The full text on this page is automatically extracted from the file linked above and may contain errors and inconsistencies.
Assessing U.S. Aggregate Fluctuations Across Time and Frequencies WP 19-06 Thomas A. Lubik Federal Reserve Bank of Richmond Christian Matthes Federal Reserve Bank of Richmond Fabio Verona Bank of Finland Assessing U.S. Aggregate Fluctuations Across Time and Frequencies Thomas A. Lubik Federal Reserve Bank of Richmondy Christian Matthes Federal Reserve Bank of Richmondz Fabio Verona Bank of Finlandx February 2019 Working Paper No. 19-06 Abstract We study the behavior of key macroeconomic variables in the time and frequency domain. For this purpose, we decompose U.S. time series into various frequency components. This allows us to identify a set of stylized facts: GDP growth is largely a high-frequency phenomenon whereby in‡ation and nominal interest rates are characterized largely by low-frequency components. In contrast, unemployment is a mediumterm phenomenon. We use these decompositions jointly in a structural VAR where we identify monetary policy shocks using a sign restriction approach. We …nd that monetary policy shocks a¤ect these key variables in a broadly similar manner across all frequency bands. Finally, we assess the ability of standard DSGE models to replicate these …ndings. While the models generally capture low-frequency movements via stochastic trends and business-cycle ‡uctuations through various frictions, they fail at capturing the medium-term cycle. JEL Classification: C32, C51, E32 Key Words: Wavelets, bandpass …lter, SVAR, sign restrictions, DSGE model We wish to thank Bob King, Annika Lindblad, and Denis Tkachenko for useful comments. We are also grateful to participants at the 2018 CEF Conference in Milan, the University of Adelaide, Carleton University, the 2019 Meeting of the Finnish Economic Association and the Bank of Finland. James Geary provided exceptional research assistance. The views expressed in this paper are those of the authors and should not be interpreted as those of the Federal Reserve Bank of Richmond, the Federal Reserve System, or the Bank of Finland. y Research Department, P.O. Box 27622, Richmond, VA 23261. Email: thomas.lubik@rich.frb.org. z Research Department, P.O. Box 27622, Richmond, VA 23261. Email: christian.matthes@rich.frb.org x Monetary Policy and Research Department. Snellmaninaukio, PO Box 160, 00101 Helsinki. Email: fabio.verona@bof.…. 1 1 Introduction Economists have often found it useful to separate long-run trends from business-cycle ‡uctuations, which generally are considered those that occur with a cycle length of between two and eight years. On the statistical side, this approach is probably best characterized by the idea of a trend-cycle decomposition as in Beveridge and Nelson (1981), where the trend is associated with permanent movements in a time series as opposed to a business cycle being driven by transitory shocks. Conceptually, this idea is also inherent in …ltering methods such as the Hodrick-Prescott (HP) …lter, which has been the dominant approach in business-cycle modeling to extract a trend from aggregate times and render them stationary. Such decompositions are convenient since they align with the idea of economic ‡uctuations as being driven by either permanent or temporary shocks that do not necessarily interact. In addition, monetary policy is often framed in terms of stabilizing the ‡uctuations of key variables around a trend that is una¤ected by policy. However, there is a growing awareness in the macroeconomics literature that this common view of economic ‡uctuations is no longer adequate to characterize the behavior of economic activity over time. For instance, Comin and Gertler (2006) argue that a substantial part of economic ‡uctuations is located in what they label a ‘medium-term cycle’, that is, ‡uctuations that are beyond a length of eight years, but fall short of a trend. Moreover, these medium-term ‡uctuations cannot be thought of in isolation from other frequency bands. Using a theoretical model, Comin and Gertler (2006) show that business cycles and medium-term cycles are intimately connected since they are driven by the same underlying temporary shock. Speci…cally, a temporary innovation to, say, productivity or the policy rate can reverberate throughout several frequency bands as they get propagated over time.1 Against this background, we aim to provide a somewhat more encompassing view of cyclical behavior across all frequencies. In particular, we study three issues. First, we compute a decomposition of key macroeconomic time series using wavelet-based …ltering. That is, we decompose a time series into several time series components, each of them ‡uctuating within a speci…c frequency band. We …nd the use of wavelets advantageous for our purposes since this …ltering approach is more ‡exible than standard Fourier analysis and more traditional bandpass …ltering. In particular, it allows di¤erent frequency movements to be more pronounced in some parts of the sample than others and thereby reveals time 1 Cogley (2001) makes a similar point for trend speci…cations where he shows the e¤ects of trend speci…cation errors are not con…ned to low frequencies, but are spread across the entire frequency domain. Researchers, therefore, have to have a clear understanding of the inter-relatedness of frequency bands for which a wavelet approach o¤ers a covenient tool. 2 variation in the importance of di¤erent frequency components. The second question looks at the e¤ects of identi…ed monetary policy shocks across di¤erent frequency bands to assess the plausibility of medium-term cycles as being generated by temporary shocks. The third question asks whether standard dynamic stochastic general equilibrium (DSGE) models that are used in monetary policy analysis can replicate the volatility of di¤erent cycles of each macroeconomic variable under consideration and are thereby useful in addressing the policy questions raised. We establish three main …ndings. First, the wavelet decomposition of key macroeconomic variables shows that the bulk of ‡uctuations in GDP growth, unemployment, and in‡ation occurs across di¤erent frequency bands. More than half of real GDP growth is explained by short-term, high-frequency components with only a third of ‡uctuations attributable to business-cycle frequencies between two and eight years. Unemployment is dominated by medium-term ‡uctuations between eight and 32 years and, to a lesser extent, by low-frequency movements while close to three-quarters of in‡ation and short-term interest rate ‡uctuations fall into the slow-moving trend component. The corollary to these results is that business cycles play only a secondary role in explaining overall aggregate ‡uctuations as real GDP growth is very much a high-frequency phenomenon, while the behavior of in‡ation is all trend. Since these variables are central to thinking about monetary policy, both in terms of target variables as well as their information content for the state of the economy, we next assess the e¤ects of monetary policy shocks on the individual frequency components. Using identi…ed structural VARs with sign restrictions, we …nd that across all frequency bands the results from an aggregate VAR carry over to individual components and short-term, business-cycle, medium-term, and long-term components. In a baseline speci…cation that includes only the overall data series, a contractionary policy shock, that is, an increase in the federal funds rate, lowers in‡ation, raises the unemployment rate, and decreases real GDP growth. We …nd similar patterns across most frequency bands, but as we increase the cycle length, the peak response moves further out, while precision of the impulse response estimates worsens and the quantitative importance declines. We take this as somewhat tentative evidence that monetary policy has an impact across all frequency bands and that a mechanism in line with interaction of endogenous growth and cycles as in Comin and Gertler (2006) is at play. In addition, we …nd that in the long run the relationship between the nominal interest rate and the in‡ation rate is positive, whereas in the short run an interest-rate increase lowers in‡ation. This relationship weakens or is non-existent over the 3 medium term, which arguably re‡ects a contrast between a demand e¤ect in the short run and the Fisher e¤ect in the long term. Our third …nding shows that standard DSGE models are in principle capable of replicating the behavior of macroeconomic variables in di¤erent frequency bands. We simulate arti…cial time series from three canonical DSGE models (Smets and Wouters, 2007; del Negro et al., 2015; and Christiano et al., 2016) and apply our wavelet decomposition to the same set of variables as before. Generally, all three models perform reasonably well for business-cycle frequencies and for long-term ‡uctuations. In a sense, this is perhaps not surprising in that the models are built as business-cycle models around the idea that such ‡uctuations are the outcome of stochastic shocks and endogenous propagation. These DSGE models also include elements such as habit formation, investment adjustment costs, and wage and price indexation to impart persistence on the variables, which helps match behavior at business-cycle frequencies.2 Long-run behavior is captured by stochastic trends and time-varying in‡ation targets, which have been introduced successively over the course of model development to capture trends. We show, however, that these models largely fail in capturing behavior at medium-term frequencies, which is particularly prevalent in the case of unemployment and a monetary DSGE model with search and matching frictions in the labor market. We interpret these …ndings as a challenge for modelers to develop frameworks capable of capturing medium-term cycles. This paper touches upon various literatures in macroeconomics and time series analysis. There has been a long-standing debate as to whether a frequency-based view of economic ‡uctuations is useful for analyzing and understanding policy. Perhaps emblematic of a critical viewpoint is Watson (1993), who argues that policy analysis at di¤erent frequencies is not relevant for policymakers and that the close relationship between a time series representation of a variable and its counterpart in the frequency domain, such as the spectrogram, invalidates the need for a separate analysis of frequency-speci…c considerations. This viewpoint is implicitly questioned by Onatski and Williams (2003), who study the e¤ects of uncertainty, broadly understood, on monetary policy decisions. They show that when uncertainty enters a policymaker’s decision problem at di¤erent frequencies it may have substantially di¤erent e¤ects on outcomes. This criticism of the Watson-critique is taken up by Brock et al. (2007), who analyze the di¤erential e¤ects of various policy rules on outcomes across frequencies. In a follow-up paper, Brock et al. (2013) demonstrate how 2 Tkachenko and Qu (2012) and Sala (2015) estimate medium-size DSGE models in the frequency domain with a focus on business-cycle frequencies. They report similar …ndings as to the ability of such models to replicate observed behavior over the cycle. 4 reductions of variance at some frequencies lead to increases in variance at others, which then creates a policy trade-o¤. Our paper informs this debate by showing empirically the contributions of di¤erent frequency bands to the overall volatility of key macroeconomic variables and how they are impacted by monetary policy shocks. Our paper also continues and contributes to the debate about the use of detrending methods in macroeconomics. Many empirical methods require the underlying data series to be stationary and thereby necessitate the use of a …lter to remove trending components. However, as Canova (1998) has demonstrated, di¤erent detrending methods extract di¤erent information from the underlying data series. This implies that the thus derived stylized facts can di¤er substantially qualitatively and quantitatively across di¤erent …ltering methods.3 This insight is extended by Gorodnichenko and Ng (2010) and to the estimation of DSGE models. When researchers apply standard data transformations, this induces biases in structural estimates and distortions in the policy conclusions. In order to address this issue Canova (2014) proposes joint modeling of the cycle and the trend within the model and the raw data. We add to this literature by establishing a set of stylized facts based on the timefrequency decomposition inherent in wavelet analysis that has certain advantages over more traditional methods. Thereby, we also highlight the importance of joint theoretical modeling of economic behavior across all frequency bands and especially the medium term as an important component of economic ‡uctuations. While the importance of the medium run has been on economists’ minds for a long time (e.g., Blanchard, 1997), there has been a ‡urry of recent research in the wake of Comin and Gertler’s (2006) contribution that studies the origin and e¤ects of medium-term cycles (e.g., Beaudry et al., 2017; Cao and Huillier, 2018). In this paper, we exploit the bene…ts of wavelet analysis as a complementary approach to classical time series and spectral analysis. We …rst use the univariate wavelet transform for exploratory data analysis of US macroeconomic variables. In addition, we use the wavelet power spectrum to analyze the evolution over time of the variance of the variable at di¤erent frequencies. We then use this approach to isolate speci…c frequency components from each variable and use those frequency components in a standard VAR regression setup. Our paper thus contributes to a growing literature on the use of alternative …ltering methods in economics and …nance, such as Aguiar-Conraria et al. (2012) and Bandi et al. (2019). The remainder of the paper is structured as follows. In the next section, we present 3 This observation is also in line with the recent criticism in Hamilton (2018) on the use and application of the HP-…lter in macroeconomic modeling. 5 our …rst set of results, namely new stylized facts based on a wavelet decomposition of aggregate data. In Section 3, we use the decomposition to assess the e¤ects and importance of monetary policy shocks across di¤erent frequency bands in a structural VAR framework. Section 4 considers the question, whether existing DSGE models are able to capture these regularities. Section 5 concludes. 2 A Frequency-Band Decomposition of Aggregate Time Series We use the wavelet methodology to decompose standard US macroeconomic time series into di¤erent components that can be associated with the scale of the underlying cycles. We regard this time-frequency decomposition, that is, a decomposition of a variable into components in the time domain with precise counterparts in the frequency domain, as a useful and informative alternative to typical trend-cycle decompositions that provides a more encompassing view of the nature of economic ‡uctuations. In what follows, we brie‡y discuss the methodology and detail the data used in our empirical exercise. We then present our baseline results, followed by an extensive robustness analysis with respect to alternative …ltering methods and choices. 2.1 Methodology and Data The analysis in this paper is based on a time-frequency decomposition of key economic time series. Our basic objective is to decompose a time series into individual components that can be cleanly and clearly associated with ‡uctuations at di¤erent frequencies or di¤erent lengths of a cycle, but are represented in the time domain. For this purpose, we employ wavelet multiresolution analysis (MRA), which performs such decomposition in a way similar to the traditional time series trend-cycle decomposition approach (e.g., Beveridge and Nelson, 1981; Watson, 1986), or other …ltering methods like the Hodrick and Prescott (1997) or the Baxter and King (1999) band-pass …lter. However, a wavelet approach aims for a more …ne-grained understanding of the di¤erent components of a time series that make up what is considered a ‘cycle’as opposed to a ‘trend’.4 Speci…cally, we employ a particular version of a wavelet transformation of a time series called the Maximal Overlap Discrete Wavelet 4 Conceptually, our line of reasoning is informed by the notion of medium-term cycles as advocated by Comin and Gertler (2006). There is a growing understanding that the neat trend-cycle view of economic ‡uctuations is inadequate to capture the nature of economic activity. 6 Transform (MODWT).5 As an example, by using the speci…c form of a Haar wavelet …lter, any time series Xt can be decomposed into a scale component SJ;t and J detail components Dj;t : J X Xt = Dj;t + SJ;t ; (1) j=1 where these coe¢ cients are given by: 0 2j 1 1 1 @ X Dj;t = Xt 2j i=0 2J SJ;t = 1 1 X Xt i : 2J i j 1 2X i=2j 1 1 Xt i A ; (2) (3) i=0 Intuitively, the wavelet …lter separates the original series Xt , which is de…ned in the time domain, into di¤erent time series components. These represent the ‡uctuations of Xt in a speci…c frequency band, that is, a range of frequencies, or length of cycles, that are grouped together.6 In this example, the smooth scale component SJ;t at time t is computed as the weighted average of lagged values of Xt at scale J, while the detail components Dj;t are overlapping weighted moving averages up to scale J. The bands are associated with di¤erent details j such that for small j, the wavelet component Dj;t captures the higher-frequency characteristics of the time series, that is, its short-term ‡uctuations. As j increases, the components represent lower frequency movements of the series. Finally, the smooth component SJ;t captures the lowest frequency dynamics, that is, the long-term behavior.7 The key parameter for the economic interpretation of the wavelet decomposition is the scale J, which determines how …ne-grained or detailed the decomposition is. For J large enough, the scale component SJ;t approximates the true underlying trend of the series. If J is small, then the scale component includes ‡uctuations of shorter duration, which one may not normally associate with a trend.8 An alternative interpretation is that SJ;t is the 5 The MODWT version of the wavelet …lter has become the standard in the empirical …nance and forecasting literature, e.g. Berger (2016) or Faria and Verona (2018). 6 The individual components, or wavelets, thus make up the overall wave in a prescribed manner. 7 As in the Beveridge and Nelson (1981) time-series decomposition into stochastic trends and transitory components, the wavelet coe¢ cients Dj;t can be viewed as components with di¤erent levels of calendartime persistence operating at di¤erent frequencies, whereas the scaling component SJ;t can be seen as the low-frequency trend of the time series under analysis. 8 The Appendix contains a simple example how the scale parameter J is related to the idea of taking various di¤erences of time series. 7 underlying scale of the time series upon which ‡uctuations of higher frequencies and shorter cycle durations are built. In that sense, our analysis falls in line with a more standard trendcycle decomposition. On a …nal note, the …lter discussed above is one-sided since future values are not needed to compute the wavelet coe¢ cients of the transform of Xt at time t. This implies that that the Dj;t and SJ;t lag Xt . Moreover, since the length of the …lters increases with j, so does the delay. Hence, the coarser the scale, the more the Dj;t and SJ;t are lagging Xt . We use this fact in our VAR analysis below. What distinguishes the wavelet decomposition is that the choice of the scale allows the researcher to hone in on and isolate speci…c frequency bands that are the objects of interest. While other …ltering methods, such as Fourier analysis, also allow a researcher to focus on speci…c frequencies, a wavelet approach has some key advantages. Traditional decomposition techniques, such as spectral analysis of a time series, tend to impose strong assumptions about the data-generating process. Speci…cally, they often require data to be stationary or pre-…ltered. However, many economic and …nancial time series are hardly stationary as they exhibit trends and patterns such as structural breaks, volatility clustering, and long memory, which the wavelet approach can handle with ease. Unlike Fourier analysis, wavelets are de…ned over a …nite window in the time domain, which is automatically and optimally resized according to the frequency of interest and the choice of the scale J. Wavelets and standard Fourier analysis are essentially approximations with basis functions, but Fourier basis functions are non-zero almost everywhere, making it harder for them to capture local phenomena. Using a short time window isolates the high-frequency features of a time series, while treating the same signal with a large time window reveals its low-frequency features. By varying the size of the time window, we can therefore capture time-varying and frequency-varying features of the time series at the same time. Wavelets are, thus, very useful when dealing with non-stationary time series, irrespective of whether the non-stationarity comes from the level of the time series (that is, from a long-term trend or jumps) or from higher-order moments (that is, from changes in volatility). Wavelet …ltering methods are similar to …ltering by a set of band-pass …lters so as to capture the ‡uctuations of a time series in di¤erent frequency bands, e.g., Christiano and Fitzgerald (2003). The band-pass …lter is a combination of a Fourier decomposition in the frequency domain with a moving average in the time domain. It applies optimal Fourier …ltering to a sliding window in the time domain with constant length regardless of the frequency being isolated. Wavelet …ltering, in contrast, provides better resolution in the 8 time domain as the wavelet basis functions are both time-localized and frequency-localized. In this paper, we use the MODWT to compute the decomposition. This version is not restricted to a particular sample size: if the data are discrete, the standard wavelet decomposition requires a sample of length 2J for the decomposition to be exact; that is, it imposes a tight restriction on which and how many frequency bands can be considered and might require dropping observations. The MODWT avoids this problem and is also translation-invariant, that is, it is not sensitive to the choice of a starting point for the examined time series. Finally, implementation of the wavelet decomposition requires choice of a speci…c functional form for the …lter that maps the original series into its components. We follow the literature and choose as a benchmark the Haar …lter, but we also consider the Daubechies …lter as an alternative. Speci…cally, we employ the …lter to decompose our time series of interest into seven individual series, labeled D1 , ..., D6 for the detail components and S6 for the scale component; that is, we choose J = 6. The individual components are such that they add up to the underlying series. Given the scale of the decomposition as powers of two we can associate the components with individual frequency bands. Speci…cally, D1 captures ‡uctuations up to four quarters, D2 between four and eight quarters, up to D6 , which covers the band between 64 and 128 quarters. The scale component S6 is associated with movements above 128 quarters. We collect quarterly data on US macroeconomic aggregates, interest rates, and prices. Speci…cally, we report results for real GDP, the unemployment rate, the in‡ation rate for the overall personal consumption price index (PCE), the federal funds rate (FFR), and a 3-month and 10-year interest rate.9 The data are described in more detail in the Appendix. The full range of our sample covers 1954Q3 to 2017Q3. We utilize data in levels and in growth rates, where growth rates are computed as quarter-over-quarter values. Although not required for the wavelet …ltering, we report results for GDP growth as it is the focus of policymakers’decisions. For our baseline decomposition, we use a one-sided Haar …lter, which is then employed in the VAR analysis. In a sense, the di¤erent scale components are generated regressors where we do not want to impart information onto the econometrician running the VAR that he could not possibly possess; that is, knowledge of the data at the end of sample should not be used to produce a decomposition for periods in the middle. For informative purposes and as a robustness check, we also provide results for two-sided 9 The 3-month Treasury rate at constant maturity is only available from 1981Q4 on. We use the 3-month Treasury rate from secondary market instead since it is available from 1947Q1. Preliminary analysis for the two series shows that they co-move extremely closely and that there is at most a level di¤erence of up to 50 basis points. 9 (smoothed) wavelet …lters, for alternative kernels, and for alternative …lters, such as the Hodrick-Prescott and the Christiano-Fitzgerald bandpass …lters. 2.2 Baseline Results We report two sets of results. For purposes of exposition, we group the seven series into four categories that we label ‘Short Term’ (D1 , D2 ), ‘Business Cycle’ (D3 , D4 ), ‘Medium Term’(D5 , D6 ), and ‘Long Term’(S6 ). The short-term category captures high-frequency ‡uctuations under two years, which in macroeconomic applications are often discarded as noise, but may contain useful information about the incidence of shocks. The business-cycle category covers ‡uctuations at frequencies between 8 and 32 quarters (2-8 years), which most macroeconomic research on the sources of aggregate movements focuses on. This frequency band is, for instance, designed to be isolated by the application of the Hodrick-Prescott …lter with a smoothing parameter of = 1; 600. We maintain this terminology for clarity, although one aspect of our paper is to argue for less rigid classi…cations in the standard trend-cycle methodology. Components D5 and D6 are grouped under ‘Medium Term’‡uctuations and cover frequencies up to 128 quarters (32 years). We note that this scale is shorter than the medium-term cycle adopted in Comin and Gertler (2006), de…ned as movements between 8 and 50 years. Finally, we associate S6 with the ‘Long Term’ or, loosely speaking, the trend. We report the grouped wavelet decompositions for real GDP growth, the unemployment rate, the in‡ation rate, the federal funds rate, the 3-month and 10-year rate, and the di¤erence between the latter two series, namely the term spread, in Figures 1-6. The decompositions into the individual wavelets are collected in the Appendix. Table 1 reports the variance decompositions by frequency. We …nd that more than 50 percent of overall ‡uctuations in real GDP growth are explained by the short-term components D1 and D2 , roughly one-third by the businesscycle components D3 and D4 , with the rest by medium- to long-term components.10 This raises the question whether and to what extent macroeconomic stabilization policy can a¤ect this short-term component, especially since it is likely to contain measurement error. At the same time, the low-frequency component S6 declines from above 4 percent to below 2 percent over the course of the sample (see Figure 1). This is in line with the secular decline in trend growth that has been found in numerous studies. However, this is not the full picture behind the recent lower growth rates, as the two medium-term components D5 and D6 essentially o¤set each other since 2000 and thereby do not contribute to the 10 The medium-frequency components as de…ned by Comin and Gertler (2006) thus make up only 12.5 percent of the overall ‡uctuations, with half falling on the band between 32-64 quarters. 10 underlying growth trend. This comes largely from the business cycle components during the recovery from the Great Recession. The Great Moderation is most visible in the short-term components and to a lesser extent in the business-cyle band.11 The wavelet decomposition shows that it is more of a higher frequency phenomenon. This observation lends support to the argument that the Great Moderation came about because of an improvement in the way monetary stabilization policy was conducted rather than a change in, for instance, inventory management. The unemployment rate decomposition in Figure 2 and Table 1 reveals a slightly different pattern. Roughly one-third of unemployment ‡uctuations are due to short-term and business-cycle movements, while medium- and longer-term frequencies (D5 -S6 ) each explain around 20 percent. Fluctuations in the unemployment rate can therefore be described as a medium-term cycle. What dominates the level of the unemployment rate is its long-term component S6 , which could be interpreted loosely as a natural rate of unemployment. A focus of the next section is the extent to which the trend components are a¤ected by monetary policy. What is striking is that the di¤erent components do not seem to comove closely. For instance, the unemployment rate is at 5.4 percent in 1990, while the long-term component S6 is at 7.2 percent, the di¤erence being made up by components D4 -D6 . In other words, the business-cycle peak produces a negative unemployment gap relative to a very high natural rate on account of strong medium-term components that might be tied to labor force participation peaking in the late 1990s. Finally, the Great Moderation is considerably less visible in the unemployment rate, if at all. We now turn to the nominal side of the economy. Figure 3 contains the results from the decomposition of the PCE in‡ation rate. 40 percent of in‡ation movements can be traced back to the long-term component S6 . The business cycle component explains around one…fth of the overall variability, while medium-term components cover 25 percent. About 15 percent of in‡ation variability can be traced back to very short-term or noise components. As in the case of the unemployment rate, the scale of the decomposition is dominated by the trend S6 . The monetary policy literature often interprets this component as the in‡ation target or the perception thereof. It can also be seen as a measure of the extent to which in‡ation expectations are anchored. In our decomposition, it shows a gradual rise from almost zero in the late 1960s to a peak of 6.2 percent in the early 1980s followed by a gradual decline until reaching the 2 percent target in the 2000s. 11 Figure C.1 in the Appendix shows that this largely due to the D3 component, indicating that the Great Moderation is essentially a high-frequency event. To this point, see also Aguiar-Conraria et al. (2012) and Pancrazi (2015). 11 A similar pattern in terms of the Volcker disin‡ation can be found in the medium-term components D5 and D6 . What is striking is the run-up in trend in‡ation over the course of the 1970s and the drawn-out, three-decade struggle to return it to 2 percent. Since the Federal Reserve arguably did not change its implicit in‡ation target over that time, this component may therefore be better described as the public’s perceived target. Our results then depict a striking loss of central bank credibility.12 In light of this aspect, it is perhaps surprising that there is not much of the Great Moderation visible when interpreted as a binary event, that is, a break in policy or a structural change before or after the early 1980s. Instead, in the graphs in Figure 3 it is possible to discern the high volatility of the 1970s, preceded and followed by the more stable 1960s and 1980s, respectively. Interestingly, in‡ation volatility seems to have gone up again in the 2000s, especially around the Great Recession. We report decompositions for the FFR and the 10-year rate in Figures 4 and 5.13 They show similar patterns as the in‡ation decompositions, whereby volatility in the 10-year rate can be attributed to almost 70 percent to the long-term component S6 , ten percentage points more than for the short rates. Presumably, this re‡ects that longer rates are less subject to the vagaries of higher-frequency ‡uctuations. Since the interest rates share common components, especially in the medium and longer run, it is therefore often instructive to consider the term spread, in our case the di¤erence between the 10-year and the 3-month rate. The term spread decomposition in Figure 6 puts most weight, almost 45 percent, on the business-cycle components. This supports the idea that at frequencies commonly associated with the business cycle the spread is a useful indicator of economic and …nancial conditions. Interestingly, the long-term component has gone up considerably since the early 1980s to a level of above 2%, implying that the di¤erence between the short and long rates has become more persistent. As a …nal exercise, we produce the power spectra from the wavelet decomposition for real GDP growth, unemployment, and the federal funds rate in Figures 7-9.14 The wavelet-based spectra are akin to classical spectra in that they decompose a time series into frequencies and measure the contribution of each frequency to the overall behavior of a time series.15 12 This interpretation is consistent both with the learning and inherent in‡ation persistence story in Primiceri (2006) or Sargent, Williams, and Zha (2006), the in‡ation misperception argument in Lubik and Matthes (2016), as well as a number of recent papers on evolving private sector beliefs, for instance, in Bianchi (2013). 13 The results for the 3-month rate are almost identical to those for the FFR. The respective decompositions are shown in the Appendix. 14 We brie‡y describe and discuss the concept behind the wavelet power spectra in the Appendix. 15 As discussed before, the wavelet power spectrum does not require stationarity. 12 Ordered by frequency, this spectral density is depicted in the right column of each …gure. In addition, we show the time-frequency decomposition in the left column. Using the wavelet …lter, the graph reports the wavelet power spectrum (WPS), a decomposition by time (on the horizontal axis) and frequency (on the vertical axis). The graph is coded as a heat map such that areas of higher activity are depicted as redder on the color spectrum. The solid black lines delineate 95 percent con…dence regions. Note that the wavelet-based spectral density can be obtained by integrating the WPS over time. The spectral peak of GDP growth in Figure 7 is just above two years, which con…rms our prior …nding that output ‡uctuations are highly concentrated among the highest frequencies. The WPS, however, shows that this observation is largely driven by the late 1950s and 1970s, which show considerably higher concentration of activity in the short term than what occurred during later periods. In the same vein, the Great Moderation is quite visible from the graph.16 The unemployment rate in Figure 8 shows two local spectral peaks associated with cycles of around 8 and 32 years. This is in line with our previous …ndings, but sharpens the observation of unemployment being subject to medium-term cycles towards the edges of that frequency band. Notably, the Great Moderation is not apparent from the WPS, while deep recessions in the mid-1970s and the Great Recession impart some higher frequency components on the decomposition. Finally, Figure 9 reveals that the FFR has a spectral peak at a very low frequency, which we associate with the presence of an explicit or implicit in‡ation target. However, the WPS shows a local peak at a frequency of 8 years, which is largely driven by the period from 1968 until the early 2000s. Overall, what emerges from these decompositions is a multifaceted picture of macroeconomic ‡uctuations. Across all variables, the business-cycle components D3 - D4 , that is, cycles between two and eight years, explain about one-third of overall ‡uctuations. There is considerable heterogeneity across variables as far as the other components are concerned: 50 percent of real GDP growth is captured by high frequency components (cycles of less than 2 years). Essentially, the majority of quarterly GDP movements occur at very high frequencies.17 In turn, short-run ‡uctuations do not seem to play much of a role for the other variables. The behavior of unemployment is dominated by medium-term movements with a cycle length of between 8 and 32 years and to a lesser extent by longer-term move16 This is in line with related research by Pancrazi (2015), who argues that the reduction of volatility of GDP after the mid-1980s is mainly a high-frequency phenomenon of cycles up to 4 years and that it is much milder, or absent, for other frequencies. 17 We use …nal data in our empirical study, that is, the last data vintage available. In contrast, policymakers operate in a real-time environment where initial data releases are subject to sometimes large measurement error. Lubik and Matthes (2016) show that this can lead to what looks like policy mistakes ex post. 13 ments with a lower frequency than that. In‡ation and interest rates have sizeable long-term components, too. These components can be interpreted as “trends”and natural or potential rates. Their behavior arguably conforms to conventional wisdom, that is, in‡ation seems to be all trend, driven by the Federal Reserve’s implicit, and then later, explicit in‡ation target. This naturally raises the question, whether stabilization policy aimed at the business cycle is misdirected or misses important aspects that policymakers should focus on.18 An immediate follow-up question is whether models that are being used to describe and analyze monetary policy are consistent with the heterogeneity in ‡uctuations. We address these two questions in turn in the following two sections. First, we investigate whether identi…ed monetary policy shocks have di¤erential e¤ects on key variables for di¤erent frequencies; and second, we study whether some standard DSGE models are capable of replicating the wavelet-based variance decompositions in this section. 2.3 A Comparison of Alternative Filters We assess the robustness of our baseline …ndings for the one-sided Haar …lter along several dimensions. First, we consider a two-sided version of the Haar …lter. The second exercise considers an alternative kernel for the wavelet decomposition, namely the Daubechies …lter. The third robustness check uses …lters that are more common in the macroeconomics literature, speci…cally the Christiano-Fitzgerald bandpass …lter and the Hodrick-Prescott …lter. As before we focus on four broad frequency bands for exposition. The decompositions are reported in Figures 10-12. Figures 10 and 11 contain the decompositions of, respectively, real GDP growth and unemployment for the one- and two-sided Haar …lter and the Daubechies …lter. By de…nition the two-sided …lter is smoother than a one-sided …lter since it uses all available information over the whole span of the sample and not just up to the data point at which the …lter is applied. This is evident by comparing the one-sided Haar …lter with its two-side counterpart in the …gures. Generally, there are no large di¤erences in terms of the overall direction and volatility for both unemployment and real GDP growth, but the one-sided …lter imparts more volatility to the short-term and business-cycle components than the other …lters. Moreover, the one-sided Haar …lter lags the other …lters in picking up general directional movements. This is especially visible in the medium- and long-term components of unemployment in Figure 11.19 The fact that the one-sided Haar is slow in picking up the 18 19 This argument has been made most succinctly by Brock et al. (2008, 2013). We …nd similar patterns in the decompositions for in‡ation and the interest rates. These results are 14 rise and subsequent fall in trend unemployment in the 1970s and 1980s is simply a feature of how it is constructed. As discussed before, we prefer a one-sided …lter since we use the individual components as variables in a VAR, which rests on the idea that the innovations are one-step-ahead forecast errors and thereby do not re‡ect the full information in the sample. The …gures also report results for the Debauchies …lter as an alternative to the twosided Haar …lter. The Haar …lter produces less volatile components than the Daubechies, but the di¤erence seems minor. There are a few episodes where the two …lters do not overlap each other. For instance, the medium-term components of in‡ation in the mid1970s di¤er noticeably, but these occurrences are the exception. We prefer the Haar over the Debauchies implementation of the wavelet decomposition since the former has a more intuitive interpretation (see the discussion in the Appendix). The di¤erences between the various implementations of the decomposition are small enough, however, not to a¤ect the conclusions drawn in the next two sections.20 In contrast, the decompositions based on two widely used …lters in macroeconomic analyses are materially di¤erent. Figure 12 compares our baseline …lter with the corresponding bandpass …lter of Christiano and Fitzgerald (2003) (CF) and the canonical Hodrick-Prescott (HP) …lter. In a sense, the CF …lter and our Haar …lter are conceptually similar in that they explicitly isolate speci…c frequency bands and represent them in the time domain. This is evident from comparing the two …lters in the …gure for unemployment rate decompositions as an illustrative example. At business-cycle frequencies, the CF …lter extracts more volatile components, but is arguably not that di¤erent from the wavelet-based …lter. The exception are the longer-term components, especially D6 and S6 , where the two …lters pick out di¤erent peaks and are generally not that well aligned.21 In contrast, the HP …lter produces quite di¤erent series. For a smoothing parameter of = 1; 600, it extracts the business-cycle frequencies corresponding to our components D3 and D4 . It is considerably more volatile than the wavelet decomposition. More striking is the pattern for lower frequencies. The …gure reports the HP trend, which is computed as the di¤erence between the original series and the business-cycle component obtained with = 1; 600. It is akin to the S6 component with wavelets, which is the “residual”part of the included in the Appendix. 20 We performed the empirical excercises in Sections 3 and 4 using alternative wavelet decompositions. The results are available on request. 21 However, recall that the CF …lter is optimized to extract business-cycle frequencies and not low frequencies. At the same time, this is another argument in favor of wavelet …lters since it treats all frequencies the same way. 15 series. This slow-moving component is quite di¤erent from the other series and thus raises concerns as to whether the HP …lter introduces spurious dynamics (see Hamilton, 2017). 3 The Frequency-Speci…c E¤ects of Monetary Policy Shocks We now study whether and to what extent monetary policy shocks a¤ect key aggregate variables across di¤erent frequencies. The previous section demonstrated that the behavior of GDP growth, unemployment, and the in‡ation rate di¤ers in terms of the contribution of various frequency bands to overall volatility. Whereas the majority of ‡uctuations in GDP growth are located among the highest frequencies, that is, the short-term components, the unemployment rate is more evenly split between a large medium-term and lower-frequency components. In turn, most of the movements in the in‡ation rate are driven by the long term, which we might associate with the in‡ation target. As we think of monetary policy as trying to stabilize movements in GDP growth and unemployment against a background of stable prices or constant in‡ation, the question is whether policy is successful in a¤ecting these variable at frequencies that are the main drivers of their overall volatility. Our approach is as follows. We assess the e¤ects of monetary policy shocks on individual frequency bands by using the …ltered series as explanatory variables in a VAR. Given a plausible identi…cation of policy shocks, we then compute impulse response functions to these shocks for the various decompositions. We begin by assessing the plausibility of our preferred identi…cation scheme in a standard model. To this end, we estimate a threevariable VAR in an activity variable, that is, either the unemployment rate or real GDP growth, in‡ation, and the federal funds rate. We then identify a structural monetary policy shock using a sign restriction approach where we assume that all restrictions are imposed only on impact. Speci…cally, we assume that a contractionary monetary policy shock - one that raises the federal funds rate on impact - lowers output, increases unemployment, and lowers in‡ation. Figure 13 reports impulse responses to an identi…ed policy shock from the two VARs. The left column shows the responses in the model with unemployment, the right those of the model with GDP growth. In this baseline speci…cation, a rise in the interest rate by 25 basis points increases unemployment by 10bp with a hump-shaped peak after three to four quarters of 20bp. It lowers in‡ation by 60bp on impact before gradually returning to its longrun level. Similarly, a contractionary monetary policy shock lowers GDP growth by almost 1.5 percentage points and in‡ation shows a similar decline as in the other speci…cation.22 22 Both the unemployment rate and GDP overshoot their long-run level in their adjustment path after 16 In the next step, we add the frequency components to the baseline speci…cation, either in terms of unemployment or real GDP growth, as an activity variable. For each speci…cation we identify the policy shock separately, which allows for the possibility that there could be di¤erences across models. We consider two alternative speci…cations. First, we add the seven frequency bands, D1 -S6 , of each variable included in the VAR one by one to the baseline speci…cation. This results in a six-variable VAR, estimated separately for each band. We report selected impulse responses for GDP growth in Figures 14-16, where the left column shows the responses of the aggregate variables and the right column the corresponding responses for a frequency band. The respective responses with unemployment as the activity variable can be found in the Appendix. We …nd that the responses of the high-frequency components D2 are signi…cant, and are in line with the baseline results and what theory would suggest; however, the responses are not large quantitatively and economically small. Nevertheless, this indicates that the monetary transmission mechanism works as theoretical reasoning and practical experience would indicate. The response of the business-cycle component D4 is not signi…cant on impact but becomes more sharply estimated a few quarters out. As before, the direction of the responses is consistent with the identi…cation scheme on the overall series. Contractionary shocks are thus likely to have their strongest impact in a few quarters, which is in line with the idea that monetary policy stabilizes business cycles with a lag. Finally, the response for the long-term component S6 is drawn out and not signi…cant over the business-cycle horizon, but it exhibits comovement between the federal funds rate response and in‡ation. In other words, at longer horizons and cycles, the Fisher e¤ect, namely that interest rates and in‡ation rates are positively correlated, comes through; whereas at higher frequencies this correlation moves in the opposite direction as the demandconstricting e¤ect of higher rates reduces in‡ation. It is clear from these …ndings that in the transition between high frequency and low frequency movements the comovement patterns for these two variables switch. The second VAR speci…cation adds the …ltered series in groups that represent broader frequency bands. Since the wavelet decomposition is fully additive, we cannot include all individual series. We therefore focus on a speci…cation that looks at the business-cycle components (D3 + D4 ), the medium-term cycles (D5 + D6 ), and the long term (S6 ). This results in a 12-variable VAR, where we identify the policy shock by imposing sign restrictions on impact on the aggregate variables only. Figure 17 contains the respective responses the shock in line with the interest rate responses. That is, monetary policy responds endogenously to the worsening economic conditions due to the unanticipated contraction by loosening policy. 17 for the model with GDP growth, while the corresponding responses for unemployment are reported in Figure 18. In each graph, the top row contains the aggregate responses, followed by the short-term, medium-term, and long-term components in separate rows. A contractionary monetary policy shock has a negative impact on real activity in each frequency band whereby the largest response is for the business-cycle component. If we just look at the short-term frequencies (not reported), the impact e¤ect is larger. This possibly re‡ects the dominant role of high-frequency movements in GDP growth (see Table 1). The responses of the business-cycle and medium-term components return to zero after 20 and 30 quarters, respectively. The response of the long-term component on the other hand remains negative over the full projection horizon of 10 years. This indicates that monetary policy shocks can have long-lasting e¤ects even on GDP growth. We …nd a similar pattern for the unemployment rate, with oscillating behavior of the higher-frequency components and a more drawn out response of the trend. In terms of the size of the policy-induced movements, the business-cycle and medium-term components are roughly similar, in contrast to the variance decompositions in Table 1. This suggests that there are other shocks that drive movements in the unemployment rate in these frequency bands. The response of the FFR and in‡ation components for both VAR speci…cations is very similar. At higher frequencies, the FFR rises and in‡ation falls, where especially the business-cycle components move together closely. The response of the respective trend components is di¤erent, however. In‡ation and the FFR do not react much on impact and in the near term, but they move together positively over the longer horizon. A contractionary policy shock thus has a long-lasting negative e¤ect on the long-term component of the FFR and in‡ation. These results con…rm the existence of a Fisher e¤ect in the long-term component, whereas in the short term the demand-constricting e¤ect of an interest rate hike dominates as in standard monetary policy models. Moreover, the results also show that contractionary policy lowers the long-term component persistently presumably through an expectations e¤ect: tightening policy gains credibility, anchors in‡ation expectations, and lowers in‡ation overall. 4 Assessing DSGE Models We now investigate whether several medium-scale DSGE models can replicate the stylized facts identi…ed above. Such models have been developed explicitly with an eye on replicating the performance for business-cycle movements and the long run. This raises the question, whether they can, in fact, capture behavior along all frequency bands identi…ed by our 18 wavelet decomposition. In a preview of the results, we …nd that the models generally do well for business-cycle frequencies and in the long term as these are frequency bands that the models are designed to replicate. However, the models generally fail at capturing medium-term frequencies. 4.1 DSGE Models and Simulation In the DSGE literature, it is well-known that various modeling devices are useful in matching persistence in the data, at least over the business cycle (see the programmatic papers by Christiano et al., 2005, and Christiano et al., 2010, and also the seminal DSGE models by Smets and Wouters, 2003, 2007). Examples are modi…cations to utility, such as habits in consumption; production, such as investment adjustment costs; and highly persistent shock processes. At the same time, stochastic trends have proved to be a ‡exible modeling component to capture drifting behavior over time. This section studies whether these modeling elements are useful across all frequencies. We select models based on their widespread use in monetary policy analysis and their consistency with the speci…c data that we have considered so far. Moreover, we want to give the chosen models a fair chance at capturing the patterns found in the wavelet decomposition. We therefore require that one of the underlying drivers of business cycles is a stochastic trend in productivity that can smoothly vary over time. This speci…cation is well-known for matching the movements in the GDP trend. We thus focus on three canonical models in the literature: Smets and Wouters (2007), del Negro et al. (2015), and Christiano et al. (2016).23 Smets and Wouters (2007) (SW) is a further development of the canonical Smets and Wouters (2003) New Keynesian DSGE model. It is the prototype of a medium-sized, optimization-based model designed to jointly capture the evolution of output and in‡ation and the monetary policy process. To this end, the model contains a variety of shocks and frictions that have come to be accepted as central to understanding aggregate ‡uctuations. The basic setup involves a representative household that makes consumption choices and supplies labor to a competitive labor market. On the production side, there are monopolistically competitive …rms that employ labor and capital to generate output, make investment decisions, and set prices. The third type of agent in this model is a policymaker who sets interest rates based on given feedback rules. The model features nominal price stickiness and sticky wages with backward in‡ation 23 We use computer codes for these models available at Volker Wieland’s Macroeconomic Model Data Base (MMB): https://www.macromodelbase.com/ and from the journal websites of the published articles. 19 indexation to capture slow-moving aspects of these variables. On the real side, there is habit formation in consumption and investment adjustment costs designed to create humpshaped responses of these aggregate demand components. The model is driven by seven structural shocks including a monetary policy disturbance. One key distinguishing feature of Smets and Wouters (2007) as opposed to Smets and Wouters (2003) is that the former does not have a time-varying in‡ation target. The model is estimated using Bayesian methods over the period 1966-2004 for seven key aggregate variables, but the set of observables does not include the unemployment rate. We can therefore not compare their model with our decomposition along this margin.24 We take their parameters estimates as given and simulate the model under this speci…cation. The second model that we consider, del Negro et al. (2015) (dNGS), is an extension of the SW model. It introduces a time-varying target in‡ation rate and incorporates …nancial frictions in the vein of Christiano et al. (2014). The model is estimated for a slightly larger dataset than the SW model and over the period 1964-2008. The key …nding of the paper is that the model is compatible with Great Recession outcomes in that it successfully predicts a sharp contraction in economic activity along with a drawn-out but modest decline in in‡ation. The third model is Christiano et al. (2016) (CET). While it is built around the same nominal structure as SW, CET introduces a much richer labor market setting governed by search and matching frictions and various wage determination mechanism. We report results both for a benchmark speci…cation with Nash bargaining and an alternative, namely alternative o¤er bargaining. What is important for our purposes is that the framework models the unemployment rate in contrast to the previous two DSGE models. Christiano et al. (2016) estimate the model over the sample period 1951-2008, with the same end date as del Negro et al. (2015). Our simulation procedure is as follows. We take the estimated models as given and …x the parameter values at the reported posterior medians. The models are then simulated by drawing from the innovation distributions over 10,000 periods. This is repeated 1,000 times, whereby we record the last observations to coincide with the length of our sample. From this sampling distribution we then compute the variance decomposition from the wavelet …lter as in section 2 and report 90-percent con…dence regions for the mean estimate of the variance decomposition. We group the individual wavelet decompositions into the categories ‘Short Term’(D1 -D2 ), ‘Business Cycle’(D3 -D4 ), ‘Medium Term’(D5 -D6 ), and ‘Long Term’(S6 ). 24 They also use the GDP de‡ator to measure in‡ation, whereas we report results for PCE in‡ation. The di¤erences between these two in‡ation measures are minor. 20 4.2 DSGE Models and Frequency-Band Decompositions We report the results of the simulation exercise in Table 2. Our focus is on the results for four key variables, namely real GDP growth, in‡ation, the federal funds rate, and the unemployment rate. We compare the simulation results to two di¤erent sets of underlying data: …rst, our original sample that covers 1954Q3 to 2017Q3; and second, the actual sample period over which the respective model was originally estimated. For all three models this excludes the Great Recession period and its aftermath. The latter results are reported separately in Table 3. The SW model is remarkably successful in replicating the overall volatility of real GDP components across all frequency groups, essentially matching the data exactly: around 60 percent is attributed to the short-term component, 30 percent to the business-cycle component and a much smaller percentage to the medium and long term. The same pattern is found for the dNGS model and with some minor di¤erences for the Nash-bargaining speci…cation of the CET model. A key driver for this …nding is the speci…cation of the exogenous productivity process as a stochastic trend, which is now standard modeling device in DSGE models. The wavelet decomposition thus con…rms the importance of this assumption. Turning to the nominal variables, in‡ation and interest rates, the performance of the SW model notably deteriorates. While the model is consistent with the short-term and medium-term components in in‡ation, the contribution of the business-cycle and long-term ‡uctuations is essentially ‡ipped. The SW model attributes only 15 percent to the long term and more than one-third to the business cycle. In contrast, the dNGS model comes much closer to the patterns in the data, although it underpredicts the contribution of the long-term component by almost 10 percentage points. The key di¤erence between the two models is that del Negro et al. (2015) incorporate a time-varying in‡ation target that is stationary but highly persistent. Over the sample period it e¤ectively pins down the trend movements in the in‡ation rate. As discussed before, trend in‡ation in the data might simply re‡ect the changing implicit or explicit in‡ation target, which in the DSGE modeling sense can be captured by such an exogenous process. Interestingly enough, the Nash-bargaining speci…cation of Christiano et al. (2016) has di¢ culty with this pattern, as it attributes considerable variability to the short-term and business-cycle components and not enough to the long-term component. Notably, their model does not feature a time-varying in‡ation target, which reinforces the point raised above. However, the alternative speci…cation of the CET model with wage determination based on an alternative o¤er bargaining mechanism is on point in capturing the behavior 21 of in‡ation across all frequency bands. It is well-known, e.g., Krause et al. (2008), that a Nash-bargaining nominal wage mechanism does not impart enough in‡ation persistence in a New Keynesian search and matching framework which these results con…rm. On the other hand, the alternative o¤er bargaining mechanism implies endogenous wage inertia that then translates into inertial prices (see the discussion in Christiano et al., 2016). The frequency-speci…c patterns of the FFR in the data do not di¤er much from that of the in‡ation rate, although the wavelet decomposition attributes 80 percent of its movements to medium-term and long-term components, as opposed to two thirds in the case of in‡ation. A similar pattern is discernible for the three DSGE models in that they cannot replicate the importance of the long-term component and the relative lack thereof in the businesscycle frequencies. Most strikingly, the trend in the interest rate is associated with almost 60 percent of movements in the data, only half of which the dNGS model can capture. As before, the alternative speci…cation of the CET model does remarkably well for the behavior of the FFR. We …nally consider the decompositions of the unemployment rate, which, of our three DSGE models, only Christiano et al. (2016) can address. In the data, half of the ‡uctuations in unemployment are captured by the medium-term component with the remainder roughly equally attributed to the business-cycle and long-term components. Under the Nash-bargaining speci…cation, the CET model attributes half of unemployment ‡uctuations to the long-term component, one third to the medium-term component and the remainder to the business cycle. The model gets the broad pattern of ‡uctuations at di¤erent frequencies right: what matters for explaining the unemployment rate are the medium- to long-term components, but not those that are arguably more directly shaped by monetary policy. At the same time, the alternative bargaining speci…cation of CET results in a considerably worse performance for the unemployment rate. It also has problems with the decompositions of GDP growth where it attributes too much volatility to the long-term and medium-term components. Yet, its performance for the two nominal variables, the in‡ation rate and the interest rate, is spot on, where the baseline speci…cation with Nash bargaining put too much weight on the business-cycle components. Comparing the two approaches to modeling wage determination, these …ndings indicate that alternative o¤er bargaining generates more persistence in the model than Nash bargaining does.25 The ‡ip 25 This is, of course, related to the Shimer (2005) puzzle who argues that the standard search and matching model cannot replicate the observed volatility and also persistence of the unemployment rate and vacancies, that is, open positions. Alternative o¤er bargaining therefore presents an attractive solution to the Shimer 22 side of this …nding is that the former imparts too much persistence, which hurts the model’s performance with respect to the medium- and long-term components of GDP growth and unemployment. Our …nal exercise considers the importance of the sampling period for the assessment of the models. We subject each of the four model speci…cations to the same test, namely whether they could replicate the behavior of the wavelet decompositions for the full length of our empirical sample from 1954Q3 to 2017Q3. However, in their published versions the estimation periods of the three model frameworks di¤er. Speci…cally, the estimation period for the SW model is 1966-2004, for the dNGS model it is 1964Q1-2008Q3, and for the CET framework the sample period is 1954Q1-2008Q4. The former two periods are similar, they both miss 10 years at the beginning of our sample and then the Great Recession and its aftermath; whereas the CET sample di¤ers from ours in that it ends at the onset of the Great Recession. Although the underlying idea of structural DSGE modeling is that structural parameters are generally invariant over these sample periods, it is also well-known that sample size and sample period can a¤ect structural parameter estimates (e.g., Canova and Ferroni, 2012). In Table 3 we therefore contrast select decompositions for the actual estimation sample with the simulated sample for the same number of observations. The decompositions for the SW and dNGS sample periods are very similar to each other for real GDP growth and in‡ation. The biggest di¤erence is the long-term in‡ation component in the dGNS sample, which includes an additional four years before the onset of the Great Recession. In our full sample, this component explains 41 percent of in‡ation movements, in the shorter sample only 34 percent. For both models, the biggest discrepancy can be found in the behavior of the FFR. The shorter sample attributes much more volatility to the business-cycle and medium-term components for the policy rate, that is, 60 percent compared to 40 percent in the full sample. The long-term component explains about one third of the volatility in the SW sample, but almost 60 percent in the full sample. This discrepancy is arguably due to di¤erences in policy across the two periods, either at the beginning of the full sample period between 1954 and 1964 or during the Great Recession. In any case, the long-term FFR component is more pronounced over the longer period. The results in Table 3 do show, however, that the SW model struggles to match these facts, while the dNGS model is closer to the data. This pattern is also evident from the CET model, where we …nd that the long-term puzzle, which does not have to rely on exogenous wage stickiness. 23 FFR component in the full sample explains more of the overall volatility. This suggests that the Great Recession period has had a noticeable e¤ect on the behavior of the FFR which may not be surprising since during this period the Federal Reserve held its policy rate essentially …xed at its e¤ective lower bound of zero. This, by itself, imparts persistence onto the FFR. Interestingly, such di¤erences are not visible in any of the other variables, GDP growth and in‡ation, with the exception of the unemployment rate. It therefore seems that the behavior of the policy rate is largely disentangled from that of other macroeconomic aggregates. Table 3 also shows that the behavior of the unemployment rate is di¤erent across the samples and that the CET model under alternative o¤er bargaining cannot capture the behavior in the di¤erent sample either. In summary, we conclude that the three canonical DSGE models are able to replicate the wavelet decomposition we found in the data. We identify a stochastic trend in productivity and a time-varying in‡ation target as the key modeling elements. The random walk component in the former and a highly persistent in‡ation target capture the long-term components in real GDP and in‡ation exceptionally well. Replicating the frequency-speci…c components of the unemployment rate proves to be more di¢ cult. While the decompositions from the simulated data go broadly in the same direction, the challenge is that the variance decomposition is more evenly distributed among frequency bands than for the other variables. Based on these …ndings, we advocate wavelet decompositions as a straightforward tool to assess the validity of a DSGE model as a data-generating process, especially with respect to the contribution of individual modeling elements.26 5 Conclusion This paper advances three main …ndings. First, we show that more than two thirds of in‡ation and unemployment ‡uctuations in the US occur at low frequencies, whereas at most a quarter are attributable to business-cycle frequencies. However, it is mainly these latter ‡uctuations that are the focus of monetary policymakers and researchers: policy objectives are normally phrased in terms of stabilizing ‡uctuations around trends or potential. This dichotomy is generally re‡ected in the DSGE models that are used to study monetary policy and its e¤ects. Frequency-speci…c decompositions, such as the one we performed 26 In a sense, we are simply con…rming the results of Sala (2015), who estimates the SW-model in the frequency domain using likelihood-based methods and working o¤ the counterpart of the time-series respresentation of a state-space model. He …nds that this DSGE model broadly performs well and matches the data at various frequencies, but fails to capture labor market data and the interactions between real and nominal variables. However, he uses stationary, thus pre-…ltered data, and can therefore not speak to the overall decomposition into the several frequency bands. 24 using wavelet methodology, thus produce information relevant for policymakers. Our second …nding shows that several standard DSGE models do a credible job of replicating behavior at business-cycle frequencies that we identi…ed in the data. However, the models need to be suitably modi…ed to account for long-term movements via stochastic trends or time-varying in‡ation targets. They generally fail to capture behavior at medium-term cycles of between 8 and 32 years. We demonstrate in a third set of results that monetary policy shocks exert in‡uence over all frequency bands and in a broadly similar manner with the exception of the relationship between short-term interest rates and in‡ation, where the Fisher e¤ect prevails in the long run. Our paper thus contributes to a growing area of research that suggests that the notion of a cycle relevant for stabilization policy should be extended to include at least the medium term. Speci…cally, the analysis in the paper indicates that temporary shocks can have longlasting e¤ects that traditional business-cycle modeling largely abstracts from. Future work could therefore study time-frequency decompositions in models with such a transmission mechanism as in, for instance, Comin and Gertler (2006). Similarly, the …ndings in this paper also support the idea that what matters for monetary policy is less the short-term response of policy rates to deviations of economic activity from some target, but rather the credible anchoring of expectations.27 Typical analyses of optimal monetary policy focus on weighted averages of the unconditional variances of policy targets. It is common to compare policies by considering, for example, a weighted average of the unconditional variances of in‡ation and unemployment. However, such computations mask the e¤ects of policies on the variance of ‡uctuations at di¤erent frequencies. Frequency-based optimal policy in the vein of Brock et al. (2013) would thus be an interesting extension based on the analysis in this paper. References [1] Aguiar-Conraria, Luis, Manuel M.F. Martins, and Maria Joana Soares (2012): “The Yield Curve and the Macro-Economy Across Time and Frequencies”. Journal of Economic Dynamics and Control, 36, pp. 1950-1970. 27 In‡ation expectations can be anchored by the execution of tough anti-in‡ationary policies. In that sense, short-term stabilization policies and commitment to a long-term target are essentially two sides of the same coin since the former helps ensure the latter. However, the joint determination of policy is rarely modeled in DSGE models where the in‡ation target is often assumed rather than chosen. 25 [2] Bandi, Federico M., Benoit Perron, Andrea Tamoni, and Claudio Tebaldi (2019): “The Scale of Predictability”. Journal of Econometrics, 208(1), pp. 120-140. [3] Baxter, Marianne, and Robert G. King (1999): “Measuring Business Cycles: Approximate Band-Pass Filters for Economic Time Series”. The Review of Economics and Statistics, 81(4), pp. 575-593. [4] Beaudry, Paul, Dana Galizia, and Franck Portier (2017): “Putting the Cycle Back into Business Cycle Analysis”. Manuscript. [5] Berger, Theo (2016): “Forecasting Based on Decomposed Financial Return Series: A Wavelet Analysis”. Journal of Forecasting, 35(5), pp. 419 433. [6] Beveridge, Stephen, and Charles R. Nelson (1981): “A New Approach to Decomposition of Economic Time Series into Permanent and Transitory Components with Particular Attention to Measurement of the ‘Business Cycle’ ”. Journal of Monetary Economics, 7(2), pp. 151-174. [7] Bianchi, Francesco (2013): “Regime Switches, Agents’Beliefs, and Post-World War II U.S. Macroeconomic Dynamics”. Review of Economic Studies, 80(2), pp. 463-490. [8] Blanchard, Olivier J. (1997): “The Medium Run”. Brookings Papers on Economic Activity, 2, pp. 89-158. [9] Brock, William A., Steven N. Durlauf, and Giacomo Rondina (2008): “FrequencySpeci…c E¤ects of Stabilization Policies”. American Economic Review, 98(2), pp. 241245. [10] Brock, William A., Steven N. Durlauf, and Giacomo Rondina (2013): “Design Limits and Dynamic Policy Analysis”. Journal of Economic Dynamics and Control, 37, pp. 2710-2728. [11] Brock, William A., James Nason, and Giacomo Rondina (2007): “Simple versus Optimal Rules as Guides to Policy”. Journal of Monetary Economics, 54(5), pp. 1372-1396. [12] Canova, Fabio (1998): “Detrending and Business Cycle Facts”. Journal of Monetary Economics, 41(3), pp. 475-512. [13] Canova, Fabio (2014): “Bridging DSGE Models and the Raw Data”. Journal of Monetary Economics, 67, pp. 1-15. 26 [14] Canova, Fabio, and Filippo Ferroni (2012): “The Dynamics of US In‡ation: Can Monetary Policy Explain the Changes?” Journal of Econometrics, 167, 47-60. [15] Cao, Dan, and Jean-Paul L’Huillier (2018): “Technological Revolutions and the Three Great Slumps: A Medium-run Analysis”. Journal of Monetary Economics, 96, pp. 93-108. [16] Christiano, Lawrence J., Mathias Trabandt, and Karl Walentin (2010): “DSGE Models for Monetary Policy Analysis”. In: Benjamin M. Friedman and Michael Woodford (eds.): Handbook of Monetary Economics, Volume 3a, Chapter 7, pp. 285-367. [17] Christiano, Lawrence J., Martin S. Eichenbaum, and Charles L. Evans (2005): “Nominal Rigidities and the Dynamic E¤ects of a Shock to Monetary Policy”. Journal of Political Economy, 113(1), pp. 1-45. [18] Christiano, Lawrence J., Martin S. Eichenbaum, and Mathias Trabandt (2016): “Unemployment and Business Cycles”. Econometrica, 84(4), pp. 1523-1569. [19] Christiano, Lawrence J., Roberto Motto, and Massimo Rostagno (2014): “Risk Shocks”. American Economic Review, 104(1), pp. 27-65. [20] Christiano, Lawrence J., and Terry J. Fitzgerald (2003): “The Band Pass Filter”. International Economic Review, 44(2), pp. 435-465. [21] Cogley, Timothy (2001): “Estimating and Testing Rational Expectations Models When the Trend Speci…cation is Uncertain”. Journal of Economic Dynamics and Control, 25, pp. 1485-1525. [22] Comin, Diego, and Mark Gertler (2006): “Medium-Term Business Cycles”. American Economic Review, 96(3), pp. 523-551. [23] Del Negro, Marco, Marc P. Giannoni, and Frank Schorfheide (2015): “In‡ation in the Great Recession and New Keynesian Models”. American Economic Journal: Macroeconomics, 7(1): pp. 168-96. [24] Faria, Gonçalo, and Fabio Verona (2018): “Forecasting Stock Market Returns by Summing the Frequency-Decomposed Parts”. Journal of Empirical Finance, 45, pp. 228242. [25] Gorodnichenko, Yuriy, and Serena Ng (2010): “Estimation of DSGE Models when the Data are Persistent”. Journal of Monetary Economics, 57, pp. 325-340. 27 [26] Hamilton, James D. (2017): “Why You Should Never Use the Hodrick-Prescott Filter”. Forthcoming, The Review of Economics and Statistics. [27] Hodrick, Robert, and Edward C. Prescott (1997): “Postwar U.S. Business Cycles: An Empirical Investigation”. Journal of Money, Credit, and Banking, 29(1), pp. 1-16. [28] Krause, Michael U., David Lopez-Salido, and Thomas A. Lubik (2008): “In‡ation Dynamics with Search Frictions: A Structural Econometric Analysis”. Journal of Monetary Economics, 55, pp. 892-916. [29] Lubik, Thomas A., and Christian Matthes (2016): “Indeterminacy and Learning: An Analysis of Monetary Policy in the Great In‡ation”. Journal of Monetary Economics, 82, pp. 85-106. [30] Onatski, Alexei, and Noah Williams (2003): “Modeling Model Uncertainty”. Journal of the European Economic Association, 1(5), pp. 1087-1122. [31] Pancrazi, Roberto (2015): “The Heterogeneous Great Moderation”. European Economic Review, 74, pp. 207-228. [32] Primiceri, Giorgio (2006): “Why In‡ation Rose and Fell: Policymakers’Beliefs and US Postwar Stabilization Policy”. The Quarterly Journal of Economics, 121, pp. 867-901. [33] Sala, Luca (2015): “DSGE Models in the Frequency Domain”. Journal of Applied Econometrics, 30, pp. 219-240. [34] Sargent, Thomas, Noah Williams, and Tao Zha (2006): “Shocks and Government Beliefs: The Rise and Fall of American In‡ation”. American Economic Review, 96(4), pp. 1193-1224. [35] Shimer, Robert (2005): “The Cyclical Behavior of Equilibrium Unemployment and Vacancies”. American Economic Review, 95(1), pp. 25–49. [36] Smets, Frank, and Raf Wouters (2003): “An Estimated Dynamic Stochastic General Equilibrium Model of the Euro Area”. Journal of the European Economic Association, 1(5), pp. 1123-1175. [37] Smets, Frank, and Raf Wouters (2007): “Shocks and Frictions in US Business Cycles: A Bayesian DSGE Approach”. American Economic Review, 97(3), pp. 586-606. 28 [38] Tkachenko, Denis, and Zhongjun Qu (2012): “Frequency Domain Analysis of Medium Scale DSGE Models with Application to Smets and Wouters (2007)”. Advances in Econometrics : DSGE Models in Macroeconomics – Estimation, Evaluation and New Developments, Volume 28, pp. 319-385. [39] Watson, Mark W. (1986): “Univariate Detrending Methods with Stochastic Trends”. Journal of Monetary Economics, 18, pp. 49-75. [40] Watson, Mark W. (1993): “Measures of Fit for Calibrated Models”. Journal of Political Economy, 101(6), pp. 1011-1041. 29 30 32.7 1.3 8.6 1.5 1.3 0.6 5.5 GDP Growth Unemployment In‡ation Federal Funds 3-Month Rate 10-Year Rate Term Spread 9.7 1.3 2.4 2.8 7.4 3.8 22.8 Short Term D1: 2-4Q D2: 4-8Q 17.1 2.5 4.8 5.6 7.7 9.6 19.0 26.7 4.0 8.9 10.2 9.9 18.5 12.6 Business Cycle D3: 8-16Q D4: 16-32Q 19.1 7.5 10.1 10.5 9.3 25.4 6.1 5.4 15.5 13.1 12.8 15.7 20.3 2.5 Medium Term D5: 32-64Q D6: 64-128Q Table 1: Variance Decomposition US 1954 - 2017 16.5 68.5 59.3 56.6 41.3 21.1 4.3 Long Term S6: >128Q Table 2: Variance Decomposition for Simulated Data Short Term Business Cycle Medium Term Long Term 56 59 32 29 8 10 4 2 (48-71) (21-38) (5-18) (1-5) RGDP Data SW dNGS CET (Nash) CET (AOB) 60 28 10 2 (50-70) (20-35) (5-17) (1-3) 65 27 6 2 (56-74) (21-34) (3-9) (1-5) 27 40 23 10 (20-35) (29-50) (14-37) (3-22) 16 20 18 35 25 29 41 15 (13-31) (22-48) (14-36) (3-36) In‡ation Data SW dNGS CET (Nash) CET (AOB) 17 20 29 34 (7-31) (8-32) (14-47) (9-66) 27 39 18 16 (18-37) (24-50) (9-29) (3-42) 13 22 20 44 (4-25) (6-41) (8-36) (10-80) 4 16 16 36 24 33 57 17 (9-23) (23-50) (16-51) (3-38) 12 24 33 31 (5-21) (10-39) (17-53) (8-61) FFR Data SW dNGS CET (Nash) CET (AOB) Unemployment Data CET (Nash) CET (AOB) 25 38 19 16 (16-34) (25-50) (11-31) (3-41) 10 23 25 43 (3-17) (7-42) (11-44) (9-79) 5 6 29 18 45 30 21 46 (2-11) (5-33) (12-52) (13-79) 1 10 30 59 (0-3) (2-21) (10-57) (23-86) 31 Table 3: Variance Decomposition for Simulated Data - Alternative Sample Short Term Business Cycle Medium Term Long Term Smets-Wouters FFR Full Sample SW Sample Simulated 4 8 16 16 25 36 24 35 33 57 32 17 dNGS FFR Full Sample dNGS Sample Simulated 4 7 12 16 25 24 24 32 33 57 36 31 CET (AOB) FFR Full Sample CET Sample Simulated 4 6 10 16 21 23 24 28 25 57 45 43 Unemployment Full Sample CET Sample Simulated 5 6 1 29 27 10 45 37 30 21 30 59 32 Figure 1: Wavelet Decompositions: Real GDP Growth Figure 2: Wavelet Decompositions: Unemployment 33 Figure 3: Wavelet Decompositions: In‡ation Figure 4: Wavelet Decompositions: Federal Funds Rate 34 Figure 5: Wavelet Decompositions: 10-Year Treasury Rate Figure 6: Wavelet Decompositions: Term Spread 35 Figure 7: Wavelet Power Spectra: Real GDP Growth Figure 8: Wavelet Power Spectra: Unemployment Figure 9: Wavelet Power Spectra: Federal Funds Rate 36 Figure 10: Wavelet Decompositions for Alternative Filters: Real GDP Growth Figure 11: Wavelet Decompositions for Alternative Filters: Unemployment 37 38 Figure 12: Decompositions with HP and CF Filters: Unemployment 39 Figure 13: Impulse Response Functions: 3-Variable Baseline VAR with Sign Restrictions. 40 Figure 14: Impulse Response Functions with D2 Components 41 Figure 15: Impulse Response Functions with D4 Components 42 Figure 16: Impulse Response Functions with D7 Components 43 Figure 17: Impulse Response Functions with All Components: GDP Growth 44 Figure 18: Impulse Response Functions with All Components: Unemployment