تغییرات آموزش و سیاست پولی

This paper estimates a dynamic stochastic equilibrium model in which monetary policy follows a nominal interest rate rule that is subject to regime switches in the target inflation rate. Two specifications are considered: agents know the current state of monetary policy (full information) and agents use Bayesian updating to infer the policy regime (learning). First, our policy regime estimates are consistent with the view that policy was marked by a shift to a high-inflation regime in the early 1970s which ended with Volcker's stabilization policy. Second, while Bayesian posterior odds favor the full-information version of the model, the fall of interest rates, actual and expected inflation in the early 1980s is better captured by the delayed response of the learning specification. Third, monetary policy shocks of up to two standard deviations essentially do not trigger the Bayesian learning mechanism. Yet due to nonlinearities, interventions that lead to small initial interest rate changes may be associated with much larger effects on output and inflation than under full information.

an unanticipated deviation—a policy shock—from the perceived policy rule, and its consequences are evaluated through impulse response functions. However, extended systematic deviations from the perceived policy rule may lead agents to change their beliefs about the conduct of monetary policy and invalidate the VAR impulse response predictions. One method to overcome this problem is to estimate a fully-specified dynamic stochastic general equilibrium (DSGE) model that can be re-solved for alternative policy rules to predict effects of fundamental changes in the policy regime. However, this approach faces some conceptual difficulties as well (e.g. Sims, 1982 and Cooley et al., 1984). First, for several periods after the regime change the agents are potentially uncertain whether the policy shift was temporary or permanent and the transition dynamics are possibly affected by the learning process. Second, the extent to which past data can be used to validate the predictions is very limited, since the policy change typically has no precedent. In this paper we estimate a basic New Keynesian monetary DSGE model, along the lines of Woodford (2003), in which monetary policy follows a nominal interest rate rule that is subject to regime switches in the target inflation rate. In the first version of the model agents take the possibility of regime shifts into account when they translate observed monetary policy into expectations about future output, prices, and interest rates. They use a Bayesian learning rule to infer the current state of monetary policy. Under the second specification agents have full information about the state of monetary policy. Unlike in the models considered by Sargent (1999), Cogley and Sargent (2004) and Primiceri (2004), our regime switching framework offers no explanation why monetary policy shifts occur over time. We simply assume that there are high-inflation and low-inflation regimes and that the transition probabilities stay constant. While firms and households make inference about the conduct of monetary policy, the central bank itself does not attempt to learn about the effectiveness of their policy and choose policy in an optimal manner. For the empirical analysis we combine a prior distribution with the likelihood functions derived from the structural model specifications and conduct Bayesian inference. First, we estimate the model parameters and the monetary policy regimes for the post-war US. Our estimates are consistent with the view that policy was marked by a shift to a high-inflation regime in the early 1970s which ended with Volcker’s stabilization policy at the beginning of the 1980s. The regime-switching model fits the data better than the standard specification in which the target inflation rate is assumed to be constant. Markov-switching models have been used, for instance, by Sims (2000) and Sims and Zha (2004) to study timevariation in monetary policy. Sims (2000) considers a univariate policy reaction function that is subject to regime shifts, whereas Sims and Zha (2004) extend the analysis to an identified VAR. While these papers find frequent oscillation between regimes and emphasize heteroskedasticity of the structural shocks, our estimation uncovers essentially two distinct shifts of monetary policy. Second, we study the empirical evidence in favor of the learning mechanism. While Bayesian posterior odds favor the full-information version of the model, the fall of interest rates, actual and expected inflation during and after the disinflation episode in the early 1980s is better captured by the delayed response of the learning specification. DSGE models with shifting policy regimes have recently been analyzed by Andolfatto et al. (2002), Andolfatto and Gomme (2003), and Erceg and Levin (2003). The first two papers consider cash-in-advance models. Andolfatto et al. (2002) show that the agents’ learning can explainthe failure of conventional tests of unbiasedness of inflation expectations. Andolfatto and Gomme (2003) use their model to study the Canadian disinflation episode. Our findings with respect to the effects of a disinflation policy in the learning and full-information environment, by and large, resemble the results reported by Andolfatto and Gomme (2003). Erceg and Levin (2003) demonstrate that the learning mechanism is able to generate inflation persistence in a DSGE model with staggered nominal contracts. However, none of these papers formally estimates the full DSGE model. Third, we examine the magnitude of the expectation-formation effect of monetary policy interventions in the learning specification by comparing its impulse responses to those under full information. Leeper and Zha (2003) consider a simple two-equation model with an exogenous money-growth rate rule that is subject to regime shifts and propose a measure of modesty of policy interventions. The authors conjecture that conditional on a particular policy regime variations of the monetary policy shock of up to two standard deviations do not trigger expectation-formation effects. Leeper and Zha claim that as long as policy interventions do not fall outside of the historical two-standard deviation band their effects can be well approximated by impulse response functions from a VAR that has been estimated based on the relevant policy regime. According to our estimates it is indeed correct that monetary policy shocks of up to two standard deviations essentially do not trigger the learning mechanism and the associated impulse responses under learning and full information are very similar. However, due to the nonlinearity introduced by the learning mechanism there often exist multiple values of the policy shock that lead to the same initial interest rate response. As a consequence we find that under learning a 25 basis point drop in interest rates is consistent with a much larger rise in output growth and inflation than under full information. This prediction cannot be generated with a linear VAR. The likelihood-based estimation of a DSGE model with policy-regime shifts in general leads to a very complicated nonlinear filtering problem that takes a long time to solve on a conventional computer.1 We make two simplifications. First, we use a log-linear approximation of the DSGE model, and second we assume that the policy rule depends only on observed variables, except for the regime indicator and the policy shock, and not on latent model variables such as potential output. We show that under these assumptions it is fairly straightforward to compute the likelihood function of the regime switching model, using a modification of the Kalman filter. Methods described in Schorfheide (2000) are employed to evaluate the posterior distributions. The paper is organized as follows. Section 2 presents the monetary DSGE model and its approximation. Section 3 describes the econometric approach and provides some details about the computation of the likelihood function. Empirical results are presented in Section 4 and Section 5 concludes.

We have estimated a simple New Keynesian monetary DSGE model that has become a popular benchmark model for the analysis of monetary policy. Unlike in earlier econometric work, monetary policy follows a rule that is subject to regime shifts. While our model provides no explanation why these regime shifts occur, we assume that the public has potentially incomplete information about the state of monetary policy and has to learn about the current regime. Our regime estimates are consistent with the popular story that monetary policy is characterized by a high-inflation regime in the 1970s which ended with Volcker’s stabilization policy in the early 1980s. The evidence on the importance of learning and uncertainty about the policy regime is mixed. Posterior probabilities of the full information versus learning specification of the DSGE model favor the former. On the other hand, a closer look at the disinflation episode in the early 1980s indicates that the fall of inflation and interest rates is better explained by the delayed response of the learning specification. The presence of a learning mechanism has potentially important consequences for the prediction of the effects of policy interventions. A prolonged intervention might be interpreted by the agents as a shift to a new policy regime and lead to changes in the agents’ expectation formation. We document these expectation formation effects with our estimated DSGE models. Monetary policy shocks of up to two standard deviations essentially do not trigger the Bayesian learning mechanism. Yet due to nonlinearities, interventions that lead to small initial interest rate changes may be associated with much larger effects on output and inflation than under full information.