A Unified Measure of Fed Monetary Policy Shocks

Finance and Economics Discussion Series Divisions of Research & Statistics and Monetary Affairs Federal Reserve Board, Washington, D.C. A Unified Measure of Fed Monetary Policy Shocks Chunya Bu, John Rogers, and Wenbin Wu 2019-043 Please cite this paper as: Bu, Chunya, John Rogers, and Wenbin Wu (2019). “A Unified Measure of Fed Monetary Policy Shocks,” Finance and Economics Discussion Series 2019-043. Washington: Board of Governors of the Federal Reserve System, https://doi.org/10.17016/FEDS.2019.043. NOTE: Staff working papers in the Finance and Economics Discussion Series (FEDS) are preliminary materials circulated to stimulate discussion and critical comment. The analysis and conclusions set forth are those of the authors and do not indicate concurrence by other members of the research staff or the Board of Governors. References in publications to the Finance and Economics Discussion Series (other than acknowledgement) should be cleared with the author(s) to protect the tentative character of these papers.

A Unified Measure of Fed Monetary Policy Shocks∗ Chunya Bu (U of Rochester), John Rogers (Federal Reserve Board), Wenbin Wu (Fudan U) May 2019 Abstract Identification of Fed monetary policy shocks is complex, in light of the distinct policymaking regimesbefore, during, andaftertheZLB periodofDecember 2008 toDecember2015. Wedevelop a heteroscedasticity-based partial least squares approach, combined with Fama-MacBeth stylecross-sectionregressions,toidentifyaUSmonetarypolicyshockseriesthatusefullybridges periodsofconventionalandunconventionalpolicymakingandiseffectivelydevoidofthecentral bank information effect. Our series has moderately high correlation with well-known shocks in the literature, but has crucially important differences. Following conventional tests, we find scantevidenceoftheinformationeffectinourmeasure. Weattributethesourceofthesedifferent findingstooureconometricprocedureandouruseofthefullmaturityspectrumofinterestrate instrumentsinconstructingourmeasure. Wethenpresentevidenceconfirminganhypothesisin the literature that the information effect can lead to the result that shocks to monetary policy have transmission effects with signs that differ from traditional theory. We find that shocks to seriesthataredevoidof(embody)theinformationeffectdisplayconventionally-signed(perverse) impulse responses of output and inflation. This provides evidence of first-order importance to staff at central banks undertaking quantitative theoretical modeling of the effects of monetary policy. ∗WethankfortheircommentsDarioCaldara,EdHerbst,ThomasLaubachandotherFederalReserveBoardMA workshop participants, Eric Swanson, Jon Steinsson, Jonathan Wright, James Hamilton, Shang-Jin Wei, Jun Qian, Yi Huang, Marek Jarocinski, Cynthia Wu, and Xu Zhang. The views expressed here are solely our own and should not be interpreted as reflecting the views of the Board of Governors of the Federal Reserve System or of any other personassociatedwiththeFederalReserveSystem. 1

1 Introduction The adoption of unconventional monetary policy tools by the Federal Reserve in the wake of the Great Financial Crisis brought policymaking into new territory and rekindled challenges for research measuring monetary policy shocks and estimating their effects. Much of the new research built on influential work that pre-dated the crisis and used bond market data at daily or intra-daily frequencies (Kuttner (2001), Cochrane and Piazzesi (2002), Rigobon and Sack (2003), Gurkaynak, Sack, and Swanson (2005)). In much of this new work, monetary policy surprises are measured as the change in interest rate futures prices in narrow windows around FOMC announcements (for examples, see Wright (2012), Gertler and Karadi (2015), Nakamura and Steinsson (2018), Rogers, Scotti and Wright (2018), Swanson (2018), and Jarocinski and Karadi (2018)). This represented a departure from traditional approaches to measurement and identification such as the use of orthogonalized innovations to the Federal Funds rate in recursive VARs (Christiano, Eichenbaum and Evans (1996)) or the narrative approach of Romer and Romer (2004). An advantage of the methods developed in the recent papers is that, under certain assumptions, the resulting shock series captures both conventional policymaking, through shocks to the target Fed Funds rate, as well as unconventional policymaking, as reflected in identified shocks to forward guidance (FG) and largescale asset purchases (LSAPs). The use of narrow time windows around FOMC announcements enhances identification, it is argued, because no other economic news is (routinely) released. The literature on the central bank private information effect has called into question this assertion, however (Romer and Romer (2000), Campbell et al. (2012 and 2016), Nakamura and Steinsson (2018), Miranda-Agrippino (2016), and Jarocinski and Karadi (2018)). Under this view, the central bank reveals in its meeting day announcements not only pure monetary policy “news” but also its private information on the state of the economy, its own preferences, or the model it uses to analyze the economy. This in turn causes the private sector to change its outlook for macroeconomic developments. Thus, conventionally-measured monetary policy surprises may be correlatedwithdevelopmentsinnon-monetarypolicyeconomicfundamentals,evenintightwindows around central bank announcements. Further confounding identification, these studies document a tendency for private sector expectations (and possibly stock prices) to go in the ”wrong“ direction. That is, following a contractionary monetary policy surprise, expectations of future GDP growth (or stock prices) rise. The empirical presence of the Fed information effect calls into question the central assumption that these surprises are appropriate to identify (pure) monetary policy shocks. 1

Thefocusofmostofthesepapers,especiallytheearlyones,isonthetransmissiontofinancial markets and expectations. For example, Nakamura and Steinsson (2018) document the effects of their policy news shock on the real interest rate, expected inflation, and expected output growth. Swanson (2018) finds that both forward guidance and LSAP shocks have highly statistically significant effects on a wide variety of assets: Treasuries, corporate bonds, stocks, exchange rates, and options-implied interest rate uncertainty. He also examines the persistence of these shocks, compares magnitudes before and during the ZLB period, and concludes with an appeal to examine the transmission to macroeconomic variables.1 In this paper, we develop a heteroscedasticity-based, partial least squares (PLS) approach to identify shocks to US monetary policy, compare our measure to those in the literature, and estimate the macroeconomic transmission effects of shocks.2 The general idea behind construction ofourmeasureistouseFamaandMacBeth(1973)two-stepregressionstoestimatetheunobservable monetarypolicyshock. Thisworksinitiallythroughthesensitivityof“outcomevariables”toFOMC announcements. Specifically,inthefirststepweruntime-seriesregressionstoestimatethesensitivity of interest rates at different maturities to FOMC announcements. This is equivalent to the asset beta in the original Fama-MacBeth method. In order to filter out non-monetary policy news, we employtheheteroskedasticity-basedestimatorofRigobonandSack(2003,2004),implementedwith instrumentalvariables(IV),intothisstep. Inthesecondstep, weregressalloutcomevariablesonto the corresponding estimated sensitivity index from step one, for each time t. In this way, we derive the new monetary policy shock as the series of estimated coefficients from the Fama-MacBeth style second step regressions. The application of this procedure to estimating monetary policy shocks is novel as far as we are aware,3 and has non-trivial effects on the resulting measure. Our approach to estimating a monetary policy shock series has a couple of conceptual advantages. One is simplicity. Our procedure has very mild data requirements and is easy to implement econometrically. Compared to the path-breaking work of Romer and Romer (2004), implementing our method involves no need to parse through Federal Reserve transcripts and forecasts. Nor does it require the use of intra-daily data, which is costly to acquire and can have spotty coverage, as in much of the newest research. Thus, a second and related advantage of our method is its greater 1“Goingforward,therearemanyimportantissuesthatcallforfurtherexploration. Firstandforemost,estimating the effects of forward guidance and LSAPson macroeconomic variables such as the unemploymentrate should be a top priority for future research. After all, the FOMC’s stated goal in pursuing these unconventional policies was to boosttheeconomy(pg. 37).” 2WuandXia(2016)andJarocinskiandKaradi(2018)alsofocusontransmissiontomacroeconomicvariables,as discussedbelow. 3SeeWold(1966,1975)andKellyandPruitt(2013,2015)forapplicationstoequityreturns. 2

applicability. Our approach can be implemented over longer sample periods and for more countries, for which data requirements often render the process untenable. To see this, we use the procedure to construct a monetary policy shock series for the European Central Bank (ECB) as well. This series has properties that are similar to the Fed series, including absence of the information effect. Wealsodemonstratetheimportanceofourseriesinpractice. Tobegin,weshowthatourshock series has moderately high correlation with the Nakamura and Steinsson (2018), Swanson (2018), and Jarocinski and Karadi (2108) monetary policy shocks. Focusing on the period surrounding liftoff in December 2015, we show that our shock series reflects the strong forward guidance delivered at the October 2015 FOMC meeting, and thus implies that a contractionary monetary policy shock took place in the meeting before the actual interest rate hike, consistent with existing measures. In addition, we show that both the short end and long end of the yield curve respond less to our shock than do medium-horizon maturities like 2-year and 5-year rates, similar to the Swanson forward guidance shock. Moreover, there are many days in which the stock market co-moves positively with our series, consistent with the Jarocinski-Karadi observations that are the focus of their paper. Similarities with existing measures notwithstanding, we show that there are important differences, beginning with evidence on the Fed information effect. Our investigation includes both testing for the presence of the information effect in the monetary policy shock series and estimating impulse responses from shocks that are purged of the estimated information effect. We follow twoprominentapproachesintheliterature: theNakamura-Steinsson(2018)expectations-basedtest and Jarocinski-Karadi (2018) “indirect” test. Using the Nakamura-Steinsson test, we do not find a statistically significant information effect in our new shock series, while we confirm its presence in theseriesestimatedbyNakamuraandSteinsson(2018)andSwanson(2018). JarocinskiandKaradi (2018) examine the high-frequency co-movement of interest rates and stock prices around FOMC announcements. Monetary policy announcements that lead to positive co-movement (within the day) are defined to be those that reveal central bank private information. Using our new measure, and even confining our analysis to observations that occur on days with positive co-movement between stock prices and interest rates, we find no evidence of an information effect in the sense of Nakamura and Steinsson (2018). We reconcile the different findings between our monetary policy shock series and existing measures by pointing to important differences in the construction of the measures. A simple “encompassing” analysis shows that differences in the econometric approach and data used to identify the monetary policy shock series both play a key role. Whereas Nakamura and Steinsson construct 3

theirshockseriesfromshort-terminterestratesupto2years,andJarocinskiandKaradi(2018)use onlyathree-monthrate,weusethewholeyieldcurve. Inclusionoflongerterminterestratesisvery important, because we find that longer term interest rates display less evidence of an information effect. Our PLS approach extracts a common component from the whole yield curve, and assigns more weight to interest rates that have greater correlation with the policy indicator (the five-year treasury rate in the benchmark case). Because the Fed information effect is essentially non-existent in maturities of five years and longer, the common factor we extract also contains less of an informationeffect.4 JarocinskiandKaradiconstructtheirproxyformonetarypolicysurprisesalsousing onlyashortrate, thethree-monthFedFundsfuturesrate(FF3). Usingtheirdata, wefindevidence of the Fed information effect, in the sense of Nakamura-Steinsson, on (JK) information effect days, butasnotedabove,wedonotfinditinourmeasure,evenondaysofpositiveco-movementbetween stock prices and our series. Finally, we present evidence confirming an hypothesis in the literature that the information effect can give rise to monetary policy shocks having transmission effects with opposite signs from thosepredictedbytraditionaltheory. Usingourseries,wefindthatapositivemonetarypolicyshock leads to significantly negative effects on output and prices, consistent with standard theory. This is true in the full sample and for sub-samples before and during the ZLB. We also find conventional signs using only those of our shocks that occur on Jarocinski-Karadi (JK) information effect days or, equally, only those that occur on non-information effect days. On the other hand, shocks to the alternative measures that embody the information effect produce non-traditional signs. This is especiallyevidentduringtheZLBperiodwhereoutputrises inresponsetoapositiveNSorSwanson monetarypolicyshock. Similarly, whenweusetheJarocinski-Karadiproxyforthemonetarypolicy shock, we replicate their finding that an announcement-day interest rate increase accompanied by a stockpriceincreaseleadstosignificantlyhigheroutputandpricelevel,andimprovementinfinancial conditions. However, when we replace their measure with the BRW monetary policy shock we find onlyminordifferencesintheimpulseresponsesoninformationeffectdaysandnon-informationeffect days: with our measure, the responses are always of the conventional sign. Theinformationeffectisanissueoffirst-orderconcerntostaffattheFederalReserveandother central banks. Should staff models be constructed to feature the information effect associated with 4The yield curve is also used as a function in contemporaneous work by Inoue and Rossi (2018). Like us, they proposeanewwaytoidentifymonetarypolicyshocks,inwhattheyrefertoas“functionalshocks”,andthenestimate transmission effects during periods of conventional and unconventional policy. We differ in several important ways: (1) we use a much simpler method involving only linear regressions; and (2) we focus on the information effects of identified shocks while Inoue and Rossi focus more on econometric issues. Conclusions concerning the transmission effectofshocksareconsistent,however. 4

monetary policy announcements? If so, how, what are the appropriate building blocks? Should the impulseresponsesthatthestaff’squantitativemodelsattempttomatchbeofthesignspredictedby traditionalmonetarytheory,oroftheunconventionalsignsconsistentwiththeevidenceininfluential recent papers on the information effect? In the next section, we describe our econometric approach and the data. In section 2, we display our new series and compare it to existing measures in the literature. In section 3, we test for the presence of the information effect in our Fed monetary policy shock series and alternatives, and reconcile the different findings. In section 4, we confirm the hypothesis in the literature that the information effect can give rise to impulse responses that have signs opposite to those predicted by conventional theory. Section 5 concludes: we provide a US monetary policy shock series that is easy to estimate, that usefully bridges periods of conventional and unconventional policymaking periods, is devoid of the information effect, and that helps substantiate an hypothesis concerning transmission effects to output and inflation. 2 A New Monetary Policy Shock 2.1 Methodology: Fama-—MacBeth Meets Rigobon-—Sack We assume that the true monetary policy shock e is unobservable. We further assume that t the (observable) changes in Treasury yields around FOMC announcement days are driven by a monetarypolicyshocke andnonmonetarypolicyshockη . Ourobjectiveistoestimatetheformer. t t We normalize the unobserved monetary policy shock to have a one to one relationship with the 5 year Treasury yield,5 ∆R =α +e +η (1) 5,t 0 t t where ∆R is the change in our policy indicator – the 5 year Treasury yield, α is a constant, e 5,t 0 t is the monetary policy shock, and η denotes factors unrelated to monetary policy news.6 We allow t η to include idiosyncratic noise specific to the 5 year interest rate as well as noise that is common t 5ThisismotivatedbythenotionthatFedpolicyaimstoaffectinterestratesataboutthishorizon,anassumption that became more compelling during the ZLB period and is used elsewhere in the literature. We examine (and confirm)robustnessofthischoiceofmonetarypolicyindicatorto2-yearand10-yearrates. 6This includes factors associated with the Fed information effect, e.g., the market interpreting an FOMC policy announcement as (also) revealing private information it has on the state of the economy, its own preferences for inflationversusoutputstabilization,etc. ThefactthatFederalReserveBoardstaffconstructtheindexofIndustrial Production is one potential source of such private information. Fed staff are situated particularly auspiciously, for example, to ascertain and report to the FOMC in private how noisy is a particular release of the IP series. See NakamuraandSteinsson(2018)forfurtherdiscussionof“backgroundnoise”. 5

to the entire yield curve. Our Fama-MacBeth two-step procedure extracts monetary policy shocks e from the common t component of the outcome variables ∆R . In the first step, we estimate the sensitivity of each i,t outcome variable to monetary policy via time-series regressions. We assume that the outcome of monetary policy decisions is reflected in the movements of zero-coupon yields with maturities of 1 yearto30years. Aswedemonstrateinsection3,ouruseofthefullmaturitystructureisimportant, most notably in producing a shock series that is devoid of the information effect. These outcome variables are also affected by background noise: ∆R =α +β e +(cid:15) (2) i,t i i t i,t where ∆R is the change in the zero-coupon yield with i-year maturity and (cid:15) is the idiosyni,t i,t cratic noise for ∆R . We assume the error term (cid:15) and the unobserved monetary shock e are i,t i,t t uncorrelated. Due to our normalization, we can rewrite (2) as, ∆R =θ +β ∆R +ξ (3) i,t i i 5,t i,t where ξ = −β η +(cid:15) and θ is a constant. Recalling that η is the error term in the policy i,t i t i,t i t indicator (see equation (1)), we see that the regressor ∆R is correlated with the error term ξ 5,t i,t due to the component “−β η ”. The OLS estimate of β is thus biased. i t i Therefore, although this step could be done using OLS and high-frequency FOMC announcementdaydata,weinsteadusetheheteroskedasticity-basedestimatorofRigobon(2003)andRigobon and Sack (2003, 2004). As demonstrated formally in Appendix A1, β in (2) can be consistently i estimated using instrumental variables (IV). Rewrite (2) as: [∆R ]=α +β [∆R ]+µ i=1,2,···,30 (4) i,t i i 5,t i,t where the independent variable [∆R ] = (∆R ,∆R∗ )(cid:48), ∆R is the 1-day movement in the 5,t 5,t 5,t 5,t policy indicator around the FOMC announcement, and ∆R∗ is the movement with the same event 5,t window length but one week before FOMC announcement day.7 The event window for [∆R ] i,t corresponds to that of [∆R ], β measures the sensitivity of ∆R to monetary policy shocks, and 5,t i i,t µ istheidiosyncraticnoiseassociatedwithionly. Weexploitthefactthatβ canbeestimatedusing i,t i 7The choice of one week follows Nakamura and Steinsson (2018). We examine (and confirm) robustness to two daysbeforetheFOMCannouncementday,whichisakintotheRigobon(2003)practice. 6

an instrumental variable ∆RIV = (∆R ,−∆R∗ )(cid:48) for the independent variable. The underlying t 5,t 5,t assumption is that, on days of FOMC meetings, the variance of the ’true’ monetary policy shock increases while that of non-monetary policy news remains unchanged. In the estimation, we use a 1-daywindow,capturingpolicysurprisesbetweenFOMCannouncementday(end)andtheprevious day (end). Because the Fed released no public statements about monetary policy decisions until 1994, we begin estimation of our shock series then. The second step of our approach, by analogy to Fama and MacBeth, is to recover the aligned monetary policy shock from cross-sectional regressions of ∆R on the estimated sensitivity index i,t βˆ for each time t, i ∆R =α +ealignedβˆ +v t=1,2,···,T (5) i,t i t i i,t where ealigned is the coefficient of interest. This series of T estimated coefficients from the second t step regressions is the BRW monetary policy shock series. 2.2 The Data We collect data on the monetary policy indicator from the Federal Reserve Board public website. As noted above, we examine 2-year, 5-year, and 10-year Treasury rates, with 5-year as benchmark. We also use data on estimated term premia, from Adrian, Crump, and Moench (2013), which are available through the New York Fed website https://www.newyorkfed.org /research/data indicators/term premia.html. Thepolicyoutcomevariables,thezerocouponyieldswith maturitiesof1to30years,areestimatedbyGurkaynak,Sack,andSwanson(2005),andavailableat https://www.federalreserve.gov/pubs/ feds/2006/200628/200628abs.html. To estimate impulse responses,weusemonthlyindustrialproductionandCPI,bothtakenfromhttps://fred.stlouisfed.org, the core commodity price index from Thompson Reuters, and the excess bond premium from Gilchrist and Zakrajsek (2012). 2.3 BRW Monetary Policy Shock Series WedisplayourmonetarypolicyshockseriesinFigure1. Therearesizablemovementsbefore, during,andaftertheZLBperiod. TheannouncementsofQE1,QE2,andQE3,whicharemarkedby navy lines, all generate large expansionary monetary policy shocks. Monetary policy shocks during OperationTwist,denotedbytheorangelines,areinsteadcontractionary. Wemarkwiththeblueline the FOMC meeting in October 2015, the meeting preceding lift-off in December. Zooming in on the 7

last three meetings of 2015, our shock series takes the values -0.080 (September), 0.115 (October), and 0.038 (December). Expectations of a lift-off had been growing throughout the summer and heading into the October meeting. For a variety of reasons, including turmoil in global equity markets, the FOMC decided to keep the target Fed Funds rate unchanged at that meeting but sent a clear signal of a likely rise in December 2015.8 Our measure indicates that this forward guidance gaverisetoasizablecontractionarymonetarypolicyshockinOctober2015, onemeetingbeforethe actual rate increase. This is consonant with the dynamic pattern of alternative measures that use intra-dailydataandestimateseparatecomponentsofFedmonetarypolicyshocks. Forexample,the corresponding values of the policy news shock of Nakamura and Steinsson (2018) are (-0.042, 0.032, 0.016), the forward guidance surprise in Rogers, Scotti, and Wright (2018) are (-0.09, 0.09, 0.03), and in Swanson (2018) (-1.50, 1.67, NA).9 We analyze this further in the next section. 2.3.1 Comparison with Shocks in the Literature MovingbeyondtheissueofplausibilityofspecificobservationsaroundliftoffandQEannouncements, we provide in Table 1 a comprehensive comparison of our shock series with well-known measures in the literature: Kuttner (2001), Romer and Romer (2004), Nakamura and Steinsson (2018), Swanson (2018), and Jarocinski and Karadi (2018). The updated R&R shock series, constructed using their same narrative method, runs through the end of 2007. Kuttner (2001) shocks are extracted from changes in Federal Funds futures rates in 30-minute windows around FOMC announcements. Nakamura and Steinsson also examine high-frequency movements around FOMC announcements. Their monetary policy shock is the first principal component of changes in the current month Federal Funds futures rate, theFederal Funds futures rateimmediately following the next FOMC meeting, and two, three and four quarter ahead euro dollar futures in the 30-minute eventwindow.10 JarocinskiandKaradi(2018)usethree-monthFedFundsfutures(FF3)changesin 30-minute windows around FOMC announcements, while Swanson (2018) separately identifies the effects of forward guidance, large-scale asset purchases, and target Federal Funds rate shocks, also using principal components.11 In Table 1 we present the correlation between our measure and the alternatives (figures are 8AsheadlinedintheFinancialTimesonOctober29,2015: “FederalReservedropswarningsonglobalriskstoUS economy: CentralbankhawkishstatementincreaseschancesofDecemberriseininterestrates.” 9Magnitudesdifferduetodifferentnormalizationchoices,especiallybySwanson,whoseseriesendswithliftoff. 10We obtain these shocks from Nakamura and Steinsson (2018) through 2014m3 (their sample period) and then follow their procedures to update to the present. For this exercise and all of our work using intra-daily data, we obtainthedatafromthe“EventStudy”databasemaintainedbyFederalReserveBoardstaff. 11Rogers, Scotti, and Wright (2018) implement an approach similar to Swanson (2018) in computing their three separatecomponentsofFedpolicyshocks. TheseriesareveryhighlycorrelatedwiththoseofSwanson,around0.96. 8

availableintheonlineAppendix). Asseenincolumn1,overthefullsample,ourshockisreasonably well correlated (around 0.5) with the NS and Swanson shocks, which themselves are relatively large before and during the ZLB. The next two columns decompose the comparison into sub-periods, before and during the ZLB. Before the ZLB, our series is correlated with NS, JK, and the Swanson FG shock at around 0.6. In the final column, we present correlations during the ZLB. The largest correlation, at 0.57, is with the Swanson FG shock. In Figure A7, we display plots of our shock seriesagainstthealternatives. Consistentwiththecorrelationsabove,priorto2008ourshockseries exhibits a similar pattern to the NS, Kuttner, and R&R shocks. After 2008, the alternative series are quite small given that the Fed Funds rate is at zero during the ZLB. In contrast, our new shock series exhibits relatively large movements, consistent with Fed monetary policy being about more than the target FFR. Our shock series is more similar to the FG and LSAP shocks of Swanson. 2.3.2 BRW Series Construction Robustness We examine several modifications to the construction of the baseline BRW shock series. As previewedabove,weconsideralternativenormalizationsofthemonetarypolicyshockseriestoeither the2-yearorthe10-yearTreasuryrateinsteadofthe5-year. AsseeninColumns1and2ofTable2, thecorrelationwithourbaselineshockseriesisabove0.97. Thusourapproachisrobusttodifferent choices of the monetary policy indicator. Our second check is to extend our monetary policy shock seriesbackwardto1969. Before1994,therewasnopublicannouncementofFOMCdecisions. Thus, for this earlier period, we use the 1-day policy window between the FOMC announcement day and the following day to capture the policy effect. From the third column of Table 2, we see that the correlation with our BRW shock is over 98%.12 Our third modification is to use only zero-coupon yields with 1-, 2-, 5-, 10-, and 30-year maturities, the more commonly-used series, as the outcome variables. The correlation with the baseline shock series, as shown in column 4 of Table 2, is over 0.95. Fourth, we assess robustness to leaving out the QE1 announcement in the alignment process. This announcement, in March 2009, was a sufficiently big event occurring at a time when financial markets were so sluggish that the market response might not represent a typical effect of monetary policy. The new shock series without QE1 is again highly correlated with our baseline series (Column 5). Next, we extend our 12Onefeatureofourmethodologyistheneedtocheckthestabilityofthesensitivitiesofinterestrateswithdifferent maturitiestomonetarypolicyshocks. Here,wedotherollingsampletestforeachperiodof15years,expandingthe samplesizeto1969-2017. Whenweusedifferentmonetarypolicyindicatorsof1-,2-,5-and10-yearTreasuryRates, thecoefficientsarenotcompletelystableuntilearly1990(figureavailableintheonlineAppendix). Thuswestartthe sample in 1994, when the Fed first released a statement about FOMC policy decisions. The sensitivity index is flat after1994,indicatingstabilityofouralignmentprocess. 9

sample to include all unscheduled FOMC meeting dates since 1995, reconstruct our shock, and find a correlation of 0.9 (Column 6). We then consider using a 2-day event window for both policy indicator and outcome variables. Doing this, we find that the correlation with the baseline shock series is 0.84 (Column 7). We also construct the instrumental variable as the daily movement in the policy indicator one day (as opposed to one week) before FOMC announcement day. As presented inColumn8ofTable2,thisalternativeshockserieshasacorrelationof0.99withthebaselineseries. 2.3.3 Real-Time U.S. Shock Series and an Application to the ECB As a final robustness check on our Fed shock, we construct real-time versions of the series.13 We use two methods: First, estimate the first step on the sample up to 2007:12, use the betas from that in the second step regression to compute the aligned monetary policy shock for 2008:1, then roll through the sample one month at a time to construct a real-time shock for 2008:2, 2008:3, ... using these rolling window sensitivity indexes. Second, estimate the first step regression only up through 2007:12 and use the estimated betas from that regression to generate the aligned monetary policy shock series for each observation beginning in 2008:1. The correlations of these two real-time measureswiththebaseline,“ex-post”BRWshockseriesare0.95and0.88,respectively(seecolumns 9 and 10 of Table 2 and the on-line Appendix figures). Finally, to provide an example of our methodology’s general applicability, we construct a new shock series for the ECB. As described in the Euro Area appendix, we use as outcome variables Euro area zero coupon yields with maturities of 3 months and 1, 2, 5, 7, and 10 years. As policy indicator,webenchmarkalternatelywiththe2-yearand5-yearOISrate. AppendixFigureB.1plots the shock series together. As detailed below, this series is also devoid of the information effect. 2.3.4 Monetary Policy Shocks and the Slope of the Yield Curve Comparisons above suggest that our shock is closely related to forward guidance, which is well captured by movements in 2- or 5-year interest rates. Table 3 provides further evidence, with estimates of the effect of our shock on interest rate spreads. Here we’ll consider the 5-year interest rate as benchmark and regress interest rate spreads of different maturities over the 5-year rate on the monetary policy shock, 13OneadvantageofusingrawsurprisesasinKuttner(2001)andJK(2018)isthattheresultingshocksareprecisely whatoccurredinrealtime. SeriessuchasNS(2018),Swanson(2018),andourbaselinemeasureaboveare(full-sample) estimation-based,donotaccountforestimationerror,andarethusnotstrictlyspeakingreal-time. 10

∆SPREAD =α +β e +(cid:15) (6) i,t i i t i,t whereSPREAD isthedifferencebetweeninterestratewithmaturityiandthe5-yearratearound i,t the FOMC announcement and e is, alternatively, the BRW, NS, Swanson, and JK monetary policy t shock series. Column (1) of Table 3 shows the regression results of the 5-year rate itself. The coefficient on BRW is 0.679 and highly statistically significant. The response of the 2-year/5-year interest rate spread -0.113 (Column 4) is significantly negative but close to zero. Thus, the 2-year interest rate responds to our shock in a similar way as does the 5-year rate. Coefficients in regressions for all of the other spreads (6 month and 1 year (Column 2 and 3), 10, and 30 year rates (Column 5 and 6)) are negative and significant, suggesting that both the short and long end of the yield curve respond to our shock by less than does the 5-year interest rate. Finally, we run the same regressions for the NS, JK, and Swanson shock series, as seen in the remaining rows of the table. Our BRW series is similartoSwanson’sforwardguidanceshockseriesinthesensethatbothmovethe2-yearand5-year interest the most. The NS shock series and Swanson’s LSAP shock series capture the movements of the yield curve at the short end and long end, respectively. The JK shock (FF3) affects spreads significantlydifferentlyoninformationeffectandnon-informationeffectdays, arguablyasexpected. As seen in the final two rows, FF3 shocks on non-information effect days affect spreads in much the same way as NS shocks, while on information effect days the shock is strongest at the very short end of the yield curve, with zero effect on the 5-year rate itself or the 2-year rate.14 3 The Fed Information Effect RomerandRomer(2000),NakamuraandSteinsson(2018),andJarocinskiandKaradi(2018), among others, advance the hypothesis of a ”Fed information effect”: monetary policy announcementscontaininformationaboutcentralbankforecastsofeconomicfundamentals. Asaby-product, macroeconomicvariablessuchasoutputandinflationmaybeinfluencednotonlybytheannounced policy itself but also by the forecasting information contained in the announcement. The opposite forcesfromthesetwosources(thepolicyandthereactiontoit)maycausepuzzlingimpulseresponses such as output rising after a contractionary policy shock. Use of even narrow windows around cen- 14Results for our ECB shock are similar to our Fed shock (see appendix Table B.1). The shock series normalized onthe2-year(5-year)ratecapturesrelativelymoreinformationattheshorttomedium(mediumtolong)endofthe yieldcurve. 11

tral bank announcements may not alleviate the issue for researchers.15 In this section, we subject our series to the same tests for the information effect used by Nakamura-Steinsson and Jarocinski- Karadi. We find scant evidence of the information effect in the BRW measure and pinpoint reasons for why our results are different from others. 3.1 A Direct Test and Implications We begin with the test of Nakamura and Steinsson (2018). We confirm their results for their seriesandexaminerobustnesstoourshockandSwanson’s(2018). Specifically,werunregressionsof monthly changes in Blue Chip survey expectations of output growth on the monetary policy shock series of that month, and test for the Fed information effect based on the sign of the estimated coefficient.16 Table 4 reports the results. While the information effect is significant in the measures of Nakamura-Steinsson and Swanson, it is insignificantly different from zero in ours (see the first three columns). For a robustness check, we also find that the two real-time BRW measures are devoid of the information effect (fourth and fifth columns). InFigure2,wedepictthedifferencebetweenFedandBlueChipforecastsofrealGDPgrowth, a standard proxy for central bank private information used in the literature.17 Noteworthy are the large negative values around September 11, 2001 and the last quarter of 2008. At these times, the Fed was significantly more bearish on the economy than the private sector.18 Table 5 reports OLS regressions of the various monetary policy shock series on these forecast differences. The coefficient is positive and significant for the NS and Swanson measures, but insignificantly different from zero intheregressionusingourseries,aregressionwithanR2ofonly0.02. Onceagain,thecentralbank information effect seems barely present in our new series. 15Campbell et. al. (2012) also provide evidence of a Fed information effect. Faust, Swanson, and Wright (2004) and Zhang (2019) find no such evidence, however, while Lunsford (2018) argues that in his sample from February 2000toMay2006theinformationeffectispresentinthefirsthalfonly. 16Inaddition,wefindrobustresultsrunningthetestsontheNSsub-samples: 1995-2014,2000-2014,and2000-2007 (see the online appendix). Extending through 2018 does not alter our conclusions. Also following NS, we exclude fromtheseregressionsallobservationswhenFOMCmeetingsoccurredinthefirstweekofthemonth, asthatlikely precedesthetimethattheBlueChipsurveyforecastwasmadeforthatmonth. 17The series is constructed as follows: (1) prior to December 2013, the average of the first four quarters ahead GreenbookforecastsminusthecorrespondingBlueChipforecasts. (2)AfterJanuary2014,forwhichtheGreenbook forecasts are not yet publicly available, we use the forecasts from the Fed summary of economic projections (SEP). Theseareavailablefourtimesayear: inMarch,June,September,andDecember. FortheotherfourFOMCmeetings each year, we use the SEP from the previous meeting. We use the current year SEP forecast if the FOMC meeting happens in the first quarter of the year. Otherwise, we use the projection for the following year. We subtract from thistheyear-aheadBlueChipforecast. 18Thesewerealsotimeswhenimportantnewseventsoccurredatahigherfrequencythantheavailableforecasts. 12

3.2 Evidence from an Indirect Approach JarocinskiandKaradi(2018)constructtheirinformationshockbyexaminingthehigh-frequency co-movement of interest rate and stock price surprises on FOMC announcement days. They argue thatwhenthestockmarketmovesinthesamedirectionasinterestrates, theFedinformationeffect dominates the monetary policy news effect of the announcement. Following Jarocinski and Karadi, we depict in the scatterplot of Figure 3 daily returns on the S&P 500 on FOMC announcement days against the BRW shock (blue dots) as well as the JK surprises – FOMC announcement day high-frequency changes in the third Fed Funds futures contract (in orange). Although the relationship is negative overall, there are clearly many points falling in the first and third quadrants. As emphasized by Jarocinski and Karadi, these are difficult to explain as purely monetary policy shocks. We re-estimate the NS information effect regressions, separately on Fed information effect days and non-information days, for both BRW and JK measures. The results are displayed in columns six and seven (BRW) and eight and nine (JK) of Table 4. In regressions with the BRW measure, the point estimates are very small and have no statistical significance. Thus, even during the “Jarocinski-Karadi” information effect days our BRW shock does not display economically or statisticallyimportantFedinformationeffectsinthesenseofNS.However, thenexttwocolumnsof Table 4 confirm that the information effect is present in the Jarocinski-Karadi data. This naturally sparks the question we address in the sub-section after next. 3.3 Evidence on the ECB Shock Series As described in the Euro Area appendix, we estimate the NS regressions by forecaster and on thetimeseriesofthe(monthly)medianforecast. AsshowninTableB.2, thereisnoevidenceofthe informationeffectformostofthe49forecasters. Onlyforoneforecasteristhereconsistentevidence of an information effect. Turning to the time series, in Figure B.2 we plot for each policy meeting date the number of forecasters whose outlook changed in the same direction as the policy surprise. Mostofeventdateshaveaverysmallfractionofforecastersexhibitingchangesintheiroutlookthat reflect an information effect. 3.4 Why Does Our Shock Series Have Less of a Fed Information Effect? In order to understand why our monetary policy shock series does not have an information effectinit,webeginbyconsideringtheimportanceoftheunderlyingdataandeconometricprocedure 13

usedtoconstructtheseries. First,wefindthattheinclusionoflong-terminterestratesisimportant because long-term interest rates are less associated with Fed information effects. Nakamura and Steinsson construct their monetary policy shock from a set of variables that contains short-term interestratesuptotwoyears. Bycontrast,weusethewholeyieldcurvetocomeupwithasummary measure of the stance of monetary policy. In Table 6, we report results of the NS information effect regressions–monthly changes in Blue Chip survey expectations of output growth on the 30-minute changesofinterestrates–withmaturitiesfrom1day(Fedfundsfuturerate)to30-yeartreasurybond yield. This table is similar to Table 4. It is clear that as the maturity of interest rates increases, the coefficientsbecomelesssignificant. ThisindicatesthatonereasonourBRWshockseriescontainsless of a Fed information effect is because we incorporate longer term interest rates than do alternative measures of Fed monetary policy shocks. Second, we find that the two-step PLS procedure (i.e. Fama-Macbeth) is equally important in reducing the Fed information effect in our shock series. To see this, we input our data into the principalcomponentsestimationproceduretoconstructanalternativemonetarypolicyshockseries, whichwelabelthe“PCAshock”. Asseenincolumn13ofTable2,thecorrelationbetweenthisshock andourbaselineBRWshockisonly0.25. Moreover,estimatingtheNSinformationeffectregressions with this PCA shock, we find that a positive shock leads to a significant increase in the Blue Chip real GDP growth rate forecast in the next quarter, consistent with Fed private information effects embedded in this alternative series (Table 4, column 12). The PCA approach does not remove the Fed information effect even when the underlying data include long-term interest rates. We conclude our encompassing analysis by inputting data in tight windows around FOMC announcements, as in NS, into our estimation procedure. This includes data on the expected 3month eurodollar interest rates with horizons of 2 to 4 quarters, the current month Fed funds futures rate and the Fed funds futures rate immediately following the next FOMC announcement. The“Tight(NS)shock”generatedinthiswayhasacorrelationof0.38withtheBRWshock(Table2, column14). TheinformationeffectregressionsofTable4indicatethatapositiveshocktothisseries isunrelatedtochangesintheBlueChiprealGDPgrowthrateforecast(column11). Whathappens when we expand the NS data set to include longer horizon maturities? The “Tight(full) shock” is generatedwithourPLSestimationprocedurebutwiththeNSdataexpandedtofurtherincludethe expected3-montheurodollarinterestrateswithhorizonsof1to8quartersandon-the-runTreasury rates of 3 months, 6 months, 2 years, 10 years and 30 years. Using this expanded data increases the correlation with the BRW shock up to 0.50 (Table 2, column 15). Again, the information effect is 14

absent from this Fama-MacBeth aligned shock (Table 4, column 12). This confirms the importance of using the Fama-Macbeth procedure in accounting for differences in results on the information effects in monetary policy shock series. The PLS and PCA approaches are similar in the sense of extracting the common component from outcome variables, but the PLS procedure we use assigns weights based on the correlation of outcome variables with the policy indicator (5-year treasury yield).19 Since the Fed information effect is not present in the 5-year interest rate or interest rates with longer maturities (Table 6), it is to be expected that the common factor we extract also contains less of a Fed information effect. We thus conclude that the inclusion of long-term interest rates and the Fama-MacBeth procedures play important roles in the construction of the BRW shock, and accounts for much of the difference in our findings concerning the information effect.20 4 Impulse Responses Asnotedabove,theexistingliteraturehasofferedtheinformationeffectasonereasonwhythe transmissioneffectsofshockstomonetarypolicycouldhavesignsthatdifferfromthosepredictedby traditional theory. In this section, we present robust evidence confirming this hypothesis, using the array of monetary policy shock series above to compute impulse responses of output, inflation, and creditconditions. Shockstoseriesthatdonotcontaintheinformationeffect,suchasbaselineBRW, display conventionally-signed impulse responses while shocks to series that contain the information effect often give rise to impulse responses with the opposite signs. 4.1 BRW Shocks Following Romer and Romer (2004), we place our cumulative shock series in a monthly VAR model to identify the transmission effects of monetary policy shocks. We allow our monetary policy shocktocontemporaneouslyaffectallvariables: output,inflation,commoditypricesandexcessbond premium.21 We include commodity prices in light of the “price puzzle” (CEE, 1996) and the excess bond premium because of its ability to explain business cycles (Gilchrist and Zakrajsek, 2012) and 19AspointedoutbyKellyandPruitt(2013,2015),thePLSforecastasymptoticallyrecoversthelatentfactorthat drivesmovementsinthepolicyindicatorasthenumberofoutcomevariablesandlengthoftimeseriesbothincrease. 20We also investigated which part of our estimation procedure, IDH or PLS, is more important in isolating the Fedinformationeffect. WeconstructedanalternativeBRWshockseriesusingtheFama-Macbethtwo-stepprocedure withouttheuseofIDH butwiththesamepolicyindicatorandoutcomevariablesasinthebaseline. Aspresentedin Table2(columnlabelledOLS),theIDH-freeshockishighlycorrelatedwiththebaselineBRWshock(0.991). 21ThisalsofollowsRomerandRomer. Ourseriesandtheirsareplausiblyexogenous,givenhowtheyareconstructed. 15

as an indicator of the price of risk (Creal and Wu, 2016). The variables in our baseline model are thusordered: cumulativemonetarypolicyshockseries,logindustrialproduction,logconsumerprice index, log commodity price index, and excess bond premium. We use 12 monthly lags.22 Figure 4a presents the impulse responses to a contractionary monetary shock using the full sample (1994-2017). Here and throughout the paper we normalize to a 100 basis point positive monetary policy shock on impact. The 68% and 90% standard error confidence intervals, displayed as deep and shallow gray areas respectively, are generated by the bootstrap. Both output and inflation decrease after a contractionary monetary policy shock. The responses reach their troughs after about 10 months. The excess bond premium increases and peaks after about 8 months. These results are conventional, in line with those of Gertler and Karadi (2015), for example. Figure 4b shows the impulse responses when the model is estimated on the post-2008 subsample. The responses are similar. Output and inflation significantly decrease for the first 10 months after a contractionary monetary policy shock, while the excess bond premium increases significantly.23 Thus, the impulse responses from a shock to the BRW series are conventional and highly stable across the ZLB sub-period. 4.2 IRF Robustness with BRW Shocks In light of standard concerns about potential dynamic mis-specification in VAR models, our first robustness check is to re-estimate using Jorda (2005) local projections.24 This constructs impulseresponsesfromtime-seriesregressionsforeachpointintime. AppendixFigureA1apresents the impulse responses to a contractionary monetary policy shock using the full sample (1994-2017). Afterapositiveshock,industrialproductionsignificantlydecreasesabout2monthslaterandreaches its trough after 15 months. Inflation immediately and sharply decreases throughout the 24 months. The excess bond premium responds positively through the first 10 months. Figure A1b shows that results for the ZLB sub-period estimated using local projections are very similar to those of the full sample and hence similar to those estimated from the VAR model. The next robustness check concerns the term premium. For this purpose, we subtract from the raw interest rates the corresponding term premium on the 5-year Treasury rate and all the 22Wealsoexaminesystemswiththe5-yearinterestrateasanadditionalvariableintheVARmodel. Thesegenerate similarimpulseresponses. 23Estimatesfromthepre-2008sub-samplearehighlysimilarandomittedforbrevity. 24AgainthisfollowsRomerandRomer(2004), whoestimateaVARwithcumulativemonetarypolicyshocksand alsoestimateaversionoflocalprojections. 16

zero-couponyieldswith1to10-yearmaturity, asestimatedbyAdrian, Crump, andMoench(2013). We then reconstruct our monetary policy shock series excluding the term premium. Inserting the cumulative values of that series into the baseline VAR model, we find that the impulse responses are quantitatively identical to the baseline results of Figure 4, although the negative effect on IP is dampened for the first few months (see online Appendix). As shown in column 9 of Table 2, the correlation between the term-premium free shock and our baseline shock is high, 0.79. 4.3 Alternatives: Nakamura-Steinsson, Swanson, and Jarocinski-Karadi Wecomparetheimpulseresponsesabovetothoseestimatedbyreplacingourshockserieswith that of, alternately, Nakamura and Steinsson (2018) and Swanson (2018), both of which embody the information effect (Table 4). Nakamura and Steinsson do not directly estimate the effects of their policy news shock on output (nor does Swanson (2018)), but rather focus on the response of expectationsoffutureoutputgrowthandrealinterestratesinanon-VARframework. Theseauthors alsodoextensivequantitativemodelingandconcludefromtheirestimationofthemodelthatroughly two-thirds of the monetary shock is due to the Fed information effect. Following Gurkaynak, Sack, and Swanson (2005), Swanson (2018) argues that monetary policy has more than one dimension. Changes in the federal funds rate are different from forward guidance announcements, and both of thesearedifferentfromLSAPannouncements,atleastintermsoftheireffectsonfinancialmarkets. ThevariousshockmeasuresfromtheSwansonpapersthusreflecttheeffectof,e.g.,a25bpdeclinein long rates that is carried out through an increase in asset purchases versus one that is accomplished viastimulativeforwardguidanceoradropinthetargetrate. Interpretationoftheeffectsofshocksto our series is different but complementary. Our estimates represent the effects of an FOMC meeting day shock that reflects the effect of, e.g., a 25bp decline in the 5-year rate following the words and actions(orinactions)undertakenbytheFOMC.Ourmeasureisbestthoughtoftellingustheeffect of an “average” 25bp loosening of the 5-year Treasury yield following the FOMC meeting, where this average is in principle a combination of Fed funds rate loosening, some expansionary forward guidance, and some LSAP increases.25 Figure 5 presents the results. The sample periods are: full (1994-2015)and during the ZLB (2008-2015). For the full sample (Figures 5a), impulse responses using any of the shocks follow the conventional monetary model. Output and inflation decrease while the excess bond premium 25This can be thought of as a “FRB-US view of the world”, in the sense that it mimics how Federal Reserve Board staff analyze monetary policy in their large scale estimated general equilibrium model of the U.S. economy (https://www.federalreserve.gov/econres/us-models-about.htm). 17

increasesafteracontractionarymonetarypolicyshock. However,duringtheZLBsub-sample(Figure 5b),theimpulseresponsesdifferacrosscases. FollowingapositiveshocktotheNakamura-Steinsson measure, both output and inflation rise significantly after about 10 months. In response to the shockidentifiedbySwanson(FGplusLSAP),output,inflationandexcessbondpremiumeffectively do not change.26 To further assess the possible role of Fed private information in accounting for differences in thetransmissioneffectsduringtheZLBperiodshowninFigure5,wereplacetheoriginalshockseries with the residual from the regression of Table 5.27 This “purged” series represents that component of the raw monetary policy shock that is not accounted for by differences in the Fed-private sector outlook. Impulse responses using the shock series of NS and Swanson are reported in Appendix Figure A2a-b, respectively. In the left panels, we depict point estimates and confidence bands from the VARs with the orthogonalized series. In the far right panels are IRFs using the original shock series. The middle column presents the comparison, omitting confidence bands for ease of viewing. ForbothNSandSwansonpurgedshocks,thepositiveresponsesofoutputtoacontractionarypolicy shockarediminishedcomparedtoIRFsfromtherawshocks. Indeed, theresponsesofshockstothe purged Swanson measure have conventional signs (Figure A2a). As noted above, Jarocinski and Karadi (2018) argue that the information effect is empirically important by showing that output, price level, and excess bond premium respond with significantly different signs to a monetary policy shock compared to the shock conditioned on stock prices and interest ratesco-movingpositively, whichtheylabel centralbankinformationshocks. InFigure 6A, we replicate the results of Jarocinski and Karadi (2018) using their monetary policy surprise FF3. InFigure6Bwere-estimateusingournewshockandfindquitedifferentresults. Wedepictimpulse responses on “non-information effect days”, points in the second and fourth quadrants of Figure 3, andon“informationeffectdays”,pointsinthefirstandthirdquadrants.28 Intheleft(right)panels, we report the point estimates and error bands for the non-information (information) day shocks. In the middle column, we display the point estimate comparison without confidence bands. ConsiderFigure6Afirst,theresultswiththeJKmeasure. Onnon-informationeffectdays,the 26WealsoperformthisexercisewitheachoftheseparateSwansonshocksandfindsimilarresults. Inaddition,we estimateimpulseresponsestoidentifiedshockstotheWu-Xiashadowrateindex. DuringtheZLB,impulseresponses are conventional and significant at first, but exhibit the opposite sign at long horizons. Wu and Xia estimate a FAVAR model, different from the basic VAR here, and report conventional responses. For example, they find that expansionaryFedmonetarypolicyshocksraiseIPandlowerunemploymentduringtheperiodJuly2009toDecember 2013,inmuchthesamewaythatshockstotheeffectiveFedFundsratedidpriortotheZLBperiod. 27Miranda-AgrippinoandRicco(2017)andKane,Rogers,andSun(2018)pursueasimilarstrategy. 28We use all available VAR data in these experiments, and simply set shocks on the other days to zero. This is equivalenttothesecondestimationprocedureusedbyJarocinskiandKaradi,labelled“poormansignrestrictions.” 18

leftpanel,theimpulseresponsesexhibittraditionalsigns. Outputandpricelevelfallinresponsetoa monetarycontraction,whilecreditconditionstighten(EBPrises). Impulseresponsesoninformationeffectdays, therightsidecolumn(inblue), producesignificantlydifferentresults, however, withthe transmission effects changing signs. The results are noticeably different when we use our new shock series, as in Figure 6B. Transmission to output, prices, and credit conditions exhibit conventional signs, irrespective of estimating on information effect days or non-information effect days.29 As a final check, we estimate impulse responses from shocks to the various measures constructed in our encompassing analysis of section 3. Results are displayed in the online appendix. Responses to the “PCA shock”, which embodies the information effect, are unconventional: muted inthefullsampleandmovinginthe“wrong”directionduringtheZLBperiod(FigureC.8). Impulse responsestoapositive“Tight(NSdata)”shock,whichisdevoidoftheinformationeffect,lookmore conventional: in the post-2008 sample, the IP and CPI responses are mostly negative, especially at intermediate horizons; the response of EBP is less negative at first and quickly turns positive after a short period of time (Figure C.9). Finally, positive shocks to the “Tight (full data)” shock series, alsodevoidoftheinformationeffect(Table4,row10),produceimpulseresponseswithconventional signs, albeit with some lagged effects compared to those with baseline BRW shocks (Figure C.10). 5 Conclusion We perform a novel application of well-known estimation procedures to derive a US monetary policy shock series that usefully bridges periods of conventional and unconventional policymaking and is effectively devoid of the information effect. Our approach has very mild data requirements and is easy to implement econometrically. As an example of the latter, we construct a new series for ECB monetary policy. It too shows essentially no evidence of the information effect. The heteroskedasticity-based estimator filters out background noise, while the monetary policy shock is aligned using Fama-MacBeth regressions. We demonstrate the importance of our procedure to the identificationofU.S.monetarypolicyshocksthroughdetailedcomparisonwithalternativemeasures in the literature, including an investigation of the Fed information effect. Overall, using the same testing and “purging” procedures as two prominent approaches in the literature, we find essentially no evidence of an information effect in our new monetary policy shock series. Wethenpresentevidenceconfirminganhypothesisintheliteraturethattheinformationeffect 29Recall from Table 4 that there is little evidence of an information effect, in the sense of NS, in the BRW series evenonJKinformationeffectdays. 19

canleadtomonetarypolicyshockshavingtransmissioneffectstooutputandinflationwithsignsthat differ from those predicted by traditional theory. We find that in response to contractionary shocks to our new measure, output and prices fall significantly, consistent with conventional theory. This result is found in samples both before the ZLB and during the ZLB sub-period with our measure. However, estimating impulse responses to monetary policy shocks that embody the information effect, we find responses that are either zero or positive. Staff at the Federal Reserve and other central banks want and need to know whether their models should be constructed to feature the information effect. Should the impulse responses associated with monetary policy announcements that the staff’s quantitative models attempt to match be of the signs predicted by traditional monetary theory, or of the unconventional signs consistent with evidence in influential papers like Nakamura-Steinsson and Jarocinski-Karadi? The evidence in this paper, and our unified measure, are useful for guiding these and other exercises in empirical and quantitative theoretical modeling of the effects of Fed monetary policy. References (1) Adrian, T., R.K. Crump, and E. Moench. 2013. Pricing the Term Structure with Linear Regressions. Journal of Financial Economics, 110(1), pp. 110-138. (2) Campbell, J. R., C. Evans, J. D. M. Fisher, and A. Justiniano. 2012. Macroeconomic Effects of Federal Reserve Forward Guidance. Federal Reserve Bank of Chicago working paper 2012-03. (3) Christiano, L., M. Eichenbaum and C. Evans. 1996. The Effects of Monetary Policy Shocks: Evidence from the Flow of Funds. The Review of Economics and Statistics, 78(1): 16-34. (4) Cochrane, H. J., and M. Piazzesi, 2002. The Fed and Interest Rates: A High-Frequency Identification. NBER Working Paper No. 8839. (5) Creal,D.,andC.Wu. 2016. BondRiskPremiainConsumption-basedModels. NBERWorking Paper, No. 22183. (6) Fama, E. F., J. D. MacBeth. 1973. Risk, Return, and Equilibrium: Empirical Test. Journal of Political Economy, 81 (3): 607-636. (7) Faust, J., E. Swanson, and J. H. Wright, 2004. Do Federal Reserve policy surprises reveal superior information about the economy? Contributions in Macroeconomics, 4(1). 20

(8) Gertler, M,andP.Karadi, 2015. MonetaryPolicySurprises, CreditCostsandEconomicActivity. American Economic Journal: Macroeconomics, 7(1): 44-76. (9) Gilchrist, S., and E. Zakrajsek. 2012. Credit Spreads and Business Cycle Fluctuations. American Economic Review, 102(4): 1692-1720. (10) Gurkayanak, R., B. Sack, and E. Swanson. 2005. Do Actions Speak Louder Than Words? The Response of Asset Prices to Monetary Policy Actions and Statements. International Journal of Central Banking, 1(1): 55-93. (11) Inoue, A., Rossi, B. 2018. The effects of conventional and unconventional monetary policy: A new approach. Manuscript, Vanderbilt University and Pompeu Fabra. (12) Jarocinski, M., and P. Karadi. 2018. Deconstructing Monetary Policy Surprises -— the Role of Information Shocks. CEPR working paper 12765. (13) Jord`a,O`scar,2005. EstimationandInferenceofImpulseResponsesbyLocalProjections. American Economic Review, 95(1): 161-182. (14) Kane, A., J. Rogers, and B. Sun. 2018. Communications Breakdown: the Transmission of Different Types of ECB Policy Shocks. Federal Reserve Board working paper. (15) Kelly, B., and S. Pruitt. 2013. Market expectations in the cross-section of present values. Journal of Finance, 68(5): 1721-1756. (16) Kelly, B., and S. Pruitt. 2015. The three-pass regression filter: A new approach to forecasting using many predictors. Journal of Econometrics, 186(2): 294-316. (17) Kuttner,K.N..2001. Monetarypolicysurprisesandinterestrates: EvidencefromtheFedfunds futures market. Journal of Monetary Economics, 47(3): 523-544. (18) Lunsford,K.G.2018. UnderstandingtheAspectsofFederalReserveForwardGuidance,working paper, Federal Reserve Bank of Cleveland. (19) Miranda-Agrippino, S., and G. Ricco. 2017. The Transmission of Monetary Policy Shocks. Centre for Macroeconomics, Discussion Paper 2017-11. (20) Nakamura, E., and J. Steinsson. 2018. High Frequency Identification of Monetary Non- Neutrality: The Information Effect. Quarterly Journal of Economics, 133(3): 1283-1330. 21

(21) Rigobon, R. 2003. Identification through Heteroscedasticity. Review of Economics and Statistics, 85(4): 777-792. (22) Rigobon, R., and B. Sack. 2003. Measuring the Reaction of Monetary Policy to the Stock Market. Quarterly Journal of Economics, 118(2): 639-669. (23) Rigobon, R., and B. Sack. 2004. The Impact of Monetary Policy on Asset Prices. Journal of Monetary Economics, 51(8), 1553-1575. (24) Rogers,J.,C.Scotti,andJ.H.Wright. 2018. UnconventionalMonetaryPolicyandInternational Risk Premia. Journal of Money, Credit, and Banking, 50(8). (25) Romer,C.andD.Romer,2000. FederalReserveInformationandtheBehaviorofInterestRates. American Economic Review 90 (2000), 429–57. (26) Romer, C. and D. H. Romer. 2004. A new measure of monetary shocks: Derivation and implications. American Economic Review, 94(4):1055-1084. (27) Swanson, E. 2018. Measuring the Effects of Federal Reserve Forward Guidance and Asset Purchases on Financial Markets. NBER Working Paper No. 23311. (28) Wright, J. H.. 2012. What does Monetary Policy do to Long-term Interest Rates at the Zero Lower Bound? The Economic Journal, 122(564): F447-F466. (29) Wold, H.. 1966. Estimation of principal components and related models by iterative least squares. In P. Krishnaiaah (eds.), Multivariate Analysis: 391-420. New York: Academic Press. (30) Wold,H.1975. Pathmodelswithlatentvariables: TheNIPALSapproach. H.Blalock,A.Aganbegian,F.Borodkin,R.Boudon,andV.Cappecchi(eds.),QuantitativeSociology: International Perspectives on Mathematical and Statistical Modeling. New York: Academic Press. (31) Wu, C. J. and F. D. Xia. 2016. Measuring the Macroeconomic Impact of Monetary Policy at the Zero Lower Bound. Journal of Money, Credit, and Banking, 48(2-3), 253-291. (32) Zhang, X. 2019. A New Measure of Monetary Policy Shocks. Working paper, UC-San Diego. 22

Figure 1: BRW Shock Series Jan 1994 to Dec 2017 Note: The BRW shock series is estimated from Equations (3) and (4). The navy vertical lines denoteannouncementsofQE1,QE2,andQE3;theorangeverticallinesdenotetheOperationTwist period; and the blue line denotes Oct. 2015, the FOMC meeting prior to liftoff. 23

Figure 2: GDP Growth Forecasts, Fed Minus Blue Chip Note: Prior to December 2013, this is the average of the first four quarters ahead Greenbook forecasts less the corresponding Blue Chip forecasts. After January 2014, we use forecasts from the FOMC summary of economic projections (SEP) because the Greenbook data is not yet publicly available. TheFedSEPareavailablefourtimesperyear-—inMarch,June, September,andDecember. For the other four FOMC meetings, we use the SEP from the previous FOMC meeting. We use the current year SEP forecast for real GDP growth rate if the FOMC meeting happens in the first quarter of the year. Otherwise, we use the next year SEP forecast for real GDP Growth. 24

Figure 3: S&P 500, the BRW Shock, and the JK Shock Note: The S&P 500 returns are computed over a 30-minute window around FOMC meeting announcements. ThebluedotsrepresenttheBRWshocks,andtheorangetrianglesarethesurprises of the 3-month federal funds futures that are used by Jarocinski and Karadi (2018). 25

Figure 4: Baseline SVAR Impulse Responses: BRW Shocks a. 1994m1-2017m12 b. 2008m1-2017m12 Note: Structural VAR with monthly data, 5 endogenous variables and 12 lags. Variables are orderedasfollows: cumulativeBRWshockseries,logindustrialproduction,logconsumerpriceindex (CPI), log commodity prices, and excess bond premium. Graphs show impulse responses estimated overdifferentsampleperiodstoa100basispointincreaseinthecumulativeBRWshockseries. Deep and shallow gray shaded areas are 68% and 90% confidence intervals produced by bootstrapping 1000 times, respectively. 26

Figure 5: SVARs with Alternative Shock Series: BRW, NS, and Swanson a. 1994m1-2015m12 b. 2008m1-2015m12 Note: BRW,NSandSwansonrefertocumulativeBRWshockseries,NakamuraandSteinsson (2018) shock series, and Swanson (2017) shock series, respectively. For these cases, variables are ordered: the cumulative shock series, log industrial production, log consumer price index (CPI), log commodity prices, and excess bond premium. Graphs show impulse response to a 100 basis point increase in the monetary policy indicator series. Deep and shallow gray shaded areas are 68% and 90% confidence intervals produced by bootstrapping 1000 times, respectively. 27

Figure 6: SVAR on Non-information Days (red) and Information Days (blue) a. Using the Jarocinski-Karadi FF3 Shock b. Using BRW Shock Note: Full sample-period estimation. FF3 is accumulated 3 month federal funds futures rate around the 30-minute FOMC announcement window according to the information day definition in Jarocinski and Karadi (2018). The BRW shock is accumulated in the same way. 28

Table 1: Correlation with BRW Shock Series Full Sample Pre-ZLB ZLB NS Shock 0.512 0.653 0.494 SS shock 0.625 0.684 0.532 R&R Shock 0.131 Kuttner Shock 0.308 SS FFR 0.373 SS FG 0.492 0.605 0.575 SS LSAP 0.365 FF3 0.395 0.593 0.336 Note: The benchmark shock is our BRW shock series estimated from Equation (3) and (4). NS Shock refers to the policy factor shock of Nakamura and Steinsson (2018), which we update to the present. SS Shock refers to the sum of the shock series of the federal funds rate, the forward guidanceandthelargeassetpurchasesinSwanson(2018). R&R Shock referstotheestimatedshock series in Romer and Romer (2004). Kuttner Shock refers to the 30-minute Fed Funds rate changes around FOMC announcements. SS FFR, SS FG, SS LASP refers to the shock series of the Federal Funds rate, forward guidance and large asset purchases in Swanson (2018). FF3 is the 30-minute changein3monthfederalfundsfuturesratearoundtheFOMCannouncement. Sampleperiodsare: Full sample 1994m1-2017m12, Pre ZLB 1994m1-2008m12, ZLB 2009m1-2015m12. 29

seireS kcohS WRB enilesaB htiw snoitalerroC :ssentsuboR seireS kcohS :2 elbaT thgiT thgiT ACP SLO PT WRB WRB 2VI 2yaD eludehcsnU EQ 5R 96WRB 01N 2N )lluF( )SN( )2TR( )1TR( 005.0 083.0 942.0 299.0 197.0 388.0 459.0 599.0 838.0 309.0 789.0 759.0 389.0 189.0 579.0 kcohS WRB 191 191 091 191 191 191 191 191 191 381 091 191 191 191 191 snoitavresbO .)4( dna )3( noitauqE morf detamitse seires kcohs WRB ruo ot srefer kcohS WRB :etoN .rotacidni ycilop sa etaR yrusaerT raey-2 eht gnisu morf dengila seires kcohs WRB eht ot srefer 2N .rotacidni ycilop sa etaR yrusaerT raey-01 eht gnisu morf dengila seires kcohs WRB eht ot srefer 01N .21m7102 ot 1m9691 revo elpmas eht morf detamitse seires kcohs WRB ruo ot srefer 96WRB .selbairav emoctuo sa seitirutam raey-03 ,01 ,5 ,2 ,1 eht ylno htiw sdleiy nopuoc-orez gnisu dengila seires kcohs WRB eht ot srefer 5R .9002 ,hcraM ni 1EQ fo tnemecnuonna eht gnidulcxe seires kcohs WRB eht ot srefer EQ .5991 ecnis setad gniteem CMOF deludehcsnu eht fo lla gnidulcni dengila seires kcohs WRB eht ot srefer eludehcsnU .syad tnemecnuonna CMOF dnuora wodniw tneve yad-2 a gnisu dengila seires kcohs WRB eht ot srefer 2yaD eno naht rehtar yad tnemecnuonna CMOF erofeb yad-1 rotacidni ycilop eht ni stnemevom yliad gnisu dengila seires kcohs WRB eht ot srefer 2VI .elbairav latnemurtsni eht sa keew .8002 erofeb kcohs WRB lanigiro dna 8002 tsop dohtem elpmas gnillor gninibmoc seires kcohs WRB ot srefer )1TR( WRB .elpmasbus 8002-erp fo sexedni ytivitisnes morf dengila seires kcohs WRB ot srefer )2TR( WRB .muimerp mret detamitse eht fo eerf tub )4( dna )3( noitauqE fo hcaorppa enilesab eht sa detareneg seires kcohs WRB eht ot srefer PT .erudecorp HDI eht tuohtiw dohtem htebcaM-amaF elpmis eht morf dengila seires kcohs WRB evitanretla eht ot srefer SLO segnahcyliad(selbairavemoctuolla,.e.i,atadgniylrednuruofotnenopmoclapicnirptsrfiehtgnitcartxemorfdetarenegseireskcohsehtotsreferACP .)gniteem CMOF dnuora etar nopuoc orez raey-03 ot 1 fo htnomtnerrucehtfosegnahcetunim-03eht ,.e.i ,)8102(nossnietSdnaarumakaNgniylrednuatadehtgnisuseireskcohsWRBehtotsrefer )SN(thgiT rallod orue daeha retrauq ruof ,eerht ,owt dna ,gniteem CMOF txen eht gniwollof yletaidemmi etar serutuf sdnuf deF eht ,etar serutuf sdnuf deF .serutuf raey 03 ,raey 01 ,raey 5 ,raey 2 ,htnom 6 ,htnom 3 eht fo segnahc etunim-03 eht dna atad SN eht gnisu seires kcohs WRB eht ot srefer )lluF(thgiT .stnemecnuonna CMOF dnuora setar tseretni 30

Table 3: Monetary Policy Shocks and the Slope of the Yield Curve 5y 6m - 5y 1y - 5y 2y - 5y 10y - 5y 30y - 5y BRW 0.679*** -0.432*** -0.351*** -0.113*** -0.232*** -0.782*** (0.05) (0.05) (0.05) (0.04) (0.02) (0.02) NS 1.102*** -0.211 -0.175 0.076 -0.366*** -0.990*** (0.14) (0.14) (0.12) (0.08) (0.05) (0.11) SS (FG) 0.508*** -0.350*** -0.284*** -0.0645** -0.111*** -0.342*** (0.05) (0.05) (0.04) (0.03) (0.02) (0.04) SS (LSAP) 0.575*** -0.588*** -0.529*** -0.346*** 0.0977*** -0.185** (0.08) (0.07) (0.06) (0.04) (0.03) (0.08) FF3 (JK info) -0.292 0.659*** 0.472*** 0.302** -0.0124 -0.0773 (0.18) (0.19) (0.16) (0.12) (0.06) (0.14) FF3 (Non JK info) 0.867*** -0.175 -0.1 0.0563 -0.350*** -0.830*** (0.16) (0.15) (0.13) (0.08) (0.06) (0.13) Note: Constant term not displayed. Robust standard errors in brackets. * p < 0.10, ** p < 0.05,***p<0.01. 5y referstothedailychangeinthe5-yeartreasurybondyieldaroundtheFOMC announcement. 6m-5y, 1y-5y, 2y-5y, 10y-5y, and 30y-5y refer to the differences between the daily changes in 6 month, 1, 2, 10, and 30 year treasury bond yields around the FOMC announcement and the 5-yr. rate. The (updated) NS Shock is the shock series updated to 2015m12 following the method in Nakamura and Steinsson (2018). The regressions are estimated over each authors’ full sampleperiods. Sampleperiodsare1994m1-2018m8forBRWshockseries,1994m1-2015m12forNS shock series, and 1994m1-2015m11 for Swanson’s FG and LSAP shock series. 31

)8102( nossnietS dna arumakaN fo snoissergeR tceffE noitamrofnI deF :4 elbaT thgiT WRB thgiT WRB ACP 3FF 3FF WRB WRB WRB WRB SN SS WRB )atad lluf( )atad SN( )ofniN KJ( )ofnI KJ( )ofniN KJ( )ofnI KJ( )2TR( )1TR( 41.0- 51.0- **35.0 73.0 ***60.1 30.0- 93.0 51.0- 10.0- ***67.0 **61.0 10.0 5102-5991 )92.0( )31.0( )62.0( )32.0( )22.0( )61.0( )04.0( )71.0( )61.0( )12.0( )70.0( )61.0( 731 731 331 67 42 48 05 431 431 531 631 431 sbO eht no desserger sretrauq 3 txen eht revo htworg tuptuo fo snoitatcepxe yevrus pihC eulB ni )txen ot htnom tnerruc( egnahc ylhtnoM :etoN p ** ,01.0 < p * .stekcarb ni rorre dradnats tsuboR .pot ta detsil era sdoirep elpmaS .)deyalpsid ton( tnatsnoc a sulp htnom taht ni seires kcohs fo seires kcohs eht fo mus eht ot srefer SS .erudecorp noitamitse SLP ruo dna atad ruo :seires kcohs ruo ot srefer WRB .10.0 < p *** ,50.0 < fo skcohs swen ycilop eht ot srefer SN .01 yb delacs ,)8102( nosnawS fo sesahcrup tessa elacs egral dna ecnadiug drawrof ,etar sdnuF laredeF eht tceffe noitamrofni deF eht no ylno gnisucof snoisserger elpmas-bus owt era )ofniN KJ(WRB dna ) ofnI KJ(WRB .)8102( nossnietS dna arumakaN htnom-3ehtfosesirprusehtera )ofniN KJ(3FFdna )ofnI KJ(3FF .)8102(idaraKdnaiksnicoraJybdenfiedsa ,syadtceffenoitamrofni-nondnasyad ACP .)8102(idaraKdnaiksnicoraJybdesusa ,syadtceffenoitamrofni-nondnasyadtceffenoitamrofnideFehtnoylnognisucofserutufsdnuflaredef segnahc yliad( selbairav emoctuo lla ,.e.i ,atad gniylrednu ruo fo tnenopmoc lapicnirp tsrfi eht gnitcartxe morf detareneg seires kcohs eht ot srefer SLP gnisu detupmoc seires kcohs WRB eht ot srefer )atad SN(thgiT WRB .001 yb delacs ,)gniteem CMOF dnuora etar nopuoc orez raey-03 ot 1 fo etar serutuf sdnuf deF eht ,etar serutuf sdnuf deF htnom tnerruc eht fo segnahc etunim-03 eht ,.e.i ,)8102( nossnietS dna arumakaN ni atad eht htiw WRB eht ot srefer )atad lluf(thgiT WRB .serutuf rallod orue daeha retrauq ruof ,eerht ,owt dna ,gniteem CMOF txen eht gniwollof yletaidemmi CMOF dnuora setar tseretni raey 03 ,raey 01 ,raey 5 ,raey 2 ,htnom 6 ,htnom 3 eht fo segnahc etunim-03 eht sulp atad SN eht gnisu seires kcohs WRB .8002 erofeb kcohs WRB lanigiro dna 8002 tsop dohtem elpmas gnillor gninibmoc seires kcohs WRB ot srefer )1TR( WRB .stnemecnuonna seires kcohs eht ot srefer )PASL( SS ,)RFF( SS ,)GF( SS .elpmasbus 8002-erp fo sexedni ytivitisnes morf dengila seires kcohs WRB ot srefer )2TR( .01 yb delacs lla ,ylevitcepser ,)8102( nosnawS fo sesahcrup tessa elacs egral dna etar sdnuF laredeF ,ecnadiug drawrof fo 32

Table 5: Shock Series Regressed on Fed minus Blue Chip GDP Growth Forecasts (1) (2) (3) (4) NS Shock Updated NS Shock BRW Shock Swanson Shock Fed - BC 2.00** 1.93*** 1.95 0.67** (0.77) (0.70) (1.53) (0.31) Observations 130 150 150 149 R-squared 0.09 0.08 0.02 0.07 Note: Constant term not displayed. Robust standard error in brackets. * p < 0.10, ** p < 0.05, *** p < 0.01. BRW Shock refers to our BRW shock series estimated from Equation (3) and (4). SS Shock refers to the sum of the shock series of the federal funds rate, the forward guidance and the large asset purchases proposed by Swanson (2018). We scale the SS shock by 100. NS Shock refers to the policy factor shocks from Nakamura and Steinsson (2018). The updated NS Shock is the shock series updated to 2015m12 following the method in Nakamura and Steinsson (2018). Fed - BC is the difference between Fed and Blue Chip GDP growth Forecasts, constructed as described above. Sample periods are: 1995m1-2014m3, 1994m1-2015m12, 1994m1-2015m12, and 1994m1-2015m11 (Swanson’s sample ends just before lift-off). Table 6: Fed Information Effect in Interest Rates with Different Maturities Kuttner 6-month 2-yr. 5-yr. 10-yr. 30-yr. Coef. 0.296*** 0.389* 0.368** 0.277 0.308 0.214 (0.11) (0.22) (0.17) (0.18) (0.22) (0.30) Observations 144 144 144 144 144 144 R-squared 0.04 0.024 0.034 0.017 0.012 0.004 Note: Constant term not displayed. Robust standard error in brackets. * p < 0.10, ** p < 0.05, *** p < 0.01. We regress the monthly change (current month to next) in survey expectations of output growth over the next 3 quarters from Blue Chip Economic Indicators on the shock series in that month. Kuttner Shock refers to monetary policy shock of Kuttner(2001). 6 month refers to the 30-minute change in 6 month treasury note yield around the FOMC announcement. 2, 5, 10, and 30 year refer to the 30-minute changes in 2, 5, 10, and 30 year treasury bond yields around the FOMC announcement. The sample period is 1994m1-2018m8. Following NS, we exclude the Great Recession period. 33

Appendix A A1. Implementation of Identification through Heteroskedasticity - IV approach We assume the monetary policy shock is unobservable. We normalize the shock to have 1-1 relationship with the changes in the 5 year interest rate, ∆R =α +e +η . (7) 5,t 0 t t The equation of interest is ∆R =θ +β ∆R +ξ (8) i,t i i 5,t i,t where ξ =−β η +(cid:15) , where (cid:15) is the idiosyncratic error associated with ∆R , (cid:15) is assumed i,t i t i,t i,t i,t i,t not to correlate with the monetary policy shock e , and ∆R is the change in i year interest rate t i,t around FOMC announcements. For simplicity and without loss of generality, we suppress the subscript i, and demean both ∆R and ∆R , i,t 5,t ∆R =β∆R +ξ . (9) t 5,t t Heteroskedasticity-based estimation – By construction, the regressor ∆R is correlated with 5,t the error term ξ due to the component −β η . The OLS estimation of β is biased due to the i,t i t i errors-in-variables problem. To deal with this problem, we need to identify two subsamples, which are denoted as M and NM. M is the sample with event windows around FOMC announcements and NM represents the non-monetary windows, which are the corresponding event windows one week before. We also need twoassumptionsregardingthesecondmomentoftheshockspresentinthemodel: ondaysofFOMC meetings, the variance of the ’true’ monetary policy shock increases while that of the background noise remains unchanged. Assumption 1: σM >σNM, σM =σNM,σM =σNM. e e η η ξ ξ Assumption 2: E[η e ]=E[ξ e ]=0. t t t t The implementation is very similar to Rigobon and Sack (2004). Denote the variance covariance matrix of each subsample as ΩM = E (cid:104)(cid:2) ∆RM ∆RM(cid:3)(cid:48) ∗ (cid:2) ∆RM ∆RM(cid:3)(cid:105) (10) 5,t t 5,t t ΩNM = E (cid:104)(cid:2) ∆RNM ∆RNM(cid:3)(cid:48) ∗ (cid:2) ∆RNM ∆RNM(cid:3)(cid:105) 5,t t 5,t t It is clear that (cid:34) (cid:0) ∆RM(cid:1)2 ∆RM∆RM (cid:35) ΩM = E 5,t 5,t t · (cid:0) ∆RM(cid:1)2 t  (cid:0) σM(cid:1)2 + (cid:0) σM(cid:1)2 β (cid:0) σM(cid:1)2  e η e =  · β2(cid:0) σM(cid:1)2 + (cid:16) σM (cid:17)2  1 e ξ The second equality follows from E[η e ]=E[ξ e ]=0. Similarly, we can write ΩNM out in terms t t t t of σNM and σNM. η ξ If we take the difference between these two covariance matrices and let (cid:0) σM(cid:1)2 − (cid:0) σNM(cid:1)2 =λ, e e we have ∆Ω = ΩM −ΩNM (cid:20) (cid:21) λ βλ = · β2λ 34

(cid:20) (cid:21) 1 β = λ · β2 Then, it is clear that β can be estimated as follows, ∆Ωˆ βˆ = 12 1 ∆Ωˆ 11 Now, ∆Ωˆ βˆ = 12 (11) 1 ∆Ωˆ 11 cov (cid:0) ∆RM,∆RM(cid:1) −cov (cid:0) ∆RNM,∆RNM(cid:1) = 5,t t 5,t t (12) (cid:0) (cid:1) (cid:0) (cid:1) var ∆RM −var ∆RNM 5,t 5,t E (cid:104)(cid:0) ∆RM,−∆RNM(cid:1)(cid:0) ∆RM,∆RNM(cid:1)(cid:48) (cid:105) 5,t 5,t t t = (13) E (cid:104)(cid:0) ∆RM,−∆RNM (cid:1)(cid:0) ∆RM,∆RNM (cid:1)(cid:48) (cid:105) 5,t 5,t 5,t 5,t According to (13), we may use an IV approach to implement this estimator. This approach rewrites (8) as: [∆R ]=α +β [∆R ]+µ i=1,2,···,30 (14) i,t i i 5,t i,t wheretheindependentvariable[∆R ]=(∆RM,∆RNM)(cid:48),theeventwindowof[∆R ]corresponds 5,t 5,t 5,t i,t to [∆R ]. β can be estimated using an instrumental variable ∆RIV =(∆RM,−∆RNM)(cid:48) for the 5,t i t 5,t 5,t independentvariable. Intuitively, (cid:0) ∆RM,−∆RNM(cid:1)(cid:48) isabletoinstrument (cid:0) ∆RM,∆RNM(cid:1)(cid:48) because, 5,t 5,t 5,t 5,t (1)itisclearthattheyarecorrelated;(2) (cid:0) ∆RM,−∆RNM(cid:1)(cid:48) doesnotcorrelatewiththeerrorterms, 5,t 5,t which follows directly from Assumption 1 & 2. 35

Figure A1: BRW Shock Series IRFs using Jorda (2005) Local Projections Method a. 1994m1-2017m12 b. 2008m1-2017m12 36

Figure A2: SVARs using shock series purged of the information effect a. Swanson Shock: Original (blue) versus Purged (red) Shock Series (table 5 residual) b. N&S Shock: Original (blue) versus Purged (red) Shock Series (table 5 residual) 37

Appendix B: A New Shock Series for the Euro Area Data We use the same methodology to construct a monetary policy shock series for the euro area. The outcome variables are Euro area zero coupon yields with maturity of 3 months, 1 year, 2 years, 5 years, 7 years and 10 years. These series are available on the ECB website http://sdw.ecb.europa.eu/browse.do?node=9691126. The policy indicator is, alternately, the 2-yr and5-yrOISrate. FigureB.1plotsthetwoshockseriestogether. Toexploretheinformationeffect, we use Consensus Forecasts for euro area GDP growth. This comprises 25-30 different forecasters every month since 2002 December. We test for the information effect by individual forecaster and for their median each month. Monetary Policy Shocks and the Slope of the Yield Curve InTableB.1wedisplayestimatesoftheeffectsofourECBshockoninterestratespreadswith different maturities (Equation (6)). The benchmark is the daily movement of the interest rate on whichtheshockseriesisnormalized. Considerfirsttheresultsbasedonthe2-yearrate. Asexpected, the benchmark 2-year interest rate responds most strongly while the 1-year interest rate behaves in a similar way. This indicates that the 2-year shock series (BRW OIS2Y) captures more information at the short to medium end. When our shock series is normalized on the 5-year OIS contract rate (BRW OIS5Y), both coefficients of 2-year/5-year and 7-year/5-year spread are numerically close to zero, indicating that this shock series captures more information at the medium to long end. Information Effect Tests We again follow the information effect tests of Nakamura and Steinsson (2018). We run these regressions by forecaster as well as the time series of the (monthly) median forecast. The data set includes two series, forecasts for the the current and the next year. We construct our monthly forecast series in two ways. The first is to use the current year forecast if the Consensus Forecast is from the first half of the year and use the next year forecast if the Consensus Forecast is made in the remaining half year. The second approach is to use the current year forecast if it is in the first nine months of the year and use the next year expectation if the forecast is made in the last quarter. Results are similar for these two approaches. Panel A of Table B.2 reports the overall test for the two shock series, focusing on the median forecast. None of the coefficients are significantly different from zero, suggesting that there is no information effect. Panel B repeats the same tests by individual forecaster. There is no evidence of the information effect for most of the 49 forecasters. Only for Lehman Brothers forecasts is there evidence of the information effect for both shock series and both forecast averaging methods. ForecastsfromBankAustriaandBancaIntesa,andthosefromCapitalEconomicshaveinformation effect for shock series normalized on 2-year and 5-year OIS contract rates, respectively. Finally, we turn to the time series, computing the percentage of forecasters with information effect at each policy event date. Figure B.2 plots the number of forecasters whose outlook changed inthesamedirectionasthepolicysurprise, ateachpolicydate. FollowingNakamuraandSteinsson (2018), we exclude the policy events happening in the first week of that month. Regardless of the interest rate on which the shock series is normalized, most of event dates have less than 60% forecasters exhibiting changes in their outlook that reflect an information effect. However, there are several policy events that have a quite large information effect percentage (around 80%): 2006m8, 2007m2,2008m5forBRW OIS2Yand2006m8,2007m2,2011m9forBRW OIS5Y.Wegothroughthe ECB statements. All of these policy decisions either increased the interest rate or left it unchanged while warning of high inflation. This common point may shed some light on the nature of the information effect, a topic for future research. 38

Figure B.1: BRW Shock Series for the Euro Area Note: Shock series estimated from Equations (3) and (4) using euro area data. The navy and gray bars are series normalized on 2-year and 5-year OIS rates, respectively. Figure B.2 Information Effect Counts a. Series normalized on 2-year OIS rate b. Series normalized on 5-year OIS rate Note: Theinformationeffectisdefinedastheco-movementsofGDPforecastersandmonetary policy surprises in the same direction. For each event, compute the percentage of forecasters that have information effect. 39

Table B.1: ECB Monetary Policy Shock and the Slope of the Yield Curve (1) (2) (3) (4) (5) (6) Panel A. BRW OIS2Y 2y 3m - 2y 1y - 2y 5y - 2y 7y - 2y 10y - 2y BRW 0.481*** -0.403*** -0.109*** -0.160*** -0.304*** -0.444*** (0.0314) (0.0233) (0.0196) (0.0185) (0.0205) (0.0215) Panel B. BRW OIS5Y 5y 3m - 5y 1y - 5y 2y - 5y 7y - 5y 10y - 5y BRW 0.396*** -0.373*** -0.102*** 0.0422** -0.0896*** -0.185*** (0.0283) (0.0203) (0.0256) (0.0197) (0.00862) (0.0177) Note: Constant term not displayed. Robust standard errors in brackets. * p < 0.10, ** p < 0.05, *** p < 0.01. BRW OIS2Y, BRW OIS5Y refers to BRW shock series normalized on 2-year and 5-year Euro OIS contract rates, respectively. 5y and 2y refers to the daily movements in the 2-year and 5-year zero coupon yield around ECB policy events. 3m-2y, 1y-2y, 5y-2y, 7y-2y, and 10y-2y refer to the differences between the daily movements in 3-month, 1-, 5-, 7-, and 10-year zero coupon yield and that of the 2-year zero coupon yield around ECB policy events. 3m-5y, 1y-5y, 2y-5y, 7y-5y, and 10y-5y refer to the differences between the daily movements in 3-month, 1-, 2-, 7-, and 10-year zero coupon yield and that of the 5-year zero coupon yield around ECB policy events . Table B.2: Information Effect (Regressions of Nakamura and Steinsson (2018)) (1) (2) (3) (4) BRW OIS2Y BRW OIS5Y Method 1 Method 2 Method 1 Method 2 Panel A. Consensus Median Median Forecasts -0.191 -0.235 -0.0738 -0.0828 (0.429) (0.426) (0.36) (0.358) Panel B. Individual Forecasters 1.ABNAmro -1.432 0.128 -1.829 -1.132 (2.266) (1.971) (1.938) (1.695) 2.AXAInvestmentManagers -0.0601 -0.0601 -0.00307 -0.00307 (0.219) (0.219) (0.179) (0.179) 3.Allianz 0.136 0.158 0.259 0.284 (0.27) (0.294) (0.221) (0.241) 4.BBVA 0.472 0.521 0.693 0.776 (0.718) (0.707) (0.571) (0.559) 5.BNPParibas 0.04 0.103 0.291 0.327 (0.585) (0.587) (0.494) (0.495) 6.BancaIMI -0.779 1.482 -0.673 -0.123 (0.703) (1.371) (0.753) (1.573) 7.BancaIntesa 6.835** 2.715 7.714 2.077 (1.361) (1.174) (7.984) (4.214) 8.BankAustria -0.567 2.466** 0.818 1.904 (1.38) (0.857) (1.624) (1.224) 9.BankJuliusBaer -0.884 -0.779 -0.365 -0.358 (0.56) (0.573) (0.478) (0.486) 10.BankVontobel 1.188 1.408 -0.177 -0.487 (0.71) (1.117) (0.99) (1.436) 40

11.BankofAmerica -0.628 -0.195 0.123 0.251 (0.791) (0.794) (0.666) (0.662) 12.Barclays -0.874 -0.874 -0.0208 -0.0208 (0.876) (0.876) (0.939) (0.939) 13.CapitalEconomics -54.87 -54.87 14.30* 14.30* (39.65) (39.65) (1.787) (1.787) 14.Citigroup -0.572 -0.581 -0.202 -0.209 (0.701) (0.7) (0.6) (0.599) 15.Commerzbank -0.569 -0.569 0.111 0.111 (0.425) (0.425) (0.358) (0.358) 16.CreditAgricole 0.703* 0.760* 0.468 0.49 (0.373) (0.386) (0.341) (0.355) 17.CreditSuisse -0.179 -0.213 -0.787 -0.813* (0.632) (0.606) (0.49) (0.465) 18.DeutscheBank -0.0257 0.787 -0.592 -0.529 (0.901) (0.907) (0.734) (0.758) 19.DresdnerBank -0.789 -0.789 -0.32 -0.32 (1.462) (1.462) (1.1) (1.1) 20.ETLA -0.324 -0.348 -0.338 -0.299 (0.662) (0.671) (0.54) (0.548) 21.EconIntelligenceUnit 0.00119 0.122 0.436 0.505 (0.726) (0.724) (0.604) (0.602) 22.EuropeanFcastNetwork -0.334 -0.338 -0.411 -0.401 (0.461) (0.461) (0.376) (0.376) 23.Exane 0 0 0 0 (0) (0) (0) (0) 24.Fortis -0.571 0.314 -0.36 -0.109 (1.243) (1.313) (1.084) (1.142) 25.GlobalInsight 0.373 0.662 0.293 0.387 (0.624) (0.594) (0.53) (0.514) 26.GoldmanSachs -1.226 -1.067 -0.483 -0.435 (0.832) (0.847) (0.688) (0.697) 27.GrupoSantander -0.694 -0.49 -0.548 -0.494 (0.717) (0.734) (0.603) (0.615) 28.HSBC 0.0385 0.302 0.0252 0.0882 (0.799) (0.832) (0.666) (0.695) 29.IHSMarkit -0.692 -0.589 -0.159 -0.127 (0.419) (0.449) (0.366) (0.386) 30.ING -0.44 -0.163 -0.595 -0.547 (0.9) (0.954) (0.876) (0.927) 31.IXISCIB 0.696 0.696 -0.969 -0.969 (2.444) (2.444) (1.827) (1.827) 32.IntesaSanpaolo -0.578 -0.579 -0.119 -0.137 (0.539) (0.537) (0.437) (0.435) 33.JPMorgan -1.201 -0.806 -0.0316 0.138 (0.802) (0.797) (0.676) (0.661) 34.LehmanBrothers 2.669** 2.669** 1.967* 1.967* (1.032) (1.032) (0.922) (0.922) 35.LloydsBankCB 0 0 0 0 (0) (0) (0) (0) 36.LloydsTSB -0.628 -0.628 -0.464 -0.464 (0.599) (0.599) (0.512) (0.512) 37.MerrillLynch 0.56 1.611 0.129 0.498 (1.922) (1.987) (1.673) (1.766) 41

38.MoodysAnalytics 0.0847 0.143 0.527 0.571 (0.507) (0.52) (0.432) (0.44) 39.MorganStanley 0.131 0.194 0.138 0.216 (0.537) (0.416) (0.453) (0.351) 40.Natixis -0.718 -0.734 -0.149 -0.267 (0.546) (0.553) (0.469) (0.475) 41.Nomura 0.0659 0.081 0.199 0.257 (0.552) (0.559) (0.436) (0.439) 42.OxfordEconomics -0.424 -0.294 -0.106 -0.0965 (0.381) (0.398) (0.323) (0.335) 43.SEB -1.953 -1.953 -1.688 -1.688 (1.163) (1.163) (0.982) (0.982) 44.Schroders 0 0 0 0 (0) (0) (0) (0) 45.SocieteGenerale 0.239 0.524 0.48 0.526 (0.476) (0.512) (0.337) (0.367) 46.SwissLife -0.148 -0.148 -1.773 -1.773 (1.351) (1.351) (1.415) (1.415) 47.UBS -1.239* -1.037 -0.835 -0.779 (0.694) (0.723) (0.592) (0.612) 48.UniCredit 0.167 0.27 0.25 0.274 (0.563) (0.567) (0.466) (0.47) 49.WestLB -1.847 -1.709 -1.807 -1.762 (1.407) (1.436) (1.175) (1.196) Note: Constant term not displayed. Robust standard errors in brackets. * p < 0.10, ** p < 0.05,***p<0.01. BRW OIS2Y,BRW OIS5YreferstoBRWshockseriesnormalizedon2-yearand 5-year Euro OIS contract rates, respectively. 5y and 2y refers to the daily movements in the 2-year and 5-year zero coupon yield around ECB policy events. Method 1 is to construct movements of forecasts between the next and the current month as movements of the current year forecasts if it is in the first half of the year and as movements of the next year forecasts in the remaining half year. Method 2 is to use movements of the current year forecasts if it is in the first three quarters of that year and movements of the next year forecasts in the last quarter. .Median Forecasts refers to the median of the forecasters in each month. 42

Appendix C: Online Appendix Table C.1: Fed Information Effect Regressions over Sub-periods 1995-2014 2000-2014 2000-2007 BRW Shock 0.09 0.10 0.33 (0.20) (0.20) (0.31) SS Shock 1.94** 1.81* 2.38*** (0.79) (0.99) (0.84) NS Shock 0.81*** 0.82*** 0.81*** (0.24) (0.29) (0.27) BRW(JK Info days) 0.69 0.89 -0.39 (0.78) (0.81) (0.92) BRW(JK non-info days) 0.00 -0.02 0.71 (0.36) (0.35) (0.59) FF3(JK Info days) 1.028*** 0.871*** 4.874*** (0.25) (0.23) (0.98) FF3(JK non-info days) 0.25 0.217 0.416** (0.20) (0.23) (0.15) PCA Shock 0.63** 0.63* 0.29 (0.29) (0.32) (0.39) BRW Tight(NS data) -0.15 -0.19 -0.17 (0.13) (0.13) (0.16) BRW Tight(full data) -0.14 -0.20 -0.28 (0.31) (0.30) (0.34) Observations 121 89 52 Note: Monthly change (current month to next) in Blue Chip survey expectations of output growth over the next 3 quarters regressed on the shock series in that month plus a constant (not displayed). Sample periods are listed at top. Robust standard error in brackets. * p < 0.10, ** p < 0.05, *** p < 0.01. BRW Shock refers to our shock series: our data and our PLS estimation procedure. SS Shock refers to the sum of the shock series of the Federal Funds rate, forward guidance and large scale asset purchases of Swanson (2018), scaled by 100. NS Shock refers to the policy news shocks of Nakamura and Steinsson (2018). BRW(JK Info days) and BRW(JK non-info days) are two sub-sample regressions focusing only on theFedinformationeffectdaysandnon-informationeffectdays,asdefinedbyJarocinskiandKaradi(2018). FF3(JKxdays)arethesurprisesofthe3-monthfederalfundsfuturesthatareusedbyJarocinskiandKaradi (2018). PCA Shock referstotheshockseriesgeneratedfromextractingthefirstprincipalcomponentofour underlying data, i.e., all outcome variables (daily changes of 1 to 30-year zero coupon rate around FOMC meeting). BRWTight(NSdata) referstotheBRWshockseriescomputedusingPLSwiththedatainNakamuraandSteinsson(2018),i.e.,the30-minutechangesofthecurrentmonthFedfundsfuturesrate,theFed funds futures rate immediately following the next FOMC meeting, and two, three, four quarter ahead euro dollar futures. BRW Tight(full data) refers to the BRW shock series using the NS data plus the 30-minute changesofthe3month,6month,2year,5year,10year,30yearinterestratesaroundFOMCannouncements. 43

Raw Shock Series Figures Figure C.1: BRW Shock Series and the Three Alternative Shock Series Note: The solid blue line represents the BRW shock series estimated from Equations (3) and (4). N&S Shock,theblackdottedline,referstothepolicyfactorshocksobtainedfromNakamuraandSteinsson (2018). Kuttner Shock, the solid black line, refers to the 30-minute fed funds rate changes around FOMC announcement obtained from Nakamura and Steinsson (2018). R&R Shock, which is the blue dashed line, refers to the estimated shock series in Romer and Romer (2004). 44

Figure C.2: BRW Shock Series & Swanson’s Shock Series Note: All navy bars are our BRW shock series estimated from Equations (3) and (4). Gray bars are benchmark shock series: SS FFR, SS FG, SS LSAP, and SS Sum, the shocks to the federal funds rate, forward guidance, large asset purchases, and the sum of the three shocks, all from Swanson (2018). 45

Figure C.3: BRW and NS Shock Series Note: All navy bars are in the graphs are our BRW shock series estimated from Equation (3) and (4). N&S Shock referstothepolicyfactorshocksobtainedfromNakamuraandSteinsson(2018),whichare extended to 2017m12. 46

Figure C.4: Rolling Sample 1969m1-2017m1 Note: rolling sample from 1969m1 to 2017m12, each of which has 15 years. 1 beta refers to the estimatedcoefficientfromusingthe1-yearTreasuryRateasmonetarypolicyindicator. 2 beta referstothe estimatedcoefficientfromusingthe2-yearTreasuryRateasmonetarypolicyindicator. 5 beta referstothe estimated coefficient from using the 5-year Treasury Rate as monetary policy indicator. 10 beta refers to the estimated coefficient from using the 10-year Treasury Rate as monetary policy indicator. 47

Figure C.5: Robustness Check: Influence of the Term Premium a. 1994m1-2017m12 b. 2008m1-2017m12 Note: Graphs show impulse responses to a 100 basis point increase in the cumulative BRW shock series. Deepandshallowgrayshadedareasare68%and90%confidenceintervalsproducedbybootstrapping 1000 times, respectively. 48

Figure C.6: SVAR Impulse Responses with alternative IV a. Sample period 1994m1-2017m12 b. Sample period 2008m1-2017m12 Note: Graphs show impulse responses to a 100 basis point increase in the cumulative shock series. Deep and shallow gray shaded areas are 68% and 90% confidence intervals produced by bootstrapping 100 times, respectively. 49

Figure C.7: SVAR Impulse Responses with Simple Fama-Macbeth Shock a. Sample period 1994m1-2017m12 b. Sample period 2008m1-2017m12 Note: AlternativeBRWshockseriesisalignedfromtheFama-MacbethprocedurewithoutIDH.The IRFs are estimated as above. 50

Figure C.8: SVAR Impulse Responses with PCA Shock a. Sample period 1994m1-2017m12 b. Sample period 2008m1-2017m12 Note: The PCA shock is constructed from applying the Nakamura-Steinsson estimation procedure to our data: extracting the first principal component of all BRW outcome variables (daily changes of 1 to 30-year zero coupon rate around FOMC announcement days). The IRFs are estimated using the same approach as above. 51

Figure C.9: SVAR Impulse Responses with Tight-window(NS data) Shock a. Sample period 1994m1-2017m12 b. Sample period 2008m1-2017m12 Note: The tight-window(NS data) shock is constructed from using the Nakamura-Steinsson (2018) data with our econometric procedure. The underlying data include the 30-minute changes of the current month Fed funds futures rate, the Fed funds futures rate immediately following the next FOMC meeting, andtwo,three,fourquarteraheadeurodollarfuturesaroundthecurrentFOMCannouncement. TheIRFs are estimated using the same approach as above. 52

Figure C.10: SVAR Impulse Responses with Tight-window(Full data) Shock a. Sample period 1994m1-2017m12 b. Sample period 2008m1-2017m12 Note: The tight-window shock is constructed using our econometric procedure with the Nakamura- Steinsson (2018) data plus some long term interest rate data.IRFs are estimated using the same approach as above. 53