The Market-Perceived Monetary Policy Rule

Board of Governors of the Federal Reserve System International Finance Discussion Papers Number 982 October 2009 The Market-Perceived Monetary Policy Rule James D. Hamilton Seth Pruitt Scott C. Borger NOTE: International Finance Discussion Papers are preliminary materials circulated to stimulate discussion and critical comment. References in publications to International Finance Discussion Papers (other than an acknowledgement that the writer has had access to unpublished material) should be cleared with the author or authors. Recent IFDPs are available on the Web at www.federalreserve.gov/pubs/ifdp/ . This paper can be downloaded without charge from Social Science Research Network electronic library at www.ssrn.com

The Market-Perceived Monetary Policy Rule∗ James D. Hamilton† Seth Pruitt‡ Scott C. Borger§ UCSD Federal Reserve Office of Board of Governors Immigration Statistics October 2009 Abstract We introduce a novel method for estimating a monetary policy rule using macroeconomic news. Market forecasts of both economic conditions and monetary policy are affected by news, and our estimation links the two effects. This enables us to estimate directly the policy rule agents use to form their expectations, and in so doing flexibly capture the particular dynamics of policy response. We find evidence that between 1994 and 2007 the market-perceived Federal Reserve policy rule changed: the output response vanished, and the inflation response path became more gradual but larger in long-run magnitude. In a standard model we show that outputsmoothingcausedbyalargerinflationresponsemagnitudeisoffsetbythemoremeasured pace of response. Our response coefficient estimates are robust to measurement and theoretical issues with both potential output and the inflation target. Keywords: monetary policy rule, market perceptions, Taylor Rule, Fed funds futures JEL codes: E43, E52, E58 ∗Firstversion: January,2009. WethankMicheleCavallo,JosephGruber,ElmarMertens,RicardoNunes,Robert Vigfusson,andparticipantsoftheInternationalFinanceWorkshop,MidwestMacroMeetings2009,andNorthAmerican Econometric Society Meetings 2009 for helpful comments, and Michael Reynolds for research assistance. The views in this paper are solely the responsibility of the authors and should not be interpreted as reflecting the views oftheBoardofGovernorsoftheFederalReserveSystemorofanyotherpersonassociatedwiththeFederalReserve System. †jhamilton@ucsd.edu ‡seth.j.pruitt@frb.gov §sborger@ucsd.edu

1 Introduction AlargeliteratureestimatesmonetarypolicyrulesoftheformproposedbyTaylor(1993)thatrelate the realized fed funds rate to past or expected future indicators of output and inflation. Examples include Evans (1998), Clarida, Gali, and Gertler (2000), Rudebusch (2002), Owyang and Ramey (2004), Boivin (2006), Ang, Dong, and Piazzesi (2007), and Smith and Taylor (forthcoming). That kind of estimation is well suited to describe what policy rule the Fed has actually followed. However, there is also considerable interest in what market participants expect the Fed to do. Expectations of future monetary policy are a key part of the monetary transmission mechanism in virtually any macroeconomic model. The Federal Reserve’s expected future policy rate influences current interest rates immediately upon the market learning about the Federal Reserve’s intentions to stimulate or curtail economic behavior (Hamilton (2008)). Moreover, Federal Open Market Committee (FOMC) statements provide guidance for the direction of future policy rates and are responded to instantaneously by the market upon their public release (Kohn and Sack (2004)). Thispaperproposesanovelmethodthatenablesustouncoverthemarket’sperceivedmonetary policy rule. Like many previous researchers (e.g., Gu¨rkaynak, Sack, and Swanson (2005), Faust, Rogers, Wang, and Wright (2007), and Bartolini, Goldberg, and Sacarny (2008)), we identify news bythedifferencebetweenamacroeconomicdatareleasevalueandthevalueexpectedbeforehandby themarket. Onthisnewsday,wemeasurethenews’effectsoneconomicfundamentals’forecastsand monetary policy forecasts, the latter coming from the change in market prices for fed funds futures contracts. Our contribution is to use a Taylor-Rule structure to link the fundamentals forecast updates with the policy forecast updates in order to estimate the market-perceived parameters for a Taylor Rule. Our methodology opens up to researchers the use of daily data, which offers three advantages we highlight. First, by looking at the response of fed funds futures prices for contracts of different horizons to a new data release, we are able to measure how long the market believes it will take the Fed to adjust interest rates in response to changing fundamentals. We can thereby obtain new measures of the nature of monetary policy inertia, something that is difficult for traditional methods to estimate.1 We document a change in the market’s perception of the Fed’s policy rule in terms of both the magnitude of the ultimate response and in the degree of inertia. Since 2000, the 1See Rudebusch (2002) and Rudebusch (2006) for alternative approaches. 1

market-perceivedmonetarypolicyruleinvolvesaneventualresponsetoinflationthatisbiggerthan that associated with perceived pre-2000 behavior. On the other hand, the market also believes that the Fed is more sluggish in making its intended adjustments. We show in simulations with a simple new-Keynesian model that the first feature would tend to stabilize output, whereas the second feature would be destabilizing. These simulations suggest that the “measured pace” of monetary tightening during 2004-2006 may have been counterproductive. Second, our approach offers a cleaner answer for how to handle real-time versus revised data sets, by focusing on market expectations formed on the basis of the information as it had actually been publicly released as of a particular calendar date. Third, our approach is robust to estimation problems engendered by potential output and the inflation target. Potential output is tricky to define and measure in real time (Orphanides and van Norden (2002)), and Orphanides (2001) argues that this can confound policy rule estimation. On the other hand, the Fed’s inflation target is unobservable, and moreover a growing literature, includingIreland(2007)andCogleyandSbordone(2008)amongothers, haspostulatedanimportant historical role for low-frequency variation in the Fed’s inflation target. The latency of potential output and the inflation target poses a problem for standard policy rule estimation methods because their values are necessary for measuring the explanatory variables. Our method uses daily data to difference out these slowly moving latent variables from the estimation equations. The remainder of the paper is structured as follows: Section 2 introduces our framework and its testable implications. Section 3 discusses the empirical strategy based on these implications and describes the data. Section 4 presents our baseline full sample results, and then shows evidence of time variation in perceived policy response and estimates parameters on subsamples. Section 5 generalizes the approach to estimation of a Taylor Rule with lagged adjustment dynamics and discussestheeconomicsignificanceofthosedynamics. Weinvestigatethesensitivityofourconclusions to various assumptions and variable decisions in Section 6. Section 7 concludes. 2 Framework We begin with a standard Taylor Rule that is assumed by the market to characterize Federal Reserve decisions. Let t represent a particular month and r the average daily effective fed funds t rate for that month. The market assumes that the Fed sets the funds rate in response to the Taylor 2

Rule variables π −π∗, the deviation from target of cumulative inflation between t−12 and t, and t t y −y∗, a measure of the real output gap in t: t t r = r+β(π −π∗)+δ(y −y∗)+u (2.1) t t t t t t where y is real output growth and y∗ is potential real output growth. t t We will be keeping careful track in this analysis of exactly when data of different sorts arrives. LetΩ denotetheinformationsetthatisactuallyavailabletomarketparticipantsasoftheithday i,t of month t; let Ω˜ denote the Fed’s information set at that time. The formulation (2.1) assumes i,t that the Fed knows the values of π −π∗ and y −y∗ at the time it sets r , even though π and t t t t t t y would not be known to market participants until some later time. The framework is readily t generalizable to a case where the Fed instead sets r on the basis of information available as of t some day j within month t: r t = r+βE (cid:16) π t (cid:12) (cid:12)Ω˜ j,t (cid:17) −βE (cid:16) π∗ t (cid:12) (cid:12)Ω˜ j,t (cid:17) +δE (cid:16) y t (cid:12) (cid:12)Ω˜ j,t (cid:17) −δE (cid:16) y t ∗(cid:12) (cid:12)Ω˜ j,t (cid:17) +u t . (2.2) Consider the expectation of (2.1) conditional on information available to the market as of the ith day of an earlier month τ = t−h: E(cid:0) r t (cid:12) (cid:12)Ω i,τ (cid:1) = r+βE(cid:0) π t (cid:12) (cid:12)Ω i,τ (cid:1) −βE(cid:0) π∗ t (cid:12) (cid:12)Ω i,τ (cid:1) +δE(cid:0) y t (cid:12) (cid:12)Ω i,τ (cid:1) −δE(cid:0) y t ∗(cid:12) (cid:12)Ω i,τ (cid:1) +E(cid:0) u t (cid:12) (cid:12)Ω i,τ (cid:1) . (2.3) Alternatively, if we take expectations of (2.2) conditional on the information set Ω , the identical i,τ equation (2.3) follows due to the Law of Iterated Expectations.2 In either case, we obtain the following expression for the change in expectations between the ith day and the previous day (i−1) of month τ: E(cid:0) r t (cid:12) (cid:12)Ω i,τ (cid:1) −E(cid:0) r t (cid:12) (cid:12)Ω i−1,τ (cid:1) (2.4) = β (cid:2)E(cid:0) π t (cid:12) (cid:12)Ω i,τ (cid:1) −E(cid:0) π t (cid:12) (cid:12)Ω i−1,τ (cid:1)(cid:3) +δ (cid:2)E(cid:0) y t (cid:12) (cid:12)Ω i,τ (cid:1) −E(cid:0) y t (cid:12) (cid:12)Ω i−1,τ (cid:1)(cid:3) −β (cid:2)E(cid:0) π∗ t (cid:12) (cid:12)Ω i,τ (cid:1) −E(cid:0) π∗ t (cid:12) (cid:12)Ω i−1,τ (cid:1)(cid:3) −δ (cid:2)E(cid:0) y t ∗(cid:12) (cid:12)Ω i,τ (cid:1) −E(cid:0) y t ∗(cid:12) (cid:12)Ω i−1,τ (cid:1)(cid:3) + (cid:2)E(cid:0) u t (cid:12) (cid:12)Ω i,τ (cid:1) −E(cid:0) u t (cid:12) (cid:12)Ω i−1,τ (cid:1)(cid:3) . Equation(2.4)isthekeytowhatfollows,statingthatupdatestothemarketforecastoffuturepolicy are linked to updates to the market forecast of future economic conditions via the market-perceived monetary policy rule. 2We assume that Ω ⊆Ω˜ . i,τ j,t 3

We will consider a set of k = 1,2,...,K different days within month τ on which particular information becomes available. Consider first k = 1, which we associate with the release of, say, the CPI. Let i(1,τ) denote the day in month τ on which a new inflation number (namely, the value of π ) is released. For example, for τ = December 2008, the CPI data reported on December 16 τ−1 (i(1,τ) = 16) was the value for November 2008 (so that π became known on i(1,τ)). Consider τ−1 then the initial report of the value of π on day i(1,τ). We will proxy the news content of this τ−1 report by comparing the actual value π with the value expected by the market, which we denote τ−1 π˜ : τ−1 E(cid:0) π τ−1 (cid:12) (cid:12)Ω i(1,τ),τ (cid:1) −E(cid:0) π τ−1 (cid:12) (cid:12)Ω i(1,τ)−1,τ (cid:1) = π τ−1 −π˜ τ−1 . The CPI announcement of π (arriving on i(1,τ)) has an implication for what market partic- τ−1 ipants would have expected π to turn out to be. We propose to model this implication by a linear t equation forecasting π on the basis of π π˜ , and x , where x denotes a vector of other t τ−1, τ−1 1,τ 1,τ variables that would have been known to market participants prior to the day i(1,τ) of month τ: π = γ π +ξ π˜ +ζ(cid:48) x +v . (2.5) t π,1 τ−1 π,1 τ−1 π,1 1,τ π,1,t The first subscript (π) on the coefficients indicates that this is a coefficient used to forecast subsequent inflation, and the second subscript (1) indicates that the forecast is formed on the day on which the first information variable (the CPI) is released. Note that the coefficients in equation (2.5) are defined as linear projection coefficients, so that v is uncorrelated with π , π˜ , and π,1,t τ−1 τ−1 x by the definition of γ , ξ , and ζ(cid:48) . The consequences of the month τ, day i(1,τ) news 1,τ π,1 π,1 π,1 release about π for market expectations of π are then given by τ−1 t E(cid:0) π t (cid:12) (cid:12)Ω i(1),τ (cid:1) −E(cid:0) π t (cid:12) (cid:12)Ω i(1)−1,τ (cid:1) = γ π,1 (π τ−1 −π˜ τ−1 ) (2.6) where we will subsume the dependence of i(1,τ) on τ when it is clear from the context. Theannouncementofπ mayalsoholdimplicationsformarketexpectationsaboutrealoutput τ−1 y , which we proxy analogously as t y = γ π +ξ π˜ +ζ(cid:48) x +v t y,1 τ−1 y,1 τ−1 y,1 1,τ y,1,t E(cid:0) y t (cid:12) (cid:12)Ω i(1),τ (cid:1) −E(cid:0) y t (cid:12) (cid:12)Ω i(1)−1,τ (cid:1) = γ y,1 (π τ−1 −π˜ τ−1 ). (2.7) Note that certain elements of ζ(cid:48) and ζ(cid:48) may be set to zero, depending on what elements of x π,1 y,1 1,τ forecast π or y . t t 4

(h) Let f denote the futures interest rate on day j of month τ for a fed funds futures contract jτ based on r , the effective fed funds rate h months ahead. We propose that these fed funds futures t offer us a direct observation on how the market expectation of r changed on day i(1): t f i ( ( h 1 ) ),τ −f i ( ( h 1 ) )−1,τ = E(cid:0) r t (cid:12) (cid:12)Ω i(1),τ (cid:1) −E(cid:0) r t (cid:12) (cid:12)Ω i(1)−1,τ (cid:1) +η r,1 +q r,1,τ . (2.8) Here η captures the average change in the risk premium on fed funds futures contracts and q r,1 r,1,τ any change in the risk premium relative to that average. In the absence of risk aversion in the fed funds futures markets, the terms η and q would be identically zero. There is certainly r,1 r,1,τ good evidence for supposing the contribution of risk aversion to daily changes in fed funds prices to be small; see Piazzesi and Swanson (2008) and Hamilton (2009).3 In the estimation strategy adopted here, any changes in the risk premium, along with changes in the market’s expectation of the residual in the Taylor Rule, changes in the market’s expectation of the inflation target, and changesinthemarket’sexpectationofpotentialoutputgrowth,areincorporatedintoaspecification error v , r,1,τ v r,1,τ = −δ (cid:2)E(cid:0) y t ∗(cid:12) (cid:12)Ω i(1),τ (cid:1) −E(cid:0) y t ∗(cid:12) (cid:12)Ω i(1)−1,τ (cid:1)(cid:3) −β (cid:2)E(cid:0) π∗ t (cid:12) (cid:12)Ω i(1),τ (cid:1) −E(cid:0) π∗ t (cid:12) (cid:12)Ω i(1)−1,τ (cid:1)(cid:3) (2.9) + (cid:2)E(cid:0) u t (cid:12) (cid:12)Ω i(1),τ (cid:1) −E(cid:0) u t (cid:12) (cid:12)Ω i(1)−1,τ (cid:1)(cid:3) +q r,1,τ . Substituting (2.6), (2.7), (2.8), and (2.9) into (2.4), we have (h) (h) f −f = η +(βγ +δγ )(π −π˜ )+v . i(1),τ i(1)−1,τ r,1 π,1 y,1 τ−1 τ−1 r,1,τ Considernextasecondnewsreleaseinmonthτ, namelytherealactivityindicatory released τ−1 on day i(2). For these days we employ the auxiliary forecasting equations π = γ y +ξ y˜ +ζ(cid:48) x +v t π,2 τ−1 π,2 τ−1 π,2 2,τ π,2,t y = γ y +ξ y˜ +ζ(cid:48) x +v t y,2 τ−1 y,2 τ−1 y,2 2,τ y,2,t where x is known prior to day i(2,τ). From these we derive 2,τ (h) (h) f −f = η +(βγ +δγ )(y −y˜ )+v . i(2),τ i(2)−1,τ r,2 π,2 y,2 τ−1 τ−1 r,2,τ 3Our method works if either the risk premium is constant, as implied by the common “expectations hypothesis” or under the implication of consumption-based asset pricing models that the risk premium would change little on a daily basis. Piazzesi and Swanson’s (2008) results indicate that “[these] risk premia seem to change primarily at business-cycle frequencies.” 5

In general, if some indicator w is released on day i(k,τ), we have the following three k,τ−1 equations: π = γ w +ξ w˜ +ζ(cid:48) x +v (2.10) t π,k k,τ−1 π,k k,τ−1 π,k k,τ π,k,t y = γ w +ξ w˜ +ζ(cid:48) x +v (2.11) t y,k k,τ−1 y,k k,τ−1 y,k k,τ y,k,t (h) (h) f −f = η +(βγ +δγ )(w −w˜ )+v . (2.12) i(k),τ i(k)−1,τ r,k π,k y,k k,τ−1 k,τ−1 r,k,τ Let z = (1,π ,π˜ ,x(cid:48) )(cid:48) denote the vector including the day i(1) release of π and the 1,τ τ−1 τ−1 1,τ τ−1 information available as of the day before, where we assume that z is uncorrelated with v , 1,τ π,1,t v , and v . Similarly, we take z = (1,w ,w˜ ,x(cid:48) )(cid:48) to be uncorrelated with v , y,1,t r,1,τ k,τ k,τ−1 k,τ−1 k,τ π,k,t v , and v , for k = 1,2,...,K. Thus our identifying assumption is that the following vector y,k,t r,k,τ has expectation zero:  (cid:0) π −γ w −ξ w˜ −ζ(cid:48) x (cid:1) z  t π,1 1,τ−1 π,1 1,τ−1 π,1 1,τ 1,τ  (cid:0) y t −γ y,1 w 1,τ−1 −ξ y,1 w˜ 1,τ−1 −ζ(cid:48) y,1 x 1,τ (cid:1) z 1,τ   (cid:104) (cid:105)   f (h) −f (h) −η −(βγ +δγ )(w −w˜ ) z   i(1),τ i(1)−1,τ r,1 π,1 y,1 1,τ−1 1,τ−1 1,τ   .    . .   . (2.13)   (cid:0) π t −γ π,K w K,τ−1 −ξ π,K w˜ K,τ−1 −ζ(cid:48) π,K x K,τ (cid:1) z K,τ      (cid:104) (cid:0) y t −γ y,K w K,τ−1 −ξ y,K w˜ K,τ−1 −ζ(cid:48) y,K x K,τ (cid:1) z K,τ (cid:105)    (h) (h) f −f −η −(βγ +δγ )(w −w˜ ) z i(K),τ i(K)−1,τ r,k π,K y,K K,τ−1 K,τ−1 K,τ Note that the ability to distinguish β from δ results from using at least K ≥ 2 different news releases during month τ. A single release such as the inflation number could in principle have implications both for future inflation (as captured by γ ) and future output (as captured by γ ). π,1 y,1 Hence any response of the fed funds futures prices to that news could come from either the policy rule inflation coefficient (β) or output coefficient (δ). However, γ and γ are each separately π,1 y,1 observable (from the differing responses of π and y to π ), so the change in the futures price t t τ−1 on i(1) tells us one linear combination (namely βγ +δγ ) of the policy rule parameters β and π,1 y,1 δ. But the separate response to the output release on day i(2) gives us a second linear combination (βγ +δγ ). Thus, the 3K equations above are sufficient to identify β and δ separately. π,2 y,2 For each month τ there are K days of interest, for K the number of economic indicators under consideration. Identification of this system is achieved so long as it is not the case that any one indicator always arrives on the same day as another indicator.4 Of course, it is all right for any two indicators occasionally to arrive on the same day. 4We make sure this is the case with the indicators we choose below. 6

3 Estimation We begin this section by describing the formal estimation strategy, which is Hansen’s (1982) generalized method of moments. Then we describe the data used. 3.1 Method Recall that τ +h = t. Denoting (cid:16) (cid:17)(cid:48) ζ (h) = 1,π ,y ,f (h) ,w ,w˜ ,x(cid:48) ,z(cid:48) ,...,f (h) ,w ,w˜ ,x(cid:48) ,z(cid:48) , t t t i(1),τ 1,τ−1 1,τ−1 1,τ 1,τ i(K),τ K,τ−1 K,τ−1 K,τ K,τ we rephrase (2.13) as the following population orthogonality condition for each θ(h), h = 1,2,..., (cid:104) (cid:16) (cid:17)(cid:105) E g θ(h),ζ (h) = 0, (3.1) t where θ collects the auxiliary forecasting parameters (γ(cid:48),ξ(cid:48),ζ(cid:48))(cid:48) along with the main parameters of (cid:16) (cid:17)(cid:48) interest, the policy rule coefficients (β,δ,η(cid:48))(cid:48). Let Y (h) ≡ ζ (h)(cid:48) ,ζ (h)(cid:48) ,...,ζ (h)(cid:48) be the vector of r T T T−1 1 all observations for each choice of horizon h. Then we have the sample average T g (cid:16) θ(h);Y (h) (cid:17) ≡ T−1 (cid:88) g (cid:16) θ(h),ζ (h) (cid:17) T t t=1 and the GMM estimator (Hansen (1982)) for each horizon h minimizes (cid:16) (cid:17) (cid:16) (cid:17)(cid:48) (cid:16) (cid:17) Q θ(h),Y (h) = g θ(h);Y (h) W (h) g θ(h);Y (h) . (3.2) T T T T (h) As usual, the optimal weighting matrix W is given by the inverse of the asymptotic variance of T (cid:16) (cid:17) the sample mean of g θ(h),ζ (h) . In turn, we calculate a heteroskedasticity and autocorrelation t robust estimate Sˆ(h) of this asymptotic variance, and the efficient GMM estimator uses the inverse T of this HAC estimate as the weighting matrix, with the following asymptotic approximations: θˆ(h) ≈ N (cid:16) θ(h),T−1Vˆ(h) (cid:17) , Vˆ(h) = (cid:16) [Dˆ(h) ][Sˆ(h) ]−1[Dˆ(h) ](cid:48) (cid:17)−1 T T T T T (cid:16) (cid:17) (h) ∂g θ;Y (cid:12) and [Dˆ(h) ](cid:48) = T (cid:12) (cid:12) . T ∂θ(cid:48) (cid:12) θ=θˆ(h) Since g(·) is nonlinear in θ(h), the minimization of (3.2) is achieved numerically. Our results are calculated by two-step GMM starting from an initial guess provided by a simple two-stage OLS procedure and with other initial conditions considered to obtain some assurance that the global 7

optimum has been found.5 The inconsistent two-stage OLS procedure would instead first estimate the auxiliary forecasting equations independently, then use these forecast parameter estimates to generate regressors for the Taylor Rule regression.6 Joint estimation by (nonlinear) two-step GMM is consistent and efficient – see Newey and McFadden (1994).7 We estimate each horizon h independently from the others so that nothing other than the original data links these estimates to one another. As mentioned, identification is achieved by considering at least two indicators w, in which case thesystem(3.2)ingeneralisjust-identified. Whenweusemorethantwoindicators,thesystemnaturally delivers overidentifying restrictions. Additionally, we can impose cross-equation restrictions that create overidentification. Our baseline specification is overidentified for both reasons. 3.2 Data OurdatasetconsistsofK particulardaysforeachmonthovertheperiod1994:M1through2007:M6. Data on fed funds futures contracts come from the Chicago Board of Trade. Fed funds futures are accurate predictors of the effective fed funds rate, as documented in numerous studies including Evans (1998), Gu¨rkaynak, Sack, and Swanson (2007), Piazzesi and Swanson (2008), and Hamilton (2009). We restrict our attention to fed funds futures traded after 1994. One reason, as noted by Gu¨rkaynak, Sack, and Swanson (2007), is that the Federal Open Market Committee began announcing the fed funds target in 1994, and this change in procedure could cause changes in the forecasting relations. In addition, the trading volumes pick up noticeably during this year.8 At the other end of the sample period, we end our data at 2007:M7 in order to avoid the period of major financial disruptions that started in the late summer of 2007 following the fund freezes by BNP Paribas in early August. We measure inflation by the year-over-year growth rate of the Core-PCE price index from the BEA. This has been the Federal Reserve’s key inflation indicator over the sample we consider. We measure output growth by the year-over-year growth rate of industrial production from the Federal Reserve Board. To use as much data as possible we stay at the monthly frequency and 5We have tried Continually-Updated GMM, but have found this makes little difference to our baseline results. 6Since our framework introduces a generated-regressor, the two-stage OLS procedure is inconsistent – see Pagan (1986). 7Our HAC estimator is that of Newey and West (1987) with 13 lags. 8See Figure A1 for a plot of these data. 8

therefore require a monthly output growth series. Industrial production growth has been used by previous studies to proxy for overall output growth (e.g. Stock and Watson (2002)) and is a natural candidate for our baseline. As a robustness check we will consider another measure for output growth, Macroeconomic Advisers’ monthly GDP. Theeconomicindicatorsweconsideraredatareleasesfromvariousgovernmentagenciesthatare followedbytheMoneyMarketSurvey(MMS).FollowingGu¨rkaynak,Sack,andSwanson(2005),the median forecast provides a proxy for each variable’s market expectation. MMS provides market expectations for several candidate economic indicators. Our choice is guided by asking which economic variables might be most helpful for forecasting output growth and core PCE inflation. It is natural for this purpose to use core CPI inflation (CPIXFE) and industrial production (INDPRD) themselves.9 In addition, the macroeconomic announcement literature has noted that market participantsscrutinizeandrespondtononfarmpayrollemployment(Gu¨rkaynak,Sack,andSwanson (2005), Bartolini, Goldberg, and Sacarny (2008)), and so we will consider that indicator (NFPAY) as well. It is worth noting that NFPAY and the unemployment rate are released on the same day eachmonth,intheBLSEmploymentSituationreport. Asmentionedabove,thisimpliesthateither one of these, but not both, can be used in estimation. Given the importance of NFPAY found by prior studies, this steers us away from the unemployment rate as an indicator. This in turn makes the unemployment rate less attractive a proxy for output growth, since we would naturally then use it as an indicator. In terms of the variables entering the auxiliary forecasting equations, we set (cid:16) (cid:17)(cid:48) (h) x = π ,y ,f ,1 . k,τ τ−2 τ−2 i(k)−1,τ The lagged values of inflation and output growth are included to control for their autoregressive nature. Forparsimony,wesettozerothefirstelementofζ ,thecoefficientonπ inindicatork’s y,k τ−2 auxiliaryforecastingequationfor y ; likewise, wezerooutthesecondelementofζ , thecoefficient t π,k on y in indicator k’s auxiliary forecasting equation for π . The fed funds futures prediction for τ−2 t the day before i(k) is included to control for the predictive content (vis-a-vis each Taylor Rule variable) of the futures price that has already been priced into the contract. 9MMS does not survey forecasts for Core-PCE inflation, hence our reliance on Core-CPI inflation. Fortunately, Core-CPI forecasts Core-PCE inflation well – further details are available from the authors upon request. 9

4 Results First we present our baseline full-sample results using three indicators. We then show that statistical tests of our overidentifying restrictions fail to reject our baseline model, along with other specifications considered for robustness. Motivated by related literature, we run tests for breaks in the policy rule parameters and find evidence of their variation over time. Placing the break around the beginning of the year 2000, we present subsample estimates suggesting the market-perceived monetary policy rule has changed over time, and repeat the overidentification tests on the separate subsamples. 4.1 Baseline Our baseline results use three indicators – CPIXFE, INDPRD, and NFPAY – and impose the crossequation restriction that the average risk premium change is identical across indicators: η = η ,k = 1,2,...,K. (4.1) r,k r This cross-equation restriction embodies the assumption that the different economic indicators systematically affect the forecasted policy rate only through changes to forecasted inflation and output, anditaddsstatisticalprecisiontoourestimates; wefurtherdiscussandtestthisrestriction in Section 6. The policy rule response coefficient estimates are presented in Table 1, and we note a few features deserving mention.10 Table 1: Market-Perceived Monetary Policy Rule Estimates, baseline h 1 2 3 4 5 6 β 0.3423 0.8723** 1.3487 1.1114** 1.1068* 1.3733 0.2046 0.1496 0.7757 0.4305 0.5118 0.7252 δ 0.0510** 0.0279 0.1603** 0.1258** 0.1429** 0.1938** 0.0092 0.0204 0.0361 0.0306 0.0438 0.0702 Notes: The policy rule coefficient on inflation is β and on the output gap is δ. HAC standard errors in italics. The markers * and ** denote significance at 5% and 1% levels, respectively. There are 160 observations for h = 1, 159 forh=2,etc. TheindicatorsareCPIXFE,INDPRD,andNFPAY.Pointestimatesandstandarderrorsfromtwo-step nonlinear GMM. Data run over 1994:M1-2007:M7. 10Estimates of the constant are reported in Table A1 of Appendix A.1. 10

First, we obtain reasonably precise estimates of the market-perceived policy response to inflation. Horizonstwo, four, andfiveallexhibitinflationresponsecoefficientsthataresignificantatthe 5% level, and coefficients for the remaining horizons are significant at the 10% level. The output response coefficient is statistically significant and positive at the 1% level for all horizons except the second. These results suggest that our empirical methodology effectively extracts information from market forecast updates that occur in response to macroeconomic news. Second, the market does not expect the Fed to implement changes immediately. The response coefficients at longer horizons tend to be larger than the response coefficients at shorter horizons, and 95% confidence intervals for β or δ often exclude the point estimates obtained for different h. Recall that the units of the inflation response coefficients are identical across horizons, as are the coefficients on output. These parameters answer the question: looking h months ahead, what is the response of the forecasted policy rate to a one unit increase in the forecasted rate of inflation or output growth? This feature of the results strongly suggests that the market believes the Fed graduallyadjustspolicyinresponsetoeconomicfundamentals,apointweexplorefurtherinSection 5. 4.2 Overidentification and Break Tests We next evaluate the appropriateness of the assumptions behind these estimates. We first investigate Hansen’s (1982) J-tests of overidentifying restrictions given by TQ (cid:16) θˆ(h) ,Y (h) (cid:17) ≈ χ2(m) (4.2) T for m the number of overidentifying restrictions. The p-values for this test are presented in Table 2. Recall that our baseline specification overidentifies the model both by using three indicators and by imposing that the policy rule specification error means are identical for these indicators (equation (4.1)). Row 1 displays the p-values associated with the J-statistics for the baseline specification. We fail to reject at the 5% level the overidentifying restrictions for every horizon h, offering some confirmation that our basic framework is consistent with the data. Papers including Clarida, Gali, and Gertler (2000), Primiceri (2006), and Boivin (2006) have argued that U.S. monetary policy has changed over time. Unfortunately, our data are not available for the period over which those papers find the most dramatic policy changes. Nonetheless, if monetary policy changed once, then it could change again – and market participants are aware of 11

Table 2: Overidentification Tests, baseline h 1 2 3 4 5 6 (1) Baseline 0.1176 0.1172 0.1099 0.1006 0.1222 0.1044 (2) Baseline, pre 0.3657 0.4600 0.3881 0.3827 0.4038 0.4068 (3) Baseline, post 0.2790 0.2879 0.2756 0.2963 0.3018 0.3045 Notes: p-valuesfromHansen’s(1982)J-testofoveridentifyingrestrictions,forthebaselinespecifications. Baseline is the baseline specification estimated over the full sample. Baseline, pre and Baseline, post are the baseline specifications estimated over the pre-2000 and post-2000 subsamples, respectively. this possibility. We therefore ask whether market participants’ perception of the monetary policy rule have changed over time. To answer this question, we test for a break in the parameters of interest. Using Andrews’s (1993)breaktest,wetestthenullhypothesisthatallparametersareconstantagainstthealternative thatthepolicyrulecoefficientsβ,δ,andη experiencedabreak.11 Lettingthepolicyrulecoefficient r vector be b = (β,δ,η )(cid:48), we test: r H : b = b ∀t ≥ 1 for some b ∈ R3 0 t 0 0 (cid:26) (cid:27) b ((cid:36)) for t = 1,...,T(cid:36) H ((cid:36)) : b = 1 for some constants b ((cid:36)),b ((cid:36)) ∈ R3 1 t b ((cid:36)) for t = T(cid:36)+1,...,T 1 2 2 for values of (cid:36) in (0.25,0.75). We use the sup-Wald statistic and tabulated critical values in Andrews (1993). For each horizon considered, there is strong evidence of a break in the policy rule coefficients b. In particular, for our application the 1% critical value is 16.6: the sup-Wald statistic is estimated to be 204.3, 190.2, 170.6, 168.4, 142.0, and 123.1 for horizons 1 through 6, respectively. Moreover, these maximal statistics occur at nearly the same time for every horizon, at the beginning of the year2000. Inlightofthisevidence,were-estimateourbaselinemodelonthepre-2000andpost-2000 subsamples. Returning to the overidentification test results of Table 2, rows 2 and 3 display the p-values for the model estimated across horizons on each subsample. We now find that the model is readily accepted for each subsample. Evidently, the break in the policy parameters was a factor in the 11Thevectorofparameterstakentobeconstantunderboththenullandthealternativeisthevectorofauxilliary forecasting parameters (γ(cid:48),ξ(cid:48),ζ(cid:48))(cid:48). 12

Table 3: Market-Perceived Monetary Policy Rule Estimates, baseline pre-2000 and post-2000 h 1 2 3 4 5 6 Pre-2000 β 0.3611 0.8309* 1.2049* 1.2125 1.1116* 1.3821 0.2173 0.3717 0.4708 0.7954 0.4630 0.7610 δ 0.1550** 0.1284** 0.2641** 0.3968** 0.3844** 0.5744** 0.0349 0.0314 0.0670 0.1142 0.1364 0.1436 post-2000 β 0.1465** 0.3797** 0.4292* 0.5913** 0.8237** 2.2240* 0.0537 0.0673 0.2110 0.1318 0.1694 0.9529 δ 0.0366** 0.0034 -0.0474 -0.0544* 0.0401 -0.0777 0.0127 0.0071 0.0248 0.0277 0.0297 0.0834 Notes: The policy rule coefficient on inflation is β and on the output gap is δ. HAC standard errors in italics. The markers*and**denotesignificanceat5%and1%levels,respectively. pre-2000,thereare69observationsforh=1, 68 for h = 2, etc.; post-2000, there are 88 observations for h = 1, etc. The indicators are CPIXFE, INDPRD, and NFPAY. Point estimates and standard errors from two-step nonlinear GMM. Data run over 1994:M1-2007:M7 lower full sample p-values in row 1. Once the parameters are allowed to differ by sub-period, we find no evidence against our framework. 4.3 Time Variation Table 3 displays the estimation results for the two subsamples. We now discuss the output and inflation response coefficients estimated in each subsample and how they differ from one another. Looking at the output response coefficients, the output response during the 1990s is moderate but tightly estimated. At all horizons the point estimates are positive and significant at the 1% level. The response is around 0.15 in the first month, rising to 0.57 by the sixth month, and hence the policy response exhibits gradual adjustment. However, during the 2000s the response to output changes dramatically. For half of the horizons the output response is a tightly estimated zero, while for the remaining horizons the response is small and insignificant. Taken together, this evidence suggests that during the 1990s the market perceived a moderate output gap response that essentially vanished during the 2000s. Lookingnowattheinflationresponses,weseetwonoteworthydifferencesacrossthesubsamples. First, the pre-2000 estimates are not as precisely estimated as the post-2000 estimates. Prior to 13

the year 2000 none are significant at 1% level. On the other hand, post-2000 all horizons are statistically significant at the 5% level, with four horizons significant at the 1% level. This suggests thatthemorerecentperiodhasseengreatersignal, relativetonoise, inthemarketforecastupdates to policy and fundamentals. Second,thepost-2000inflationresponseisloweratshorthorizonsandhigheratthelonghorizon. For the first four months after an inflation forecast increase, the forecasted policy response during the 2000s is half what it was during the 1990s. For these early horizons we can reject at the 1% level that the post-2000 inflation responses are equal to their pre-2000 estimated values. At the long horizon, the post-2000 six-month horizon response (2.22) is about 50% greater than the pre- 2000 six-month horizon response (1.38). During the 1990s, policy followed the Taylor Principle (a more-than one-for-one response of nominal rates to inflation) by the third month after a shock to forecasted inflation; during the 2000s, policy has met this principle only by the sixth month. Together, these observations suggest that the market-perceived policy response to inflation changed over time in two distinct ways: during the 1990s the response adjusted at a quicker pace withamoderatelong-runmagnitude, whileduringthe2000stheresponseadjustedataslowerpace with a larger long-run magnitude. 5 Dynamic Analysis of the Policy Response UptothispointinthepaperwehavebeeninvestigatingastaticTaylorRuleoftheformofequation (2.1). WefoundthattheimpliedmarketexpectationsofhowtheFedwouldrespondtonewsturned outtobeafunctionofthetimehorizonh,meaningthattheFedisimplicitlyassumedbythemarket to implement the policy changes warranted by the news only gradually. We next formulate a more detailed specification of the nature of that lagged response that is consistent with the observed market behavior, developing and calibrating a dynamic Taylor rule. Section 5.3 then explores the implications of these changed dynamics using a simple new-Keynesian model. 5.1 Dynamic Forecasting Equations We first modify the earlier notation to make the dependence on the horizon h explicit, rewriting the h-period-ahead forecasting equations (2.11) and (2.10) as (h) (h) (h)(cid:48) (h) y = γ w +ξ w˜ +ζ x +v (5.1) t y,k k,t−h y,k k,t−h y,k k,t−h+1 y,k,t 14

(h) (h) (h)(cid:48) (h) π = γ w +ξ w˜ +ζ x +v . (5.2) t π,k k,t−h π,k k,t−h π,k k,t−h+1 π,k,t We will also now need a version of equations (5.1) and (5.2) for the case h = 0, in order to keep track of the implication of the release of one indicator for the values of other indicators to be released later that month. Suppose that the first indicator released in month t+1 is NFPAY, denoted here as w . That release could cause us to update our expectation of the values for 1,t INDPRD (y = w ) and CPIXFE (π = w ) that will be reported later that same month t+1 t 2,t t 3,t according to (0) (0) (0)(cid:48) (0) y = γ w +ξ w˜ +ζ x +v (5.3) t y,1 1,t y,1 1,t y,1 1,t+1 y,1,t (0) (0) (0)(cid:48) (0) π = γ w +ξ w˜ +ζ x +v . (5.4) t π,1 1,t π,1 1,t π,1 1,t+1 π,1,t (0) (0) Thusforexampleestimatesofγ andγ couldbeobtainedbyOLSestimationof(5.3)and(5.4). y,1 π,1 Later in month t+1 when the output indicator w is released, that allows us to know the value 2,t (0) of y with certainty, which to preserve the general notation we would represent by γ = 1, and t y,2 would also induce an update to the forecast for w , 3,t (0) (0) (0)(cid:48) (0) π = γ w +ξ w˜ +ζ x +v (5.5) t π,2 2,t π,2 2,t π,2 2,t+1 π,2,t (0) When w is finally released, it has no implications for w which is already known (γ = 0) and 3,t 2,t y,3 (0) changes our forecast of inflation one-for-one (γ = 1). π,3 5.2 A Dynamic Taylor Rule Consider now the following dynamic generalization of (2.1): r = r+β (π −π∗ )+β (π −π∗ )+··· +δ (y −y∗ )+δ (y −y∗ )+···+u . (5.6) t 1 t−1 t−1 2 t−2 t−2 1 t−1 t−1 2 t−2 t−2 t Unlike our earlier expression (2.2), equation (5.6) is strictly a backward-looking formulation, presuming that the Fed responds dynamically to the history of available information; note that π t−1 and y are the most recent values available as of the end of month t. t−1 (h) Recall that the value of w is released on day i(k,t − h), and let f denote the k,t−h−1 i(k),t−h interest rate implied by a futures contract for settlement based on the value of r , and quoted as of t (0) the end of trading on day i(k,t−h). For example, f would reflect an expectation of the current i(k),t month’sfedfundsrateonthedaythattheindicatorw isreleased. Taketheexpectationof(5.6) k,t−1 15

conditional on market information available on day i(k,t−h) and subtract from it the expectation formed the day before: (h) (h) (h) f −f = η + (5.7) i(k),t−h i(k)−1,t−h r,k (cid:104) (cid:105) (h) (h) (h−1) (h−1) (0) (0) (h) β γ +δ γ +β γ +δ γ +···+β γ +δ γ (w −w˜ )+v . 1 π,k 1 y,k 2 π,k 2 y,k h+1 π,k h+1 y,k k,t−h−1 k,t−h−1 r,k,t−h For comparison, recalling that τ = t−h, we can rewrite equation (2.12) as f (h) −f (h) = η (h) +(β(h)γ (h) +δ(h)γ (h) )(w −w˜ )+v (h) (5.8) i(k),t−h i(k)−1,t−h r,k π,k y,k k,t−h−1 k,t−h−1 r,k,t−h where β(h) and δ(h) denote the original parameters whose estimates we reported in column h of Tables 1 or 3. Comparing equations (5.7) and (5.8), the values of the dynamic parameters {β ,δ } j j in (5.6) are related to our baseline estimates {β(h),δ(h)} according to β γ (h) +δ γ (h) +β γ (h−1) +δ γ (h−1) +···+β γ (0) +δ γ (0) = β(h)γ (h) +δ(h)γ (h) . (5.9) 1 π,k 1 y,k 2 π,k 2 y,k h+1 π,k h+1 y,k π,k y,k To arrive at estimates of the dynamic parameters, we chose {β ,δ }6 so as to minimize the j j j=1 equally-weighted sum of squared differences between the LHS and RHS of (5.9) across indicators k = 1,2,3 and horizons h = 0,1,2,...,6. On the RHS, the values for {β(h),δ(h),γ (h) ,γ (h) } for π,k y,k h = 1,...,6 were taken from the earlier split-sample GMM estimation reported in Table 3, while values for h = 0 were obtained from GMM estimation of β(0), δ(0), γ (0) , γ (0) , and γ (0) based on the y,1 π,1 π,2 moment conditions  (cid:16) (cid:17)  (0) (0) (0)(cid:48) y −γ w −ξ w˜ −ζ x z t y,1 1,t y,1 1,t y,1 1,t+1 1,t+1  (cid:16) (cid:17)   π −γ (0) w −ξ (0) w˜ −ζ (0)(cid:48) x z   t π,1 1,t π,1 1,t π,1 1,t+1 1,t+1   (cid:104) (cid:105)   f (0) −f (0) −η(0)−(β(0)γ (0) +δ(0)γ (0) )(w −w˜ ) z   i(1),t+1 i(1)−1,t+1 π,1 y,1 1,t 1,t 1,t+1   (cid:16) (cid:17)  (5.10)  π −γ (0) w −ξ (0) w˜ −ζ (0)(cid:48) x z   t π,2 2,t π,2 2,t π,2 2,t+1 2,t+1   (cid:104) (cid:105)   f (0) −f (0) −η(0)−(β(0)γ (0) +δ(0)γ (0) )(w −w˜ ) z   i(2),t+1 i(2)−1,t+1 π,2 y,2 2,t 2,t 2,t+1   (cid:104) (cid:105)  f (0) −f (0) −η(0)−(β(0)γ (0) +δ(0)γ (0) )(w −w˜ ) z i(3),t+1 i(3)−1,t+1 π,3 y,3 3,t 3,t 3,t+1 whereasbeforez denotesinformationavailablethedaypriortoreleaseofw . ThislastGMM k,t+1 k,t estimation resulted in the estimates βˆ(0) = 0.3629,ˆδ (0) = 0.1745 for the pre-2000 subsample, and βˆ(0) = .0424,ˆδ (0) = 0.0031 after 2000. For all the above calculations, the values γ (0) = 1, γ (0) = 0, y,2 y,3 (0) and γ = 1 were imposed throughout. π,3 The resulting values of β and δ are reported in Table 4. In the last column is the sum of the j j 16

Table 4: Dynamic Taylor Rule Parameters j 1 2 3 4 5 6 7 sum Pre-2000 β 0.3629 –0.0009 0.8859 –0.0994 0.2395 –0.0542 –0.0464 1.2874 j δ 0.1760 –0.0102 –0.0147 0.0399 0.1692 –0.0674 –0.0343 0.2585 j post-2000 β 0.0848 0.0830 0.3642 0.2760 0.4918 –0.1552 0.4330 1.5780 j δ 0.0060 0.0140 –0.0040 –0.0414 0.0442 0.0302 0.0102 0.0592 j Notes: from minimum-distancemethoddescribedin text, usingsubsample parameter estimatesacross allhorizons. parameter values across all j, which gives the long-run response to the inflation or output pressure. Recall from Section 3.1 that the parameter vector θ(h) for horizon h was estimated completely independentlyfromanyotherhorizon. Thisapproachofleavingthedynamicsimpliedby{θ(h)}6 h=0 completely unrestricted offers at least two benefits. First, nothing in our procedure requires that the long-horizon responses should be bigger than the short-horizon responses. The fact that we nonetheless find them to be increasing in h is strong evidence that the market perceives policy to respond only gradually to changing conditions. Second, our procedure allows the adjustment to inflationary pressures to differ from the adjustment to real activity, similar to the policy rules of Christiano, Eichenbaum, and Evans (1996, 2005). Table 4 implies different paths in the response of policy to inflation and output. For example, the pre-2000 response to inflation jumps up at the three-month lag (j = 3) while the response to output stays relatively steady until the 5-month lag (j = 5). This flexibility in the rule’s process is greater than that permitted by including only lags of the policy rate itself, and our estimates suggest this greater flexibility is warranted by the data. 5.3 Implications of changes in the dynamics We now explore the implications of the estimated changes in the Taylor Rule for the consequences of monetary policy. Following Clarida, Gali, and Gertler (2000), we use a standard sticky-price, rational expectations model whose equilibrium conditions, log-linearized around a zero inflation 17

steady state, are π = λ E (π )+λ (y −z ) (5.11) t 1 t t+1 2 t t y = E (y )−λ−1(r −E (π ))+g (5.12) t t t+1 3 t t t+1 t r = β(L)π +δ(L)(y −z ) (5.13) t t t t Thefirstequation(5.11)saysthatinflationtodayisafunctionoftheoutputgapandtheexpectation of next period’s inflation, which in turn can be derived from an underlying Calvo pricing structure. With relative risk aversion measured by λ , equation (5.12) is an IS schedule where today’s output 3 depends on the ex ante real rate and the expectation of next period’s output gap. Equation (5.13) is a dynamic Taylor Rule that closes the model. The model is driven by autocorrelated demand shocks g and supply shocks z with the same unconditional variance. We take parameter values t t from Clarida, Gali, and Gertler (2000) and set λ = 0.9967, λ = 0.3, λ = 1, and the shocks’ 1 2 3 autocorrelation to 0.9655 in our monthly model. Our goal is to characterize what difference the inflation-response parameters β(L) might make for the volatility of macro variables according to this model. To do so, we fix δ(L) at the pre-2000 values, and calculate the difference in volatilities using pre-2000 and post-2000 values for β(L). We find that the post-2000 dynamics imply a 41.2% reduction in the variance of inflation, a 0.2% reduction in the variance of output, and a 33.5% reduction in the variance of the fed funds rate. Alternatively, we fixed the output dynamics δ(L) at the post-2000 values, and calculated how much difference the change in inflation dynamics β(L) made for that specification, with very similar results. These comparisons are reported in the last column of Table 5. We next wanted to see what it was about the post-2000 inflation response that helped stabilize inflation. Was it the overall magnitude of the inflation response, as reflected in the sum of the β j coefficients, or was it the more gradual post-2000 response, as reflected in the shape of the dynamic response? To find out, we explored the consequences of changing just one of these two elements at pre post a time. Let β denote the pre-2000 inflation responses and β the post-2000 responses. We j j calculated what would happen if the inflation responses were given by pre pre pre [β +β +...+β ] β = β post 0 1 6 j j [β post +β post +...+β post ] 0 1 6 so that the sum of the coefficients β was restricted to be the same as for the pre-2000 estimates, j while the shape of β(L) was that for the post-2000 estimates. These results are reported in the 18

Table 5: Effects of Changing Inflation Policy Response Inflation Coefficients Pre Path, Post Path, Pre Path, Post Path, Variable Output Coefficients Pre LR Pre LR Post LR Post LR Percentage Change in Volatility from Benchmark Pre 0 −8.7 −39.3 −41.2 π Post 0 −7.9 −40.6 −41.4 Pre 0 +24.2 −17.4 −0.2 y Post 0 +30.9 −21.0 +0.7 Pre 0 −14.0 −26.3 −33.5 r Post 0 −15.7 −27.6 −35.6 Notes: volatilities of endogenous variables of the sticky-price model described in text. Dynamic Taylor Rule coefficientsinTable4. Pre Path, Pre LRisbenchmarkusingthedynamicTaylorRulederivedfrompre-2000estimates. Post Path, Pre LR uses the dynamic Taylor Rule derived from post-2000 estimates, with inflation coefficients multiplied by the ratio of the pre-2000 long-run inflation response to the post-2000 long-run inflation response. Pre Path, Post LRusesthedynamicTaylorRulederivedfrompre-2000estimates,withinflationcoefficientsmultiplied by the ratio of the post-2000 long-run inflation response to the pre-2000 long-run inflation response. Post Path, Post LR uses the dynamic Taylor Rule derived from the post-2000 estimates. These inflation coefficient modifications are considered holding constant the output coefficients at the pre-2000 values in rows marked Pre, and at the post-2000 values in rows marked Post. column labeled “post-path, pre-LR” in Table 5. Such a change would have only modestly improved the variance of inflation, and would have resulted in a significant deterioration in the variability of output. On the other hand, if we change just the long-run response, but leave the dynamics the same as for the pre-2000 rule, post post post [β +β +...+β ] β = β pre 0 1 6 , j j [β pre +β pre +...+β pre ] 0 1 6 as reported in the “pre-path, post-LR” column of Table 5, we would have achieved the full benefits of inflation stabilization as well as additional benefits of output stabilization. Thus the key improvement in perceived monetary policy was a stronger long-run response to inflation. The fact that the market also perceives these responses to come more slowly in the post-2000 data has in fact been counterproductive. A lesson from this basic New Keynesian analysis is the following. Increasing the long-run magnitude of inflation response, as the market perceives the Fed to have done, had a stabilizing effect both on inflation and output. Implementing the response more slowly, as the market also perceives the Fed to have done, counteracted what otherwise would have been a benefit for output volatility of the stronger eventual inflation response. The “measured pace” of monetary tightening 19

Table 6: Specification Tests h 1 2 3 4 5 6 (4) K=2, pre 0.0939 0.1061 0.1113 0.0975 0.0957 0.0956 (5) K=2, post 0.0781 0.0707 0.0721 0.0649 0.0650 0.0705 (6) MGDP, pre 0.3422 0.3681 0.3723 0.3725 0.3765 0.3974 (7) MGDP, post 0.3015 0.2816 0.2719 0.2850 0.2857 0.2859 (8) Cross, pre 0.9878 0.9668 0.9185 0.9961 0.9829 0.9985 (9) Cross, post 0.9999 0.9869 0.9488 0.9999 0.9999 0.9999 Notes: p-valuesfromHansen’s(1982)J-testofoveridentifyingrestrictions,forvariousalternatespecifications. K=2, pre and K=2, post use only two indicators (CPIXFE and INDPRD). MGDP, pre, and MGDP, post use monthly GDPinsteadofIndustrialProductionastheoutputvariable. Cross, preandCross, postteststhecross-equation restriction that the average risk premium change is identical across indicators. during 2004-2006 could thus have been a factor contributing to unnecessary volatility of output – doing the same thing more quickly might have produced a better result. 6 Sensitivity Analysis We next investigate the sensitivity of our results to using alternative economic indicators, test the cross-equation restrictions imposed, and look for corroboration of the identifying assumptions from other data sources. 6.1 Results for Alternative Economic Indicators Hereweanalyzethesensitivityofourbaselineresultstothedataused. First,wereportresultsusing onlytwoindicators. Second,wereportresultsusingmonthlyGDP,ascalculatedbyMacroeconomic Advisers, instead of industrial production to measure monthly output. Given the strong evidence of a parameter break, we consider these alternate specifications estimated separately over the two subsamples. Table 6 displays the overidentification test results while Table 7 presents the response coefficient estimates. Specifications (4) and (5) of Table 7 suggest that the nonfarm payrolls indicator provides useful variation to the estimation. In its absence, the parameter estimates are less precisely estimated at several horizons. We still estimate a pre-2000 steeply adjusting inflation response, nearly identical to the baseline pre-2000 estimates. And again the pre-2000 output response, when statistically significant, looks to be modest. However, the longer horizon estimates are accompanied by large 20

Table 7: Market-Perceived Monetary Policy Rule Estimates, alternate specifications h 1 2 3 4 5 6 (4) K=2, pre β 0.3655* 0.8756** 1.1693** 1.1302* 1.0811 1.3718 0.1451 0.2245 0.3877 0.4897 3.6476 1.9668 δ 0.0628** 0.0487** 0.1386 0.1338 0.1875 0.1984 0.0136 0.0166 0.0762 0.1227 0.6804 0.3485 (5) K=2, post β -0.0737** 4.7939 0.9579 -0.0881 1.0009 2.6046 0.0255 60.3866 1.5374 0.1503 1.5029 18.1027 δ 0.0111** 0.1660 -0.0357 0.0253 0.0643 0.2120 0.0038 2.0437 0.1054 0.0208 0.0329 1.2092 (6) MGDP, pre β 0.4059 0.9063** 1.1353** 1.1681** 1.1832* 1.4515 0.2170 0.1916 0.2419 0.2716 0.4921 0.9977 δ 0.2140** 0.0848** 0.2163** 0.3608** 0.3795** 0.5833** 0.0615 0.0197 0.0418 0.0593 0.1101 0.1910 (7) MGDP, post β 0.0012 0.3985** 0.3923** 0.1520* 0.9583** 1.8190** 0.2929 0.0818 0.0984 0.0667 0.1467 0.5357 δ 0.2986 0.0063 -0.1860** -0.1570* 0.0923 0.0533 0.1678 0.0448 0.0706 0.0721 0.0752 0.1134 (8) No cross, pre β 0.3606 0.8534* 1.2100* 1.2124 1.1126 1.3821 0.2435 0.4101 0.5341 0.8613 0.5899 0.9823 δ 0.1548** 0.1316** 0.2615** 0.3970** 0.3857* 0.5746** 0.0369 0.0364 0.0841 0.1220 0.1509 0.2012 (9) No cross, post β 0.1425** 0.3656* -1.1842 0.5738** 0.8037** 2.0393 0.0542 0.1455 0.6257 0.1802 0.2048 1.0932 δ 0.0368** 0.0037 0.1361* -0.0559 0.0409 -0.0691 0.0126 0.0091 0.0621 0.0337 0.0306 0.0793 Notes: The policy rule coefficient on inflation is β and on the output gap is δ. HAC standard errors in italics. The markers * and ** denote significance at 5% and 1% levels, respectively. See the notes for Table 6 or the text of Section6fordescriptionsofthealternatespecifications. Pointestimatesandstandarderrorsfromtwo-stepnonlinear GMM. Data run over 1994:M1-2007:M7 standard errors, rendering them statistical insignificant. Still, the overidentifying tests in Table 6 fail to reject at the 5% level. Specifications (6) and (7) of Table 7 give evidence that our baseline results are robust to other measuresofoutput. Thepre-2000outputresponsecoefficientsareverysimilartothebaseline, with a slightly larger 0.21 response at the one-month horizon and virtually unchanged 0.58 response at the six-month horizon. The profile for the post-2000 response is similar to the baseline results, againwiththethree-andfour-monthhorizonresponseswronglysigned. Testsoftheoveridentifying restrictions in Table 6 are quite supportive of the assumptions. 21

6.2 Tests of Cross-Equation Restrictions In addition to the average change in the risk premium on fed funds futures contracts, the constant term η in equation (2.12) would incorporate any non-zero mean for the specification error that r,k represented day-to-day changes in the market forecasts of potential output, the inflation target, and the policy rule residual (see expression (2.9)). If this constant terms turned out to be different for different indicators k, that could be evidence of general mis-specification. For example, if the indicators were in part providing signals about changes in potential output, and if the value of this signal differed across indicators, that might show up as differences in η across different k. r,k It is easy to conduct tests of the restriction (4.1) that the policy rule constant is identical across indicators based again on Hansen’s J-statistic TQ (cid:16) θˆ(h) ,Y (h) (cid:17) −TQ (cid:16) θˆ(h) ,Y (h) (cid:17) ≈ χ2(2) R T U T whereθˆ istheGMMparameterestimatesubjecttothecross-equationrestrictionη = η = η R r,1 r,2 r,3 and θˆ is the unrestricted estimate. The p-values for this test are reported in rows 8 and 9 of Table U 6. The restrictions are quite consistent with the data. The unrestricted policy parameter estimates from θˆ are reported in rows 8 and 9 of Table 7. U Nothing substantive is lost, and statistical precision is noticeably gained, by imposing the crossequationrestrictionthatthepolicyruleconstantisidenticalacrosseconomicindicators. Comparing rows 8 and 9 of Table 7 to Table 3, one sees that estimating separate policy rule constants reduces the statistical precision with which we estimate the policy rule response coefficients, in particular the inflation response coefficients at longer horizons. 6.3 Potential Output and the Inflation Target A challenge for standard methods of estimating monetary policy rules is the difficulty in measuring potential output y∗ and the inflation target π∗. We have argued that our approach can avoid these t t problemstotheextentthatthedailynewsitemsofwhichwemakeusehavenegligibleconsequences fory∗ orπ∗. Hereweprovideadditionalevidenceonwhywebelievethatisareasonableassumption. Toexplorethisissueempirically,wewillbelookingatthepropertiesoftheCongressionalBudget Office’s series for quarterly potential real GDP growth, denoted y∗ where q indexes quarters. If q one looks at the historical values of this series as reported in the January 2009 vintage, y∗ is an q 22

extremely smooth and highly predictable series. However, over time the CBO will make many revisions to its estimate of the value of y∗ for a given historical quarter q. For example, on April q 17, 1996, CBO estimated the growth rate of potential GDP for q = 1995:Q4 to be 1.98% (at an annual rate), whereas by January 8, 2009, they had revised the estimate for y∗ up to 2.76%. 1995:Q4 Orphanides (2001) and Orphanides and van Norden (2002) demonstrated that such revisions can pose a big problem for traditional Taylor Rule estimates. Is it reasonable to assert that the daily news events exploited in our analysis had negligible implications for these subsequent revisions of potential GDP? Let Ω(q) denote the information set available to the public as of the 20th calendar day of the first month of quarter q+1. For example, for q = 1995:Q4, Ω(q) would represent information publiclyreportedasofJanuary20, 1996. Bythisdate, valuesforthepercentagegrowthinnonfarm payrolls for each month of quarter q would have been reported, denoted x , x , and 1q|Ω(q) 2q|Ω(q) x , though the actual GDP growth rate for quarter q would not yet be known. Thus for 3q|Ω(q) example for q = 1995:Q4, x would be the growth rate of seasonally adjusted nonfarm payroll 1q|Ω(q) employment during the month of October 1995 as reported by the Bureau of Labor Statistics on January 6, 1996, while x would be the November 1995 growth rate as reported on January 2q|Ω(q) 6. Let {y∗ ,...,y∗ } denote the four most recent quarterly growth rates for potential q−1|Ω(q) q−4|Ω(q) GDP as they would have been reported by CBO prior to date Ω(q); for example, for q = 1995:Q4, y∗ is the potential growth rate for 1995:Q3 as estimated by CBO on February 1, 1995 (the q−1|Ω(q) mostrecentCBOestimatereleasedpriortoJanuary20, 1996). Finally, lety∗ denotethepotential q|T GDP growth rate for quarter q as reported on January 8, 2009. Vintage values for x and iq|Ω(q) y∗ were obtained from ALFRED, the real-time archived data set maintained by the Federal q−j|Ω(q) Reserve Bank of St. Louis. We then estimated the following regression by OLS for q = 1994:Q1 to 2007:Q3: 3 4 (cid:88) (cid:88) y∗ = α + α x + γ y∗ +(cid:15) . q|T 0 j jq|Ω(q) j q−j|Ω(q) q j=1 j=1 The coefficients α can tell us the extent to which the values of nonfarm payroll growth that j arrive during quarter q could help predict the potential GDP growth rate for quarter q as it would ultimately be reported, relative to information about potential GDP that had arrived prior to the quarter’s actual GDP report. We fail to reject the null hypothesis that α = α = α = 0 1 2 3 23

(F(3,46) = 0.27,p = 0.85). On the other hand, a parallel regression for predicting the actual real GDP growth rates as eventually reported, 3 4 (cid:88) (cid:88) y = α˜ + α˜ x + γ˜ y +˜(cid:15) , q|T 0 j jq|Ω(q) j q−j|Ω(q) q j=1 j=1 leads to rejection of H : α˜ = α˜ = α˜ = 0 (F(3,46) = 3.37,p = 0.03). Nonfarm payrolls contain 0 1 2 3 useful information about the current quarter’s actual GDP growth but little information about the current quarter’s potential GDP growth. We repeated the same calculations using monthly industrial production growth rates in place of nonfarm payroll employment growth.12 We again found that industrial production is of no use in predicting potential GDP (F(3,46) = 0.98,p = 0.41), but is helpful for predicting actual GDP (F(3,46) = 4.06,p = 0.01). Our maintained assumption that markets are responding to news about near-term economic conditions y and not potential output y∗ is thus fully consistent t+h t+h with these hypothesis tests. As far as the inflation target is concerned, the validity of our identifying assumption seems even more compelling. Although there may be changes in the Fed’s inflation objectives over time, the suggestion that the FOMC is changing its long-run inflation target on a daily basis in response to the latest economic news would seem quite strange to those who actually implement monetary policy. Apart from the discrete effects of personnel changes, the Fed’s long-run inflation target should be by definition an even smoother series than potential GDP. 7 Conclusion It is important to be able to measure market participants’ beliefs, manifest through their behavior, about how monetary policy is conducted. Previous work has identified futures contract prices as powerful predictors of their underlying; in particular, fed funds futures contracts are good predictors of future Federal Reserve policy. This paper proposed that market participants forecast future policy along with future economic conditions, and linked the two by the Taylor Rule. This enabled us to measure the market’s beliefs about how the Federal Reserve responds to inflation and the output gap. Additionally, by focusing on daily forecast updates, we are able to nearly eliminate 12Release of the December 1995 value for industrial production was delayed until January 24, 1996. We used this January 24, 1996 release for q=1995:Q4. 24

the impact of potential output and the inflation target on our main focus: the market-perceived monetary policy response to inflation and output. Our baseline results for the 1994–2007 sample suggest the market perceives that the Federal Reserve gradually responds to inflation and real activity. Similar to previous literature working on post-Volcker data, we find the Federal Reserve follows the Taylor Principle, a greater than one-forone response to inflation. We also find evidence that the market-perceived monetary policy rule changed over our sample. During the 1990s market-perceived policy responded robustly to output and quickly to inflation; during the 2000s market-perceived policy doesn’t respond to output and responds at a more measured pace to inflation, though its long-run inflation response is greater than before. We quantify the importance of the inflation response path and long-run magnitude in a standard model, and find that raising the long-run magnitude is effective at lowering inflation volatility while making the path more gradual is counterproductive. Our baseline results were found to be robust to alternative possible specifications. 25

References Andrews, D. W. K. (1993): “Tests for Parameter Instability and Structural Change with Unknown Change Point,” Econometrica, 61, 821–856. Ang, A., S. Dong, and M. Piazzesi (2007): “No-Arbitrage Taylor Rules,” NBER Working Papers 13448, National Bureau of Economic Research, Inc. Bartolini, L., L. Goldberg, and A. Sacarny (2008): “How Economic News Moves Markets,” Current Issues in Economics and Finance, 14(6), 1–7, Federal Reserve Bank of New York. Boivin, J. (2006): “Has US Monetary Policy Changed? Evidence from Drifting Coefficients and Real-Time Data,” Journal of Money, Credit, and Banking, 38(5), 1149–1173. Christiano, L., M. Eichenbaum, and C. Evans (2005): “Nominal Rigidities and the Dynamic Effects of a Shock to Monetary Policy,” Journal of Political Economy, 113(1), 1–45. Christiano, L. J., 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. Clarida, R., J. Gali, and M. Gertler (2000): “Monetary Policy Rules And Macroeconomic Stability: Evidence And Some Theory,” The Quarterly Journal of Economics, 115(1), 147–180. Cogley, T., andA.M.Sbordone(2008): “TrendInflation,Indexation,andInflationPersistence in the New Keynesian Phillips Curve,” American Economic Review, 98(5), 2101–26. Evans, C. L. (1998): “Real-time Taylor rules and the federal funds futures market,” Economic Perspectives, pp. 44–55. Faust, J., J. H. Rogers, S.-Y. B. Wang, and J. H. Wright (2007): “The High-Frequency Response of Exchange Rates and Interest Rates to Macroeconomic Announcements,” Journal of Monetary Economics, 54(4), 1051–68. Gu¨rkaynak, R. S., B. Sack, and E. Swanson (2005): “The Sensitivity of Long-Term Interest Rates to Economic News: Evidence and Implications for Macroeconomic Models,” American Economic Review, 95(1), 425–436. Gu¨rkaynak, R. S., B. T. Sack, and E. P. Swanson (2007): “Market-Based Measures of Monetary Policy Expectations,” Journal of Business & Economic Statistics, 25, 201–212. Hamilton, J.D.(2008): “Dailymonetarypolicyshocksandnewhomesales,”JournalofMonetary Economics, 55, 1171–1190. (2009): “Daily Changes in Fed Funds Futures Prices,” Journal of Money, Credit, and Banking, 41, 567–582. Hansen, L. P.(1982): “LargeSamplePropertiesofGeneralizedMethodofMomentsEstimators,” Econometrica, 50, 1029–1054. Ireland, P. N. (2007): “Changes in the Federal Reserve’s Inflation Target: Causes and Consequences,” Journal of Money, Credit and Banking, 39(8), 1851–1882. 26

Kohn, D. L., and B. P. Sack (2004): “Central bank talk: does it matter and why?,” in Macroeconomics, Monetary Policy, and Financial Stability, pp. 175–206. Bank of Canada. Newey, W. K., and D. McFadden (1994): “Large sample estimation and hypothesis testing,” in Handbook of Econometrics, ed. by R. F. Engle, and D. McFadden, vol. 4 of Handbook of Econometrics, chap. 36, pp. 2111–2245. Elsevier. Newey, W. K., and K. D. West (1987): “A Simple, Positive Semi-Definite, Heteroskedasticity and Autocorrelation Consistent Covariance Matrix,” Econometrica, 55(3), 703–708. Orphanides, A.(2001): “MonetaryPolicyRulesBasedonReal-TimeData,” American Economic Review, 91(4), 964–985. Orphanides, A., and S. van Norden (2002): “The Unreliability of Output-Gap Estimates in Real Time,” Review of Economics and Statistics, 84(4), 569–583. Owyang, M. T., and G. Ramey (2004): “Regime switching and monetary policy measurement,” Journal of Monetary Economics, 51(8), 1577–1597. Pagan, A. (1986): “Two Stage and Related Estimators and Their Applications,” Review of Economic Studies, 53(4), 517–38. Piazzesi, M., and E. T. Swanson (2008): “Futures prices as risk-adjusted forecasts of monetary policy,” Journal of Monetary Economics, 55(4), 677–691. Primiceri, G. E. (2006): “Why Inflation Rose and Fell: Policy-Makers’ Beliefs and U.S. Postwar Stabilization Policy,” Quarterly Journal of Economics, 121(3), 867–901. Rudebusch, G. D. (2002): “Term structure evidence on interest rate smoothing and monetary policy inertia,” Journal of Monetary Economics, 49(6), 1161–1187. (2006): “Monetary Policy Inertia: Fact or Fiction?,” International Journal of Central Banking, 2(4), 85–135. Smith, J., and J. B. Taylor (forthcoming): “The Long and Short End of the Term Structure of Policy Rules,” Journal of Monetary Economics. Stock, J.H., andM.W.Watson(2002): “MacroeconomicForecastingUsingDiffusionIndexes,” Journal of Business & Economic Statistics, 20(2), 147–62. Taylor, J.B.(1993): “DiscretionversusPolicyRulesinPractice,”Carnegie-RochesterConference Series on Public Policy, XXXIX, 195–214. 27

A Appendix A.1 Tables and Figures Table A1: Policy Rule Constant Estimates h 1 2 3 4 5 6 Baseline η -0.4024** -0.4750** -0.6223** -0.5060** -0.6844** -0.6056** r 0.0787 0.1031 0.1351 0.1223 0.1041 0.1046 Baseline, pre η -0.4085** -0.4744** -0.6187** -0.5288** -0.7056** -0.6310* r 0.0897 0.0693 0.1050 0.1459 0.2027 0.2691 Baseline, post η -0.4228** -0.5163** -0.6179** -0.5194** -0.6075** -0.6850** r 0.1174 0.1545 0.0924 0.1440 0.1945 0.1587 Notes: η is the average risk premium change. HAC standard errors in italics. The markers * and ** denote r significance at 5% and 1% levels, respectively. Point estimates and standard errors from two-step nonlinear GMM. Data run over 1994:M1-2007:M7 28

Figure A1: Trading Volume on Fed Funds Futures Contracts 5000 4000 3000 2000 1000 0 1988 1990 1992 1994 1996 1998 2000 2002 Notes: Data from Chicago Board of Trade. As quarterly average. 29