The full text on this page is automatically extracted from the file linked above and may contain errors and inconsistencies.
PRSRT STD U.S. POSTAGE PAID ST. LOUIS MO PERMIT NO. 444 P.O. Box 442 St. Louis, MO 63166-0442 Change Service Requested REVIEW FEDERAL RESERVE BANK OF ST. LOUIS SECOND QUARTER 2019 VOLUME 101 | NUMBER 2 Gauging Market Responses to Monetary Policy Communication Kevin L. Kliesen, Brian Levine, and Christopher J. Waller International Trade Openness and Monetary Policy: Evidence from Cross-Country Data Fernando Leibovici The Real Term Premium in a Stationary Economy with Segmented Asset Markets YiLi Chien and Junsang Lee Racial Gaps, Occupational Matching, and Skill Uncertainty Limor Golan and Carl Sanders REVIEW Volume 101 • Number 2 President and CEO James Bullard Director of Research Christopher J. Waller Chief of Staff Cletus C. Coughlin Deputy Directors of Research iii In Memoriam: Keith M. Carlson B. Ravikumar David C. Wheelock Review Editor-in-Chief Carlos Garriga Research Economists David Andolfatto Subhayu Bandyopadhyay YiLi Chien Riccardo DiCecio William Dupor Maximiliano Dvorkin Miguel Faria-e-Castro Sungki Hong Kevin L. Kliesen Julian Kozlowski Fernando Leibovici Oksana Leukhina Fernando M. Martin Michael W. McCracken Alexander Monge-Naranjo Christopher J. Neely Michael T. Owyang Paulina Restrepo-Echavarría Juan M. Sánchez Ana Maria Santacreu Don Schlagenhauf Guillaume Vandenbroucke Yi Wen Christian M. Zimmermann 69 Gauging Market Responses to Monetary Policy Communication Kevin L. Kliesen, Brian Levine, and Christopher J. Waller 93 International Trade Openness and Monetary Policy: Evidence from Cross-Country Data Fernando Leibovici 115 The Real Term Premium in a Stationary Economy with Segmented Asset Markets YiLi Chien and Junsang Lee 135 Racial Gaps, Occupational Matching, and Skill Uncertainty Limor Golan and Carl Sanders Managing Editor George E. Fortier Editors Jennifer M. Ives Lydia H. Johnson Designer Donna M. Stiller Federal Reserve Bank of St. Louis REVIEW Second Quarter 2019 i Review Review is published four times per year by the Research Division of the Federal Reserve Bank of St. Louis. Complimentary print subscriptions are available to U.S. addresses only. Full online access is available to all, free of charge. Online Access to Current and Past Issues The current issue and past issues dating back to 1967 may be accessed through our Research Division website: http://research.stlouisfed.org/publications/review. All nonproprietary and nonconfidential data and programs for the articles written by Federal Reserve Bank of St. Louis staff and published in Review also are available to our readers on this website. Review articles published before 1967 may be accessed through our digital archive, FRASER: http://fraser.stlouisfed.org/publication/?pid=820. Review is indexed in Fed in Print, the catalog of Federal Reserve publications (http://www.fedinprint.org/), and in IDEAS/RePEc, the free online bibliography hosted by the Research Division (http://ideas.repec.org/). Authorship and Disclaimer The majority of research published in Review is authored by economists on staff at the Federal Reserve Bank of St. Louis. Visiting scholars and others affiliated with the St. Louis Fed or the Federal Reserve System occasionally provide content as well. Review does not accept unsolicited manuscripts for publication. The views expressed in Review are those of the individual authors and do not necessarily reflect official positions of the Federal Reserve Bank of St. Louis, the Federal Reserve System, or the Board of Governors. Subscriptions and Alerts Single-copy subscriptions (U.S. addresses only) are available free of charge. Subscribe here: https://research.stlouisfed.org/publications/ review/subscribe/. Our monthly email newsletter keeps you informed when new issues of Review, Economic Synopses, Regional Economist, and other publications become available; it also alerts you to new or enhanced data and information services provided by the St. Louis Fed. Subscribe to the newsletter here: http://research.stlouisfed.org/newsletter-subscribe.html. Copyright and Permissions Articles may be reprinted, reproduced, republished, distributed, displayed, and transmitted in their entirety if copyright notice, author name(s), and full citation are included. In these cases, there is no need to request written permission or approval. Please send a copy of any reprinted or republished materials to Review, Research Division of the Federal Reserve Bank of St. Louis, P.O. Box 442, St. Louis, MO 63166-0442; STLS.Research.Publications@stls.frb.org. Please note that any abstracts, synopses, translations, or other derivative work based on content published in Review may be made only with prior written permission of the Federal Reserve Bank of St. Louis. Please contact the Review editor at the above address to request this permission. Economic Data General economic data can be obtained through FRED® (Federal Reserve Economic Data), our free database with over 500,000 national, international, and regional data series, including data for our own Eighth Federal Reserve District. You may access FRED through our website: https://fred.stlouisfed.org. © 2019, Federal Reserve Bank of St. Louis. ISSN 0014-9187 ii Second Quarter 2019 Federal Reserve Bank of St. Louis REVIEW In Memoriam: Keith M. Carlson For several decades, the St. Louis Fed was known as the “Monetarist Fed.” Several St. Louis Fed economists— and a few Bank presidents—were instrumental in developing and burnishing this reputation. Keith M. Carlson, who passed away on January 22, 2019, was one of those economists. Most economists familiar with this history likely knew Keith through his important work on the St. Louis equation. His most famous article was “A Monetarist Model for Economic Stabilization,” co-authored with Leonall (Andy) Andersen and published in the St. Louis Fed Review in 1970 (Andersen and Carlson, 1970). This article described the first version of the St. Louis model used to assess how monetary policy affects macroeconomic activity through Keith M. Carlson changes in the money stock. 1934-2019 For at least 25 years, the St. Louis model guided St. Louis Fed Bank presidents and their research directors as they prepared for Federal Open Market Committee meetings. Moreover, the model was central to the debate in the 1970s and 1980s over the relative effectiveness of monetary and fiscal policy. As with any economic model, though, it went through numerous iterations. Keith would publish the final version of the model in 1986 (Carlson, 1986). Although best known for his work on the St. Louis model, Keith’s earliest contributions focused on fiscal and federal budget policy. He studied the “crowding out” effects of budget deficits; developed a measure of the “high employment budget” (Carlson, 1967); and elaborated on a key result in the Anderson-Jordan equation with St. Louis Fed colleague Roger W. Spencer (Carlson and Spencer, 1975). Keith was also known as a great teacher. His Review articles span a variety of topics and explain difficult material in a manner that is accessible to students of economics and the lay public alike. A few articles stand out. In 1988, Keith discussed variables that influence the natural rate of unemployment, a topic that remains highly relevant today for anyone trying to understand the perils facing a central bank that gauges the stance of policy by comparing the unemployment rate to the unknown natural rate (Carlson, 1988). A year later, a Review article on price indexes revisited the role that asset prices should play in a cost of living index (Carlson, 1989). Keith was known for his friendly ease, openness, honesty, and sense of humor. He always had an excellent and timely smile and a one-liner suitable for the moment. He was extremely Federal Reserve Bank of St. Louis REVIEW Second Quarter 2019 iii In Memoriam: Keith M. Carlson helpful in giving of his time and attention to his colleagues in the support of their data, economic theories, econometric findings, and policy and institutional history. Not everyone knew this, but he was also a walking encyclopedia of Major League Baseball, especially the St. Louis Cardinals. He spent much of his free time studying and honing his knowledge of the game through statistical analysis. At one point, a request came from the leading authorities in this field to collect obscure data for their compilation of baseball statistics. Keith methodically went through old issues of the St. Louis Globe-Democrat to assemble hitby-pitch data from the 1890s. He was duly thanked for his contribution and added another, unexpected accomplishment: Keith reported that his life felt full now that he had shaken hands with both Milton Friedman and Pete Palmer.1 Keith was born in Mitchell, South Dakota, on July 3, 1934. He received his undergraduate degree from Gustavus Adolphus College and his M.A. in economics from the University of Nebraska. After graduation, he moved back to Minnesota to teach at St. Olaf College. In 1963, he moved to St. Louis to take a position at the Federal Reserve Bank of St. Louis that would serve as the culmination of his career. He will be missed. n NOTE 1 Pete Palmer co-authored The Hidden Game of Baseball, which is generally viewed as creating the foundation for sabermetrics, a tool used to evaluate a player’s ability and featured in the movie “Moneyball.” REFERENCES Carlson, Keith M. “Estimates of the High-Employment Budget: 1947-1967,” Federal Reserve Bank of St. Louis Review, June 1967, pp. 6-14. Andersen, Leonall C. and Keith M. Carlson, “A Monetarist Model for Economic Stabilization,” Federal Reserve Bank of St. Louis Review, April 1970, pp. 7-25; https://doi.org/10.20955/r.68.45-66.djw. Carlson, Keith M. and Roger W. Spencer, “Crowding Out and its Critics,” Federal Reserve Bank of St. Louis Review, December 1975, pp. 2-17. Carlson, Keith M. “A Monetarist Model for Economic Stabilization: Review and Update,” Federal Reserve Bank of St. Louis Review, October 1986, pp. 18-28; https://doi.org/10.20955/r.68.18-28.yls. Carlson, Keith M. “How much lower can the unemployment rate go?” Federal Reserve Bank of St. Louis Review, July 1988, pp. 44-57; https://doi.org/10.20955/r.70.44-57.fgu. Carlson, Keith M. “Do Price Indexes Tell us about Inflation? A Review of the Issue,” Federal Reserve Bank of St. Louis Review, November 1989, pp. 12-30; https://doi.org/10.20955/r.71.12-30.hhd. iv Second Quarter 2019 Federal Reserve Bank of St. Louis REVIEW Gauging Market Responses to Monetary Policy Communication Kevin L. Kliesen, Brian Levine, and Christopher J. Waller The modern model of central bank communication suggests that central bankers prefer to err on the side of saying too much rather than too little. The reason is that most central bankers believe that clear and concise communication of monetary policy helps achieve their goals. For the Federal Reserve, this means to achieve its goals of price stability, maximum employment, and stable longterm interest rates. This article examines the various dimensions of Fed communication with the public and financial markets and how Fed communication with the public has evolved over time. We use daily and intraday data to document how Fed communication affects key financial market variables. We find that Fed communication is associated with changes in prices of financial market instruments such as Treasury securities and equity prices. However, this effect varies by type of communication, by type of instrument, and by who is doing the speaking. (JEL E52, E58, E61, G10) Federal Reserve Bank of St. Louis Review, Second Quarter 2019, 101(2), pp. 69-91. https://doi.org/10.20955/r.101.69-91 KEYNES: Arising from Professor Gregory’s questions, is it a practice of the Bank of England never to explain what its policy is? HARVEY: Well, I think it has been our practice to leave our actions to explain our policy. KEYNES: Or the reasons for its policy? HARVEY: It is a dangerous thing to start to give reasons. KEYNES: Or to defend itself against criticism? HARVEY: As regards criticism, I am afraid, though the Committee may not all agree, we do not admit there is a need for defense; to defend ourselves is somewhat akin to a lady starting to defend her virtue. Exchange between John Maynard Keynes and Bank of England Deputy Governor Sir Ernest Harvey, December 5, 1929.1 Kevin L. Kliesen is a business economist and research officer, Brian Levine was a senior research associate, and Christopher J. Waller is executive vice president and director of research at the Federal Reserve Bank of St. Louis. © 2019, Federal Reserve Bank of St. Louis. The views expressed in this article are those of the author(s) and do not necessarily reflect the views of the Federal Reserve System, the Board of Governors, or the regional Federal Reserve Banks. Articles may be reprinted, reproduced, published, distributed, displayed, and transmitted in their entirety if copyright notice, author name(s), and full citation are included. Abstracts, synopses, and other derivative works may be made only with prior written permission of the Federal Reserve Bank of St. Louis. Federal Reserve Bank of St. Louis REVIEW Second Quarter 2019 69 Kliesen, Levine, Waller INTRODUCTION Central bank communication has come a long way since the Bank of England’s motto ostensibly was “Never explain, never apologize.”2 Today, the motto of central bankers might instead be “Can you hear me now?” The modern model of central bank communication suggests that central bankers prefer to err on the side of saying too much rather than too little. In this vein, central bank communication takes many forms, from economic forecasts and official reports, to speeches, interviews, testimonies before governmental bodies, and policy statements and press conferences immediately after policy meetings. In the United States, enhancements in central bank communication are most pronounced in the realm of speeches and other remarks (e.g., television interviews) by Federal Reserve (hereafter, Fed) governors and Reserve Bank presidents. These forms of communication have become more prominent since the recession and Financial Crisis. In an era of increased communication by Federal Open Market Committee (FOMC) participants, one may ask whether additional information is useful for financial market participants who carefully monitor monetary policy developments. Indeed, some economists and analysts have argued that Fed officials talk too much.3 There are many nuances to this argument, but the primary claim is that more information increases the probability of market mispricing. Shin (2017) discusses some of these issues. There are at least two counterarguments to the market mispricing view. The first, as enunciated by Kocherlakota (2017), is that the price of an independent central bank is a set of independent voices to insure against group think. The second counterargument is that the pricing of financial instruments in markets is more efficient with more, not less, information. Regardless, central bank communication is important because individuals’ economic decisions are based on expectations of future policies. Thus, clear communication of its policies and actions may help the Fed achieve its mandated goals of stable prices, maximum employment, and moderate long-term interest rates. The purpose of this article is twofold. The first part examines the various dimensions of Fed communication with the public and financial markets. This includes documenting how communication with the public has evolved over time. The second part empirically analyzes the economic effects of Fed communication on key financial market variables. Our analysis uses daily and intraday data. We find that Fed communication can affect prices of financial market instruments such as equities and Treasury securities. However, this effect varies by type of communication, by type of instrument, and by who is doing the speaking. We also find that larger financial market reactions tend to be associated with communication from the Fed Chair, non-Chair Fed governors, and FOMC meetings without an associated press conference. We further find that financial market reactions following press conferences after FOMC meeting statements are not significant. HOW DOES THE FED COMMUNICATE? As the exchange between John Maynard Keynes and Bank of England Deputy Governor Sir Ernest Harvey demonstrated, the principles of central bank communication have evolved 70 Second Quarter 2019 Federal Reserve Bank of St. Louis REVIEW Kliesen, Levine, Waller Table 1 Types of Fed Communication Type Communicator Frequency Release timing Policy statement FOMC 8 times per year After each FOMC meeting, ~2 PM EST Minutes FOMC 8 times per year 3 weeks after each FOMC meeting, ~2 PM EST Press conference Chair 8 times per year* After designated FOMC meeting, ~2:30 PM EST Summary of Economic Projections FOMC 4 times per year After designated FOMC meeting, ~2 PM EST Monetary Policy Report to Congress Chair 2 times per year ~February and July of each year Speeches and other public remarks FOMC Continuous† NA Statement of Longer-Run Goals and Policy Strategy FOMC 1 time per year Reaffirmed each January Policy Normalization Principles and Plans‡ FOMC Updated periodically After associated FOMC meeting, ~2 PM EST NOTE: Table reflects the present-day FOMC procedure. The timing and frequency of each event has changed over the past 20 years. ~Indicates times are approximations and may differ slightly from event to event. *During the period analyzed, press conferences were held only four times per year. Beginning in January 2019, press conferences are held after every FOMC meeting. †Excludes FOMC “blackout periods,” which begin the second Saturday preceding an FOMC meeting and end the Thursday following the meeting. ‡Initially released in September 2014. An addendum was adopted in March 2015 and augmented in June 2017. For a history of revisions, see https://www.federalreserve.gov/monetarypolicy/timeline-policy-normalization-principles-and-plans.htm. over time. A modern comparison describing the evolution of Fed communication was noted in 2003 by then Fed Governor Janet Yellen when she said that the FOMC “had journeyed from ‘never explain’ to a point where sometimes the explanation is the policy.”4 Some have termed this policy “open-mouth operations.”5 Although views may differ between policymakers and across central banks, the fundamental principles of central bank communication are founded on the dual notions that increased transparency enhances the effectiveness of policy and the accountability of policymakers in a democratic society.6 In this article, we focus on Fed communication, though the principles and practices are similar among many of the world’s central banks. When analyzing central bank communication, the following questions come to mind: First, who should do the talking; second, what should the central bank talk about; and, third, who should the central bank talk to? There is a vast economic literature that attempts to answer these questions. One notable early effort was a cross-country study by Blinder et al. (2001), who surveyed communication methods and tactics, among other things. A subsequent article by Blinder et al. (2008) argued that there was large variation in strategies but no consensus on the best-practice approach to communicating monetary policy to the public. Woodford (2001) was an early proponent of using communication to influence market expectations. This view influenced several subsequent Fed officials, most notably former Fed Chairman Ben Bernanke.7 Finally, in the aftermath of the Financial Crisis of 2008, several event studies were published that analyzed the FOMC’s unconventional policy actions on prices of financial market instruments, macroeconomic outcomes, and the expectations about future monetary policy actions.8 Federal Reserve Bank of St. Louis REVIEW Second Quarter 2019 71 Kliesen, Levine, Waller In sum, the academic literature offers more support for the modern view of central bank communication: More is generally better. Table 1 lists the primary methods that the Fed uses to communicate its policies, procedures, and policy expectations to the public.9 These methods include the policy statement released at the end of the eight regularly scheduled FOMC meeting, the minutes released three weeks after each of the eight regularly scheduled FOMC meetings, the Chair’s quarterly press conference, along with speeches, testimonies, and media interviews by Fed governors and Reserve Bank presidents. Some of these innovations are long standing, such as the FOMC minutes, while others are more recent, such as the Chair’s press conferences.10 Given the prominence of FOMC policy statements as a communication instrument, the following discussion will first briefly focus on their history and role. Policy Statements: Length and Readability The Fed’s principle medium of communication is the policy statement released after each FOMC meeting. The policy statement has evolved over time. From 1967 to 1992, the FOMC issued a “Record of Policy Actions” (ROPA), which were initially released with a 90-day lag.11 Beginning under Chairman Alan Greenspan, the FOMC began to issue policy statements immediately after the February 4, 1994, meeting. The first policy statement was rather short, at 99 words, and made no mention of the intended federal funds target rate. Instead, the inaugural statement indicated that the Committee decided to “increase slightly the degree of pressure on reserve positions” in financial markets. In taking this action, the FOMC noted that they expected an “associated small increase in short-term money market interest rates.”12 Following the release of the inaugural statement, the FOMC released a post-meeting statement four additional times in 1994. Three post-meeting statements were released in 1995, including the statement released after the July 6, 1995, meeting, which was the first instance that the FOMC specifically mentioned the federal funds rate. The FOMC continued to issue post-meeting statements over the next few years, but only at meetings where a policy change occurred. However, beginning with the May 18, 1999, meeting, statements were released after each FOMC meeting.13 The public focus on the policy statement was such that the financial press developed a “briefcase barometer.”14 The post-meeting FOMC statements have evolved over time. Prior to the Financial Crisis, the post-meeting policy statement mostly focused on the state of the economy and the Com mittee’s rationale for raising or lowering the policy rate or reasons why the policy rate was not changed. In general, less was said about the future path of interest rate changes. The policy statement evolved to take on a larger role in communicating the stance of monetary policy during the Financial Crisis after the federal funds rate reached the zero lower bound (ZLB) on December 16, 2008.15 Figure 1 shows that the word count of the policy statements began to increase steadily in 2007 during the early stages of the Financial Crisis. The word count continued to increase during the adoption of quantitative easing (QE) policies that both increased the size of the balance sheet and changed its composition. Prior to the ZLB period, the number of words in each statement averaged 223. During the ZLB period, the count was more than twice as much, averaging 580 words. After the nominal federal funds target rate reached the ZLB in December 2008, the Fed provided the largest amount of monetary accommodation through balance sheet adjustments and 72 Second Quarter 2019 Federal Reserve Bank of St. Louis REVIEW Kliesen, Levine, Waller Figure 1 FOMC Statement Word Count Number of Words in Each Statement 1,000 900 800 700 600 500 400 300 200 100 0 1994 1999 2001 2002 QE1 2004 2006 2008 QE2 MEP QE3 2009 2011 2013 2015 2017 NOTE: Shaded area indicates the period of the FOMC’s unconventional monetary policy with interest rates at the effective ZLB. MEP, Maturity Extension Program. Under the MEP, the Fed sold or redeemed shorter-term Treasury securities and used the proceeds to buy longer-term Treasury securities, thereby extending the average maturity of the securities in the Fed’s portfolio. Updated through 2017. SOURCE: Board of Governors of the Federal Reserve System. other unconventional policies.16 But as the U.S. economy transitioned from recession to a slower-than-average recovery, the Fed’s policy approach also changed. The new approach focused instead on influencing the public’s expectations of the future direction and level of the federal funds target rate. This approach, in its current form, is referred to as forward guidance.17 For example, following the August 9, 2011, meeting, the policy statement stated the following: The Committee currently anticipates that economic conditions—including low rates of resource utilization and a subdued outlook for inflation over the medium run—are likely to warrant exceptionally low levels for the federal funds rate at least through mid-2013. In this case, the FOMC’s intent was to signal to the public that its policy rate would remain low for a long time in order to spur the economy’s recovery. This signal was meant to be taken as a public commitment, what Campbell et al. (2012) termed “Odyssean” policy. Using language from Greek mythology, Odyssean policy is meant to convey a public commitment not to change policy for a certain period—in this case, for more than two years. Instead, the public appeared to view this statement as a forecast, what Campbell et al. (2012) termed “Delphic” policy. In effect, the Delphic statement strongly suggested that, in the FOMC’s view, the economic weakness would persist for more than two years. However, at the June 2011 meeting two months earlier, the Summary of Economic Projections (SEP) indicated that real gross domestic product (GDP) would increase by 3.5 percent in 2012 and by 3.9 percent in 2013 (each measure is the midpoint of the central tendency).18 Thus, by August, the Committee appeared to have concluded that it, like most private sector forecasters, had been much too optimistic about the pace of real GDP growth during the early stages of the expansion. Indeed, by the January 2012 meeting, forecasts for real GDP growth in 2012 and 2013 had been marked down to 1.7 percent and 2.5 percent, respectively. Federal Reserve Bank of St. Louis REVIEW Second Quarter 2019 73 Kliesen, Levine, Waller Figure 2 FOMC Statement Complexity Flesch-Kincaid Reading Grade Level 24 QE1 QE2 MEP QE3 22 20 18 16 14 12 10 8 6 1994 1999 2001 2002 2004 2006 2008 2009 2011 2013 2015 2017 NOTE: Shaded area indicates the period of the FOMC’s unconventional monetary policy with interest rates at the effective ZLB. MEP, Maturity Extension Program. Under the MEP, the Fed sold or redeemed shorter-term Treasury securities and used the proceeds to buy longer-term Treasury securities, thereby extending the average maturity of the securities in the Fed’s portfolio. Updated through December 2017. SOURCE: Board of Governors of the Federal Reserve System (FOMC statements) and Educational Testing Service (word count). To accomplish the Fed’s goals and objectives in a slow-growth economy, the post-meeting statement changed in two dimensions. The first change, as noted above, was that the length increased. The statements included more discussion of the economic situation and its implication for the near-term direction of policy (changes in the federal funds target rate).19 Second, the statements incorporated more complex economic terms and analysis. This is shown in Figure 2, which uses text evaluation software to measure the Flesch-Kincaid reading grade level of the policy statement. A higher grade level is assumed to reflect increased complexity of the statement. Prior to the ZLB period, the median grade level was 13.5, indicating comprehension accessible to someone reading at a college undergraduate level. But by late 2013, when the FOMC was in the midst of increasing the size of its balance sheet through asset purchases, the grading level rose to 20, which is commensurate to a graduate school reading level. For the entire ZLB period, the grade level rose to 16 (median), but then fell to 15 (median) during the post-ZLB period.20 Researchers find that the readability of central bank policy statements and remarks are an important factor in how they are received by financial markets. Not surprisingly, clearer statements lead to lower volatility.21 This section has highlighted how the FOMC changed the length and composition of the policy statement during the period of unconventional monetary policy. But the policy statement is only one form of central bank communication. Speeches and other public remarks are another form of communication that policymakers have deployed to increase the public’s knowledge of the prevailing monetary policy regime. The next two sections will delve into monetary policy communication strategies by Fed officials, both old and new. 74 Second Quarter 2019 Federal Reserve Bank of St. Louis REVIEW Kliesen, Levine, Waller Figure 3 Number of Public Remarks by Type of Fed Official Total Remarks Per Year 250 ZLB Period Begins (December 2007) 200 Bank Presidents 150 100 Non-Chair Governors 50 0 FOMC Chair 1998 2000 2002 2004 2006 2008 2010 2012 2014 2016 NOTE: Through 2017. SOURCE: Board of Governors of the Federal Reserve System, the 12 Federal Reserve Banks, Bloomberg, and authors’ calculations. Public Remarks by Fed Officials Fed officials have long used other forms of public communication besides policy statements.22 Public remarks can take many forms, including formal speeches, Congressional testimonies, interviews with the financial media, or published articles and commentaries. Sometimes, Fed officials do not comment on monetary policy issues that may be discussed at recent or upcoming FOMC meetings. In those instances, policymakers may instead choose to focus on other issues, such as local economic conditions, economic education, community development, or banking and financial market regulation. The ZLB period witnessed an unprecedented rate of spoken and written communication with the public by Fed governors and Reserve Bank presidents. Figure 3 shows the annual number of public remarks by the Fed Chair, non-Chair governors, and Reserve Bank presidents since 1998.23 From 1998 to 2004, the total number of public remarks by Reserve Bank presidents remained roughly constant at about 150 per year. A slightly different pattern occurred with governors and the Fed Chair. Total remarks over this period steadily fell, but then rebounded, so that the numbers of public remarks in 2004 were close to the 1998 totals. Beginning in 2005, the total number of public remarks by Reserve Bank presidents began to increase, reaching a peak in 2013 of a little more than 220 public remarks. Interestingly, though, the FOMC Chair and governors delivered public remarks slightly less frequently over the ZLB period. Some of the reduced frequency of public remarks by members of the Board of Governors (excluding the Chair) reflects the fact that the Board has rarely operated with a full complement of Governors (seven). From 1998 to 2017, there has only been four years when there were seven governors present at the last formal meeting of the year. Indeed, at the end of 2017, there were only four governors at the December meeting. At the March 2018 meeting, the number of governors had dwindled to three. Federal Reserve Bank of St. Louis REVIEW Second Quarter 2019 75 Kliesen, Levine, Waller Figure 4 Number of Times More Than One Bank President or Governor Spoke on the Same Day, 1998-2017 Frequency 80 70 Bank Presidents 60 60 50 40 30 20 Governors 10 0 3 1998 2000 2002 2004 2006 2008 2010 2012 2014 2016 SOURCE: Board of Governors of the Federal Reserve System, Bloomberg, and authors’ calculations. Speeches have become important communication events. Chairman Greenspan’s new economy speech in 1995 and his “irrational exuberance” speech in 1996 were among his more notable speeches. Chairman Ben Bernanke also gave notable speeches during his tenure. Two that standout are his “Deflation: Making Sure ‘It’ Doesn’t Happen Here” speech in 2002 and his global saving glut speech in 2005. Days with multiple Fed communication events have become more numerous over time—particularly since the Financial Crisis. Figure 4 shows the increase in multiple Fed communication events on the same day stems from an increase in more than one Reserve Bank president speaking on the same day. For example, in 2017, there were 60 days when more than one Reserve Bank president spoke. In 2004, it was about half as much. By contrast, in 2017 there were only three days when more than one Fed governor spoke publicly on the same day. This is down sharply from 2003, when there were 19 days when multiple Fed governors spoke on the same day.24 In separate analysis, we looked at the annual number of public remarks by Reserve Bank presidents from January 1998 to December 2017. We separated the sample into roughly two 10-year periods: January 1998 to August 2008 (pre-Financial Crisis) and September 2008 to December 2017 (post-Financial Crisis). The number of public remarks by Reserve Bank presidents increased in all but three Fed Districts (Chicago, New York, and Richmond). The average increase in volume across these nine Districts was 46 percent. We did not examine whether the nature of the remarks by Reserve Bank presidents has changed over time. We did, however, analyze the number of speeches and public remarks given by presidents of the Fed Bank of St. Louis since January 1929. We have documented this in the boxed insert. Other Forms of Fed Communication In the past several years, chiefly under the Bernanke regime, the FOMC has adopted several new forms of communication to further increase transparency. As noted earlier, the Chair’s quarterly press conference, beginning under Chairman Bernanke’s term in January 76 Second Quarter 2019 Federal Reserve Bank of St. Louis REVIEW Kliesen, Levine, Waller Volume and Subject Matter Across Time: An Example Using St. Louis Fed Presidents The table lists the number of speeches (including public remarks) given by presidents of the Federal Reserve Bank of St. Louis since January 1929. The table also lists the primary subject matter of the speeches under three broad headings: The economic outlook (which includes monetary policy related topics), banking and finance, and all other subjects. In general, St. Louis Fed presidents prior to the 1960s tended to give fewer speeches than presidents after the 1960s. Moreover, presidents prior to the 1960s tended to talk more about non-economic outlook topics and proportionately more about banking and finance issues relative to their modern-era counterparts. Speeches by St. Louis Fed Presidents, 1929 to 2017 Other % of total speeches about economic outlook % of total speeches about banking and finance % of total speeches about other Average number of speeches per year Start End Months in office Total speeches Economic outlook Banking and finance William McChesney Martin Sr. Jan-1929 Feb-1941 145 4 0 3 1 0 75 25 0.3 Chester C. Davis Apr-1941 Feb-1951 118 64 38 11 15 59 17 23 6.5 Delos C. Johns Feb-1951 Feb-1962 132 25 15 8 2 60 32 8 2.3 Harry A. Shuford Oct-1962 Jan-1966 39 9 4 5 0 44 56 0 2.8 Darryl R. Francis Jan-1966 Feb-1976 121 110 94 11 5 85 10 5 10.9 Lawrence K. Roos Mar-1976 Jan-1983 82 37 33 4 0 89 11 0 5.4 Theodore H. Roberts Feb-1983 Dec-1984 22 8 6 2 0 75 25 0 4.4 Thomas C. Melzer Jun-1985 Jan-1998 151 100 82 17 1 82 17 1 7.9 William Poole Mar-1998 Mar-2008 120 140 133 4 3 95 3 2 14.0 James Bullard Apr-2008 Dec-2017* 116 179 173 5 1 97 3 1 18.5 President NOTE: *End date for President Bullard reflects our window for analysis rather than his tenure in office. SOURCE: Presidents prior to Poole, FRASER®, Federal Reserve Bank of St. Louis. Poole and Bullard, Bloomberg. As shown in the table, the focus in recent years has shifted toward a greater emphasis on the economic outlook, which includes macroeconomic conditions and monetary policy developments. Indeed, the economic outlook (and monetary policy related topics) comprised a very large percentage (more than 95 percent) of the speeches of the most recent two St. Louis Fed presidents—William Poole and James Bullard. Finally, consistent with the findings of Figure 3, the last column of the table shows that these two St. Louis Fed presidents have given the highest number of speeches per year of all St. Louis Fed presidents. 2012, is one key innovation. Current Chairman Jerome Powell expanded on this innovation, announcing that press conferences will be held after every FOMC meeting beginning in January 2019. Other innovations include the FOMC’s “Statement of Longer-Run Goals and Monetary Policy Strategy,” “Policy Normalization Principles and Plans,” and “Summary of Economic Projections” (SEP). These are also listed in Table 1. The first two are meant to provide clarity on the Fed’s dual mandate and balance sheet, respectively, while the SEP conveys projections for four key macroeconomic variables. In addition, the SEP conveys each FOMC participant’s assessment of appropriate monetary policy, as indicated by their federal funds rate projections over short-, medium-, and longer-term horizons. Federal Reserve Bank of St. Louis REVIEW Second Quarter 2019 77 Kliesen, Levine, Waller Grading Fed Communication The Hutchins Center on Fiscal and Monetary Policy at Brookings conducted a survey of academics and private sector Fed watchers to assess the effectiveness of different forms of Fed communication.25 Survey participants viewed the FOMC policy statement, speeches by the FOMC Chair, and quarterly press conferences as the most useful forms of Fed communication. On net, academics generally found these forms of communication more useful than did the private sector economists and Fed watchers. One of the key communication innovations during the Bernanke tenure was the public release of individual FOMC participants’ expectations of the future level of the federal funds rate. Once a quarter, with the release of the SEP, each FOMC participant—anonymously— indicates their preference for the level of the federal funds rate at the end of the current year, at the end of the next two to three years, and over the “longer run.” According to the survey, these projections are often termed the FOMC “dot plots.” Both academics and those in the private sector found the dot plots of limited use as an instrument of Fed communication (more “useless” than “useful”). One-third of the respondents found the dot plots “useful or extremely useful,” 29 percent found them “somewhat useful,” and 38 percent found them “useless or not very useful.” The limited usefulness of the dot plots probably reflects many factors. First, each participant’s projection is conditioned on the highly restrictive assumption of “appropriate monetary policy.” Each participant’s appropriate monetary policy stance is conditioned on their view of the outlook for real GDP growth, inflation, and the unemployment rate over the medium term. Moreover, the range of participants’ views may not dovetail with the policy path outlined in the FOMC statement, which can further complicate the communicated outlook and diminish the tool’s effectiveness. The regular presence of dissents suggests that appropriate policy can differ sharply across the Committee. Second, the participants may have other vastly different assumptions that influence their outlook, such as the equilibrium real interest rate, the future path of crude oil prices, the foreign exchange value of the dollar, or their outlook for foreign economic growth. For these reasons and more, FOMC participants persistently over-projected the federal funds target rate path during the early years of the current expansion. [See earlier discussion on page 73.] These persistent one-sided forecast errors may have impaired the credibility of the dot plots to the extent that the projections were important inputs in establishing expectations about future monetary policy. Finally, the Brookings study revealed that survey participants believe that Reserve Bank presidents’ speeches are slightly less useful than the dot plots, but still more useful than Fed reports to Congress, such as the semi-annual Monetary Policy Report.26 This finding is perhaps striking given that the number of public remarks by Reserve Bank presidents has been trending up over time, especially during the ZLB period, while the number of public remarks by the Chair and non-Chair governors has been trending down. 78 Second Quarter 2019 Federal Reserve Bank of St. Louis REVIEW Kliesen, Levine, Waller EMPIRICAL ANALYSIS The final section of the article assesses how financial market participants respond to various forms of Fed communication. Admittedly, this is a difficult empirical exercise for many reasons. First, public remarks by Fed senior officials are often context- and perspectivedependent. Each individual brings their own perspective, model of the economy, and view of the monetary policy transmission mechanism. These views naturally inform their assessments of appropriate monetary policy going forward, which are then conveyed in public remarks. For their part, financial market participants may become familiar with a given policymaker’s view or assume a given outcome for a particular FOMC meeting. If so, markets may react only to views that are sufficiently different from expectations. Past research has demonstrated that monetary policy surprises can have significant effects on high-frequency asset prices.27 We acknowledge the importance of monetary policy surprises, but use a different approach to assess the significance of Fed communication events. Second, when attempting to gauge the significance of public remarks, markets do not usually assign equal weights to all FOMC participants. Certainly, markets carefully parse remarks by the Chair, who is typically viewed as the public voice of the FOMC and the one who sets the policy agenda. Moreover, while the Chair’s views often convey the consensus view of the Committee, the Chair nonetheless also has a policy preference. Although the Chair’s preference invariably prevails, dissents still occur periodically. Indeed, Reserve Bank presidents sometimes use their public remarks, or dissents, with the intention of signaling future policy preferences or advocating for alternative frameworks.28 Still, markets may discount the views of the presidents, on average, because they believe their views unnecessarily distort market signals or future policy intentions. For example, Lustenberger and Rossi (2017) claim that remarks by Reserve Bank presidents worsen the accuracy of private sector forecasts. With these caveats in mind, we adopt a two-pronged empirical exercise. The first exercise uses daily data to examine whether Fed communication events are associated with significant movements in key financial market variables. Admittedly, this approach has some drawbacks. First, daily financial market data tend to be more volatile compared with monthly or quarterly data. Second, this volatility arises, in part, because financial markets trade on many types of information, such as macroeconomic data or global financial or geopolitical developments. Thus, while Fed communication comprises one set of information the market uses to price assets, there are potentially many other sources of information that the market uses that we can’t readily account for. Our intent is to assess market reactions to Fed communication events and not to model changes in asset price movements at a high frequency. The second empirical exercise uses intraday data at 5-minute frequencies. Using intraday data allows us to more closely match the timing of Fed communication events with the responses in financial markets. This is the approach adopted by most of the aforementioned event studies. Our intent is to determine if the empirical results using the daily data are consistent with those from the intraday data. Before presenting the results, we provide a detailed description of our data sources and approach. Federal Reserve Bank of St. Louis REVIEW Second Quarter 2019 79 Kliesen, Levine, Waller Data Sources and Approach We study the effects of seven types of Fed communication events: FOMC meeting statements;29 FOMC minutes; Fed Chair press conferences; public remarks by the Fed Chair, non-Chair Fed governors, and Reserve Bank presidents; and unconventional monetary policy announcements.30 Five of the seven categories are included in the Brookings study. It is important to note that there is an overlap between FOMC meetings and six of the seven unconventional policy announcements we include.31 Initially, our data set included public remarks made after market hours and on weekends. Consistent with some of the literature, we initially moved an after-hours communication event to the following trading day to gauge the market’s reaction to the remark. However, this approach ended up producing large reactions that were probably not tied to the public remark itself. For example, many key data releases are often issued before the market opens.32 In this case, it is difficult to determine whether the market is responding to the public remarks by a Fed official or to economic data releases that may be a surprise.33 Empirical Analysis: Daily Data We create a series of dummy variables for the Fed communication events. Because the Brookings study found that survey participants viewed the Fed Chair press conferences as a useful form of communication, we identify regularly scheduled FOMC meetings with and without an associated press conference. In recent years, FOMC press conferences have occurred after the March, June, September, and December meetings. Since the liftoff from the ZLB at the December 2015 meeting, increases in the FOMC’s federal funds target rate have occurred at meetings with an associated press conference by the Fed Chair. Our sample period is January 6, 1998, to December 29, 2017. There are nine types of communication events: • • • • • • • • • Non-press conference FOMC meeting statements Press conference FOMC meeting statements Releases of FOMC minutes Remarks by the FOMC Chair Remarks by all other Fed governors Remarks by Reserve Bank presidents Days when there are multiple Fed communication events (e.g., speeches) Unconventional policy actions (e.g., large-scale asset purchases) Key macroeconomic data releases (e.g., industrial production) We evaluate the market reaction for three financial instruments: the absolute value of the daily change in the yield on 2-year Treasury notes, the yield on 10-year Treasury notes, and the Chicago Board Options Exchange equity market volatility index (VIX). Changes in 2-year Treasury yields are widely viewed as being sensitive to expected changes in FOMC policy. The 10-year Treasury yield is the most liquid, long-term, risk-free interest rate in the financial markets. It is also sensitive to changes in inflation expectations and longer-term expectations about short-term interest rates. Finally, the VIX, which is often termed the mar80 Second Quarter 2019 Federal Reserve Bank of St. Louis REVIEW Kliesen, Levine, Waller ket’s “fear gauge,” is sometimes viewed as signaling changes in economic uncertainty. This exercise can be represented by the following equation: ΔYi ,t = α + β1ΔYi ,t−1 + β 2 NPCi ,t + β3 PCi ,t + β 4 MIN i ,t + β5CHAIRi ,t + β6GOVi ,t + β7 PRESi ,t + β8 MULTi ,t + β9UNCONVi ,t + β10 MACROi ,t , where ∆Yi,t represents the absolute value of the daily change in financial variable i (either the 2-year Treasury yield, 10-year Treasury yield, or VIX) on day t. The independent variables include a constant, a one-day lag of the dependent variable, and a series of dummy variables (specified earlier in this section) that take the value of 1 if that event occurs on day t or are zero if the event does not occur on day t. We analyze daily data with three ordinary least-squares regressions. We use the absolute value of the daily changes because some communication events will cause yields to increase or decrease, while others will generate no market response. Using absolute values are a more effective way to gauge the effects of communication events on financial market activity.34 We also include another dummy variable (MACRO) on days when key economic statistics are released. The motivation for this is that the market trades on information contained in these reports. Our economic statistic dummy variable takes the value of 1 when the following monthly economic reports are released (and is zero on all other days): the consumer price index, monthly employment situation, industrial production, retail sales, the Institute for Supply Management Report on Manufacturing, and the three GDP releases (advance, second, and third estimates). Table 2 shows the results of our analysis using daily data. Daily data allow us to make a few noteworthy observations. First, for the change in the 2-year Treasury yield, markets react significantly (at the 1 or 5 percent level) to Fed Chair and Fed governor communication events and also to FOMC statements at non-press conference meetings. Table 2 further indicates that changes in 2-year Treasury securities do not react significantly on days when one Reserve Bank president speaks, but they do react significantly on days when there are multiple Fed speakers. (Recall from Figure 4 that the number of days with multiple Fed speakers has increased since the Financial Crisis). Finally, 2-year yields also react significantly to macroeconomic data releases. Unconventional policy actions are marginally significant (at the 8 percent level). With the exception of days with multiple Fed speakers, the signs of the coefficients on the significant variables are positive. The second and third sets of regressions in Table 2 show results for the change in the 10-year Treasury yield and in the VIX. Traders of longer-term Treasury securities react broadly similarly to Fed communication events and data releases as traders of 2-year Treasury securities. For instance, 10-year yields react significantly to the Fed Chair’s remarks, on days when there are multiple Fed speakers, and to macroeconomic data releases; the coefficients generally have the same signs and magnitudes as those from the regression using 2-year yields. However, there are some differences between the 2-year and 10-year responses. For example, the change in the 10-year yield is significantly associated with unconventional policy actions. Moreover, 10-year yields do not react significantly to remarks by Fed governors or to non-press conference FOMC statements. Federal Reserve Bank of St. Louis REVIEW Second Quarter 2019 81 Kliesen, Levine, Waller Table 2 Federal Reserve Communication Events and Financial Market Responses Using Daily Data Dependent variables Independent variables 2-year Treasury 10-year Treasury VIX Constant 0.023 (0.000)** 0.035 (0.000)** 0.641 (0.000)** Lagged dependent variable 0.248 (0.000)** 0.132 (0.000)** 0.321 (0.000)** Non-press conference FOMC meetings 0.013 (0.001)** 0.006 (0.187) 0.059 (0.588) Press conference FOMC meetings 0.004 (0.594) 0.001 (0.880) 0.092 (0.626) FOMC minutes 0.004 (0.154) 0.004 (0.309) –0.078 (0.315) FOMC Chair remarks 0.006 (0.001)** 0.005 (0.012)* 0.018 (0.787) Fed governor remarks 0.004 (0.003)** 0.001 (0.347) 0.020 (0.645) Fed president remarks 0.000 (0.775) 0.000 (0.877) 0.093 (0.111) Multiple Fed speakers –0.004 (0.032)* –0.004 (0.018)* –0.068 (0.265) Unconventional policy actions 0.036 (0.080) 0.044 (0.021)* 0.626 (0.176) Macroeconomic data releases 0.009 (0.000)** 0.009 (0.000)** 0.077 (0.031)* Adjusted R-squared 0.089 0.040 0.109 Durbin-Watson statistic 2.107 2.043 2.264 NOTE: p-values listed in parentheses. The sample period is January 6, 1998 to December 29, 2017. * and ** indicate significance at the 5 percent and 1 percent levels, respectively. Dependent variables are expressed as the absolute value of their one-day changes. Column 3 presents the results for the change in the VIX. Equity market volatility does not react significantly to Fed communication events. The closest variable of significance (p = 0.11) are remarks by Reserve Bank presidents. Equity market volatility does, however, react significantly to macroeconomic data releases. Finally, in all three regressions, the constant and the lagged dependent variable are significant at the 1 percent level. Figure 5 provides some visual evidence for the behavior of equity market volatility around FOMC meetings: From January 1994 to December 2017, the VIX begins to rise about a week before an FOMC meeting. The VIX then drops relatively sharply (nearly 3 percent) on the day the FOMC statement is released. This finding suggests that equity markets appear to be increasingly uncertain about the meeting outcome, or its effects on financial markets, in the run-up to FOMC meetings.35 Likewise, we see a noticeable reduction in market volatility 82 Second Quarter 2019 Federal Reserve Bank of St. Louis REVIEW Kliesen, Levine, Waller Figure 5 Relative Changes in the VIX Near FOMC Announcement Days VIX = 100 on FOMC Announcement Day 104 103 102 101 100 99 98 97 Mean drop in VIX on day of FOMC Statement release: 2.8% –10 –9 –8 –7 –6 –5 –4 –3 –2 –1 FOMC 1 2 3 4 5 6 7 8 9 10 NOTE: Sample includes all regularly scheduled FOMC meetings between January 1994 and December 2017. SOURCE: Haver Analytics and authors’ calculations. after the policy announcement (statement), perhaps indicating a decline in uncertainty and clearer understanding of the Fed’s reaction function. Finally, other than the lagged dependent variable, the dummy variable that accounts for the release of key economic reports is the only other independent variable that is statistically significant. We now turn to the second approach of our empirical exercise, namely, examining the effects of communication events on financial market outcomes using intraday data. Empirical Analysis: Intraday Data We use intraday data to estimate the effects of Fed communication on key financial market variables. Many researchers have used intraday data to gauge market reactions to monetary policy surprises or to the Fed’s announcements of unconventional polices after the Financial Crisis. These event studies, as they are often called, are intended to measure the financial market’s response to news at intervals measured in minutes. Our analysis of the market’s response to Fed communication events generally follows the form and practice of the event study literature. Event studies can be criticized for many reasons. First, the studies gauge only the initial announcement responses rather than the responses across time. Second, the results can be sensitive to the choice of window size—that is, responses evaluated over a 1-minute window versus a 5- or 10-minute window. Third, responses could be affected by non-announcement effects, such as from economic data releases or geopolitical events. In view of these concerns, we tested several different window sizes for robustness and used minute-by-minute asset price data for the S&P 500 stock prices and 10-year Treasury futures prices.36 For FOMC meeting statements, FOMC minutes, and unconventional policy announcements (non-speaker events), a window of plus or minus 15 minutes is used. For FOMC press conferences and other public remarks (speaker events), a window of 15 minutes before to 60 minutes after the event is used. We do not find that the interpretation of the results meaningfully changes when the event window is adjusted.37 Federal Reserve Bank of St. Louis REVIEW Second Quarter 2019 83 Kliesen, Levine, Waller Since these two variables are both prices, we calculate the percent change in each series over each event window.38 We then summarize this information via two metrics: mean absolute change and cumulative change. For non-speaker events, where the event window is +/–15 minutes, the mean absolute change can be represented as j MACNon-Speaker = j 1 N ⎛ Yi ,t+15 ⎞ −1⎟ * 100 , ∑ j N i=1 ⎜⎝ Yi ,t−15 ⎠ where Yi,tj represents, for each non-speaker event category j, the asset price associated with observation i at time t, and N represents the total number of observations for each nonspeaker event category j over the sample. Likewise, for speaker events, where the event window is –15/+60 minutes, the mean absolute change can be represented as j MACSpeaker = j 1 N ⎛ Yi ,t+60 ⎞ −1⎟ * 100 , ∑ j N i=1 ⎜⎝ Yi ,t−15 ⎠ using the same notation as before. The cumulative change is calculated similarly, but we are now summing (instead of averaging) over our sample, and we are not taking the absolute value beforehand. We represent this as and N ⎡⎛ Y j ⎤ ⎞ j CCNon-Speaker = ∑ ⎢⎜ i ,t+15 −1⎟ * 100 ⎥ j i=1 ⎢ ⎥⎦ ⎣⎝ Yi ,t−15 ⎠ N ⎡⎛ Y j ⎤ ⎞ j CCSpeaker = ∑ ⎢⎜ i ,t+60 −1 * 100 ⎥ j ⎟ i=1 ⎢ ⎥⎦ ⎣⎝ Yi ,t−15 ⎠ for non-speaker and speaker events, respectively, again using the same notation as before. The results are shown in Figure 6A, Figure 6B, and Figure 7. The grouping on the left side of Figure 6A shows the mean absolute changes in the S&P 500 index in response to Fed communication events not associated with an individual Fed official (non-speaker events), while the grouping on the right side of Figure 6A shows those is response to events with public remarks by a Fed official (speaker events).39 On the left side, we find that stock prices react most strongly to unconventional policy actions—indeed, twice as strong as the next-largest event (FOMC meeting statements). This finding appears consistent with the event study literature cited earlier. On the right side, stock prices react the most to the Chairs’ press conferences and their remarks. In contrast, stock price changes in response to Fed communication events by Reserve Bank presidents and Governors are similar in magnitude. Figure 6B shows the same calculation for 10-year Treasury bond futures prices. The results in Figure 6B are broadly similar to those in Figure 6A. In particular, responses to unconventional policies are substantially larger than to other forms of Fed communication, such as FOMC meeting statements. As with stock prices, bond markets appear to react more strongly 84 Second Quarter 2019 Federal Reserve Bank of St. Louis REVIEW Kliesen, Levine, Waller Figure 6A Mean Absolute Changes in S&P 500 Index Percent 0.70 0.7 Non-Speakers 0.6 Speakers 0.5 0.4 0.35 0.35 0.3 0.21 0.2 0.33 0.28 0.26 0.25 0.16 0.1 #2 ts ts en en id Ev es n- Pr ai Go rR ve em rn ar or s ks es No Ch sC es Pr Un No co n- on Ev fe en re ts nc #1 es ut in tin gs ee M nv M en tio na l 0.0 NOTE: Underlying data are expressed as a percent change in the S&P 500 index over an event window of –15/+15 minutes (non-speaker events) or of –15/+60 minutes (speaker events). The absolute values of these percent changes are then averaged, for each event category, over the full sample. Non-event days are days with no Fed communication event. Non-speaker events and speaker events have different non-event day controls because they are associated with different window sizes. Figure 6B Mean Absolute Changes in 10-Year Treasury Futures Prices Percent 0.7 0.68 Non-Speakers 0.6 Speakers 0.5 0.4 0.3 0.22 0.2 0.18 0.09 0.1 0.14 0.12 0.05 0.10 0.08 Ev en t n- es id Pr s# 2 ts s or ve rn Go en No Ch ai rR em ar ks es on sC es Pr No n- Ev e nt fe re nc s# 1 es ut in M ee t M Un co nv en tio na in gs l 0.0 NOTE: Underlying data are expressed as a percent change in 10-year Treasury futures price over an event window of –15/+15 minutes (non-speaker events) or of –15/+60 minutes (speaker events). The absolute values of these percent changes are then averaged, for each event category, over the full sample. Non-event days are days with no Fed communication event. Non-speaker events and speaker events have different non-event day controls because they are associated with different window sizes. Federal Reserve Bank of St. Louis REVIEW Second Quarter 2019 85 Kliesen, Levine, Waller Figure 7 Cumulative Changes for Fed Communication Events Percentage Points 4.0 S&P 500 10-Year Treasury Futures 2.0 3.19 1.32 1.32 0.0 –0.54 –2.0 3.39 1.76 –0.07 –1.38 Meetings Minutes Press Conferences Unconventional NOTE: Underlying data are expressed as a percent change in the index (S&P 500) or price (10-year Treasury futures) over the event window. These percent changes are then summed, by category, over the full sample. For FOMC meetings, minutes, and unconventional policy measures, the window is –/+15 minutes. For press conferences, the window is –15/+60 minutes. For illustrative purposes, other public remarks were removed from the figure because of very high cumulative change values. to meeting statements than the release of FOMC minutes. The right side of Figure 6B shows that the bond market’s responses to the Chair’s press conferences and the Chair’s remarks are appreciably larger than to non-Chair Fed governors and Reserve Bank presidents. Figure 7 plots the cumulative changes for FOMC meeting statements, minutes, press conferences, and unconventional monetary policy announcements. We exclude other events for illustrative purposes, as they exhibit very high cumulative change values. Similar to the findings in Figures 6A and 6B, unconventional policies are associated with large stock and bond market responses during our sample. The cumulative change in stock prices associated with FOMC press conferences is also relatively large and positive. However, for FOMC meeting statements and the release of FOMC minutes, the cumulative response of stock prices is negative, with the response of the latter more than double the former. The response of bond futures prices to FOMC meeting statements is of the same magnitude as the minutes, but, again, far smaller than to unconventional policies. For Chair press conferences, the near-zero cumulative change is not a function of the bond futures market ignoring this information; rather, it is the result of large, positive price reactions negating large, negative price reactions over the sample. In summary, the empirical analysis presented in this article suggests that stock and bond markets respond to a variety of Fed communication events, especially FOMC meeting statements, FOMC press conferences, and remarks by the Fed Chair. CONCLUSION Clear and concise communication of monetary policy helps the Fed achieve its congressionally mandated goals of price stability, maximum employment, and stable long-term 86 Second Quarter 2019 Federal Reserve Bank of St. Louis REVIEW Kliesen, Levine, Waller interest rates. It does so by helping to reduce uncertainty about the future direction of policy. This helps to reduce distortions in market pricing, thereby improving the efficient allocation of resources by firms, households, and governments. This article has examined the various dimensions of Fed communication with the public and financial markets. This includes documenting how the Fed’s communication with the public has evolved over time. Using both daily and intraday data, our empirical analysis documents how Fed communication affects key financial market variables. We find that Fed communication is associated with changes in prices of financial market instruments such as Treasury securities and equity prices. How ever, this effect varies by type of communication, by type of instrument, and by who is doing the speaking. Perhaps not surprisingly, we find that the largest financial market reactions tend to be associated with communication by Fed Chairs rather than by other Fed governors and Reserve Bank presidents and with FOMC meeting statements rather than FOMC minutes. n NOTES 1 The occasion was a hearing of the Committee on Finance and Industry. According to Ahamed (2009), this was a select committee to investigate the British banking system in the aftermath of the 1929 collapse in stock prices and the poor performance of the British economy. See Ahamed (2009, pp. 371-72). 2 Ahamed (2009, p. 371). 3 For example, see Cochrane (2017), Cogan and Shultz (2017), and Derby (2017). 4 From a 2003 speech by Governor Yellen, as quoted in Holmes (2013). Holmes argues that central bankers, both in the United States and elsewhere, have increasingly (even before the Financial Crisis) moved away from traditional instruments, such as interest rates or exchange rates, toward “communicative experiments” designed to influence public sentiments and expectations. 5 For an early discussion of this phenomenon applied to the Reserve Bank of New Zealand and the FOMC, see Guthrie and Wright (2000) and Thornton (2004), respectively. 6 See Blinder et al. (2001). 7 See Bernanke, Reinhart, and Sack (2004). A synthesis of Bernanke’s views was presented in a 2013 speech, “Communication and Monetary Policy.” 8 See, for example, Neely (2015) or Bauer and Rudebusch (2014). 9 We define the public as anyone who uses expectations about future monetary policy actions as an input into their decisionmaking process. 10 Current Chairman Jerome Powell expanded on this innovation, announcing that press conferences will be held after every FOMC meeting beginning in January 2019. 11 For example, the ROPA for the January 15, 1970, meeting was released on April 15, 1970, a three-month lag. The FOMC ceased publication of the ROPA after the December 22, 1992, meeting. Beginning in 1993, the ROPA was effectively folded into the FOMC minutes and released with a much shorter lag. For more historical detail, see https://www.federalreserve.gov/monetarypolicy/fomc_historical.htm. 12 This statement, and subsequent policy statements, can be found on the Board of Governors of the Federal Reserve System website: https://www.federalreserve.gov/monetarypolicy/fomc_historical_year.htm. 13 Wynne (2013) provides a short history of the FOMC’s communication practices. 14 See Gavin and Mandal (2000). 15 The ZLB is the period when the target range for the intended federal funds rate was 0 percent to 0.25 percent. The ZLB period ended at the December 2015 FOMC meeting. Federal Reserve Bank of St. Louis REVIEW Second Quarter 2019 87 Kliesen, Levine, Waller 16 The monetary easing commenced in August 2007, when the Board of Governors voted to reduce the discount rate by 50 basis points. See https://www.stlouisfed.org/financial-crisis/full-timeline. 17 Wynne (2013) documented that the Fed used forward-looking language to shape expectations before the Financial Crisis. For example, in 2003, the FOMC noted that “policy accommodation can be maintained for a considerable period” in its post-meeting statement. Most economists and policymakers, though, would probably agree that the use was most pronounced during the ZLB era. The FOMC’s forward guidance policy was influenced importantly by Woodford (2001) and Eggertsson and Woodford (2003). 18 The midpoint of the central tendency excludes the three highest and three lowest projections for each variable in each year. 19 Moreover, following the November 3, 2010, meeting, the policy statements crafted under the leadership of Chair man Bernanke began to emphasize the economy’s current performance and expected outcome relative to the Fed’s “statutory mandate” of price stability and maximum employment. This was a departure from the Greenspan era, when the statement rarely—if ever—mentioned the Fed’s statutory mandate. The November 2010 statement was also noteworthy because it announced the second round of the large-scale asset purchase program (QE2). 20 The average Flesch-Kincaid scores during this period were very close to the reported medians. 21 See also Jansen (2011). Others have found similar findings for other major central bank communications. See Coenen at al. (2017), Haldane (2017), and Ehrmann and Talmi (2017). 22 Meltzer (2009) documents a 1962 FOMC meeting where communication with the public was discussed. Then- Chairman Martin favored increased communication with the public as a way to counter academic critics of Fed policy who he believed were mistaken in their analysis. However, Martin opposed regular (quarterly) policy reviews because there were instances where the FOMC would not wish to explain its decision. See discussion on p. 337 of Meltzer (2009). 23 The source of this repository is Bloomberg. More detail on this source, and its limitations, is provided in the empirical analysis section. 24 As noted above, the declining number of Fed governors speaking on the same day reflects to some extent the dwindling number of years when there was full complement of governors (seven) serving on the FOMC. 25 See Olson and Wessel (2016). 26 See https://www.federalreserve.gov/monetarypolicy/mpr_default.htm. 27 See Fawley and Neely (2014). 28 See Bullard (2016), Evans (2017), and Kashkari (2017). 29 Conference calls and unscheduled FOMC meetings were excluded from the analysis. 30 For simplicity, we only focus on announcements directly related to a large-scale asset purchase program. These include the following: QE1 announcement and expansion, QE2 announcement, Maturity Extension Program announcement and expansion, and QE3 announcement and expansion. 31 The initial QE1 announcement, which was made on November 25, 2008, did not coincide with an FOMC meeting. 32 For example, the release of nonfarm payroll employment, CPI inflation, and GDP (advance, second, and third esti- mates) all occur before or at the market open. 33 As previously mentioned, our database for public remarks comes from Bloomberg; it begins in 1998. For consis- tency, we start all Fed communication event categories at this date, where applicable. Only public remarks made during market hours are included in the event study. If Bloomberg did not provide a time for an event, and this time could not be identified by other sources, the event was removed from the sample. We considered merging Bloomberg’s repository with other databases, but since there was not a consistent time horizon or speaker overlap, we did not proceed with this approach. In particular, we examined databases from the Board of Governors of the Federal Reserve System and the Federal Reserve Bank of St. Louis’s “FOMC Speak.” The Board’s database does not include public remarks made by Bank presidents, while “FOMC Speak” only begins in 2010. Merging either database with Bloomberg’s would result in an upward estimate of governors’ remarks (for the former scenario) or an upward estimate of remarks over the 2010-17 period (for the latter scenario), which would also affect Figure 3. 88 Second Quarter 2019 Federal Reserve Bank of St. Louis REVIEW Kliesen, Levine, Waller Nevertheless, we acknowledge that the Bloomberg database is only a proxy for public remarks when presenting this analysis. 34 Our dependent variable is very similar to the approach used by Andersson (2010), who used intraday data to analyze financial market responses to Federal Reserve and European Central Bank monetary policy decisions. 35 Andersson (2010) studied intraday volatility in the bond futures market and in the equity market (S&P 500 index) around FOMC statement releases from April 1999 to May 2006. He found that intraday volatility rises sharply at the time of the release of FOMC meeting statements. 36 TickWrite is the source for the intraday data used in this analysis. 37 The one exception is for press conferences, where expanding our event window noticeably increased the market reaction relative to other events. One possible explanation is that press conferences are often more than an hour long. However, a closer inspection reveals that the press conferences driving this jump in magnitude are those on June 22, 2011, and June 19, 2013. The latter was noteworthy because this is when Chairman Bernanke discussed the so-called taper tantrum that had developed in the markets in response to his Congressional testimony a month earlier. In that testimony, he raised the possibility of the FOMC beginning to taper asset purchases later that year. 38 It is not our intent to examine whether stock and bond prices may react differently to Fed communication events. We refer the reader to numerous studies on the effects of these dynamics in the interactions with monetary policy actions. For example, see Campbell and Ammer (1993), Bernanke and Kuttner (2005), Andersen et al. (2007), and Connolly, Stivers, and Sun (2005). 39 The non-event day controls in Figures 6A and 6B are constructed to have similar response windows to the events they are compared with. For example, in Figure 6A, we use a rolling event window of 30 and 75 minutes to calculate a benchmark for non-speakers and speakers, respectively. Windows that either include an event or overlap days are removed before calculating the benchmark mean absolute changes. We follow the same procedure for Figure 6B.The authors thank Chris Neely for helpful comments in this regard. REFERENCES Ahamed, Liaquat. Lords of Finance: The Bankers Who Broke the World. Penguin Books, 2009 (paperback version). Andersen, Torben G.; Bollerslev, Tim; Diebold, Francis X. and Vega, Clara. “Real-Time Price Discovery in Global Stock, Bond and Foreign Exchange Markets.” Journal of International Economics, November 2007, 73(2), pp. 251-77; https://doi.org/10.1016/j.jinteco.2007.02.004. Andersson, Magnus. “Using Intraday Data to Gauge Financial Market Responses to Federal Reserve and ECB Monetary Policy Decisions.” International Journal of Central Banking, June 2010, 6(2), pp. 117-46. Bauer, Michael D. and Rudebusch, Glenn D. “The Signaling Channel for Federal Reserve Bond Purchases.” International Journal of Central Banking, September 2014, 10(3), pp. 233-89. Bernanke, Ben S. “Deflation: Marking Sure ‘It’ Doesn’t Happen Here.” Remarks before the National Economists Club, Washington, D.C., November 21, 2002; https://www.federalreserve.gov/boarddocs/speeches/2002/20021121/. Bernanke, Ben S. “The Global Saving Glut and the U.S. Current Account Deficit.” Remarks at the Sandridge Lecture, Virginia Association of Economists, Richmond, Virginia, March 10, 2005; https://www.federalreserve.gov/boarddocs/speeches/2005/200503102/. Bernanke, Ben S. and Kuttner, Kenneth N. “What Explains the Stock Market’s Reaction to Federal Reserve Policy?” Journal of Finance, May 2005, 60(3), 1221-257; https://doi.org/10.1111/j.1540-6261.2005.00760.x. Bernanke, Ben S.; Reinhard, Vincent R. and Sack, Brian P. “Monetary Policy Alternatives at the Zero Bound: An Empirical Assessment.” Brookings Papers on Economic Activity, January 2004; https://www.brookings.edu/wp-content/uploads/2004/06/2004b_bpea_bernanke.pdf. Blinder, Alan; Ehrmann, Michael; Fratzscher, Marcel; De Haan, Jakob and Jansen, David-Jan. “Central Bank Communication and Monetary Policy: A Survey of Theory and Evidence.” ECB Working Paper No. 898, European Central Bank, May 2008; https://www.ecb.europa.eu/pub/pdf/scpwps/ecbwp898.pdf. Federal Reserve Bank of St. Louis REVIEW Second Quarter 2019 89 Kliesen, Levine, Waller Blinder, Alan; Goodhart, Charles; Hildebrand, Philipp; Lipton, David and Wyplosz, Charles. “How Do Central Banks Talk?” Geneva Report on the World Economy, No. 3, International Center for Monetary and Banking Studies, 2001. Bullard, James. “A New Characterization of the U.S. Macroeconomic and Monetary Policy Outlook.” Remarks delivered at the Society of Business Economists Annual Dinner, London, United Kingdom, June 30, 2016; https://www.stlouisfed.org/from-the-president/speeches-and-presentations/2016/a-new-characterization. Campbell, Jeffrey R.; Evans, Charles L.; Fisher, Jonas D.M. and Justiano, Alejandro. “Macroeconomic Effects of Federal Reserve Forward Guidance.” Brookings Papers on Economic Activity, Spring 2012, pp. 1-63; https://doi.org/10.1353/eca.2012.0004. Campbell, John Y. and Ammer, John. “What Moves the Stock and Bond Markets? A Variance Decomposition for Long-Term Asset Returns.” Journal of Finance, March 1993, 48(1), pp. 3-37; https://doi.org/10.1111/j.1540-6261.1993.tb04700.x. Cochrane, John H. “Taylor for Fed.” The Grumpy Economist Blog, October 20, 2017; https://johnhcochrane.blogspot.com/2017/10/taylor-for-fed.html. Coenen, Günter; Ehrmann, Michael; Gaballo, Gaetano; Hoffmann, Peter; Nakov, Anton; Nardelli, Stefano; Persson, Eric and Strasser, Georg. “Communication of Monetary Policy in Unconventional Times.” ECB Working Paper No. 2080, European Central Bank, June 2017; https://www.ecb.europa.eu/pub/pdf/scpwps/ecb.wp2080.en.pdf. Cogan, John F. and Shultz, George P. “The Fed Chief America Needs.” Commentary, Wall Street Journal, October 18, 2017. Connolly, Robert; Stivers, Chris and Sun, Licheng. “Stock Market Uncertainty and the Stock-Bond Return Relation.” Journal of Financial and Quantitative Analysis, March 2005, 40(1), pp. 161-94; https://doi.org/10.1017/S0022109000001782. Derby, Michael S. “Fed’s Fischer Laments ‘The Cacophony’ of Too Much Fed Speak.” WSJ Pro, April 25, 2017. Eggertsson, G. and Woodford, Michael. “The Zero Bound on Interest Rates and Optimal Monetary Policy.” Brookings Papers on Economic Activity, 2003, 34(1), pp. 139-235; https://www.brookings.edu/wp-content/uploads/2003/01/2003a_bpea_eggertsson.pdf. Ehrmann, Michael and Talmi, Jonathan. “Starting from a Blank Page? Semantic Similarity in Central Bank Communication and Market Volatility.” ECB Working Paper No. 2023, European Central Bank, February 2017; https://www.ecb.europa.eu/pub/pdf/scpwps/ecbwp2023.en.pdf?561f7f26fea2285effb36dae431188d1. Evans, Charles L. “Rationale for My Dissent at the December 2017 Meeting.” Federal Reserve Bank of Chicago, December 15, 2017. https://www.chicagofed.org/publications/speeches/2017/12-15-2017-evans-rationale-fordissent-december-fomc. Fawley, Brett W. and Neely, Christopher J. “The Evolution of Federal Reserve Policy and the Impact of Monetary Policy Surprises on Asset Prices.” Federal Reserve Bank of St. Louis Review, First Quarter 2014, 96(1), pp. 73-109; https://files.stlouisfed.org/files/htdocs/publications/review/2014/q1/fawley.pdf. Gavin, William T. and Mandal, Rachel J. “Inside the Briefcase: The Art of Predicting the Federal Reserve.” Federal Reserve Bank of St. Louis Regional Economist, July 2000, pp. 4-9; https://www.stlouisfed.org/publications/regional-economist/july-2000/inside-the-briefcase-the-art-of-predicting-the-federal-reserve. Greenspan, Alan. “Remarks before the Economic Club of Chicago.” Chicago, Illinois, October 19, 1995; https://fraser.stlouisfed.org/scribd/?item_id=8549&filepath=/files/docs/historical/greenspan/ Greenspan_19951019.pdf. Greenspan, Alan. “The Challenge of Central Banking in a Democratic Society.” Remarks at the Annual Dinner and Francis Boyer Lecture of the American Enterprise Institute for Public Policy Research, Washington, D.C., December 5, 1996; https://fraser.stlouisfed.org/scribd/?item_id=8581&filepath=/files/docs/historical/greenspan/Greenspan_19961205.pdf. Guthrie, Graeme and Wright, Julian. “Open Mouth Operations.” Journal of Monetary Economics, October 2000, 46(2), pp. 489-516; https://doi.org/10.1016/S0304-3932(00)00035-0. Haldane, Andrew G. “A Little More Conversation, A Little Less Action.” Remarks at the Macroeconomics and Monetary Policy Conference, Federal Reserve Bank of San Francisco, March 2017; https://www.bankofengland.co.uk/speech/2017/a-little-more-conversation-a-little-less-action. 90 Second Quarter 2019 Federal Reserve Bank of St. Louis REVIEW Kliesen, Levine, Waller Holmes, Douglas R. Economy of Words: Communicative Imperatives in Central Banks. University of Chicago Press, 2013; https://doi.org/10.7208/chicago/9780226087764.001.0001. Jansen, David-Jan. “Does the Clarity of Central Bank Communication Affect Volatility in Financial Markets? Evidence from Humphrey-Hawkins Testimonies.” Contemporary Economic Policy, 2011, 29, pp. 494-509; https://doi.org/10.1111/j.1465-7287.2010.00238.x. Kashkari, Neel. “Why I Dissented a Third Time.” Federal Reserve Bank of Minneapolis, December 18, 2017; https://www.minneapolisfed.org/news-and-events/messages/why-i-dissented-a-third-time. Kocherlakota, Narayana. “Maybe Central Banks Are Too Independent.” Bloomberg View, August 7, 2017; https://www.bloomberg.com/view/articles/2017-08-07/maybe-central-banks-are-too-independent. Lustenberger, Thomas and Rossi, Enzo. “Does Central Bank Transparency and Communication Affect Financial and Macroeconomic Forecasts?” Swiss National Bank Working Paper, December 2017; https://www.snb.ch/n/mmr/reference/working_paper_2017_12/source/working_paper_2017_12.n.pdf. Meltzer, Allan H. A History of the Federal Reserve. Volume 2, Book 1: 1951-1969. University of Chicago Press, 2009. Neely, Christopher J. “Unconventional Monetary Policy Had Large International Effects.” Journal of Banking and Finance, March 2015, 52, pp. 101-11; https://doi.org/10.1016/j.jbankfin.2014.11.019. Olson, Peter and Wessel, David. “Federal Reserve Communications: Survey Results.” Hutchins Center on Fiscal and Monetary Policy at Brookings, November 2016; https://www.brookings.edu/wp-content/uploads/2016/11/fed-communications-survey-results.pdf. Shin, Hyun Song. “Can Central Banks Talk Too Much?” Speech at the European Central Bank conference “Communications Challenges for Policy Effectiveness,” Frankfurt, November 14, 2017. Thornton, Daniel L. “The Fed and Short-Term Rates: Is It Open Market Operations, Open Mouth Operations or Interest Rate Smoothing?” Journal of Banking and Finance, 2004, 28(4), pp. 475-98; https://doi.org/10.1016/S0378-4266(02)00409-0. Woodford, Michael. “Monetary Policy in the Information Economy.” Presented at the Federal Reserve Bank of Kansas City symposium “Economic Policy for the Information Economy,” August 30–September 1, 2001; https://www.kansascityfed.org/publicat/sympos/2001/papers/S02wood.pdf. Wynne, Mark A. “A Short History of FOMC Communication.” Federal Reserve Bank of Dallas Economic Letter, September 2013, 8(8); https://www.dallasfed.org/research/eclett/2013/el1308.cfm#n8. Federal Reserve Bank of St. Louis REVIEW Second Quarter 2019 91 92 Second Quarter 2019 Federal Reserve Bank of St. Louis REVIEW International Trade Openness and Monetary Policy: Evidence from Cross-Country Data Fernando Leibovici This article studies the extent to which open economies conduct monetary policy differently from economies that are relatively closed to international trade. I first estimate country-specific Taylor rules for 26 economies, following the approach of Clarida, Galí, and Gertler (1998 and 2000). Then, I examine the extent to which open economies assign systematically different weights to changes in economic outcomes, such as inflation and the output gap, than their closed economy counterparts do. I find that open economies respond less strongly to changes in expected inflation than relatively closed economies do and that the response to changes in the output gap is independent of the degree of trade openness. Moreover, I find that this difference between closed and open economies may be accounted for by the higher weight open economies give to changes in the real exchange rate, whereby these economies are more likely to decrease the nominal interest rate when the real exchange rate is relatively appreciated. (JEL E5, F1, F41) Federal Reserve Bank of St. Louis Review, Second Quarter 2019, 101(2), pp. 93-113. https://doi.org/10.20955/r.101.93-113 1 INTRODUCTION Open economies are typically exposed to different sources of shocks than economies relatively closed to international trade1: To the extent that a country trades goods with the rest of the world, economic conditions in its trade partners and changes in international relative prices may affect domestic economic activity. Insofar as central banks design monetary policy to moderate business cycle fluctuations, the different nature of business cycles in open economies has led many to ask, to what extent should central banks in open economies conduct monetary policy differently from their closed-economy counterparts? In a recent study, Leibovici and Santacreu (2015) find that openness should indeed be an important consideration for the design of monetary policy. In particular, we show that international trade fluctuations play a key role in accounting for the optimal monetary policy that central banks in open economies should conduct. This finding suggests that trade openness should be a key factor in optimal monetary policy design. Fernando Leibovici is an economist at the Federal Reserve Bank of St. Louis. The author thanks Jonas Crews for excellent research assistance and Ana Maria Santacreu for helpful discussions. © 2019, Federal Reserve Bank of St. Louis. The views expressed in this article are those of the author(s) and do not necessarily reflect the views of the Federal Reserve System, the Board of Governors, or the regional Federal Reserve Banks. Articles may be reprinted, reproduced, published, distributed, displayed, and transmitted in their entirety if copyright notice, author name(s), and full citation are included. Abstracts, synopses, and other derivative works may be made only with prior written permission of the Federal Reserve Bank of St. Louis. Federal Reserve Bank of St. Louis REVIEW Second Quarter 2019 93 Leibovici More generally, while theoretical studies have also largely concluded that open economies should indeed conduct monetary policy differently, studies differ in their policy recommendations. On the one hand, Clarida, Galí, and Gertler (2002) and Corsetti, Dedola, and Leduc (2010) show that in a specific class of economic models the design of monetary policy in open economies is “isomorphic” to the conduct of monetary policy in a closed economy: Central banks should only respond to changes in inflation and the output gap, but they may respond differently to these depending on the degree of trade openness. On the other hand, morerecent studies have shown that this need not necessarily be the case in more realistic environments (Faia and Monacelli, 2008, De Paoli, 2009, and Lombardo and Ravenna, 2014) in which central banks in open economies may also want to respond to changes in other variables such as the real exchange rate. Yet, while much work has been devoted to understanding the normative question about whether and how central banks in open economies should design monetary policy differently from their closed-economy counterparts, much less is known about the positive question on whether they do indeed conduct policy differently. Therefore, in this article I ask, to what extent do central banks of open economies conduct monetary policy differently from those of closed economies? The answer to this question has important implications. To the extent that the relationship between trade openness and monetary policy observed in the data differs from the relationship implied by the optimal policy analysis of structural economic models, this may suggest that some countries are indeed conducting monetary policy suboptimally. Alternatively, to the extent that central banks of open economies conduct monetary policy differently from that implied by standard models, this may reflect economic channels or concerns of policymakers that may not be explicitly considered in the economic models, but which may yet be important for understanding the link between trade openness and monetary policy. The goal of this article is, thus, to investigate the empirical relationship between trade openness and the design of monetary policy using cross-country time-series data for the period 1980 to 2006. In the first step of the analysis, I compute empirical measures that allow me to characterize and compare the nature of monetary policy across different countries over this period. Then, I use these empirical measures to examine whether open economies conduct monetary policy differently from closed economies. My starting point to characterizing the nature of monetary policy across countries is the standard Taylor rule (1993), which specifies a link among nominal interest rates, inflation, and the output gap. While an increasing number of countries use the nominal interest rate as their preferred instrument for the conduct of monetary policy combined with some sort of inflation and/or output-related targets, the approach I take here is more broad. In particular, in this article I do not interpret the Taylor rule coefficients as structural parameters that govern the response of interest rates to changes in inflation and output but, instead, use the Taylor rule as a device to summarize the statistical properties of how interest rates, inflation, and output behave in the time series across countries regardless of the particular underlying monetary policy instruments, outcome-targeting regimes, or exchange rate regimes in place in each country. 94 Second Quarter 2019 Federal Reserve Bank of St. Louis REVIEW Leibovici The goal of this approach is to characterize monetary policy across countries through a common lens in order to facilitate cross-country comparability. This advantage comes at the cost of forcing the analysis to abstract from differences in monetary policy across countries not captured by differences in the empirical Taylor rule coefficients. Thus, I first estimate country-specific Taylor rules for 26 countries with differing degrees of openness to international trade (as measured by the ratio of aggregate exports to gross domestic product [GDP]) using time-series data on interest rates, inflation, and output. I follow the generalized method of moments (GMM) approach of Clarida, Galí, and Gertler (1998 and 2000), which allows me to obtain estimates of the explicit or implicit response of interest rates to changes in inflation and the output gap. Then, I examine whether the statistical relationship among interest rates, inflation, and output differs systematically based on the degree of trade openness. I begin the analysis by considering a baseline specification of the Taylor rule, which specifies the relationship among nominal interest rates, expected inflation, and the current output gap. In addition, I include lagged nominal interest rates, following a large literature that has observed that central banks adjust interest rates gradually over time. I find considerable dispersion across countries in the empirical relationship between nominal interest rates and expected inflation, as well as between nominal interest rates and the output gap. Moreover, I find that these relationships differ systematically across countries based on their degree of international trade openness: Nominal interest rates in open economies respond systematically less to changes in expected inflation than they do in closed economies. I find no systematic relationship between the response of a country’s nominal interest rate to changes in its output gap and the degree of the country’s openness. While the lower response of nominal interest rates in open economies to changes in inflation may reflect that such countries are less concerned about inflation, it may also reflect that these countries actually respond to changes in variables other than inflation. To investigate this possibility, I reconduct the analysis, extending the Taylor rule to include a trade-related variable that may affect how open economies conduct monetary policy, as suggested by previous studies (Faia and Monacelli, 2008, and De Paoli, 2009, among others)—the real exchange rate. I find that, indeed, open economies are systematically more likely to adjust their nominal interest rate in response to changes in the real exchange rate. These findings suggest that open economies do conduct monetary policy differently from their closed-economy counterparts. First, I find that open economies respond relatively less to changes in inflation. Second, I find that open economies respond relatively more to changes in the real exchange rate. And, finally, I find that the degree of interest rate smoothing and the response of nominal interest rates to changes in the output gap do not vary systematically with the degree of international trade openness. This article contributes to a growing empirical literature that studies the relationship between trade openness and monetary policy, such as Lubik and Schorfheide (2007), Berument, Konac, and Senay (2007), and Basilio (2013). This article is also related to empirical papers aimed at estimating Taylor rules across countries. I follow very closely the estimation approach of Clarida, Galí, and Gertler (1998 and 2000), who apply it to the United States, Japan, Germany, Federal Reserve Bank of St. Louis REVIEW Second Quarter 2019 95 Leibovici France, Italy, and the United Kingdom. Also related are Torres (2003), Hayo and Hofmann (2006), Yazgan and Yilmazkuday (2007), and Kahn (2012). More broadly, this article is also related to a large theoretical and quantitative literature that investigates the extent to which open economies should conduct monetary policy differently. Corsetti, Dedola, and Leduc (2010) provide a broad discussion of many of the studies in this literature. More recently, Faia and Monacelli (2008), De Paoli (2009), Lombardo and Ravenna (2014), and Leibovici and Santacreu (2015) investigate this question in richer and more realistic economic environments. The rest of this article is structured as follows. Section 2 presents the economic framework, and Section 3 presents my approach to estimating it. Section 4 presents the data that I use to estimate the economic framework and describes the details of the implementation. Sections 5 and 6 present the results, and Section 7 concludes. 2 THE TAYLOR RULE The starting point of the analysis is the Taylor rule, an equation that specifies the nominal short-term interest rate target as a function of three variables: (i) the long-run equilibrium nominal rate, (ii) deviations of inflation from an inflation target, and (iii) deviations of output from a target level of output. In addition, I consider an extension of the standard Taylor rule to allow for the possibility that nominal interest rates also respond to other variables. Mathe matically, we have (1) Rt* = R + β ⎡⎣ E (π t+1 | Ωt ) − π * ⎤⎦ + γ ⎡⎣ E ( yt | Ωt ) − yt* ⎤⎦ + ξ z t , – where R is the long-run equilibrium nominal rate; πt+1 is the rate of inflation between periods t and t+1; yt is real output; π* is the target level of inflation; yt* is the target level output; and zt is a vector of additional variables that the monetary authority may respond to. Moreover, E is the expectation operator and Ωt is the information set available to the central bank at the time it sets interest rates. A variety of specifications of the Taylor rule have been considered in the literature; here, I focus on the specification studied by Clarida, Galí, and Gertler (1998 and 2000). I assume that, every period, the effective short-term nominal interest rate Rt adjusts partially to the nominal interest rate target Rt* according to the following AR(2) process: (2) Rt = (1− ρ1 − ρ 2 ) Rt* + ρ1Rt−1 + ρ 2 Rt−2 + ν t , where ρ1 (0,1) and ρ2 (0,1) capture the degree of interest rate smoothing, vt is a zeromean independent and identically distributed (i.i.d.) shock to the nominal interest rate, and ρ1 + ρ2 < 1. Then, given the above expressions for the target short-term nominal interest rate (equation (1)) and the nominal interest rate (equation (2)), the actual nominal short-term interest rate can be expressed as 96 Second Quarter 2019 Federal Reserve Bank of St. Louis REVIEW Leibovici { Rt = (1− ρ1 − ρ 2 ) R + β ⎡⎣ E (π t+n | Ωt ) − π * ⎤⎦ + γ ⎡⎣ E ( yt | Ωt ) − yt* ⎤⎦ + ξ z t + ρ1Rt−1 + ρ 2 Rt−2 + ν t . (3) } Thus, to the extent that a country’s nominal interest rate follows equation (3), its evolution is characterized by – (i) a long-run value R for the nominal interest rate, (ii) a set of outcome variables X t = E (π t+1 | Ωt ) − π *, E ( yt | Ωt ) − yt* , Rt−1, Rt−2,z t on which the nominal interest rate target depends, and (iii) a set of coefficients θ = {β ,γ , ρ1, ρ 2 , ξ } that dictate the response of the nominal interest rate target to changes in the outcome variables Xt . { } Throughout the rest of the article, I use this framework as a lens to characterize differences in the design of monetary policy across countries. In particular, given a set of outcome variables X t = E (π t+1 | Ωt ) − π *, E ( yt | Ωt ) − yt* , Rt−1, Rt−2,z t on which the nominal interest rate is assumed to depend, I estimate the set of coefficients θ = {β ,γ , ρ1, ρ 2 , ξ } that dictates its response to changes in the outcome variables. From the lens of this approach, differences in the design of monetary policy across countries boil down to differences in the set of estimated coefficients θ. Then, I examine the extent to which open economies conduct monetary policy differently from closed ones by investigating whether the set of coefficients θ varies systematically with the degree of international trade openness across countries. Importantly, note that while I use the framework above to characterize differences in the design of monetary policy across countries, I do not necessarily restrict attention to countries with explicit interest rate targeting rules. In particular, I do not interpret the Taylor rule coefficients θ as structural parameters that govern the response of nominal interest rates to changes in inflation and output but, instead, use the specification above as a device to summarize the statistical properties of how interest rates, inflation, and output behave in the time series across countries regardless of the particular underlying monetary policy instruments, outcome- targeting regimes, or exchange-rate regimes in place in each country. { } 3 ESTIMATION APPROACH To estimate the Taylor rule described above (equation (3)) for a given country, I follow the approach of Clarida, Galí, and Gertler (1998 and 2000). The objective is to estimate the vector of coefficients θ = {β ,γ , ρ1, ρ 2 , ξ } that dictates the response of the nominal interest rate to changes in the outcome variables Xt . The first problem I face in estimating θ using equation (3) is that E (π t+1 | Ωt ) and E ( yt | Ωt ) are unobservable. While inflation and output are regularly measured by statistical agencies across countries, this is not typically the case for the expected value of inflation and output at the time that monetary policy decisions are made. To address this challenge, I rewrite equation (3) in terms of observable variables, following Clarida, Galí, and Gertler (1998 and 2000): Federal Reserve Bank of St. Louis REVIEW Second Quarter 2019 97 Leibovici ( ) ( ) Rt = (1− ρ1 − ρ 2 ) ⎡⎣ R − βπ * + βπ t+1 + γ yt − yt* + ξ z t ⎤⎦ + ρ1Rt−1 + ρ 2 Rt−2 + ε t , (4) where I refer to yt – yt* as the output gap and the error term εt is now given by ε t = vt − (1− ρ1 − ρ 2 ) β ⎡⎣π t+1 − E (π t+1 | Ωt ) ⎤⎦ (5) { } − (1− ρ1 − ρ 2 )γ ⎡⎣ yt − yt* ⎤⎦ − ⎡⎣ E ( yt | Ωt ) − yt* ⎤⎦ . Having addressed the first challenge, the question now is how to estimate equation (4) given data on inflation πt+1 and the output gap yt – yt*. Estimating it by ordinary least squares (OLS) would result in biased estimates since we would be violating the exogeneity assumption, which in this case requires that E ⎡⎣ε t | π t+1,yt − yt* , z t ,Ωt ⎤⎦ = 0. To see this, note that even though vt may be a mean-zero i.i.d. variable, equation (5) implies that E ⎡⎣ε t | π t+1 , yt − yt* , z t ,Ωt ⎤⎦ = − (1− ρ1 − ρ 2 ) β ⎡⎣π t+1 − E (π t+1 | Ωt ) ⎤⎦ { } − (1− ρ1 − ρ 2 )γ ⎡⎣ yt − yt* ⎤⎦ − ⎡⎣ E ( yt | Ωt ) − yt* ⎤⎦ . That is, E ⎡⎣ε t | π t+1,yt − yt* , z t ,Ωt ⎤⎦ is a function of the forecast errors for inflation and the output gap, which generically need not be equal to zero in all states of the world. The violation of the exogeneity assumption required by OLS to produce unbiased estimates can equivalently be expressed as { ( ) ( } ) % | Ω ≠ 0, E ⎡ Rt − (1− ρ1 − ρ 2 ) ⎡⎣ R − βπ * + βπ t+1 + γ yt − yt* + ξ z t ⎤⎦ − ρ1Rt−1 − ρ 2 Rt−2 ⎤ X ⎣ ⎦ t t { } % t = π t+1,yt − yt* ,z t ,Rt−1,Rt−2 . where X To address this second challenge, I follow Clarida, Galí, and Gertler (1998 and 2000) in pursuing an instrumental-variables approach. This approach requires one to find a vector of % t but that are uncorrelated with ε or, equivalently, variables ut that are correlated with X t uncorrelated with νt , π t − E (π t | Ωt ), and ⎡⎣ yt − yt* ⎤⎦ − ⎡⎣ E ( yt | Ωt ) − yt* ⎤⎦ . In other words, the objective of this approach is to find variables ut such that the following two properties are satisfied: { } % t ut | Ωt ≠ 0, (i) non-zero correlation between instruments and outcome variables: E X and (ii) zero correlation between instruments and the error term εt : { ( ) ( ) } (6) E ⎡ Rt − (1− ρ1 − ρ 2 ) ⎡⎣ R − βπ * + βπ t+1 + γ yt − yt* + ξ z t ⎤⎦ − ρ1Rt−1 − ρ 2 Rt−2 ]ut | Ωt = 0. ⎣ Following Clarida, Galí, and Gertler (1998), possible elements of ut include any lagged variable that may help to forecast inflation or the output gap. Why? By definition, the forecast error consists of the difference between the realization of a variable and its expected value conditional on the information set at the time that the nominal interest rate is determined. Then, to the extent that one identifies a lagged variable that may help to forecast either πt+1 or yt – yt*, such a variable should not be associated with the forecast error. Other candidate instruments ut 98 Second Quarter 2019 Federal Reserve Bank of St. Louis REVIEW Leibovici include any contemporaneous variables that are uncorrelated with the current interest rate shock vt . Given a set of instruments ut that satisfy conditions (1) and (2) above, I estimate equation (4) following a GMM approach. Let j = 1,…,J index the different instruments in vector ut . First, I define the population moment corresponding to instrument j as { ( ) ( } ) M j = E ⎡ Rt − (1− ρ1 − ρ 2 ) ⎡⎣ R − βπ * + βπ t+1 + γ yt − yt* + ξ z t ⎤⎦ − ρ1Rt−1 − ρ 2 Rt−2 ⎤ ut,j | Ωt = 0. ⎣ ⎦ Second, I construct its empirical counterpart: { } T % j = 1 ∑ Rt − (1− ρ1 − ρ 2 ) ⎡ R − βπ * + βπ t+1 + γ yt − yt* + ξ z t ⎤ − ρ1Rt−1 − ρ 2 Rt−2 ut,j . M ⎣ ⎦ T t=1 ( ) ( ) Third, I construct the moment condition mj corresponding to instrument j: % j − Mj mj = M % j −0 =M % . =M j Fourth, I stack all moment conditions mj for j = 1,…,J into a vector m. Finally, given a symmetric positive-definite weighting J×J matrix W,2 the GMM estimator is given by the vector θ = {β ,γ , ρ1, ρ 2 , ξ } of parameters that solves the following problem: min {β ,γ , ρ1, ρ2 ,ξ } m T Wm, where a variable with superscript T denotes the transpose of the variable. 4 DATA To implement the estimation procedure described in the previous section, I use quarterly cross-country time-series data collected by the International Monetary Fund (IMF) and the Organisation for Economic Co-operation and Development (OECD) and accessed through Haver Analytics.3 I select the variables and sources used as well as the details of the specification that I estimate under two guiding principles. First, my goal is to obtain country-specific estimates of the Taylor rule coefficients θ that can be compared across countries. Therefore, I restrict attention to variables and data sources that maximize the number of countries available while ensuring that variables are measured under a methodology that is as similar as possible across countries.4 My second goal is, to the extent possible, to follow the estimation approach of Clarida, Galí, and Gertler (1998 and 2000), with the intention of maximizing the comparability of my findings with previous estimates from the literature. In this section, I present the data used to estimate equation (3) under the constraint ξ = zt = 0, which is the standard Taylor rule formulation. I present the corresponding results in Section 5 and consider extensions of this specification in Section 6. Federal Reserve Bank of St. Louis REVIEW Second Quarter 2019 99 Leibovici 4.1 Taylor Rule Variables I measure the nominal interest rate Rt using data on central bank policy rates collected by the IMF (Haver code: C###IC@IFS). As described by the IMF (n.d.), “The central bank policy rate (CBPR) is the rate that is used by the central bank to implement or signal its monetary policy stance. It is most commonly set by the central banks’ policy making committees (e.g., the Federal Open Market Committee).”5 I measure the realized inflation rate πt+1 as the quarterly log change in the GDP deflator between period t+1 and period t; I obtain data on the GDP deflator from the IMF and OECD (Haver codes: C###GJ@IFS for the IMF series and C###GPI@OECDNAQ for the OECD series).6 To measure the output gap yt – yt*, I compute the log difference of real GDP from its country-specific quadratic trend.7 I obtain data on real GDP from the IMF and OECD (Haver codes: C###GDPC@IFS for the IMF series and E###GDPC@OECDNAQ for the OECD series). 4.2 Instruments Following Clarida, Galí, and Gertler (1998 and 2000), the set of instruments that I use consists of two types of variables. On the one hand, I include the first four lags of each of the Taylor rule variables as part of the instrument set. In particular, I include the following variables as instruments: ( )( )( )( ) * * * * * * * * Rt−1,Rt−2,Rt−3,Rt−4, yt−1 − yt−1 , yt−2 − yt−2 , yt−3 − yt−3 , yt−4 − yt−4 ,π t−1 , π t−2 , π t−3 , π t−4 . On the other hand, I include lags of other variables that may forecast changes in the Taylor rule variables but that are unlikely to be correlated with forecast errors. Specifically, I include lags of two variables not included in the baseline specification of the Taylor rule: an index of world commodity prices ωt and country-specific effective real exchange rates Qt . To measure world commodity prices, I use the S&P Goldman Sachs Commodity Index (S&P GSCI Com modity Nearby Index; Haver code GSCI@USECON). To measure country-specific effective real exchange rates, I use data from the IMF and OECD (Haver codes: C###EIRC@IFS for the IMF series and C###FXEF@OECDMEI for the OECD series) and the Bank for International Settlements (BIS).8 In particular, I include the following variables as instruments: ΔlnQt−1,ΔlnQt−2,ΔlnQt−3 ,ΔlnQt−4,Δlnω t−1,Δlnω t−2,Δlnω t−3,Δlnω t−4 , where Δlnxt−k = lnxt−k − lnxt−k−1 for any variable x and integer k. I adopt the convention throughout that real exchange rates are measured as the relative price of a domestic consumption basket relative to a foreign consumption basket (both in domestic units), so that an increase in the real exchange rate consists of a real appreciation. 4.3 Cleaning Up the Data Before using these variables to estimate country-specific Taylor rules following the GMM approach described in the previous section, I apply a few filters to ensure that there is sufficient data available for each country as well as to clean the variables used. 100 Second Quarter 2019 Federal Reserve Bank of St. Louis REVIEW Leibovici First, I restrict attention to the period 1980:Q1 to 2006:Q4.9 For each of the variables that has data available from both the OECD and the IMF, I examine for each country whether there are data available from the OECD for at least 36 consecutive quarters. If so, then I use data from the OECD for that country. Otherwise, I examine whether there are data available from the IMF for at least 36 consecutive quarters. If so, then I use data from the IMF for that country; otherwise, I exclude the country from the analysis.10 Second, I seasonally adjust the level of all of these variables except Rt, using the X13 seasonal adjustment procedure developed by the U.S. Census Bureau. I adjust this procedure to avoid controlling for U.S.-specific seasonality patterns that may not be applicable to other countries (e.g., U.S. holidays). Finally, I restrict attention to countries with at least 36 consecutive quarters of data in which all of the Taylor rule variables and instruments are available (Rt ,yt ,πt ,ωt ,Qt ). The countries and time periods used throughout the analysis are displayed in Table 1. 4.4 Trade Openness and GDP Per Capita Table 1 Countries and Periods Country First period Last period Australia 1980:Q2 2006:Q4 Austria 1980:Q2 1998:Q4 Bolivia 1996:Q2 2006:Q4 Canada 1993:Q1 2006:Q4 Chile 1995:Q1 2006:Q4 Czech Republic 1996:Q1 2006:Q4 Denmark 1980:Q2 2006:Q4 Germany 1980:Q2 1998:Q4 Hungary 1995:Q2 2006:Q4 Iceland 1994:Q1 2006:Q4 India 1997:Q2 2006:Q4 Indonesia 1990:Q2 2006:Q4 Israel 1995:Q2 2006:Q4 Italy 1980:Q2 1998:Q4 Japan 1980:Q2 2006:Q4 Latvia 1995:Q2 2006:Q4 Netherlands 1980:Q2 1993:Q4 Norway 1980:Q2 2006:Q4 Poland 1998:Q1 2006:Q4 Portugal 1980:Q2 1998:Q4 In order to examine the extent to which monetary Slovenia 1996:Q4 2006:Q4 policy in open economies is systematically different, South Africa 1980:Q2 2006:Q4 I measure openness to international trade as the Spain 1984:Q1 1998:Q4 average ratio of nominal exports to nominal GDP Switzerland 1980:Q2 2006:Q4 across the sample period (Haver codes: C###GE@IFS and C###GDP@IFS for the IMF series and United Kingdom 1980:Q2 2006:Q4 A###X@OECDNAQ and A###GDP@OECDNAQ United States 1982:Q4 2006:Q4 for the OECD series), both in domestic units. How ever, all the findings are robust to alternative measures of trade openness, such as the average ratio of nominal exports plus nominal imports to nominal GDP. In the next section, I also sometimes control for the level of economic development, as proxied by GDP per capita, which I obtain from Penn World Tables 9.0 (output-side real GDP, purchasing-power-parity adjusted, in 2011 U.S. dollars). 4.5 Implementation Throughout the next sections, I execute the GMM estimation approach described above by applying the two-step GMM estimator implemented by Stata’s “gmm” command with correction for heteroskedasticity and serial correlation. That is, first I obtain parameter estimates Federal Reserve Bank of St. Louis REVIEW Second Quarter 2019 101 Leibovici based on an initial weighting matrix. Then, I compute a new weighting matrix based on those estimates. Finally, I reestimate parameters based on that weighting matrix. I correct for hetero skedasticity and serial correlation using the Newey-West kernel with four lags. 5 BASELINE TAYLOR RULE In this section, I begin to analyze the extent to which open economies conduct monetary policy differently from their closed-economy counterparts. To do so, I use the Taylor rule as the lens to measure differences in the conduct of monetary policy across countries. In particular, in this section I restrict attention to a Taylor rule where the nominal interest rate responds to changes in expected inflation, the output gap, and lagged values of the nominal interest rate (as in Clarida, Galí, and Gertler, 1998 and 2000). 5.1 Taylor Rule Coefficients The first step of the analysis consists of following the estimation approach described above to estimate a Taylor rule for each of the countries in the sample. Table 2 presents the estimated Taylor rule coefficients corresponding to equations (3) and (4). In addition, I also report the average degree of international trade openness as measured by the aggregate exports-to-GDP ratio. 5.1.1 Validation: Estimates for the United States. To begin with, I contrast my estimated Taylor rule coefficients for the United States with estimates from previous studies in the literature. Consistent with previous studies, I find that monetary policy in the United States places a higher weight on changes in inflation than on the output gap, and nominal interest rate adjustments are smoothed over time. For instance, in the specification that uses data series closest to the ones that I use, Clarida, Galí, and Gertler (2000) estimate a weight on expected inflation equal to 1.97 (vs. 1.35), a weight on the output gap equal to 0.55 (vs. 0.62), and an impact of lagged interest rates equal to 0.76 (vs. 0.93).11 Importantly, I find that most differences between my estimates and those of Clarida, Galí, and Gertler (2000) are explained by differences in the data series used. Using exactly the same data series used by Clarida, Galí, and Gertler (2000) to estimate the second row of Table II (their baseline estimates) in their paper, I estimate a weight on expected inflation equal to 2.17 (vs. 2.15), a weight on the output gap equal to 1.23 (vs. 0.93), and an impact of lagged interest rates equal to 0.80 (vs. 0.79 ). While these data series lead to estimates that are closest to those in Clarida, Galí, and Gertler (2000), some of them are not available across a large number of countries, preventing me from conducting the cross-country analysis using these particular series. Recall that, in this article, my choice of data series is significantly driven by my twofold goal of conducting the analysis for as many countries as possible and simultaneously using data series that are as comparable across countries as possible. An implication of this approach is that the data series that I use need not always line up exactly with the variables that the central bank in each country responds to, leading to estimated Taylor rule coefficients different from those estimated under the best possible data, as illustrated above for the United States. 102 Second Quarter 2019 Federal Reserve Bank of St. Louis REVIEW Table 2 Taylor Rule Coefficients: The Baseline Taylor Rule Taylor rule coefficients Country Australia Austria Bolivia Canada Chile Czech Republic Denmark Germany Hungary Iceland India Indonesia Israel Italy Japan Latvia Netherlands Norway Poland Portugal Slovenia South Africa Spain Switzerland United Kingdom United States Expected inflation 1.486*** (0.268) 0.135 (0.187) 0.326*** (0.111) 0.088 (0.138) –0.163** (0.083) 1.078*** (0.129) 0.195 (0.538) 0.394 (0.311) 0.531*** (0.075) 1.004*** (0.287) –0.165*** (0.027) 0.876*** (0.334) –1.543 (1.247) 1.244*** (0.214) 1.466*** (0.181) 0.0637*** (0.016) 0.735*** (0.082) –0.629 (0.471) –0.427*** (0.111) 1.223*** (0.253) 1.267*** (0.110) 0.354 (0.253) 1.672*** (0.195) 0.614*** (0.172) 0.873*** (0.337) 1.354** (0.584) Output gap 0.612 (0.430) 1.033*** (0.251) 3.354*** (0.466) 1.072*** (0.173) 0.904*** (0.105) 1.985*** (0.485) 0.677 (0.620) 0.678*** (0.159) –2.010* (1.157) 0.574* (0.333) 0.281*** (0.037) –0.735*** (0.268) –1.218 (0.956) 2.031*** (0.766) –0.257** (0.114) 0.0221 (0.039) 1.248*** (0.203) 1.927** (0.770) 6.484*** (0.526) 0.964** (0.435) 0.621 (0.491) 1.006*** (0.327) –0.205 (0.145) 0.494*** (0.170) 0.903* (0.545) 0.621*** (0.228) Lagged interest rates 0.893*** (0.016) 0.919*** (0.019) 0.901*** (0.018) 0.851*** (0.016) 0.758*** (0.036) 0.942*** (0.008) 0.976*** (0.008) 0.887*** (0.022) 0.890*** (0.010) 0.942*** (0.015) 0.823*** (0.013) 0.788*** (0.052) 0.962*** (0.020) 0.915*** (0.030) 0.919*** (0.013) 0.753*** (0.021) 0.841*** (0.024) 0.970*** (0.011) 0.855*** (0.014) 0.920*** (0.034) 0.874*** (0.018) 0.897*** (0.022) 0.822*** (0.036) 0.915*** (0.012) 0.903*** (0.022) 0.932*** (0.016) Constant 1.673 (1.204) 4.287*** (0.540) 5.628*** (0.939) 3.819*** (0.396) 5.509*** (0.424) –0.123 (0.583) 4.208** (1.769) 3.470*** (0.863) 5.877*** (0.600) 6.493*** (1.004) 6.931*** (0.157) 5.307* (3.101) 8.186*** (2.785) 2.844 (1.754) 1.619*** (0.225) 3.254*** (0.133) 4.726*** (0.232) 10.25*** (2.216) 7.930*** (0.438) –1.327 (3.639) 1.444* (0.737) 9.633*** (2.771) 1.619 (1.063) 1.860*** (0.366) 3.846*** (1.305) 1.707 (1.610) Overidentification test Trade openness X/GDP J = 11.77, χ 2(16) p-value = 0.760 J = 10.30, χ 2(16) p-value = 0.851 J = 7.63, χ 2(16) p-value = 0.959 J = 8.47, χ 2(16) p-value = 0.934 J = 6.81, χ 2(16) p-value = 0.977 J = 7.46, χ 2(16) p-value = 0.963 J = 10.68, χ 2(16) p-value = 0.829 J = 10.10, χ 2(16) p-value = 0.862 J = 7.82, χ 2(16) p-value = 0.954 J = 7.49, χ 2(16) p-value = 0.963 J = 6.21, χ 2(16) p-value = 0.986 J = 6.72, χ 2(16) p-value = 0.978 J = 7.97, χ 2(16) p-value = 0.950 J = 10.37, χ 2(16) p-value = 0.847 J = 11.43, χ 2(16) p-value = 0.782 J = 8.12, χ 2(16) p-value = 0.945 J = 6.76, χ 2(16) p-value = 0.978 J = 11.73, χ 2(16) p-value = 0.762 J = 6.73, χ 2(16) p-value = 0.978 J = 8.38, χ 2(16) p-value = 0.937 J = 6.88, χ 2(16) p-value = 0.976 J = 11.98, χ 2(16) p-value = 0.745 J = 9.30, χ 2(16) p-value = 0.901 J = 12.68, χ 2(16) p-value = 0.696 J = 7.61, χ 2(16) p-value = 0.960 J = 10.83, χ 2(16) p-value = 0.820 0.172 0.339 0.249 0.377 0.328 0.490 0.390 0.220 0.567 0.338 0.146 0.299 0.334 0.205 0.115 0.384 0.548 0.396 0.304 0.257 0.521 0.260 0.196 0.463 0.235 0.093 NOTE: *, **, and *** denote 10 percent, 5 percent, and 1 percent statistical significance, respectively. The coefficients on expected inflation, the – output gap, and lagged interest rates correspond to β, γ, and ρ1 + ρ2 , respectively, from equation (4). The constant corresponds to R – βπ *. Standard errors are in parentheses. Federal Reserve Bank of St. Louis REVIEW Second Quarter 2019 103 Leibovici Therefore, throughout the rest of the article, I interpret the Taylor rule coefficients as moments that characterize salient features of monetary policy in each country rather than as structural parameters. 5.1.2 Cross-Country Estimates. I now examine salient features of my estimated Taylor rule coefficients across countries (Table 2). First, I find that there is substantial heterogeneity across countries in the Taylor rule coefficient on expected inflation. This coefficient is estimated to be negative for three countries (Chile, India, and Poland), not statistically different from zero12 for seven countries (Austria, Canada, Denmark, Germany, Israel, Norway, and South Africa), statistically higher than zero but lower than 1 for seven countries (Bolivia, Hungary, Indonesia, Latvia, the Netherlands, Switzerland, and the United Kingdom), and higher than 1 for nine countries (Australia, the Czech Republic, Iceland, Italy, Japan, Portugal, Slovenia, Spain, and the United States).13 Second, I find that there is a similar degree of heterogeneity across countries in the Taylor rule coefficient on the output gap. This coefficient is estimated to be negative for three countries (Hungary, Indonesia, and Japan), not statistically different from zero for six countries (Australia, Denmark, Israel, Latvia, Slovenia, and Spain), statistically higher than zero but lower than 1 for eight countries (Chile, Germany, Iceland, India, Portugal, Switzerland, the United Kingdom, and the United States), and higher than 1 for nine countries (Austria, Bolivia, Canada, the Czech Republic, Italy, the Netherlands, Norway, Poland, and South Africa). Third, I find that the sum of the coefficients on lagged interest rates is statistically significant for all countries, suggesting that all countries engage in some degree of nominal interest rate smoothing. Therefore, while countries differ in the response to changes in expected inflation and the output gap, in all cases lagged interest rates are an important factor in determining current interest rates. Finally, the last column of Table 2 reports the average degree of trade openness, measured through the exports-to-GDP ratio across the 26 countries under analysis. I find that there is substantial heterogeneity in the degree of trade openness across these countries, ranging from relatively closed economies such as the United States and Japan, with exports-to-GDP ratios equal to 0.093 and 0.115, respectively, to relatively open economies such as Hungary, the Netherlands, and Slovenia, with exports-to-GDP ratios equal to 0.567, 0.548, and 0.521, respectively. 5.2 Taylor Rule Coefficients and Trade Openness I now ask, to what extent do open economies conduct monetary policy differently from closed economies? To answer this question, I examine the relationship between the cross- country Taylor rule coefficients reported in Table 2 and the countries’ degree of openness to international trade. 5.2.1 Expected Inflation. The top panel of Table 3 reports the results of regressing the country-specific Taylor rule coefficients on each country’s aggregate exports-to-GDP ratio. The first two columns of the table report the results from conducting the analysis using the expected- inflation Taylor rule coefficients for all countries regardless of the degree of statistical significance obtained in the previous section. The last two columns of this table report the 104 Second Quarter 2019 Federal Reserve Bank of St. Louis REVIEW Leibovici Table 3 Monetary Policy and Trade Openness: The Baseline Taylor Rule With Taylor rule coefficient significant at 10% All countries (1) (2) (3) (4) –0.895 (0.826) –1.182** (0.529) Dependent Variable: The Taylor Rule Coefficient on Expected Inflation X/GDP –1.284 (1.033) –1.471 (0.955) 0.184 (0.159) GDP per capita 0.434*** (0.115) Constant 0.947** (0.358) –0.799 (1.523) R2 0.049 0.082 0.046 0.372 26 26 19 19 –0.963 (2.745) –0.91 (2.943) No. of countries 1.071*** (0.333) –3.034** (1.165) Dependent Variable: The Taylor Rule Coefficient on the Output Gap X/GDP –0.806 (2.039) –0.674 (2.134) –0.13 (0.408) GDP per capita –0.0396 (0.437) Constant 1.142 (0.682) 2.374 (4.075) 1.428 (0.877) 1.796 (4.261) R2 0.005 0.009 0.006 0.006 26 26 20 20 No. of countries Dependent Variable: The Taylor Rule Coefficient on Interest Rate Lags X/GDP 0.000228 (0.0668) –0.0434 (0.0546) 0.000228 (0.0668) 0.0428*** (0.0139) GDP per capita –0.0434 (0.0546) 0.0428*** (0.0139) Constant 0.886*** (0.0220) 0.482*** (0.140) 0.886*** (0.0220) 0.482*** (0.140) R2 0.000 0.284 0.000 0.284 26 26 26 26 No. of countries NOTE: *, **, and *** denote 10 percent, 5 percent, and 1 percent statistical significance, respectively. Standard errors are in parentheses. GDP per capita denotes the natural logarithm of the GDP per capita variable. results when restricting attention to the subset of expected-inflation Taylor rule coefficients that are statistically significant at the 10 percent level.14 I find that all specifications imply that there is a negative relationship between the degree of a country’s trade openness and the Taylor rule coefficient on expected inflation: In open economies the nominal interest rate responds relatively less to changes in inflation. This relationship, however, is only statistically significant when one controls for GDP per capita while Federal Reserve Bank of St. Louis REVIEW Second Quarter 2019 105 Leibovici restricting attention to countries with significant Taylor rule coefficients. That is, among the set of countries that respond to changes in expected inflation, they seem to respond relatively less if they are open. As observed in Table 2, the Taylor rule coefficient on expected inflation is not statistically significant for all countries. This is because either their Taylor rule coefficients are very close to zero or because the coefficients are estimated with a significant amount of error. Therefore, in columns 3 and 4, I restrict attention to countries in which the Taylor rule coefficients are significant at the 10 percent level. Yet, only once I control for the countries’ level of GDP per capita does the relationship between trade openness and the Taylor rule coefficient on expect inflation become significant. This is not necessarily very surprising, given developed economies are simultaneously more likely to be open as well as to be inflation targeters in their conduct of monetary policy. To quantify the economic importance of this relationship, consider the average Taylor rule coefficient on expected inflation is equal to 0.54. Then, changing the aggregate exportsto-GDP ratio from its lowest to highest value across countries (from 0.093 to 0.567) is associated with a decrease in the value of the Taylor rule coefficient equal to 0.56. This evidence suggests that open economies assign a lower weight on expected inflation when conducting monetary policy, a relationship that is both statistically and economically significant. 5.2.2 Output Gap. The middle panel of Table 3 reports the results of regressing the country-specific output-gap Taylor rule coefficients on each country’s aggregate exports-toGDP ratio. I find that there is no systematic relationship between the degree of a country’s trade openness and the weight it assigns to changes in the output gap for the conduct of monetary policy. This finding is robust to restricting attention to Taylor rule coefficients that are statistically significant at the 10 percent level as well as to controlling for GDP per capita (see columns 1 to 4). In all cases, the relationship between trade openness and the weight on the output gap is negative, but the degree of error is substantial, making these coefficients statistically insignificant. This evidence suggests that monetary policy in open and closed economies responds similarly to changes in the output gap. 5.2.3 Lagged Interest Rates. Finally, the bottom panel of Table 3 reports the results of regressing the country-specific sum of the interest rate lags of the Taylor rule on each country’s aggregate exports-to-GDP ratio. As with the output gap, I find that there is no systematic relationship between the degree of a country’s trade openness and the extent to which that country smooths its interest rate adjustments over time. This finding is robust to restricting attention to Taylor rule coefficients that are statistically significant at the 10 percent level as well as to controlling for GDP per capita (see columns 1 to 4). Yet, what I do find is that richer economies smooth their nominal interest rate adjustments relatively more than poorer ones, suggesting that richer economies conduct their monetary policy decisions in a more predictable way. 6 TAYLOR RULE WITH REAL EXCHANGE RATE The evidence presented in the previous section suggests that monetary policy in open economies is conducted differently from monetary policy in closed economies, but only along 106 Second Quarter 2019 Federal Reserve Bank of St. Louis REVIEW Leibovici certain dimensions. In particular, while open economies typically assign a lower weight to changes in expected inflation, they respond to changes in the output gap to a similar extent as closed economies do. I now examine whether, in addition to inflation, open economies also respond differently to changes in other variables not included in the baseline specification of the Taylor rule. Following previous studies from the literature, in this article I restrict attention to the real exchange rate as an additional variable that central banks in open economies may respond to when conducting monetary policy. 6.1 Taylor Rule Coefficients As in the previous section, I begin by estimating a Taylor rule for each of the countries in the sample. In this section, however, I extend the Taylor rule by including the real exchange rate as an additional variable that nominal interest rates may respond to.15 In particular, I compute the real exchange rate variable that nominal interest rates respond to as the log difference of the real exchange rate from its country-specific quadratic trend.16 Table 4 presents the estimated Taylor rule coefficients corresponding to this extended specification of the Taylor rule. I find that there is significant heterogeneity in the estimated Taylor rule coefficients on the real exchange rate. This coefficient is estimated to be negative for eight countries (the Czech Republic, Germany, Hungary, Indonesia, Japan, the Netherlands, Poland, and Switzerland), not statistically different from zero for 13 countries (Australia, Austria, Chile, Denmark, Iceland, India, Israel, Italy, Norway, Slovenia, South Africa, the United Kingdom, and the United States), and statistically higher than zero for five countries (Bolivia, Canada, Latvia, Portugal, and Spain).17 Moreover, I also find that including the real exchange rate as an additional variable in the Taylor rule significantly affects the estimated coefficients on expected inflation and the output gap, as can be readily observed by comparing Tables 2 and 4. These findings suggest that the real exchange rate is a variable that many central banks across the world might respond to when executing monetary policy decisions, even if they sometimes respond to it to different extents and in qualitatively different ways. 6.2 Taylor Rule Coefficients and Trade Openness I now ask, to what extent do open economies conduct monetary policy differently from closed economies once the real exchange is included as an additional variable in the Taylor rule? To answer this question, I examine the relationship between the cross-country Taylor rule coefficients reported in Table 4 and the countries’ degree of openness to international trade. 6.2.1 Expected Inflation. The top panel of Table 5 reports the results of regressing the country-specific expected-inflation Taylor rule coefficients on each country’s aggregate exports-to-GDP ratio. I find that, while the relationship between trade openness and the weight on expected inflation is negative, as in the previous section, this relationship is statistically insignificant in all the specifications considered. In particular, note that I no longer find that open economies assign a lower weight on expected inflation once I restrict attention Federal Reserve Bank of St. Louis REVIEW Second Quarter 2019 107 Table 4 Taylor Rule Coefficients: The Baseline Taylor Rule Taylor rule coefficients Country Australia Austria Bolivia Canada Chile Czech Republic Denmark Germany Hungary Iceland India Indonesia Israel Italy Japan Latvia Netherlands Norway Poland Portugal Slovenia South Africa Spain Switzerland United Kingdom United States Expected inflation 1.569*** (0.286) –0.0278 (0.342) 0.208* (0.121) 0.0427 (0.132) –0.178* (0.095) 0.960*** (0.153) 0.0876 (0.586) 0.176 (0.373) 0.635*** (0.116) 1.294** (0.503) –0.184*** (0.039) 0.266*** (0.092) –1.245 (1.287) 1.336*** (0.239) 1.309*** (0.194) 0.0686*** (0.019) 0.554*** (0.082) –0.651 (0.516) –1.803*** (0.407) 0.985*** (0.136) 1.267*** (0.111) 0.239 (0.294) 1.528*** (0.238) 0.925*** (0.203) 0.846** (0.343) 1.701 (1.035) Output gap 0.614 (0.460) 1.020*** (0.283) 2.408*** (0.498) 1.260*** (0.211) 1.404*** (0.250) 1.691** (0.843) 0.755 (0.755) 0.810*** (0.262) 0.364 (1.616) 0.481 (0.383) 0.256*** (0.060) 1.101** (0.441) –0.690 (1.241) 2.717** (1.222) –0.220** (0.110) 0.0801 (0.058) 1.221*** (0.183) 1.939** (0.841) 10.91*** (1.449) 0.313 (0.215) 0.578 (0.619) 1.038*** (0.383) –2.405*** (0.462) 0.197 (0.148) 0.909 (0.579) 0.491 (0.359) Real exchange rate –0.0799 (0.148) –0.415 (0.337) 1.173*** (0.272) 0.186*** (0.066) –0.186 (0.132) –0.984*** (0.298) –0.705 (0.552) –0.355** (0.159) –0.654*** (0.191) 0.270 (0.212) 0.0546 (0.071) –0.874*** (0.227) –0.410 (0.610) –0.203 (0.222) –0.0939** (0.041) 0.128*** (0.022) -0.346*** (0.065) –0.178 (0.716) –0.653*** (0.177) 0.421*** (0.158) –0.00940 (0.152) –0.110 (0.153) 1.154*** (0.220) –0.489*** (0.138) 0.00500 (0.123) 0.0587 (0.059) Lagged interest rates 0.898*** (0.015) 0.932*** (0.020) 0.892*** (0.025) 0.846*** (0.021) 0.799*** (0.032) 0.951*** (0.008) 0.976*** (0.010) 0.916*** (0.029) 0.913*** (0.013) 0.953*** (0.018) 0.820*** (0.012) 0.733*** (0.032) 0.959*** (0.023) 0.927*** (0.029) 0.916*** (0.015) 0.759*** (0.020) 0.828*** (0.022) 0.972*** (0.014) 0.898*** (0.016) 0.848*** (0.055) 0.873*** (0.019) 0.904*** (0.027) 0.832*** (0.030) 0.908*** (0.018) 0.905*** (0.022) 0.931*** (0.016) Constant 1.330 (1.342) 4.537*** (1.060) 6.457*** (1.355) 3.781*** (0.327) 5.376*** (0.530) –0.349 (0.935) 4.605** (1.801) 3.787*** (0.864) 4.182*** (1.020) 5.767*** (0.212) 7.042*** (0.237) 13.38*** (1.748) 7.569** (3.116) 1.959 (2.000) 1.458*** (0.304) 3.262*** (0.148) 4.929*** (0.190) 10.30*** (2.380) 10.08*** (0.894) 2.792* (1.627) 1.454** (0.737) 10.84*** (3.143) 2.126* (1.148) 1.529*** (0.517) 3.950*** (1.326) 0.846 (2.688) Overidentification test Trade openness X/GDP J = 11.61, χ 2(15) p-value = 0.708 J = 10.65, χ 2(15) p-value = 0.777 J = 7.55, χ 2(15) p-value = 0.951 J = 8.33, χ 2(15) p-value = 0.910 J= 6.59, χ 2(15) p-value = 0.968 J = 7.81, χ 2(15) p-value = 0.931 J = 9.89, χ 2(15) p-value = 0.827 J = 9.09, χ 2(15) p-value = 0.873 J = 7.51, χ 2(15) p-value = 0.942 J = 7.00, χ 2(15) p-value = 0.958 J = 5.95, χ 2(15) p-value = 0.981 J = 8.83, χ 2(15) p-value = 0.886 J = 7.94, χ 2(15) p-value = 0.926 J = 10.10, χ 2(15) p-value = 0.814 J = 11.15, χ 2(15) p-value = 0.742 J = 7.38, χ 2(15) p-value = 0.946 J = 6.68, χ 2(15) p-value = 0.966 J = 11.69, χ 2(15) p-value = 0.703 J = 6.39, χ 2(15) p-value = 0.972 J = 7.30, χ 2(15) p-value = 0.949 J = 6.86, χ 2(15) p-value = 0.962 J = 11.22, χ 2(15) p-value = 0.74 J = 8.53, χ 2(15) p-value = 0.901 J = 9.53, χ 2(15) p-value = 0.848 J = 7.68, χ 2(15) p-value = 0.936 J = 10.73, χ 2(15) p-value = 0.772 0.172 0.339 0.249 0.377 0.328 0.490 0.390 0.220 0.567 0.338 0.146 0.299 0.334 0.205 0.115 0.384 0.548 0.396 0.304 0.257 0.521 0.260 0.196 0.463 0.235 0.093 NOTE: *, **, and *** denote 10 percent, 5 percent, and 1 percent statistical significance, respectively. The coefficients on expected inflation, the – output gap, and lagged interest rates correspond to β, γ, and ρ1 + ρ2 , respectively, from equation (4). The constant corresponds to R – βπ *. Standard errors are in parentheses. 108 Second Quarter 2019 Federal Reserve Bank of St. Louis REVIEW Table 5 Monetary Policy and Trade Openness: The Taylor Rule with the Real Exchange Rate With Taylor rule coefficient significant at 10% All countries (1) (2) (3) (4) –0.332 (0.992) –1.123 (0.642) Dependent Variable: The Taylor Rule Coefficient on Expected Inflation X/GDP –1.200 (1.163) –1.541 (1.042) 0.335** (0.140) GDP per capita 0.574*** (0.0940) Constant 0.837* (0.418) –2.331* (1.280) 0.751 (0.442) R2 0.033 0.118 0.003 0.302 26 26 18 18 4.148 (3.288) 4.731 (3.831) No. of countries –4.509*** (0.906) Dependent Variable: The Taylor Rule Coefficient on the Output Gap X/GDP 0.919 (1.611) 1.269 (1.655) –0.343 (0.303) GDP per capita –0.234 (0.371) Constant 0.834 (0.790) 4.080 (3.256) 0.440 (1.291) 2.505 (3.072) R2 0.003 0.016 0.031 0.036 26 26 15 15 No. of countries Dependent Variable: The Taylor Rule Coefficient on the Real Exchange Rate X/GDP –1.630*** (0.579) –1.548** (0.558) –2.480** (1.113) –0.0808 (0.175) GDP per capita –2.348* (1.083) –0.122 (0.356) Constant 0.389 (0.229) 1.154 (1.765) 0.746 (0.512) 1.877 (3.494) R2 0.171 0.185 0.251 0.268 26 26 13 13 No. of countries Dependent Variable: The Taylor Rule Coefficient on Interest Rate Lags X/GDP 0.0109 (0.0732) –0.0396 (0.0648) 0.0109 (0.0732) 0.0496*** (0.0163) GDP per capita –0.0396 (0.0648) 0.0496*** (0.0163) Constant 0.885*** (0.0241) 0.415** (0.164) 0.885*** (0.0241) 0.415** (0.164) R2 0.001 0.348 0.001 0.348 26 26 26 26 No. of countries NOTE: *, **, and *** denote 10 percent, 5 percent, and 1 percent statistical significance, respectively. Standard errors are in parentheses. GDP per capita denotes the natural logarithm of the GDP per capita variable. Federal Reserve Bank of St. Louis REVIEW Second Quarter 2019 109 Leibovici to countries with Taylor rule coefficients that are statistically significant at the 10 percent level and control for GDP per capita. This evidence shows that including the real exchange rate as part of the Taylor rule fundamentally affects the differential response to changes in expected inflation between open and closed economies estimated in the previous section. This evidence also suggests that differences in the responses to the real exchange rate may be a fundamental dimension along which these economies differ in the conduct of monetary policy. 6.2.2 Output Gap. The second panel of Table 5 reports the results of regressing the country-specific output gap Taylor rule coefficients on each country’s aggregate exports-toGDP ratio. In contrast to the results presented in the previous section, I now find that open economies are estimated to assign a relatively higher weight on changes in the output gap; however, these estimates are statistically insignificant at the 10 percent level in all the specifications considered. Thus, I conclude that open economies do not systematically respond differently from closed economies to changes in the output gap. 6.2.3 Real Exchange Rate. The third panel of Table 5 reports the results of regressing the country-specific real exchange rate Taylor rule coefficients on each country’s aggregate exports-to-GDP ratio. First, I find that open economies assign a lower (or more negative) weight on changes in the real exchange rate than closed economies do. And, moreover, I find that this relationship is statistically significant in all the specifications considered. To quantify the economic importance of this relationship, I restrict attention to the results reported in column 4, where I control for GDP per capita and consider only Taylor rule coefficients that are statistically significant at the 10 percent level. On the one hand, consider that the average Taylor rule coefficient on the real exchange rate in this specification is equal to –0.11. On the other hand, note that changing the aggregate exports-to-GDP ratio from its lowest to highest value across countries (from 0.093 to 0.567) is associated with a decrease in the value of the Taylor rule coefficient equal to –1.11. This decrease suggests that open economies respond to a significantly larger extent to deviations of the real exchange rate from its trend than closed economies do. In particular, open economies are more likely to decrease the nominal interest rate when the real exchange rate is relatively appreciated (that is, when the real exchange rate is above trend). 6.2.4 Lagged Interest Rates. Finally, the bottom panel of Table (5) reports the results of regressing the country-specific sum of the interest rate lags of the Taylor rule on each country’s aggregate exports-to-GDP ratio. As in the previous section, I find that there is no systematic relationship between the degree of a country’s trade openness and the extent to which that country smooths interest rate adjustments over time in any of the specifications considered. Yet, as in the previous section, I find that richer economies smooth their nominal interest rate adjustments relatively more than poorer ones. 7 CONCLUSION In this article, I study the extent to which open economies conduct monetary policy differently from closed economies. To do so, I apply the estimation approach of Clarida, Galí, and Gertler (1998 and 2000) to estimate country-specific Taylor rules for 26 economies and 110 Second Quarter 2019 Federal Reserve Bank of St. Louis REVIEW Leibovici then examine whether open economies assign systematically different weights to changes in inflation and the output gap than their closed-economy counterparts do. I find that, indeed, open economies respond less strongly to changes in expected inflation than closed economies do; in contrast, I find that the response to changes in the output gap is independent of the degree of trade openness. Moreover, I find that this difference between closed and open economies may be accounted for by the higher response of open economies to changes in the real exchange rate. Recomputing the analysis by extending the Taylor rule to allow nominal interest rates to respond to movements in the real exchange rate, I find that open economies no longer assign a systematically lower weight to inflation as closed economies do. Instead, I find that open economies respond more strongly to deviations of the real exchange rate from its trend. It is important to remark that the analysis conducted in this article is subject to several caveats. An important one is that I restrict attention to differences in monetary policy as measured from the lens of the Taylor rule. To the extent that central banks may conduct monetary policy using instruments and targets that have no impact on the joint time-series dynamics of nominal interest rates, inflation, and the output gap, such policies would not be captured by my approach. One question raised by these findings concerns the optimality of these differences in policymaking. To what extent should open economies indeed conduct monetary policy differently from closed economies along the dimensions documented in this article? And, if the observed differences in monetary policy are indeed suboptimal, then to what extent can open economies achieve better economic outcomes by conducting monetary policy in an optimal fashion? These are important questions left for further research. n Federal Reserve Bank of St. Louis REVIEW Second Quarter 2019 111 Leibovici NOTES 1 Practically no country in the world can be accurately described to be a fully closed economy, since most countries trade with other countries to some extent. Nevertheless, throughout the rest of the paper, I sometimes loosely refer to economies that trade relatively less than others as “closed economies.” 2 A symmetric J × J matrix W is positive definite if the scalar xTWx is positive for every non-zero vector x, where xT is the transpose of x. 3 All variables used are based on the latest revised data; that is, the analysis is not conducted using real-time data and is subject to the limitations discussed in Orphanides (2001). Similarly, all filters (e.g., seasonal adjustment, detrending) applied to the data are based on the full sample and are not estimated on a real-time basis. 4 Thus, the variables that I use need not correspond exactly to the variables targeted by each of the country-specific central banks, even for countries in which monetary policy may be characterized as following a Taylor rule. 5 For more information, see IMF (n.d.). 6 The GDP deflator is typically not the main variable used by central banks across the world, such as the Federal Reserve or European Central Bank, to measure inflation. Yet, in this article I restrict attention to measuring inflation based on the GDP deflator to maximize the cross-country comparability of the inflation measure. 7 Results are robust to other detrending procedures, such as computing the output gap as the cyclical deviation of real GDP from a Hodrick-Prescott trend with a smoothing parameter of 1,600. 8 Downloaded directly from the BIS website: https://www.bis.org/statistics/eer. 9 While there are data available extending beyond 2006, I restrict attention to the pre-Great-Recession period to abstract from the measurement and modeling issues that would be introduced by having to deal with monetary policy at the zero lower bound. 10 For each country, I select the real exchange rate series used in the analysis as follows: (i) I use IMF data if there are at least 36 consecutive quarters available; (ii) otherwise, I use the BIS narrow real exchange rate data if there are at least 36 consecutive quarters available, (iii) otherwise, I use the OECD data if there are at least 36 consecutive quarters available; (iv) otherwise, I use the BIS broad real exchange rate data if there are at least 36 consecutive quarters available, and (v) otherwise, I exclude the country from the analysis. 11 These values correspond to the estimates from the second row of Table III in Clarida, Galí, and Gertler (2000). 12 Throughout the rest of the paper I refer to “not statistically different from zero” if a coefficient is not statistically significant at the 10 percent level. 13 I interpret the estimated Taylor rule coefficients as empirical moments informative about the comovement among nominal interest rates with expected inflation, the output gap, and lagged nominal rates. Thus, for instance, I interpret negative and insignificant values of the Taylor rule coefficients as informative about the way in which a country conducts monetary policy, regardless of whether the country follows a Taylor rule or some other monetary policy regime. 14 Note that the Taylor rule coefficients estimated in the previous section are estimated with uncertainty. I abstract from this source of uncertainty when computing the standard errors reported in Table 3. 15 I measure the real exchange rate using the variables described in Section 4.2. 16 I also include four lags of this variable as additional instruments. 17 Note that a negative coefficient on the real exchange rate implies that a depreciated real exchange rate (a low value of the real exchange rate relative to trend) is associated with a higher nominal interest rate; similarly, a positive coefficient implies that a depreciated real exchange rate is associated with a lower nominal interest rate. 112 Second Quarter 2019 Federal Reserve Bank of St. Louis REVIEW Leibovici REFERENCES Basilio, J.R. “Empirics of Monetary Policy Rules: The Taylor Rule in Different Countries.” Ph.D. thesis, University of Illinois at Chicago, 2013. Berument, H.; Konac, N. and Senay, O. “Openness and the Effectiveness of Monetary Policy: A Cross-Country Analysis.” International Economic Journal, 2007, 21(4), pp. 577-91; https://doi.org/10.1080/10168730701699018. Clarida, R.; Galí, J. and Gertler, M. “Monetary Policy Rules in Practice Some International Evidence.” European Economic Review, 1998, 42(6), pp. 1033-67; https://doi.org/10.1016/S0014-2921(98)00016-6. Clarida, R.; Galí J. and Gertler, M. “Monetary Policy Rules and Macroeconomic Stability: Evidence and Some Theory.” Quarterly Journal of Economics, 2000, 115(1), pp. 147-80; https://doi.org/10.1162/003355300554692. Clarida, R.; Galí J. and Gertler, M. “A Simple Framework for International Monetary Policy Analysis.” Journal of Monetary Economics, 2002, 49(5), pp. 879-904; https://doi.org/10.1016/S0304-3932(02)00128-9. Corsetti, G.; Dedola L. and Leduc, S. “Optimal Monetary Policy in Open Economies,” in B.M. Friedman and M. Woodford, eds., Handbook of Monetary Economics. Volume 3. North-Holland, 2010, pp. 861-933; https://doi.org/10.1016/B978-0-444-53454-5.00004-9. De Paoli, B. “Monetary Policy and Welfare in a Small Open Economy.” Journal of International Economics, 2009, 77(1), pp. 11-22; https://doi.org/10.1016/j.jinteco.2008.09.007. Faia, E. and Monacelli, T. “Optimal Monetary Policy in a Small Open Economy with Home Bias.” Journal of Money, Credit, and Banking, 2008, 40(4), pp. 721-50; https://doi.org/10.1111/j.1538-4616.2008.00133.x. Hayo, B. and Hofmann, B. “Comparing Monetary Policy Reaction Functions: ECB versus Bundesbank.” Empirical Economics, 2006, 31(3), pp. 645-62; https://doi.org/10.1007/s00181-005-0040-7. International Monetary Fund. “What Is the Central Bank Policy Rate?” N.d.; http://datahelp.imf.org/knowledgebase/articles/484375-what-is-the-central-bank-policy-rate, accessed December 10, 2018. Kahn, G.A. “Estimated Rules for Monetary Policy.” Federal Reserve Bank of Kansas City Economic Review, 2012, p. 5; https://www.kansascityfed.org/publicat/econrev/pdf/12q4Kahn.pdf. Leibovici, F. and Santacreu, A.M. “International Trade Fluctuations and Monetary Policy.” Working paper, 2015. Lombardo, G. and Ravenna, F. “Openness and Optimal Monetary Policy.” Journal of International Economics, 2014, 93(1), pp. 153-72; https://doi.org/10.1016/j.jinteco.2014.01.011. Lubik, T.A. and Schorfheide, F. “Do Central Banks Respond to Exchange Rate Movements? A Structural Investigation.” Journal of Monetary Economics, 2007, 54(4), pp. 1069-87; https://doi.org/10.1016/j.jmoneco.2006.01.009. Orphanides, A. “Monetary Policy Rules Based on Real-Time Data.” American Economic Review, 2001, 91(4), pp. 964-85; https://doi.org/10.1257/aer.91.4.964. Taylor, J.B. “Discretion versus Policy Rules in Practice.” Carnegie-Rochester Conference Series on Public Policy, 1993, 39, pp. 195-214; https://doi.org/10.1016/0167-2231(93)90009-L. Torres, A. “Monetary Policy and Interest Rates: Evidence from Mexico.” North American Journal of Economics and Finance, 2003, 14(3), pp. 357-79; https://doi.org/10.1016/j.najef.2003.08.001. Yazgan, M. Ege and Yilmazkuday, H. “Monetary Policy Rules in Practice: Evidence from Turkey and Israel.” Applied Financial Economics, 2007, 17(1), pp. 1-8; https://doi.org/10.1080/09603100600606206. Federal Reserve Bank of St. Louis REVIEW Second Quarter 2019 113 114 Second Quarter 2019 Federal Reserve Bank of St. Louis REVIEW The Real Term Premium in a Stationary Economy with Segmented Asset Markets YiLi Chien and Junsang Lee This article proposes a general equilibrium model to explain the positive and sizable term premia implied by the data. The authors introduce a slow mean-reverting process of consumption growth and a segmented asset-market mechanism with heterogeneous trading technologies into an otherwise standard heterogeneous agent general equilibrium model. First, the slow mean-reverting consumption growth process implies that the expected consumption growth rate is only slightly countercyclical and the process can exhibit near-zero first-order autocorrelation, as observed in the data. This slight countercyclicality suggests that long-term bonds are risky, and hence the term premia should be positive. Second, the segmented asset-market mechanism amplifies the magnitude of the term premia because aggregate risk is highly concentrated in a small fraction of marginal traders who demand high compensation for taking risk. For sensitivity analysis, the role of each assumption is further investigated by removing each factor one at a time. (JEL G11, G12, E30) Federal Reserve Bank of St. Louis Review, Second Quarter 2019, 101(2), pp. 115-34. https://doi.org/10.20955/r.101.115-34 1 INTRODUCTION The positive and sizable term premia observed in the data have been hard to reconcile using a standard structural macroeconomic model. Backus, Gregory, and Zin (1989) demonstrate the failure of a standard model in accounting for the sign and the magnitude of real bond risk premia. Campbell (1986), Donaldson, Johnsen, and Mehra (1990), and den Haan (1995) also experience the same difficulty with standard macroeconomic models.1 Although equilibrium models are difficult to work with and have limited success, it is still important to try to understand the fundamental mechanisms behind positive and sizable term premia. For macroeconomists, the disconnect between the observed term premia in the data and what a standard structural macroeconomic model predicts is often referred to as the “term premium puzzle.” The issue is also important to central bankers. As pointed out by YiLi Chien is an economist and research officer at the Federal Reserve Bank of St. Louis. Jungsang Lee, the corresponding author, is an associate professor at Sungkyunkwan University, Seoul, Korea. © 2019, Federal Reserve Bank of St. Louis. The views expressed in this article are those of the author(s) and do not necessarily reflect the views of the Federal Reserve System, the Board of Governors, or the regional Federal Reserve Banks. Articles may be reprinted, reproduced, published, distributed, displayed, and transmitted in their entirety if copyright notice, author name(s), and full citation are included. Abstracts, synopses, and other derivative works may be made only with prior written permission of the Federal Reserve Bank of St. Louis. Federal Reserve Bank of St. Louis REVIEW Second Quarter 2019 115 Chien and Lee Wright (2011), the term premia represent the relationship between the short rate, which is controlled by central banks, and the long rate, which relates more deeply to real economic activities. Hence, understanding the term premia helps central bankers evaluate the effectiveness of monetary policy and the mechanism behind its effects on the real economy. Finally, for investors, it is of utmost importance to understand term premia—to hedge against interest rate risk. A standard macroeconomic model with the pricing kernel or the stochastic discount factor derived from a utility maximization problem generally has great difficulty matching the slope and the level of the term structure. Campbell (1986) shows that the term premium depends on the nature of the consumption growth process. If the consumption growth process is positively autocorrelated, then the expected future growth rate falls and bond prices rise in a recession. The long-term bond then becomes a good hedge, and hence the term premium is negative. On the other hand, if the consumption growth rate is negatively autocorrelated, then the model predicts a positive term premium since the long-term bond becomes risky because of its procyclical pricing. This intuition together with near-zero autocorrelation of consumption growth, that is, a random walk in empirical studies, implies that the term premium should be close to zero when the pricing kernel is derived from a standard macroeconomic model. In addition, it is also well known that the pricing kernel of a standard model, which relies purely on the expected aggregate consumption growth rate, is not volatile enough to deliver a high market price of risk. Therefore, the standard model not only fails to match the sign of the term premium but also fails to generate the correct magnitude of the term premium. In this article, we assume that the aggregate consumption process is trend stationary with a long memory process, which shows near zero but slightly negative autocorrelation of the consumption process. This consumption process alone generates positive term premia but with very small magnitude. This process is not easily statistically distinguished from a difference-stationary process such as the random walk. This view is supported by Christiano and Eichenbaum (1990,) who argue that no clear statistical evidence exits to support either a trend-stationary or a difference-stationary process of aggregate consumption. More specifically, we consider a slow trend-reverting aggregate consumption process in our model economy and hence the level of consumption can be well above or below its long-run trend for an extended period. With this process, when a bad shock is realized, the expected growth rate of consumption is only slightly higher because of its slow mean-reverting property. Therefore, the expected growth rate of consumption is only slightly countercyclical and the autocorrelation of consumption growth between two consecutive periods could be very close to zero but slightly negative (only –0.02 in our calibrated model), which is consistent with the randomwalk-like consumption process in the data. Following Chien, Cole, and Lustig (2011), there is a segmented asset-market mechanism in our model. Specifically, the model features a large fraction of households who do not participate in the equity market and hence do not bear any aggregate risk. There is, however, a small fraction of households who do participate in the equity market and hence bear a great amount of aggregate risk, which in this article results in a high market price of risk. In equi116 Second Quarter 2019 Federal Reserve Bank of St. Louis REVIEW Chien and Lee librium, those households demand high risk compensation. As a result, high risk premia are obtained not only in equities but also reflected onto long-term bonds. Therefore, the segmented market mechanism amplifies the size and the magnitude of the term premia. Our calibrated model considers not only the segmented asset-market mechanism but also the asymmetric bond positions of U.S. households. The data on U.S. households show that a large fraction of households carry long-term mortgage loans but save in short-term risk-free assets, such as checking or savings accounts. In other words, these households essentially borrow in long-term bonds by using housing as a collateral and save in short-term bonds. In our calibrated model, we also evaluate the extent to which this asymmetric bond position of households matters for term premia quantitatively. The assumptions in our model are built with solid support from empirical evidence. The first assumption, of a mean-reverting consumption growth process, is prevalent in the macro economics literature. The growth of aggregate variables, such as output or consumption, is often decomposed into trend components and cyclical components (business cycles). Such a decomposition is consistent with the mean-reverting assumption. The second assumption, of a segmented market mechanism, is firmly grounded in empirical evidence from the household finance literature. The evidence shows that most households do not purchase most of the assets available to them (Guiso and Sodini, 2012). In fact, the composition of household asset holdings varies greatly across households, even in a developed country such as the United States. Only 50 percent of U.S. households participate in the equity market, according to the 2010 Survey of Consumer Finance (SCF hereafter) data. Moreover, even among the participants in the equity market, many investors still hold low-risk portfolios and do not adjust their portfolios frequently.2 On the other hand, a small fraction of households actively adjust their portfolios and earn a higher return by taking more aggregate risk. The SCF data also show that a large fraction of households carry mortgage loans and save in short-term safe assets. These households effectively have a long position in short-term bonds and a short position in long-term bonds. As the data also show, wealthier households, a relatively small fraction of all households, tend to hold a higher fraction of long-term bonds in their portfolios. Only a handful of structural models in the literature are able to deliver an average upward- sloping nominal and/or real yield curve. Many of them modify household preferences into various forms in the standard macroeconomic model. Piazzesi and Schneider (2007) demonstrate that the nominal yield curve can be upward sloping even with a flat or downward-sloping real yield curve since a low-frequency negative correlation between consumption growth and the inflation rate causes inflation risk. They assume a recursive preference, and hence agents are very willing to substitute consumption over time even though they are risk averse. The recursive preference plays a critical role in the low-frequency correlation mattering for the current price. Bansal and Shaliastovich (2013) also generate a positive nominal term premium with inflation risks and recursive preferences. Rudebusch and Swanson (2012) further extend the endowment economy model to a production economy general equilibrium model. By introducing inflation ambiguity into a representative agent model, Ulrich (2013) explains the upward-sloping nominal yield curve with a log utility function. Our work is complementary to the existing papers discussed above since we focus on the real term premia rather than Federal Reserve Bank of St. Louis REVIEW Second Quarter 2019 117 Chien and Lee the nominal premia. Wachter (2006) uses a habit-persistence model to explain both the positive real and nominal bond premia. To maintain the consumption level, investors tend to sell long-term bonds during recessions and vice versa during expansions. Namely, the demand for long-term bonds is procyclical, which makes the bond price procyclical and hence the longterm bond itself a risky asset. Rudebusch and Swanson (2012) find that the habit-formation mechanism in Wachter (2006) fails to generate a sizable term premia without distorting the behavior of other macroeconomics variables. Our benchmark model generates a high and volatile equity premium with a 7.26 percent mean and a 15.63 percent standard deviation, as well as a low and stable risk-free return with a 0.95 mean and a 1.45 percent standard deviation—estimates quite close to those in the asset- pricing literature. Most importantly, our quantitative result also predicts a high real term premium: 1.92 percent for 30-year zero-coupon bonds. This article delivers a reasonable term premium result, with a risk aversion coefficient of 4. For the sensitivity analysis, we further investigate the role of our assumptions by removing each factor one by one. Our main contribution to the literature is to provide a simple and intuitive story that can reconcile the puzzling disconnect between asset prices, equity and term premia in particular, and aggregate macroeconomic variables. The model in this article integrates the empirical facts of heterogeneous portfolios across households, as found in the household finance literature, and a mean-reverting aggregate consumption process, as found in the macroeconomics literature, to explain the real term-premia puzzle. Our model successfully delivers a positive sign for and significant magnitude of the real term premia. Specifically, we demonstrate the importance of the household portfolio heterogeneity documented in the macro-finance literature, while the majority of asset-pricing models rely on a representative agent framework with modifications to preferences. 2 THE MODEL We consider an endowment economy in which households sequentially trade assets and consume. Two features distinguish our model from the standard model. First, our endowment (consumption) growth follows a slow mean-reverting process. After the realization of a bad endowment shock, the expected consumption growth rate edges up only slightly because of the trend-reverting property, which makes the autocorrelation of the consumption growth process slightly negative but very close to zero. Hence, our shock process is consistent with the empirical fact that consumption growth is well approximated by the random walk. The second key feature of our model is that it exhibits ex-ante heterogeneity in the trading technologies. The trading technologies are modeled on the menu of assets, specifically by exogenously restricting the portfolios a household can trade and hold. The goal of these restrictions is to capture the observed portfolio behavior of most households. In our calibrated model, this form of ex-ante heterogeneity delivers a high market price of risk and hence helps to deliver sizable term premia. 118 Second Quarter 2019 Federal Reserve Bank of St. Louis REVIEW Chien and Lee 2.1 The Environment There is a unit measure of households subject to both aggregate endowment growth risks and idiosyncratic income shocks. Households are ex-ante identical except for the trading technologies that they are endowed with. Ex post, these households differ in terms of their idiosyncratic income-shock realizations. All households face the same stochastic process for idiosyncratic income shocks, and all households start with the same present value of wealth. In the model, time is discrete, infinite, and indexed by t = 0,1,2,… The first period, t = 0, is a planning period in which all trading takes place. We assume constant average growth of the endowment process and a transitory shock that makes the actual level of aggregate consumption deviate from its long-term trend. More specifically, let mt be the percentage deviation of aggregate endowment from the growth trend. Then, the total endowment in period t, denoted by Yt , is lnYt = tln g + mt , where g– is the average growth rate of the endowment. The output growth process is therefore affected by the evolution of mt, which is assumed to follow an AR(1) process: ( ) mt+1 = ρmt + ε t+1 , ε t+1 ∼ N 0,σ ε2 . With this specification of the endowment shock process, the growth rate of output, denoted Y by g t+1 ≡ t+1 , is therefore given by Yt (1) ln Yt+1 ≡ ln g t+1 = ln g + ( ρ −1)mt + ε t+1 . Yt If ρ is 1, then the endowment process follows a random walk with drift, a difference-stationary process. If ρ is less than 1 but close to 1, then the endowment slowly reverts to trend, a trend- stationary process (a long-memory property). As mentioned in the Introduction, there is no clear evidence in favor of either a trend-stationary or difference-stationary process for macro economic variables, such as consumption or output. Our model follows the view of a trend- stationary endowment process with long memory. Hence, the value of ρ is set to 0.95 in the calibration. Let zt denote the history of aggregate states up to period t, and hence let Yt(zt ) denote the aggregate endowment is period t. In addition, aggregate endowment each period is divided into two parts: diversifiable income and nondiversifiable income. Claims to diversifiable income can be traded in financial markets, while claims to nondiversifiable income cannot. We assume a constant share of nondiversifiable income, denoted by γ (0,1). The nondiversifiable component is subject to idiosyncratic stochastic shocks, denoted by ηt . Nondiversifi able household income is denoted by γYt(zt )ηt . Similarly, let ηt denote the history of idiosyncratic shocks up to period t. In addition, we use π(zt ,ηt ) to denote the unconditional probability of state (zt ,ηt ) being realized. The idioFederal Reserve Bank of St. Louis REVIEW Second Quarter 2019 119 Chien and Lee syncratic shock events are governed by a first-order Markov process, and their probabilities are assumed independent between z shocks and η shocks: ( ) π z t+1 ,η t+1 z t ,η t = π ( zt+1 z t )π (ηt+1 η t ) . Since we can appeal to the law of large numbers, π(ηt ) also denotes the fraction of agents in state zt that have drawn the history ηt . We introduce some additional notation: z t+1 zt or η t+1 ηt means that the left-hand-side node is a successor node to the right-hand-side node. We denote by {zτ zt } the set of successor aggregate histories for zt , including those many periods in the future; ditto for {ητ η t }. When we use , we include the current nodes zt or η t in the set. All households live for infinite periods and rank a stream of consumption according to the following criterion: (2) U ({c}) = 1 ∞ ∑ β t 1− α ct ( z t ,η t ) π ( z t , η t ) , ( t ≥1, z t , η t 1−α ) where α denotes the coefficient of relative risk aversion, β denotes the time discount factor, and ct(zt ,η t ) denotes household consumption in state (zt ,η t ). In this economy, there are four type of assets available: state-contingent claims on aggregate shocks, a long-term bond (consol) with a constant stream of payments, risky equities, and one-period risk-free bonds. Note that the market is incomplete in our environment since there is no state-contingent claims available for idiosyncratic shocks. Equity is assumed to be a leveraged aggregate output process, with dividend growth determined by ΔlnDt+1 = Et ( ΔlnYt+1 ) + φ ⎡⎣ ΔlnYt+1 − Et ( ΔlnYt+1 ) ⎤⎦ , where ϕ is the leverage ratio, which is assumed constant over time. Finally, we denote the value of total equity by Vt (zt ). The gross returns of leveraged equity, or Ret,t–1(zt ), are given by (3) e Rt,t−1 ( z ) = (V )( z () ) . t Dt z t +Vt z t t−1 t−1 2.2 Heterogeneity in Trading Technologies To match the size of the term premium, we introduce the segmented market mechanism, in particular portfolio heterogeneity at the household level. As mentioned in the Introduction, heterogeneity in portfolio choices is widely supported by the data. As we demonstrate later, the concentration of a large portion of aggregate risk in a relatively small fraction of households amplifies the price of risk in the calibrated model. Without such a mechanism, the model fails to match the size of the term premium quantitatively. To capture such portfolio heterogeneity, we adopt the approach by Chien, Cole, and Lustig (2011), which exogenously imposes different restrictions on investors’ portfolio choices. These restrictions apply to the menu of assets that households can trade as well as to the composition of households’ portfolios. 120 Second Quarter 2019 Federal Reserve Bank of St. Louis REVIEW Chien and Lee There are two classes of investors in terms of their asset-trading technologies: Mertonian traders and non-Mertonian traders. Mertonian traders face no restrictions on their portfolio choices and hence a menu of tradable assets. Specifically, they are capable of trading a complete set of contingent claims on the aggregate endowment. They optimally adjust their portfolio choices in response to changes in the set of investment opportunities. Therefore, they act as market arbitrageurs and price aggregate risk in our model. Non-Mertonian traders face restrictions on their portfolio choices. Specifically, their portfolio composition is restricted to be constant over time. There are two types of non- Mertonian traders: The first type, non-Mertonian equity investors, can trade equities, riskfree bonds, and long-term bonds but not state-contingent claims on aggregate shocks. The second type, nonparticipants, do not hold equity but invest only in risk-free bonds and console bonds. Even though the portfolio composition of non-Mertonian traders is exogenously given, they can still choose how much to save and consume. Non-Mertonian investors deviate from the optimal portfolio choice in two dimensions: First, they cannot change the share of equities, long-term bonds, or short-term bonds in their portfolios in response to changes in the market price of risk, which indicates missed market timing. Second, their portfolio share in each asset might deviate from the optimal share on average, implying that their average exposure to aggregate risk might not be optimal. We denote the measure of different types of households by μj , where j {me,et,np} represents Mertonian investors, non-Mertonian equity investors, and non-Mertonian nonparticipants, respectively. 2.3 The Household’s Problem 2.3.1 Budget Constraints of Mertonian Traders. Consider a Mertonian trader entering the period with net financial wealth at(zt ,η t–1) given the event history (zt ,η t–1). Note that net financial wealth is not spanned by the realization of idiosyncratic shocks, ηt , since there are no contingent claims on idiosyncratic shocks. At the end of the period, Mertonian traders buy shares of equities st(zt ,η t ), one-period risk-free bonds bt(zt ,η t ), long-term consol bonds bct (zt ,η t ), and state-contingent claims, ât(zt ,η t–1) in financial markets, and consumption ct(zt ,η t ) in the goods markets is subject to the following one-period budget constraint: (4) ( ) ( ) ( ) ( ) ( ) ( ) ( st z t ,η t Vt z t + bt z t ,η t + btc z t ,η t + ∑ Q z t+1 z t ât+1 z t+1 ,η t + ct z t ,η t ( ) ( ) z t+1 ) ≤ at z t ,η t−1 + γ Yt z t η t , for all z t ,η t , where Q(z t+1|zt ) denotes the state-contingent price of a unit contingent claim to the consumption good in aggregate state z t+1 acquired in aggregate state zt . The agent’s net financial wealth, at(zt ,η t–1), in state (zt ,η t ) is given by the payoff from the agent’s portfolio last period: (5) ( ) ( ( z )b ( z ) ( ) ( ) ) + â ( z ,η ), ( ) ( f at zt ,η t−1 = st−1 z t−1 ,η t−1 ⎡⎣ Dt z t +Vt z t ⎤⎦ + Rt,t−1 z t−1 bt−1 z t−1 ,η t−1 c +Rt,t−1 t Federal Reserve Bank of St. Louis REVIEW c t−1 t−1 ,η t−1 t t ) t−1 Second Quarter 2019 121 Chien and Lee f where Rct,t–1(zt ) and R t,t–1 (zt–1) denote the return of a long-term consol bond and a one-period risk-free bond in period t, respectively. Note that the total equity share of this economy, st(zt ,η t ), is normalized to 1. 2.3.1 Budget Constraints of Non-Mertonian Traders. The non-Mertonian traders have no access to state-contingent claims on aggregate shocks and are restricted to fixed portfolio weights among the equity, short-term risk-free bonds, and long-term consol bonds. At the end of period t, the households buy equity shares, one-period risk-free bonds, and long-term consol bonds, subject to a fixed target portfolio equity share and long-bond share, denoted –e and ω –c, respectively. As a result, in addition to equations (4) and (5), their constraints by ω also include a portfolio restriction, ) ( ) s ( z ,η )V ( z ) + b ( z ,η ) + b ( z ,η ) b ( z ,η ) = , s ( z ,η )V ( z ) + b ( z ,η ) + b ( z ,η ) t t ω c ( st z t ,η t Vt z t ωe = t t t t c t t t t t t t t t t c t t t t c t t t t t and no access to state-contingent claims ( ) ât z t ,η t−1 = 0 for all z t and η t−1 . –e – ω –c. The portfolio share of short-term bonds is therefore 1 – ω Alternatively, we can simplify the budget constraint of non-Mertonian traders as follows: ( ) ( ) ( ) ( ) ( ) p ŝt z t ,η t + c z t ,η t ≤ Rt,t−1 z t ŝt−1 z t−1 ,η t−1 + γ Yt z t ηt , p where ŝt denotes the asset holdings at the end of period t. R t,t–1 (zt ) represents the gross return on the fixed portfolio imposed on the non-Mertonian traders and is given by ( ) ( ) ( ) ( ) ( ) p f e c Rt,t−1 z t ,η t = ω e Rt,t−1 z t + ω c Rt,t−1 z t + 1− ω e − ω c Rt,t−1 zt . –e is zero. In the case of nonparticipants, ω Finally, all households are subject to nonnegative net wealth constraints, given by at(zt ,η t–1) ≥ 0 for Mertonian and ŝt(zt ,η t ) ≥ 0 for non-Mertonian traders. The details of the household problem and its associated optimal conditions are provided in Appendix A.1. 2.4 The Competitive Equilibrium The competitive equilibrium for this economy is defined in the standard way. It consists of a consumption allocation, allocations of state-contingent claims, one-period risk-free bonds, long-term consol bonds, and equity choices as well as a list of prices such that (i) given these prices, households’ assets and consumption choices maximize the households’ expected utility subject to the budget constraints, the solvency constraints, and the constraints on their portfolio choices and (ii) all asset markets clear. 122 Second Quarter 2019 Federal Reserve Bank of St. Louis REVIEW Chien and Lee 3 QUANTITATIVE RESULTS This section performs a quantitative exercise of our model. The next two subsections explain how we calibrate idiosyncratic shocks and aggregate shocks as well as the pool of traders in the benchmark case. Next, we briefly describe the real Treasury yields observed in the data. Finally, we report our benchmark asset-pricing quantitative results, especially the size of the real term premia, in Subsection 3.4. 3.1 Calibration The calibration of aggregate shocks is critical to our results. The aggregate endowment process is assumed to have a constant growth trend and an innovation term that makes the realization of output deviate from its trend: lnYt = t ln g + mt , where g– is the average growth rate of the endowment and the deviation from trend is captured by a variable m, which is assumed to follow the AR(1) process ( ) mt+1 = ρmt + ε t+1 , ε t+1 ∼ N 0,σ ε2 . We use a two-state Markov process to approximate the independent and identically distributed innovation εt . More specifically, since expansions occur more often than recessions, the probability of a good innovation shock is set to 27.4 percent, as in Alvarez and Jermann (2001). However, the expected endowment growth rate in each period depends on how far the current consumption level has deviated from its trend, which depends on the whole past history of innovation shocks. In computation, we therefore have to keep track of one extra state variable, m, in order to compute the conditional expected growth rate. Our model operates at an annual frequency. The average aggregate consumption growth rate g– is set to 1.8 percent with a standard deviation of 3.15 percent. Given that the aggregate consumption growth data are well approximated by the random walk in the short run, the persistency of m has to be high. We set ρ = 0.95, which makes the consumption growth autocorrelation of –0.02 sufficiently close to zero. We also consider a two-state first-order Markov chain for idiosyncratic shocks. The first state is low and the second state is high. Following Alvarez and Jermann (2001) and Storesletten, Telmer, and Yaron (2004), we calibrate the shock process by two moments: the standard deviation of idiosyncratic shocks and the first-order autocorrelation of the shocks, except we eliminate the countercyclical variation in idiosyncratic risk. The Markov process for the log of the nondiversified income share, lnη, has a standard deviation of 0.71 and an autocorrelation of 0.89. The transition probability is denoted by ⎡ 0.945 0.055 ⎤ π (η ′ η ) = ⎢ ⎥. ⎣ 0.055 0.945 ⎦ Federal Reserve Bank of St. Louis REVIEW Second Quarter 2019 123 Chien and Lee The two states of the idiosyncratic shock, for which the mean is normalized to 1, are ηL = 0.3894 and ηH = 1.6106. All households have the identical constant relative risk aversion preference. In our calibration, there are strong incentives for household to save because of idiosyncratic shocks in an incomplete-market environment, which causes the risk-free rate to be lower than the reciprocal of the preference discount factor β, even in a growing economy As a result, we set the time discount factor to β = 0.95 to match the low risk-free rate in our benchmark model. The risk-aversion rate α is set to 4 to produce a high risk premium in our benchmark calibration. Following Mendoza, Quadrini, and Rios-Rull (2009), the fraction of nondiversifiable output is set to 88.75 percent. As shown in Section 2, equity in our model is simply a leveraged claim to the diversifiable income process. Following Abel (1999) and Bansal and Yaron (2004), the leverage ratio parameter is set to 3. 3.2 The Composition of Traders In the model, 50 percent of U.S. households are stock market nonparticipants, as in the 2010 SCF data. The remaining 50 percent do hold equities, and we divide them into non- Mertonian equity traders and Mertonian traders as discussed. To match the high risk premia, a small fraction of Mertonian traders must absorb a large amount of aggregate risk. We therefore set the fraction of Mertonian traders to 5 percent for our benchmark economy. The remaining fraction of households, 45 percent, are classified as non-Mertonian equity traders, who can own constant portfolio shares in short risk-free bonds, long-term risky bonds, and equities. In addition to the equity market participation rate, the portfolio shares of non-Mertonian traders are also important parameters. Again, we use the 2010 SCF data to calibrate the portfolio share of non-Mertonian equity traders and non-participants in our model, which account for 45 percent and 50 percent of the population, respectively. To identify the portfolio choice of the non-Mertonian equity traders, we must first sort by equity position the 50 percent of households in the data that hold equities and then compute the average equity share excluding the top 5 percent of equity holders. The average computed equity share is 21.1 percent, which we use as the equity share of non-Mertonian equity traders in the benchmark case. This calibration reflects the observations both from the data and from our model that more sophisticated households tend to hold larger amounts of equities. 3.3 The Real Yield Curve in the Data Using the constructed international data on zero-coupon yields, Wright (2011) demonstrates that nominal term premia are estimated to be positive among 10 industrialized countries. However, we do not observe the real term premium directly from the data and the positive nominal term premia do not necessarily imply positive real term premia, because of inflation risk. Strong empirical evidence supporting the real term premia comes from the real Treasury yield. The average real Treasury yields for 5-year, 7-year, and 10-years maturities are listed in Table 1 for different sample periods. For 2003 to 2007, the average 5-year, 7-year, and 10-year real Treasury yields are 1.646 percent, 1.871 percent, and 2.061 percent, respectively, which indicates a positive real term premium. The real yield data decrease after 2009 as the sample 124 Second Quarter 2019 Federal Reserve Bank of St. Louis REVIEW Chien and Lee Table 1 The Real Treasury Yield Period/Maturity 5-year 7-year 10-year 2003 to 2007 1.646 1.871 2.061 2003 to 2009 1.513 1.757 1.962 2003 to 2011 1.160 1.452 1.715 2003 to 2013 0.772 1.082 1.367 2003 to 2015 0.658 0.968 1.226 SOURCE: U.S. Department of Treasury; https://www.treasury.gov/resource-center/data-chart-center/interest-rates/Pages/TextView.aspx?data=realyield. Table 2 Quantitative Results Case Benchmark No mortgage No NME No HTT RA economy Mertonian 5% 5% 50% 100% 100% Non-Mertonian equity 45% 45% 0% 0% 0% Nonparticipant 50% 50% 50% 0% 0% ωet (0.211,–0.195) (0.211,0) NA NA NA ωnp (0,–0.537) (0,0) NA NA NA E(R zc30 – R f ) 1.924 1.713 0.678 0.427 0.532 σ (Q ) E (Q ) 0.475 0.464 0.204 0.135 0.133 ⎛ σ (Q ) ⎞ Std ⎜ t ⎝ Et ( Q ) ⎟⎠ 9.766 9.494 1.077 0.033 0.000 E(R e – R f ) 7.262 6.937 2.602 1.697 1.741 σ (Re − Rf ) E (Re − Rf ) 0.465 0.455 0.204 0.135 0.132 E(R f ) 0.949 1.206 3.485 3.923 11.714 σ(R f ) 1.449 1.296 1.326 1.365 2.049 NOTE: Parameter settings: risk aversion rate, γ = 4; discount factor, β = 0.95; nondiversifiable share of income, γ =0.885; and leverage ratio: ф = 3. NME, non-Mertonian equity trader. HTT, heterogeneous trading technologies. RA, representative agent. The simulation results are generated by an economy with 3,000 agents and 10,000 periods. Federal Reserve Bank of St. Louis REVIEW Second Quarter 2019 125 Chien and Lee Figure 1 The Average Real Yield Curve in the Benchmark Case, the Zero-Coupon Bond Percent 3.5 3.0 2.5 2.0 1.5 1.0 0.5 5 10 15 20 25 30 35 40 45 50 Year periods become longer. This is because the short-term real rate becomes negative. However, the sizable positive real term premia implied by the data should remain robust to the different sample periods. 3.4 Benchmark Results The benchmark asset-pricing results are shown in the “Benchmark” column in Table 2. First, we report the real term premia of our model, defined as the average yield difference between a 30-year and a 1-year zero-coupon bond, denoted by E(R zc30 – R f ). In addition, Table 2 also includes the market price of risk, σ(Q)/E(Q); the conditional standard deviation of the market price of risk, std(σ(Q)/E(Q)); the equity premium E(R e – R f ), the Sharpe ratio on equity returns E(R e – R f )/σ(R e – R f ); the average risk-free rate E(R f ); and the standard deviation of the risk-free rate σ(R f ). As explained earlier, the long bond is risky because its price tends to fall during recessions, which is simply a result of higher expected growth after recessions. In our benchmark case, the average term premium of a 30-year zero-coupon bond is 1.924 percent. To illustrate the upward-sloping real yield curve, Figure 1 plots the yield curve of zero-coupon bonds. In addition, our benchmark economy produces a high and volatile market price of risk as well as a low and stable risk-free rate. These asset-pricing statistics are hard to match in standard macroeconomic model, as indicated by Mehra and Prescott (1985). In the benchmark 126 Second Quarter 2019 Federal Reserve Bank of St. Louis REVIEW Chien and Lee case of Table 2, the market price of risk is high, 0.475, and volatile, with a standard deviation of 9.766 percent. The equity premium reaches 7.262 percent, and the Sharpe ratio on equity is 0.465. The average risk-free rate is low at 0.949 percent, and its volatility is only 1.449 percent. Hence, our calibrated model is capable of producing reasonable asset-pricing results. In our model, the success of matching high-risk premia and low risk-free rates relies on two key frictions. The first friction is the incomplete market with respect to idiosyncratic risk. It is well known that incomplete-market models can produce reasonable risk-free rates in a growing economy. The second friction, which is limited participation combined with a relatively small fraction of Mertonian traders, produces a high equity premium by concentrating aggregate risk on Mertonian traders. Our results show the mechanism of our model is able to deliver positive term premia of sufficient magnitude. This success comes from imposing the trend-reverting consumption process and heterogeneous portfolio choices into an otherwise standard macroeconomic model. In the next section, we explore the relative importance of these assumptions. 4 TREND REVERTING VERSUS A RANDOM WALK The trend-reverting endowment process is important to our results. If the endowment growth process truly follows a random walk, then long-bond returns do not necessarily fall during recessions and hence the standard model might fail to generate even a positive bond premium. In this subsection, we demonstrate this point analytically in the representative agent economy. Given the assumption of our shock process, the following lemma describes the expression for the term premia as well as its property in a representative agent economy. Lemma 1. The unconditional expectation of the term premium for a k-period zero-coupon bond is (6) ⎡ 1 1− ρ 2k ⎤ α 2σ ε2 E ⎡⎣rtk − rt1 ⎤⎦ = ⎢1− . 2 ⎥ ⎣ k 1− ρ ⎦ 2 In addition, the term premium, E[r tk – rtl ] is increasing in k given 0 < ρ < 1. Proof. Please refer to Appendix Section A.2. With 0 < ρ < 1 in the representative agent economy, Lemma 1 not only shows a positive term premium but also shows that the term premium is increasing in k, an indication of an upward-sloping real yield curve. However, if ρ = 1, the random-walk case, the average term premium shown in equation (6) becomes zero and independent of k. The independence implies a flat yield curve. This result is not surprising in the sense that the autocorrelation of the consumption growth rate is negative when 0 < ρ < 1 and becomes zero when ρ = 1. This can be seen clearly ⎛ 1− ρ ⎞ 2 from the fact that cov ( ln g t+1, ln g t ) = − ⎜ σ < 0 if ρ < 1. However, if ρ is sufficiently close ⎝ 1+ ρ ⎟⎠ to 1, then the consumption growth correlation is close to zero, which is not far from what we observe in the data. Federal Reserve Bank of St. Louis REVIEW Second Quarter 2019 127 Chien and Lee The quantitative results of a representative agent economy with a trend-reverting process are reported in the “RA economy” column of Table 2. The term premium is positive, as shown in this subsection. However, the term premium is not sizable, at only 0.487 percent, because of the absence of a high market price of risk, which drops significantly to only 0.133 from 0.475 in our benchmark economy. 5 INSPECTION OF THE MECHANISM In our benchmark economy, two features quantitatively contribute to the sizable real term premia: the heterogeneous trading technologies and the mortgage effects. In this section, we decompose the contribution of each feature by removing each one from our benchmark economy. 5.1 No Mortgage Effects Part of the real bond risk premia could be from the asymmetric bond portfolio holdings across households, which is motivated by the heterogeneous amounts of mortgages held across households. Here, we simply shut down this channel by assuming that both nonparticipants and non-Mertonian equity holders do not have a position on long-term bonds. The “No mortgage” column of Table 2 reports the results. We find that this asymmetric bond portfolio channel has minor effects on the market price of risk and equity risk premia. The market price of risk decreases slightly, from 0.475 to 0.464, as does the standard deviation of the market price of risk, from 9.766 percent to 9.494 percent. The equity premium decreases 32.5 basis points to 6.937 percent. The risk-free rate increases to 1.206 percent from 0.949 percent, and the standard deviation of the risk-free rate decreases to 1.296 percent from 1.449 percent. As for the impact on the real term premia, the 30-year term premium drops by 21 basis points. This exercise revels that the asymmetric portfolios in terms of bond maturity play only a minor role in increasing bond risk premia. 5.2 No Non-Mertonian Equity Traders In this subsection, we highlight the importance of distinguishing between Mertonian traders and non-Mertonian equity traders. Both types of traders hold equity, and hence it is not easy to distinguish between them in the data. In fact, most of the work in the segmented market literature does not consider the possibility of different types of equity market participants. We remove the differences between these two types of equity market participants and assume all are Mertonian traders. The assumption of a trend-reverting process for the endowment growth rate remains unchanged. The results are reported in the “No NME” column of Table 2. With half of total households now Mertonian traders and able to absorb the residual risk created by nonparticipants, aggregate risk is not concentrated enough quantitatively. The market price of risk and the term premium are greatly reduced. Quantitatively, the market price of risk drops significantly to 0.204 from 0.475. The equity premium is only 2.602 percent, 128 Second Quarter 2019 Federal Reserve Bank of St. Louis REVIEW Chien and Lee and the term premium for a 30-year zero-coupon bond drops significantly, to only 0.678 percent. This sensitivity analysis suggests that a high market price of risk is essential to our results. The large fraction of nonparticipants and few equity market participants not only helps to match the high and volatile equity risk premia, but also goes a long way to increasing the real bond risk premia. 5.3 No Heterogeneous Trading Technologies In this subsection, we remove totally the heterogeneous trading technologies assumed in the model. All households are now Mertonian traders and face no restriction on their portfolio choices. The results are reported in the “No HTT” column of Table 2. With the force of hetero geneous risk loading completely absent, the model acts similar to a Bewley-type model and exhibits low risk premia. The market price of risk, the equity premium, and the term premium are all small and close to those in a representative agent model.3 6 CONCLUSION We find that a slow mean-reverting shock process of consumption growth and a segmented asset-market mechanism with a heterogeneous trading technology can quantitatively account for the positive and sizable term premium for bonds as suggested by the data. The slow mean-reverting consumption process explains the positive term premia, although the size of the premia is still very small quantitatively. Our quantitative exercise shows that with this slight modification of the aggregate shock process, the long-term bond is risky because the risk-free rate is slightly countercyclical, even in the representative agent model. The segmented market mechanism with a heterogeneous trading technology and an asymmetric bond position across households can amplify the size and magnitude of the term premia while raising the market price of aggregate risk. We think our model is the first step in resolving the inconsistency between theoretical macroeconomic models and the empirical asset-pricing findings of the yield curve, with no modification of preferences. There are two obvious directions for future research that can improve the model. For the asset-pricing literature, one can enrich the model by additionally introducing a long-term consumption growth shock in order to match more properties of real bond premia found in the empirical literature. For macroeconomics, our mechanism does not rely on the modification of preferences, and the behavior of asset pricing is pinned down by a relatively small set of marginal investors in a segmented market mechanism. Therefore, the consumption and saving behaviors of most households stay close to those in a standard macroeconomic model. It is more likely that our results extend to a general equilibrium production economy without comprising dynamics of other macroeconomic variables. n Federal Reserve Bank of St. Louis REVIEW Second Quarter 2019 129 Chien and Lee APPENDIX A.1 THE TIME-ZERO TRADING HOUSEHOLD PROBLEM This appendix describes an equivalent version of this economy in which all households trade at time zero. The time-zero price of a claim that pays one unit of consumption in node zt can be constructed recursively from the one-period-ahead Arrow prices: ( ) ( ) ( ) ( ) ( ) P z t π z t = Q zt z t−1 Q zt−1 z t−2 …Q z1 z 0 Q ( z0 ) . The net financial wealth position of any trader given the trader’s history can be stated as ( ) −at z t , η t = ∑ ( ) ( ( s s ∑ )( t s≥t z , η ± z , η t ( ) ( ) ( ) P% z s , η s ⎡⎣γ Y z s ηs − c z s , η s ⎤⎦ , ) ) ( ) where P% z t , η t = π z t , η t P z t . From the above equation, we are able to write the household problem in the form of a time-zero trading problem, as shown in the next subsection. A.1.1 The Household Optimization Problem Following Chien, Cole, and Lustig (2011), we state the household problem in this Arrow- Debreu economy. We start with the Mertonian traders’ problem in the home country. There are two constraints. Let χ denote the multiplier on the present value budget constraint and φ(zt ,η t ) denote the multiplier on debt constraints. The saddle-point problem of a Mertonian trader can be stated as follows: ∞ max t L = {min χ ,υ ,ϕ } {c} ∑ β t=1 1 ∑ 1− α c ( z t , η t ) π ( z t , η t ) 1−α z t ,η t ⎫⎪ ⎧⎪ ∞ + χ ⎨∑ ∑ P% z t , η t ⎡⎣γ Y z t ηt − c z t , η t ⎤⎦ + a0 z 0 ⎬ t t ⎪⎭ ⎩⎪ t=1 ( z ,η ) ( ( ) ( ⎪ ∑ ϕ t ( z t , η t ) ⎨∑ t=1 z t ,η t ( ) ( ) ⎫⎪ P% z s , η s ⎡⎣γ Y z t ηs − c z s , η s ⎤⎦ ⎬. ⎪⎩ s≥t ( z s ,η s )±( z t ,ηt ) ⎪⎭ ⎧ ∞ −∑ ) ) ∑ ( ) ( ) ( ) The first-order condition with respect to consumption is given by (7) t ( β c z t ,ηt ) −α ( ) ( ) ( ) = ζ z t , η t P z t for all z t , η t , where ζ(zt ,η t ) is defined recursively as ( ) ( ) ( ζ t z t , η t = ζ t−1 z t−1, η t−1 − ϕt z t ,η t ) with initial ζ 0 = χ. It is easy to show that this is a standard convex constraint maximization problem. Therefore, the first-order conditions are necessary and sufficient. Non-Mertonian traders face additional restrictions on their portfolio choices. Let vt(zt ,η t ) denote the multiplier on portfolio restrictions. Given the same definition of other multipliers 130 Second Quarter 2019 Federal Reserve Bank of St. Louis REVIEW Chien and Lee as in the active-trader problem, the saddle-point problem of a nonparticipant trader whose assets at the end of the period are ât–1(zt–1,η t–1) can be stated as ∞ max t L = { min χ ,υ ,ϕ } {c, ŝ} ∑ β t=1 1 ∑ 1− α ct ( z t , η t ) π ( z t , η t ) 1−α z t , ηt ⎫⎪ ⎧⎪ ∞ χ ⎨∑ ∑ P% z t , η t ⎡⎣γ Y z t ηt − c z t , η t ⎤⎦ + a0 z 0 ⎬ t t ⎪⎭ ⎩⎪ t=1 ( z , η ) ⎧ Σ s≥t Σ s s t t P% z s , η s ⎡γ Y z s ηs − c z s , η s ⎤ ⎫ ∞ ⎣ ⎦⎪ ( z , η )±( z , η ) t t ⎪ + ∑ ∑ νt z ,η ⎨ ⎬ p t=1 ( z t , η t ) ⎪⎩ ⎪⎭ z t ŝt−1 z t−1, η t−1 + P% z t , η t Rt,t−1 ∞ ⎫⎪ ⎧⎪ − ∑ ∑ ϕ t z t , η t ⎨∑ P% z s , η s ⎡⎣γ Y z s ηs − c z s , η s ⎤⎦ ⎬. ∑ s s t t t=1 ( z t , η t ) ⎩⎪ s≥t ( z , η )±( z , η ) ⎭⎪ ( ) ( ) ( ) ( ) ( ) ( ) ( ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) ) The first-order condition with respect to consumption is given by t ( ) β c z t ,ηt −α ( ) ( ) ( ) = ζ z t , η t P z t for all z t , η t , where ζ(zt ,η t ) is defined as ( ) ( ) ( ) ( ) ζ t z t , η t = ζ t−1 z t−1, η t−1 + ν t z t , η t − ϕt z t , η t . Therefore, the first-order condition with respect to consumption is independent of trading restrictions. The first-order condition with respect to total asset holdings at the end of period t–1, ŝt–1(zt–1,η t–1) is p ( z t ) v t ( z t , η t ) P ( z t )π ( z t , η t ) = 0 ∑ Rt,t−1 t z ,η t for all z t , η t . This condition varies according to the different trading restrictions. A.1.2 Stochastic Discount Factor By summing the first-order condition with respect to consumption, equation (7), across all domestic households at period t, we obtain the consumption sharing rule as follows: ( c z t ,ηt ( ) C z ( ) t )= ( where ht(zt ) is defined as ht z t ≡ ∑ζ ( z t , η t ) η t ) h (z ) ζ z t ,ηt t − 1 α t − 1 α , ( ) π η t . In addition, by plugging the consump- tion sharing rule back into to the first-order condition with respect to consumption, equation (7), we obtain the price of the home consumption basket at state zt : ( ) P z t = β t C ( z t ) ht ( z t ) . Federal Reserve Bank of St. Louis REVIEW −α α Second Quarter 2019 131 Chien and Lee Therefore, the stochastic discount factor is given by the Breeden-Lucas stochastic discount factor with a multiplicative adjustment: ( Qt+1 z t+1 ( ) = β ⎛ C(z )⎞ z )≡ ⎜ ⎟ P(z ) ⎝ C(z ) ⎠ P z t+1 t t+1 t t −α ( ) ( ) α ⎛ ht+1 z t+1 ⎞ ⎜ ⎟ . t ⎝ ht z ⎠ A.2 PROOF OF LEMMA 1 Given that our assumed growth rate of output is g t+1 = ge ( ρ −1)mt +εt+1 the one-period-ahead − α ρ −1 m − αε pricing kernel is Mt,t+1 = β g −α e ( ) t e t +1 . The price of a one-period bond is therefore Pt1 = Et Mt,t+1 = β 2 2σ ε − α − α ( ρ −1)mt α 2 g e e , which is a function of the current deviation from trend mt. The one-period yield is then rt1 = −lnPt1 = −ln β + α ln g + α ( ρ −1)mt − α 2σ ε2 . 2 We derive the price of a k-period zero-coupon bond and its yields as follows: ⎡ k−1 ⎤ Ptk = Et ⎡⎣ Mt,t+k ⎤⎦ = Et ⎢ ∏ Mt+τ ,t+τ +1 ⎥ ⎣τ +0 ⎦ k α 2σ ε2 k=1 τ ⎡ k−1 2 τ ⎤ −kα − α ( ρ −1)⎡⎣ ∑τ =0 ρ ⎤⎦mt ⎣ ∑τ =0 ρ ⎦ 2 g e e =β 1 rtk = lnPtk k = −lnβ + α ln g + α ( ρ −1) , 1 1− ρ k t 1− ρ 2k α 2σ ε2 m − . 1− ρ 2 2 k 1− ρ The term premium at period t for a k-period zero-coupon bond is ⎡ 1 1− ρ k ⎤ t ⎡ 1 1− ρ 2k ⎤ α 2σ ε2 rtk − rt1 = α (1− ρ ) ⎢1− ⎥ m +1 ⎢1− k 1− ρ 2 ⎥ 2 , ⎣ k 1− ρ ⎦ ⎣ ⎦ which is again a function of mt. Taking the unconditional expectation of rtk – rtl gives equation (6) because of E(mt ) = 0. Moreover, we want to show that the term premium is increasing with k, that is, for ∂E ⎡⎣rtk − rt1 ⎤⎦ > 0 for k > 1. First, notice that ρ 2k 1− ln ρ 2k <1 because (i) ∂k lim ρ 2 k →1ρ 2k 1− ln ρ 2k = 1 and (ii) ( ( ) ) ( ) ∂ ⎡⎣ ρ 2k 1− ln ρ 2k ⎤⎦ = −ln ρ 2k > 0 if ρ 2k <1. ∂ ρ 2k τ Second, by using the fact that ∑τk−1 =0 ρ = 132 Second Quarter 2019 1− ρ 2k , the average term premium can be rewritten as 1− ρ 2 Federal Reserve Bank of St. Louis REVIEW Chien and Lee E ⎡⎣rtk − rt1 ⎤⎦ = α 2σ ε2 α 2σ ε2 1− ρ 2k − × 2 k 2 1− ρ 2 ( ) and hence the derivative of E[rtk – rtl ] with respect to k is ∂E ⎡⎣rtk − rt1 ⎤⎦ ∂k = 1− ρ α 2σ ε2 × 2 2 1− ρ ( ) 2k (1− ln ρ ) 2k k2 > 0 if ρ < 1 and k >1. The last inequality uses the fact that ρ2k(1 – lnρ2k) < 1 if ρ2k < 1. NOTES 1 The standard equilibrium macroeconomic model refers to the classical representative agent complete-market model, such as the one in Mehra and Prescott (1985). See also Grkaynak and Wright (2012) for a review of issues related to term premia. 2 In this article, the terms “household,” “trader,” and “investor” are used interchangeably. 3 The statistics for the representative agent economy are reported in the “RA economy” column of Table 2. REFERENCES Abel, A.B. “Risk Premia and Term Premia in General Equilibrium.” Journal of Monetary Economics, 1999, 43, pp. 3-33; https://doi.org/10.1016/S0304-3932(98)00039-7. Alvarez, F. and Jermann, U. “Quantitative Asset Pricing Implications of Endogenous Solvency Constraints.” Review of Financial Studies, 2001, 14, pp. 1117-52; https://doi.org/10.1093/rfs/14.4.1117. Backus, D.K.; Gregory, A.W. and Zin, S.E. “Risk Premiums in the Term Structure: Evidence from Artificial Economies.” Journal of Monetary Economics, 1989, 24(3), pp. 371-99; https://doi.org/10.1016/0304-3932(89)90027-5. Bansal, R. and Shaliastovich I. “A Long-Run Risks Explanation of Predictability Puzzles in Bond and Currency Markets.” Review of Financial Studies, 2013, 26(1), pp. 1-33; https://doi.org/10.1093/rfs/hhs108. Bansal, R. and Yaron A. “Risks for the Long Run: A Potential Resolution of Asset Prizing Puzzles.” Journal of Finance, 2004, 59, 1481-509; https://doi.org/10.1111/j.1540-6261.2004.00670.x. Campbell, J.Y. “Bond and Stock Returns in a Simple Exchange Model.” Quarterly Journal of Economics, 1986, 101, pp. 785-804; https://doi.org/10.2307/1884178. Chien, Y.; Cole H. and Lustig H. “A Multiplier Approach to Understanding the Macro Implications of Household Finance.” Review of Economic Studies, 2011, 78(1), pp. 199-234; https://doi.org/10.1093/restud/rdq008. Christiano, L.J. and Eichenbaum M. “Unit Roots in Real GNP: Do We Know, and Do We Care?” Working Paper Series, Macroeconomic Issues 90-2, Federal Reserve Bank of Chicago, 1990. den Haan, W.J. “The Term Structure of Interest Rates in Real And Monetary Economies.” Journal of Economic Dynamics and Control, 1995, 19(5-7), pp. 909-40; https://doi.org/10.1016/0165-1889(94)00813-W. Donaldson, J. B.; Johnsen, T. and Mehra, R. “On the Term Structure of Interest Rates.” Journal of Economic Dynamics and Control, 1990, 14(3-4), pp. 571-96; https://doi.org/10.1016/0165-1889(90)90034-E. Grkaynak, R.S. and Wright, J.H. “Macroeconomics and the Term Structure.” Journal of Economic Literature, 2012, 50(2), pp. 331-67; https://doi.org/10.1257/jel.50.2.331. Guiso, L. and Sodini, P. “Household Finance. An Emerging Field.” EIEF Working Paper, 2012. Federal Reserve Bank of St. Louis REVIEW Second Quarter 2019 133 Chien and Lee Mehra, R. and Prescott E.C. “The Equity Premium: A Puzzle.” Journal of Monetary Economics, 1995, 15(2), pp. 145-61; https://doi.org/10.1016/0304-3932(85)90061-3. Mendoza, E.G.; Quadrini, V. and Rios-Rull, J.-V. “Financial Integration, Financial Development, and Global Imbalances.” Journal of Political Economy, 2009, 117(3), pp. 371-416; https://doi.org/10.1086/599706. Piazzesi, M. and Schneider, M. “Equilibrium Yield Curves,” in NBER Macroeconomics Annual 2006. Volume 21. National Bureau of Economic Research, Inc., 2007, pp. 389-472. Rudebusch, G.D. and Swanson, E.T. “The Bond Premium in a DSGE Model with Long-Run Real and Nominal Risks.” American Economic Journal: Macroeconomics, 2012, 4(1), pp. 105-43; https://doi.org/10.1257/mac.4.1.105. Storesletten, K.; Telmer, C.I. and Yaron, A. “Cyclical Dynamics in Idiosyncratic Labor Market Risk.” Journal of Political Economy, 2004, 112(3), pp. 695-717; https://doi.org/10.1086/383105. Ulrich, M. “Inflation Ambiguity and the Term Structure of U.S. Government Bonds.” Journal of Monetary Economics, 2013, 60(2), pp. 295-309; https://doi.org/10.1016/j.jmoneco.2012.10.015. Wachter, J. “A Consumption-Based Model of the Term Structure of Interest Rates.” Journal of Financial Economics, 2006, 79, pp. 365-99; https://doi.org/10.1016/j.jfineco.2005.02.004. Wright, J.H. “Term Premia and Inflation Uncertainty: Empirical Evidence from an International Panel Dataset.” American Economic Review, 2011, 101(4), pp. 1514-34; https://doi.org/10.1257/aer.101.4.1514. 134 Second Quarter 2019 Federal Reserve Bank of St. Louis REVIEW Racial Gaps, Occupational Matching, and Skill Uncertainty Limor Golan and Carl Sanders White workers in the United States earn almost 30 percent more per hour on average than Black workers, and this wage gap is associated with large racial differences in occupational assignments. In this article, we theoretically and empirically examine the Black-White disparity in occupations. First, we present a model based on Antonovics and Golan (2012) that relates occupational assignments to the incentives workers face while learning about their own unknown ability. Second, we document differences between Black and White workers in both the complexity of skills required in their initial occupations and the growth rates of this complexity over time. To do this, we match panel data from the National Longitudinal Survey of Youth 1979 with the Dictionary of Occupational Titles measures of occupational characteristics and find that, compared with White workers, Black workers start in occupations requiring less-complex skills, see slower growth in job complexity over time, and are relatively more likely to transition to jobs with lower complexity. Finally, we consider the relationship between our model and our empirical findings; for example, discrimination in hiring early in the career can have long-term consequences on the ability of Black workers to learn their best occupa tional match and explains part of their lower wage growth. We conclude with suggestions for policy and future research directions. (JEL J01, J24, J31) Federal Reserve Bank of St. Louis Review, Second Quarter 2019, 101(2), pp. 135-53. https://doi.org/10.20955/r.101.135-53 1 INTRODUCTION The labor market experiences of Black and White workers in the United States are dramati cally different. A first-order difference is the well-documented racial wage gap: The average hourly wage for White workers is 30 percent higher than for Black workers. This racial wage gap has been shown to reflect differences in both socioeconomic backgrounds and discrimina tory practices in the labor market and has created a sizable literature across multiple disciplines. But differences in earnings do not exhaust the racial differences in labor market experiences. In this article, we consider racial gaps in workers’ occupations, that is, differences in what Limor Golan is an associate professor at Washington University in St. Louis and a research fellow at the Federal Reserve Bank of St. Louis. Carl Sanders is an assistant professor at Washington University in St. Louis. © 2019, Federal Reserve Bank of St. Louis. The views expressed in this article are those of the author(s) and do not necessarily reflect the views of the Federal Reserve System, the Board of Governors, or the regional Federal Reserve Banks. Articles may be reprinted, reproduced, published, distributed, displayed, and transmitted in their entirety if copyright notice, author name(s), and full citation are included. Abstracts, synopses, and other derivative works may be made only with prior written permission of the Federal Reserve Bank of St. Louis. Federal Reserve Bank of St. Louis REVIEW Second Quarter 2019 135 Golan and Sanders White and Black workers do rather than what they earn. Even the most basic descriptive statistics show large differences between Black and White workers in the types of work they perform. For example, Black workers were 12 percent of the working population in 2016 and made up 26 percent of the occupation “truck and tractor operators,” while making up only 3 percent of “chief executives.” White workers showed the opposite pattern, making up 45 per cent of the working population in 2016 and 85 percent of chief executives.1 Our focus on racial differences in occupations is consistent with recent work in empirical labor economics, which links differences in occupations and occupational mobility to workers’ wage growth. It is often more informative to know someone’s occupational title than their current wage: taking two young workers who each earn $15 an hour, the knowledge that one has the occupation “accountant” and the other “refrigeration mechanic” helps to make the (on average correct) prediction that the accountant will make significantly more than the mechanic 10 years later. Understanding the reasons Black and White workers take different occupations can provide insight into differences in wage levels and wage growth. The primary contribution of this article is documenting and interpreting the differences in the relative occupational assignments of Black and White workers. In the first part of the study, we present an economic model to use as a framework for interpreting our empirical results on occupational choice, occupational turnover, and wage growth. In the second section, we document the racial gaps in occupational choice in representative U.S. data, looking both at aggregate trends in occupational choice within a worker’s career and at occupation-tooccupation transition rates. Finally, in our empirical results, we compare the predictions of the model to our descriptive findings and discuss the implications of the underlying economic mechanisms for policy and future research. Section 2 of our article gives a framework to interpret occupational mobility across races as the result of different economic circumstances. Our model is a learning model following Antonovics and Golan (2012) that is capable of generating occupational mobility and wage growth across the career. For simplicity, we present analysis of a two-period model. Each period, workers choose an occupation to maximize the expected present discounted value of lifetime income, but occupational choice is complicated by the fact that workers do not know their own ability and thereby their best occupational match. Working allows a worker to learn about his ability over time, but different jobs give information about the worker’s skills at different rates. The amount that workers learn about their skills depends on the intensity of their job. Different jobs require performance of tasks that require varying intensities of unobservable skills. The more output depends on the unobserved skills, the more information the job reveals about those skills. For example, workers learn more about their ability as a manager in management jobs. Information about skill levels may increase future earnings because it allows a better assignment of workers to jobs. Thus, workers experiment, forgoing expected current-period output in order to learn about their skills by taking jobs they would not take otherwise. Antonovics and Golan’s (2012) results show that the optimal level of experimen tation is initially small, increases as workers gain experience, and then declines as workers become increasingly certain about their skills. 136 Second Quarter 2019 Federal Reserve Bank of St. Louis REVIEW Golan and Sanders Mapping the theoretical concept of “occupational intensity” to the data requires nonstandard measures, since in Census-like data sets only the occupational title is recorded rather than any specific on-the-job activities. To overcome this data limitation, we use data on occu pational characteristics from the Dictionary of Occupational Titles (DOT) merged with worker- level panel data from the National Longitudinal Survey of Youth 1979 (NLSY79). We use the detailed occupation-level characteristics from the DOT to reduce the unordered list of occu pational titles into a single-dimensional index that ranks occupations by the degree to which output depends on skills that are difficult to observe directly, e.g., creativity. This index ranks occupations with respect to the dependence of output on skills that are hard to observe, which we call “complexity” throughout for brevity.2 In Section 3, we analyze the merged data sets and show that, as expected, the average White worker’s first occupation tends to be more-complex than the average Black worker’s first occupation. Moreover, over a career, the average White worker’s occupational complexity grows faster than the average Black worker’s. Further empirical analysis considers whether these differences are driven by rates of occupational switching, differences in the promotion rates of less- versus more-complex occupations, and the role of demographic characteristics and education in explaining these occupational gaps. When we look more closely at the rates of occupational switching, we find that the slower pattern of occupational upgrading by Black workers relative to White workers is not driven by lower occupational mobility. Rather, Blacks are marginally more likely to switch occupa tions than Whites, but a greater proportion of Black occupational transitions are “downgrades,” that is, switches toward occupations characterized by lower levels of complexity. Given an occupational switch, White workers make an occupational upgrade 54 percent of the time, compared with 51 percent for Black workers. Regarding potential average demographic differences between races, we examine to what extent differences in workers’ first occupations and the growth rate of occupational complexity are driven by race alone rather than other explanations. For example, Black workers may begin in less-complex occupations on average more often than White workers due simply to average differences in education levels. We find that racial differences in the speed of occupa tional upgrading persist even if we consider White and Black workers who are originally in the same occupation. This finding suggests that race-specific factors such as discrimination might partially explain differences in occupational transition rates. Additionally, we consider the first job that workers have and find that controlling for measurable demographics such as education level and test scores does not eliminate the effect of race on the initial occupational assignment: Black workers with seemingly very similar skills as White workers tend to work in less-complex occupations. Our economic model suggests that if Black workers are discriminated against in hiring for high-complexity occupations, it can have long-term effects on their occupational complexity and wage growth relative to White workers. From a policy perspective, one of the major issues surrounding racial labor market gaps is the question of mismatch. Workers may not always be well matched to their job or occu pation, and if demographic background differences and racial discrimination make this problem Federal Reserve Bank of St. Louis REVIEW Second Quarter 2019 137 Golan and Sanders more severe, there may be large productivity gains by improving the match of Black workers with their best occupations. In this article, we focus on the role of mismatch induced by infor mational frictions in the Black-White wage gap. Even if there were no barriers to hiring or finding jobs, workers and employers do not always have complete information about the worker’s skills or the suitability of those skills to the tasks required by the job. Over time, as a worker comes to know his or her own skills and the employer observe the worker’s performance, both the worker and the employer both could learn about the worker’s ability and compara tive advantage. Based on this information, workers change jobs and, over time, wages grow because workers work in occupations and jobs in which they are better matched. Policies that aim to improve the information available to both workers and employers could potentially reduce the costs of mismatch and, and if discrimination in occupational attainment leads to Black workers receiving less information about their skills over their careers, it may be nec essary to target these policies to Black workers. The remainder of the article is organized as follows: In Section 2 we review the literature on labor learning models and racial labor market gaps; in Sections 3 and 4 we set up and solve the two-period learning model; in Section 5 we describe worker and occupational data con struction; in Section 6 we present the empirical results; and in Section 7 we present conclusions. 2 RELATED LITERATURE The theoretical and empirical literature on uncertainty in the labor market primarily focuses on models of matching (see, for example, Jovanovic, 1979, and Miller, 1984) and models of learning (see, for example, Farber and Gibbons, 1996; Gibbons and Waldman, 1999; Neal, 1999; and Gibbons et al., 2005). There is also a set of empirical labor papers that analyze how workers and firms learn about unobserved ability and how quickly this happens (see Miller, 1984; Pastorino, 2009; Papageorgiou, 2014; James, 2012; Sanders, 2017; and Golan et al., 2017). There is a large literature on Black-White pay gaps as well as other racial differences in labor market outcomes. The economics literature emphasizes the importance of pre-market factors in these differences in outcomes. See Altonji and Blank (1999), Cameron and Heckman (2001), Carneiro et al. (2005), and Fryer (2011) for surveys of economic analyses of racial labor market gaps, including the empirical relationships between pre-market characteristics such as education, test scores, and family background, with a variety of labor market outcomes. There has been less recent focus on the role of post-market entry differences in the experiences of Black and White workers. The fact that similar Black and White workers are employed at different rates was originally discussed in the context of long-term trends in racial wage gaps: Brown (1984), Chandra (2000), Juhn (2003), and Western and Pettit (2005) all found evidence that more Black workers than White workers dropped out of the labor force between 1940 and 1990, and higher-wage Black workers were more likely to drop out than higher-wage White workers, increasing the measured wage gap. In the context of a single cohort, Eckstein and Wolpin (1999) emphasize the difference between actual and potential wage-offer distribu tions, making the point that observed wages can either under- or over-estimate discrimination. 138 Second Quarter 2019 Federal Reserve Bank of St. Louis REVIEW Golan and Sanders Antecol and Bedard (2004) find that including measures of actual (rather than potential) labor market experience closes even more of the gap in the Neal and Johnson (1996)-type wage specification. There are other economic theories of racial wage gaps that do not require systematic differences in average skills. For example, Oettinger (1996) develops and tests a model of statistical discrimination, where assuming Blacks have less-precise signals of ability predicts an increasing wage gap with experience, even if both races have equal mean ability. He finds that Blacks have lower gains to mobility, which causes the Black-White wage gap to rise with experience. Altonji and Pierret (2001) develop a test for statistical discrimination and do not find evidence supporting the hypothesis that employers discriminate based on easily observed characteristics such as education. This article focuses on describing the patterns in the occupational choices and job transi tions of Black and White individuals. While there is a growing literature relating occupational choices of workers, tasks performed, and data on occupational skill requirements (see Sanders and Taber, 2012, for a review of the literature on heterogeneous/multidimensional human capital, which often uses task-based data), the labor literature on job tasks has not typically focused on cross-race differences in labor force outcomes. Golan et al. (2017), a complemen tary paper, develops a generalized life-cycle model that includes occupational sorting, job turnover, multidimensional skill gaps, and taste differences, and quantifies the magnitude of the different factors accounting for the life-cycle racial gaps in earnings, participation, and occupational choices. 3 MODEL The model presented was developed in Antonovics and Golan (2012). We consider a two-period economy with risk-neutral workers and firms with a common discount factor δ. Workers differ in the sets of skills they possess. We examine a simple scenario in which each worker has only two skills: a known skill, k, and an unknown skill, θ, both of which are time invariant. For simplicity we assume that each firm offers one job. Each job differs in the extent to which output depends on k and θ. There is a continuum of jobs, j, each completely charac terized by a given value of α j [0,1], where α denotes the degree to which output depends on θ relative to k. Thus, choosing a job in period t is equivalent to choosing a value of α. Given this choice, we assume output in period t is given by (1) yt = α tθ + (1− α t ) k + ε t , where εt are independent and identically distributed productivity shocks and αt is the value of α chosen by the worker at time t. Note that there is one job in which output is sensitive only to θ(α = 1) and one job in which output is sensitive only to k(α = 0). For the rest of the jobs, the higher is α j, the more output depends on θ. Information in the model is symmetric; firms and workers have common priors on θ, k is known to everyone, and output is commonly observed. Workers and firms acquire addi Federal Reserve Bank of St. Louis REVIEW Second Quarter 2019 139 Golan and Sanders tional information about a worker’s unknown skill through successive observations of output. Thus, having observed output, workers and firms calculate xt = (2) yt − (1− α t ) k ε =θ + t , αt αt where xt serves as a signal of the worker’s unobserved skill, θ. The noise in xt is not indepen dent of a worker’s job choice. In particular, the higher is αt , the higher is the signal-to-noise ratio and the more information about θ the market is able to extract from xt. Under the assumption that the prior distribution of θ at time t is normal with mean μt and variance σt2 and the distribution of εt is normal with mean zero and variance σε2, the posterior distribution 2 of θ is known to be normal with mean μt+1 and variance σt+1 , where µt+1 = (3) and σ (4) 2 t+1 µtσ ε2%,t + xtσ t2 σ ε2%,t + σ t2 σ ε2%,tσ t2 = 2 σ t + σ ε2%,t σ ε2 . In addition, μt+1 is itself normally distributed with mean mt+1 and α t2 variance st2+1 given by and where σ ε2%,t = mt+1 = µt (5) 2 st+1 = (6) σ t4 . σ t2 + σ ε2%,t Thus, the posterior mean of θ follows a martingale and the more information xt reveals about θ (the higher is α), the higher is the variance of the posterior mean. We assume competitive markets and free entry into the labor market. Thus, wages are the expected productivity in each period. Workers’ current-period utility is given by Ut = wt so workers choose αt to maximize the expected present discounted value of lifetime wages. 3.1 Optimal Job Choice Workers work for two periods and then retire. Thus, the worker’s problem can be written as (7) ( ) V µ1 ,σ 12 = max α 1µ1 + (1− α 1 ) k + δ E1 ⎡⎣α 2 µ2 + (1− α 2 ) k ⎤⎦ , α t∈[ 0,1] which we can solve recursively beginning from the second period. The second-period optimal choice of job is given by (8) 140 Second Quarter 2019 ⎧⎪ 1, if µ2 ≥ k, α 2 ( µ2 ) = ⎨ ⎩⎪ 0, otherwise. Federal Reserve Bank of St. Louis REVIEW Golan and Sanders The second-period job assignment is simply to choose the job in which the wage is the highest. Solving backwards, we first solve for the optimal assignment in period one. Note that expected productivity in period two depends on the second-period belief, μ2, which in turn depends on α1 through x1 (see equations (2) and (3)). We therefore rewrite the first-period problem as (9) ( ) ∞ V µ1 ,σ 12 = max α 1µ1 + (1− α 1 ) k + δ ⎡ Φ ( r ) k + ∫k µ2 f ( µ2 )d µ2 ⎤ , ⎣ ⎦ α1∈[ 0,1] k − µ1 , Φ(.) is the standard normal cumulative density function and f is the normal s2 α 2σ 4 probability density function with mean m2 = μ1 and variance s22 = 2 1 2 1 . The above equation α 1 σ 1 +1 makes clear that when μ1 < k, there is a cost associated with selecting α1 > 0 since expected current-period output will be less than k. Thus, when μ1 < k, workers must weigh the benefit of increasing α1 in terms of expected second-period output against the cost in terms of expected current-period output. The second-period output is increasing in α1. The expected value of any left-truncated normal random variable is increasing in the variance of that random variable. Thus, since s22 is increasing in α1, we know that expected second-period output must also be increasing in α1. Intuitively, information is valuable because workers can truncate the loss if they receive a negative signal about θ by selecting future jobs with α = 0, but can take advantage of the arrival of positive information about θ by selecting jobs with α = 1. We say that a worker experiments if a worker chooses to forego expected current-period wages in order to gain information about θ, that is, if a worker with μ1 < k chooses α1 > 0. Proposition 2 in Antonovics and Golan (2012) establishes that for every value of beliefs μ1 < k, there exists a prior variance σ 2, so it is optimal to choose α > 0. That is, experimentation is beneficial if there is sufficient uncertainty about a worker’s skill. Intuitively, even when θ is believed to be very low, if there is sufficient uncertainty about θ, then the probability that θ > k is high enough that it is worth foregoing current-period output to gain additional information about θ’s true value. The first-order necessary condition for an interior solution is where r = (10) ∂φ ( r ) ∂s ∂Φ ( r ) k − µ1 s2 + 2 φ ( r ) + ( k − µ1 ) = , ∂α 1 ∂α 1 ∂α 1 δ where ϕ(r) is the standard normal probability density function. Thus, in an interior solution, the marginal benefit of experimenting with α in terms of second-period output is equal to the marginal cost in terms of first-period output. Note that when μ1 > k, the cost of increasing α is negative (the right-hand side of equation (10) is nega tive). Thus, when μ1 > k, both first-period and second-period expected output are increasing in α1, and there will be a corner solution at α1 = 1. As shown in Antonovics and Golan (2012), the optimal level of experimentation is increasing in μ; however, it is non-monotonic in the variance σ. This non-monotonicity is counterintutitive: We would expect that if you had less information you would always want to experiment more, but that is not the case. Federal Reserve Bank of St. Louis REVIEW Second Quarter 2019 141 Golan and Sanders To understand why there is a non-monotonic relationship between α1 and σ1, note that current-period expected output does not depend on σ12; this finding implies that the nonmonotonic relationship between σ12 and the optimal choice of α1 must depend solely on how increases in α1 affect expected future output. In particular, the effect of increasing α1 on expected future output must be low both when σ12 is small and when σ12 is large. When σ12 is small, the option value of new information is low because new information on θ is unlikely to have a large impact on the posterior mean of θ. Moreover, the expected loss of output due to incorrect future job assignments is small because the likelihood that θ is much different from μ1 is small. To see why the benefit of increasing α1 is also small when σ12 is large, recall that an increase in α1 increases expected future output through its effect on s22, the spread of μ2. As is clear from equation (6), however, s22 is also increasing in σ12 and the marginal effect of an increase in α1 on s22 is small when σ12 is large. Thus, when there is considerable uncertainty about a worker’s skill, the spread of μ is large and experimentation has little value on the margin since increased information has little effect on the optimal job assignment in the second period. While the two-period model cannot be generalized directly to a life-cycle model, intui tively, over the life-cycle the uncertainty about a worker’s skill decreases, which implies, hold ing μ constant, that experimentation may initially increase and then decrease (for some values of μ and σ) based on expectations (see details and analysis of a full dynamic problem in Section 4 of Antonovics and Golan, 2012). In our empirical work, we will use the intuition gained from the model to understand patterns of occupational choices and transitions in the data. 4 DATA AND EMPIRICAL IMPLEMENTATION In our empirical work, we merge the occupational work histories from the NLSY79 to occupational characteristics from the DOT in order to construct patterns of occupational complexity and wages over workers’ careers. 4.1 The Dictionary of Occupational Titles The model relates the occupational returns to unobservable skills to worker experimen tation; however, these returns are typically not directly observable. To create an empirical analogue of these returns, the α from the model, we use an index derived from occupationlevel characteristics that ranks occupations by the degree to which output depends on unob served skills. This index was derived in Antonovics and Golan (2012) and here corresponds to our idea of occupational “complexity.” The construction of α relies on information in the DOT. The DOT provides information on the primary tasks performed in a given occupation and the worker characteristics necessary for successful job performance. The occupational characteristics given in the DOT are linked to the 1970 Census three-digit occupation codes in an augmented version of the April 1971 Current Population Survey compiled by the Committee on Occupational Classification and Analysis at the National Academy of Sciences. The data in the DOT are both comprehensive and detailed, describing over 12,000 occupations along 44 dimensions. 142 Second Quarter 2019 Federal Reserve Bank of St. Louis REVIEW Golan and Sanders It is important to use a list of job characteristics that captures the importance of the unknown (hard-to-observe) skill to job performance. In our model, there are several key fea tures that characterize the unknown skill: First, there must be uncertainty about the skill prior to a worker’s entry into the labor market. Second, observing output only gradually reveals a worker’s skill and the more important the unobservable skill is to successful job performance, the more quickly the skill is revealed. In order to identify occupations in which hard-toobserve skill is important to job performance, we select occupational characteristics that indicate the importance of complex tasks. We define complex tasks as those for which it is hard to write down an explicit algorithm for successful completion. This is similar to the definition of “nonroutine” tasks in Autor et al. (2003). The reasoning is that if a task can be broken into an ordered list of well-defined actions, then a worker’s ability can be quickly learned by observing how the worker performs each separate action. In contrast, if it is difficult to explicitly describe how to successfully complete a task, then it will be difficult to determine a worker’s skill without observing his or her on-thejob performance.3 Using this list of variables, the index is created using principal component analysis. To ease comparison with the theoretical model, the index is converted into percent ages. Each normalized predicted score takes on a value between zero and 1, with higher values indicating a higher level of required skill. The index is then matched to the occupation data from the NLSY79. 4.2 The National Longitudinal Survey of Youth 1979 In order to construct this occupational work history, we use the NLSY79, which follows individuals born between 1957 and 1964. We focus our empirical analysis on males in the cross-sectional sample. Although the NLSY79 contains information on individuals’ labor force activities for each week from 1978 through the most recent year in which a respondent was interviewed, we rely only on labor market data from 1978 through 2000 because of a switch in occupational coding that occurred after 2000. If a respondent is not interviewed in a given year (or years), then at the next interview date, the respondent is asked to go back and retro spectively report their labor force activities. As a result, the NLSY79 allows us to construct relatively complete work histories. The work history data include information on each of up to five jobs a respondent may have held in a given week, and we define an individual’s occu pation in a given week to be their occupation in the job at which they worked the most hours. We follow individuals’ occupational histories starting with their first transition to full-time work after the completion of their highest degree. In particular, following completion of a degree, we identify the first week in which an individual works at least 10 hours per week and from which he continues to work at least 10 hours per week for at least 39 of the next 52 weeks. We then keep a running tab of the individual’s actual labor market experience and occupation in each week in which he works.4 In our empirical analysis, we focus on the first 350 weeks (about 6.7 years) of each individual’s actual experience in the labor force because attrition from the sample makes it difficult to construct complete work histories for longer horizons. We take as our initial sample the cross-sectional sample of White males and the crosssectional plus supplemental sample of Black males. We lose 829 respondents because we Federal Reserve Bank of St. Louis REVIEW Second Quarter 2019 143 Golan and Sanders Table 1 Summary Statistics Whites mean Standard deviation Blacks mean Standard deviation AFQT 62 (26) 29 (23) Percent with only high school or below 56 (.) 77 (.) Hourly wage ($) 14.58 (7.95) 11.86 (7.05) Occupational α 0.54 (0.27) 0.42 (0.27) Total no. of workers 1,601 Whites 1,048 Blacks 553 NOTE: AFQT, Armed Forces Qualification Test. cannot identify either their highest degree or the date at which they received their highest degree. We additionally drop 420 respondents who completed their highest degree prior to the start date of the work-history record and 305 respondents who completed their highest degree relatively late in life, because we worry that these workers already may have accumu lated substantial labor market experience that could influence employers’ beliefs about skills. We also drop 36 respondents who lack information on their first week in the labor market as well as 370 respondents whose occupational histories are relatively incomplete. In particular, we drop those who either have more than 150 weeks in which they did not work or have miss ing occupation information during the first 500 weeks following the transition to full-time work. In other words, we give individuals 500 weeks in which to accumulate 350 weeks of valid occupation information, otherwise we drop them from the sample. We additionally drop 72 respondents who ever reported an hourly wage of either over $100 or under $2 and 38 respondents with missing Armed Forces Qualification Test (AFQT) scores. After these restric tions, we are left with 1,601 individuals: 553 Black and 1,048 White. Relative to the initial sample, these individuals are young and have strong attachment to the labor force. 5 EMPIRICAL ANALYSIS 5.1 Racial Wage and Occupation Gaps Over the Career Sample summary statistics are shown in Table 1. The left (right) two columns contain mean and standard deviations for White workers’ (Black workers’) scores on the AFQT, the percentage of those workers with only a high school diploma or below, their hourly wages, the percentile rank of their occupation in terms of α, and the complexity of tasks performed there. Black workers’ socioeconomic background characteristics fall below White workers’ for both the AFQT score and completed years of education. Black workers earn around 25 percent lower hourly wages than White workers, and controlling for education and the AFQT 144 Second Quarter 2019 Federal Reserve Bank of St. Louis REVIEW Golan and Sanders Figure 1 Wage as a Function of Experience Mean Wage (in 2000 dollars) 16 14 12 10 Black 8 0 50 100 150 200 250 White 300 350 Actual Experience (in weeks) Figure 2 Occupational Complexity as a Function of Experience, By Race Mean Alpha 0.6 0.5 0.4 Black 0.3 0 50 100 150 200 250 300 White 350 Actual Experience (in weeks) Federal Reserve Bank of St. Louis REVIEW Second Quarter 2019 145 Golan and Sanders score reduces (but does not eliminate) this wage gap (see Neal and Johnson, 1996, and Lang and Modove, 2011). Less documented in the literature, Black workers tend to work in lesscomplex occupations, with the average Black worker’s occupation about 10 percentage points lower in the occupation distribution than the average White worker’s. As a reference point, note that the modal occupational category near the α = 0.42 average for Black workers is “trade, industrial, and technical teachers” and near the α = 0.54 average for White workers is “salesmen and sales clerks.” Differences between White and Black average wages are driven by both cross-sectional differences (at one point in the careers) and different growth rates of wages across the careers. Figure 1 shows wage levels of White and Black workers as a function of their actual labor market experiences; that is, we take the average wages of workers who we have observed in a job for exactly 50 weeks regardless of their age, the time since labor market entry, etc. As the figure shows, White workers start with hourly wages around $10.30 compared with around $9.00 for Black workers; and by 350 weeks of work, White workers are earning nearly $16.00, compared with slightly less than $13.00 for Black workers.5 Figure 2 summarizes a less well-known pattern in labor force outcomes: Black workers begin their careers in less-complex occupations, and even by 350 weeks, the average Black worker is not even at the same job complexity as the average White worker in his first job. As in the wage-gaps graph (see Figure 1), there are differences between races in both the levels and growth rates of occupational complexity: White workers both perform higher levels of complex tasks and move over time to higher levels of complex tasks more quickly than Black workers do. 5.2 Patterns in Occupational Mobility There are many possible explanations for the faster occupational-upgrading rates of White workers relative to Black workers. First, consider explanations within the class of racial discrimination: White and Black workers could face the same potential promotions and out side offers, but Black workers could be passed over more often due to discrimination at the screening phase; White and Black workers could get the same total number of opportunities to move, but White workers move up more at each transition; or Black workers could be more likely to be fired than White workers. An explanation that doesn’t require any racial discrim ination motive at all is simply that the growth rate of occupational complexity for workers in less-complex jobs is lower, and the lower average level of job complexity for Black workers also then explains the lower growth rate. To get some idea about the relative merits of some of these potential explanations for the racial gap in occupational complexity, we first decompose complex-task growth over time into (i) occupation and wage changes conditional on moving jobs and (ii) the probability of moving jobs. Table 2 summarizes our results. In Panels A and B of the table, we take every week-to-week occupational transition of White workers and Black workers, respectively, and divide these into those who “move down” (move to an occupation with a lower α), those who “stay” (do not change occupations or move to an occupation with an identical α), and those who “move up” (move to an occupation with a higher α). We also then split these weeks into 146 Second Quarter 2019 Federal Reserve Bank of St. Louis REVIEW Golan and Sanders Table 2 Occupation and Wage Changes by Mover Category A. Whites Weeks 1-50 51-100 101-150 151-200 201-250 251-300 310-350 –0.22 –0.21 –0.21 –0.22 –0.21 –0.21 –0.21 E[ΔW] ($) 1.38 1.11 1.38 1.52 1.37 1.33 1.15 σΔW ($) 4.28 5.60 4.52 6.65 6.91 5.50 5.67 Count 309 318 317 296 264 262 238 Move down E[Δα] Stay E[ΔW] ($) 0 0 0 0 0 0 0 σΔW ($) 0.42 0.47 0.48 0.59 0.76 0.54 0.60 44,455 46,837 47,205 47,355 47,779 47,657 47,764 E[Δα] 0.25 0.24 0.24 0.22 0.21 0.20 0.21 E[ΔW] ($) 1.98 1.04 1.28 1.11 1.74 1.92 0.56 σΔW ($) 4.80 4.91 4.87 4.70 4.68 6.38 4.67 Count 344 385 352 338 301 303 251 Weeks 1-50 51-100 101-150 151-200 201-250 251-300 310-350 –0.20 –0.21 –0.21 –0.21 –0.23 –0.22 –0.24 E[ΔW] ($) 0.58 1.10 0.30 0.18 0.68 1.19 0.81 σΔW ($) 3.20 3.59 4.63 4.14 7.23 4.57 3.82 Count 164 209 174 152 174 172 141 Count Move up B. Blacks Move down E[Δα] Stay E[ΔW] ($) 0 0 0 0 0 0 0 σΔW ($) 0.34 0.54 0.74 0.73 0.51 0.67 0.90 23,256 24,494 24,805 24,704 25,116 25,089 24,925 E[Δα] 0.23 0.22 0.23 0.21 0.24 0.21 0.22 E[ΔW] ($) 1.48 0.46 1.31 1.03 0.75 0.15 0.86 σΔW ($) 6.12 7.72 4.04 4.38 7.58 6.00 3.58 Count 220 244 196 205 155 166 134 Count Move up NOTE: For White workers, the weekly probability of a move is 1.20 percent: 0.64 percent upward and 0.56 percent downward. For Black workers, the weekly probability of a move is 1.25 percent: 0.65 percent upward and 0.60 percent downward. Federal Reserve Bank of St. Louis REVIEW Second Quarter 2019 147 Golan and Sanders bins based on total experience: workers in weeks 1-50, 51-100, etc. The “move down” category in the respective panels contains the average size of α for workers that make a downward occu pational move (to a lower α, which must by construction be negative), the average change in wages for downward occupational moves, and the standard deviation of wages for the down ward occupational moves. The same statistics are computed for those who stayed in the same occupation and those who moved to a higher occupation. Comparing Black and White workers, the two panels of Table 2 are surprisingly similar in terms of occupational moves: Given a downward occupational move, both White and Black workers move down an average of about –.20 α (20 percentage points) regardless of their career stage. Similarly, for those who move to a higher α, the average size of an occupational change is about 0.21. However, the two tables do differ in terms of the wage changes from occupational moves. Even for downward occupational moves, White workers see an average increase of around $1.30, compared with $0.80 for Black workers. For upward occupational moves, White workers see an average increase in wages of about $1.50, compared with $1.00 for Black workers. Given that the sizes of the occupational moves are the same, Black workers’ wages are seemingly less sensitive to the complex tasks in a job. The types of occupational moves Black and White workers make are similar, and the factor that explains the differences in life-cycle patterns of occupational moves is noted with Table 2: Black and White workers have almost identical weekly probabilities of switching occupations (1.20 percent for Whites and 1.25 percent for Blacks), but the relative rates of upward and down ward moves differ dramatically: Given a move, White workers are 14 percent more likely to move up versus down, while Black workers are only 8 percent more likely to move up versus down. To examine the upward versus downward mobility rates of Black and White workers in more detail, we consider transition tables across occupations, conditional on moves. Table 3 is constructed to show these rates: We take every occupational transition (and ignore “stayers”), group the “pre-move” and “post-move” occupations into 10 bins, and calculate the conditional probability of a worker being in in some post-move bin given the worker’s pre-move occu pation. So by construction, the rows sum to 1 and reading across a row gives the probabilities of moving into that bin given the row. This table helps deal with differences in the average occupations between White and Black workers: Now we can compare the occupational moves of a White worker and a Black worker who begin in the same occupation, which will help control for potential demographic and unobserved differences between them. The overall results from Table 3 show that conditional on a worker’s current occupation, the race of the worker matters strongly when predicting to which occupations he will move to next. Take a White worker and a Black worker both in the occupation “sales,” which has 0.42 α, putting them in the “0.5” row. At his next move, the White worker ends up in a morecomplex-task job 37 percent of the time. The Black worker, however, ends up at a higher-α job only 25 percent of the time.6 This pattern is repeated across almost every row: low-occupation White workers move up faster than low-occupation Black workers, and high-occupation White workers downgrade less than high-occupation Black workers. Interpreting these transition matrices within our model, given a worker is in a particular occupation in one period, there are two reasons he can be in a different occupation the next 148 Second Quarter 2019 Federal Reserve Bank of St. Louis REVIEW Golan and Sanders Table 3 Transition Matrices 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 Count 0.1 0.18 0.16 0.14 0.16 0.17 0.07 0.06 0.04 0.01 0.02 539 0.2 0.20 0.12 0.17 0.14 0.14 0.09 0.08 0.04 0.02 0.01 406 0.3 0.16 0.13 0.15 0.09 0.16 0.09 0.11 0.06 0.02 0.02 418 0.4 0.16 0.11 0.10 0.09 0.16 0.12 0.09 0.11 0.03 0.04 448 0.5 0.15 0.12 0.12 0.11 0.14 0.10 0.12 0.09 0.03 0.03 485 0.6 0.08 0.05 0.08 0.10 0.12 0.13 0.09 0.15 0.11 0.09 415 0.7 0.06 0.05 0.09 0.07 0.07 0.13 0.12 0.19 0.08 0.15 430 0.8 0.06 0.01 0.05 0.06 0.09 0.13 0.13 0.21 0.11 0.16 467 0.9 0.03 0.01 0.02 0.04 0.05 0.12 0.09 0.22 0.14 0.28 292 1.0 0.01 0.02 0.01 0.02 0.03 0.10 0.17 0.19 0.19 0.27 378 0.1 0.27 0.14 0.19 0.12 0.14 0.05 0.03 0.04 0 0.01 560 0.2 0.29 0.11 0.20 0.13 0.10 0.03 0.07 0.06 0 0.01 269 0.3 0.28 0.11 0.15 0.14 0.12 0.08 0.07 0.03 0.01 0.01 375 0.4 0.17 0.09 0.12 0.16 0.19 0.09 0.05 0.07 0.03 0.03 329 0.5 0.23 0.14 0.19 0.15 0.07 0.06 0.05 0.07 0.03 0.02 296 0.6 0.16 0.08 0.13 0.18 0.10 0.05 0.09 0.11 0.08 0.02 171 0.7 0.10 0.06 0.19 0.06 0.11 0.05 0.13 0.11 0.14 0.06 145 0.8 0.09 0.04 0.07 0.14 0.10 0.09 0.09 0.17 0.07 0.14 162 0.9 0.02 0.02 0.05 0.09 0.10 0.13 0.13 0.15 0.15 0.16 93 1.0 0.02 0.03 0.05 0.04 0.04 0.07 0.10 0.16 0.16 0.34 106 Whites Blacks NOTE: Bolded entries represent the conditional median of post-move α given pre-move α. period: First, he learned that another occupation is a better fit, and/or second, the value of experimentation has changed after learning in the first period. For example, if Black workers in high-complexity occupations are in those occupations because they are informative, whereas they are the best matches for the average White worker, you are more likely to see Black workers leave those occupations as they learn more and have less incentive to experiment further. Discrimination adds a further level of complexity because Black workers may be restricted from entering their preferred occupations, so the incentives to learn may be the same but they may be being restricted from finding their best match as quickly as White workers. Using a structural life-cycle model, we further investigate the roles of productivity, learning, and discrimination in Golan et al. (2017). Federal Reserve Bank of St. Louis REVIEW Second Quarter 2019 149 Golan and Sanders Table 4 Table 5 The Relationship Between Wage Growth and Initial Job Assignment Determinants of an Initial Job Assignment (1) Blacks (2) Whites Initial wage 0.596*** (0.044) 0.674*** (0.038) Initial α Initial α 0.880 (0.747) 0.358 (0.634) Years of schooling (0 to 7) 0.057*** (0.004) 0.572*** (0.111) Actual experience (in years) 0.508*** (0.149) 0.954*** (0.127) Age 0.014*** (0.002) 0.168*** (0.055) Initial wage × actual experience –0.032** (0.016) –0.045*** (0.015) Black –0.026* (0.015) –0.312 (0.324) Initial α × actual experience 1.267*** (0.264) 1.130*** (0.190) AFQT 0.068*** (0.008) 0.440*** (0.175) AFQT 1.532*** (0.286) 1.330*** (0.185) Constant –0.128 (0.049) 3.340*** (1.070) Constant 2.320*** (0.346) 0.805*** (0.410) Dependent variable: wages Observations R2 181,424 343,620 0.310 0.347 Dependent variable: wages (1) Initial α (2) Initial wage 3.477*** (0.555) Observations 1,593 1,593 R2 0.376 0.189 NOTE: Robust standard errors are in parentheses. *** p < 0.01 and * p < 0.1. NOTE: Robust standard errors are in parentheses. *** p < 0.01 and ** p < 0.05. 5.3 Initial Job Assignment In the context of a learning model, the initial job assignment has an outsized effect on the career path. If workers start their careers in a job or occupation where they learn nothing about themselves, their wages would grow only as a result of human capital growth or other factors. On the other hand, workers beginning in very informative jobs will see their wages grow on average as they become better matched. If more-complex jobs also reveal more about a worker’s innate ability, there should be a positive relationship between the measure of com plexity in a worker’s initial job and future wage growth. In Table 4 we consider the relationship between characteristics of a worker’s first job and future wage growth. The dependent variable is the pooled cross section of wages in all jobs after the first job, while the independent variables include both initial wages and initial α, actual labor market experience, and interactions between initial job characteristics and actual expe rience. The model has no particular predictions for the relationship between the initial α and wages, but it predicts that the coefficient on the interaction of the initial α and experience will be positive, which is what the data show. To interpret the estimates, relative to the wage growth of a worker at the Black-worker sample average of α = 0.42, a worker at the White-worker 150 Second Quarter 2019 Federal Reserve Bank of St. Louis REVIEW Golan and Sanders sample average of α = 0.54 would see faster average wage growth of about $0.13 per year. That is, there is around a 1 percent difference in the average yearly rates of wage growth that can be attributed to racial differences in occupational characteristics. Differences in information about one’s own ability can be reflected as well in the choice of initial occupation: Workers with more to learn about themselves would tend toward occu pations that provide more information. In Table 5 we document the relationship between worker characteristics and the complexity measure α and the wage of the worker’s first job. In the first column, the dependent variable is the complexity of the initial occupation, and we condition on education, age, race, and AFQT score. Even after the controls for ability, provided by controlling for education and the AFQT score, we find that Black workers start their careers in about 2 percentage point less-complex-task occupations. While this may seem small, it can be compared with the effect of lowering a worker’s AFQT score by 1/3 of a stan dard deviation, which is a substantial drop. On the other hand, in the second column of Table 5 we find little evidence of a direct effect of race on the wage of the initial job once α and the ability measures have been controlled for. The finding that Black workers end up in less-complex jobs than comparable White workers all while earning effectively the same as those White workers can be interpreted in multiple ways consistent with our model of worker learning. For one, say discrimination is the reason that Black workers don’t have access to more-complex jobs. This will reduce the amount of learning Black workers have about themselves relative to what White workers— in more-informative occupations—have about themselves. The long-run effect of the initial discrimination would be that Black workers’ wage growth is lower relative to White workers’ wage growth than it would otherwise be if both races were assigned the same α jobs initially. 6 CONCLUSION This article documents racial gaps in occupational assignment, turnover, and wages. We use a learning model in which there is uncertainty about skills early in a worker’s career to interpret the empirical evidence on the economic forces behind occupation mobility and the Black-White pay gap and its evolution over a worker’s career. We first document that both wage and occupation racial gaps increase with worker expe rience. Occupational turnover is on average associated with an increase in pay. This is true even for transitions from high- to low-skill-requirement occupations. Our learning model is consistent with this pattern because it implies that workers sort into jobs that better match their skills. This pattern holds for both Black and White individuals. Furthermore, we find no significant differences in the number of occupational moves for Black and White workers. However, Black workers are more likely to transition into occupations with lower skill require ments than White workers and therefore experience smaller wage growth. This pattern is also consistent with the widening gap in occupational skill requirements between Black and White workers. To further explore the Black-White wage gaps, we focus on differences in initial occupa tional assignments. The learning model presented predicts that workers assigned initially to Federal Reserve Bank of St. Louis REVIEW Second Quarter 2019 151 Golan and Sanders more-complex jobs will experience faster wage growth than those assigned initially to less- complex jobs, because those in more-complex jobs will have more ability to learn about their best occupation. Again, our analysis demonstrates that this pattern empirically holds for both Black and White individuals. However, Black workers are assigned initially to jobs with lower complexity partly because they have different demographic characteristics when they enter the labor market, such as lower educational attainment and lower AFQT scores. Nevertheless, we show that AFQT scores and educational attainment do not explain the entire gap in initial occupational assignments. This finding can be interpreted both as differences in beliefs and learning, but can at the same time be consistent with discriminatory hiring practices. For future research, we suggest that further analyses should focus on understanding the interaction between discrimination and learning, since restricting workers from entering their preferred occupation as young workers can have lifetime effects on earnings and potentially be an important source of racial inequality. n NOTES 1 2016 Current Population Survey, Bureau of Labor Statistics. 2 This index was created in Antonovics and Golan (2012). 3 For details and examples, see Antonovics and Golan (2012). 4 In this tabulation, after making the sample selection rules discussed below, we treat individuals with missing occupation information as being out of the labor force. 5 An alternative specification of these graphs uses worker age on the horizontal axis, which makes the gaps marginally larger and increasing over time, since differences in actual experience between White and Black workers grow over time. See Golan et al. (2017) for a discussion of the effects of labor market participation rates on wage and occupation gaps. 6 This calculation is done by summing all elements of the row to the right of the “0.5” column. REFERENCES Altonji, J.G. and Blank, R.M. “Race and Gender in the Labor Market,” in O. Ashenfelter & D. Card, eds., Handbook of Labor Economics. Edition 1, Volume 3. Elsevier, 1999, pp. 3143-259. Altonji, J.G. and Pierret, C.R. “Employer Learning and Statistical Discrimination.” Quarterly Journal of Economics, 2001, 116(1), pp. 313-50; https://doi.org/10.1162/003355301556329. Antecol, H. and Bedard, K. “The Racial Wage Gap the Importance of Labor Force Attachment Differences Across Black, Mexican, and White Men.” Journal of Human Resources, 2004, 39(2), pp. 564-83; https://doi.org/10.2307/3559027. Antonovics, K. and Golan, L. “Experimentation and Job Choice.” Journal of Labor Economics, 2012, 30(2), pp. 333-66; https://doi.org/10.1086/663356. Autor, D.H.; Levy, F. and Murnane, R.J. “The Skill Content of Recent Technological Change: An Empirical Exploration.” Quarterly Journal of Economics, 2003, 118(4), pp. 1279-333; https://doi.org/10.1162/003355303322552801. Brown, C. “Black-White Earnings Ratios Since the Civil Rights Act of 1964: The Importance of Labor Market Dropouts.” Quarterly Journal of Economics, 1984, 99(1), pp. 31-44; https://doi.org/10.2307/1885719. Cameron, S. V. and Heckman, J.J. “The Dynamics of Educational Attainment for Black, Hispanic, and White Males.” Journal of Political Economy, 2001, 109(3), pp. 455-99; https://doi.org/10.1086/321014. 152 Second Quarter 2019 Federal Reserve Bank of St. Louis REVIEW Golan and Sanders Carneiro, P.; Heckman, J.J. and Masterov, D.V. “Labor Market Discrimination and Racial Differences in Premarket Factors.” Journal of Law and Economics, 2005, 48(1), p. 1; https://doi.org/10.1086/426878. Chandra, A. “Labor-Market Dropouts and the Racial Wage Gap: 1940-1990.” American Economic Review, 2000, 90(2), pp. 333-38; https://doi.org/10.1257/aer.90.2.333. Eckstein, Z. and Wolpin, K.I. “Estimating the Effect of Racial Discrimination on First Job Wage Offers.” Review of Economics and Statistics, 1999, 81(3), pp. 384-92; https://doi.org/10.1162/003465399558319. Farber, H. and Gibbons, R. “Learning and Wage Dynamics.” Quarterly Journal of Economics, 1996, 111(4), pp. 1007-47; https://doi.org/10.2307/2946706. Fryer, R.G. “Racial Inequality in the 21st Century: The Declining Significance of Discrimination.” Handbook of Labor Economics, 2011, 4, pp. 855-971; https://doi.org/10.1016/S0169-7218(11)02408-7. Gibbons, R.; Katz, L.F.; Lemieux, T. and Parent, D. “Comparative Advantage, Learning, and Sectoral Wage Determination.” Journal of Labor Economics, 2005, 23(4), pp. 681-724; https://doi.org/10.1086/491606. Gibbons, R. and Waldman, M. “A Theory of Wage and Promotion Dynamics Inside Firms.” Quarterly Journal of Economics, 1999, 114(4), pp. 1321-58; https://doi.org/10.1162/003355399556287. Golan, L.; Sanders, C. and James, J. “The Black-White Wage Gap Over the Life Cycle.” Working Paper, 2017. James, J. “Learning and Occupational Sorting.” Working paper, California State Polytechnic, 2012. Jovanovic, B. “Job Matching and the Theory of Turnover.” Journal of Political Economy, 1979, 87(5, Part 1), pp. 972-90; https://doi.org/10.1086/260808. Juhn, C. “Labor Market Dropouts and Trends in the Wages of Black and White Men.” Industrial and Labor Relations Review, 2003, 56(4), pp. 643-62; https://doi.org/10.1177/001979390305600406. Lang, K. and Manove, M. “Education and Labor Market Discrimination.” American Economic Review, 2011, 101(4), pp. 1467-96; https://doi.org/10.1257/aer.101.4.1467. Miller, R.A. “Job Matching and Occupational Choice.” Journal of Political Economy, 1984, 92(6), pp. 1086-120; https://doi.org/10.1086/261276. Neal, D. “The Complexity of Job Mobility Among Young Men.” Journal of Labor Economics, 1999, 17(2), pp. 237-61; https://doi.org/10.1086/209919. Neal, D.A. and Johnson, W.R. “The Role of Premarket Factors in Black-White Wage Differences.” Journal of Political Economy, 1996, 104(5), pp. 869-95; https://doi.org/10.1086/262045. Oettinger, G.S. “Statistical Discrimination and the Early Career Evolution of the Black-White Wage Gap.” Journal of Labor Economics, 1996, 14(1), pp. 52-78; https://doi.org/10.1086/209803. Papageorgiou, T. “Learning Your Comparative Advantages.” Review of Economic Studies, 2014, 81(3), pp. 1263-95; https://doi.org/10.1093/restud/rdt048. Pastorino, E. “Learning in Labor Markets and Job Mobility.” Unpublished manuscript, Department of Economics, University of Iowa, 2009. Sanders, C. “Skill Accumulation, Skill Uncertainty, and Occupational Choice.” Working paper, Washington University in St. Louis, 2017. Sanders, C. and Taber, C. “Life-Cycle Wage Growth and Heterogeneous Human Capital.” Annual Review of Economics, 2012, 4(1), pp. 399-425; https://doi.org/10.1146/annurev-economics-080511-111011. Western, B. and Pettit, B. “Black-White Wage Inequality, Employment Rates, and Incarceration.” American Journal of Sociology, 2005, 111(2), pp. 553-78; https://doi.org/10.1086/432780. Federal Reserve Bank of St. Louis REVIEW Second Quarter 2019 153 154 Second Quarter 2019 Federal Reserve Bank of St. Louis REVIEW