcement drift compared to reactions to earnings surprises on other days. Overall,my findings suggest a complementary relationship between macro and firm-level news.I also find this complementary relationship exists on days with many macro announcements. Using a fulllist of macro announcements from Bloomberg, I identify days with a large number of macro announcements.I find that investors’ immediate reaction to earnings announcements is 12% stronger and the drift is 46%weaker when earnings announcements are released on days with many macro announcements.The complementary relationship between macro and micro news is further confirmed by a trading strat-egy. The post-earnings announcement drift suggests that firms that have positive (negative) earnings sur-prises experience an increase (decrease) in stock prices following announcements. Thus, a trading strategybased on the drift buys stocks with positive earnings surprises and shorts stocks with negative earnings sur-prises. Such a strategy does not have abnormal returns for earnings released on macro-news days, although itstill has about 1% abnormal return per month among earnings announcements released on non-macro-newsdays. In short, the trading strategy approach confirms the effect of macro news on investors’ reactions toearnings announcements.The complementary relationship also displays some persistence and heterogeneity. The macro-newseffect is most pronounced when macro-news and earnings announcements are released on the same day, andit still lasts when macro news is released one or two days after earnings announcements. The macro-newseffect completely disappears if macro news is three or more days before/after earnings announcements. Theresults are more pronounced among large firms and firms with high analyst coverage.Why is the processing of earnings news more efficient on macro-news days? Why do my findingscontradict the existing theories? I suggest extensions of existing theories that could be consistent with thesefindings. Existing models focus on allocating attention between two dimensions: macro and micro news.However, these theories miss some important dimensions when modeling investors’ decisions. Investors’daily lives involve not only processing different kinds of financial news, but also devoting time and effort toother activities unrelated to the stock market.27 Investors may hierarchically allocate their attention amongthese activities first, and then condition on a stock market focus, divide the stock-market attention betweenmacro and micro news. Macro news, like the Federal Open Market Committee (FOMC) decisions, areusually attention-grabbing events, drawing investors’ attention to the stock market.28 As a result, investors’attention is market focused, allowing them to pay more attention to earnings announcements on macro-newsdays. Although this third dimension is economically intuitive and compelling, it has been overlooked in the27Retail investors spend time on their jobs and leisure activities. Even institutional investors, like fund managers, have dutiesthat, in addition to managing portfolios, also include networking with their clients (e.g., playing golf) or researching new tradingstrategies. There is evidence that even institutional investors are distracted by various factors (Corwin and Coughenour, 2008;Kempf et al., 2016; Schmidt, 2016).28For example, consider a framework where an investor can allocate his attention (i.e., time and other resources) to two areas:a) following the stock market, b) playing golf. On a regular day, 10% of his attention is allocated to following the stock market,and 90% attention is spent on playing golf. The 10% attention to the stock market is fully used for earnings news since there isonly earnings news on a regular day. On a macro-news day, he assigns 50% of attention to both the stock market and golf. Amongthe 50% of his attention to the stock market, 20% is used to process earning news, and 30% is use to analyze macro news. So, theattention to earnings news increases on days with macro-news.39existing literature.29To test this explanation, I use a measure of abnormal institutional investor attention (AIA) from Bloomberg(Ben-Rephael et al., 2017). I find that AIA is higher on macro-news days in general and that AIA to firmswith earnings announcements is significantly higher when macro news is released on the same day. Theimportance of institutional investors is further confirmed by the finding that the macro-news effect is con-centrated among firms with high institutional ownership. Investors not only pay more attention to the stockmarkets on macro-news days but also trade more on such days. I find that investors’ trading volume reactionto earnings announcements is substantially higher on macro-news days. Together, these results suggest thatthe impact of macro news on investors’ reactions to earnings announcements is strongly related to investors’attention.I also rule out several alternative explanations. One concern is that firms’ exposure to the market (i.e.,market beta) is higher when they release earnings announcements, and their betas decrease from the secondtrading day after news release (Patton and Verardo, 2012). Thus the stronger immediate reactions andweaker drift may be driven by changes in firms’ risk exposure. However, I only find weak evidence thatfirms’ market risk loadings are associated with the macro-news effect. Also, the risk-based explanationcannot explain why the immediate reaction to earnings announcements is still higher even if the macro newsis released one or two days after the earnings news. According to Patton and Verardo (2012), firms’ betashave already fallen by the time macro news is released, meaning there should be lower immediate reactions.Thus, the macro-news effect is associated with something other than risk.Another concern is that the macro-news effect may be driven by a firm’s liquidity premium. Prior studiesshow that a firm’s earnings announcement premium is associated with its liquidity (Sadka, 2006; Frazziniand Lamont, 2007). It is possible that firms with a positive earnings surprise have more liquidity on macro-news days. However, I do not find evidence supporting this hypothesis. Using two measures of liquidity, Ifind that liquidity is higher on macro-news days in general. However, firms with high earnings surprises donot have more liquidity on days with macro news compared to firms with low earnings surprises. Moreover,I find that the macro-news effect is mainly concentrated among large firms and firms with high analystcoverage (i.e., they are liquid firms). Overall, these findings make liquidity a less likely explanation.One may concern that there is information transmission from macro news to earnings announcements.The presence of macro news may help investors process earnings news in a better or faster manner, result-ing in stronger reactions to earnings announcements. Macro news is one type of information that affectsinvestors’ expectations of firm values. When investors face both macro news and firm-specific news, thepresence of macro news may affect how investors process firm-specific information. Although the informa-tion spillover explanation is appealing, I do not find any supporting evidence for this theory. I find that themacro-news effect does not depend on the content of macro news.One may also be concerned that if firm managers are aware of the increase in investors’ reactions toearnings news on macro-news days, they may strategically choose dates to announce their earnings. Thisstrategic manipulation may bias the results if firms tend to advance their earnings announcements dates to29Goldstein and Yang (2015) provide theoretical evidence that the presence of complementarities between two firm-level signalsfacilitates information acquisition and improves price informativeness.40macro-news days and delay those dates to non-macro-news days. However, I find that the macro-news effectis concentrated among firms that do not significantly change their earnings announcements, suggesting thatstrategic timing is unlikely to drive the effect.3.1.1 Related literatureThis paper contributes to several strands of literature, including the literature on rational inattention asdiscussed above.30 This study is also related to several recent studies on the determinants of investors’reactions to earnings announcements.31 For instance, prior research finds that investors can get distractedand decrease their reactions to an earnings announcement when there is a greater number of same-dayearnings announcements from other firms (Hirshleifer et al., 2009), when earnings are announced on Fridays(DellaVigna and Pollet, 2009), when market return is low (Gulen and Hwang, 2012), and when the earningsnews is released after larger earnings surprises are announced by other large firms on the same day or theprevious day (Hartzmark and Shue, 2017). This paper extends this line of research by studying the impactof macro news on investor reactions to earnings announcements. Macro news is distinct from any factorconsidered by prior literature.By studying the returns of earnings announcements on macro-news days, this paper adds to the literatureon the impacts of macro news on stock markets in two ways (e.g., Andersen et al., 2003; Boyd et al., 2005;Gilbert, 2011; Gilbert et al., 2017).32 First, I provide a more direct test of the effect of macro news onmarket efficiency. Savor and Wilson (2014) find that an asset pricing model like CAPM works better forstock returns on macro-news days, suggesting that the stock market is more efficient on macro-news days.Their conjecture of the impact of macro-news on market efficiency is from the joint hypothesis in the sensethat it measures the fit of an asset-pricing model. This paper directly tests the effect of macro news onmarket efficiency through an event study. Second, this paper offers a comprehensive examination of theunderlying mechanisms through which macro news affects asset prices. While some studies propose a risk-based explanation (Savor and Wilson, 2013), others suggest that alternative channels need to be considered(Lucca and Moench, 2015; Cieslak et al., 2016; Bernile et al., 2016). My findings suggest that a factorbeyond risk − investor attention − is very important for understanding the impact of macro news on stockprices; the risk-based explanation is not enough to explain this impact.30The literature on rational inattention dates back to Kahneman and Tversky (2013) and Sims (2003).31Other studies that examine earnings announcements include Garfinkel and Sokobin (2006) and Schmalz and Zhuk (2018).32More generally, this paper also relates to the literature on investors’ attention and stale news (Huberman and Regev, 2001),media coverage (Chan, 2003; Barber and Odean, 2008; Hillert et al., 2014; Kaniel and Parham, 2017), sports news (Edmans et al.,2007; Schmidt, 2013), related firms’ news (Cohen and Frazzini, 2008), and the level and volatility of the stock market (Sichermanet al., 2016).413.2 Data3.2.1 Macroeconomic announcementsThere are many types of macroeconomic announcements, and almost every day there is at least one macroannouncement.33 However, not every announcement is important for the stock market. Thus, I first selecta set of important macro announcements from a list of 40 macro announcements by Bloomberg Econoday.I define a day to be a macro-news day (hereafter, Macroday) if one of the following four announcementshappens on this day: the Federal Open Market Committee (FOMC) decision, Nonfarm Payroll, ISM PMI,and Personal Consumption. These days make up 23% of all trading days. The rationale for selecting thesefour announcements is given below. In section 3.3.4, I show that the results are robust to an alternativedefinition of macro-news days.Following the method in Savor and Wilson, 2013, I test whether the market excess return (market returnminus risk-free rate) is significantly higher on announcement days for each macroeconomic announcement.I find that announcements that have statistically and economically significant impacts on the market excessreturn include FOMC, Nonfarm Payroll, ISM PMI, and Personal Consumption. The results are providedin the Appendix. The importance of the FOMC announcement is well documented (see, e.g., Lucca andMoench, 2015; Cieslak et al., 2016). Gilbert et al. (2017) find that announcements, including Nonfarmpayroll, ISM PMI, and Personal consumption, are important for financial markets.3.2.2 Earnings newsI obtain quarterly earnings release data from Compustat and I/B/E/S as micro news from 1998 to 2014.Following Hirshleifer et al. (2009), I measure earnings surprise (ES) using Equation (3.1). It is the differencebetween actual earnings (Actual) for the quarter recorded by I/B/E/S and the median forecast (Forecast)included in the I/B/E/S detail file during the 30 days before the quarterly earnings announcements scaled bythe stock price at the end of the corresponding quarter.ES =Actual−ForecastPrice(3.1)Stock price response to earnings news is measured by cumulative abnormal return (CAR) for each stock,which is the raw buy-and-hold return adjusted using estimated beta from the market model. For each earn-ings announcement date τ of quarter t, I define the cumulative abnormal return over time period (τ + h,τ+H) CAR[h,H] as followsCAR[h,H] =[τ+H∏j=τ+h(1+R j,k)−1]− βˆt,k[τ+H∏j=τ+h(1+R j,m)−1](3.2)where R j,k is the stock return of company k on day j, R j,m is the market return on day j, and βˆt,k is obtainedfrom the market model regression R j,m = αt,k+βt,kR j,k for days j from τ−300 to τ−46. For the immediate33For example, there are more than 130 different macro announcements according to Bloomberg’s Econoday calendar(http://mam.econoday.com/).42stock price reaction, I use CAR over a 2-trading-day window [0, 1]. For drift, I use CAR over a 60-trading-day window [2, 61].34 I exclude the penny stocks, observations in which actual or forecast earnings aregreater than stock price, and those with a missing earnings surprise. The final sample includes 158,399observations.3.2.3 Summary statisticsTable 3.1 Panel A reports summary statistics based on the full sample. It shows that, on average, thereare 118 earnings announcements per day. The mean immediate reaction to an earnings announcement(CAR[0,1]) is 0.1 %, and the mean of the drift (CAR[2,61]) is 1%. Panel B shows the same statistics, con-ditional on being on a Macroday, compared to all other days. On average, Macrodays have a significantlyfewer number of earnings announcements and higher market return. Firms that release their earnings an-nouncements on macro-new days have significantly higher immediate reaction to earnings news (CAR[0,1]),abnormal trading volume (AVOL[0,1]), and lower drift (CAR[2,61]).3.3 The effect of macro news on the processing of earnings news3.3.1 Main resultsIn this section, I first test whether reactions to earnings announcements on days with macro news are signif-icantly different from reactions on other days by focusing on firms that have the most positive and negativeearnings surprise. Following DellaVigna and Pollet (2009), I rank firms’ earnings surprise and assign theminto 11 quantiles for each year. Firms with negative surprises are equally assigned to quantiles 1 to 5, andfirms with positive surprises are equally assigned to quantiles 7 to 11.35 Firms with zero surprises are labeledas quantile 6. In this section, I focus on the top and bottom groups, quantiles 1 and 11, because this makesit easy to interpret the magnitude of the effect. To empirically test this effect, I run the following regressionCAR = a0+a1ESTOP+a2Macroday+a3(ESTOP×Macroday)+n∑i=1[biXi+ ci(ESTOP×Xi)]+ e (3.3)where CAR is either CAR[0,1] for immediate reaction, or CAR[2,61] for drift. ESTOP equals to 1 if theearnings surprise quantile is 11 and 0 if the earnings surprise quantile is 1. Macroday is a dummy variableequaling 1 if that day is an announcement day for any FOMC, Nonfarm payroll, ISM PMI, or Personalconsumption news. Xi contains various control variables. Previous research shows that stock response toearnings news varies with firm size, analyst coverage, day of the week, and the number of the same-day34In Section 3.1, I show that the results are robust to alternative choices of windows.35In Section 6, I show that the results are robust to alternative choices of earnings surprise groups, like deciles. In general,the earnings announcement literature uses earnings surprise quantiles rather than the raw value of earnings surprise because itsdistribution is nonlinear. Thus, it is problematic to use the raw earnings surprises in a linear regression (Bernard and Thomas,1989).43earnings announcements (e.g., Bernard and Thomas, 1989; DellaVigna and Pollet, 2009; Hirshleifer et al.,2009). Thus, I include size deciles, analyst coverage, share turnover, day of week/month/year dummies, andthe number of earnings announcements per day as control variables.The key coefficient of interest is a3. If the relationship between macro news and earnings announcementsis complementary, then the investors’ immediate reaction to earnings announcements is stronger, and thedrift is weaker when macro news is released on the same day. Thus, I expect that a3 > 0 for CAR[0,1] anda3 < 0 for CAR[2,61]. In contrast, if the relationship between announcements is substituted, then I expectthat a3 < 0 for CAR[0,1] and a3 > 0 for CAR[2,61].Table 3.2 reports the results of this test. Column (1) presents the result from a parsimonious specificationwithout including any control variables. The coefficient on the interaction term (ESTOP×Macroday) ispositive and significant at the 1% level (1.277), suggesting that the price reaction to an earnings announce-ment with a big surprise is stronger on macro-news days than on other days. The economic magnitude is alsosignificant. Compared to the coefficient on the stock reaction to a top earnings surprise (ESTOP) on otherdays (8.352), the reaction on Macroday is greater by 15% (1.277/8.352). The economic magnitude increasesto 17% (1.373/8.127) if control variables are included. This is comparable to the 15% reduction for Fridayannouncements documented in DellaVigna and Pollet (2009) , and the 13% reduction for high-news-dayearnings announcements documented in Hirshleifer et al. (2009).For the drift, the coefficient on the interaction term is negative (-3.458 without controls, -3.682 withcontrols) and significant at the 5% level, suggesting that post-earning announcement drift is smaller fortop surprise earnings announcements released on macro-news days compared to other days’ earnings news.Column (4) shows that my estimates indicate 71% (3.458/4.846) smaller drift for earnings announcementsreleased on macro-news days. Again, the economic magnitude is significant and comparable to prior studies.Hirshleifer et al. (2009) report that the post-earnings announcement drift is 75% greater for high-news-day earnings announcements compared to low-news day announcements. DellaVigna and Pollet (2009)find that the drift is 69% greater for Friday earnings announcements compared to other weekday earningsannouncements.To further understand the nature of this differential drift, I compare the drift differences over varioushorizons in Figure 3.1. The drift is defined as the difference between average cumulative abnormal returns ofthe top group and those of the bottom group. The difference in drifts between Macroday and non-Macrodayannouncements becomes clear from the 10th trading day after the earnings announcement and continues toincrease during the next 60 trading days. The drift on Macroday announcements increases quickly duringthe first 10 trading days after announcements but decreases slightly until the 50th trading day. However, thedrift on non-Macroday announcements displays a totally different pattern. It increases quickly during thefirst 10 trading days and continues to increase until the 60th trading day. These patterns suggest that earningsnews released on macro-news days is almost fully incorporated in prices within 10 trading days followingthe announcement. However, earnings news released on non-Macroday requires significantly more time tobe incorporated into the stock price.I then examine how macro news affects investors’ reactions to earnings announcements across all earn-ings surprises quantiles. To empirically test this effect, I estimate the following regression44CAR = d0+d1ES+b2Macroday+d3(ES×Macroday)+n∑i=1[ fiXi+gi(ES×Xi)]+ ε (3.4)where most variables are similarly defined as in Equation (3.4). ES is the earnings surprise quantile, whichequals 1 to 11. Again, the coefficient on the interaction term (d3) is the key parameter of interest.Table 3.3 reports the regression results. Consistent with Panel A, the coefficient on the interactionterm (ES×Macroday) is positive and significant for CAR [0,1], suggesting that immediate stock responseto earnings news is stronger on Macrodays than on other days. As for the economic magnitude, comparedto the coefficient on the stock reaction to earnings surprise on other days (0.842), the reaction is greaterby 11% (0.092/0.848) to earnings news released on Macrodays (Column (2)). For the drift, the coefficienton the interaction term is negative and significant at the 1% level, suggesting that the drift is smaller forearnings news released on Macrodays compared to earnings news on other days. Column (4) shows that myestimates indicate 52% (0.201/0.388) smaller drifts for earnings announcements released on Macrodays.It is also informative to pay attention to the controls. First, I find that the immediate price reaction toearnings announcements is much smaller if the news is released on Friday, which is consistent with DellaV-igna and Pollet (2009). Second, earnings announcements released on days with a high number of earningsnews releases experience much weaker immediate reaction and much stronger drift, which is consistentwith Hirshleifer et al. (2009). Third, earnings released on days with high market returns have much strongerimmediate reactions, which is consistent with Gulen and Hwang (2012). Macro news is distinct from anyfactor considered in the prior literature. Macro-news days are clearly different from Fridays as macro newscan be announced on any day of the week. Compared to the number of earnings news, macro news is an al-together different type of information than earnings news. Although both the market return and macro newsare market-wide variables, macro news is pre-scheduled and is associated with information release, whilethe market return is unpredictable ex-ante. Thus, the asset pricing implications and channels through whichmacro news affects investors’ reactions to earnings announcements are different compared to the existingexplanations. Several robustness tests controlling for these factors are provided in Section 3.6.While the measure of immediate price reaction to earnings announcements is consistent across dif-ferent studies (i.e., CAR[0,1]), different studies use different measures to capture drift. Most studies useCAR[2,61] as the measure of drift as suggested by Bernard and Thomas (1989), and some studies use longerhorizons like CAR[2,75] (see, e.g., DellaVigna and Pollet, 2009). To address concerns that the findings inTable 3.3 depend on the choice of window to measure the drift, I conduct the same exercise as Equation(3.4) using different measures of the drift. Table 3.4 presents the result and demonstrates that the drift issignificantly lower on macro-news days for alternative definitions of the drift horizon, which is similar tothe main finding in Table 3.3.453.3.2 Lead and lags effects and heterogeneitySo far, I have focused on days in which both macro news and earnings announcements are released. Whatabout when macro news is released several days before or after earnings announcements? The lead and lageffects are particularly interesting as the key element of this paper is the timing of the news release.To test whether the lead and lag of macro news is important or not, I conduct similar tests as in Equation(3.4) but with a different definition of Macroday. For cases where macro-news days are one-day beforethe earnings announcements, Macroday equals to 1 if there is macro news on day t − 1 for an earningsannouncements released on day t. A similar definition applies to other lead and lag windows. I examinecases where macro news is released one to three days before and after earnings announcements.Table 3.5 presents the results of these tests. Panel A shows that the immediate reaction is significantlyhigher if macro news is released one day or two days after the earnings announcement. The economicmagnitude is smaller than that of the main result in Table 3.3, but there is no such effect on the drift. Thereis no statistical significance in either the immediate reaction or the drift if macro news is released threedays after earnings news. Panel B shows that if macro news is released two or three days before earningsannouncements, reactions to these earnings announcements are not statistically different from reactions toearnings announcements on other days. The immediate reaction to earnings announcements is higher ifthere is macro news released one day before, but the effect is marginally significant. Again, the effect on thedrift is insignificant. These results suggest that the macro-news effect on immediate reactions is quite broadand applied to cases where macro news is released one to two days around earnings announcements. Thisfinding also helps identify explanations for the impact of macro news on investors’ reactions to earningsannouncements, which is discussed in Section 3.5.In addition, I also examine whether the macro-news effect varies with firms’ size, analyst coverage, andinstitutional ownership, all of which have been shown to be important for the processing of earnings news.In particular, I examine the effect separately for small, medium, and large firms. Table 3.6 Panel A showsthat the effect is most pronounced for large firms. Similarly, I find that the effect is concentrated amongfirms with high analyst coverage (Panel B) and high institutional ownership (Panel C).3.3.3 The number of macroeconomic announcementsThus far, I have focused on important macro news and have established that the relationship between macronews and earnings news is complementary in the sense that immediate (delayed) reactions to earnings newsincrease (decrease) on macro-news days. This finding suggests that the quantity of information does not pre-vent investors from processing it. To further test this idea, I examine whether investors’ reactions to earningsannouncements are different on days with many macroeconomic announcements. Using a full list of macroe-conomic announcements from Bloomberg Econoday, I identify days with a large number of macroeconomicannouncements. The cutoff point for the top 10% of the number of macroeconomic announcements is 7.Thus, I define a “High Macro News” day as one that has 7 or more macro announcements.Table 3.7 Panel A presents the results of this test. The coefficients on the interaction terms are positiveand significant for immediate reaction, and negative and significant for the delayed reaction, for both the46full sample and the sample of the top and bottom earnings surprise groups. These results suggest thatinvestors’ immediate reactions to earnings announcements increase and delayed reactions decrease whena large number of macro announcements are released on the same day. The economic magnitudes aresignificant as well. Thus, this confirms that the relationship between macro news and earnings news iscomplementary.Investors’ immediate reactions to earnings news are stronger and the drift is weaker when there areimportant macro announcements or a significant number of macro announcements. What about days withimportant macro news and a significant number of macro information releases? I expect to find an evenstronger pattern on those days, which is confirmed by Table 3.7 Panel B. Economic and statistical signif-icance of the coefficients on the interaction term increase. Compared to the immediate price reaction toearnings surprise on other days (0.847), the reaction is greater by 17% (0.140/0.847) when important macronews and a significant number of macro news are released on the same day. The drift is smaller by 84%(0.341/0.380).3.3.4 Portfolio trading strategyAnother way to test the impact of macro news on investors’ reaction to earnings announcements is to usea trading strategy designed to capture this impact. I have shown that the drift is substantially smaller formacro-day announcements than for non-macro-day announcements. In a typical drift portfolio, I go longstocks with good earnings news and short stocks with bad earnings news. If investors underact to earningsnews, then stocks with good (bad) earnings news will enjoy an increase (decrease) in returns within thefollowing quarter. However, as shown earlier, the profit of this type of trading strategy is far lower for stockswith earnings announcements on macro-news days because the drift is 70% weaker.The new drift trading strategy based on macro news is as follows. In month t, it purchases firms that,in month t -1 made announcements on a non-macro-day in the top decile and sells short firms that made anannouncement on a non-macro-day in the bottom decile. Therefore, the return for the non-macro-day driftportfolio is RDNM = R11NM −R1NM. I construct the macro-day drift portfolio for month t following a similarprocedure except that I only include firms that made an earnings announcement on a macro-news day inprevious month. The return for this portfolio is RDM = R11M −R1M. The long-short portfolio of buying thenon-macro-day drift portfolio and selling the macro-day portfolio has return, RDNM−M = RDNM −RDM. Theintuition here is that conducting the traditional drift trading strategy on a macro-news day is not profitableor has negative profit. Thus, shorting the macro-day drift portfolio and longing the non-macro day driftportfolio will be profitable, if macro-news indeed impacts investors’ reactions to earnings announcements.Table 3.8 presents the results of this trading strategy. Column (1) shows that a non-macro-day driftportfolio earns a return of 0.970% per month, while the return on the macro-day portfolio is much smallerand statistically insignificant (column (2)). The long-short portfolio earns 0.891% per month (column (3)).The results are similar if portfolios are constructed by a value-weighted method. Standard risk factors,such as Fama-French three-factor are controlled in the regression. A similar conclusion is reached using anequally-weighted method for portfolio construction.Overall, consistent with the complementary relationship, macro news has positive effects on the sensi-47tivity of stock returns’ immediate reaction to earnings surprises and negative effects on the sensitivities ofstock returns’ delayed reaction to earnings surprises.3.4 Investor attentionSo far, this study has established that immediate price reactions to earnings announcements increase anddrifts decrease when macro news is released on the same day. This section tests the explanation of investorattention.As discussed in the literature review, one major reason for the drift is that investors do not pay fullattention to earnings news. Thus, one possible explanation for increased reactions to earnings announce-ments when macro news is released is that investors pay more attention to earnings news on macro-newsdays. Macro news, like FOMC, is usually an attention-grabbing event, leading investors to financial markets.Given that investors have limited attention, they may rationally choose to pay more attention to the stockmarket when important macro news is released. As a result, the fraction of investors who react to earningsnews increases.The economic rationale is as follows. Investors care about macro news because the information frommacro news is one of the most important pieces of information for their investment decisions. As shownearlier, the immediate price reaction is much stronger and the drift is much weaker when macro news andearnings news are in the same direction. This suggests that the presence of macro news may affect theinformativeness of earnings news, leading to its improved processing. When investors choose to allocatetheir attention across time, days with important macro news are among their top choices. They may alsotrade more on these days because macro news has an important impact on stocks (with or without earningsannouncements) in their portfolios.To test the attention explanation, I use two direct measures of attention. One measure is abnormal insti-tutional investor attention (AIA), which captures the news-searching and news-reading activity for specificstocks on Bloomberg terminals. Bloomberg assigns a raw score based on the number of ticker searches andthe number of clicks on related articles for each firm. The AIA is a relative index compared to the previousmonth’s average of the raw score and has a value from 0 to 4 (see Ben-Rephael et al., 2017, for more details).The other measure is Google Search Volume Index (SVI), which captures the ticker-searching activity foreach firm. Prior studies show that SVI is more informative about the attention of retail investors (Da et al.,2011; Drake et al., 2012).I first examine whether investors pay more attention to stocks with earnings announcements on macro-news days. Table 3.9 presents the result of this test. In Column (1), the coefficient on Macroday is positiveand significant, indicating that institutional investor attention to all firms is significantly higher on macro-news days compared to other days. I define Eday as a dummy for whether there is an earnings announcementfor each firm. The coefficient on Eday is positive and significant, suggesting that attention to firms is signifi-cantly higher when firms have earnings announcements. Most importantly, the coefficient on the interactionterm is positive and significant. This indicates that institutional investors pay more attention to firms whenearnings announcements are released on macro-news days compared to when earnings announcements are48released on non-macro-news days. Interestingly, I find no evidence that retail investor attention to the stockmarket is higher on macro-news days in general (Column (3)) and attention to firms with earnings announce-ments is even lower on macro-news days (Column (4)). Overall, these findings suggest that the macro-newseffect is strongly related to institutional investors’ attention.I further test the attention explanation by looking at the volume reaction as in Hirshleifer, Lim, andTeoh (2009) and DellaVigna and Pollet (2009). It is well known that trading volume increases on days withinformation releases or large price moves (Karpoff, 1987), and aggregate trading volume increases whenthere is a market-wide attention-grabbing event (Yuan, 2015) and when media attention on macroeconomicfundamentals is high (Fisher et al., 2017). If the significant difference in immediate price reactions betweenmacro-day and non-macro-day earnings announcements is caused by investor attention, I expect tradingvolume reaction to earnings announcements on macro-news days to be substantially higher than on otherdays because trading is the mechanism that causes the price to adjust.Following DellaVigna and Pollet (2009), I measure the abnormal trading volume as follows:AVOL[ j] = log(Vt+ j +1)− 110t−11∑k=t−20log(Vk +1) (3.5)where Vt+ j is the dollar value of trading volume at t + j. Immediate abnormal volume response is overa 2-day window (AVOL[0,1]) and defined as the average of abnormal trading volumes on the earningsannouncement date (AVOL[0]) and on the following day (AVOL[1]). 36Table 3.10 tests whether the abnormal trading volume to firms with earnings announcements is higheron macro-news days. Since both extreme positive and negative surprises cause changes in trading volume,I use the absolute earnings surprise decile (AES) here. Columns (1)-(2) show that firms’ abnormal tradingvolume is significantly higher on macro-news days. This indicates that there is more trading of firms thatannounce earnings on macro-news days compared to the trading of firms that release their earnings news onother days.3.5 Alternative explanationsThe complementary relationship between macro and micro news may be related to some factors other thaninvestor attention. This section tests several alternative explanations.3.5.1 RiskAnother possible explanation is related to risk. A firm’s abnormal return (AR) is defined as the differencebetween its realized return and expected return under the certain risk factor model. Under the market model,ARt = rett−βˆ×mktrett , where rett is a firm’s stock returns on announcement date t, βˆ is the estimated marketbeta during the estimating window [t−46, t−300], and mktrett is the market return on announcement date t.The realized return can be expressed as rett = β ×mktrett + εt , where the β on the earnings announcementdate. Prior research finds that the market beta of an announcing firm is higher on earnings-announcement36The results are similar if I use the measure of abnormal trading volume as in Hirshleifer et al. (2009) (see Appendix).49days (Patton and Verardo, 2012), which means β > βˆ . Thus, the stronger abnormal return (i.e., strongerimmediate reaction) may be driven by a firm’s greater exposure to market risk on earnings announcementdays. Similarly, since a firm’s market beta decreases following earnings announcement days (Patton andVerardo, 2012), the weaker abnormal return following a news release (i.e., weaker drift) may be caused bythe lower beta after earnings-news release. Additionally, since Savor and Wilson (2014) find that the CAPMworks better on Macrodays, this effect of a higher market beta may be more pronounced on Macrodays.To test this hypothesis, I augment Equation (3.4) by adding a set of additional variables: the Fama-French three-factor, the momentum factor (Fama and French, 1993; Carhart, 1997) and their interactionswith ES, Macroday, and ES×Macroday. If changes in risk matter, then we expect the coefficients on thethree-way interaction term to be significant. Table 3.11 reports the results for this test. The coefficientson the three-way interaction of market risk factor are positive and significant for immediate reaction andnegative and significant for the delayed reaction, both at a 10% level. Other risk factors are not important inexplaining the macro-news effect here.Another way to test the risk-based explanation is to look at price reactions when macro-news is releasedseveral days before or after earnings announcements. Patton and Verardo (2012) find that firms’ beta de-creases starting from the second day following the earnings announcements. If we still find that investors’reactions to earnings announcements are stronger on macro-news days even when earnings announcementsare released several days before macro news, then changes in beta are unlikely to explain the complemen-tarity. However, as shown earlier, I find the opposite evidence (Table 3.3). Immediate price reaction to anearnings announcement is significantly higher when macro news is released one or two days later. This find-ing cannot be explained by the risk-based story. According to Patton and Verardo (2012), firms’ betas havealready fallen by the time the macro news is released, meaning there should be lower immediate reactions.Thus, the stronger immediate reactions are associated with something other than risk.3.5.2 Trading frictionsAlternatively, the macro-news effect may be driven by firms’ liquidity premium. It is possible that firmswith positive earnings surprises are more liquid or have lower trading costs on Macrodays. Prior literatureshows that the earnings announcement premium is associated with liquidity risk (Sadka, 2006; Frazzini andLamont, 2007).Table 3.12 tests for such a channel. I use two measures of liquidity: bid-ask spread and turnover. Fora firm, greater bid-ask spread means it is less liquid, while higher turnover means it is more liquid. If theliquidity story holds, we would expect that the coefficient on the interaction term ES×Macroday is significantand positive. However, Table 3.12 shows that the coefficients are not significant. Thus, liquidity is unlikelyto account for the macro-news effect.3.5.3 Information transmissionAnother possible explanation for the effects of macro news on investors’ reaction to earnings news is thatmacro news provides additional information and therefore helps investors to have a better understanding50of earnings news. Grossman and Stiglitz (1980) show that if factors other than genuine information affectasset prices, then rational agents will collect information. In reality, investors collect information all thetime, suggesting that factors other than genuine information are important for investment decisions. Macronews is clearly one type of information that investors care about. The presence of macro news may affectthe informativeness of earnings news. Thus, it possible that investors learn from macro news and then reactto earnings news. If the information content of macro news matters, then investors have stronger reactionswhen earnings and macro announcements are in the same direction (i.e., both are positive or negative),compared to reactions when these two sources of news have different directions.In order to test this hypothesis, I use a similar regression as in Equation (3.4) with additional variables(Macro positive) capturing the content of macro news and its interaction with earnings surprise quantiles.Macro positive equals to 1 if the market return is positive on that macro-news day. Among all macro-newsdays, market returns are either positive or negative. One dummy variable is sufficient.This regression compares whether the effect of macro news depends on the direction of two types ofnews. Notice that ES is earnings surprise quantiles, which has the greatest positive surprise in quantile 11.If this hypothesis holds, then I expect that the coefficient on the interaction term ES×Macro positive orES×Macronegative will be significant. Table 3.13 shows the results for this test. The coefficients on theinteraction term are not significant for both immediate and delayed reactions, suggesting that the sign ofmacro news surprise does not matter. Thus, I find little evidence to support this hypothesis.3.5.4 Strategic timing of earnings announcementsA further possible explanation is that firms may strategically choose the dates to announce their earningsif they are aware that the macro news is salient. Prior research finds that firms advance or delay theirearnings announcements relative to the schedule used in the previous year. They tend to advance the earningsannouncement date with good news and delay the date with bad news (Boulland and Dessaint, 2017; Johnsonand So, 2017). Following Hartzmark and Shue (2017) and Johnson and So (2017), I identify firms that movetheir earnings announcement dates by comparing their current earnings announcement dates to the previousyear’s earnings announcement dates. Specifically, I categorize firms as having advanced or delayed theirearnings dates if they differ from their previous same-quarter date by five or more days. I find that roughly80% of firms do not significantly change (five day or more) their earnings announcement dates, 15% advancethem by five or more days, and 5% delay them by five or more days.A direct way to test strategic timing is to perform subsample analysis. If firms indeed strategicallytime macro-news days, then one would expect that the macro-news effect is concentrated among firms thatsignificantly change their earnings announcements dates. Table 3.14 Panel B rejects this hypothesis. Column(1) shows that firms that did not greatly change their announcement dates have a large positive coefficient of0.097 on the immediate reaction that is statistically and economically significant. Firms that changed theirearnings announcements forward or backward have insignificant estimates of the effects of macro news onreaction to earnings news. Columns (3) and (4) reach similar conclusions for the drift. Thus, the evidencedoes not support the strategic timing hypothesis.The strategic manipulation may bias our results only if firms with efficient trade (i.e., weaker drift) tend51to advance their dates to days with macro news and firms with less efficient trade (i.e., stronger drift) delaydates to days without macro news. If this were the case, then one would expect that for firms that changetheir earnings announcements day to an earlier date, the average earnings surprise on macro-news dayswould be more positive. One would expect the opposite for firms that delayed their earnings announcementdates. Table 3.14 Panel A tests this. Among firms that significantly advance their earnings announcementdates, about 14% (1137/(1137+7202)) move the new announcement date to a macro-news day. Among firmsthat significantly delay their earnings announcement dates (more than 5 days), about 15% of them move thenew announcement date to a macro-news day. Thus, there is no evidence that firms consider or are awareof the fact that reactions to earnings are significantly different on macro-news days when they change theirearnings announcement dates. Also, the t-statistics suggest that there is no significant difference in earningssurprises between firms that change earnings announcement dates to macro-news days and firms that changedates to other days.3.6 RobustnessThis section provides several robustness tests. One test is designed to address the concern that the set of firmsthat announce their earnings news on macro-news days are always the same. Two other tests address factorsthat affect investors’ reactions to earnings announcements in the prior literature. One factor is the number ofearnings news documented by Hirshleifer et al. (2009), while the other one is the market return as in Gulenand Hwang (2012). These factors are included as control variables in all the tests I conducted above, and donot affect the findings discussed above. Thus, they are not likely to drive the results. Nevertheless, I providefurther tests by excluding observations that can potentially contribute to the macro-news effect.3.6.1 Firms with strong preference of announcements datesOne concern is that firms that choose to announce on macro-news days are always the same set of firms.If this is the case, the macro-news effect of reactions to earnings news is just the difference between thisset of firms and other firms. To address this concern, I calculate the fraction of firms that always issue theirearnings announcements on macro-news days. Specifically, I create an Abnormal Announcement Preference(AAP) ratio for each firm, which is the number of earnings announcements on macro-news day divided bythe total number of its announcements. There are no firms where all earnings announcements are releasedon macro-news days (AAP ratio=1). Among firms that release earnings news on macro-news days at leastonce, less than 3% (114) of firms release more than 50% of their earnings news on macro-news days. Thisaccounts for only 13% even if I count firms that issue more than 33% of their earnings announcements onmacro-news days.I then formally conduct a test by re-estimating Equation (3.4) without these firms. Table 3.15 Panel Areports the results of this test. It shows that the macro-news effect on reactions to earnings news remainsstatistically and economically similar as in Table3.3. Thus, my results cannot be driven by a small set offirms that have strong preference of announcement dates.523.6.2 The number of earnings announcementsHirshleifer et al. (2009) find that investors’ immediate reactions to earnings announcements are much weakerand drift is much stronger when a large number of earnings are issued by other firms on the same day.Given that macro-news days have slightly fewer earnings announcements (Table 3.1 Panel B), one may beconcerned that the macro-news effect is driven by days with a low number of earnings news. I address thisconcern by removing days with a low number of earnings news (bottom quantile) and present the resultsin Table 3.15 Panel B. It shows that the macro-news effect is the same as in Table 3.3 at both statisticaland economic levels. Thus, the macro-news effect on reactions to earnings announcements is a distinctcontributor that cannot be explained by the number of earnings news.3.6.3 Stock market swingsGulen and Hwang (2012) show that investors’ immediate reactions to the corporate event, including earningsannouncements, are much stronger and delayed reactions are much weaker when earnings are released ondays with high market returns and the earnings surprises are positive. To the extent that both macro-newsand market returns are aggregate variables, one may be concerned about the new implications from macro-news compared to market returns. The fact that market returns and macro-news are correlated (Savor andWilson, 2013) and market returns affect investors’ reactions to earnings news does not mean that macronews is not a distinct phenomenon for studying investor behaviors. Macro-news is different from marketreturns for at least two reasons. First, macro-news affects stock market returns, but not the opposite. Also,many factors move stock market returns. Thus, the impact of market returns on investor behavior can comefrom factors other than macro news. Second, macro news is associated with information release and itsimpact on reactions to earnings news provides a unique setting to study the interaction between two typesof information. This is crucial in understanding the channels through which macro-news affects investors’behavior. Macro-news announcement dates are pre-scheduled. This makes the investor attention explanationmore plausible as investors can plan to allocate their attention beforehand.To address the concern that macro-news and market returns are the same driving force for the changesin investors’ reactions to earnings news, I re-estimate Equation (3.4) by excluding days with high marketreturns (top quantile). Table 3.15 Panel B reports the results of this test. The macro-news effect is barelyaffected by removing these observations, suggesting that market return swings cannot explain this effect.3.6.4 Alternative measuresI also test whether the results are robust to alternative measures of investor reactions and earnings surprisegroups. First, instead of using the market model, I use the Fama-French Three-Factor model when calculat-ing CAR[0,1] and CAR[2,61] and re-estimate Equation (3.4). Table 3.15 Panel C presents the results. Thecoefficient on the interaction term is positive and significant for CAR[0,1] (Column (1)), and negative andsignificant for CAR[2,61] (Column (2)). Thus, the results are similar to the main findings in Table 3.3. Theeconomic magnitudes of the coefficients are also similar. Moreover, I use 10 groups of earnings surprise andre-estimate Equation (3.4) and the results remain qualitatively and quantitatively similar (Columns (3)-(4)).53Overall, the macro-news effect is robust to the choice of model in calculating the reaction measures.3.7 ConclusionHow does the interaction of two types of information affect stock prices? The answer to this question iscrucial for understanding the functioning of stock markets. Ever increasing access to information has madethe problem of how investors process information when they face multiple sources of news at the same timemore relevant. I examine the interaction between macro news and earnings announcements and documenta novel complementary relationship between these two types of information. The presence of macro newsimproves investors’ information-processing of earnings announcements. Their immediate price reaction toearnings announcements is 17% higher and the drift is 71% lower when macro news is released on the sameday.The impact of macro news on investors’ reaction is strongly related to increased attention to earningsannouncements by institutional investors on macro-news days. Existing literature, however, assumes thatinvestors are 100% attentive to the stock market and the attention allocation is just between two types ofnews. My findings suggest a new framework where investors allocate attention not only between differenttypes of financial news, but also between financial news and activities unrelated to the stock market. Newtheories along this line are a promising area for future research.Furthermore, these results provide new evidence that the stock market is more efficient on macro-newsdays because the post-earnings-announcement drift is significantly smaller. This finding furthers our un-derstanding of the time-varying properties of market efficiency. While the concept of market efficiencydates to Fama (1970), the dynamics of market efficiency for individual stocks have not been studied untilrecently (Savor and Wilson, 2014; Engelberg et al., 2017; Birru, 2017). The improved price efficiency onmacro-news days indicates that market efficiency varies through time in a predictable way.54Figure 3.1: Performance of drift at different horizonsThis figure plots the cumulative abnormal returns over different horizons. The sample covers January 1997 to Decem-ber 2014. Cumulative abnormal return for each stock is the based market model. In event time, day 0 is the day ofearnings announcement. X-axis is the event time window, and Y-axis is average cumulative abnormal returns (Quantile11 minus Quantile 1).55Table 3.1: Summary statisticsThis table reports summary statistics. The sample covers January 1997 to December 2014. SUE is earnings surprise,# Earnings news is number of earnings announcements per day, # Analyst is the number of analysts following thefirm, Market cap is the market capitalization, Share turnover is the turnover of a firm’s share, and Market return is thedaily value-weighted market return from CRSP, CAR[0,1] is the cumulative abnormal return based on market modelover days [0,1], CAR[2,61] is the cumulative abnormal return based on market model over days [2,61], AVOL[0,1] isthe average abnormal trading volume on the earning announcement day AVOL[0] and on the following day AVOL[1],where abnormal trading volume on day t is the difference between log dollar volume and the average log dollar volumeover days [-20,-11]. Macroeconomic news days (Macroday) include days with announcements of Federal Open MarketCommittee (FOMC) decision, Employment situation, ISM PMI, or personal consumption.Panel A. Full sampleCount Mean SD P25 P50 P75ES % 158399 -0.01 1.10 -0.05 0.04 0.21# Earnings news 158399 118 79.44 46 107 180# Analyst 158399 6.03 5.78 2 4 8Market cap($ml) 158399 5187 20513 238 735 2617Share turnover % 158399 2.42 4.01 0.48 1.22 2.83Market returns % 158399 0.04 1.31 -0.60 0.09 0.67CAR[0,1] % 158399 0.10 8.54 -3.77 0.02 3.96CAR[2,61] % 158399 1.05 27.16 -12.44 -0.68 11.44AVOL[0,1] 158018 0.64 0.73 0.19 0.62 1.07Panel B. Sample of Macroday vs. sample of other daysCount Mean Mean comparisonMacroday Other days Macroday Other days Macroday Other daysES % 18876 139523 -0.004 -0.010 0.006 0.76# Earnings news 18876 139523 110 119 -9 -13.92# Analyst 18876 139523 6.12 6.02 0.10 2.16Market cap($ml) 18876 139523 4895 5227 -332 -2.09Share turnover % 18876 139523 2.63 2.39 0.24 7.78Market returns % 18876 139523 0.25 0.01 0.24 23.75CAR[0,1] % 18876 139523 0.24 0.08 0.16 2.34CAR[2,61] % 18876 139523 0.70 1.09 -0.40 -1.88AVOL[0,1] 18833 139185 0.70 0.63 0.07 12.8956Table 3.2: The macro-news effect–top groupsThis table reports the macro-news effect among top groups based on earnings surprises. The sample covers January1997 to December 2014. The dependent variable is cumulative abnormal return and is indicated under each columnheading. ES is earnings surprise quantile (11 groups), ES Top equals to 1 if earnings surprise quantile is 11 and 0if the earnings surprise quantile is 1. Macroday is a dummy variable equaling 1 if day t is an announcement dayfor Federal Open Market Committee (FOMC) decision, Employment situation, ISM PMI, or personal consumption.Control variables include the number of earnings announcements, the number of analysts following the firm, analystdispersion, market capitalization, share turnover, market return, and dummy variables for year, month, and day ofweek. Standard errors are adjusted for heteroscedasticity and clustered by the day of earnings announcement. ***, **,and * indicate statistical significance at the 1%, 5%, and 10% levels, respectively.(1) (2) (3) (4)CAR[0,1] CAR[2,61]ES Top 8.352*** 8.127*** 4.846*** 5.190***(0.175) (0.179) (0.527) (0.535)Macroday -0.667* -0.514 1.912 1.468(0.369) (0.377) (1.349) (1.358)(ES Top)×Macroday 1.277*** 1.373*** -3.458** -3.682**(0.446) (0.450) (1.504) (1.504)Friday -0.614 1.452(0.385) (1.298)Size 0.255*** -0.322***(0.036) (0.101)# Analyst -0.917*** -0.572(0.147) (0.426)# Earnings news -0.193* 0.804**(0.109) (0.316)Turnover 0.128*** 0.022(0.043) (0.059)Market return top 0.352** 1.314**(0.165) (0.533)Constant -4.491*** -3.299*** -0.385 3.328(0.137) (0.633) (0.446) (2.275)Controls N Y N YObservations 26,460 26,460 26,460 26,460Adj. R2 0.119 0.124 0.004 0.01857Table 3.3: The macro-news effect–all firmsThis table reports the macro-news effect for all firms. The sample covers January 1997 to December 2014. The de-pendent variable is cumulative abnormal return and is indicated under each column heading. ES is earnings surprisequantile (11 groups). Macroday is a dummy variable equaling 1 if day t is an announcement day for Federal OpenMarket Committee (FOMC) decision, Employment situation, ISM PMI, or personal consumption. Control variablesinclude the number of earnings announcements, the number of analysts following the firm, analyst dispersion, mar-ket capitalization, share turnover, market return, and dummy variables for year, month, and day of week. Standarderrors are adjusted for heteroscedasticity and clustered by the day of earnings announcement. ***, **, and * indicatestatistical significance at the 1%, 5%, and 10% levels, respectively.(1) (2) (3) (4)CAR[0,1] CAR[2,61]ES 0.848*** 0.842*** 0.357*** 0.388***(0.011) (0.011) (0.029) (0.029)Macroday -0.459** -0.354* 1.056* 0.776(0.186) (0.183) (0.586) (0.588)ES×Macroday 0.089*** 0.092*** -0.192** -0.201***(0.026) (0.025) (0.076) (0.076)Friday -0.263** 0.342(0.112) (0.359)Size 0.138*** -0.306***(0.011) (0.030)# Analyst -0.213*** 0.286**(0.044) (0.112)# Earnings news -0.184*** 0.155*(0.032) (0.091)Turnover -0.235*** -0.003(0.017) (0.021)Market return top 0.183*** 0.967***(0.051) (0.154)Constant -5.737*** -5.019*** -1.513*** 0.981(0.079) (0.226) (0.212) (0.728)Controls N Y N YObservations 158,399 158,399 158,399 158,399Adj. R2 0.086 0.100 0.002 0.00858Table 3.4: Drift over different horizonsThis table reports the impact of macro news on drift over different horizons. The sample covers January 1997 toDecember 2014. The dependent variable is cumulative abnormal return and is indicated under each column heading.ES is earnings surprise quantile (11 groups). Macroday is a dummy variable equaling 1 if day t is an announcement dayfor Federal Open Market Committee (FOMC) decision, Employment situation, ISM PMI, or personal consumption.Control variables include the number of earnings announcements, the number of analysts following the firm, analystdispersion, market capitalization, share turnover, market return, and dummy variables for year, month, and day ofweek. Standard errors are adjusted for heteroscedasticity and clustered by the day of earnings announcement. ***, **,and * indicate statistical significance at the 1%, 5%, and 10% levels, respectively.(1) (2) (3) (4) (5) (6)CAR[2,30] CAR[2,45] CAR[2,61] CAR[2,75] CAR[2,90] CAR[3,61]ES 0.250*** 0.306*** 0.388*** 0.372*** 0.376*** 0.333***(0.018) (0.023) (0.029) (0.035) (0.038) (0.028)Macroday 0.135 1.262** 0.785 0.572 0.538 0.679(0.397) (0.571) (0.589) (0.660) (0.685) (0.575)ES×Macroday -0.100* -0.213*** -0.201*** -0.186** -0.206** -0.183**(0.051) (0.072) (0.076) (0.087) (0.090) (0.074)Constant 0.506 0.714 1.152 1.541* 2.039** 1.584**(0.462) (0.553) (0.727) (0.819) (0.850) (0.719)Controls Y Y Y Y Y YObservations 158,399 158,399 158,399 158,399 158,399 158,399Adj. R2 0.005 0.005 0.008 0.006 0.005 0.00859Table 3.5: Lead and lag effectsThis table presents the lead and lag effect of macro news and earnings news. The sample covers January 1997 toDecember 2014. The dependent variable is cumulative abnormal return and is indicated under each column heading.ES is earnings surprise quantile (11 groups). “One day before” indicates that the macro news announcement is one daybefore the earnings announcement. The same definition applies to other lead and lag windows. For cases where themacro-news day is one-day before the earnings announcements, Macroday equals to 1 if there is macro-news on dayt-1 for an earnings announcement released on day t. Macro announcements include Federal Open Market Committee(FOMC) decision, Employment situation, ISM PMI, or personal consumption. Control variables include the numberof earnings announcements, the number of analysts following the firm, analyst dispersion, market capitalization, shareturnover, market return, and dummy variables for year, month, and day of week. Standard errors are adjusted forheteroscedasticity and clustered by the day of earnings announcement. ***, **, and * indicate statistical significanceat the 1%, 5%, and 10% levels, respectively.Panel A. Macro news is released after earnings newsOne day after Two days after Three days afterCAR[0,1] CAR[2,61] CAR[0,1] CAR[2,61] CAR[0,1] CAR[2,61]ES×Macroday 0.061*** -0.028 0.063*** -0.032 0.023 -0.040(0.021) (0.061) (0.022) (0.063) (0.022) (0.068)Panel B. Macro news is released before earnings newsOne day before Two days before Three days beforeCAR[0,1] CAR[2,61] CAR[0,1] CAR[2,61] CAR[0,1] CAR[2,61]ES×Macroday 0.045* -0.064 0.019 -0.102 0.006 -0.039(0.023) (0.070) (0.023) (0.066) (0.023) (0.071)60Table 3.6: HeterogeneityThis table reports how the macro-news effect varies by institutional ownership, size, and analyst coverage. The samplecovers January 1997 to December 2014. The dependent variable is cumulative abnormal return and is indicatedunder each column heading. ES is earnings surprise decile (11 groups), Macroday is a dummy variable equaling1 if day t is an announcement day for Federal Open Market Committee (FOMC) decision, Employment situation,ISM PMI, or personal consumption. Panel A reports the tests on three subsamples partitioned based on firm sizedecile. Small, medium, and large firms are in size decile 1 to 3, 4 to 7, and 8 to 10, respectively. Panel B reportsthe tests on three subsamples partitioned based on analyst coverage. Low, medium, and high coverage firms are firmsin decile 1 to 3, 4 to 7, and 8 to 10, respectively. Panel C reports the tests on three subsamples partitioned based oninstitutional ownership (Instown) decile calculated from Thomson Reuters Institutional (13f) Holdings data. Firmswith low, medium, and high institutional ownership are in Instown decile 1 to 3, 4 to 7, and 8 to 10, respectively.Control variables include the number of earnings announcements, the number of analysts following the firm, analystdispersion, market capitalization, share turnover, market return, and dummy variables for year, month, and day ofweek. Standard errors are adjusted for heteroscedasticity and clustered by the day of earnings announcement. ***, **,and * indicate statistical significance at the 1%, 5%, and 10% levels, respectively.Panel A. Firm size(1) (2) (3) (4) (5) (6)Small firms Medium firms Large firmsCAR[0,1] CAR[2,61] CAR[0,1] CAR[2,61] CAR[0,1] CAR[2,61]ES 0.739*** 0.594*** 0.950*** 0.169*** 0.825*** 0.307***(0.017) (0.049) (0.019) (0.049) (0.017) (0.044)Macroday 0.117 -1.770* -0.889*** 1.200 -0.621* 3.660***(0.285) (1.001) (0.343) (1.058) (0.320) (0.971)ES×Macroday 0.045 -0.006 0.155*** -0.254* 0.118*** -0.502***(0.039) (0.132) (0.048) (0.132) (0.044) (0.126)Constant -4.788*** -0.424 -10.904*** 5.118*** -5.273*** -2.670**(0.449) (1.544) (0.467) (1.467) (0.341) (1.214)Controls Y Y Y Y Y YObservations 43,623 43,623 45,720 45,720 69,056 69,056Adj. R2 0.091 0.016 0.138 0.007 0.097 0.007Panel B. Analyst coverage(1) (2) (3) (4) (5) (6)Low analyst coverage Medium analyst coverage High analyst coverageCAR[0,1] CAR[2,61] CAR[0,1] CAR[2,61] CAR[0,1] CAR[2,61]ES 0.764*** 0.611*** 0.936*** 0.190*** 0.852*** 0.257***(0.015) (0.044) (0.018) (0.047) (0.020) (0.057)Macroday -0.381 0.134 -0.226 -0.285 -0.564 3.272**(0.263) (0.927) (0.316) (0.917) (0.425) (1.274)ES×Macroday 0.094** -0.176 0.077* -0.107 0.114** -0.423**(0.037) (0.121) (0.043) (0.119) (0.058) (0.165)Constant -4.458*** -1.184 -6.322*** 2.557** -5.188*** 1.192(0.371) (1.146) (0.417) (1.269) (0.417) (1.640)Controls Y Y Y Y Y YObservations 54,792 54,792 53,710 53,710 49,897 49,897Adj. R2 0.093 0.012 0.120 0.008 0.090 0.01161Table 3.7: Many macroeconomic announcementsThis table presents results with many macroeconomic announcements on earnings days. The sample covers January1997 to December 2014. The dependent variable is cumulative abnormal return and is indicated under each columnheading. ES is earnings surprise quantile (11 groups), ES Top equals to 1 if earnings surprise quantile is 11 and 0 if theearnings surprise quantile is 1. High Macro News equals to 1 if that day has 7 or more macroeconomic announcements.Macroday is a dummy variable equaling 1 if that day is an announcement day for Federal Open Market Committee(FOMC) decision, Employment situation, ISM PMI, or personal consumption. Macroday High is a dummy variableequaling 1 if that day has the listed announcement and has more than 7 macro announcements at the same time.Control variables include the number of earnings announcements, the number of analysts following the firm, analystdispersion, market capitalization, share turnover, market return, and dummy variables for year, month, and day ofweek. Standard errors are adjusted for heteroscedasticity and clustered by the day of earnings announcement. ***, **,and * indicate statistical significance at the 1%, 5%, and 10% levels, respectively.Panel A. High number of macro news(1) (2) (3) (4)CAR[0,1] CAR[2,61] CAR[0,1] CAR[2,61]High Macro News -0.028 2.797** -0.581*** 1.188**(0.353) (1.417) (0.172) (0.583)ES Top 8.176*** 5.240***(0.176) (0.529)(ES Top)×(High Macro News) 1.021** -4.085***(0.426) (1.541)ES 0.841*** 0.385***(0.011) (0.029)ES×(High Macro News) 0.101*** -0.179**(0.024) (0.075)Controls Y Y Y YObservations 26,460 26,460 158,399 158,399Adj. R2 0.124 0.018 0.100 0.008Panel B. Days with important macro news & high number of macro news(1) (2) (3) (4)CAR[0,1] CAR[2,61] CAR[0,1] CAR[2,61]High Macro News -0.843 6.141** -0.865*** 2.539**(0.572) (2.495) (0.265) (0.998)ES Top 8.215*** 5.107***(0.172) (0.511)(ES Top)×(High Macro News) 1.699** -7.372***(0.677) (2.665)ES 0.847*** 0.380***(0.011) (0.027)ES×(High Macro News) 0.140*** -0.341***(0.037) (0.125)Controls Y Y Y YObservations 26,460 26,460 158,399 158,399Adj. R2 0.124 0.018 0.099 0.00862Table 3.8: Trading strategy on drift portfoliosThis table presents the results from a post-earning announcement drift trading strategy. The stock returns data is fromCRSP and is matched with firms’ characteristics from Compustat and I/B/E/S from January 1997 to December 2014.The trading strategy portfolio based on non-macro-day drift is constructed as following. In month t, it purchases firmsthat, in month t -1 made an announcement on a non-macro-day in the top quantile; sells short firms that made anannouncement on a non-macro-day in the bottom quantile. Therefore, the return for the non-macro-day drift portfoliois RDNM = R11NM−R1NM . I construct the macro-day drift portfolio for month t following a similar procedure except thatI only include firms that made an earnings announcement on a macro-day in a previous month. The return for thisportfolio is RDM = R11M −R1M . The long-short (LS) portfolio of buying the non-macro-day drift portfolio and sellingmacro-day portfolio has return, RDNM−M = RDNM −RDM . The Fama-French three-factor returns are from Ken French’swebsite. Standard errors are adjusted for heteroscedasticity and clustered by the day of earnings announcement. ***,**, and * indicate statistical significance at the 1%, 5%, and 10% levels, respectively.(1) (2) (3) (4) (5) (6)Value-weighted Equally-weightedOther days Macroday LS Other days Macroday LSConstant 0.970** 0.157 0.891** 1.150*** 0.350*** 0.804**(0.387) (0.478) (0.437) (0.397) (0.108) (0.406)Market Excess Return 0.078 -0.092 0.170 0.263 -0.027 0.290*(0.200) (0.148) (0.234) (0.166) (0.065) (0.164)SMB 0.180 -0.336** 0.517** 0.015 -0.172** 0.187(0.235) (0.142) (0.259) (0.168) (0.072) (0.180)HML 0.020 -0.211 0.231 0.059 -0.154 0.213(0.227) (0.215) (0.276) (0.206) (0.122) (0.220)Observations 179 179 179 179 179 179Adj. R2 0.011 0.025 0.018 0.002 0.019 0.01463Table 3.9: Investor attentionThis table presents the results of investor attention. The sample periods depend on the data availability of attentionmeasures. Abnormal institutional investor attention (AIA) is the news-searching and news-reading activity for Russell3000 firms from Bloomberg terminal from 2010 to 2014. AIA is a dummy variable if AIA index is higher than 2.The regression for AIA test is a probit test and the reported coefficient is marginal effects (there is no constant termreported and Pseudo R-squared is reported). Both measures are at daily frequency. Eday is dummy variable equaling1 if that has one or more earnings announcements. Google search volume index (SVI) is the ticker-searching activityfor S&P 500 firms from 2005 to 2008. Control variables include dummy variables for year, month, and day of week.Macroday is a dummy variable equaling 1 if day t is an announcement day for Federal Open Market Committee(FOMC) decision, Employment situation, ISM PMI, or personal consumption. Control variables include the numberof earnings announcements, the number of analysts following the firm, analyst dispersion, market capitalization, shareturnover, market return, and dummy variables for year, month, and day of week. Standard errors are adjusted forheteroscedasticity and clustered by the day of earnings announcement. ***, **, and * indicate statistical significanceat the 1%, 5%, and 10% levels, respectively.(1) (2) (3) (4)Attention measure AIA SVIMacroday 0.011*** 0.012*** 0.000 0.001(0.002) (0.002) (0.002) (0.002)Eday 0.522*** 0.525*** 0.098*** 0.106***(0.008) (0.008) (0.035) (0.035)Macroday×Eday 0.055*** -0.025***(0.017) (0.008)Constant 0.008*** 0.008***(0.001) (0.001)Controls Y Y Y YObservations 1,173,450 1,173,450 632,494 632,494Adj. R2/Pseudo R2 0.039 0.039 0.003 0.00364Table 3.10: Volume reactionThis table tests whether the stock volume response to earnings news is different on macro-news days. The samplecovers January 1997 to December 2014. The dependent variables are two measures of abnormal trading volume. AESis absolute earnings surprise quantile, and Macroday is a dummy variable equaling 1 if day t is an announcement dayfor Federal Open Market Committee (FOMC) decision, Employment situation, ISM PMI, or personal consumption.Following Hirshleifer, Lim, and Teoh (2009), I define abnormal trading on earnings announcement day AVOL[0] as thedifference between log dollar volume on day 0 and the average log dollar volume over days [-20,-11]. Similar definitionapplies to the abnormal trading volume on the following day AVOL[1]. AVOL[0,1] is the average of AVOL[0] andAVOL[1]. Control variables include the number of earnings announcements, the number of analysts following thefirm, analyst dispersion, market capitalization, market return, and dummy variables for year, month, and day of week.Standard errors are adjusted for heteroscedasticity and clustered by the day of earnings announcement. ***, **, and *indicate statistical significance at the 1%, 5%, and 10% levels, respectively.(1) (2)Macroday 0.050*** 0.048***(0.017) (0.016)AES 0.015*** 0.015***(0.001) (0.001)Constant 0.537*** 0.713***(0.007) (0.025)Controls No YesObservations 158,018 158,018Adj. R2 0.004 0.18265Table 3.11: Changes in riskThis table tests whether the effects of macro news on stock price response to earnings news is driven by changes inrisk. The sample covers January 1997 to December 2014. The dependent variable is cumulative abnormal return andis indicated under each column heading. ES is earnings surprise quantile (11 groups), and Macroday is a dummy vari-able equaling 1 if day t is an announcement day for Federal Open Market Committee (FOMC) decision, Employmentsituation, ISM PMI, or personal consumption. Mkt-rf, SMB, HML, and UMD are Fama-French 3 factors and mo-mentum factor, respectively. Control variables include the number of earnings announcements, the number of analystsfollowing the firm, analyst dispersion, market capitalization, share turnover, market volatility, and dummy variablesfor year, month, and day of week. Standard errors are adjusted for heteroscedasticity and clustered by the day ofearnings announcement. ***, **, and * indicate statistical significance at the 1%, 5%, and 10% levels, respectively.CAR[0,1] CAR[2,61]ES 0.846*** 0.382***(0.011) (0.029)Macroday -0.409** 0.880(0.183) (0.588)ES×Macroday 0.072*** -0.164**(0.026) (0.077)Mkt-rf×ES×Macroday 0.054* -0.153*(0.032) (0.087)SMB×ES×Macroday 0.048 0.024(0.036) (0.092)HML×ES×Macroday -0.032 -0.076(0.034) (0.087)UMD×ES×Macroday 0.058* 0.016(0.032) (0.090)Constant -5.098*** -0.504(0.244) (0.399)Controls Y YObservations 158,399 158,399Adj. R2 0.102 0.00366Table 3.12: Trading frictionsThis table tests whether the effects of macro news on stock price response to earnings news are driven by a firm’sliquidity. The sample covers January 1997 to December 2014. The dependent variables are abnormal trading volumeand bid-ask spread and are indicated under each column heading. ES is earnings surprise decile (11 groups), andMacroday is a dummy variable equaling 1 if day t is an announcement day for Federal Open Market Committee(FOMC) decision, Employment situation, ISM PMI, or personal consumption. Bid-ask is bid-ask spread and Turnoveris the firm’s trade volume divided by number of share outstanding. Control variables include the number of earningsannouncements, the number of analysts following the firm, analyst dispersion, market capitalization, market return,and dummy variables for year, month, and day of week. Standard errors are adjusted for heteroscedasticity andclustered by the day of earnings announcement. ***, **, and * indicate statistical significance at the 1%, 5%, and 10%levels, respectively.(1) (2) (3) (4)Bid-ask TurnoverES -0.182*** -0.201*** 0.002*** 0.002*(0.028) (0.041) (0.000) (0.001)Macroday -0.008*** -0.008*** 0.000*** 0.000***(0.001) (0.002) (0.000) (0.000)ES×Macroday 0.003 0.000(0.004) (0.000)Constant 1.703*** 1.706*** 0.031*** 0.031***(0.064) (0.064) (0.001) (0.001)Controls Yes Yes Yes YesObservations 127,045 127,045 158,399 158,399Adj. R2 0.105 0.105 0.006 0.00667Table 3.13: Information spillover from macro newsThis table reports the results of testing on information spillover from macro news. The sample covers January 1997 toDecember 2014. The dependent variable is cumulative abnormal return and is indicated under each column heading.ES is earnings surprise decile (11 groups), and Macroday is a dummy variable equaling 1 if day t is an announcementday for Federal Open Market Committee (FOMC) decision, Employment situation, ISM PMI, or personal consump-tion. Macro positive equals to 1 if the market return is positive on that macro-news day. Among all macro-news days,market returns are either positive or negative. One dummy variable is enough. Control variables include the numberof earnings announcements, the number of analysts following the firm, analyst dispersion, market capitalization, shareturnover, market return, and dummy variables for year, month, and day of the week. Standard errors are adjusted forheteroscedasticity and clustered by the day of earnings announcement. ***, **, and * indicate statistical significanceat the 1%, 5%, and 10% levels, respectively.CAR[0,1] CAR[2,61]ES 0.842*** 0.388***(0.011) (0.029)Macroday -0.675** 2.292**(0.267) (1.076)ES×Macroday 0.106*** -0.315**(0.037) (0.137)Macro positive 0.513 -2.481**(0.332) (1.215)ES×(Macro positive) -0.025 0.190(0.046) (0.157)Constant -4.984*** 1.128(0.226) (0.727)Controls Y YObservations 158,399 158,399Adj. R2 0.100 0.00868Table 3.14: Strategic timing of earning announcementsThis table tests whether the effects of macro news on stock price response to earnings news is driven by a firm’s strate-gic timing of earning announcements. The sample covers January 1997 to December 2014. Panel A presents resultsof testing the difference between average earnings surprise (Avg.ES) on macro-news days and Avg. ES on other days.Panel B presents regression results. The dependent variable is cumulative abnormal return and is indicated under eachcolumn heading. ES is earnings surprise quantile (11 groups), and Macroday is a dummy variable equaling 1 if day tis an announcement day for Federal Open Market Committee (FOMC) decision, Employment situation, ISM PMI, orpersonal consumption. ∆date is the difference between the day of the current earnings announcements and the previousyear’s same-quarter earnings announcement. Control variables include the number of earnings announcements, thenumber of analysts following the firm, analyst dispersion, market capitalization, share turnover, market volatility, anddummy variables for year, month, and day of week. Standard errors are adjusted for heteroscedasticity and clusteredby the day of earnings announcement. ***, **, and * indicate statistical significance at the 1%, 5%, and 10% levels,respectively.Panel A. Earnings date change and surprise∆ date<-5Count Mean SD Min MaxAvg.SUE (%) on macro days 1137 -0.105 6.682 -148.168 120.000Avg.SUE (%) on other days 7202 -0.040 4.206 -143.077 159.927Differences -0.065t-stat -0.442∆ date>5Count Mean SD Min MaxAvg.SUE (%) on macro days 2758 -0.283 3.855 -95.652 35.338Avg.SUE (%) on other days 16068 -0.526 11.084 -1077.576 195.906Differences 0.243t-stat 1.140Panel B: Earning announcement date change and the impact of macro news(1) (2) (3) (4)CAR[0,1] CAR[2,61]ES×Macroday if abs(∆ date)<=3 0.097*** -0.263***(0.029) (0.093)ES×Macroday if abs(∆ date)>3 0.072 -0.044(0.048) (0.135)ES×Macroday if abs(∆ date)<=5 0.095*** -0.263***(0.029) (0.090)ES×Macroday if abs(∆ date)>5 0.077 -0.010(0.051) (0.146)69Table 3.15: RobustnessThis table reports several robustness tests. The sample covers January 1997 to December 2014. The dependentvariable is cumulative abnormal return and is indicated under each column heading. ES is earnings surprise quantile(11 groups). Macroday is a dummy variable equaling 1 if day t is an announcement day for Federal Open MarketCommittee (FOMC) decision, Employment situation, ISM PMI, or personal consumption. Panel A reports the testexcluding firms that have strong preference to issue their earnings on macro-news days. Abnormal AnnouncementPreference (AAP) ratio for a firm is the number of earnings announcements on macro-news day divided by the totalnumber of its announcements. Panel B reports the test excluding days with a low number of earnings announcements(bottom quantile) and days with high S&P market returns (top quantile). Panel C reports the same test as in Table3.3 with CAR calculated based on Fama-French Three-Factor model and use earnings surprise deciles (10 groups).Control variables include the number of earnings announcements, the number of analysts following the firm, analystdispersion, market capitalization, share turnover, market return, and dummy variables for year, month, and day ofweek. Standard errors are adjusted for heteroscedasticity and clustered by the day of earnings announcement. ***, **,and * indicate statistical significance at the 1%, 5%, and 10% levels, respectively.Panel A. Exclude firms with strong prefernce(1) (2) (3) (4)exclude AAP-ratio>0.5 exclude AAP-ratio>0.33CAR[0,1] CAR[2,61] CAR[0,1] CAR[2,61]ES 0.843*** 0.386*** 0.844*** 0.383***(0.011) (0.029) (0.011) (0.029)Macroday -0.350* 0.759 -0.573*** 0.686(0.185) (0.592) (0.199) (0.616)ES×Macroday 0.091*** -0.194** 0.120*** -0.173**(0.026) (0.077) (0.027) (0.080)Panel B. Exclude certain days(1) (2) (3) (4)Days with low # of earnings Days with top S&P returnsCAR[0,1] CAR[2,61] CAR[0,1] CAR[2,61]ES 0.830*** 0.397*** 0.841*** 0.366***(0.012) (0.031) (0.011) (0.030)Macroday -0.324 0.529 -0.438** 1.047(0.208) (0.640) (0.199) (0.663)ES×Macroday 0.095*** -0.178** 0.091*** -0.207**(0.028) (0.083) (0.028) (0.084)Panel C. Alternative measures(1) (2) (3) (4)Fama-French CAR ES 10 DecilesCAR[0,1] CAR[2,61] CAR[0,1] CAR[2,61]ES 0.843*** 0.386*** 0.844*** 0.383***(0.011) (0.029) (0.011) (0.029)Macroday -0.350* 0.759 -0.573*** 0.686(0.185) (0.592) (0.199) (0.616)ES×Macroday 0.091*** -0.194** 0.120*** -0.173**(0.026) (0.077) (0.027) (0.080)70Chapter 4Media Attention, MacroeconomicFundamentals, and the Stock Market4.1 IntroductionClassical theories of asset pricing, based on exogenous information flows and efficient market pricing (e.g.,Merton, 1973), provide no explicit role for investor attention. A growing literature establishes however thatinvestor attention plays an important role in financial markets. For example, Da et al. (2011) show thatinvestor attention to individual stocks positively predicts subsequent short-run returns for those stocks.37Focusing on the aggregate stock market, Andrei and Hasler (2014) develop theoretical and empirical linksbetween attention and conditional moments of returns, in particular the equity risk premium and volatility.38While these findings are broad, they do not address attention to systematic risk factors other than theaggregate market. In this paper, we study attention to macroeconomic fundamentals. We propose newmeasures of attention to separate categories of macroeconomic fundamentals such as unemployment, outputgrowth, and policy announcements of the Federal Open Market Committee (FOMC).We focus on attention to macroeconomic fundamentals for several reasons. First, the finance literaturehas long sought to connect asset prices to underlying macroeconomic factors (Chen et al., 1986). Second,scheduled macroeconomic announcements have strong effects on asset prices (e.g., Andersen et al., 2003;Savor and Wilson, 2013),39 and such announcements should also impact attention. Third, while the assetpricing literature often tends towards stock-market based factors in describing the cross-section of returns(e.g., Fama and French, 1993), casual observation of news media coverage suggests that attention to system-atic risks is more frequently framed in terms of macroeconomic factors such as unemployment and inflationas opposed to stock-market based factors like size and value.Focusing on attention to macroeconomic fundamentals also allows us to shed light on theory. First, atheoretical prediction of Andrei and Hasler (2016) is that attention should increase whenever fundamentalsdeviate from their expectations, in either direction. Their theory also implies that attention changes should beasymmetric, increasing more when fundamentals move towards a “bad” state than toward a “good” state. Werelate each of our macroeconomic attention indices to a relevant series of macroeconomic fundamentals, andconfirm these predictions. Second, Ai and Bansal (2016) show that the risk premia associated with scheduled37For further evidence regarding attention to individual stocks, see Huberman and Regev (2001); Barber and Odean (2008);DellaVigna and Pollet (2009); Hirshleifer et al. (2009), and Hillert et al. (2014).38Kacperczyk et al. (2016) study interactions between firm-level and aggregate attention. Additional theoretical studies of atten-tion include Sims (2003), Peng and Xiong (2006), and Andrei and Hasler (2016).39See also Andersen et al. (2007) and Brusa et al. (2016).71macroeconomic announcements should be driven by the resolution of uncertainty. Although their modeldoes not explicitly consider attention, it is natural to interpret attention prior to a scheduled announcementas costly information acquisition. Following this interpretation, attention can provide a useful instrumentfor the risk premium (and anticipated resolution of uncertainty) associated with an announcement. We showthat increasing attention prior to key macroeconomic announcements predicts higher risk premia and greaterresolution of uncertainty on those announcement dates.Our measures of attention are based on media coverage of different types of fundamental news. The cat-egories we consider are unemployment, output growth, inflation, credit ratings, the housing market, interestrates, monetary policy, oil, and the U.S. dollar. We create lists of search words that capture attention to eachof these fundamentals. For example, to capture attention to U.S. output growth, we use the following setof words: “gross domestic product”, “GDP”, “gross national product”, “GNP”, and “industrial production”.We count the number of articles in the Wall Street Journal (WSJ) and New York Times (NYT) that includeany of these search terms, starting in June 1980 for NYT and January 1984 for WSJ, and ending in Decem-ber 2016. Scaling by the total number of articles published gives a measure of attention to each category ofmacroeconomic fundamental.Our indices most directly measure media attention, but the media clearly has strong incentives to coverissues of interest to their readers (Mullainathan and Shleifer, 2005), and prior literature often uses mediaattention as a proxy for investor attention (e.g., Barber and Odean, 2008;Yuan, 2015). A separate line ofresearch, which we do not contribute to, investigates the causal role of media attention (e.g., Tetlock (2007,2010); Engelberg and Parsons (2011); Peress (2014)). We view media coverage as a useful proxy for investorattention because of the long time-series it permits, beginning in 1980. More direct measures of investorattention, such as Google search (e.g., Da et al., 2011) have other advantages but provide shorter time series.In the remainder of the paper, we do not distinguish between media and investor attention. We provideseparate measures of attention for the NYT and WSJ, as well as combined measures, which give insightsinto the commonalities and differences in attention across the readerships of these outlets.Our macroeconomic attention indices (“MAI”) show interesting empirical properties. We first addresscomovement in attention and show that the indices are not driven by a single factor. They are imperfectlycorrelated with one another and with prior measures such as the Economic Policy Uncertainty (EPU) indexof Baker et al. (2016). Over time attention shifts across inflation, employment, monetary policy, and theother fundamentals. If these shifts in attention reflect changes in investor concerns, then only in very specialcases could efforts to price assets reduce to a single factor representation of risk.We next address the duration of cycles in attention. For the macroeconomic fundamentals we consider,the attention indices are stationary but persistent. The conservative Bayesian Information Criterion suggestsat most four lags in a monthly autoregression framework. However, when we aggregate the attention indicesover different window lengths, similar to the MIDAS framework of Ghysels et al. (2006), we find that mostof the series show evidence of cycles at multiple frequencies, ranging from one day to as long as one year.These “long-memory” characteristics of attention are properties also observed in aggregate stock marketvolume and volatility in prior literature.4040See, for example, Andersen et al. (2001) and Bollerslev and Mikkelsen (1996).72We next seek to relate attention to movements in economic fundamentals. We associate each of theattention indices with a related macroeconomic variable, and, where possible, at least one scheduled an-nouncement. As expected, high-frequency variations in attention relate to scheduled news announcements,and we document which announcements have the most impact on attention. Lower frequency movementsin attention relate to movements in economic fundamentals. We decompose each of the economic series(e.g., unemployment, inflation) into simple moving averages over different window sizes. Attention relatesto variations and squared variations in simple moving averages of fundamentals over different horizons. Allsignificant squared terms on variations are positive, consistent with the idea that changes in fundamentalslead to increased attention. Attention also increases when fundamentals move in more “negative” direc-tions (e.g., attention to unemployment increases when unemployment increases, and attention to housingincreases when house prices decrease.)In some cases, the relation between attention and fundamentals is very strong. For example, over 50%of the variation in the unemployment attention index is explained by unemployment fundamentals, and thecomovement is strong enough to be apparent in a simple plot (see Figure 4.1). We also document differencesbetween the WSJ and NYT in the strength of the relation between their attention indices and fundamentals.We show that attention to macroeconomic news relates to the stock market in two ways. First, financialeconomists have long sought to understand the drivers of time-varying volatility Schwert (1989), and aleading explanation is time variation in the rate of information flow (e.g., Andersen, 1996). Followingthis explanation, we might expect attention to act as a proxy for information flow, and therefore positivelyrelate to both volume and volatility. We test this hypothesis and show that controlling for macroeconomicannouncements, increasing attention to macroeconomic news correlates positively with changes in bothaggregate volume and aggregate volatility.Second, changes in attention help to predict future macroeconomic news, as well as returns and volatilitychanges on future scheduled macroeconomic announcement dates. Following prior evidence (Savor andWilson, 2013; Ai and Bansal, 2016) that a significant proportion of the market risk premium is earned onannouncement dates of employment reports and FOMC decisions, we focus on these announcements.Boyd et al. (2005) document that the impact of employment announcements on stock prices dependson the business cycle. In particular, positive unemployment surprises are usually associated with highermarket returns on the announcement date, but this is reversed during recessions. We add to their results inseveral dimensions. First, increasing attention to employment news before the unemployment announcementpositively predicts the announcement “surprise”, where we measure surprise in a variety of ways, includinga random walk model, the statistical model proposed by Boyd et al. (2005), and Bloomberg consensusforecasts. Second, increasing attention before the announcement also positively predicts market returns onthe announcement date, controlling for the employment surprise. This is consistent with the idea that higherrisk premia are earned on employment announcement days, and that attention provides a useful instrumentfor the magnitude of the risk premium. To provide further support for this interpretation, we show that thefall in implied volatility on the announcement date, commonly attributed to the resolution of uncertainty, islarger when attention before the announcement has increased more.We show similar findings on FOMC announcement dates. Stock returns are larger on FOMC announce-73ment days when attention to monetary policy before the announcement has risen more, and higher pre-announcement attention also predicts a greater reduction in implied volatility on the FOMC announcementdate, especially in the later part of the sample. Our results are also robust to announcement surprises and theEPU index of Baker et al. (2016).Altogether these findings reinforce the idea that macroeconomic news impacts financial markets, anddemonstrates the role of attention to macroeconomic news. The contents and dates of macroeconomicannouncements are certainly important, as explored in the prior literature. The attention that investors payto different categories of macroeconomic fundamentals provides additional market-relevant information andsheds light on theoretical explanations of how investors allocate attention.414.2 Macroeconomic Attention IndicesWe create indices of news-media attention to the following macroeconomic risks: output growth, inflation,employment, interest rates, monetary policy, housing, credit conditions, oil, and the U.S. dollar. For eachfundamental, we create a list of related words and phrases, shown in Table 4.1. We aim for the lists to beobjectively reasonable.We search articles in the Wall Street Journal (WSJ) and New York Times (NYT). These publicationscover general news, economic news, and financial news, and have been used in numerous prior studies.We use two different publications to provide a sense of the robustness, and also to illuminate differences inattention across outlets with different audiences. WSJ is generally regarded as having a tighter focus on theeconomy and financial markets as well as a more conservative editorial slant, while NYT provides broadercoverage of general news and has a more politically liberal reputation.42 For the NYT, the sample periodis from June 1, 1980 to December 31, 2016. For the WSJ, the sample period is from January 1, 1984 toDecember 31, 2016. During these sample periods broad digital coverage of the publications is available.We consider only the newspaper print editions. Table 4.2 presents MAI and reports the data sources forassociated fundamentals to each MAI.4.2.1 Construction of the Attention IndicesEach day in the sample period, we count the number of articles in each publication that satisfy the searchcriteria for each macro fundamental. This provides 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 normalize thesecounts by dividing by the average number of articles per day Nˆp,t for publication p during the calendar monthincluding observation t.43 The“unadjusted\" macroeconomic attention index for each individual publicationP is:MAI-p f ,t =Np, f ,tNˆp,t. (4.1)41In other applications, Liu and Matthies (2017) use media attention as a conditioning variable for cross-sectional asset pricing.42The differences in media slant and its economic impact are well-documented in the literature (see e.g., DellaVigna and Kaplan,2007; Gentzkow and Shapiro, 2010).43We have the total number of articles in each publication per month, which we divide by the number of business days in themonth.74The unadjusted attention indices measure the percentage of articles on a given day that has content relatedto the macroeconomic fundamental of interest.We define related measures that are demeaned, or alternatively demeaned and standardized. Let µp, fand σp, f denote respectively the time-series means and standard deviations of the daily unadjusted attentionindices MAI-p f ,t . The demeaned measures are denotedMAI-pd f ,t = MAI-p f ,t −µp, f ,and the standardized measures are denotedMAI-ps f ,t = MAI-pd f ,t/σp, f .We also define two composite indexes of attention. The first composite index, denoted MAI-C1, is anaverage of the demeaned NYT and WSJ indices in time periods when both are available, and the NYT indexonly in the 1980-1983 period:MAI-C1 f ,t ={(MAI-WSJd f ,t +MAI-NYTd f ,t)/2 from Jan. 1, 1984, to Dec. 31, 2016,MAI-NYTd f ,t from June 1, 1980, to Dec. 31, 1983.(4.2)Demeaning the individual publication indices before averaging ensures that we will not induce a level effectdriven simply by the change in composition that occurs in 1984 when the WSJ data becomes available.The second composite index denoted MAI-C2, is an average of the standardized NYT and WSJ indiceswhen both are available:MAI-C2 f ,t ={(MAI-WSJs f ,t +MAI-NYTs f ,t)/2 from Jan. 1, 1984, to Dec. 31, 2016,MAI-NYTs f ,t from June 1, 1980, to Dec. 31, 1983.(4.3)Standardizing ensures that both publications contribute equally to the variation of MAI-C2.44 While theweighting of the two composite indices is different, neither is superior in any sense. The publication withmore variation in its own attention index will be weighted more heavily in MAI-C1 relative to MAI-C2. Ifone believes that greater variation in attention over time reflects more information, then the weighting ofMAI-C1 may be preferred to MAI-C2. In this paper, we report in the regression tables our results usingMAI-C1. In the Internet Appendix, we show the results using MAI-C2.All of the indices build on simple counts of the number of articles related to a macroeconomic fun-damental, as a proportion of all articles. Many elaborations of this approach are possible, for example,weighting articles by their number of words, or attempting to measure the intensity of relevance rather thana simple binary coding. We take a basic approach for simplicity and expect other measurement methods tobe 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 negative sentiment.44We drop weekend observations when calculating the demeaned and demeaned and standardized MAI.754.2.2 Empirical Properties of the Attention IndicesTable 4.3, Panel A provides summary statistics for the unadjusted daily attention indices for both NYT andWSJ. NYT coverage of macroeconomic fundamentals is uniformly lower as a proportion of all coveragethan WSJ. The NYT index means have the lowest value of 0.08% for U.S. dollar coverage, and the highestindex means are inflation and GDP (0.89%), unemployment (0.80%), and monetary (0.92%). For the WSJ,the index averages range from a low of about 0.5% of articles for credit rating to a high of over 2% forinflation, GDP, and oil. Consistent with the higher mean attention levels in the WSJ, the standard deviationof attention is also uniformly higher for the WSJ than the NYT. This implies that the weight of the WSJ inthe composite indices MAI-C1 will be higher than in the composite indices MAI-C2.Table 4.3, Panel A also provides index means by day of the week. For NYT, the Saturday edition appearsto have roughly similar content to other days, while the large Sunday edition offers more coverage than otherdays. The Saturday edition of WSJ generally has less coverage of macro fundamentals than other days of theweek. While the effects of weekend news coverage are interesting and potentially important, for simplicityin the remainder of our analysis, we discard all non-trading days (weekends and holidays). To account forpotential day-of-week seasonalities in news coverage, all of our empirical results use day-of-week dummyvariables.Figure 4.2 plots the attention indices.45 For reference, each attention index is associated with a seriesof macroeconomic fundamentals that seems relevant.46 For example, the output growth attention index isplotted on the same axes with the log quarter-to-quarter growth in real GDP. The full list of attention indicesversus the associated macroeconomic fundamentals plotted in Figure 4.2 is given in Table 4.2.We emphasize several properties of the attention indices. First, the indices do not appear to be driven bya single factor. They are imperfectly correlated, and over time attention shifts across different fundamentals.Second, attention is highly persistent. All series show fluctuations that last over periods at least as longas several years, including both gradual trends and sharp changes. Third, the indices also show cycles at arange of higher frequencies, including short bursts of attention. Finally, attention seems to be at least looselyrelated to underlying fundamentals. This is seen most clearly in the plot for employment (see Figure 4.1),where broad patterns in attention seem to match closely with the level of the unemployment rate. We nowinvestigate each of these aspects of these plots using statistical analyses.Table 4.3 shows daily (Panel B) and monthly (Panel C) correlations among the composite attentionindices MAI-C1, as well as correlations with other series of interest: implied volatility (VIX) from theChicago Board Options Exchange47, Economic Policy Uncertainty (EPU) from Baker et al. (2016)48, anddetrended S&P 500 trade volume (Volume) from the Center for Research in Security Prices. The resultsconfirm the imperfect correlation of the attention indices. In daily data, the highest inter-MAI correlationsare between monetary and inflation (0.45), monetary and interest rates (0.56), oil and inflation and oil andmonetary (0.32), U.S. dollar and oil (0.38), and inflation and interest rates (0.34). Not all correlations are45We show plots for all nine MAI in the Internet Appendix.46The approach follows Carroll (2003), who plots a monthly news count index of inflation from the New York Times and theWashington Post against CPI, from 1981 to 2001.47Data source: https://www.cboe.com/micro/vix/historical.aspx.48Data source: http://www.policyuncertainty.com/.76positive. For example, in monthly data, the MAI for GDP and inflation are negatively correlated (-0.06)and credit rating and inflation (-0.17). We also are interested in correlations between the attention indicesand other variables. In the monthly data, the highest correlations with EPU are unemployment (0.51), creditrating (0.39), GDP (0.33), and monetary (0.30). The highest correlations with VIX are credit rating (0.41)and unemployment (0.47).To address stationarity, we estimate AR(p) models for each attention index from monthly data. Fol-lowing Campbell and Yogo (2006), we use the lag length that minimized the Bayesian information criteria(BIC). The minimum BIC for all of our MAI occurs at four lags or less. Table 4.5 shows these AR estimates,controlling for monthly fixed-effects. The table also reports Dickey-Fuller p-values for the null hypothesisthat each series has a unit root. The DF statistics reject the presence of unit roots except for the U.S. dollarMAI.49To further explore time-series dependence, Figure 4.4 shows autocorrelation plots of each compositeseries MAI-C1 for lag lengths from 1 to 250 trading days. We plot the autocorrelations for residuals aftercontrolling for day-of-the-week dummies and month-of-the-year dummies. These plots show very slowdecay in this range of frequencies, and the autocorrelations are significantly larger than zero at 250 lags forall series. Several of the autocorrelation plots show apparent cycles in dependence. For example, GDP showsstrong increases in correlations at each monthly interval. Other series (housing, U.S. dollar) have increasesin autocorrelations at weekly intervals. These cycles are consistent with the importance of periodic newsannouncements.To account for potential long-memory dependence as well as multiple cycles in news variation, we useregressions that aggregate attention indices over different horizons similarly to MIDAS regression (Ghyselset al., 2006, see). Specifically, we construct simple moving averages of attention indices over window sizesof one day, five days, 21 days (monthly), 62 days (quarterly), and 250 days (annual), and 1000 days (businesscycle).Panel B of Table 4.5 shows results of regressing each attention index on lagged simple moving averagesof its own history, for the full set of different window sizes. All of the series show persistence at multiplefrequencies, with the majority having significant positive persistence in daily, weekly, monthly, quarterly,and annual-length moving averages in the multiple regression frameworks. One exception is credit ratingattention, which does not show significant persistence beyond monthly horizons. A separate monthly cycleis not present in GDP attention, although it does show significant persistence at all other cycle lengthsbetween daily and annual. This result seems intuitive given the quarterly reporting cycle for GDP growth.These results are consistent with slow, approximately hyperbolic decay in the persistence of attention toeach of the fundamental factors. The presence of multiple frequencies in attention to the financial newsis also broadly consistent with the motivation and theoretical framework in Calvet and Fisher (2007), whohypothesize fractal patterns in the news about fundamentals impacting asset prices. We next determinewhether fluctuations of the individual attention indices can be related to macroeconomic fundamentals.49The U.S. dollar MAI-C2 rejects the unit root with a p-value of 0.07.774.3 Attention and Macroeconomic FundamentalsIntuition suggests that high-frequency fluctuations in attention could be driven by economic announcements,while lower frequency variations might be related to movements in economic fundamentals. We test theseideas.4.3.1 Macroeconomic AnnouncementsPrior literature has established links between economic announcements and returns and volatility for theforeign exchange and stock market (Andersen et al., 2003, 2007). We now investigate the relationshipbetween macroeconomic announcements and attention to macroeconomic fundamentals. Attention couldbe limited to simply reporting on announcements. Alternatively, attention might be high in advance ofannouncements as news media strive to anticipate the content of announcements or to put the potentialoutcomes of an announcement into a broader context for the benefit of their readers.Cross-sectionally, our analysis can tell us which types of announcements have the largest impacts onmacroeconomic attention. If the media play an important role in the transmission of economic news, thenunderstanding the allocation of media resources to covering different types of announcements should beinformative about which announcement matters most to readers.The economic announcements we consider are the Consumer Price Index (CPI), Employment Situation,the Federal Open Market Committee (FOMC), and the quarterly GDP report. The announcement datesspan the entire sample length of our indices. The CPI and Employment Situation announcement dates arefrom the Bureau of Labor Statistics website, FOMC announcement dates are from the Federal ReserveBoard website, and the GDP report dates are from the Bureau of Economic Analysis. As in Savor andWilson (2013), we assume that FOMC decisions before 1994 became public one day after the meeting.Macroeconomic attention can be influenced by multiple announcements, hence we study the most intuitivelinks between the macroeconomic attention indices and macroeconomic announcements as shown in Table4.2. The specification we use is:MAI-C1d f ,t = α+δ=4∑δ=−4βδAnn j,t+δ + εt (4.4)where MAI-C1d f ,t is the composite index MAI-C1 detrended by its own 60-day simple moving average.The variables Ann j,t+δ are equal to 1 if there is an announcement on day-t + δ , 0 otherwise, and we let δtake integer values from -4 to 4. We show the regression coefficients, βδ , and their 95% confidence intervalsin Figure 4.5.Two main findings stand out from Figure 4.5. First, our measures of attention captures what they aresuppose to measure, that is, attention to fundamentals. The figure shows that attention to inflation, unem-ployment, monetary, and GDP rises before their respective announcements, spikes on the day following theannouncement, and followed by gradual decay.Second, our measures of attention show which type of announcements have the most impact on attention.For example, only after 1994 do we observe significant changes in attention to monetary policy around78FOMC announcements. Before 1994, the FOMC members did not announce publicly their monetary policydecisions (e.g., changes to the Fed fund rate). Also, we observe that attention to GDP rises only beforethe first GDP report, i.e., GDP-advance. It is also the first GDP report that impacts attention the most onannouncement day and on the following day. This finding support the finding of Gilbert et al. (2017) whoshow that among GDP report announcements, only the first GDP announcement has significant impact onasset prices.4.3.2 Macroeconomic FundamentalsBeyond the link between economic announcements and daily spikes in attention, what accounts for thelower-frequency fluctuations in attention indices? Figure 4.1 and 4.2 suggests attention dynamics couldreflect changing economic conditions.Prior literature has attempted to establish links between macroeconomic variables and financial marketvariables such as volatility (Schwert, 1989). We expect that macroeconomic attention connects economicnews with financial markets, serving an intermediary function. A benefit of measuring macroeconomicattention is that we can measure not just aggregate interest in financial and economic news, we can alsotell what writers are talking about. Hence the low-frequency variations in our different MAI should pick upchanging patterns in concerns for different macroeconomic fundamentals.To study how variations in macroeconomic fundamentals impact macroeconomic attention, we decom-pose the macro variables into detrended moving averages over different window sizes as in Ortu et al.(2013).50 That is, given a particular macroeconomic fundamental Ft (e.g., unemployment rate, change inlog CPI, change in log house price index), we can decompose the fundamental into a set of detrended movingaverages: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 to t. The components on the right-hand side of the equation, each in parentheses, are detrended moving averages over window sizes that areexpanding approximately geometrically. These could be capable of capturing the low-frequency patterns inautocorrelations documented for the attention indices in Table 4.5. We regress the monthly attention indiceson these detrended moving averages, and their squared values:MAI f ,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+ εt , (4.6)Table 4.6 reports results for this regression for the macroeconomic attention indices for the New YorkTimes (MAI-NYT) in Panel A and the Wall Street Journal (MAI-WSJ) in Panel B. The results show thatattention responds to changes in macro fundamentals. Adjusted R2 range from 0 to over 50%, with most of50Ortu et al. (2013) use a similar decomposition to document the existence of persistent components in consumption growth.79the regressions having at least one significant coefficient on fundamentals.To help synthesize the results, we first focus on aspects that are similar to Panels A and B, or acrossattention in both MAI-NYT and MAI-WSJ. Confirming the idea that change raises attention, many of thecoefficients on squared changes in fundamentals are significant and positive in both panels. For the MAI-NYT, of the 16 significant coefficients on squared changes in fundamentals, 14 are positive. For the MAI-WSJ, all 15 of the 15 squared changes on fundamentals are positive. These results are consistent withtheories where changes in fundamentals raise attention, such as in Andrei and Hasler (2014, 2016).A second intuitive idea is that for a given magnitude of the absolute change, attention will be higherwhen the change is in a direction that is associated with “bad” versus “good” times. Focusing on thesignificant coefficients on signed changes in fundamentals, many of the series show consistent results acrossthe MAI-NYT and MAI-WSJ in the intuitive direction suggesting that bad news raises attention: Attentionto credit rises when relative credit spreads rise; attention to housing rises when house prices fall; attentionto unemployment rises when unemployment increases.We also see interesting differences across MAI-NYT and MAI-WSJ. In general, the R2 for the WSJattention index regressions on fundamentals are higher than for the NYT. One notable exception is unem-ployment. More than 50% of the variation of the NYT attention index is explained by movements in theunemployment rate, consistent with the very strong comovement apparent in Figure 1, compared to the lowerR2 of 34% for explaining WSJ attention to unemployment. Why do unemployment fundamentals have lessexplanatory power for WSJ attention than for NYT attention? Examining the plots in Figure 1, the NYThas shown a consistently positive relation between unemployment and attention to unemployment. For theWSJ, in the 1980’s and 1990’s, attention moved almost inversely with the unemployment level. Starting inthe 2000’s and certainly, by the financial crisis, WSJ coverage of unemployment began to comove positivelywith changes in unemployment, similar to the NYT. This is consistent with the idea that the readershipand editorial policy of the NYT have been more consistently focused on unemployment than the WSJ overtime; however, following the financial crisis, the WSJ became more attentive to unemployment in a mannersimilar to NYT.51Consistent with this idea of different focuses and audiences between the NYT and WSJ, we also see adifference in how inflation impacts attention. An increase in inflation tends to raise attention to inflation atthe WSJ but reduces attention at the NYT. This is again consistent with the idea that the WSJ tends to bemore politically conservative and associated with monetarist views on inflation than the NYT, which tendstowards more Keynesian views on the economy.4.4 Attention and Stock Market ActivityBeber et al. (2011) conjecture that market participants are continually digesting news about the macroecon-omy, which impacts their preferences, expectations, and risk tolerances. As a result, macroeconomic newsinduces them to trade. The authors show that market trade volume segmented by economic sectors contain51Another contributing factor could be the retirement of conservative editor Robert Bartley, who retired from the WSJ in 2000after serving for thirty years.80important macroeconomic information and in turn predict important macroeconomic announcements.We study the link between daily macroeconomic attention and stock market activity. Let V lmdt be thelogarithm of the daily aggregate trade volume of S&P 500 firms, detrended by its 60-day moving average,following Tetlock (2007). We run the regression:V lmdt = α f +β f MAI5−20, f ,t + γ f Annt +δ f Annt ·MAI5−20, f ,t + ε f ,t , (4.7)where MAI5−20,t, f is the difference between the 5-day and 20-day moving average of MAI-C1 to macrofundamental f . Annt is equal to one if there is an announcement on day t, zero otherwise.52Table 4.7 shows that for all MAI, rising attention is associated with an increase in market volume. Whenwe include macro announcements in the regressions, many of the announcements have significant impactson volume, but the inclusion of these variables does not alter inferences about the importance of attention.Interaction terms do not have a consistent sign and do not alter inference about the effects of attention orannouncements on trading volume.Another way to look at the impact of macroeconomic attention on stock market activity is to investi-gate the relationship between macroeconomic attention and implied volatility, measured by the VIX index,starting from 1990. We implement the following regression for each attention index:V IXt = α f +β f MAI20−250, f ,t + γ f Annt +δ f Annt ·MAI20−250, f ,t + ε f ,t , (4.8)where MAI20−250, f ,t is the difference between the 20-day and 250-day moving average of MAI-C1 to macrofundamental f . Table 4.8 shows that increases in macroeconomic attention on monetary, GDP, unemploy-ment, credit ratings and U.S. dollar positively relate to increases in implied volatility. The R2 are highestfor unemployment (17%) and GDP (6%). Results are similar if we detrend VIX using a 250-day movingaverage. Thus, controlling for macroeconomic announcements, increases in attention is associated with anincrease in both aggregate volume and volatility.Overall the results of this section provide strong evidence that increases in attention to macro fundamen-tals is positively correlated with the aggregate stock market activities.4.5 Using Attention for ForecastingGiven the links between media attention and macroeconomic fundamentals, it is natural to consider whethermedia attention might help to predict fundamentals on macroeconomic announcement days. Moreover, ourmeasures of attention might also predict risk premium and resolution of uncertainty on these days. Investi-gating the relationship between attention and resolution of uncertainty will shed light on recent theory as towhether the amount of resolution of uncertainty on macroeconomic days is related to the level of uncertaintyand investor attention before announcements. As suggested in Andrei and Hasler (2016), attention rises withuncertainty to fundamentals.52To simplify the analysis, we do not differentiate between all GDP announcements (advance, preliminary, and final).81We are particularly interested to understand the link between the MAI to unemployment and the Employ-ment Situation announcements and the MAI to monetary policy and FOMC announcements. Our decision tofocus on unemployment is partly motivated by the plots in Figure 4.1 which suggest that the unemploymentattention indices might act as a leading indicator and partly motivated by findings in the prior literature thatthe unemployment report is important for stock market returns (Boyd et al., 2005).We also focus on FOMC announcements because recent literature shows that risk premium is largelyearned on FOMC announcement days (Lucca and Moench, 2015; Ai and Bansal, 2016). It is natural, there-fore, to consider the link between attention to monetary policy and risk premium and resolution of uncer-tainty on FOMC announcement days. The Employment Situation and FOMC announcements are, if not,the two most important macroeconomic announcements. As shown in Ai and Bansal (2016), since 1997,virtually all risk premium is earned on days with Employment Situation and FOMC announcements.4.5.1 Employment situation announcementsWe construct measures of “surprises” in the monthly employment report in three ways. First, we con-sider a simple random walk model of unemployment, under which the prediction for the following month’sunemployment rate is the prior month’s unemployment rate, and the surprise is defined as the change inunemployment. Second, we calculate a measure of surprise by taking the difference between the actualunemployment rate and the expected rate from Bloomberg’s surveys starting from 1997. Third, we use theregression model of Boyd et al. (2005) to generate the unemployment forecasts, which we call the “BHJsurprise”. The authors’ forecasting model uses information from related macroeconomic variables, includ-ing industrial production, Treasury-bill rate, corporate bond yield spreads, and past unemployment rate. Thesurprise is defined as the difference between the announced unemployment rate and the unemployment fore-cast. The date of reference for the actual unemployment rate is the release date of the Employment Situationannouncement released by the U.S. Bureau of Labor Statistics.For predictor variables, we carry out separate analyses using detrended levels of the composite indicesMAIC. Specifically, to capture very short run movements, we use the difference between the 5-day simplemoving average and the 20-day simple moving average of the attention indices (MAI5−20). To capture arange of other movements, we similarly calculate 20- and 60-day moving averages detrended by the 250-day moving average (MAI20−250, MAI60−250). Following Boyd et al. (2005), we also interact each of thepredictor variables with NBER recession dummies. Since the NBER dummies are not known in advance,regressions using these interactions are not predictive. Boyd et al. (2005) hypothesize that “bad news” forunemployment means different things in expansions and contractions, and the interaction variables allow usto see whether the predictive ability of attention if it exists, concentrates in contractions.To investigate the link between unemployment surprises and our attention index to unemployment, weestimate the following regression:Surpriset = c+β1MAIt−1+β2NBER ·MAIt−1+ et , (4.9)where Surpriset is the unemployment announcement surprise, MAIt−1 is the detrended MAI-C1 for unem-82ployment, and NBER is an indicator variable for NBER recession.Table 4.9 shows that the detrended unemployment attention variables are significantly related to sur-prises in the unemployment report and that the interaction variables are often important. Under the randomwalk model, Panel A shows that attention indices positively predict future surprises in unemployment, andvariables are statistically significant when interacted with the NBER recession dummies. Hence, increasesin macroeconomic attention to unemployment positively predict future changes in unemployment, and thisrelationship is strong during recessions. Panel B confirms that changes in macroeconomic attention retainthe ability to explain future changes in employment relative to the Boyd et al. (2005) regression model.Figure 4.6 provides visual evidence of how attention dynamics relate to the employment surprises. Thereare six panels, corresponding to all combinations of the main three unemployment surprises, and the twounemployment attention indices for the New York Times (MAI-NYT) and the Wall Street Journal (MAI-WSJ). For each unemployment surprise, we separate the data into three equal-sized bins of small, medium,and large surprises. We then plot in event time the 60-day moving average in attention one year beforethe surprise out to one year after the surprise. The results show strong and very similar patterns. Whenthe unemployment surprise is particularly low, on average attention to unemployment in the media hasbeen declining over the past year, and continues to decline over the following year. Conversely, when theunemployment surprise is large and positive, on average attention has been increasing over the prior year,and continues to increase over the following year. When the unemployment surprise is in the middle tercile,on average attention is approximately flat over the prior and following years, and at a lower level than forlarge positive or negative surprises. These findings are consistent with the regression results, and confirmthat attention moves both before and after changes in reported fundamentals.We next investigate the relationship between risk premium on Employment Situation announcementdays and attention to unemployment using a similar regression framework as in Boyd et al. (2005). Weuse the authors’ measure of unemployment surprise defined previously and adding measures of attention tounemployment. We also control for the Economic Policy Uncertainty (EPU) measure by Baker et al. (2016)because other papers have use EPU as a driver to risk premium (Brogaard and Detzel, 2015). We specify:Rett = c+β1MAIt−1+β2NBER ·MAIt−1+β3BHJt +β4NBER ·BHJt (4.10)+β3EPU t−1+β4NBER ·EPUt−1+ et ,where Returnt is the daily log return of the S&P 500 index on the Employment Situation announcementday, and MAIt−1 and EPUt−1 are the one-day lag detrended measure of attention to unemployment andEPU, respectively.53 We detrend using the 5-day moving average minus the 20-day moving average. Wefocus specifically on this detrending variable because Figure 4.5 that attention to unemployment rises pri-marily five days before the announcement and there are approximately 20 trading days between EmploymentSituation announcements.5453When the Employment Situation announcement occurs on Good Friday (U.S. holiday) we use the stock return on the followingtrading.54In Figure C.3 of the appendix, we show that uncertainty, using the VIX, also increases five days before Employment Situation83We present the results of the above regression in the first three columns of Table 4.10.55 Our measureof attention to unemployment positively predicts returns on Employment Situation announcement days andis statistically significant at the 5% and 10% level and the predictability of attention raises during NBERannouncements.56 The EPU is also statistically significant but predicts a negative risk premium on an-nouncement days. The BHJ surprises are not statistically significant, however, the interaction BHJ×NBERis negative and statistically significant at the 1% level, consistent with the findings of Boyd et al. (2005).It is important to distinguish between the trend in attention, which reflects anticipation, and the surpriseitself, which reflects a realization. Consistent with the results of Boyd et al. (2005), during a recession ahigher realization of unemployment on the announcement date leads to lower market returns. We add to thisthat rising attention before the announcement date tends to be associated with higher market returns on theannouncement date, as uncertainty is resolved (Ai and Bansal, 2016).To further elaborate on the interpretation of these regressions, we note that since attention predicts theunemployment announcement itself, we could find that attention predicts market returns on the announce-ment date through two separate channels. First, there is the possibility of a risk premium channel, throughwhich high attention predicts high expected market returns on the announcement date. This channel is es-sentially an elaboration of the mechanism proposed by Savor and Wilson (2013) and Ai and Bansal (2016)and others who have recently studied the importance of macroeconomic announcements on asset prices. Theelaboration is that attention before an individual employment announcement provides a potentially usefulinstrument for its importance and in turn the level of the risk premium for that individual announcement.The second channel through which pre-announcement attention could impact announcement date returnsis through unexpected market returns, given that attention helps to predict the employment surprise, andsupposing that market participants are unaware of this predictability so that market prices react fully to theannouncement surprise on the announcement date.Since the significance of attention remains in our regressions when controlling for the announcementsurprise itself, the return predictability from attention is more likely to be working through the risk pre-mium channel. To further confirming this interpretation, Table 4.10, Column 4 to Column 6, reports OLSregression results from using the change in VIX (∆V IX) as dependent variable to the model specificationsof Equation (4.10). The decrease in VIX on the announcement date is larger when attention is higher beforethe announcement. These VIX regressions also provide further support to the Boyd et al. (2005) interpreta-tion that the impact of employment announcements depends on the market state. During normal times, anincrease in unemployment generates a statistically insignificant drop in the VIX. During recessions, positiveunemployment surprises cause strong and statistically significant increases in implied volatility.Altogether, these results confirm that, as documented in the prior literature, employment announcementscontain important information for financial markets. Additionally, we learn that changes in attention beforean announcement provide useful information about both the contents and the importance of individual em-ployment announcements. Increasing attention before the employment announcement positively predicts theannouncements. Table C.7 of the Appendix reports that the change in VIX from t− 6 to t− 1 where t is the trading day with theEmployment Situation announcement day is statistically significant and positive.55Our results are robust to the exclusion of the top and bottom 1% for both returns and ∆VIX.56We show in the Appendix the results using MAI-C2. The results remain quantitatively similar.84unemployment surprise, with predictability concentrating during NBER recessions when unemployment hasa stronger association with “bad news” for markets.57 Increasing attention before an announcement also pos-itively predicts market returns on the announcement date, controlling for the surprise itself. This suggestsa risk premium channel through which attention provides an instrument for the importance of an individualannouncement and accordingly the risk premium associated with it.4.5.2 FOMC AnnouncementsWe carry out a similar investigation of whether pre-announcement attention provides a useful instrumentfor the importance of individual FOMC announcement days. We test the hypothesis that attention providesan instrument for the level of the risk premium and the resolution of uncertainty associated with FOMCannouncements. We estimate:Rett = c+β1MAIt +β2FFSurpt +β3EPU t + et , (4.11)where Returnt is the daily log return of the S&P 500 index on the FOMC announcement day, and MAIt isthe contemporaneous detrended attention index to monetary policy.58 We detrend using the 3-day movingaverage minus the 30-day moving average because Figure 4.5 shows that attention to monetary policy beforeFOMC announcements rise significantly three days before FOMC announcements.59 We demean using a30-day average since 30 days is approximately the length of a FOMC cycle (Cieslak et al., 2016). Similarly,we detrend the EPU index using the 3-day minus 30-day moving average. We also include Fed Fund surprise(FFSurp) as calculated in Kuttner (2001) and Bernanke and Kuttner (2005).60We carry out our study on the relationship between attention to FOMC announcements and risk premiumafter 1994 when the Federal Reserve committee members began to announce publicly their monetary policydecisions (e.g., changes in the Fed fund rate). Figure 4.5 shows that attention to monetary policy changessignificantly around FOMC announcements only after 1994.Table 4.11, Panel A, shows the regression results. The first two columns focus on the period of 1994 to2016 and Column 1 and 2 report that the short-run trend in attention (MAI3−30) positively predicts marketreturn on FOMC announcement days. Results remain unchanged after controlling for EPU and the Fed Fundsurprise, which are both not statistically significant.61One concern may be that our results are driven by the recent financial crisis. Moreover, the fact that theFed Fund surprise is not significant in Column 2 may be explained by the zero-lower bound era of December2009 to December 2015 or because of the recent financial crisis. We present regression results over different57Garcia (2013) shows that the content of news better predicts future returns in recessions. Also, Kacperczyk et al. (2016) showthat investors focus on aggregate news in recession and idiosyncratic news in expansions.58We use MAIt rather than MAIt−1 because the regular print editions of the Wall Street Journal and the New York Times arereleased in the morning while FOMC announcements occur in the afternoon, usually around 2 p.m.59Figure C.3 of the appendix shows that uncertainty, using the VIX, also increases from three days before FOMC announcements.In Table C.7 of the Appendix, we show that the change in VIX from t−4 to t−1 where t is the FOMC announcement is statisticallysignificant and positive.60We winsorize the Fed Fund surprise at the top and bottom 1 percent.61In the Appendix, we repeat the same analysis using MAI-C2. Our results remain the same.85time periods, from 1994 to 2007 in Column 3 and 4, from 2008 to 2016 in Column 5 and 6, and from 2010and 2016 in Column 7 and 8.From 1994 to 2007, attention to monetary policy does not predict risk premium on announcement days.The Fed Fund surprise, on the other hand, is positive and statistically significant at the 1% level. This resultis consistent with the work of Bernanke and Kuttner (2005). The EPU is also positive and statisticallysignificant at the 1% level.62For the period of 2008 to 2016 and 2010 to 2016, our attention measure to monetary policy predictspositively returns on FOMC announcements and is statistically significant at the 1% and 5% level whereasthe Fed fund surprise and the EPU measure are not statistically significant. The fact that the Fed fund surpriseis not statistically significant is not surprising given that markets did not anticipate changes in interest ratesduring the zero-lower bound era.As shown in Ai and Bansal (2016), risk premium is linked with the resolution of uncertainty on FOMCannouncement days, therefore, when attention predicts risk premium, it should also predict resolution ofuncertainty. To investigate this idea, we test the dynamics of implied volatility around FOMC announce-ments using similar regression specifications defined in Equation (4.11) with the change in implied volatility(∆VIX) as the dependent variable. Table 4.11, Panel B, shows that when attention predicts risk premium, italso predicts significant resolution in uncertainty on FOMC announcement days. Consistent with results ofPanel A, we find no predictability from 1994 to 2007.63Overall, these results are consistent with the idea that pre-announcement attention provides an instru-ment for the level of the risk premium associated with individual FOMC announcement. These resultsare also consistent with theoretical models where attention allocation is costly, and investors pay greaterattention to more important announcements (Peng and Xiong, 2006; Kacperczyk et al., 2016) and whenuncertainty to fundamentals increases (Andrei and Hasler, 2016).4.6 ConclusionWe build indices of investor attention to macroeconomic fundamentals using news articles from WSJ andNYT. Attention indices rise around macroeconomic announcements and following changes in fundamentalsover quarterly, annual, and business cycle horizons. The effect of announcements and changes in fundamen-tals on indices is asymmetric, with bad news raising attention more than good news. Attention indices haveimportant implications for financial markets, and we show that aggregate trade volume and volatility coin-cide with rising attention, controlling for announcements. We further show that attention predicts surprisesas well as stock returns and resolution of uncertainty on unemployment and FOMC announcement days.Our paper adds to the growing literature documenting the importance of investor attention in financialmarkets (Da et al., 2011; Sicherman et al., 2016; Boguth et al., 2017; Sheng, 2017). Future work could go62We cannot confirm whether news articles used in the construction of EPU includes news that postpones FOMC announcements.If we lag the EPU by one day, the EPU variable is not statistically significant.63In non-tabulated results, we show that if we repeat the analysis presented in Table C.7 of the Appendix between 1994 and 2007,the VIX does not increase in a 5- or 3-day window prior to FOMC announcements. This may explain why attention in the daysleading to FOMC announcements does not predict resolution of uncertainty, and therefore, risk premium during that period.86in many directions. We find evidence of time-varying attention to different macroeconomic fundamentalsin the news media. In the spirit of the Merton (1980) Intertemporal Capital Asset Pricing Model, suchattention dynamics could be related to time-variation in the risks or risk premia associated with differenttypes of macroeconomic fundamentals. Another possible extension is to combine both investors’ sentimentand attention to macroeconomic fundamentals and relate to stock market returns.87Figure 4.1: Attention to Unemployment and the Unemployment RateThis figure shows monthly unemployment attention indices for the New York Times (MAI-NYT) and the Wall Street Journal (MAI-WSJ) and the monthly unemployment rate. The blue line is the attention index (MAI) and the dotted red line is the unemploymentrate. Units are in percentage. The gray vertical bars are NBER recessions. The sample period for MAI-NYT and MAI-WSJ is June1, 1980, to December 31, 2016, and January 1, 1984, to December 31, 2016, respectively.0.00.51.01.52.02.5UnemploymentMAIMAI-NYT1981 1985 1989 1993 1997 2001 2005 2009 2013 201712345UnemploymentMAIMAI-WSJUnemployment MAI Unemployment rate46810Unemploymentrate(%)46810Unemploymentrate(%)88Figure 4.2: Macro Attention and Macroeconomic FundamentalsThis figure shows the monthly macroeconomic attention indices (MAI) for the New York Times (MAI-NYT) and the Wall StreetJournal (MAI-WSJ) against their related macroeconomic fundamentals described in Table 4.2. All figures are at the monthlyfrequency except for the GDP MAI-Real GDP, which is at the quarterly frequency. Blue lines represent macroeconomic attentionindices (left y-axis) and dotted red lines (right y-axis) are MAI related macroeconomic fundamentals (see Table 4.2). Units arein percentage. The gray vertical bars are NBER recessions. The sample period for MAI-NYT and MAI-WSJ is June 1, 1980, toDecember 31, 2016, and January 1, 1984, to December 31, 2016, respectively.1981 1985 1989 1993 1997 2001 2005 2009 2013 20170.00.51.01.52.02.53.0CreditratingMAIMAI-NYT1981 1985 1989 1993 1997 2001 2005 2009 2013 20170.00.51.01.52.02.53.03.54.0MAI-WSJCredit rating MAI Corporate relative spread1020304050607010203040506070Corporaterelativespread1981 1985 1989 1993 1997 2001 2005 2009 2013 20170.40.60.81.01.2GDPMAIMAI-NYT1981 1985 1989 1993 1997 2001 2005 2009 2013 20172.02.53.03.54.04.55.05.5MAI-WSJGDP MAI Real GDP quarterly growth rate−2−1012−2−1012RealGDPquarterlygrowthrate(%)1981 1985 1989 1993 1997 2001 2005 2009 2013 20170.00.20.40.60.81.01.21.41.6HousingMAIMAI-NYT1981 1985 1989 1993 1997 2001 2005 2009 2013 20170123456MAI-WSJHousing MAI 12-month moving average log nominal home price return−1.0−0.50.00.51.0−1.0−0.50.00.51.0Loghomepricereturn(%)89Figure 4.3: Macroeconomic Attention and Macroeconomic Fundamentals (cont.)1981 1985 1989 1993 1997 2001 2005 2009 2013 20170.00.51.01.52.02.5InflationMAIMAI-NYT1981 1985 1989 1993 1997 2001 2005 2009 2013 20171234567MAI-WSJInflation MAI Change in CPI−1.0−0.50.00.51.0−1.0−0.50.00.51.0ChangeinCPI(%)1981 1985 1989 1993 1997 2001 2005 2009 2013 20170.00.20.40.60.8InterestMAIMAI-NYT1981 1985 1989 1993 1997 2001 2005 2009 2013 20170.00.51.01.52.02.53.03.54.0MAI-WSJInterest MAI Fed funds rate0.02.55.07.510.012.515.017.520.00.02.55.07.510.012.515.017.520.0Fedfundsrate(%)1981 1985 1989 1993 1997 2001 2005 2009 2013 201701234OilMAIMAI-NYT1981 1985 1989 1993 1997 2001 2005 2009 2013 20172468MAI-WSJOil MAI Oil log price2.53.03.54.04.55.02.53.03.54.04.55.0Oillogprice90Figure 4.4: Autocorrelation in Macroeconomic AttentionThis figure shows the autocorrelations (ρk) for residuals after controlling for day-of-the-week dummies and month-of-the-yeardummies for each of the composite macroeconomic attention index MAI-C1 for k lags ranging from 1 to 250 trading days. Thedashed line represents the 95% critical value for the test ρk ≤ 0, where we use the “large-lag” standard errors of Anderson (1976).These standard errors account for the observed autocorrelations for lags less than k.0.00.20.40.6Credit Rating GDP Housing0.00.20.40.6Inflation Interest Monetary0 50 100 150 200 250Lags0.00.20.40.6Oil0 50 100 150 200 250LagsUnemployment0 50 100 150 200 250LagsU.S. Dollar91Figure 4.5: Macroeconomic Attention around Macroeconomic AnnouncementsThis figure shows lag and forward estimated coefficients βδ from OLS regressions of detrended macroeconomic attention indicesMAI-C1 on announcement dummies as specified in Equation (4.4). The shaded area corresponds to the 95% confidence intervalaround the estimated coefficients. The x-axis corresponds to the number of days since the announcement. The announcements arethe Consumer Price Index (CPI) (first row), the Employment Situation (second row), the Federal Open Market Committee (FOMC)(third row), and the U.S. Gross Domestic Product (GDP) (fourth row). The vertical line represents the day of the announcement.−4 −2 0 2 4−0.20.00.20.40.6ResponseInflation MAI around CPI−4 −2 0 2 4Monetary MAI around CPI−4 −2 0 2 4Oil MAI around CPI−4 −2 0 2 4−0.20.00.20.40.6ResponseInflation MAI around Employment−4 −2 0 2 4Monetary MAI around Employment−4 −2 0 2 4Unemp. MAI around Employment−4 −2 0 2 4Days relative to announcement0.00.51.01.52.0ResponseMonetary MAI around FOMC−4 −2 0 2 4Days relative to announcementMonetary MAI - Pre-1994 FOMC−4 −2 0 2 4Days relative to announcementMonetary MAI - Post-1994 FOMC92Figure 4.6: Attention to Unemployment around Employment Situation AnnouncementsThis figure shows the average daily 60-day moving average of the unemployment attention index for the New York Times (MAI-NYT) and the Wall Street Journal (MAI-WSJ) around Employment Situation announcements, 250 trading days before and afterannouncements, conditioned on the unemployment surprise. Unemployment surprises are the random-walk and the Boyd et al.(2005) (BHJ) surprises, which we split into terciles. The MAI around low surprises is the solid blue line, medium surprises is thedotted red line, and high surprises is the dashed black line.Panel A: Random-walk surprise−200 −100 0 100 200Days relative to announcement0.600.650.700.750.80MAI-NYT−200 −100 0 100 200Days relative to announcement1.92.02.12.22.32.4MAI-WSJPanel B: BHJ surprise−200 −100 0 100 200Days relative to announcement0.550.600.650.700.75MAI-NYT−200 −100 0 100 200Days relative to announcement1.92.02.12.22.32.4MAI-WSJLow Medium High93Table 4.1: Newspapers Search WordsThis table presents the search words used to select the articles related to nine specific macroeconomic funda-mentals in the Wall Street Journal (WSJ) and New York Times (NYT). The nine macroeconomic fundamen-tals are credit 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)94Table 4.2: Macroeconomic Attention and Macroeconomic FundamentalsThis table presents the macroeconomic attention indices (MAI) for credit ratings, gross domestic product (GDP), housing market, inflation, interestrate, monetary, oil, US dollar, and unemployment and its related macroeconomic fundamentals and announcements. The table also reports the datasources for the fundamentals. The announcement dates are from Bloomberg except for the historical GDP announcements (pre-1997) that are from theU.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.95Table 4.3: Descriptive StatisticsPanel A shows daily descriptive statistics for unadjusted macro attention indices for the Wall Street Journal (MAI-WSJ) and the New YorkTimes (MAI-NYT), the Economic Policy Uncertainty (EPU) index, the implied volatility (VIX), and the 60-day detrended log trade volumefor the S&P 500. Columns Mon to Sun are the daily averages for each MAI. There are no MAI-WSJ on Sundays. The units for MAI indicesare in percentage. Panels B shows the correlation between daily demeaned macroeconomic attention composite indices (MAI-C1), EPU,VIX, and the detrended S&P 500 trade volume. Panel C shows the correlation at the monthly frequency. The sample period for MAI-NYTand MAI-WSJ is June 1, 1980, to December 31, 2016, and January 1, 1984, to December 31, 2016, respectively.Panel A: Daily MAI-NYT and MAI-WSJ descriptive statisticsObs. Mean St. Dev. Min Max Mon Tues Wed Thur Frid Sat SunNew York TimesCredit Rating 13,363 0.20 0.43 0.00 10.06 0.11 0.20 0.24 0.23 0.20 0.16 0.23GDP 13,363 0.89 0.76 0.00 8.15 0.64 0.81 0.82 0.86 0.83 0.72 1.52Housing 13,363 0.29 0.57 0.00 7.23 0.12 0.19 0.28 0.28 0.27 0.20 0.67Inflation 13,363 0.89 0.90 0.00 12.26 0.65 0.69 0.90 0.88 0.92 0.81 1.34Interest 13,363 0.25 0.39 0.00 4.75 0.19 0.20 0.27 0.29 0.26 0.25 0.33Monetary 13,363 0.92 0.77 0.00 8.68 0.60 0.78 0.98 1.04 1.06 0.95 1.03Oil 13,363 0.76 0.83 0.00 8.94 0.49 0.74 0.81 0.85 0.82 0.71 0.88Unemployment 13,363 0.80 0.89 0.00 10.53 0.57 0.54 0.70 0.67 0.91 0.77 1.44U.S. Dollar 13,363 0.08 0.20 0.00 3.34 0.01 0.08 0.07 0.08 0.07 0.07 0.18Wall Street JournalCredit Rating 12,053 0.46 0.88 0.00 9.67 0.50 0.58 0.72 0.57 0.62 0.23 0.00GDP 12,053 2.48 2.21 0.00 15.25 3.39 3.09 3.51 3.07 3.28 1.01 0.00Housing 12,053 0.70 1.44 0.00 17.18 0.62 0.68 1.40 0.82 0.99 0.42 0.00Inflation 12,053 2.23 2.05 0.00 15.71 3.23 2.45 2.97 2.85 3.15 0.95 0.00Interest 12,053 0.95 1.24 0.00 13.54 1.21 1.01 1.39 1.30 1.30 0.43 0.00Monetary 12,053 1.94 1.97 0.00 18.62 2.63 2.12 2.62 2.66 2.56 1.00 0.00Oil 12,053 2.33 2.55 0.00 19.47 2.79 2.95 3.35 3.02 3.14 1.04 0.00Unemployment 12,053 1.43 1.62 0.00 14.07 1.97 1.46 2.05 1.58 2.15 0.78 0.00U.S. Dollar 12,053 0.78 1.08 0.00 9.60 0.96 1.06 1.08 1.02 1.10 0.26 0.00Other VariablesEPU 11688 101.31 69.58 3.32 719.07 109.77 101.08 94.97 88.74 91.95 90.37 132.35VIX 6,802 19.67 7.85 9.31 80.86 19.93 19.70 19.64 19.57 19.54Volume 9,227 20.26 1.50 16.52 23.16 20.16 20.27 20.30 20.30 20.2596Table 4.4: Descriptive Statistics (cont.)Panel B: Daily MAI-C1 correlationCredit Rating GDP Housing Inflation Interest Monetary Oil Unemployment U.S. Dollar EPU VIX VolumeCredit Rating 1.00 0.16 0.16 -0.02 0.12 0.16 0.14 0.14 0.15 0.13 0.26 0.28GDP 0.16 1.00 0.14 0.19 0.21 0.26 0.25 0.24 0.22 0.08 0.12 0.19Housing 0.16 0.14 1.00 0.08 0.23 0.26 0.13 0.16 0.06 0.04 0.08 0.37Inflation -0.02 0.19 0.08 1.00 0.34 0.45 0.32 0.23 0.19 0.03 0.01 -0.23Interest 0.12 0.21 0.23 0.34 1.00 0.56 0.32 0.15 0.28 0.07 0.17 0.14Monetary 0.16 0.26 0.26 0.45 0.56 1.00 0.29 0.27 0.24 0.15 0.19 0.21Oil 0.14 0.25 0.13 0.32 0.32 0.29 1.00 0.03 0.38 0.03 0.09 0.02Unemployment 0.14 0.24 0.16 0.23 0.15 0.27 0.03 1.00 -0.02 0.22 0.26 0.13U.S. Dollar 0.15 0.22 0.06 0.19 0.28 0.24 0.38 -0.02 1.00 0.02 0.20 0.03EPU 0.13 0.08 0.04 0.03 0.07 0.15 0.03 0.22 0.02 1.00 0.34 0.05VIX 0.26 0.12 0.08 0.01 0.17 0.19 0.09 0.26 0.20 0.34 1.00 0.21Volume 0.28 0.19 0.37 -0.23 0.14 0.21 0.02 0.13 0.03 0.05 0.21 1.0097Table 4.5: Persistence of Macroeconomic AttentionPanel A of this table presents AR(p) models of the monthly demeaned macroeconomic attention compositeindices (MAI-C1), controlling for monthly time-fixed effects. DF(p-value) are p-values for the Dickey-Fuller statistics that test the null of a unit root in each time series. Panel B reports estimates from an OLSregression of daily MAI-C1 on various moving average lags of itself. L1 corresponds to the lag of itselfand L5, L21, L62, L250, and L1000 are the moving average for 5, 21, 62, 250, and 1000 days precedingthe observed values at time t. We control for day-of-week fixed effects. The standard errors are reportedin parenthesis and are calculated using Newey-West standard errors (10 lags). *, **, and *** denote thestatistical 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 Unemployment U.S. Dollarconst 0.02 0.05 -0.01 0.12*** 0.02 0.12** 0.17** 0.03 -0.02(0.03) (0.04) (0.04) (0.04) (0.03) (0.06) (0.08) (0.04) (0.03)AR(1) 0.70*** 0.42*** 0.47*** 0.51*** 0.59*** 0.48*** 0.71*** 0.61*** 0.66***(0.08) (0.04) (0.10) (0.05) (0.05) (0.04) (0.05) (0.06) (0.06)AR(2) -0.01 0.26*** 0.10 0.20*** 0.15** 0.16*** 0.17*** 0.18*** 0.09(0.10) (0.05) (0.08) (0.04) (0.06) (0.05) (0.04) (0.05) (0.06)AR(3) -0.00 0.16*** 0.29*** 0.06 -0.05 0.14** 0.01 0.11** 0.00(0.07) (0.05) (0.10) (0.05) (0.07) (0.07) (0.07) (0.05) (0.05)AR(4) 0.14** 0.05 0.01 0.10** 0.14** 0.05 0.02 0.01 0.18***(0.07) (0.05) (0.06) (0.05) (0.06) (0.05) (0.04) (0.04) (0.04)DF (p-value) 0.00 0.01 0.03 0.00 0.00 0.00 0.00 0.00 0.11Obs. 435 435 435 435 435 435 435 435 435Adj-R2 0.58 0.65 0.63 0.66 0.58 0.52 0.78 0.78 0.81Panel B: Daily MAI-C1 regressions on lagged attentionCredit Rating GDP Housing Inflation Interest Monetary Oil Unemployment U.S. DollarIntercept -0.08*** 0.09** -0.21*** 0.11** -0.04 -0.08* -0.22*** 0.04 -0.08***(0.02) (0.04) (0.02) (0.05) (0.03) (0.05) (0.05) (0.04) (0.02)L1 0.07*** 0.05*** 0.06** 0.04*** 0.12*** 0.18*** 0.06*** 0.01 -0.00(0.02) (0.01) (0.03) (0.01) (0.02) (0.02) (0.02) (0.02) (0.02)L5 0.27*** 0.11*** 0.55*** 0.13*** 0.18*** 0.19*** 0.37*** 0.23*** 0.18***(0.05) (0.03) (0.06) (0.03) (0.03) (0.03) (0.04) (0.03) (0.04)L21 0.44*** 0.04 0.05 0.29*** 0.31*** 0.20*** 0.35*** 0.21*** 0.50***(0.07) (0.06) (0.09) (0.06) (0.06) (0.05) (0.05) (0.07) (0.07)L62 0.02 0.34*** 0.12** 0.34*** 0.08 0.13* 0.14*** 0.31*** 0.14*(0.07) (0.10) (0.06) (0.07) (0.08) (0.07) (0.05) (0.08) (0.07)L250 0.12* 0.43*** 0.20*** 0.09 0.26*** 0.23*** 0.03 0.26*** 0.18***(0.06) (0.10) (0.08) (0.06) (0.07) (0.07) (0.03) (0.07) (0.06)L1000 0.02 -0.03 -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. 8545 8545 8545 8545 8545 8545 8545 8545 8545Adj-R2 0.28 0.15 0.42 0.17 0.23 0.26 0.53 0.31 0.4098Table 4.6: Macroeconomic Attention and Macroeconomic FundamentalsThis table presents results of OLS regressions of monthly macroeconomic attention indices (MAI) on different macroeconomic fundamentals. Panel A and Panel Breport results for the New York Times macroeconomic attention indices (MAI-NYT) and the Wall Street Journal (MAI-WSJ), respectively. The general regressionis specified in Equation (4.6). F corresponds to the associated fundamental to each MAI as described in Table 4.2 and Ft is the moving average over t days ofthe respective macroeconomic fundamental. All observations are at the monthly frequency except for the GDP MAI-Real GDP growth, which is at the quarterlyfrequency. We control for monthly fixed effects. The standard errors are reported in parenthesis and are calculated using Newey-West standard errors (10 lags). *,**, *** denote the statistical significance at the 10%, 5%, 1% levels, respectively. The sample period for MAI-NYT and MAI-WSJ is June 1, 1980, to December31, 2016, and January 1, 1984, to December 31, 2016, respectively.Panel A: MAI-NYTMAI: Credit Rating GDP Housing Inflation Interest Monetary Oil Unemployment U.S. 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 0.021 -0.227* -0.176*** -0.018 -0.021 -0.004 0.041 -0.000(0.013) (0.119) (0.067) (0.017) (0.035) (0.004) (0.149) (0.001)Ft,t−3−Ft,t−12 -0.001 0.055* -0.309*** -0.542*** 0.005 -0.010 0.002 0.064 -0.001(0.004) (0.032) (0.108) (0.157) (0.012) (0.034) (0.009) (0.090) (0.003)Ft,t−12−Ft,t−48 -0.012 0.129 -0.021 0.644 -0.018*** -0.041* 0.035 0.129*** -0.019(0.011) (0.101) (0.104) (0.744) (0.006) (0.021) (0.024) (0.042) (0.012)(Ft −Ft,t−3)2 0.000 0.540*** -0.460*** 0.029*** 0.058*** 0.002*** 0.643 0.000(0.001) (0.117) (0.166) (0.007) (0.017) (0.001) (0.721) (0.001)(Ft,t−3−Ft,t−12)2 -0.000 -0.023 0.242*** -0.271 0.013** 0.048*** 0.003*** 0.229** -0.004*(0.000) (0.048) (0.086) (0.176) (0.005) (0.014) (0.001) (0.104) (0.002)(Ft,t−12−Ft,t−48)2 0.001 0.332** 0.399** 6.615*** 0.006*** -0.005 -0.005 0.071*** -0.015(0.001) (0.134) (0.196) (2.170) (0.002) (0.007) (0.006) (0.023) (0.011)Intercept 0.186*** 0.753*** 0.017 0.638*** 0.207*** 0.836*** 0.505*** 0.563*** 0.067***(0.035) (0.061) (0.042) (0.073) (0.028) (0.063) (0.082) (0.064) (0.017)Obs. 439 130 439 439 439 439 396 439 439Adj-R2 0.05 0.06 0.34 0.16 0.13 0.09 0.26 0.51 -0.0099Panel B: MAI-WSJMAI: Credit Rating GDP Housing Inflation Interest Monetary Oil Unemployment U.S. 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 0.050** -0.266 -0.279 -0.298 -0.435 -0.018* -0.166 0.002(0.021) (0.295) (0.181) (0.240) (0.371) (0.010) (0.263) (0.013)Ft,t−3−Ft,t−12 0.024** 0.133 -0.670*** 0.525 0.163 0.174 0.005 0.122 -0.026(0.011) (0.141) (0.252) (0.435) (0.163) (0.242) (0.020) (0.250) (0.041)Ft,t−12−Ft,t−48 0.020 0.077 -0.287 4.800*** 0.130 0.125 0.157* 0.265*** -0.332***(0.021) (0.380) (0.306) (1.314) (0.090) (0.118) (0.095) (0.087) (0.128)(Ft −Ft,t−3)2 -0.002 0.487 -0.256 0.515 0.082 0.005*** 3.112** 0.013*(0.003) (0.473) (0.339) (0.621) (0.822) (0.001) (1.398) (0.008)(Ft,t−3−Ft,t−12)2 0.001 0.005 0.673*** 0.927* 0.355*** 0.298* 0.006*** 0.208 0.052**(0.001) (0.195) (0.232) (0.471) (0.123) (0.177) (0.001) (0.187) (0.021)(Ft,t−12−Ft,t−48)2 0.001 1.195*** 2.359*** 14.999** 0.071* 0.055 0.001 0.072* 0.305**(0.001) (0.384) (0.438) (6.180) (0.038) (0.064) (0.018) (0.039) (0.141)Intercept 0.557*** 3.195*** 0.173* 3.018*** 1.049*** 2.553*** 2.719*** 1.878*** 0.845***(0.079) (0.185) (0.104) (0.100) (0.099) (0.203) (0.344) (0.132) (0.152)Obs. 396 116 396 396 396 396 396 396 396Adj-R2 0.11 0.08 0.47 0.19 0.12 0.02 0.07 0.34 0.13100Table 4.7: Media Attention and Aggregate Trade VolumeThis table presents results of OLS regressions of the daily detrended S&P 500 trade volume on the differencebetween the 5-day and 20-day moving 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 log trade volume using the movingaverage of the log trade volume of the past 60 trading days. For all model specifications, we control forday-of-week fixed effects. The standard errors are reported in parenthesis and are calculated using Newey-West standard errors (250 lags). *, **, *** denote the statistical significance at the 10%, 5%, 1% levels,respectively. The sample period is from June 1, 1980, to December 31, 2016.MAI: Inflation Monetary InterestAnn: CPI and PPI FOMC FOMC(1) (2) (3) (4) (5) (6) (7) (8) (9)MAI5−20 0.053*** 0.052*** 0.059*** 0.064*** 0.064*** 0.064*** 0.057*** 0.056*** 0.058***(0.009) (0.009) (0.009) (0.008) (0.008) (0.008) (0.012) (0.012) (0.012)Ann 0.024*** 0.026*** 0.034*** 0.035** 0.038*** 0.044***(0.006) (0.006) (0.012) (0.015) (0.012) (0.012)MAI5−20×Ann -0.075*** -0.009 -0.077*(0.017) (0.046) (0.044)Intercept 0.003 -0.001 -0.001 0.003 0.003 0.003 0.003 0.003 0.003(0.006) (0.006) (0.006) (0.006) (0.006) (0.006) (0.006) (0.006) (0.006)Obs. 9208 9208 9208 9208 9208 9208 9208 9208 9208Adj-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 Employment(1) (2) (3) (4) (5) (6) (7) (8) (9)MAI5−20 0.035*** 0.035*** 0.036*** 0.034*** 0.034*** 0.035*** 0.069*** 0.036*** 0.103***(0.008) (0.008) (0.008) (0.010) (0.010) (0.010) (0.018) (0.009) (0.019)Ann 0.009 0.008 0.012 0.017(0.007) (0.007) (0.010) (0.012)MAI5−20×Ann -0.026 -0.035(0.025) (0.033)Intercept 0.003 0.003 0.003 0.004 0.001 0.001 0.004 0.003 0.003(0.006) (0.006) (0.006) (0.006) (0.007) (0.007) (0.006) (0.006) (0.006)Obs. 9208 9208 9208 9208 9208 9208 9208 9208 9208Adj-R2 0.05 0.05 0.05 0.05 0.05 0.05 0.05 0.05 0.05101Table 4.8: Media Attention and Implied VolatilityThis table presents results of OLS regressions of the daily implied volatility proxied by VIX regressed onthe difference between the 20- and 250-day moving average MAI-C1 and a dummy (Ann) equal to one ifthere is a related announcement specified in Table 4.2, zero otherwise. For all model specifications, wecontrol for day-of-week fixed effects. The standard errors are reported in parenthesis and are calculatedusing Newey-West standard errors (250 lags). *, **, *** denote the statistical significance at the 10%, 5%,1% levels, respectively. The sample period is from June 1, 1980, to December 31, 2016.MAI: Inflation Monetary InterestAnn: CPI and PPI FOMC FOMC(1) (2) (3) (4) (5) (6) (7) (8) (9)MAI20−250 -4.183 -4.182 -4.184 2.749** 2.749** 2.768** 2.830 2.830 2.868(3.215) (3.215) (3.175) (1.371) (1.371) (1.368) (2.274) (2.273) (2.268)Ann 0.077 0.078 0.304 0.306 0.306 0.305(0.102) (0.101) (0.426) (0.427) (0.434) (0.434)MAI20−250×Ann 0.026 -0.725 -1.277(0.649) (0.607) (1.032)Intercept 19.472*** 19.460*** 19.460*** 19.547*** 19.547*** 19.547*** 19.553*** 19.553*** 19.553***(1.107) (1.107) (1.105) (1.173) (1.173) (1.173) (1.176) (1.176) (1.176)Obs. 6802 6802 6802 6802 6802 6802 6802 6802 6802Adj-R2 0.03 0.03 0.03 0.02 0.02 0.02 0.01 0.01 0.01MAI: GDP Unemployment Credit Rating Oil USDAnn: GDP Report Employment(1) (2) (3) (4) (5) (6) (7) (8) (9)MAI20−250 6.814** 6.814** 6.823** 10.476** 10.478** 10.519** 7.097*** 0.716 4.991**(2.951) (2.952) (2.952) (4.172) (4.172) (4.175) (2.647) (1.207) (2.303)Ann 0.047 0.044 0.215 0.210(0.137) (0.139) (0.157) (0.165)MAI20−250×Ann -0.175 -0.847(0.545) (0.684)Intercept 19.538*** 19.534*** 19.534*** 19.477*** 19.428*** 19.427*** 19.541*** 19.544*** 19.564***(1.093) (1.092) (1.092) (0.989) (0.989) (0.988) (1.135) (1.163) (1.159)Obs. 6802 6802 6802 6802 6802 6802 6802 6802 6802Adj-R2 0.06 0.06 0.06 0.17 0.17 0.17 0.06 0.00 0.02102Table 4.9: Forecasting Unemployment SurprisesThis table presents results of OLS regressions of the unemployment surprise on the one-day lag detrendeddemeaned daily composite MAI-C1 for unemployment at different frequencies and an interaction term be-tween MAI-C1 and a NBER dummy. For example, MAI5−20 is the difference between the five-day and thetwenty-day moving average of MAI-C1. The NBER dummy is equal to one if the unemployment surpriseoccurs during a NBER recession, zero otherwise. The surprise is calculated as the difference between theactual unemployment for the month t reported in the month t + 1 and the previous month (t-1) unemploy-ment rate (i.e., the random-walk) in Panel A, the forecasted unemployment rate as in Boyd et al. (2005)(BHJ) in Panel B, and the average unemployment rate from economist surveys in Bloomberg in Panel C.The standard errors are reported in parenthesis and are calculated using Newey-West (5 lags). *, **, ***denote the statistical significance at the 10%, 5%, 1% levels, respectively. The sample period is June 1, 1980to December 31, 2016 for the random-walk and BHJ surprise and from January 1, 1997 to December 31,2016 for the Bloomberg surprise.Panel A: Random-walk surpriseMAI: MAI5−20 MAI20−250 MAI60−250(1) (2) (3) (4) (5) (6)MAI 0.036 0.016 0.142*** 0.089** 0.217*** 0.111*(0.024) (0.022) (0.045) (0.042) (0.066) (0.061)MAI×NBER 0.303* 0.186** 0.376***(0.169) (0.089) (0.077)Intercept -0.011 -0.011 -0.003 -0.010 -0.003 -0.014(0.011) (0.011) (0.010) (0.010) (0.010) (0.009)Obs. 438 438 427 427 427 427R2 0.00 0.03 0.06 0.08 0.08 0.12Panel B: BHJ surpriseMAI: MAI5−20 MAI20−250 MAI60−250(1) (2) (3) (4) (5) (6)MAI 0.034 0.029 0.075** 0.071** 0.112** 0.070(0.023) (0.022) (0.030) (0.034) (0.045) (0.051)MAI×NBER 0.084 0.013 0.149*(0.118) (0.080) (0.078)Intercept -0.032*** -0.032*** -0.027*** -0.027*** -0.027*** -0.031***(0.009) (0.009) (0.008) (0.008) (0.008) (0.008)Obs. 438 438 427 427 427 427R2 0.00 0.01 0.02 0.02 0.02 0.03103Table 4.10: Forecasting Returns and Changes in VIX on Employment Situation AnnouncementsThis table presents results of OLS regressions of the daily S&P 500 log return (in %) and changes in impliedvolatility (∆VIX) on Employment Situation announcement days on attention to unemployment. Specifically,we regress log returns in Columns 1 to 3 and ∆VIX in Columns 4 to 6 on the one-day lag detrended demeanedunemployment attention composite index MAI-C1, the Boyd et al. (2005) (BHJ) unemployment surprise,the one-day lag detrended Economic Policy Uncertainty (EPU) index, and each of independent variablesinteracted with a NBER dummy equal to one if the Employment Situation announcement day occurs duringa NBER recession, zero otherwise. MAI5−20,t and EPU5−20,t is the difference between the five-day and20-day moving average of MAI-C1 for unemployment and the EPU index, respectively. The standard errorsare reported in parenthesis and are calculated using Newey-West (5 lags). *, **, *** denote the statisticalsignificance at the 10%, 5%, 1% levels, respectively. The sample period for the return regression is June 1,1980 to December 31, 2016 and for the VIX regression is January 1990 to December 31, 2016.S&P 500 Returns ∆VIX(1) (2) (3) (4) (5) (6)MAI5−20 0.35** 0.30* 0.38** -0.86*** -0.70** -0.72***(0.16) (0.16) (0.17) (0.28) (0.28) (0.28)BHJ 0.01 0.26 0.36 0.95 0.21 0.17(0.32) (0.31) (0.34) (0.70) (0.64) (0.64)EPU5−20 -0.01*** 0.00(0.00) (0.00)NBER×MAI5−20 0.74 1.58* -2.01 -2.52*(0.75) (0.90) (1.40) (1.46)NBER×BHJ -1.60 -3.38*** 5.74** 6.28**(1.15) (1.26) (2.79) (2.61)NBER×EPU5−20 -0.01 0.01(0.01) (0.01)Intercept -0.02 -0.00 0.02 -0.19* -0.23** -0.26**(0.06) (0.06) (0.07) (0.10) (0.10) (0.10)Obs. 428 428 374 317 317 317Adj-R2 0.01 0.01 0.06 0.04 0.08 0.10104Table 4.11: Forecasting Returns and Changes in VIX on FOMC AnnouncementsThis table presents results of OLS regressions of the daily S&P 500 log return (in %) in Panel A andchanges in the implied volatility (∆VIX) in Panel B on the detrended demeaned monetary policy attentioncomposite index MAI-C1, the Fed Fund surprise, and the detrended Economic Policy Uncertainty (EPU)index. MAI3−30,t and EPU3−30,t is the difference between the 3- and 30-day moving average of MAI-C1 andEPU, respectively. The standard errors are reported in parenthesis and are calculated using the Newey-Weststandard errors (5 lags). *, **, *** denote the statistical significance at the 10%, 5%, 1% levels, respectively.The sample period is January 1, 1994 to December 31, 2016.Panel A: S&P 500 Returns1994-2016 1994-2007 2008-2016 2010-2016(1) (2) (3) (4) (5) (6) (7) (8)MAI3−30 0.40** 0.39** 0.01 0.04 0.73** 0.85** 0.83** 0.85***(0.19) (0.19) (0.17) (0.17) (0.31) (0.33) (0.33) (0.31)FFSurp -4.36 -4.35 -5.83** -5.54** 3.33 3.88 -1.67 -1.86(3.21) (3.24) (2.28) (2.19) (10.26) (9.44) (8.20) (8.39)EPU3−30 0.00 0.01*** -0.01 -0.00(0.00) (0.00) (0.01) (0.00)Intercept 0.26*** 0.25*** 0.23** 0.13 0.36** 0.37** 0.09 0.09(0.08) (0.09) (0.09) (0.10) (0.15) (0.15) (0.13) (0.13)Obs. 183 183 111 111 72 72 56 56Adj-R2 0.05 0.04 0.06 0.09 0.12 0.14 0.21 0.20Panel B: ∆VIX1994-2016 1994-2007 2008-2016 2010-2016(1) (2) (3) (4) (5) (6) (7) (8)MAI3−30 -0.87** -0.80** -0.03 -0.09 -1.60** -1.53*** -1.90** -1.70**(0.39) (0.35) (0.18) (0.17) (0.64) (0.57) (0.86) (0.72)FFSurp 5.11 4.99 7.92** 7.46** -8.90 -8.56 9.11 5.87(4.40) (4.41) (3.78) (3.73) (9.76) (10.13) (12.76) (13.13)EPU3−30 -0.01* -0.01*** -0.00 -0.01(0.01) (0.00) (0.01) (0.01)Intercept -0.49*** -0.40*** -0.52*** -0.36*** -0.56*** -0.55*** -0.29 -0.31(0.10) (0.11) (0.11) (0.12) (0.20) (0.20) (0.20) (0.21)Obs. 183 183 111 111 72 72 56 56Adj-R2 0.10 0.13 0.08 0.14 0.26 0.26 0.29 0.30105Chapter 5ConclusionThis thesis studies the relationship between information and stock returns, which is an important and long-lasting question in finance. I employ novel datasets and methods to shed new lights on this relationship.With the arrival of the internet and social media, there are more and more new types of information. Yet,their values to the stock markets are not clear. Is online information novel to existing information? Doesonline information capture firm’s fundamentals? Answering these questions are important but challenging.The first essay studies the value of online employee reviews to stock markets, using a novel dataset of nearlyone million employee reviews. I also show that why online employee reviews matter for stock market: theycan predict firms’ cash flow news. Thus, online reviews do capture a firm’s fundamentals. Moreover, thisessay is the first paper decodes the information hierarchies within firms. I find that relatively high-levelemployees’ expectations are better in predicting future stock returns. Also, in this essay, I employ machinelearning tools to perform some analysis, which is new to this literature.Nowadays, we often face multiple information sources. Thus, it crucial to understand the interactionbetween different types of information. In the second essay, I investigate interactions between macro-announcements and the processing of earnings news. The main contribution of this essay is to challengethe existing theories that suggest that macro-news should crowd out attention to firm-level news, implyingless efficient pricing. However, I find the opposite. To explain these results, I show that institutional investorattention is higher on macro-news days. Hence, macro-news appears not to be a distraction from firm-levelnews, but instead serves to enhance overall attention to financial markets. I suggest extensions of existingtheories that could be consistent with these findings. My empirical evidence suggests new direction forrational attention theories to model the relationship between two types of news.News media plays a more and more important role in financial markets. In the third essay, my co-authorsAdlai Fisher and Charles Martineau and I study the impacts of news media to stock markets by constructingindices of media attention to macroeconomic risks including employment, growth, and monetary policy.Our findings have two important implications for stock markets. First, increases in aggregate trade volumeand volatility coincide with rising attention, controlling for announcements. Second, changes in attentionbefore unemployment and FOMC announcements predict announcement surprises, stock returns, and im-plied volatility changes on the announcement day. We conclude that media attention to macroeconomicfundamentals provides market-relevant information beyond the contents and dates of macroeconomic an-nouncements.In the future work, I plan to extend my thesis in several dimensions. For instance, the first essay studiesthe value of online reviews to stock markets. It would be interesting to study the real impacts of onlinereviews. In particular, whether online company reviews can affect corporate decisions, such as investments106and innovations. In the past, corporations relied on stock market and traditional information sources to maketheir decisions, but now there are new types of information. It is important to know how the new informationenvironment can affect firms’ activities. The second direction is to use machine learning and deep learningto better understand the links between news and stock returns based on a comprehensive dataset of DowsNews Archives. It is still not clear what type of news moves stock prices, and advanced tools like machinelearning and deep learning allows us to explore this question.107BibliographyVikas Agarwal, Wei Jiang, Yuehua Tang, and Baozhong Yang. Uncovering hedge fund skill from the port-folio holdings they hide. Journal of Finance, 68(2):739–783, 2013.Hengjie Ai and Ravi Bansal. Risk preferences and the macro announcement premium. Working paper,2016.Torben G. Andersen. Return volatility and trading volume: An information flow interpretation of stochasticvolatility. Journal of Finance, 51(1):169–204, 1996.Torben G. Andersen, Tim Bollerslev, Francis X. Diebold, and Heiko Ebens. The distribution of realizedstock return volatility. Journal of Financial Economics, 61(1):43–76, 2001.Torben G. Andersen, Tim Bollerslev, Francis X. Diebold, and Clara Vega. Micro effects of macro announce-ments: real-time price discovery in foreign exchange. American Economic Review, 93:38–62, 2003.Torben G. Andersen, Tim Bollerslev, Francis X. Diebold, and Clara Vega. Real-time price discovery inglobal stock, bond and foreign exchange markets. Journal of International Economics, 73:251–277,2007.Oliver Duncan Anderson. Time series analysis and forecasting: the Box-Jenkins approach. ButterworthsLondon, 1976.Daniel Andrei and Michael Hasler. Investor attention and stock market volatility. Review of FinancialStudies, 28:33–72, 2014.Daniel Andrei and Michael Hasler. Dynamic attention behavior under return predictability. Working paper,2016.Werner Antweiler and Murray Z. Frank. Is all that talk just noise? the information content of internet stockmessage boards. Journal of Finance, 59(3):1259–1294, 2004.Ilona Babenko and Rik Sen. Do nonexecutive employees have valuable information? evidence from em-ployee stock purchase plans. Management Science, 62(7):1878–1898, 2015.Jennie Bai, Thomas Philippon, and Alexi Savov. Have financial markets become more informative? Journalof Financial Economics, 122(3):625–654, 2016.Scott R. Baker, Nicholas Bloom, and Steven J. Davis. Measuring economic policy uncertainty. QuarterlyJournal of Economics, 131(4):1593–1636, 2016.108Ray Ball and Philip Brown. An empirical evaluation of accounting income numbers. Journal of AccountingResearch, pages 159–178, 1968.Brad M. Barber and Terrance Odean. All that glitters: The effect of attention and news on the buyingbehavior of individual and institutional investors. Review of Financial Studies, 21(2):785–818, 2008.Alessandro Beber, Micheal W. Brandt, and Kenneth A. Kavajecz. What does equity sector orderflow tell usabout the economy? Review of Financial Studies, 24:3688–3730, 2011.Azi Ben-Rephael, Zhi Da, and Ryan D. Israelsen. It depends on where you search: A comparison ofinstitutional and retail attention. Review of Financial Studies, forthcoming, 2017.Shlomo Benartzi. Excessive extrapolation and the allocation of 401 (k) accounts to company stock. Journalof Finance, 56(5):1747–1764, 2001.Ben S. Bernanke and Kenneth N. Kuttner. What explains the stock market’s reaction to federal reservepolicy? Journal of Finance, 60(3):1221–1257, 2005.Victor L Bernard and Jacob K Thomas. Post-earnings-announcement drift: delayed price response or riskpremium? Journal of Accounting Research, pages 1–36, 1989.Gennaro Bernile, Jianfeng Hu, and Yuehua Tang. Can information be locked up? informed trading ahead ofmacro-news announcements. Journal of Financial Economics, 121(3):496–520, 2016.Justin Birru. Day of the week and the cross-section of returns. Journal of Financial Economics, Forthcom-ing, 2017.Nicholas Bloom, Raffaella Sadun, and John Van Reenen. The organization of firms across countries. Quar-terly Journal of Economics, 127(4):1663–1705, 2012.Oliver Boguth, Ran Duchin, and Mikhail Simutin. Dissecting conglomerates. Working Paper, 2016a.Oliver Boguth, David Newton, and Mikhail Simutin. The fragility of organization capital. Working Paper,2016b.Oliver Boguth, Vincent Grégoire, and Charles Martineau. Shaping expectations and coordinating attention:The unintended consequences of press conferences. Working Paper, 2017.Tim Bollerslev and Hans O. Mikkelsen. Modeling and pricing long memory in stock market volatility.Journal of Econometrics, 73(1):151–184, 1996.Romain Boulland and Olivier Dessaint. Announcing the announcement. Journal of Banking and Finance,82:59–79, 2017.John H. Boyd, Jian Hu, and Ravi Jagannathan. The stock market’s reaction to unemployment news: Whybad news is usually good for stocks. Journal of Finance, 60:649–672, 2005.109Jonathan Brogaard and Andrew Detzel. The asset-pricing implications of government economic policyuncertainty. Management Science, 61:3–18, 2015.Francesca Brusa, Pavel G. Savor, and Mungo I. Wilson. One central bank to rule them all. Working paper,2016.Lorenzo Caliendo, Ferdinando Monte, and Esteban Rossi-Hansberg. The anatomy of french productionhierarchies. Journal of Political Economy, 123(4):809–852, 2015.Laurent E. Calvet and Adlai J. Fisher. Multifrequency news and stock returns. Journal of Financial Eco-nomics, 86:178–212, 2007.John L Campbell, Matthew D DeAngelis, and James Moon. The credibility of the crowds: Personal stockholdings and investorsâ response to stock opinions on social media. Working paper, 2017.John Y. Campbell and Motohiro Yogo. Efficient tests of stock return predictability. Journal of FinancialEconomics, 81:27–60, 2006.Mark M Carhart. On persistence in mutual fund performance. Journal of Finance, 52(1):57–82, 1997.Christopher D. Carroll. Macroeconomic expectations of households and professional forecasters. QuarterlyJournal of Economics, 118:269–298, 2003.Wesley S Chan. Stock price reaction to news and no-news: drift and reversal after headlines. Journal ofFinancial Economics, 70(2):223–260, 2003.Jonathan Chang, Sean Gerrish, Chong Wang, Jordan L Boyd-Graber, and David M Blei. Reading tea leaves:How humans interpret topic models. In Advances in neural information processing systems, pages 288–296, 2009.Hailiang Chen, Prabuddha De, Yu Jeffrey Hu, and Byoung-Hyoun Hwang. Wisdom of crowds: The valueof stock opinions transmitted through social media. Review of Financial Studies, 27(5):1367–1403, 2014.Nai-Fu Chen, Richard Roll, and Stephen A. Ross. Economic forces and the stock market. Journal ofBusiness, 59:383–403, 1986.Chin-Han Chiang, Wei Dai, Jianqing Fan, Harrison Hong, and Jun Tu. Robust measures of earnings sur-prises. Working paper, 2016.Anna Cieslak, Adair Morse, and Annette Vissing-Jorgensen. Stock returns over the fomc cycle. WorkingPaper, 2016.Lauren Cohen. Loyalty-based portfolio choice. Review of Financial Studies, 22(3):1213–1245, 2008.Lauren Cohen and Andrea Frazzini. Economic links and predictable returns. Journal of Finance, 63(4):1977–2011, 2008.110Lauren Cohen and Dong Lou. Complicated firms. Journal of Financial Economics, 104(2):383–400, 2012.Shane A Corwin and Jay F Coughenour. Limited attention and the allocation of effort in securities trading.Journal of Finance, 63(6):3031–3067, 2008.Joshua D Coval and Tobias J Moskowitz. Home bias at home: Local equity preference in domestic portfo-lios. Journal of Finance, 54(6):2045–2073, 1999.Joshua D Coval and Tobias J Moskowitz. The geography of investment: Informed trading and asset prices.Journal of Political Economy, 109(4):811–841, 2001.Zhi Da, Joseph Engelberg, and Pengjie Gao. In search of attention. Journal of Finance, 66:1461–1499,2011.Stefano DellaVigna and Ethan Kaplan. The fox news effect: Media bias and voting. Quarterly Journal ofEconomics, 122:1187–1234, 2007.Stefano DellaVigna and Joshua M. Pollet. Investor inattention and friday earnings announcements. Journalof Finance, 64(2):709–749, 2009.Rick Di Mascio, Anton Lines, and Narayan Y Naik. Alpha decay. Working paper, 2017.Casey Dougal, Joseph Engelberg, Diego Garcia, and Christopher A Parsons. Journalists and the stockmarket. Review of Financial Studies, 25(3):639–679, 2012.Michael S Drake, Darren T Roulstone, and Jacob R Thornock. Investor information demand: Evidencefrom google searches around earnings announcements. Journal of Accounting Research, 50(4):1001–1040, 2012.Alex Edmans. Does the stock market fully value intangibles? employee satisfaction and equity prices.Journal of Financial Economics, 101(3):621–640, 2011.Alex Edmans, Diego Garcia, and Øyvind Norli. Sports sentiment and stock returns. Journal of Finance, 62(4):1967–1998, 2007.Andrea L Eisfeldt and Dimitris Papanikolaou. Organization capital and the cross-section of expected returns.Journal of Finance, 68(4):1365–1406, 2013.Joseph. Engelberg and Christopher A. Parsons. The causal impact of media in financial markets. Journal ofFinance, 66:67–97, 2011.Joseph Engelberg, R David McLean, and Jeffrey Pontiff. Anomalies and news. Journal of Finance, Forth-coming, 2017.EF Fama and KR French. A five-factor asset pricing model. Journal of Financial Economics, 116(1):1–22,2015.111Eugene F Fama. Efficient capital markets: A review of theory and empirical work. Journal of Finance, 25(2):383–417, 1970.Eugene F. Fama and Kenneth R. French. Common risk factors in the returns on stocks and bonds. Journalof Financial Economics, 33(1):3–56, 1993.Eugene F Fama and James D MacBeth. Risk, return, and equilibrium: Empirical tests. Journal of PoliticalEconomy, 81(3):607–636, 1973.Lily H. Fang and Joel Peress. Media coverage and the cross-section of stock returns. Journal of Finance,64:2023–2052, 2009.Maryam Farboodi, Adrien Matray, and Laura Veldkamp. Where has all the big data gone? Working paper,2017.Adlai Fisher, Charles Martineau, and Jinfei Sheng. Media attention, macroeconomic fundamentals, and thestock market. Working paper, 2017.Andrea Frazzini and Owen Lamont. The earnings announcement premium and trading volume. Workingpaper, 2007.Jerome Friedman, Trevor Hastie, and Robert Tibshirani. The elements of statistical learning, volume 1.Springer series in statistics New York, 2001.Diego Garcia. Sentiment during recessions. Journal of Finance, 68:1267–1300, 2013.Jon A Garfinkel and Jonathan Sokobin. Volume, opinion divergence, and returns: A study of post–earningsannouncement drift. Journal of Accounting Research, 44(1):85–112, 2006.Luis Garicano. Hierarchies and the organization of knowledge in production. Journal of Political Economy,108(5):874–904, 2000.Matthew Gentzkow and Jesse M. Shapiro. What drives media slant? evidence from us daily newspapers.Econometrica, 78:35–71, 2010.Matthew Gentzkow, Bryan T Kelly, and Matt Taddy. Text as data. Technical report, National Bureau ofEconomic Research, 2017.Eric Ghysels, Pedro Santa-Clara, and Rossen Valkanov. Predicting volatility: getting the most out of returndata sampled at different frequencies. Journal of Econometrics, 131(1):59–95, 2006.Thomas Gilbert. Information aggregation around macroeconomic announcements: Revisions matter. Jour-nal of Financial Economics, 101:114–131, 2011.Thomas Gilbert, Chiara Scotti, Georg Strasser, and Clara Vega. Is the intrinsic value of a macroeconomicnews announcement related to its asset price impact? Journal of Monetary Economics, 92:78–95, 2017.112T Clifton Green, Ruoyan Huang, Quan Wen, and Dexin Zhou. Crowdsourced employer reviews and stockreturns. Working paper, 2018.Jillian Popadak Grennan. A corporate culture channel: How increased shareholder governance reduces firmvalue. Working paper, 2013.Sanford J Grossman and Joseph E Stiglitz. On the impossibility of informationally efficient markets. Amer-ican Economic Review, 70(3):393–408, 1980.Huseyin Gulen and Byoung-Hyoun Hwang. Daily stock market swings and investor reaction to firm-specificnews. Working paper, 2012.Stephen Hansen, Michael McMahon, and Andrea Prat. Transparency and deliberation within the fomc: acomputational linguistics approach. Quarterly Journal of Economics, Forthcoming, 2017.Samuel M Hartzmark and Kelly Shue. A tough act to follow: Contrast effects in financial markets. Journalof Finance, Forthcoming, 2017.Alexander Hillert, Heiko Jacobs, and Sebastian Müller. Media makes momentum. Review of FinancialStudies, 27:3467–3501, 2014.David Hirshleifer, Sonya Seongyeon Lim, and Siew Hong Teoh. Driven to distraction: Extraneous eventsand underreaction to earnings news. Journal of Finance, 64:2289–2325, 2009.Kewei Hou, Chen Xue, and Lu Zhang. Digesting anomalies: An investment approach. Review of FinancialStudies, 28(3):650–705, 2015.Jiekun Huang. The customer knows best: The investment value of consumer opinions. Journal of FinancialEconomics, Forthcoming, 2018.Gur Huberman and Tomer Regev. Contagious speculation and a cure for cancer: A nonevent that madestock prices soar. Journal of Finance, 56(1):387–396, 2001.Narasimhan Jegadeesh, Joonghyuk Kim, Susan D Krische, and Charles Lee. Analyzing the analysts: Whendo recommendations add value? Journal of Finance, 59(3):1083–1124, 2004.Travis L Johnson and Eric C So. Time will tell: Information in the timing of scheduled earnings news.Journal of Financial and Quantitative Analysis, Forthcoming, 2017.Marcin Kacperczyk, Stijn Van Nieuwerburgh, and Laura Veldkamp. A rational theory of mutual funds’attention allocation. Econometrica, 84(2):571–626, 2016.Ohad Kadan, Leonardo Madureira, Rong Wang, and Tzachi Zach. Conflicts of interest and stock recom-mendations: The effects of the global settlement and related regulations. Review of Financial Studies, 22(10):4189–4217, 2009.113Daniel Kahneman and Amos Tversky. Prospect theory: An analysis of decision under risk. In Handbook ofthe Fundamentals of Financial Decision Making: Part I, pages 99–127. 2013.Ron Kaniel and Robert Parham. Wsj category kings–the impact of media attention on consumer and mutualfund investment decisions. Journal of Financial Economics, 123(2):337–356, 2017.Jonathan M Karpoff. The relation between price changes and trading volume: A survey. Journal of Financialand Quantitative Analysis, 22(1):109–126, 1987.Elisabeth Kempf, Alberto Manconi, and Oliver Spalt. Distracted shareholders and corporate actions. Reviewof Financial Studies, 30(5):1660–1695, 2016.Peter Koudijs. Those who know most: Insider trading in eighteenth-century amsterdam. Journal of PoliticalEconomy, 123(6):1356–1409, 2015.Kenneth N. Kuttner. Monetary policy surprises and interest rates: Evidence from the fed funds futuresmarket. Journal of Monetary Economics, 47(3):523–544, 2001.Augustin Landier, David Sraer, and David Thesmar. Optimal dissent in organizations. Review of EconomicStudies, 76(2):761–794, 2009.Yann LeCun, Yoshua Bengio, and Geoffrey Hinton. Deep learning. Nature, 521(7553):436, 2015.Ulf Von Lilienfeld-Toal and Stefan Ruenzi. Ceo ownership, stock market performance, and managerialdiscretion. Journal of Finance, 69(3):1013–1050, 2014.Tim Liu, Christos Makridis, Paige Ouimet, and Elena Simintzi. Is cash still king: Why firms offer non-wagecompensation and the implications for shareholder value. Working paper, 2017.Yukun Liu and Ben Matthies. Long run risk: Is it there? Working paper, 2017.Michael Luca and Georgios Zervas. Fake it till you make it: Reputation, competition, and yelp review fraud.Management Science, 62(12):3412–3427, 2016.David O. Lucca and Emanuel Moench. The pre-fomc announcement drift. Journal of Finance, 70:329–371,2015.Christopher J Malloy. The geography of equity analysis. Journal of Finance, 60(2):719–755, 2005.Robert C. Merton. Theory of rational option pricing. The Bell Journal of Economics and ManagementScience, 4:141–183, 1973.Robert C. Merton. On estimating the expected return on the market: An exploratory investigation. Journalof Financial Economics, 8(4):323–361, 1980.Roni Michaely and Kent L Womack. Conflict of interest and the credibility of underwriter analyst recom-mendations. Review of Financial Studies, 12(4):653–686, 1999.114Roni Michaely, Amir Rubin, Dan Segal, and Alexander Vedrashko. Lured by the consensus: The implica-tions of treating all analysts as equal. Working paper, 2017.Sendhil Mullainathan and Andrei Shleifer. The market for news. American Economic Review, 95(4):1031–1053, 2005.Fulvio Ortu, Andrea Tamoni, and Claudio Tebaldi. Long-run risk and the persistence of consumption shocks.Review of Financial Studies, 26(11):2876–2915, 2013.L’uboš Pástor and Robert F Stambaugh. Liquidity risk and expected stock returns. Journal of PoliticalEconomy, 111(3):642–685, 2003.Andrew J Patton and Michela Verardo. Does beta move with news? firm-specific information flows andlearning about profitability. Review of Financial Studies, 25(9):2789–2839, 2012.Lin Peng and Wei Xiong. Investor attention, overconfidence and category learning. Journal of FinancialEconomics, 80(3):563–602, 2006.Joel Peress. The media and the diffusion of information in financial markets: Evidence from newspaperstrikes. Journal of Finance, 69:2007–2043, 2014.Richard Roll. R-squared. Journal of Finance, 43:541–566, 1988.Ronnie Sadka. Momentum and post-earnings-announcement drift anomalies: The role of liquidity risk.Journal of Financial Economics, 80(2):309–349, 2006.Pavel Savor and Mungo Wilson. How much do investors care about macroeconomic risk? evidence fromscheduled economic announcements. Journal of Financial and Quantitative Analysis, 48(02):343–375,2013.Pavel Savor and Mungo Wilson. Asset pricing: A tale of two days. Journal of Financial Economics, 113(2):171–201, 2014.Martin C Schmalz and Sergey Zhuk. Revealing downturns. Review of Financial Studies, 2018.Daniel Schmidt. Investors’ attention and stock covariation: Evidence from google sport searches. Technicalreport, 2013.Daniel Schmidt. Distracted institutional investors. Working paper, 2016.G William Schwert. Anomalies and market efficiency. Handbook of the Economics of Finance, 1:939–974,2003.William G. Schwert. Why does stock market volatility change over time? Journal of Finance, 44:1115–1153, 1989.Jinfei Sheng. Macro news, micro news, and stock prices. Working paper, 2017.115Nachum Sicherman, George Loewenstein, Duane J. Seppi, and Stephen Utkus. Financial attention. Reviewof Financial Studies, 29:863–897, 2016.Christopher A. Sims. Implications of rational inattention. Journal of Monetary Economics, 50(3):665–690,2003.Avanidhar Subrahmanyam and Sheridan Titman. The going-public decision and the development of financialmarkets. Journal of Finance, 54(3):1045–1082, 1999.Paul C. Tetlock. Giving content to investor sentiment: The role of media in the stock market. Journal ofFinance, 62:1139–1168, 2007.Paul C. Tetlock. Does public financial news resolve asymmetric information? Review of Financial Studies,23:3520–3557, 2010.Robert Tumarkin and Robert F Whitelaw. News or noise? internet postings and stock prices. FinancialAnalysts Journal, 57(3):41–51, 2001.Kent L Womack. Do brokerage analysts’ recommendations have investment value? Journal of Finance, 51(1):137–167, 1996.Yu Yuan. Market-wide attention, trading, and stock returns. Journal of Financial Economics, 116(3):548–564, 2015.116Appendix AAppendix to Chapter 2A. Information about GlassdoorIn this appendix, I provide detailed background information about Glassdoor. Compared to other employeereview websites such as Indeed, Vault, and CareerBliss, Glassdoor is not only the largest website for em-ployee reviews but also has the most diverse set of users, as discussed in Grennan (2013).Give to get policy: Glassdoor uses a “Give to get” policy to encourage users to post information. Areport by Glassdoor released in October 2017 shows that the “Give to get” policy helps reduce polarizationbias and encourages more neutral and balanced company ratings on Glassdoor.64Here is the full text of the “Give to get” policy: “Glassdoor’s give to get policy requires that you submita contribution in order to receive unlimited access to other content on our website. It ensures that we haveenough salaries, company reviews, interview reviews, & benefits reviews to share with the community. Itonly takes a minute to submit, and your contribution is anonymous. Your contribution will also help others,as their contributions will help you. When you contribute a review, salary, interview, or benefit reviewyou will be granted 12 months of unlimited access to our content. After that period you may be asked tocontribute another salary, review, interview or benefit review to extend your unlimited access for another 12months. If you are not ready to contribute, you can create a Glassdoor account without posting. You willhave full access to salary, review, interview, and benefit content for 10 days.”65Seasonality. One may be concerned about seasonality in the reviews. In particular, Glassdoor runsEmployees’ Choice Awards: “Best Places to Work” in each year. To be considered for the awards, severalminimum requirements for eligibility must be met. For instance, the Glassdoor 2017 Employees’ ChoiceAwards: “Best Places to Work” uses company reviews and ratings from current and former employeesbetween November 2, 2015 and October 30, 2016. It requires a firm to have at least 75 reviews/ratings, anoverall rating of at least 3.51, workplace factor ratings of at least 2.85 during the eligibility period, and atleast 1000 employees at the end of the eligibility timeframe. These awards have been in place for severalyears and the timeframe is similar from year to year. Thus, it is possible that companies give employeesincentives to write more good reviews so as to reach these requirements before the timeframe ends. If thisis the case, then one would expect to see more reviews close to the deadline.I first check whether there is seasonality in the number of reviews. Figure A.4 plots the fraction ofreviews in each month. The number of reviews is relatively greater in August, September, and October,which is consistent with the idea that employees tend to post more reviews leading up to the deadline for the64https://www.glassdoor.com/research/studies/give-to-get/65See http://help.glassdoor.com/article/Give-to-get-policy/en_US117Awards. However, the difference between months is not statistically significant. Even the difference betweenOctober (most reviews submitted) and December (least reviews submitted) is not statistically different fromzero (the t-value is 1.00). Even if the number of reviews is not statistically different, is it possible that thereview content is quite different? Table A.1 reports the mean values of employee outlook and star ratingsfor each month. Although there is some small variation among months, the differences are small and notsignificantly different from zero.Overall, despite concerns for potential issues, the quality of the Glassdoor data seems to be high. Thepolarization bias is less severe than other online reivew websites, and the distributions of the number ofreview and mean values of review variables across different months are stable.B. Machine learningB1. Using machine learning to extend the sample to 2008Although using a short-sample period for online data in finance is common, the sample period shouldbe extended for a robustness test. The outlook variable starts from 2012, while other variables (recommendthis company, overall ratings, 5 subcategory ratings) start from 2008. Because outlook is correlated withother variables (Table A.3), one can infer employees’ opinions of outlook based on their responses to otherquestions.The goal is to predict employee outlook in each review from June 2008 to February 2012. This problemis a classification exercise where each review is labeled as either positive or negative outlook in the end.66 Iuse machine learning, which is a method of data analysis that automates analytical model building to performthe prediction. The machine learning methods I use include k-nearest mean (KNN), logistic classificationwith L1 and L2, linear SVC, decision tree, random forest, and gradient boosted regression trees. I also usea deep learning method, which is an advanced machine learning method with hidden layers. The advantageof deep learning is that it allows more complex and non-linear functions between output and input variables.A detailed discussion of these methods can be found in Friedman et al. (2001) and LeCun et al. (2015).The machine learning procedure is as follows. I first use all reviews from March 2012 to December2016 as training and test samples. The split between training and test samples is random. Typically, 75%of observations are assigned as training samples (0.5 million reviews), and the remaining 25% are assignedas test samples. I also use various cross-validation with different splits of samples for robustness tests. Foreach method, the relationship between outlook and other variables is formed using training samples. Theout-of-sample test uses these relationships to predict outlook in test samples. Test samples also have the truevalues for employee outlook. The out-of-sample accuracy is calculated by comparing the predicted outlookand the true outlook. Each machine learning method has parameters that can affect its performance. Forexample, for KNN, the number of neighbors is important; for random forest, the number of trees and thedepth of each tree are important. To avoid overfitting, I use cross-validation to ensure the performance isrobust; only the robust performance is reported.Table A.4 presents the results of out-of-sample performance for different methods. Regular machine66For simplicity, I merge neutral and negative outlook into one class and just call it negative outlook.118learning methods such as KNN, logistic classification with L1 and L2, linear SVC, decision tree, randomforest, and gradient boosted regression trees have decent performance with accuracy scores of 0.76-0.77.Using deep learning significantly improves the out-of-sample performance, and can reach an accuracy scoreof 0.91. Deep learning performs better for two reasons. First, the training sample is relatively large with0.5 million observations, and deep learning often performs better with large data. Second, the deep learningmethod with a neural network is more likely to capture complex nonlinear functions, such as the underlyingrelationships between outlook and other variables.B2. Machine learning with LDAIn Section 2.5.3, I use LDA to detect the topics in review texts. This type of machine learning is called“Natural language processing” (NLP). As LDA here is an unsupervised learning, evaluating its performanceis challenging. One popular method is “word intrusion,” suggested by Chang et al. (2009). For each trainedtopic, it takes the first ten words and substitutes one with another, randomly chosen word (intruder!) and seeswhether a human can reliably tell which word is the intruder. If so, the trained topic is topically coherent;if not, the topic has no discernible theme. In the end, I find ten topics in LDA: business growth, personneldevelopment, leadership, salary, flexibility, benefits, career opportunities, work hours, culture, and workenvironment. The top five words for each topic based on reviews from all firms are reported in Table A.5.119Figure A.1: Review Form120Figure A.2: Salaries information121Figure A.3: Histogram of reviews by star ratingsThis figure presents the histogram of star ratings in employee reviews. The star rating is the mean of six numericalratings in each review. The curve shows the normal distribution.122Figure A.4: The distribution of reviews over different monthsThis figure plots the distribution of the number of reviews in each month. For each month, I calculate the total number of reviewsposted for S&P 1500 firms. The fraction of reviews is the number of reviews in each month divided by the total number of reviews.123Table A.1: Mean values of positive outlook and star ratings in each monthThis table presents the mean values of outlook and star ratings in each month. Positive outlook is the number ofreviews with positive outlook divided by the total number reviews in each month for each firm. Recommend is thenumber of reviews that state “recommend to a friend” divided by the total number of reviews in each month for eachfirm. Overall is the average overall star rating (on a scale of 1 to 5) in each month for each firm. The monthly meansof culture and values rating (Culture), work-life balance rating (WorkLife), senior management rating (Management),compensation and benefits (Compensation), and career opportunities (Career) are calculated in the same way. Thereported values are the mean of all of these variables in each month across all firms.Month Positive outlook Recommend Overall Culture WorkLife Management Compensation CareerJanuary 0.44 0.61 3.30 3.28 3.22 2.87 3.28 3.13February 0.41 0.59 3.25 3.23 3.20 2.83 3.25 3.07March 0.42 0.58 3.24 3.23 3.18 2.81 3.23 3.07April 0.42 0.59 3.25 3.23 3.18 2.80 3.24 3.07May 0.42 0.59 3.24 3.22 3.18 2.80 3.24 3.07June 0.43 0.60 3.27 3.26 3.19 2.81 3.23 3.10July 0.42 0.59 3.25 3.23 3.18 2.81 3.25 3.09August 0.42 0.59 3.25 3.23 3.18 2.81 3.25 3.09September 0.42 0.59 3.27 3.26 3.20 2.84 3.27 3.11October 0.43 0.60 3.29 3.27 3.21 2.86 3.29 3.12November 0.43 0.60 3.28 3.26 3.19 2.84 3.27 3.10December 0.44 0.60 3.28 3.27 3.20 2.86 3.27 3.12124Table A.2: Employee outlook and other components in reviewsThis table reports results of regression Equation (1), where the dependent variable is a dummy variable (one forpositive outlook) and independent variables include employee characteristics, star ratings, and variables constructedfrom texts. Current employee is equal to one if the reviewer is a current worker, and zero otherwise; Headquarterstate is equal to one if that reviewer works in the headquarter state, and zero otherwise; and Recommend is equal toone if a reviewer states “recommend to a friend”, and zero otherwise. Overall is the overall star rating (on a scaleof 1 to 5). Similarly, work-life balance rating (WorkLife), senior management rating (Management), compensationand benefits (Compensation), and career opportunities (Career) have subcategory ratings with a scale of 1 to 5. Theindependent variables in Panel B are number of words calculated from different sections of a review text. Numbersin parentheses are t-statistics that are adjusted for heteroscedasticity and clustered by firms. ***, **, and * indicatestatistical significance at the 1%, 5%, and 10% levels, respectively.Panel A. Employee characteristics and star ratingsVARIABLES (1) (2) (3) (4)Current employee 0.13***(34.60)Headquarter state 0.00(0.49)Recommend 0.51***(138.89)Overall 0.10***(79.71)Career 0.05***(66.11)Compensation 0.01***(10.73)Management 0.08***(56.92)WorkLife 0.00***(3.27)Observations 656,204 455,111 656,204 656,204Adjusted R2 0.082 0.063 0.294 0.351Panel B. Review textsVARIABLES (1) (2)# of words in review title -0.62***(-20.18)# of words in review pros 0.32***(51.85)# of words in reivew cons -0.16***(-48.66)# of words in reivew advice -0.17***(-31.83)# of words in the whole reivew -0.09***(-48.34)Observations 656,204 656,204Adjusted R2 0.117 0.086125Table A.3: Correlations and PCA analysisPanel A reports correlations between outlook and other variables from employee reviews. Outlook is the positiveoutlook defined as the number of reviews with positive outlook divided by the total number of reviews in each monthfor each firm. Recommend is the fraction of reviews that state “recommend to a friend” divided by the total numberof reviews in each month for each firm. Overall is the average overall star rating (on a scale of 1 to 5) in each monthfor each firm. The monthly mean of culture and values rating (Culture), work-life balance rating (WorkLife), seniormanagement rating (Management), compensation and benefits (Compensation), and career opportunities (Career).Panel B reports the principal component analysis (PCA) results for the top three components. Eigenvalues for eachcomponent are also reported.Panel A. CorrelationsOutlook Recommend Overall Culture WorkLife Management Compensation CareerOutlook 1.00Recommend 0.61 1.00Overall 0.62 0.75 1.00Culture 0.57 0.68 0.78 1.00WorkLife 0.41 0.52 0.62 0.59 1.00Management 0.59 0.67 0.77 0.75 0.59 1.00Compensation 0.42 0.49 0.61 0.51 0.44 0.51 1.00Career 0.55 0.62 0.74 0.65 0.48 0.67 0.57 1.00Panel B. PCAComponent 1 Component 2 Component 3Outlook 0.64 -0.28 -0.15Recommend 0.43 0.07 -0.03Overall 0.32 0.19 0.16Culture 0.30 0.30 0.00WorkLife -0.08 0.85 -0.04Management 0.33 0.25 0.01Compensation -0.05 -0.03 0.91Career 0.30 -0.04 0.35Eigenvalues 5.24 0.62 0.60% Variance Explained 0.66 0.08 0.08126Table A.4: Machine learning methods and their performanceThis table presents the out-of-sample accuracy of different machine learning methods. For each method, the rela-tionships between outlook and other variables are formed using training samples. The out-of-sample test uses theserelationships to predict outlook in test samples. Test samples also have the true value for the outlook. The accuracy iscalculated by comparing the predicted outlook and the true outlook.Method AccuracyKNN 0.76Logistic (L1) 0.77Logistic (L2) 0.77Linear SVC 0.76Decision tree 0.77Random forest 0.76Gradient boosted trees 0.77Deep learning 0.91Table A.5: Top five words in each topic from LDAThis table presents top five words of 10 topics from LDA method.Topics Top five wordsbusiness growth business, growth, development, industry, opportunitypersonnel development coworkers, training, learn, friendly, teamleadership leadership, supervisors, listen, upper-level, communicationsalary salary, paid, advancement, compensation, moneyflexibility flexible, balance, work, home, familybenefits vacation, health, insurance, 401k, holidayscareer opportunities career, growth, opportunity, promotions, movework hours shift, busy, hour, workers, workloadculture performance, goals, clear, focus, expectationswork environment treat, respect, care, hr, hire127Appendix BAppendix to Chapter 3A. Details about Macro Announcements1. FOMC: The Federal Open Market Committee (FOMC) is the policy-making arm of the Federal Reserve.It determines short-term interest rates in the U.S. when it decides the overnight rate that banks pay eachother for borrowing reserves when a bank has a shortfall in required reserves. The Fed announces its policydecision at the end of each FOMC meeting. This is the FOMC announcement, which happens eight timesa year. The announcement also includes brief comments on the FOMC’s views on the economy and howmany FOMC members voted for and how many voted against the policy decision.2. Employment situation: Employment situation include non-farm payroll, unemployment rate, averageworkweek, and average hourly earnings. The data is released monthly, usually on the first Friday of themonth, by Bureau of Labor Statistics, U.S. Department of Labor.3. ISM PMI: ISM manufacturing index is a diffusion index calculated from five of the eleven sub-components of a monthly survey of purchasing managers at roughly 300 manufacturing firms nationwide.It is a leading indicator of output.4. Personal Consumption: Personal consumption expenditures are the monthly analogues to the quar-terly consumption expenditures in the GDP report, available in nominal and real (inflation-adjusted) dollars.B. The impact of macro news on market risk premiumThis section tests the impact of individual macro news on market risk premium. I find important macroannouncements for stock markets by running the following regression over a sample period of January 1997to December 2014.Mktt = γ0+ γ1Macrodayt + γ2Mktt−1+ γ3(Mktt−1)2+ etwhere Mktt is the CRSP value-weighted market return minus the risk-free rate. Macrodayt is a dummyvariable equaling 1 if day t is an announcement day for a specific type of macro news, and 0 otherwise. Forexample, if my focus is on ISM PMI, then Macroday equals 1 if that day has an ISM announcement, and 0otherwise. I also include dummy variables for the day of week.Due to limited space, I only listed macro announcements that have statistically and economically sig-nificant impact on market risk premium. Table B.1 presents results for macro announcements that havestatistically and economically significant impact on market risk premium. Columns (1)-(2) show the resultsfor FOMC news. Column (1) is parsimonious specification without including any control variables. Thecoefficient on Macroday is positive and significant, suggesting that the market risk premium is higher on128FOMC days than other days. I include the market excess return lagged 1 day, squared market return andthe day of week as control in column (3). The Macroday effect remains positive and highly significant inall specifications. Columns (3)-(8) show similar macro-day effects for announcements of Nonfarm Payroll,ISM PMI, and Personal Consumption. Columns (9)-(10) show results on all of these four macro announce-ments. The coefficient on Macroday is also positive and significant. Overall, Table B.1 shows that thesefour important macro announcements are market-moving indicators and therefore investors care about thesetypes of macro news.C. Additional resultsThis section provides additional result for the volume reaction. Instead of using the measure used in DellaV-igna and Pollet (2009), I now use the measure in Hirshleifer et al. (2009), which is calculated this way:AVOL[ j] = log(Vt+ j +1)− 130t−11∑k=t−40log(Vk +1)where Vt + j is the dollar value of trading volume at t + j. Immediate abnormal volume response is overa 2-day window (AVOL[0,1]) and defined as the average of abnormal trading volumes on the earningsannouncement date (AVOL[0]) and on the following day (AVOL[1]).Table B.2 reports the result, which issimilar to that of Table 3.10.129Table B.1: Characteristics of Macroeconomic AnnouncementsThis table presents four important macroeconomic announcements used in analysis: days with announcements ofFederal Open Market Committee (FOMC) decision, Employment situation, ISM PMI, or personal consumption. Therelease time is Eastern Time. The sample covers January 1997 to December 2014.Announcement Source Frequency Unit/Type # of eventsFederal Funds Rate FOMC 8/yr % level 144Employement situation BLS M K, change 216ISM PMI ISM M index 216Personal consumption BEA M % change 216Table B.2: Volume reaction with an alternative measureThis table tests whether stock volume response to earnings news is different on macro-news days. The sample coversJanuary 1997 to December 2014. The dependent variables are two measures of abnormal trading volume. AES isabsolute earnings surprise quantile, and Macroday is a dummy variable equaling 1 if day t is an announcement dayfor Federal Open Market Committee (FOMC) decision, Employment situation, ISM PMI, or personal consumption.Following Hirshleifer, Lim, and Teoh (2009), I define abnormal trading on earnings announcement day AVOL[0]as the difference between log dollar volume on day 0 and the average log dollar volume over days [-40,-11]. Asimilar definition applies to the abnormal trading volume on the following day AVOL[1]. AVOL[0,1] is the averageof AVOL[0] and AVOL[1]. Control variables include the number of earnings announcements, the number of analystsfollowing the firm, analyst dispersion, market capitalization, market return, and dummy variables for year, month, andday of week. Standard errors are adjusted for heteroscedasticity and clustered by the day of earnings announcement.***, **, and * indicate statistical significance at the 1%, 5%, and 10% levels, respectively.(1) (2)Macroday 0.070*** 0.049***(0.017) (0.016)AES 0.019*** 0.017***(0.001) (0.001)Constant 0.526*** 0.745***(0.008) (0.024)Controls No YesObservations 158,018 158,018Adj. R2 0.006 0.170130Table B.3: Macro announcements and market risk premiumThis table reports the results of OLS regressions of daily stock market excess return on a macro announcement day (Macroday) dummy variable and control variables. The samplecovers January 1997 to December 2014. The dependent variable MKT is the CRSP value-weighted market return minus the risk-free rate. Macro-day is a dummy variable equaling 1if day t is an announcement day for FOMC, Employment situation, ISM PMI, Personal Consumption, and all these four respectively, and 0 otherwise. Monday-Thursday are dummyvariables for the corresponding days of the week. ***, **, and * indicate statistical significance at the 1%, 5%, and 10% levels, respectively.(1) (2) (3) (4) (5) (6) (7) (8) (9) (10)FOMC Employment ISM PMI Personal consumption All 4 newsMacroday 0.25** 0.23** 0.14* 0.18** 0.27*** 0.27*** 0.21** 0.21** 0.25*** 0.26***(2.56) (2.36) (1.82) (2.17) (3.67) (3.59) (2.29) (2.30) (5.63) (5.63)MKTEXRETt-1 0.01 0.01 0.01 0.01 0.01(0.52) (0.55) (0.59) (0.67) (0.55)(MKTEXRETt-1)2 39.00 38.60 46.57 18.17 40.28(0.45) (0.44) (0.53) (0.21) (0.46)Monday -0.01 0.03 -0.02 0.00 0.04(-0.13) (0.61) (-0.46) (0.07) (0.79)Tuesday -0.02 0.04 -0.01 0.00 0.04(-0.39) (0.68) (-0.13) (0.09) (0.75)Wednesday 0.02 0.08 0.04 0.03 0.08(0.45) (1.45) (0.76) (0.60) (1.62)Thursday -0.00 0.04 -0.00 -0.01 0.05(-0.05) (0.70) (-0.02) (-0.13) (0.90)Constant 0.03 0.02 0.03 -0.01 0.02 0.01 0.02 -0.05(1.58) (0.68) (1.59) (-0.36) (1.18) (0.40) (0.46) (-1.24)Observations 4,357 4,289 4,357 4,289 4,357 4,289 4,289 4,289Adj. R-squared 0.10% 0.15% 0.20% 0.90% 0.30% 1.60% 1.10% 2.60%131Appendix CAppendix to Chapter 4Figure C.1: Media Attention and Macroeconomic FundamentalsThis figure shows monthly media attention indices for the New York Times (MAI-NYT), the Wall Street Journal (MAI-WSJ),the demeaned composite index (MAI-C1), and the demeaned and standardized composite index (MAI-C2) against their relatedmacroeconomic fundamentals. All figures are at the monthly frequency except for the GDP MAI - Real GDP, which is at thequarterly frequency. Blue lines are macroeconomic attention indices (left y-axis) and dotted red lines (right y-axis) are MAI relatedmacroeconomic fundamentals (see Table 4.2). MAI units are in percentage. The gray vertical bars are NBER recessions. Thesample period for MAI-NYT, MAI-C1, and MAI-C2 is June 1, 1980 to December 31, 2016 and MAI-WSJ is January 1, 1984 toDecember 31, 2016.1981 1985 1989 1993 1997 2001 2005 2009 2013 20170.00.51.01.52.02.53.0CreditratingMAIMAI-NYT1981 1985 1989 1993 1997 2001 2005 2009 2013 20170.00.51.01.52.02.53.03.54.0MAI-WSJ1981 1985 1989 1993 1997 2001 2005 2009 2013 20170.00.51.01.5MAI-C11981 1985 1989 1993 1997 2001 2005 2009 2013 2017−0.50.00.51.01.52.02.53.03.5MAI-C2Credit rating MAI Corporate relative spread10203040506070102030405060701020304050607010203040506070Corporaterelativespread1981 1985 1989 1993 1997 2001 2005 2009 2013 20170.40.60.81.01.2GDPMAIMAI-NYT1981 1985 1989 1993 1997 2001 2005 2009 2013 20172.02.53.03.54.04.55.05.5MAI-WSJ1981 1985 1989 1993 1997 2001 2005 2009 2013 2017−0.75−0.50−0.250.000.250.500.751.001.25MAI-C11981 1985 1989 1993 1997 2001 2005 2009 2013 2017−1.5−1.0−0.50.00.51.01.52.0MAI-C2GDP MAI Real GDP quarterly growth rate−2−1012−2−1012−2−1012−2−1012RealGDPquarterlygrowthrate(%)1981 1985 1989 1993 1997 2001 2005 2009 2013 20170.00.51.01.52.02.5InflationMAIMAI-NYT1981 1985 1989 1993 1997 2001 2005 2009 2013 20171234567MAI-WSJ1981 1985 1989 1993 1997 2001 2005 2009 2013 2017−1.0−0.50.00.51.01.52.0MAI-C11981 1985 1989 1993 1997 2001 2005 2009 2013 2017−0.50.00.51.01.52.02.5MAI-C2Inflation MAI Change in CPI−1.0−0.50.00.51.0−1.0−0.50.00.51.0−1.0−0.50.00.51.0−1.0−0.50.00.51.0ChangeinCPI(%)132Figure C.1: Media Attention and Macroeconomic Fundamentals (cont.)1981 1985 1989 1993 1997 2001 2005 2009 2013 20170.00.20.40.60.8InterestMAIMAI-NYT1981 1985 1989 1993 1997 2001 2005 2009 2013 20170.00.51.01.52.02.53.03.54.0MAI-WSJ1981 1985 1989 1993 1997 2001 2005 2009 2013 2017−0.50.00.51.01.5MAI-C11981 1985 1989 1993 1997 2001 2005 2009 2013 2017−0.50.00.51.01.5MAI-C2Interest MAI Fed funds rate0.02.55.07.510.012.515.017.520.00.02.55.07.510.012.515.017.520.00.02.55.07.510.012.515.017.520.00.02.55.07.510.012.515.017.520.0Fedfundsrate(%)123456MonetaryMAI1982 1986 1990 1994 1998 2002 2006 2010 2014 20180.02.55.07.510.012.515.017.520.01.01.52.02.53.03.54.04.5MonetaryMAIMAI-WSJ0.51.01.52.01982 1986 1990 1994 1998 2002 2006 2010 2014 20180.02.55.07.510.012.515.017.520.01.01.52.02.53.03.54.04.5MAI-NYT−1.0−0.50.00.51.01.52.01982 1986 1990 1994 1998 2002 2006 2010 2014 20180.02.55.07.510.012.515.017.520.01.01.52.02.53.03.54.04.5MAI-C1−0.50.00.51.01.51982 1986 1990 1994 1998 2002 2006 2010 2014 20180.02.55.07.510.012.515.017.520.01.01.52.02.53.03.54.04.5FedFundsRate(first)BalanceSheet(second)MAI-C2Monetary MAI Fed Fund Rate Fed Assets1981 1985 1989 1993 1997 2001 2005 2009 2013 20170.00.20.40.60.81.01.21.41.6HousingMAIMAI-NYT1981 1985 1989 1993 1997 2001 2005 2009 2013 20170123456MAI-WSJ1981 1985 1989 1993 1997 2001 2005 2009 2013 2017−0.50.00.51.01.52.02.53.03.5MAI-C11981 1985 1989 1993 1997 2001 2005 2009 2013 2017−0.50.00.51.01.52.02.53.0MAI-C2Housing MAI Log nominal home price return−1.0−0.50.00.51.0−1.0−0.50.00.51.0−1.0−0.50.00.51.0−1.0−0.50.00.51.0Loghomepricereturn133Figure C.1: Media Attention and Macroeconomic Fundamentals (cont.)1981 1985 1989 1993 1997 2001 2005 2009 2013 201701234OilMAIMAI-NYT1981 1985 1989 1993 1997 2001 2005 2009 2013 20172468MAI-WSJ1981 1985 1989 1993 1997 2001 2005 2009 2013 2017−101234MAI-C11981 1985 1989 1993 1997 2001 2005 2009 2013 2017−10123MAI-C2Oil MAI Oil log price2.53.03.54.04.55.02.53.03.54.04.55.02.53.03.54.04.55.02.53.03.54.04.55.0Oillogprice1981 1985 1989 1993 1997 2001 2005 2009 2013 20170.00.51.01.52.02.5UnemploymentMAIMAI-NYT1981 1985 1989 1993 1997 2001 2005 2009 2013 201712345MAI-WSJ1981 1985 1989 1993 1997 2001 2005 2009 2013 2017−1.0−0.50.00.51.01.52.02.5MAI-C11981 1985 1989 1993 1997 2001 2005 2009 2013 2017−0.50.00.51.01.52.02.5MAI-C2Unemployment MAI Unemployment rate4567891011456789101145678910114567891011Unemploymentrate(%)1981 1985 1989 1993 1997 2001 2005 2009 2013 20170.00.10.20.30.4USDMAIMAI-NYT1981 1985 1989 1993 1997 2001 2005 2009 2013 20170.00.51.01.52.02.53.03.5MAI-WSJ1981 1985 1989 1993 1997 2001 2005 2009 2013 2017−0.50−0.250.000.250.500.751.001.25MAI-C11981 1985 1989 1993 1997 2001 2005 2009 2013 2017−0.50−0.250.000.250.500.751.001.25MAI-C2USD MAI USD log price index4.24.34.44.54.64.74.84.95.04.24.34.44.54.64.74.84.95.04.24.34.44.54.64.74.84.95.04.24.34.44.54.64.74.84.95.0USDlogpriceindex134Figure C.2: Macroeconomic Attention around Macroeconomic AnnouncementsThis figure shows lag and forward estimated coefficients βδ from OLS regressions of detrended macroeconomic attention indicesMAI-C1 on announcement dummies as specified in Equation (4.4). The shaded area corresponds to the 95% confidence intervalaround estimated coefficients. The x-axis is the number days since the announcement. The announcements are the Producer PriceIndex (PPI). The vertical line represents the day of the announcement. The PPI announcement dates are from the U.S. BureauLabour of Statistics.−4 −2 0 2 4−0.20.00.20.40.6ResponseInflation MAI around CPI−4 −2 0 2 4Monetary MAI around CPI−4 −2 0 2 4Oil MAI around CPI135Figure C.3: Detrended VIX before Employment Situation and FOMC announcementsThis figure shows the average 250-day detrended VIX five days before the Employment Situation announcement in Panel A andthe FOMC announcement in Panel B until the announcement day (day 0).−0.3−0.2−0.10.0DetrendedVIXPanel A: Employment Situation Announcements−5 −4 −3 −2 −1 0Days before the announcement−0.10.00.10.20.30.4DetrendedVIXPanel B: FOMC Announcements136Table C.1: Descriptive StatisticsPanel A of this table shows the correlation between the daily demeaned and standardized macroeconomic attention composite indices (MAI-C2), the Economic Policy Uncertainty (EPU) index, the implied volatility (VIX), and the 60-day detrended S&P 500 log trade volume.Panel B shows the correlation at the monthly frequency.Panel A: Daily MAI-C2 correlationCredit Rating GDP Housing Inflation Interest Monetary Oil Unemployment U.S. Dollar EPU VIX VolumeCredit Rating 1.00 0.12 0.14 -0.05 0.06 0.12 0.04 0.12 0.09 0.15 0.24 0.26GDP 0.12 1.00 0.12 0.14 0.14 0.21 0.15 0.21 0.11 0.10 0.09 0.16Housing 0.14 0.12 1.00 0.03 0.17 0.21 0.05 0.15 0.02 0.06 0.08 0.39Inflation -0.05 0.14 0.03 1.00 0.35 0.44 0.39 0.29 0.03 0.06 -0.01 -0.35Interest 0.06 0.14 0.17 0.35 1.00 0.51 0.26 0.17 0.11 0.11 0.14 0.02Monetary 0.12 0.21 0.21 0.44 0.51 1.00 0.24 0.29 0.11 0.18 0.18 0.10Oil 0.04 0.15 0.05 0.39 0.26 0.24 1.00 0.10 0.11 0.06 0.09 -0.18Unemployment 0.12 0.21 0.15 0.29 0.17 0.29 0.10 1.00 -0.04 0.25 0.25 0.07U.S. Dollar 0.09 0.11 0.02 0.03 0.11 0.11 0.11 -0.04 1.00 -0.01 0.24 0.07EPU 0.15 0.10 0.06 0.06 0.11 0.18 0.06 0.25 -0.01 1.00 0.34 0.05VIX 0.24 0.09 0.08 -0.01 0.14 0.18 0.09 0.25 0.24 0.34 1.00 0.21Volume 0.26 0.16 0.39 -0.35 0.02 0.10 -0.18 0.07 0.07 0.05 0.21 1.00Panel B: Monthly MAI-C2 correlationCredit Rating GDP Housing Inflation Interest Monetary Oil Unemployment U.S. Dollar EPU VIX VolumeCredit Rating 1.00 0.46 0.27 -0.23 0.15 0.26 0.01 0.25 0.18 0.39 0.24 0.04GDP 0.46 1.00 0.26 -0.05 0.28 0.38 0.18 0.37 0.12 0.19 0.12 -0.01Housing 0.27 0.26 1.00 -0.09 0.31 0.38 0.02 0.20 -0.02 0.15 0.08 0.06Inflation -0.23 -0.05 -0.09 1.00 0.43 0.41 0.59 0.20 -0.12 -0.05 0.00 0.09Interest 0.15 0.28 0.31 0.43 1.00 0.69 0.50 0.17 0.22 0.24 0.05 0.09Monetary 0.26 0.38 0.38 0.41 0.69 1.00 0.37 0.38 0.11 0.32 0.20 0.12Oil 0.01 0.18 0.02 0.59 0.50 0.37 1.00 0.09 0.14 0.10 0.06 0.09Unemployment 0.25 0.37 0.20 0.20 0.17 0.38 0.09 1.00 -0.21 0.43 0.32 0.00U.S. Dollar 0.18 0.12 -0.02 -0.12 0.22 0.11 0.14 -0.21 1.00 0.40 -0.06 0.04EPU 0.39 0.19 0.15 -0.05 0.24 0.32 0.10 0.43 0.40 1.00 0.36 0.08VIX 0.24 0.12 0.08 0.00 0.05 0.20 0.06 0.32 -0.06 0.36 1.00 -0.01Volume 0.04 -0.01 0.06 0.09 0.09 0.12 0.09 0.00 0.04 0.08 -0.01 1.00137Table C.2: Persistence of Macroeconomic AttentionPanel A of this table presents AR(p) models of the monthly demeaned and standardized macroeconomicattention composite indices (MAI-C2), controlling for monthly time-fixed effects. DF (p-value) are p-values for the Dickey-Fuller statistics that test the null of a unit root in each time series. Panel B reportsestimates from an OLS regression of the MAI-C2 on various moving average lags of itself. L1 correspondsto the lag of itself and L5, L21, L62, L250, and L1000 are the moving average for 5, 21, 62, 250, and 1000days preceding the observed values at time t. We control for day-of-week fixed effects. The standard errorsare reported in parenthesis and are calculated using Newey-West standard errors (10 lags). *, **, and ***denote the statistical significance at the 10%, 5%, 1% levels, respectively.Panel A: Monthly MAI-C2 AR(4) Coefficients and DF statisticsCredit Rating GDP Housing Inflation Interest Monetary Oil Unemployment U.S. Dollarconst 0.02 0.06* -0.00 0.09*** 0.03 0.09** 0.13** 0.01 -0.03(0.04) (0.04) (0.04) (0.03) (0.03) (0.04) (0.05) (0.04) (0.03)AR(1) 0.65*** 0.32*** 0.60*** 0.49*** 0.54*** 0.47*** 0.66*** 0.66*** 0.53***(0.07) (0.05) (0.10) (0.05) (0.05) (0.04) (0.05) (0.07) (0.06)AR(2) 0.02 0.25*** 0.09 0.24*** 0.16*** 0.15*** 0.18*** 0.13** 0.21***(0.07) (0.05) (0.08) (0.04) (0.06) (0.05) (0.04) (0.06) (0.05)AR(3) 0.05 0.23*** 0.15 0.09* -0.07 0.12* 0.07 0.11** 0.11**(0.05) (0.05) (0.09) (0.05) (0.05) (0.06) (0.10) (0.05) (0.05)AR(4) 0.09 0.06 0.03 0.09** 0.16*** 0.03 -0.01 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 0.01 0.02 0.00 0.00 0.00 0.00 0.00 0.07Obs. 435 435 435 435 435 435 435 435 435Adj-R2 0.54 0.55 0.64 0.76 0.50 0.45 0.75 0.78 0.77Panel B: Daily MAI-C2 Frequency RegressionsCredit Rating GDP Housing Inflation Interest Monetary Oil Unemployment U.S. DollarIntercept -0.15*** 0.00 -0.21*** -0.01 -0.10*** -0.19*** -0.19*** -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* 0.07*** 0.13*** 0.19*** 0.11*** 0.04** 0.01(0.02) (0.01) (0.02) (0.01) (0.02) (0.02) (0.02) (0.02) (0.01)L5 0.28*** 0.11*** 0.45*** 0.12*** 0.14*** 0.18*** 0.39*** 0.22*** 0.15***(0.05) (0.03) (0.07) (0.03) (0.03) (0.03) (0.04) (0.03) (0.03)L21 0.40*** 0.08 0.23*** 0.26*** 0.31*** 0.21*** 0.29*** 0.25*** 0.39***(0.09) (0.06) (0.08) (0.06) (0.06) (0.05) (0.06) (0.06) (0.06)L62 0.07 0.32*** 0.06 0.36*** 0.12* 0.14** 0.14*** 0.27*** 0.30***(0.06) (0.10) (0.07) (0.07) (0.07) (0.07) (0.05) (0.08) (0.07)L250 0.08 0.41*** 0.17** 0.08 0.20*** 0.20*** 0.01 0.24*** 0.13**(0.06) (0.10) (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 0.01 0.03 -0.08*** -0.02(0.05) (0.06) (0.06) (0.04) (0.04) (0.05) (0.02) (0.03) (0.03)Obs. 8545 8545 8545 8545 8545 8545 8545 8545 8545Adj-R2 0.28 0.17 0.41 0.20 0.18 0.24 0.51 0.35 0.32138Table C.3: Macroeconomic Attention and Macroeconomic FundamentalsThis table presents results of OLS regressions of monthly macroeconomic attention indices MAI-C2, on different macroeconomic fundamentals. The general regression is specified in Equation(4.6). F corresponds to the associated fundamental to each MAI as described in Table 4.2 and Ft is the moving average over t days of the respective macroeconomic fundamental. All observationsare at the monthly frequency except for the GDP MAI - Real GDP growth, which is at the quarterly frequency. We control for monthly fixed effects. The standard errors are reported in parenthesisand are calculated using Newey-West standard errors (10 lags). *, **, *** denote the statistical significance at the 10%, 5%, 1% levels, respectively.MAI: Credit Rating GDP Housing Inflation Interest Monetary Oil Unemployment U.S. 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 0.047** -0.321* -0.225*** -0.051 -0.013 -0.006* -0.011 0.001(0.022) (0.175) (0.085) (0.048) (0.043) (0.003) (0.168) (0.006)Ft,t−3−Ft,t−12 0.010 5.911 -0.497*** -0.408** -0.001 -0.016 0.002 0.182 -0.009(0.008) (6.162) (0.163) (0.195) (0.031) (0.032) (0.008) (0.115) (0.018)Ft,t−12−Ft,t−48 -0.006 9.655 -0.059 1.759** -0.007 -0.015 0.053* 0.166*** -0.209***(0.021) (18.738) (0.178) (0.697) (0.023) (0.025) (0.028) (0.047) (0.066)(Ft −Ft,t−3)2 -0.000 0.709*** -0.439** 0.050** 0.038** 0.002*** 0.946 0.006(0.003) (0.227) (0.186) (0.021) (0.016) (0.001) (0.795) (0.004)(Ft,t−3−Ft,t−12)2 0.000 7.194 0.456*** -0.064 0.031* 0.050*** 0.003*** 0.240** 0.008(0.001) (6.731) (0.135) (0.182) (0.017) (0.015) (0.001) (0.120) (0.011)(Ft,t−12−Ft,t−48)2 0.002* 10.457 1.163*** 9.918*** 0.014** -0.001 -0.003 0.072*** 0.087(0.001) (20.172) (0.339) (1.939) (0.006) (0.008) (0.006) (0.023) (0.071)Intercept -0.043 -4.011 -0.435*** -0.096 -0.039 0.010 -0.209** -0.052 -0.084(0.056) (11.200) (0.057) (0.067) (0.061) (0.070) (0.084) (0.069) (0.077)Obs. 439 145 439 439 439 439 396 439 439Adj-R2 0.08 -0.00 0.47 0.23 0.11 0.07 0.22 0.54 0.08139Table C.4: Macroeconomic Attention and Aggregate Trade VolumeThis table presents results of OLS regressions of the daily detrended S&P 500 log trade volume on thedifference between the 5-day and 20-day moving average of the demeaned and standardized media attentionindex (MAI-C2) and a dummy (Ann) equal to one if there is a related announcement specified in Table 4.2,and zero otherwise. We detrend the log trade volume using the moving average of the log trade volume ofthe past 60 trading days. For all model specifications, we control for day-of-week fixed effects. The standarderrors are reported in parenthesis and are calculated using Newey-West standard errors (250 lags). *, **,*** denote the statistical significance at the 10%, 5%, 1% levels, respectively.MAI: Inflation Monetary InterestAnn: CPI and PPI FOMC FOMC(1) (2) (3) (4) (5) (6) (7) (8) (9)MAI5−20 0.061*** 0.060*** 0.068*** 0.084*** 0.083*** 0.083*** 0.051*** 0.050*** 0.051***(0.013) (0.013) (0.014) (0.010) (0.010) (0.010) (0.011) (0.011) (0.011)Ann 0.024*** 0.026*** 0.036*** 0.036** 0.039*** 0.045***(0.006) (0.006) (0.012) (0.014) (0.012) (0.012)MAI5−20×Ann -0.079*** -0.006 -0.086*(0.022) (0.061) (0.047)Intercept 0.004 -0.001 -0.001 0.003 0.003 0.003 0.004 0.004 0.004(0.006) (0.006) (0.006) (0.006) (0.006) (0.006) (0.006) (0.006) (0.006)Obs. 9208 9208 9208 9208 9208 9208 9208 9208 9208Adj-R2 0.05 0.06 0.06 0.07 0.07 0.07 0.05 0.05 0.05MAI: GDP Unemployment Credit Rating Oil USDAnn: GDP Report Employment(1) (2) (3) (4) (5) (6) (7) (8) (9)MAI5−20 0.043*** 0.043*** 0.043*** 0.039*** 0.038*** 0.040*** 0.044*** 0.056*** 0.051***(0.011) (0.011) (0.011) (0.012) (0.012) (0.012) (0.013) (0.014) (0.015)Ann 0.008 0.008 0.013 0.017(0.007) (0.007) (0.010) (0.012)MAI5−20×Ann -0.010 -0.036(0.028) (0.038)Intercept 0.003 0.003 0.003 0.004 0.001 0.001 0.004 0.004 0.004(0.006) (0.006) (0.006) (0.006) (0.007) (0.007) (0.006) (0.006) (0.006)Obs. 9208 9208 9208 9208 9208 9208 9208 9208 9208Adj-R2 0.05 0.05 0.05 0.05 0.05 0.05 0.05 0.05 0.05140Table C.5: Media Attention and Implied VolatilityThis table presents results of OLS regressions of the daily implied volatility proxied by VIX regressed onthe difference between the 20-day and 250-day moving average demeaned and standardized index (MAI-C2) and a dummy (Ann) equal to one if there is a related announcement specified in Table 4.2, and zerootherwise. For all model specifications, we control for day-of-week fixed effects. The standard errors arereported in parenthesis and are calculated using Newey-West standard errors (250 lags). *, **, *** denotethe statistical significance at the 10%, 5%, 1% levels, respectively.MAI: Inflation Monetary InterestAnn: CPI and PPI FOMC FOMC(1) (2) (3) (4) (5) (6) (7) (8) (9)MAI20−250 -4.515 -4.513 -4.509 4.706** 4.706** 4.739** 3.978* 3.979* 4.005*(4.670) (4.671) (4.614) (2.105) (2.105) (2.098) (2.311) (2.311) (2.302)Ann 0.086 0.086 0.325 0.319 0.347 0.337(0.105) (0.104) (0.424) (0.425) (0.439) (0.441)MAI20−250×Ann -0.038 -1.241 -0.932(0.954) (0.884) (0.983)Intercept 19.490*** 19.477*** 19.477*** 19.551*** 19.551*** 19.551*** 19.553*** 19.552*** 19.552***(1.125) (1.125) (1.124) (1.170) (1.170) (1.170) (1.174) (1.174) (1.174)Obs. 6802 6802 6802 6802 6802 6802 6802 6802 6802Adj-R2 0.02 0.02 0.02 0.03 0.03 0.03 0.02 0.02 0.02MAI: GDP Unemployment Credit Rating Oil USDAnn: GDP Report Employment(1) (2) (3) (4) (5) (6) (7) (8) (9)MAI20−250 11.047** 11.048** 11.068** 13.309*** 13.311*** 13.374*** 5.120*** 1.619 3.151*(4.688) (4.689) (4.686) (4.930) (4.930) (4.939) (1.575) (1.936) (1.627)Ann 0.071 0.067 0.235 0.230(0.143) (0.144) (0.159) (0.170)MAI20−250×Ann -0.406 -1.297(0.780) (0.881)Intercept 19.522*** 19.516*** 19.516*** 19.464*** 19.411*** 19.411*** 19.543*** 19.543*** 19.561***(1.068) (1.065) (1.065) (0.962) (0.962) (0.961) (1.131) (1.163) (1.159)Obs. 6802 6802 6802 6802 6802 6802 6802 6802 6802Adj-R2 0.08 0.08 0.08 0.20 0.20 0.20 0.06 0.01 0.01141Table C.6: Forecasting Unemployment SurprisesThis table presents results of OLS regressions of the unemployment surprise on the one-day lag detrendeddaily MAI for unemployment at different frequencies and an interaction term between MAI and a NBERdummy. For example, MAI5−20 is the difference between the five-day and the twenty-day moving averageof MAI for unemployment. The NBER dummy is equal to one if the unemployment surprise occurs duringa NBER recession, zero otherwise. The surprise is calculated as the difference between the actual unem-ployment for the month t reported in the month t + 1 and the previous month unemployment rate (i.e., therandom-walk), the average unemployment rate from economist surveys in Bloomberg, and the forecastedunemployment rate as in Boyd et al. (2005) (BHJ). Panel A, B, and C shows the results for MAI-WSJ, MAI-NYT, and MAI-C2, respectively. The standard errors are reported in parenthesis and are calculated usingNewey-West (5 lags). *, **, *** denote the statistical significance at the 10%, 5%, 1% levels, respectively.The sample period is June 1, 1980 to December 31, 2016 for the random-walk and BHJ surprise and fromJanuary 1, 1997 to December 31, 2016 for the Bloomberg surprise.Panel A: MAI-NYTRandom-walk surpriseMAI: MAI5−20 MAI20−250 MAI60−250(1) (2) (3) (4) (5) (6)MAI -0.008 -0.008 0.186*** 0.131*** 0.296*** 0.180***(0.035) (0.033) (0.046) (0.042) (0.073) (0.066)MAI×NBER 0.005 0.226** 0.503***(0.165) (0.106) (0.125)Intercept -0.007 -0.007 -0.004 -0.011 -0.004 -0.014*(0.012) (0.012) (0.010) (0.009) (0.010) (0.009)Obs. 438 438 427 427 427 427R2 0.00 0.00 0.06 0.08 0.09 0.13BHJ surpriseMAI: MAI5−20 MAI20−250 MAI60−250(1) (2) (3) (4) (5) (6)MAI 0.006 0.006 0.093*** 0.099*** 0.164*** 0.129**(0.033) (0.033) (0.035) (0.038) (0.052) (0.059)MAI×NBER 0.008 -0.026 0.153(0.137) (0.102) (0.122)Intercept -0.029*** -0.029*** -0.027*** -0.027*** -0.027*** -0.031***(0.008) (0.008) (0.008) (0.008) (0.007) (0.008)Obs. 438 438 427 427 427 427R2 0.00 0.00 0.02 0.02 0.03 0.03142Table C.6: Forecasting Unemployment Surprises (cont.)Panel B: MAI-WSJRandom-walk surpriseMAI: MAI5−20 MAI20−250 MAI60−250(1) (2) (3) (4) (5) (6)MAI 0.029* 0.014 0.053 0.006 0.096 0.003(0.015) (0.013) (0.039) (0.025) (0.063) (0.039)MAI×NBER 0.202** 0.175** 0.319***(0.093) (0.072) (0.052)Intercept -0.014 -0.015 -0.005 -0.013 -0.005 -0.016*(0.010) (0.010) (0.011) (0.010) (0.011) (0.009)Obs. 395 395 384 384 384 384R2 0.01 0.04 0.02 0.06 0.03 0.10BHJ surpriseMAI: MAI5−20 MAI20−250 MAI60−250(1) (2) (3) (4) (5) (6)MAI 0.024* 0.021* 0.026 0.011 0.038 -0.004(0.013) (0.013) (0.024) (0.023) (0.041) (0.034)MAI×NBER 0.043 0.053 0.145***(0.052) (0.059) (0.052)Intercept -0.033*** -0.033*** -0.027*** -0.029*** -0.027*** -0.031***(0.008) (0.008) (0.008) (0.008) (0.008) (0.008)Obs. 395 395 384 384 384 384R2 0.01 0.01 0.00 0.01 0.01 0.02143Table C.6: Forecasting Unemployment Surprises (cont.)Panel C: MAI-C2 Random-walk surpriseMAI: MAI5−20 MAI20−250 MAI60−250(1) (2) (3) (4) (5) (6)MAI 0.030 0.012 0.157*** 0.109*** 0.233*** 0.136**(0.027) (0.027) (0.040) (0.038) (0.061) (0.056)MAI×NBER 0.232 0.180** 0.379***(0.161) (0.091) (0.091)Intercept -0.010 -0.010 -0.003 -0.010 -0.003 -0.013(0.012) (0.012) (0.010) (0.009) (0.010) (0.009)Obs. 438 438 427 427 427 427R2 0.00 0.01 0.07 0.09 0.09 0.13BHJ surpriseMAI: MAI5−20 MAI20−250 MAI60−250(1) (2) (3) (4) (5) (6)MAI 0.031 0.025 0.082*** 0.082** 0.125*** 0.088*(0.028) (0.027) (0.028) (0.032) (0.042) (0.049)MAI×NBER 0.075 0.000 0.145*(0.146) (0.080) (0.086)Intercept -0.031*** -0.031*** -0.027*** -0.027*** -0.027*** -0.031***(0.009) (0.009) (0.008) (0.008) (0.008) (0.008)Obs. 438 438 427 427 427 427R2 0.00 0.00 0.02 0.02 0.03 0.04144Table C.7: Changes in VIX before and on Announcement DaysColumn 1 of this table shows the result of a five-day change, from t − 6 to t − 1 in VIX, calculated onevery trading day regressed on announcement dummies. Similarly, column 2 shows the result for a three-day change, from t − 4 to t − 1 in VIX, and column 3 for a one-day change, from t − 1 to t in VIX. Theannouncement dummy variables are Empl. and FOMC, and are equal to one if day t occurs on the Employ-ment Situation and FOMC announcements, respectively, zero otherwise. *, **, *** denote the statisticalsignificance at the 10%, 5%, 1% levels, respectively. The sample period is January 1, 1990 to December 31,2016.∆V IX[t−6,t−1] ∆V IX[t−4,t−1] ∆V IX[t−1,t](1) (2) (3)Intercept -0.027 -0.014 0.004***(0.056) (0.038) (0.001)Empl 0.320** 0.016 -0.019***(0.136) (0.120) (0.004)FOMC 0.263 0.381** -0.026***(0.183) (0.152) (0.005)Obs. 6728 6728 6732145Table C.8: Forecasting Returns and Changes in VIX on Employment Situation AnnouncementsThis table presents results of OLS regressions of the daily S&P 500 log return (in %) and changes in impliedvolatility (∆V IX) on Employment Situation announcement days on attention to unemployment. Specifically,we regress log returns in Columns 1 to 3 and ∆V IX in Columns 4 to 6 on the one-day lag detrended de-meaned and standardized unemployment attention composite index MAI-C2, the Boyd et al. (2005) (BHJ)unemployment surprise, the one-day lag detrended Economic Policy Uncertainty (EPU) index, and each ofindependent variables interacted with a NBER dummy equal to one if the Employment Situation announce-ment day occurs during a NBER recession, zero otherwise. MAI5−20,t and EPU5−20,t is the differencebetween the five-day and 20-day moving average of MAI-C2 for unemployment and the EPU index, respec-tively. The standard errors are reported in parenthesis and are calculated using Newey-West (5 lags). *,**, *** denote the statistical significance at the 10%, 5%, 1% levels, respectively. The sample period forthe return regression is June 1, 1980 to December 31, 2016 and for the VIX regression is January 1990 toDecember 31, 2016.S&P 500 Returns ∆VIX(1) (2) (3) (4) (5) (6)MAI5−20 0.38** 0.34** 0.47** -0.96*** -0.82** -0.84**(0.17) (0.17) (0.21) (0.34) (0.34) (0.34)BHJ 0.02 0.27 0.37 0.90 0.18 0.14(0.32) (0.31) (0.34) (0.70) (0.64) (0.64)EPU5−20 -0.01*** 0.00(0.00) (0.00)NBER×MAI5−20 0.39 1.96 -1.72 -2.63(0.74) (1.26) (1.72) (1.89)NBER×BHJ -1.55 -3.23** 5.49* 6.00**(1.15) (1.30) (2.86) (2.69)NBER×EPU5−20 -0.01* 0.01(0.01) (0.01)Intercept -0.01 0.00 0.02 -0.20* -0.24** -0.27***(0.06) (0.06) (0.07) (0.10) (0.10) (0.10)Obs. 428 428 374 317 317 317Adj-R2 0.01 0.01 0.06 0.03 0.07 0.08146Table C.9: Forecasting Returns and Changes in VIX on FOMC AnnouncementsThis table presents results of OLS regressions of the daily S&P 500 log return (in %) in Panel A and changesin the implied volatility (∆V IX) in Panel B on the detrended demeaned and standardized monetary policyattention composite index MAI-C2, the Fed Fund surprise, and the detrended Economic Policy Uncertainty(EPU) index. MAI3−30,t and EPU3−30,t is the difference between the 3- and 30-day moving average ofMAI-C2 and EPU, respectively. The standard errors are reported in parenthesis and are calculated usingthe Newey-West standard errors (5 lags). *, **, *** denote the statistical significance at the 10%, 5%, 1%levels, respectively. The sample period is January 1, 1994 to December 31, 2016.Panel A: S&P 500 Returns1994-2016 1994-2007 2008-2016 2010-2016(1) (2) (3) (4) (5) (6) (7) (8)MAI3−30 0.65** 0.64** 0.07 0.11 1.06*** 1.20*** 1.06** 1.06***(0.26) (0.26) (0.26) (0.26) (0.40) (0.42) (0.42) (0.37)FFSurp -4.55 -4.53 -5.90*** -5.60** 2.77 3.24 -0.50 -0.54(3.15) (3.18) (2.27) (2.18) (10.14) (9.39) (7.75) (7.89)EPU3−30 0.00 0.01*** -0.01 -0.00(0.00) (0.00) (0.01) (0.00)Intercept 0.27*** 0.26*** 0.23** 0.13 0.38** 0.39*** 0.12 0.12(0.08) (0.09) (0.09) (0.10) (0.15) (0.15) (0.14) (0.14)Obs. 183 183 111 111 72 72 56 56Adj-R2 0.06 0.06 0.06 0.09 0.13 0.15 0.17 0.15Panel B: ∆VIX1994-2016 1994-2007 2008-2016 2010-2016(1) (2) (3) (4) (5) (6) (7) (8)MAI3−30 -1.31** -1.21** -0.07 -0.14 -2.23*** -2.11*** -2.45** -2.13**(0.55) (0.49) (0.28) (0.27) (0.85) (0.75) (1.13) (0.90)FFSurp 5.38 5.25 7.95** 7.48** -7.87 -7.48 6.51 3.24(4.31) (4.32) (3.77) (3.72) (9.45) (9.93) (11.73) (12.52)EPU3−30 -0.01* -0.01*** -0.01 -0.01(0.01) (0.00) (0.01) (0.01)Intercept -0.52*** -0.42*** -0.53*** -0.36*** -0.60*** -0.59*** -0.36* -0.37(0.10) (0.11) (0.11) (0.12) (0.20) (0.20) (0.22) (0.23)Obs. 183 183 111 111 72 72 56 56Adj-R2 0.10 0.14 0.08 0.14 0.25 0.25 0.25 0.26147"@en ; edm:hasType "Thesis/Dissertation"@en ; vivo:dateIssued "2018-09"@en ; edm:isShownAt "10.14288/1.0366908"@en ; dcterms:language "eng"@en ; ns0:degreeDiscipline "Business Administration - Finance"@en ; edm:provider "Vancouver : University of British Columbia Library"@en ; dcterms:publisher "University of British Columbia"@en ; dcterms:rights "Attribution-NonCommercial-NoDerivatives 4.0 International"@* ; ns0:rightsURI "http://creativecommons.org/licenses/by-nc-nd/4.0/"@* ; ns0:scholarLevel "Graduate"@en ; dcterms:title "Essays on information and stock returns"@en ; dcterms:type "Text"@en ; ns0:identifierURI "http://hdl.handle.net/2429/65931"@en .