Monetary Policy, Incomplete Information, and the Zero Lower Bound

Finance and Economics Discussion Series Divisions of Research & Statistics and Monetary Affairs Federal Reserve Board, Washington, D.C. Monetary Policy, Incomplete Information, and the Zero Lower Bound Christopher J. Gust, Benjamin K. Johannsen, and David Lopez-Salido 2015-099 Please cite this paper as: Gust, Christopher J., Benjamin K. Johannsen, and David Lopez-Salido (2015). “Monetary Policy, Incomplete Information, and the Zero Lower Bound,” Finance and Economics Discussion Series 2015-099. Washington: Board of Governors of the Federal Reserve System, http://dx.doi.org/10.17016/FEDS.2015.099. 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.

Monetary Policy, Incomplete Information, and the Zero Lower Bound Christopher Gust Benjamin K. Johannsen David L(cid:243)pez-Salido (cid:3) November 3, 2015 Abstract In the context of a stylized New Keynesian model, we explore the interaction between imperfect knowledge about the state of the economy and the zero lower bound. We show that optimal policy under discretion near the zero lower bound responds to signals about an increase in the equilibrium real interest rate by less than it would when far from the zero lower bound. In addition, we show that Taylor-type rules that either include a time-varying intercept that moves with perceived changes in the equilibrium real rate or that respond aggressively to deviations of in(cid:135)ation and output from their target levels perform similarly to optimal discretionary policy. Our analysis of (cid:133)rst-di⁄erence rules highlights that rules with interest rate smoothing terms carry forward current and past misperceptions about the state of the economy and can lead to suboptimal performance. JEL classi(cid:133)cation: E32, E52 Key words: imperfect information, zero lower bound, discretionary policy, monetary policy rules. Prepared for the IMF Sixteenth Jacques Polak Annual Research Conference - November 5-6, 2015. We (cid:3) thank Fran(cid:231)ois Gourio for comments; all remaining errors are our own. We also thank Kathryn Holston for outstandingresearchassistance. Theviewsexpressedinthispaperaresolelythoseoftheauthorsanddonot necessarily re(cid:135)ect the views of the Board of Governors of the Federal Reserve System, the Reserve Banks, or any of their sta⁄s.

1 Introduction Policymakers routinely have to make decisions based on incomplete knowledge about the state of the economy. Once source of incomplete knowledge is the level of the equilibrium real rate.1 As former Chairman Bernanke recently wrote:2 If the Fed wants to see full employment of capital and labor resources (which, of course, it does), then its task amounts to using its in(cid:135)uence over market interest rates to push those rates toward levels consistent with the equilibrium rate, or(cid:150) more realistically(cid:150)its best estimate of the equilibrium rate, which is not directly observable. Imperfect knowledge about the equilibrium real rate raises the possibility that policymakers might misperceive its value, which has important implications for the design of monetary policy. In this paper, we discuss the ways in which optimal policy (under discretion) should take account of variation in the level of the equilibrium real interest rate in the context of a simple New Keynesian (NK) model in which policymakers have imperfect information about the level of the equilibriumreal rate and are constrained by the zero lower bound. We (cid:133)nd that uncertainty about the current state of the economy a⁄ects optimal policy in a number of important ways. First, there is an attenuation of the policy response to perceived changes in the equilibrium real rate coming from the (cid:133)ltering problem facing the monetary authority, as in Aoki (2003). Second, the zero lower bound non-linearly alters the distribution of equilibrium prices and quantities for both the private sector and the central bank, as has been recently discussed by Evans et al. (2015). Third, in our setup, the private sector has access to di⁄erent information than the monetary authority and thus forms di⁄erent beliefs about both current and future prices and quantities. Because the central bank has to form expectations about those private-sector expectations, they are not ex-post equal and the optimal policy response to perceived changes in the equilibrium real rate is further attenuated by the information processing problem. This muted response re(cid:135)ects asymmetric risks faced by the central bank near the zero lower bound. Finally, the private sector understands that the monetary authority may in the future respond to noise in the signals it receives. Near the zero lower bound, the private sector internalizes that 1In the context of the New Keynesian model that we use, the equilibrium real rate corresponds to the rate of interest associated with the e¢ cient economy that is undistorted by nominal or real rigidities. 2Ben Bernanke(cid:146)s Blog, (cid:147)Why are interest rates so low?(cid:148) March 30, 2015, http://www.brookings.edu/blogs/ben-bernanke/posts/2015/03/30-why-interest-rates-so-low 1

the monetary authority can only respond in one direction, which increases the mean of the private sector(cid:146)s expected future paths for short-term nominal (and real) interest rates. This tightening bias a⁄ects prices and quantities in the current period. Optimal monetary policy takes this tightening bias into account by attenuating the response of the nominal interest rate to positive signals about the equilibrium real rate. In particular, when the policy rate is constrained by the zero lower bound and the policymaker receives a noisy signal that the equilibrium rate has increased, the optimal policy is to make a mistake on the side of keeping the real policy rate below the perceived e¢ cient rate so as to reduce the likelihood that an adverse shock will weaken the economy enough, to require a returntothezerolowerbound. Thatis, policymakers(cid:147)takeoutinsurance(cid:148)againstsituations in which the information about the equilibriumreal rate might lead themto mistakenly raise the policy rate. Optimal policy under discretion in the neighborhood of the zero lower bound prescribes a complicated, nonlinear reaction function for the policy rate that may, in practice, be di¢ cult to communicate and implement. An alternative approach would be to commit to a policy rule for the short-term nominal interest rate. Thus, we also study the potential bene(cid:133)ts of committing to a rule-based approach in setting monetary policy within the same incomplete information framework. In particular, we compare the performance of simple Taylor-type rules with a constant intercept term to alternatives in which the intercept term is time-varying and changes based on a central bank(cid:146)s estimate of the equilibrium real rate. We also compare their performance to (cid:133)rst-di⁄erence rules (e.g., Orphanides and Williams (2002)) in which there is no intercept term and it is thus unnecessary to take a stand on how the equilibrium real rate should enter this rule. The performance of simple Taylor rules, using conventional parameter values, are substantially improved when the intercept in the rule is adjusted for movements in the equilibrium real rate. However, if the intercept of the rule is held constant, we (cid:133)nd that simple Taylor rules can perform almost as well as rules with a time-varying equilibrium real rate when they respond much more aggressively to in(cid:135)ation and the output gap. Similarly, we (cid:133)nd that (cid:133)rst-di⁄erence rules with an aggressive response to in(cid:135)ation tend to perform well. Most of the existing literature studying optimal monetary policy at the zero lower boundfocusesonthecaseinwhichpolicymakershavecompleteinformation. Thus,oneofour contributionsistoassumethatthemonetaryauthorityhaslessinformationaboutthestateof theeconomythantheprivatesector. Ajusti(cid:133)cationforthisassumptionisthatprivateagents make decisions after seeing shocks that a⁄ect them directly. By contrast, policymakers have 2

to learn about the shocks that cause households and (cid:133)rms to make demand, production, and pricing decisions, which we model as the monetary authority receiving noisy signals about the state of the economy. If there were no noise in the signals, the monetary authority would know the underlying structural shocks exactly, which would mean it had perfect information aboutequilibriumpricesandquantitiesaswell. Inthissense,thecentralbank(cid:146)sinformational disadvantage relative to the private sector captures that, in reality, central banks need to now-cast prices and quantities when they make their policy decisions. While we prefer to interpret our setup from the perspective of a now-casting problem faced by the central bank, another way to rationalize our framework is from a sequential structure to the decisions made in the model. In particular, a model economy in which nominal interest rate decisions have to be made before private-sector decisions in the same period requires that the central bank base its decisions on less information than the private sector has available at that time. This type of timing assumption captures the fact that, in reality, policy rate decisions are not made on a daily basis, but rather during meetings that occur infrequently. While policymakers may be able to react to incoming data received during the intermeeting periods, historically policy decisions have only been infrequently made outside of regularly scheduled meetings. Thus, policymakers regularly set short-term nominal interest rates based on information available only at the beginning of the period in which that policy decision will be operative. Moreover, with lags in o¢ cial data releases and revisions to real-time data, the information available even about the recent past is likely to be a⁄ected by noise, further emphasizing the informational problem faced by central banks. This paper is related to a growing literature in NK models exploring the implications of the zero lower bound for optimal policy. Eggertsson and Woodford (2003) emphasized that optimal commitment policy can be very e⁄ective in stabilizing the economy at the zero lower bound. However, as noted in Levin et al. (2010), optimal commitment policies at the zero lower bound su⁄er from a severe time inconsistency problem. Accordingly, we focus on optimal discretion to abstract from how the central bank can have access to a credible commitment technology. Adam and Billi (2007) and Nakov (2008) consider optimal discretionary policy at the zero lower bound with full information, while Aoki (2003) studies the optimal discretionary policy without the zero lower bound but when the monetary authority has imperfect knowledge about the state of the economy. We consider both imperfect information and the zero lower bound in a uni(cid:133)ed framework and thus, unlike Aoki (2003), optimal policy deviates from certainty equivalence. Several other studies have focused on the use of simple 3

feedback rules for conducting monetary policy near the zero lower bound (e.g., Reifschneider and Williams (2000) and Coenen and Wieland (2003)). We also consider simple rules but in the context of a central bank with imperfect knowledge about the state of the economy. By considering optimal simple rules in the context of imperfect information, our paper builds on work by Boehm and House (2014). Our paper is also closely related to a recent contribution by Evans et al. (2015). In their complete information setup, they focus on uncertainty about future values of the equilibrium real rate. By contrast, in our incomplete information setup, the central bank does not observe the current equilibriumreal rate, and private-sector beliefs about equilibrium prices and quantities di⁄er from those of the central bank. The remainder of this paper is organized as follows. Section 2 discusses the basic elements of the model that we use to characterize the determinants of the equilibrium real rate. Section 3 reviews optimal policy under discretion in the presence of the zero lower bound. Italsodescribestheinformationstructureofthemodelandcharacterizestheoptimal (cid:133)ltering problem. This section presents the main results regarding optimal discretionary policy near the zero lower bound when the central bank has imperfect information about the true state of the economy. Section 4 then discusses the performance of simple rules in this environment. Finally, section 5 o⁄ers some concluding remarks. 2 A Simple New Keynesian Model Our analysis builds on Woodford (2003) and Gal(cid:237) (2008) who use the simple NK model to characterize optimal monetary policy. The model has two key equations: an IS equation that relates output to the real interest rate (that is, the nominal short-term rate adjusted by expected in(cid:135)ation) and expected future output, and a Phillips curve relationship that relates in(cid:135)ation to the output gap and expected future in(cid:135)ation. We leave the details of the model for an appendix and devote this section to de(cid:133)ning the natural and the e¢ cient levels of the real interest rates using a log-linear approximation of the model economy. Asdiscussedintheappendix,theprivatesector(log-linearized)equilibriumconditions around the non-stochastic steady state can be written as follows: y = E y E i (cid:25) (cid:17) (1) t t t+1 t t t+1 t f g(cid:0) f (cid:0) (cid:0) g (cid:25) = (cid:12)E (cid:25) +(cid:20)[y +(cid:22) ] (2) t t t+1 t t f g where E represents the expectation conditional on time t information; y and (cid:25) refer to the t t t log deviations of output and in(cid:135)ation from their values in nonstochastic steady state; and i t represents the short-term nominal interest rate in deviations from its non-stochastic steady 4

state. The parameter (cid:12) represents the household(cid:146)s discount factor. The parameter (cid:20) = (cid:15) 1 (cid:0)’ is inversely related to (cid:133)rms(cid:146)costs of changing prices (’) and directly related to the elasticity of demand ((cid:15)). The variables (cid:17) and (cid:22) represent exogenous shocks to the discount rate and t t the (cid:133)rm(cid:146)s marginal cost, respectively.3 A decrease in (cid:17) represents an exogenous factor that t induces a temporary rise in households(cid:146)propensity to save, and reduces current aggregate household demand for goods. An increase in (cid:22) corresponds to an exogenous cost-push shock t that increases in(cid:135)ation for given levels of output and expected future in(cid:135)ation; therefore, it introduces a tradeo⁄between in(cid:135)ation and output. The natural rate of output and the natural real interest rate are de(cid:133)ned as the levels of output and the real interest rate in the economy without price rigidities. As shown in the appendix, deviations of the natural real rate from its steady state value can be expressed as follows: rn = (cid:17) E (cid:22) (cid:22) t t (cid:0) t f t+1 (cid:0) t g Thus, there are two sources of exogenous disturbances to the natural rate of interest: shocks to the discount factor and shocks to the marginal cost. A decrease in the desire to save (an increase in (cid:17) ) and expected decrease in marginal costs cause the natural rate of t interest to rise. The e¢ cient equilibrium corresponds to that of the economy with (cid:135)exible prices and no exogenous variation in marginal costs, as those shocks push equilibrium quantities and prices from their e¢ cient values. Formally, as shown in the appendix, the log-linear approximation around the e¢ cient non-stochastic steady state implies that:4 re = (cid:17) ; t t where re denotes the deviation of the e¢ cient real interest rate from its steady state value. t As discussed in Woodford (2003) and Gal(cid:237) (2008), the presence of a gap between the natural and the e¢ cient levels of output due to exogenous changes in real marginal costs implies that the relevant concept for welfare is the di⁄erence between the observed level and the e¢ cient level of output. Formally, (cid:135)uctuations in output relative to its e¢ cient level, x , t can be decomposed as follows: x = y ye = (y yn)+(yn ye) = (y yn) (cid:22) t t (cid:0) t t (cid:0) t t (cid:0) t t (cid:0) t (cid:0) t 3The relationship of these shocks with the speci(cid:133)cation of preferences and technology is described in the appendix. Both shocks follow (cid:133)rst-order autoregressive processes. 4AsshownintheAppendix,thenon-stochastice¢ cientlevelofoutputisconstantandnormalizedtoone, that is ye =0. t 5

Thus, x can be decomposed in the sum of deviations of output from the natural rate t (due to the presence of nominal frictions) plus deviations of the natural level of output from its e¢ cient level (due to changes in real marginal costs, (cid:22) , exogenous to monetary policy). t Using this de(cid:133)nition, the aggregate demand equation can be written in terms of deviation of output from its natural rate as follows: y yn = E y yn E i (cid:25) rn t (cid:0) t t f t+1 (cid:0) t+1g(cid:0) t f t (cid:0) t+1 (cid:0) tg The in(cid:135)ation equation can also be written in terms of the output gap: (cid:25) = (cid:12)E (cid:25) +(cid:20)[x +(cid:22) ]; (3) t t t+1 t t f g given that there are not exogenous variation in productivity as described in the appendix ye = 0. We can also rewrite equation (1) in terms of the output gap as follows: t x = E x E i (cid:25) re (4) t t f t+1 g(cid:0) t f t (cid:0) t+1 (cid:0) tg Because of the presence of the cost-pushshocks, (cid:22) , the (cid:147)divine coincidence(cid:148)property t discussed in Blanchard and Gal(cid:237) (2007) is not present in our model. Thus, it will also not be optimal for the policymaker to pursue a policy in which the ex-ante real rate equals the e¢ cient real interest rate. 3 Optimal Discretionary Policy We now discuss the optimal monetary policy when the policymaker is unable to credibly commit to future policy actions, and when the nominal interest rate is constrained by the zero lower bound. We consider the cases in which the central bank has perfect information (cid:133)rst and later imperfect information regarding the current state of the economy. 3.1 The Optimal Policy Problem WebuildonEggertssonandWoodford(2003)who(cid:133)rstintroducedthezerolowerboundconstraint into the linear-quadratic framework originally studied by Rotemberg and Woodford (1997) and Clarida et al. (1999).5 Following Woodford (2003) we assume that the monetary authority seeks to minimize the following quadratic loss function, 1 1 E (cid:12)t (cid:25)2 +(cid:21)x2 ; 2 0 t t t=0 X (cid:8) (cid:9) 5Adam and Billi (2007) and Nakov (2008) also o⁄er discussions of the e⁄ect of the e⁄ective zero lower bound on optimal monetary policy under discretion. 6

where (cid:12) is the discount factor, (cid:21)is a trade-o⁄parameter and E represents the expectation t conditional on time t information. The monetary authority is constrained by the private sector equilibrium conditions, equations (4) and (3), as well as by the zero lower bound on the nominal interest rate: (cid:12) i log : (5) t (cid:21) (cid:5) (cid:18) (cid:19) In this paper we focus on the optimal stabilization policy at the zero lower bound and abstract from the issue on how the zero lower bound will a⁄ect the optimal average rate of in(cid:135)ation. Instead, (cid:25) in the loss function represents the deviation of in(cid:135)ation from a two t percent target which is consistent with the longer-run objective of the Federal Reserve.6 This allows us to focus on how the nonlinearity introduced by the zero lower bound limits the optimal real rate from being equal to the e¢ cient real rate. At the zero lower bound, any shock creates a tradeo⁄between in(cid:135)ation and output stabilization for the policymaker because of the constraint it places on monetary policy actions. Weassumethatthemonetaryauthoritychoosespolicyunderdiscretiontakingprivatesectorexpectationsasgiven. Undertheseassumptionsandundertheassumptionofcomplete information, the (cid:133)rst-order conditions characterizing optimal policy are: (cid:21)x = (cid:20)(cid:18) (cid:30) t t t (cid:0) (cid:25) = (cid:18) t t (cid:0) (cid:13) = (cid:30) ; t t where (cid:30) , (cid:18) , and (cid:13) denote the Lagrange multipliers on equations (3), (4), and (5), respect t t tively. The complementary slackness condition is: (cid:12) [i log( )] (cid:13) = 0: t t (cid:0) (cid:5) (cid:1) In normal circumstances (e.g., far away from the zero lower bound), the previous (cid:133)rst-order conditions correspond to those analyzed by Clarida et al. (1999). In this case (cid:30) = (cid:13) = 0, and hence the optimal discretionary policy gives rise to the following optimality t t condition or targeting rule: (cid:20) x = (cid:25) : t t (cid:0)(cid:21) Without exogenous shocks to marginal costs and under the assumption that the economy will never be at the zero lower bound, the optimal discretionary policy consists of 6See Coibion et al. (2012) and the references therein, for a discussion of how the optimal steady state in(cid:135)ation is a⁄ected by the presence of the zero lower bound. 7

setting the actual real rate equal to the e¢ cient real rate, re. In these circumstances, this t policy implies that in(cid:135)ation is equal to target and the output gap is closed at every date (i.e., (cid:25) = x = 0). t t However, the presence of the zero lower bound will change this prescription. At the zero lower bound, the equilibrium under discretion implies that the targeting rule characterizing optimal policy is a⁄ected by the Lagrange multiplier associated with that constraint: (cid:20) 1 x = (cid:25) (cid:30) : (6) t t t (cid:0)(cid:21) (cid:0) (cid:21) In normal times, the output gap moves proportionately with the in(cid:135)ation gap; however, at the zero lower bound, a tightening in the constraint implies a larger decline in the output gap relative to the decline in the in(cid:135)ation gap, since (cid:30) > 0. t 3.2 Information Structure Our novel modeling contribution is to describe how optimal policy at the zero lower bound changes because the monetary authority faces imperfect information about the state of the economy. Importantly, we assume that the monetary authority has less knowledge about the state than the private sector. A justi(cid:133)cation for this assumption in the context of our model is that private agents (households and (cid:133)rms) make decisions after seeing shocks that a⁄ect them directly. By contrast, policymakers have to learn about the nature of (cid:135)uctuations in the economy that cause private agents to make demand, production, and pricing decisions. Formally, at every time t households and (cid:133)rms know the entire history of exogenous shocks and endogenous prices and quantities up to date t. However, for the monetary authority, information about the history of exogenous shocks and endogenous prices and quantities is available up to date t 1, and the monetary authority only receives noisy (cid:0) signals about current values. In particular, at time t, the monetary authority observes: s(cid:17) = (cid:17) +e ; where e N(0;(cid:27)2 ) t t (cid:17)t (cid:17)t (cid:24) e(cid:17) s(cid:22) = (cid:22) +e ; where e N(0;(cid:27)2 ): t t (cid:22)t (cid:22)t (cid:24) e(cid:22) That is, the monetary authority observes noisy signals about the discount rate and the marginal cost shocks at time t. If there were no noise in the signals, the monetary authority would know (cid:17) and (cid:22) exactly, which would mean that it had perfect information about t t (cid:25) and x as well. In this sense, uncertainty about (cid:17) and (cid:22) is a stand-in for uncertainty t t t t about the equilibrium values of (cid:25) and x . In a very general sense, the values (cid:25) and x are t t t t 8

determined in equilibrium by a mapping from states and shocks to equilibrium prices and quantities: (cid:25) t = f((cid:17) ;(cid:22) ;e ;e ;" ;" ); t 1 t 1 (cid:17)t (cid:22)t (cid:17)t (cid:22)t (cid:20) x t(cid:21) (cid:0) (cid:0) where the time-invariant function f represents the equilibrium mapping. When (cid:27)2 = (cid:27)2 = e(cid:17) e(cid:22) 0, s(cid:17) and s(cid:22) reveal " and " , which in turn reveal equilibrium quantities and prices, x and t t (cid:17)t (cid:22)t t (cid:25) . In our setup, the monetary authority has to form beliefs about e , e , " and " : t (cid:17)t (cid:22)t (cid:17)t (cid:22)t We assume that the central bank has an informational disadvantage relative to the private sector so as to capture that, in reality, the central banks need to now-cast prices and quantities when they make policy decisions. There is a long literature about the now-casting problems that central banks face and their associated policy implications. For example, a string of work initiated by Orphanides and Williams (2002) emphasizes that real-time measurement errors can lead to persistent policy mistakes if monetary policy rules do not confront this now-casting problem. In the context of NK models, the literature on monetary policy with imperfect information (typi(cid:133)ed by Svensson and Woodford (2003), Svensson and Woodford (2004), and Aoki (2003)) often speci(cid:133)es signals as noisy indicators of endogenous variables. Typically, signals take the form: s(cid:25) = (cid:25) +e t t (cid:25)t sy = y +e : t t yt When signals are speci(cid:133)ed in this way, the zero lower bound on the short-term nominal interest rate complicates the joint signal extraction problem and optimal policy problem dramatically. Our information structure allows us to model central bank beliefs using a Kalman (cid:133)lter (discussed below) to form beliefs, which is optimal because (cid:17) and (cid:22) evolve t t independently of endogenous prices andquantities. That is, there is no simultaneity problem in our setup because the conditioning set of the monetary authority does not contain any quantities that are functions of endogenous values that are determined in response to the policy decision. This leads to a separation between the optimal (cid:133)ltering problem and the optimal policy problem, which greatly simpli(cid:133)es the information processing problem. We view this imposition of separation as a compromise that embodies an information friction while keeping the problem tractable in the context of the zero lower bound. An alternative interpretation of our approach is that the distributions of e and e (cid:25)t yt are not normal and instead are such that the signals s(cid:25) and sy embody the same information t t as s(cid:17) and s(cid:22). That is, rather than e and e being independently and normally distributed t t (cid:25)t yt around the endogenous variables that they simultaneously determine, we assume that they 9

have a complex and unknown distribution that makes our assumed signal structure (and the separation between the (cid:133)ltering problem and the optimal policy problem) exactly correct. While we prefer to interpret our setup from the perspective of a now-casting problem faced by the central bank, another way to rationalize our framework is through a sequential decision structure. In particular, a model economy in which nominal interest rate decisions for period t have to be made before private agents make decisions in period t requires that the central bank bases them on less information than the private sector has available at time t. If, for example, the central bank makes its policy rate decision at the beginning of period t, that decision will be based on information available at time t 1 and potentially, (cid:0) any signals the central bank may have regarding current shocks that a⁄ect private sector decisions. This type of timing assumption captures the fact that policy rate decisions are not made on a daily basis, but rather during meetings that occur infrequently. While policymakers may be able to react to incoming data received during the intermeeting periods, historically policy decisions have only been infrequently made outside of regularly scheduled meetings. Thus, most central banks regularly set short-term nominal interest rates based on information available only at the beginning of the period in which that policy decision will be operative. Moreover, with lags in o¢ cial data releases and frequent revisions to real-time data, even the information available to the central bank about the recent past is likely to be a⁄ected by noise. 3.3 Information Processing To de(cid:133)ne the signal-extraction problem of the central bank, it is convenient to denote the vector of current state variables by: (cid:22) (cid:24) = t : t (cid:17) t (cid:20) (cid:21) Given our assumptions in the model, these exogenous variables evolve according to: (cid:24) = F(cid:24) +v ; (7) t t 1 t (cid:0) where (cid:26) 0 0 (cid:27)2 0 F = (cid:22) and v ~N ; (cid:22) = N (0;Q): 0 (cid:26) t 0 0 (cid:27)2 (cid:20) (cid:17) (cid:21) (cid:18) (cid:17) (cid:19) As described above, the monetary authority sees all prices and quantities, as well as all state variables, up through time t 1. In addition, the monetary authority has access to noisy (cid:0) 10

signals about the current values of the state variables, which we speci(cid:133)ed above, and can be written as s(cid:22) s = t = (cid:24) +w ; (8) t s(cid:17) t t t (cid:20) (cid:21) where w ~N (0;R); t where R is de(cid:133)ned accordingly given the de(cid:133)nitions above. In terms of our notation, the central bank information set is s ;(cid:24) ;(cid:24) ::: : t t 1 t 2; f (cid:0) (cid:0) g Weassumethattheprivatesectorhascompleteinformation,meaningthatitsinformationset includes everything the monetary authority sees, as well as the true values of current-period state variables. Conditional on this information structure, the problem of the monetary authority is to determine the optimal policy rate, given its expectations about equilibrium prices and quantities. The linearity of equations (7) and (8), the normality of the innovations, and the normality of the noise in the signals implies that the beliefs of the central bank about (cid:24) will t be normally distributed. In fact, given a signal, s , the central bank optimally forms beliefs t about (cid:24) using the Kalman (cid:133)lter so that t E((cid:24) s ;(cid:24) ;(cid:24) ::: ) = F(cid:24) +Q(Q+R) 1(s F(cid:24) ) t t t 1 t 2; t 1 (cid:0) t t 1 jf (cid:0) (cid:0) g (cid:0) (cid:0) (cid:0) Var((cid:24) s ;(cid:24) ;(cid:24) ::: ) = I Q(Q+R) 1 Q t t t 1 t 2; (cid:0) jf (cid:0) (cid:0) g (cid:0) (cid:0) (cid:1) wheretheKalmangainisgivenbyQ(Q+R) 1. Theseobjectsarethemeanandthevariance (cid:0) of the normal posterior distribution of the central bank. The monetary authority can use this distribution to form beliefs about about in(cid:135)ation and the output gap by integrating over the distribution of (cid:24) . The distribution of (cid:24) can then be used to solve the optimal t t policy problem of the central bank. Under incomplete information, the (cid:133)rst-order necessary condition associated with minimizing the central bank loss function is: E[(cid:21)x +(cid:25) (cid:20) s ;(cid:24) ;(cid:24) ::: ]+(cid:30) = 0: t t t t 1 t 2; t jf (cid:0) (cid:0) g The expectations operator in this (cid:133)rst-order condition re(cid:135)ects policymakers use their best estimate of in(cid:135)ation and the output gap in their optimal choice of the nominal interest rate. This condition when coupled with the expressions for x and (cid:25) , the zero lower bound t t constraint, and the complementary slackness condition de(cid:133)nes the equilibrium conditions 11

under imperfect information. To see how this (cid:133)rst order condition a⁄ects the central bank(cid:146)s choice of i , it is useful to substitute for x in order to get t t 1 (cid:20) i = (cid:30) +E E x +E (cid:25) +re + (cid:25) s ;(cid:24) ;(cid:24) ::: : t (cid:21) t t f t+1 g t f t+1 tg (cid:21) t jf t t (cid:0) 1 t (cid:0) 2; g h i Written in this way, it is clear that even if the private sector has perfect information at time t, the monetary authority needs to form expectations about private sector expectations in order to solve its optimal policy problem. Uncertainty about the state of the economy in our setup a⁄ects optimal policy in a number of ways. As we will show below, there is an attenuation coming from imperfect information. Additionally, thezerolowerboundaltersthedistributionofequilibriumobjects for both the private sector and the central bank. Furthermore, because of its informational disadvantage, the monetary authority has to form beliefs about private sector expectations. As such, the private sector expectations about future in(cid:135)ation and output can be thought of as time t variables about which the central bank is uncertain. In our setup, the private sector has di⁄erent beliefs about both current and future prices and quantities than the monetary authority. Because the central bank has to form expectations about those private sector expectations, they are not ex-post equal. This additional uncertainty about private sector expectations arises because of our informational setup, and a⁄ects the optimal policy response near the zero lower bound. Finally, the private sector understands that the monetary authority may in the future respond to noise in the signals it receives. Near the zero lower bound, the private sector internalizes that the monetary authority can only respond in one direction, which shifts the mean of the private sector(cid:146)s expected future paths for the short-term nominal interest rates. This tightening bias a⁄ects prices and quantities in the current period, which the central bank internalizes. 3.4 Results As discussed in the appendix, we solve the model using a global solution method that allows ustocharacterizehowimperfectinformationnearthezerolowerbounda⁄ectsoptimalpolicy. Wepickrelativelystandardvaluesfortheparametersofthemodel(see,e.g.,Woodford(2003) andmorerecentlyEvansetal.(2015)). Thediscountfactor,(cid:12),ischosentobeconsistentwith a quarterly model and is set equal to 0:995. As a result, the steady state real rate equals 2 (cid:22) percent on an annual basis. The central bank(cid:146)s in(cid:135)ation target, (cid:5), is chosen to be consistent with a 2 percent annual rate. The parameter (cid:20) which governs the slope of the Phillips curve, issetequalto0:01. FollowingWoodford(2003), wechoose(cid:21) = 0:0156. Thisquarterlyweight 12

translates into an annualized weight of (cid:21) 16 0:25. We set the parameters describing (cid:2) (cid:25) the shocks as follows. The discount factor shocks has an autocorrelation, (cid:26) , equal to 0:8 (cid:17) with a standard deviation, (cid:27) , of 0:45. Likewise, the markup shock has an autocorrelation (cid:17) coe¢ cient, (cid:26) , equal to 0.3 with a standard deviation, (cid:27) , of 0:1. The standard deviations (cid:22) (cid:22) of the noise in the signals, (cid:27) and (cid:27) , are calibrated to be 0:9 and 0:1, respectively. This e(cid:17) e(cid:22) implies signal-to-noise ratios of 1:25 and 2, respectively. Figure 1 illustrates how incomplete information about the e¢ cient real interest rate interacts with the zero lower bound to a⁄ect optimal policy. Each panel in Figure 1 shows the optimal response to a positive signal about the e¢ cient real rate. That is, for each panel, we show how the policymaker optimally sets the nominal short-term interest rate (on the vertical axes, measured in percentage points) in response to an innovation that signals that the natural rate has increased (the horizontal axes show the magnitude of that positive signal, measuredinpercentagepoints). Thethreepanelsdi⁄erintheirassumedinitialvalues for the e¢ cient rate. The top panel assumes that the e¢ cient rate has been well below zero and, as result, the zero lower bound is binding. Even for relatively large positive signals about the e¢ cient rate, the optimal policy leaves the nominal rate unchanged. By contrast, if the lower bound were irrelevant, as shown by the dashed blue line, the optimal policy would set the nominal rate below zero for small, positive signals about the level of the e¢ cient rate. For signals larger than forty basis points the optimal unconstrained policy rate would be positive. Notably, once the zero lower bound is taken into account, the presence of uncertainty keeps the nominal rate at the zero lower bound even in situations where the unconstrained policy would be positive. These lower policy prescriptions re(cid:135)ect that optimal policy opts to insure against particular bad outcomes associated with the zero lower bound. The middle panel repeats the exercise assuming that the e¢ cient rate has been low enough to put the short-term nominal interest rate at the cusp of the zero lower bound. If the monetary authority acts as if policy is not constrained by the zero lower bound, the nominal rate rises about ten basis point for a positive signal of (cid:133)fty basis points. When we account for the lower bound constraint, the optimal policy reacts less, rising only (cid:133)ve basis points for a signal of (cid:133)fty basis points, because the policy rate remains at the lower bound until the signal rises above thirty basis points. As the signal grows larger, the reaction of the optimal policy non-linearly approaches the reaction of the unconstrained optimal policy. The bottom panel shows the e⁄ects in normal times corresponding to an initial value of the e¢ cient real rate that is high enough so that the zero lower bound is very unlikely 13

Figure 1: Optimal Discretionary Policy with Incomplete Information Reaction Function at the ZLB 0.2 0.1 0 −0.1 −0.2 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Signal about r e )ylretrauQ( etaR lanimoN Unconstrained − Incomplete Info Constrained − Incomplete Info Reaction Function at the Cusp of the ZLB 0.5 0.4 0.3 0.2 0.1 0 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Signal about r e )ylretrauQ( etaR lanimoN Reaction Function in Normal Times 1.25 1.2 1.15 1.1 1.05 1 0.95 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Signal about r e )ylretrauQ( etaR lanimoN 14

to bind in the near future. In this case, the policymaker translates any change in beliefs about the natural rate into the nominal interest rate in the same way as a policymaker who is unconstrained by the lower bound. In Figure 2, we repeat the above experiments for the cases of incomplete and complete information while accounting for the zero lower bound on nominal short-term interest rates. As shown in the bottom panel, policymakers operating in a world of incomplete information downweight the signal they receive about the e¢ cient rate(cid:151)because they do not know that the signal is correct(cid:151)whereas policymakers under complete information do not. As can be seen, the optimal policy response in the presence of incomplete information is substantially muted relative to its complete-information counterpart in which the monetary authority translates one-for-one any change in beliefs about the natural rate into the shortterm nominal interest rates. Thus, as discussed in Aoki (2003), the optimal policy is muted under incomplete information relative to complete information. The top panel shows that when the zero lower bound is binding, incomplete information can keep optimal policy at the zero lower bound even when optimal policy with complete information would increase the short-term nominal rate. In sum, the e⁄ects of imperfect information are magni(cid:133)ed at the zero lower bound. This can be seen by comparing the distance between the lines in the top panel and the distance between the lines in the bottom panel at the far right of each chart. More generally, the combination of the zero lower bound and imperfect information makes the optimal policy response a complicated, non-linear function of signals about the state of the economy. For small revisions to the equilibriumreal rate, the policy rate remains at the zero lower bound regardless of whether information is complete or incomplete. But if the estimate of e¢ cient rate is known to be imprecise, a larger perceived increase is required to move away from the zero lower bound, and the policy responses under these conditions are even more attenuated than they are in normal times. The reason is as follows. When the policy rate is constrained by the zero lower bound and the policymaker receives a noisy signal that the e¢ cient rate has increased, the optimal policy is to make a mistake on the side of keeping the real policy rate below the perceived e¢ cient rate so as to reduce the likelihood that an adverse shock will weaken the economy enough, to require a return to the zerolowerbound. That is, policymakers (cid:147)take out insurance(cid:148)against situations inwhichthe information about the equilibrium real rate might lead them to mistakenly raise the policy rate. In contrast to the model studied by Aoki (2003) and Boehm and House (2014), the 15

Figure 2: Optimal Discretionary Policy under Alternative Information Sets Reaction Function at the ZLB 1.2 1 0.8 0.6 0.4 0.2 0 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Signal about r e )ylretrauQ( etaR lanimoN Constrained − Complete Info Constrained − Incomplete Info Reaction Function at the Cusp of the ZLB 1.2 1 0.8 0.6 0.4 0.2 0 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Signal about r e )ylretrauQ( etaR lanimoN Reaction Function in Normal Times 2 1.8 1.6 1.4 1.2 1 0.8 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Signal about r e )ylretrauQ( etaR lanimoN Complete Info Incomplete Info 16

Figure 3: Optimal Discretionary Policy near the ELB: More Volatile Economy Reaction Function at the Cusp of the ZLB 0.5 0.45 0.4 0.35 0.3 0.25 0.2 0.15 0.1 0.05 0 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Signal about r e )ylretrauQ( etaR lanimoN Unconstrained Constrained − Benchmark Volatility Constrained − More Volatile 17

zero lower bound makes the optimal policy deviate from certainty equivalence. Figure 3 showstheimplicationsforoptimalpolicyofincreasingthevarianceofshocksintheeconomy.7 In a more uncertain environment, optimal policy becomes even less responsive to positive signals about the state of the economy when the economy is at the cusp of the lower bound. Figure 4 shows the outcomes for in(cid:135)ation and the output gap when ex-post the signal about the change in the e¢ cient real rate is either correct or entirely noise. Ex-ante the optimal policy reacts to the signal itself, so the policy rate is identical whether the signal turns out to be correct or not (as shown in the top-right panel). Because the economy is at the cusp of the zero lower bound, with a su¢ ciently positive signal the central bank increases the short-term interest rate. For the same ex-ante policy, ex-post if the signal turns out to be entirely noise (the red dot-dashed line), the output gap is relatively small and in(cid:135)ation is near target. If the signal turns out to re(cid:135)ect an actual change in the natural rate, then the output gap and in(cid:135)ation move in the same direction as the natural rate. However, responses of the output gap and in(cid:135)ation are much larger when the e¢ cient rate has ex-post declined as opposed to increased. That is, it is more costly to (ex-ante) overestimate the natural rate, and to raise the nominal rate without an actual increase in the natural rate, than it is to (ex-ante) underestimate the natural rate and leave the nominal rate at the lower bound. This captures the essence of a risk-management approach at the cusp of the lower bound: Optimal policy preserves the option to maintain the zero short-term rate so that if the (exante) signal about the e¢ cient real rate is proven to be wrong the costs of misinterpreting the signal (deferring the rate increase) are low. Another way to grasp the e⁄ects of imperfect information about the e¢ cient real rate near the zero lower bound is to look at the distribution of economic outcomes. Figure 5 shows the distribution of the e¢ cient rate, the nominal rate, the output gap, and in(cid:135)ation when we simulate the economy at the cusp of the lower bound (dot-dashed lines) and in normal times (solid lines).8 As re(cid:135)ected in top-left panel, in normal times the mean of this distribution is higher as compared to when the economy is at the cusp of the lower bound. Accordingly, during normal times the distribution of nominal interest is centered to the right of the distribution of the nominal rates when the economy is at the cusp of the zero lower bound. When the economy is at the cusp of the lower bound there is substantial probability that the nominal interest rate will remain at its lower bound. For the output gap, there is 7When we increase the volatility of the underlying shocks, we keep the same signal-to-noise ratio as in the previous exercises. 8For the simulations, the innovations (including the noisy ones) for the e¢ cient real rate and marginal cost shocks are simulated according to their exogenous stochastic distributions, while initial values for the marginal cost shocks is set to zero. 18

Figure 4: The E⁄ect of Actual and Perceived Innovations to the Natural Rate Short−Term Real Rate Nominal Rate −1.1 0.2 Incorrect −1.2 Correct 0.18 0.16 −1.3 0.14 −1.4 0.12 −1.5 0.1 −1.6 0.08 −1.7 0.06 −1.8 0.04 −1.9 0.02 −2 0 −1 −0.5 0 0.5 1 −1 −0.5 0 0.5 1 Signal about r e Signal about r e Output Gap Inflation 1 2.05 0.5 2 0 1.95 −0.5 1.9 −1 1.85 −1.5 1.8 −2 −2.5 1.75 −1 −0.5 0 0.5 1 −1 −0.5 0 0.5 1 Signal about r e Signal about r e 19

Figure 5: Distributions Conditional on Alternative Natural Real Rates Distribution of Current r e Distribution of Nominal Rate 0.3 0.8 Conditional on Low r e(−1) Conditional on Average r e(−1) 0.7 0.25 0.6 0.2 0.5 0.15 0.4 0.3 0.1 0.2 0.05 0.1 0 0 −10 −5 0 5 10 0 2 4 6 8 Distribution of Output Gap Distribution of Inflation 1 0.8 0.9 0.7 0.8 0.6 0.7 0.5 0.6 0.5 0.4 0.4 0.3 0.3 0.2 0.2 0.1 0.1 0 0 −4 −3 −2 −1 0 1 2 −1 0 1 2 3 4 20

a notable di⁄erence in the two distributions, as the distribution with the economy near the cusp of the lower bound is markedly skewed downward. These adverse outcomes are what leads to the muted response we discussed above. In e⁄ect, an optimal policy that accounts for the imperfect knowledge about the e¢ cient rate will provide precautionary stimulus to o⁄set the downward bias in economic outcomes caused by the zero lower bound. 4 Simple Rules The analysis up to now has focused on optimal discretionary policy, which consists of a strategy such that the policymaker sets the nominal rate in response to incoming data on a period-by-period basis without committing itself to future actions as a way of a⁄ecting private sector expectations. The results highlight the interplay of the zero lower bound and imperfect knowledge about the e¢ cient rate in shaping optimal discretion as a basis for pursuing a risk-management approach that insures the economy against undesirable outcomes. In this section, we study the e⁄ectiveness of committing to a rule-based approach in setting monetary policy within the same economic framework. 4.1 Taylor-type Rules Common simple rules in the literature on monetary policy are Taylor-type rules which take the form of the rules studied in Taylor (1993) and Taylor (1999). We compare the performance of Taylor-type rules to optimal policy. The Taylor-type rule that we consider has the form: i = max log((cid:12)(cid:5)(cid:0) 1 );E[(cid:11)re +(cid:13) (cid:25) +(cid:13) x s ;(cid:24) ] : t f t (cid:25) t x t jf t t (cid:0) 1 g g The rule studied in Taylor (1999) serves as a baseline case and corresponds to parameter values of (cid:11) = 0, (cid:13) = 1:5, and (cid:13) = 1=4. When (cid:11) = 0, the rule does not includes a (cid:25) x time-varying intercept to re(cid:135)ect movements in the e¢ cient real rate (re). Notably, because t policymakers have imperfect information about the exact values of the underlying state of the economy, we assume that the use their best estimate of in(cid:135)ation, the output gap, and the e¢ cient real rate as arguments to the rule. Even if policymakers were to commit to a simple policy rule, operationalizing a rule for the time t interest rate that is written in terms of time t in(cid:135)ation and the time t output gap would require measurement of in(cid:135)ation and output and estimation of the deviation of output from its potential level. Our imperfect information setting is meant to capture that process. Figure 6 shows the response of the nominal interest rate to a positive signal about 21

Figure 6: Taylor Rules and Optimal Discretion at the Cusp of the ZLB Reaction Function at Cusp of the ZLB 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Signal about r e )ylretrauQ( etaR lanimoN Discretion Taylor (1999) Rule TR with Time−Varying r e 22

and increase in re. We show the policy response assuming that the economy is initially in t a situation where the equilibrium real rate is low and the zero lower bound is just binding for optimal policy under discretion. The policy response under discretion is shown as the dot-dashed red line. The policy response for the Taylor-type rule where (cid:11) = 0 corresponds to the blue solid line, which is notably higher than the optimal policy under discretion. That the Taylor-type rule prescribes higher interest rates than the optimal policy under discretion is reminiscent of results reported in Boehm and House (2014), who show that the parameter values considered by Taylor (1999) systematically under-react to movements in endogenous variables relative to the optimal policy rule. The blue dotted line shows the reaction of the Taylor-type rule with (cid:11) = 1, that is, the case when the intercept term of the Taylor rule adjusts with perceived movements in the e¢ cient real rate. The prescriptions of this rule are notably lower than the case in which (cid:11) = 0 because the central bank(cid:146)s estimate of the e¢ cient real rate directly enters the rule. That the prescriptions of this rule are similar to optimal discretionary policy might not be entirely surprising given that optimal policy in a model that ignores the zero lower bound (in the absence of markup shocks) is to set the nominal interest rate equal to the e¢ cient real rate. The fact that the rule does mimic optimal discretion when (cid:11) = 1 re(cid:135)ects the presence of the zero lower bound which makes optimal policy even more accommodative than the policy prescribed by the rule. Figure7showsthedistributionsofoutcomes, conditionalontheeconomybeingatthe cuspofthezerolowerbound(asassumedintheprevious(cid:133)gure)fortheTaylor-typeruleswith (cid:11) = 0 and (cid:11) = 1. The top panel shows the distribution of the time t nominal interest rate. Because of our assumed initial conditions for the equilibrium real rate, there is a signi(cid:133)cant probability that the nominal rate remains at the zero lower bound for optimal policy under discretion as well as under the Taylor rule with a time-varying intercept term that tracks perceived movements in the natural real rate. The Taylor rule with a constant intercept equal to the steady state real rate produces a distribution of higher nominal interest rates, with much less mass at the zero lower bound. This result is a direct e⁄ect of the relatively high prescriptions for the policy rate coming from a Taylor rule with a constant intercept (shown in the previous (cid:133)gure). TherelativelyhighnominalinterestratefromtheTaylorrulewithaconstantintercept leads to lower average outcomes for the output gap and in(cid:135)ation, shown in the middle and bottompanelsofFigure7. Thatthisruleunderreactstoperceivedchangesintheequilibrium real rate is re(cid:135)ected in the relatively dispersed distributions for the output gap and in(cid:135)ation 23

Figure7: ConditionalDistributions(attheCuspoftheZLB)underAlternativeTaylorRules Nominal Rate 1.2 Discretion 1 Taylor (1999) Rule TR with Time−Varying r e 0.8 0.6 0.4 0.2 0 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 Output Gap 0.8 0.6 0.4 0.2 0 −4 −3 −2 −1 0 1 2 Inflation 0.8 0.6 0.4 0.2 0 0 0.5 1 1.5 2 2.5 3 3.5 4 24

when compared to the outcomes under optimal discretionary policy. By contrast, if the intercept of the Taylor rule re(cid:135)ects perceived changes in the equilibrium real rate (the case of (cid:11) = 1), the distributions of the nominal interest rate, the output gap, and in(cid:135)ation closely resemble the outcomes under optimal discretion. The di⁄erences between the two speci(cid:133)cations of these Taylor-type rules are also apparent in an unconditional sense. Figure 8 shows the unconditional distributions of the nominalinterestrate, theoutputgap, andin(cid:135)ationunderoptimalpolicyandtheTaylor-type ruleswhen(cid:11) = 0andwhen(cid:11) = 1. Thedistributionsofoutcomeswhen(cid:11) = 1areverysimilar tothoseunderoptimalpolicy. Giventhecloserelationshipbetweenthetwopolicieswhenthe zero bound is ignored, this is perhaps not surprising. The Taylor-type rule with a constant intercept performs notably worse than the others in that the distributions of in(cid:135)ation and output are much more dispersed than under optimal discretionary policy. These dispersed outcomes are the result of the rule systematically under-responding to movements in the e¢ cient real interest rate, which can be seen by the relatively thin tails of the distribution of the nominal interest rate, shown in the top panel. In our simulations, the nominal interest rate under optimal policy spends more time, on average, at the zero lower bound than either speci(cid:133)cation of the Taylor-type rule. This re(cid:135)ects the nonlinearity in the optimal policy response that causes policymakers to ease policy in the vicinity of the zero lower bound as a risk-management approach to policy. That is, deviating from a Taylor-type rule with a constant intercept term by keeping policy lower near the zero lower bound is an optimal response, which can be approximated by adjusting the Taylor-rule intercept to changes in the central bank(cid:146)s best estimate of the e¢ cient real rate. 25

Figure 8: Unconditional Distributions under Alternative Taylor Rules Nominal Rate 0.3 Discretion 0.25 Taylor (1999) Rule TR with Time−Varying r e 0.2 0.15 0.1 0.05 0 0 1 2 3 4 5 6 7 8 9 10 Output Gap 1 0.8 0.6 0.4 0.2 0 −3 −2 −1 0 1 2 3 Inflation 0.8 0.6 0.4 0.2 0 0 0.5 1 1.5 2 2.5 3 3.5 4 26

Table 1. Unconditional Welfare Losses for Alternative Monetary Policy Rules Contributions of Welfare In(cid:135)ation Output Policy Rules Losses ((cid:25) E((cid:25)))2 E((cid:25)) (cid:25) (x E(x))2 E(x) (cid:0) (cid:0) (cid:0) Optimal 5.375 0.02233 -0.0038 0.2877 -0.051 Taylor-type Constant Intercept 14.720 0.0265 0.00017 3.0203 0.0018 Time-Varying Intercept 5.529 0.0221 -0.0077 0.3356 -0.1211 Constant Intercept, Optimal 5.654 0.0224 -0.0050 0.3747 -0.0327 First-Di⁄erence, Optimal 19.830 0.0153 -0.0032 5.3746 -0.0228 The unconditional distributions of the in(cid:135)ation and the output gap determine the relative performance of these rules as evaluated by the monetary authority(cid:146)s unconditional loss function. Table 1 shows the relevant moments as well as the unconditional losses across rules. Notably, the zero lower bound shifts the means of the distributions down for almost all of the rules, although by small amounts.9 Di⁄erences in relative performance across rules are almost entirely driven by di⁄erences in the dispersion of the distributions of the output gap and in(cid:135)ation. The more dispersed outcomes under the standard Taylor rule ((cid:11) = 0) lead to greater losses than the other speci(cid:133)cation of the Taylor rule ((cid:11) = 1). Interestingly, in our model neither speci(cid:133)cation performs better than optimal discretion. While the Taylor-type rule with a constant intercept performs relatively poorly for the parameter values chosen above, we also consider an optimal version of the rule, where the parameters (cid:13) and (cid:13) are chosen to minimize the unconditional loss function of the (cid:25) x monetary authority. The optimal parameter values (which are about (cid:13) = 18 and (cid:13) = 10) (cid:25) x are much higher than those commonly used in the literature on monetary policy rules.10 Notably, they imply that the monetary authority should optimally respond more strongly to deviations of in(cid:135)ation and output from their targets than prescribed by more-traditional parameter values.11 This aggressive response is evidenced by the fatter tails of the unconditional distribution of the nominal interest rate with the optimal Taylor-type rule with a 9All of the numbers in Table 1 are expressed in percent. 10The welfare loss is relatively (cid:135)at for relatively high values of these coe¢ cients. 11These results are reminiscent of results reported in Boehm and House (2014), who do not consider the interest-rate lower bound. 27

Figure 9: Unconditional Distributions under Optimal Taylor Rules Nominal Rate 0.3 Discretion 0.25 Taylor (1999) Rule Optimal Taylor Rule 0.2 0.15 0.1 0.05 0 0 1 2 3 4 5 6 7 8 9 10 Output Gap 1 0.8 0.6 0.4 0.2 0 −3 −2 −1 0 1 2 3 Inflation 0.8 0.6 0.4 0.2 0 0 0.5 1 1.5 2 2.5 3 3.5 4 28

constant coe¢ cient in Figure 9, relative to the Taylor-type rule with a constant intercept and our baseline parameters. Interestingly, as shown in Figure 9, the unconditional distribution of the nominal interest rate under the optimal Taylor-type rule is similar to the distribution under optimal policy, howeveroptimalpolicyhasasomewhatfatterlefttail, eventhoughitspendslesstime at the zero lower bound, re(cid:135)ecting the complex nonlinear reaction function of the monetary authority under optimal policy. Notably, the distributions of in(cid:135)ation and output under the optimal Taylor-type rule are similar to the distributions under discretion. That is, the outcomesunderoptimalpolicy, inourmodel, canbewellapproximatedbyaTaylor-typerule that either adjusts its intercept to changes in the equilibrium real rate or by a Taylor-type rule that responds aggressively to deviations of in(cid:135)ation and output from their target values. Although the unconditional distributions are similar, the optimal Taylor rule still slightly underperforms optimal discretionary policy as shown in Table 1. 4.2 First-di⁄erence Rules Because of the uncertainty surrounding the equilibrium real rate, some researchers have argued that rules that depend on it are (cid:135)awed (e.g., Orphanides and Williams (2002)). As an alternative, they have advocated for (cid:133)rst-di⁄erence rules as a guide for the conduct of policy, as these rules do not require responding to changes in the equilibrium real rate because they are rules for the change in the nominal interest rate, rather than the level. We revisit this debate by evaluating the performance of (cid:133)rst-di⁄erence rules when there is imperfect information about the equilibrium real rate and compare their performance to the Taylor-type rules discussed above, as well as optimal discretionary policy. We consider (cid:133)rst-di⁄erence rules of the form: 1 i = max log((cid:12)(cid:5)(cid:0) );i +E[(cid:13) (cid:25) +(cid:13) (y y ) s ;(cid:24) ] : t t 1 (cid:25) t y t t 1 t t 1 f (cid:0) (cid:0) (cid:0) jf (cid:0) g g Although this rule eliminates the equilibrium rate from its speci(cid:133)cation, policymakers still need to form expectations about (now-cast) current in(cid:135)ation and the current level of output. We specify the rule in terms of output growth, rather than the output gap as in the Taylortype rules above, because Orphanides and Williams (2002) have emphasized that responding totheoutputgaprequiresestimatingthepotentiallevelofoutput, whichisunobservableand potentially di¢ cult to learn about. However, unlike in Orphanides and Williams (2002), we assume that the monetary authority is not perfectly informed about the output and in(cid:135)ation terms that enter the rule at time t. 29

Figure 10: Unconditional Distributions under Optimal First-Di⁄erence Rules Nominal Rate Discretion 0.3 Optimal First−Difference Rule 0.25 0.2 0.15 0.1 0.05 0 0 1 2 3 4 5 6 7 8 9 10 Output Gap 1 0.8 0.6 0.4 0.2 0 −3 −2 −1 0 1 2 3 Inflation 1 0.8 0.6 0.4 0.2 0 0 0.5 1 1.5 2 2.5 3 3.5 4 30

Figure 10 shows the unconditional distributions of the nominal interest rate, the output gap, and in(cid:135)ation under the (cid:133)rst-di⁄erence rule in which the coe¢ cients have been chosen to minimize the unconditional loss function facing the monetary authority. Even when we optimally choose the parameters, the unconditional distribution of the nominal interest rate from the (cid:133)rst-di⁄erence rule is less dispersed than under optimal policy. While we (cid:133)nd that the optimal (cid:133)rst-di⁄erence rule responds aggressively to in(cid:135)ation (cid:13) = 2:1, we (cid:25) (cid:133)nd that the optimal coe¢ cient on output growth is almost zero. This muted response to changesinoutputgrowthleadstothewidedispersionoftheoutputgap, showninthemiddle panel of (cid:133)gure. While the unconditional distribution of in(cid:135)ation under the (cid:133)rst-di⁄erence rule is less dispersed than the distribution of in(cid:135)ation under optimal policy, this reduction in variance does not lead to better welfare outcomes (as shown in the last row of Table 1) than under discretionary optimal policy because of the wide distribution of the output gap. The reason that the optimal (cid:133)rst-di⁄erence rule does not respond more strongly to output or in(cid:135)ation is that doing so would also respond more strongly to incorrect signals. When noise in the monetary authority(cid:146)s information set at time t enters the (cid:133)rst-di⁄erence rule, the noise is incorporated into the level of the nominal interest rate and is not discarded in the next period. That is, the (cid:133)rst-di⁄erence rules carries imperfect information forward, magnifying the e⁄ects. By contrast, Taylor-type rules (without interest rate smoothing) do not carry forward past misperceptions about the state of the economy directly in the rule. While we could of course lag the in(cid:135)ation and output terms in the (cid:133)rst-di⁄erence rule by one period so that there would no longer be any quantities in the rule about which the monetary authority has imperfect information, that solution would be speci(cid:133)c to the informationstructureassumedinourmodel. Ifmeasuresofin(cid:135)ationandoutputwillneverbe free of noise (e.g. measurement error), even after many subsequent periods of measurement, the optimal policy problem will always have to confront the imperfect information problem, and if imperfectly observed quantities enter a (cid:133)rst-di⁄erence rule, misperceptions will be carried forward. 5 Conclusions Our analysis has shown that, in a simple NK model, imperfect knowledge about the equilibrium real interest rate interacts with the zero lower bound to cause optimal policy under discretion to respond to signals about an increase in the equilibrium real rate by less than it would during more-normal circumstances. Because optimal policy under discretion in the neighborhood of the zero lower bound prescribes a complicated, nonlinear reaction function 31

for the policy rate that may, in practice, be di¢ cult to communicate and implement, we also study simple policy rules. In our model, we show that Taylor-type rules that either include a time-varying intercept that moves with perceived changes in the equilibriumreal rate or that respond aggressively to deviations of in(cid:135)ation and output from their target levels perform similarly to optimal discretionary policy. While we do not explicitly model credibility and communications concerns for the monetary policy authority, these simple rules may be easier to implement and communicate to the public. Our analysis of (cid:133)rst-di⁄erence rules highlights that rules with smoothing terms carry forward current and past misperceptions about the state of the economy and can lead to suboptimal performance. 32

Appendix A Detailed Description of the Model The economy consists of a representative household,acontinuumof(cid:133)rmsproducingdi⁄erentiatedintermediategoods,perfectlycompetitive (cid:133)nal goods (cid:133)rms, and a central bank in charge of monetary policy. Households There is a representative household that values streams of consumption C and hours worked H according to preferences given by the utility function: t t 1 E (cid:14) (log[C ] (cid:22) H ) (9) t t 1 t t t (cid:0) (cid:0) t=1 X where E represents the mathematical expectation conditional on all exogenous shocks and t endogenous prices and quantities up to time t. The household maximizes expected utility (cid:135)ows subject to a sequence of budget constraints, P C +R 1B = W H +B +T : (10) t t t(cid:0) t t t t 1 t (cid:0) where household nominal expenditures on consumption at date t are given by P C , where t t P denotes the aggregate price level, B are nominal bonds sold at the price R 1, and t t t(cid:0) R = (1+i ) denotesthegrossnominal returnonthesebonds. Thenominal bondspurchased t t by a household pay one unit of the numØraire in the next period with certainty. A household receives income from any bonds carried over from last period in addition to its labor income, W H , from supplying its labor services to the economy(cid:146)s (cid:133)rms. A household also pays lumpt t sumtaxes and receives dividends fromintermediate goods (cid:133)rms, the sumof which is denoted by T . The labor supply shock, (cid:22) , is assumed to follow an AR(1) process in logs, such that: t t log((cid:22) ) = (cid:26) log((cid:22) )+" (11) t (cid:22) t 1 (cid:22);t (cid:0) where (cid:26) denotes the persistence of the shock and (cid:27) denotes the standard deviation of the (cid:22) (cid:22) innovations, " . (cid:22);t To capture exogenous changes in the household(cid:146)s desire to save, we allow for a shock t 1 (cid:0) to the discount rate. The rate of time discounting at time t is given by (cid:14) = (cid:12)t (cid:17) ; t s (cid:18)s=0 (cid:19) where 0 < (cid:12) < 1. Hence, (cid:14)t+1 = (cid:12) , and (cid:17) is the shock to the discount ratQe (or the (cid:14)t (cid:17)t+1 t+1 natural rate of interest, as will be clear below). The discount rate shock follows an AR(1) process in logs, such that: log((cid:17) ) = (cid:26) log((cid:17) )+" (12) t (cid:17) t 1 (cid:17);t (cid:0) 33

where (cid:26) denotes the persistence of the shock and (cid:27) denotes the standard deviation of the (cid:17) (cid:17) innovations, " . A decrease in (cid:17) represents an exogenous factor that induces a temporary (cid:17);t t rise in a household(cid:146)s propensity to save, and reduces current aggregate household demand for goods. The (cid:133)rst order necessary conditions for optimality from the household(cid:146)s problem can be written as follows: (cid:12) C 1 = R E t (cid:5) 1 ; (13) (cid:17) t t C (cid:0)t+1 t (cid:26) t+1 (cid:27) (cid:22) w = C (14) t t t where w = Wt denotes the real wage, and (cid:5) = Pt+1 is the in(cid:135)ation rate between t and t Pt t+1 Pt t+1. Thelinearspeci(cid:133)cationforlaborservicesinequation(9)alongwithacompetitivelabor market implies that there is a perfectly elastic supply of labor available to the economy(cid:146)s (cid:133)rms. Firms There is a continuum of monopolistically competitive (cid:133)rms producing differentiated intermediate goods. The latter are used as inputs by perfectly competitive (cid:133)rms that produce a single (cid:133)nal good, Y , using a constant returns to scale production technolt (cid:15) ogy, Y t = 0 1 Y t (j) (cid:15) (cid:0)(cid:15) 1 dj (cid:15) (cid:0) 1, where Y t (j) is the quantity of intermediate good j used as an input, a(cid:16)nd (cid:15) > 1 is the(cid:17)elasticity of substitution. Pro(cid:133)t maximization and perfect com- R (cid:15) petition yield the set of demand schedules, Y (j) = Pt(j) (cid:0) Y and aggregate price index, t Pt t 1 (cid:16) (cid:17) P = 1 P (j)1 (cid:15) dj 1 (cid:0) (cid:15). t 0 t (cid:0) (cid:16)The productio(cid:17)n function for intermediate good j is given by: R Y (j) = H (j): (15) t t Since intermediate goods are imperfect substitutes, the intermediate goods-producing (cid:133)rms sell their output in a monopolistically competitive market. During period t, the (cid:133)rm sets its nominal price P (j), subject to the requirement that it satis(cid:133)es the demand of the represent tative (cid:133)nal goods producer at that price. As in Rotemberg (1982), the intermediate good producer faces a quadratic cost of adjusting its nominal price between periods, measured in 2 terms of the (cid:133)nished good and given by: ’ Pt(j) 1 Y , where ’ governs the obstacles 2 (cid:5)(cid:22)Pt 1(j) (cid:0) t (cid:0) to price adjustment and (cid:5) denotes the cent(cid:16)ral bank(cid:146)s in(cid:17)(cid:135)ation target. The cost of price adjustment makes the problem of the intermediate good producer dynamic; that is, it chooses (cid:22) P (j), taking as given (cid:5), to maximize its present discount value of expected pro(cid:133)ts: t 1 (cid:14) t 1 P t (j) 1 (cid:0) (cid:15) P t (j) (cid:0) (cid:15) ’ P t (j) 2 E 0 C (cid:0) P Y t (cid:0) P (1+(cid:28))(cid:22) t w t Y t (cid:0) 2 (cid:5) (cid:22) P (j) (cid:0) 1 Y t ; X t=1 t ( (cid:18) t (cid:19) (cid:18) t (cid:19) (cid:20) t (cid:0) 1 (cid:21) ) 34

where (cid:14)t 1 measures the marginal value(cid:151)in utility terms(cid:151)to the representative household C(cid:0) t of an additional unit of real pro(cid:133)ts received in the form of dividends during period t, w is t the (cid:133)rm(cid:146)s real marginal cost. (cid:28) is an employment subsidy. The (cid:133)rst order condition for this problem is: P (j) (cid:15) P (j) (cid:15) 1 P (j) P (j) t (cid:0) t (cid:0) (cid:0) t t (1 (cid:15)) +(cid:15) (1+(cid:28))(cid:22) w ’ 1 (16) (cid:0) P P t t (cid:0) (cid:5) (cid:22) P (j) (cid:0) (cid:5) (cid:22) P (j) (cid:18) t (cid:19) (cid:18) t (cid:19) (cid:20) t 1 (cid:21) t 1 (cid:0) (cid:0) (cid:12) C P (j) P (j) P Y t t+1 t+1 t t+1 +’ E 1 = 0: (cid:17) t C (cid:5) (cid:22) P (j) (cid:0) (cid:5) (cid:22) P (j) P (j) Y t (cid:26) t+1 (cid:20) t (cid:21) t t t (cid:27) Notice that under (cid:135)exible prices, i.e. ’ = 0, then expression (16) leads to the well-known expression where (cid:133)rm set prices as a time-varying mark-up, (cid:15) , over marginal cost, i.e. (cid:15) 1 (cid:0) P = (cid:15) (1+(cid:28))(cid:22) w . t (cid:15) 1 t t (cid:0) Market Clearing and Symmetric Equilibrium Conditions In a symmetric equilibrium, all intermediate goods producing (cid:133)rms make the same decisions so that aggregate output satis(cid:133)es Y = Y (j) and aggregate labor satis(cid:133)es H = H (j) j. It then follows t t t t 8 that Y = H and goods market clearing can be written in aggregate terms as: t t ’ (cid:5) 2 t Y = C + 1 Y : (17) t t 2 (cid:5) (cid:22) (cid:0) t (cid:18) (cid:19) Households trade the nominal bond among themselves so that bond market clearing implies B = 0. t The symmetric equilibrium conditions from households and (cid:133)rms as well as the goods market clearing condition are: (cid:12)R C 1 = t E t (cid:5) 1 ; (18) G(cid:17) t C (cid:0)t+1 t (cid:26) t+1 (cid:27) (cid:5) (cid:5) ’(cid:12) C (cid:5) (cid:5) Y t t t t+1 t+1 t+1 ’ 1 = E 1 +(cid:15)C (1+(cid:28))(cid:22) +(1 (cid:15)) (19) (cid:5) (cid:22) (cid:0) (cid:5) (cid:22) (cid:17) t C (cid:5) (cid:22) (cid:0) (cid:5) (cid:22) Y t t (cid:0) (cid:18) (cid:19) t (cid:26) t+1 (cid:18) (cid:19) t (cid:27) ’ (cid:5) 2 t Y = C + 1 Y (20) t t 2 (cid:5) (cid:22) (cid:0) t (cid:18) (cid:19) These conditions along with the monetary policy rule and the stochastic processes for the shocks must be satis(cid:133)ed in equilibrium. Equation (18) is the consumption Euler equation that jointly with the next equation, the aggregate resource constraint, constitute the basis for the aggregate demand relationship. Equation (19) is the nonlinear version of the New Keynesian Phillips curve that relates current in(cid:135)ation to past and expected future in(cid:135)ation and output. 35

Natural and E¢ cient Levels The natural rate of output and the real interest rate are de(cid:133)ned as the levels in an associated economy without price rigidities. The (cid:135)exible price equilibrium is characterized by the following two equations: (cid:12)RRn Yn 1 = t E t (cid:17) t Yn t (cid:26) t+1(cid:27) (cid:15) 1 (1+(cid:28))(cid:22) Yn = (cid:0) t t (cid:15) We assume that, in the non-stochastic steady state, the employment subsidy is set such that it o⁄sets the monopoly distortion. A log-linearized approximation around the natural equilibrium steady state yields: rn = (cid:17) E (cid:22) (cid:22) t t (cid:0) t f t+1 (cid:0) t g yn = (cid:22) t (cid:0) t Thus, there are two sources of exogenous disturbances to the natural rate of interest: shocks to the discount factor and shocks to the marginal cost. A decrease in the desired to save (an increase in (cid:17) ) and expected decrease in marginal costs cause the natural rate of t interest to rise. The natural level of output moves with exogenous changes in the marginal costs. The e¢ cient equilibrium corresponds to that of (cid:135)exible prices and no exogenous variation in marginal costs. Formally, the e¢ cient allocations can be described as follows: (cid:12)RRe Ye 1 = tE t (cid:17) t Ye t (cid:26) t+1(cid:27) Ye = 1 t The log-linear approximation around the e¢ cient non-stochastic steady state implies that: re = (cid:17) t t and the e¢ cient level of output is constant and normalized to one. Thus, shocks to the marginal costs reduce output relative to its e¢ cient level. Solution Method To solve the nonlinear model, under optimal discretionary policy, we use a projection method similar to Gust et al. (2012) and Christiano and Fisher (2000). The solution algorithm involves parameterizing the unknown functions E x = fx((cid:24) ;s ;(cid:24) ), t t+1 t t t 1 (cid:0) E (cid:25) = f(cid:25)((cid:24) ;s ;(cid:24) ); where E denotes rational expectations based on private sector t t+1 t t t 1 t (cid:0) 36

information. Wealsoparameterizethefunctionfr(s ;(cid:24) ),whichsatis(cid:133)esthecentralbank(cid:146)s t t 1 (cid:0) (cid:133)rst order condition. In particular, this function can be derived by substituting the private sector equilibrium conditions into the central bank(cid:146)s (cid:133)rst order condition and rewriting it as: (cid:30) (cid:20)2 i t = fr(s ;(cid:24) )+E re + (cid:22) s ;(cid:24) t (cid:0) (cid:21)+(cid:20)2 t t (cid:0) 1 f t (cid:21)+(cid:20)2 t jf t t (cid:0) 1 gg (cid:18) (cid:19) where i is the steady state value of the nominal interest rate. To solve for the optimal policy we use the fact that (cid:30) = 0 when i is greater than or equal to the lower bound constraint. t t To be consistent with the information structure, the function fr does not take (cid:24) t as an argument. This ensures that the solution imposes that the monetary authority is unable to use information encoded in the private sector decisions in the current period. To solve for the unknown parameters of these functions, we conjecture a guess and iterate until the parameters satisfy the private sector equilibrium conditions and the central bank(cid:146)s (cid:133)rst order condition at a (cid:133)nite number of points. Note that the above expression can be used to determine the short-term interest rate and the Lagrange multiplier (cid:30) as follows. t If the right-hand side is positive (fr(s ;(cid:24) ) + E re + (cid:20)2 (cid:22) s ;(cid:24) > 0) then t t (cid:0) 1 f t (cid:21)+(cid:20)2 t jf t t (cid:0) 1 gg i = fr(s ;(cid:24) )+E re + (cid:20)2 (cid:22) s ;(cid:24) ; otherwise(cid:16)i = 0(cid:17)and (cid:30) = [fr(s ;(cid:24) )+ t t t (cid:0) 1 f t (cid:21)+(cid:20)2 t jf t t (cid:0) 1 gg t t (cid:0) t t (cid:0) 1 E re + (cid:20)2 (cid:22) s ;(cid:24) (cid:16) ]. (cid:17) f t (cid:21)+(cid:20)2 t jf t t (cid:0) 1 ggg (cid:16) (cid:17) 37

References Klaus Adam and Roberto M Billi. Discretionary monetary policy and the zero lower bound on nominal interest rates. Journal of monetary Economics, 54(3):728(cid:150)752, 2007. Kosuke Aoki. On the optimal monetary policy response to noisy indicators. Journal of monetary economics, 50(3):501(cid:150)523, 2003. OlivierBlanchardandJordi Gal(cid:237). Real wagerigiditiesandthenewkeynesianmodel. Journal of Money, Credit and Banking, 39(s1):35(cid:150)65, 2007. ChristophEBoehmandChristopherLHouse. Optimaltaylorrulesinnewkeynesianmodels. Technical report, 2014. Lawrence J Christiano and Jonas DM Fisher. Algorithms for solving dynamic models with occasionallybindingconstraints. Journal of Economic Dynamics and Control, 24(8):1179(cid:150) 1232, 2000. Richard Clarida, Jordi Gal(cid:237), and Mark Gertler. The science of monetary policy: Evidence and some theorj. Journal of Economic Literature, 37:1661(cid:150)1707, 1999. G(cid:252)nter Coenen and Volker Wieland. The zero-interest-rate bound and the role of the exchange rate for monetary policy in japan. Journal of Monetary Economics, 50(5):1071(cid:150) 1101, 2003. Olivier Coibion, Yuriy Gorodnichenko, and Johannes Wieland. The optimal in(cid:135)ation rate in new keynesian models: should central banks raise their in(cid:135)ation targets in light of the zero lower bound? The Review of Economic Studies, page rds013, 2012. Gauti B Eggertsson and Michael Woodford. Zero bound on interest rates and optimal monetary policy. Brookings Papers on Economic Activity, 2003(1), 2003. Charles Evans, Jonas Fisher, Francio Gourio, and Spencer Krane. Risk management for monetary policy near the zero lower bound. Brookings Papers on Economic Activity, March 2015. Jordi Gal(cid:237). Monetary Policy, In(cid:135)ation and the Business Cycle: An Introduction to the New Keynesian Framework. Princeton University Press, 2008. Christopher J Gust, David L(cid:243)pez-Salido, and Matthew E Smith. The empirical implications of the interest-rate lower bound. 2012. 38

Andrew T. Levin, David L·łpez-Salido, Edward Nelson, and Tack Yun. Limitations on the e⁄ectivenessofforwardguidanceatthezerolowerbound. International Journal of Central Banking, March 2010. Anton Nakov. Optimal and simple monetary policy rules with zero (cid:135)oor on the nominal interest rate. International Journal of Central Banking, 4(2):73(cid:150)127, June 2008. Athanasios Orphanides and John C Williams. Robust monetary policy rules with unknown natural rates. Brookings Papers on Economic Activity, (2), 2002. David Reifschneider and John C Williams. Three lessons for monetary policy in a lowin(cid:135)ation era. Journal of Money, Credit and Banking, pages 936(cid:150)966, 2000. Julio Rotemberg and Michael Woodford. An optimization-based econometric framework for the evaluation of monetary policy. NBER Macroeconomics Annual, 12:297(cid:150)361, 1997. Julio J Rotemberg. Sticky prices in the united states. The Journal of Political Economy, pages 1187(cid:150)1211, 1982. Lars EO Svensson and Michael Woodford. Indicator variables for optimal policy. Journal of monetary economics, 50(3):691(cid:150)720, 2003. Lars EO Svensson and Michael Woodford. Indicator variables for optimal policy under asymmetric information. Journal of Economic Dynamics and Control, 28(4):661(cid:150)690, 2004. John B Taylor. Discretion versus policy rules in practice. Carnegie-Rochester conference series on public policy, 39:195(cid:150)214, 1993. John B Taylor. A historical analysis of monetary policy rules. In Monetary policy rules, pages 319(cid:150)348. 1999. Michael Woodford. Interest rate and prices. Princeton University Press, 2003. 39