The full text on this page is automatically extracted from the file linked above and may contain errors and inconsistencies.
USING NOISY INDICATORS TO M EASURE PO T EN TIA L O U TPUT Kenneth N. Kuttner Working Paper Series Macro Economic Issues Research Department Federal Reserve Bank of Chicago June, 1991 (WP-91-14) Using Noisy Indicators to Measure Potential Output Kenneth N. Kuttner* Introduction Potential output, like the real interest rate and the natural rate of unemployment, is one of the key unobservable economic variables of macroeconomics and mon etary policy. While there is little agreement on which observable quantity best substitutes for potential output, most economists would endow it with a variety of distinct, but not mutually exclusive, attributes. Arthur Okun’s original definition emphasizes inflation. According to this interpretation, potential output is the level of production that does not add to inflationary pressures. A second characteriza tion is as the sustainable level of output — that is, a level consistent with stable future rates of output growth. A third corresponds to the full-employment level of production, where unemployment is equal to the natural rate.1 This paper synthesizes these definitions by modeling real GNP growth, inflation, and the unemployment rate as three distinct indicators of a single underlying unob servable variable, which we call potential output or potential GNP. The resulting model is an example of a dynamic multiple indicator specification, which opti mally combines data from these imperfect sources in an estimate of the underlying level of potential output. 2 This multiple-indicator approach to measuring potential output has two chief ad vantages relative to more traditional measures. The first is that by incorporating a number of alternative definitions of potential, it constructs a best-fit measure based on a variety of indicators. Such a composite promises to be more precise and reliable than measures based on only a single proxy variable. A second advantage is that this technique delivers natural measures of the standard error associated with the estimates of potential output. Although Okun himself acknowledged that it was “at best an uncertain estimate, and not a firm, precise measure,” the estimates in this paper are the first to include a set of confidence * Senior economist, Economic Research Department, Federal Reserve Bank o f Chicago, 230 South LaSalle Street, Chicago, 1L 60604. A previous version of this paper bore the title “A N ew Approach to Non-Inflationary Potential GNP.” This version has benefitted from discussions with Marty Eichenbaum, Steve Strongin, and Mark Watson, and from seminars at Northwestern University and the Federal Reserve Bank of Chicago. FRB CHICAGO Working Paper June 1991, WP-1991-14 1 bounds. Some gauge of the precision of the estimates are essential to policymakers acting on the basis of a measure of the economy’s potential. A b rief history o f potential output The origins of the potential output idea can be traced to Arthur Okun’s work from the 1960s. He viewed it as essentially an aggregate supply measure, or an index of productive capacity, describing it as “the maximum production without infla tionary pressure, ...or more precisely... the point of balance between more output and greater stability.”3 Deviations of output from potential were attributed to an inappropriate level of aggregate demand. While his definition of potential output rests on its relationship to inflation, as an empirical matter, Okun used the unem ployment rate as the sole indicator of the gap between output X and potential X*, according to the equation that became known as “Okun’s Law”: X* = X[1 + 0.032 (t/-t/* )], where the natural rate of unemployment, U*, was assumed to be a constant 4 per cent Until the 1970s, a simple log-linear trend served as an excellent approximation to the potential output series derived from Okun’s Law with a time-invariant natural rate. (In fact, Okun himself recommended the trend method, preferring to ignore the “wiggles and jiggles” in the measure introduced by random fluctuations in the unemployment rate.) After the supply shocks of the 1970s, however, economists interested in the relationship between aggregate demand and inflation recognized the importance of changes in the growth rate of potential output, and replaced the linear trend with some time-varying measure of potential, as in Gordon (1975, 1977). Most traditional measures of potential output have employed one of three alter native techniques in their construction: segmented trends, adjustments to Okun’s Law, or an aggregate production function. The segmented trend group includes some of the more familiar measures: the high-employment and mid-expansion se ries of the Bureau of Economic Analysis, for example, which use business cycle dates as the basis for the segments’ endpoints. Other examples are the Federal Re serve Board’s new potential GNP series,4 and the “trend output” series in Gordon (1990a, 1990b). The best example of the modem approach to Okun’s Law is Clark (1983), who suggests using the smoothed difference between the observed unemployment and a time-varying natural rate as the basis for potential output Those who have pro posed aggregate production function measures include Perry (1977), Clark (1979), FRB CHICAGO Working Paper June 1991, WP -1991-14 2 Perloff and Wachter (1979), and (indirectly) Boschen and Mills (1990). Among these studies, the Perloff and Wachter paper deserves note; while based on an ag gregate production function, their measure is unique for its use of inflation data as an indicator of an unobserved time-varying production function parameter. The technique described in this paper bears a resemblance to that of Perloff and Wachter, in that it uses the inflation rate to extract information on an unobserv able. However, here potential output is itself the unobservable of interest, and the latent-variable econometric framework explicitly addresses the statistical noise as sociated with its indicators. As decomposition of aggregate output fluctuations into persistent and transitory components, it is also related to the unobserved com ponent model of Watson (1986), and the structural VARs of Blanchard and Quah (1989) and Galf (1989). Potential output as an unobserved stochastic trend Constructing a meaningful potential output series first requires specifying the form of its time variation. An attractive alternative to the traditional segmented-trend method is to model potential as a stochastic trend, along the lines suggested by Stock and Watson (1988). Specifically, (log) potential output, x*, is assumed to follow an integrated stochastic process — a random walk with drift: Ax* = p* + £,. (1) This specification allows its growth rate to depart from its long-run growth rate p* over the short run, generating persistent deviations from a deterministic linear trend. Measured real GNP can now be modeled as an error-correction equation in which (log) real output, x, reverts to potential over time: 8x(L)Axt = p* + 0(x* - xt) + ^(L)^*. (2) The S*(L) and XX(L) lag polynomials and the coefficient 0 are to be estimated; the constant term p* is set equal to 8*(l)p* to ensure that potential and measured output grow at the same rate on average. The innovation is uncorrelated with e( by assumption. In the context of the unit roots literature, (1) and (2) imply that potential and mea sured real output are 1(1) processes, while their difference — the “output gap” — is stationary. Despite the fact that x* is unobservable, the interpretation of errorcorrection mechanism is straightforward: when output exceeds potential, real GNP FRB CHICAGO Working Paper June 1991, WP-1991-14 3 growth is expected to fall below its mean. Conversely, output below potential im plies higher-than-average future growth rates. The disturbance term rtf induces purely transitory deviations in output from potential.5 Viewed as a latent variable model, (2) says that Ax, (real GNP growth) may be used as an indicator of the unobserved x*. Thus, with the appropriate restrictions on the dynamics of the error-correction equation, the two equations may be estimated, and an estimate of potential output may be extracted. This is particularly easy to verify in the one-indicator case, as the two equations may be rearranged to express Ax, in canonical form6 as an ARMA process with a composite error term: [5X(L)A + 0]Ax, = 6e, + AX*(Z,)itf. Therefore, while the model’s spectral density is that of an ARMA, estimating (1) and (2) parameterizes the spectrum of Ax in terms of two distinct shocks; inno vations to the extracted potential output “signal” correspond to the low-frequency potion of the spectrum of real GNP growth. Further discussion of the models’ spectral properties appears in a subsequent section of the paper. The real payoff to the latent variable methodology comes from the use of addi tional variables as indicators of the level of potential output. This improves the precision of the estimates, and it adds to the economic interpretation of the model, associating the unobserved x* with alternative descriptions of potential output We begin by explicitly linking potential output to the rate of inflation by way of a simple dynamic aggregate supply relation: 8*(L)Ajt, = a (L)(x, - x*) + X”(L)q? + 7 v„ (3) where An, is the change in the inflation rate, the vector v, includes exogenous de terminants of inflation not related to the current state of aggregate demand, and q* is a disturbance term. Specification in terms of the change in inflation is ap propriate given the apparent non-stadonarity of the inflation rate over the 1960-90 sample period. It also means that an output gap equal to zero is consistent with any constant rate of inflation, rather than zero inflation; in other words, potential GNP corresponds to the “non-accelerating-inflation” (or, more precisely, the nonincrearing-inflation) level of output Excluding Ax, from direct inclusion in the equation for Art, imposes a factor structure on the co-movements of output and inflation growth; all covariation between the two is assumed to come from the common factor, (x* -x ,). With the addition of inflation as a second indicator of the unobservable level of potential output, equations 2-4 form what is referred to as a dynamic multiple in dicator model.7 The specifications discussed in the following section of the paper FRB CHICAGO Working Paper June 1991, WP-1991-14 4 are one-indicator output-only and two-indicator output-inflation versions. A sub sequent section of the paper introduces the unemployment rate as a third indicator of potential. Estimating the model For the purposes of estimation, the indicator equations (2) and (3) need to be re stricted; 8X(L) is taken to be a second-order lag polynomial, while the \ X(L) and XK(L) moving-average polynomials are both assumed to be first-order. The §”(L) polynomial is set to unity, so that inflation changes are entirely a function of the output gap, plus a moving-average error term. These restrictions are consistent with (unreported) Wald tests of more profligate specifications. Next, setting a (L) equal to aiL + a 2D 2 specifies inflation rate changes as a func tion of two lags of the output gap. Finally, one exogenous variable is chosen for inclusion in v(: the lagged difference between the growth rates of M2 and nominal income, or the opposite of the differenced velocity of M2. The idea, which is con sistent with an error-correction mechanism between the growth rates of M2 and nominal income, is that this difference represents the inflationary pressure from any monetary expansion not yet incorporated into nominal income. One convenient way to estimate multiple indicator models like the one presented here is through the application of the Kalman Filter to its state-space representa tion, z, = Az,_i + B e, + Gw, y, = C z, + D u, + Hwt, where z, y and w are a vectors of state, observable and exogenous variables, re spectively. The model in equations 1-3 can be expressed by including current and lagged values of x*, qx and q* in the state vector. The observable y, obviously includes An, and Ax,, while w, includes a vector of ones, lagged observables, and the monetary term discussed above. As described in Harvey (1981, 1989), the Kalman filter computes the minimum mean square error forecast of the period t state vector and its error variance, con ditional on period t - 1 information, according to the prediction equations, z<|»-1 = Azm m + Gwm fltt'-i = + B B '. FRB CHICAGO Working Paper June 1991, WP -1991-14 5 Conditional on period t data, the filter extracts the optimal estimate of the state vec tor, and computes the period t signal-extraction error, by way of the updating equations, z<|t = ztp— 1 + Kt £y<—( C z ^ i + Qn, = -K tC O m -u where K t^ a ^ iC 'ic a ^ iC '+ D D '}. If the elements of the matrices A, B, C, D , G, and H were known, extracting esti mates of the unobserved z( and its variance would be trivial. As they are not known, the series of uncorrelated prediction errors generated by the Kalman filter can be used to evaluate a likelihood function under the assumption of normally distributed disturbances. Standard optimization routines8 can compute maximum-likelihood estimates of the parameters. The one additional piece of information required for estimation is the initial value of the state vector, zo, and the variance associated with that value. The Bayesian interpretation is that zo embodies prior information on the value of the state vector at the beginning of the sample. The estimates are not generally sensitive to the choice of z0: as of 1959:4, log potential GNP is assumed to equal the Federal Reserve Board’s estimate of7.428, with a standard error of 5%.9 The initial values of the q disturbances are set to zero, with standard errors of 3%. The lack of consistent pre-1959 money stock data requires beginning the estima tion in 1960; the sample ends in 1991:1, for a total of 125 observations. While the obvious measure of real output is standard NIPA real GNP series, the choice between alternative price series is less clear. Because it is less volatile and less subject to fluctuations in oil prices, the fixed-weight GNP deflator is used as the basis for the inflation rate.10 Relative to the implicit deflator, the fixed-weight version is preferable as a better measure of pure price changes.11 Param eter estim ates Maximum-likelihood estimates of the one- and two- indicates' models appear in Ta ble 1. The estimated 6 and oe (respectively, the coefficient on the error-correction term and the standard deviation of the potential output innovation) generally suppot the idea of time-varying potential GNP. Although these two parameters are not individually significant in the single-indicator model (specification I), they are much more precisely estimated when the inflation rate is included as a second in dicator in specification II; their asymptotic standard errors fall by one-half. As oe FRB CHICAGO Working Paper June 1991, WP-1991-14 6 is expressed in percentage terms, its specification n estimate of 0.73 corresponds to a quarterly potential GNP shock with a variance of roughly half a percent Both models yield estimates of p*. the mean growth rate of potential output close to 0.74% on a quarterly basis (2.96% annualized). The estimates of the standard deviation of the transitory qx shocks are both roughly 0.9S, again in quarterly per centage terms. In contrast the estimated lag polynomials 8* and Xx change some what from specification I to n. However, these changes are small relative to the coefficients standard errors, and have little impact on the models’ spectral density. Because specification I (except for the effect of the initial conditions) is essentially a reparameterized ARMA(3,2), its maximized likelihood is virtually equal to that of the estimated ARMA. The estimate of aggregate supply relation in specification II produces no surprises. The a coefficients indicate a strong effect of aggregate demand on inflation. The estimates of cti and ot2 suggest a specification in toms of the first difference of the output gap. However, including the level of the output gap allows aggregate demand to contribute to some of the non-stationarity in the inflation rate, which the first-difference specification would not allow. Finally, y, the coefficient on lagged adjusted M2 growth, appears as a significant additional determinant of inflation. Extensions to the basic m odel The models discussed above used output growth and the inflation rate as indicators of the latent potential GNP. This section presents two extensions to the basic model, both of which include the unemployment rate as a third indicator variable. Including unemployment data The foundation for the unemployment indicator is a modified version of Okun’s Law: 8«(L)i/r=t/ ; + -*)+%, (4) where <t>(L) and 8U(L) are lag polynomials specific to the unemployment equation, and is the disturbance term. The XJ*t term represents the natural rate of unem ployment. There are two problems with using this equation as an indicator of the output gap. The first is the problem of time variation in the natural rate. While it was thought to be roughly constant at 4% in the 1960s, its apparent upward trend in the 1970s led to considerable research on time-variation in its structural component.12 A related problem is the non-stationarity of the unemployment rate. Perhaps due to highly persistent changes in its structural component, the usual statistical tests are FRB CHICAGO Working Paper June 1991, WP-1991-14 7 Table 1 Basic Estimation Results Specification I Estimate Std Error Specification n Estimate Std Error n* Oe 0.730 0.553 0.090 0.618 0.740 0.728 0.075 0.278 6 0.064 0.276 0.215 -0.001 0.940 0.070 0.442 0.158 0.456 0.069 0.106 0.435 0.200 -0.165 0.948 0.054 0.314 0.148 0.318 0.074 0.117 -0.092 -0.680 0.058 0.256 0.030 0.028 0.224 0.028 0.038 8! 85 M ox ai 0-2 *? Y 0* LLF SEax SEA* ., . . . • 4.2103 0.900 *•• 9.562 0.875 0.296 Specification I: single indicator Ax* = p* + e, A x , = (1 - 8* - SDn* + 6(x* - x ,) + 8* + SjAxrf + rtf + Xi Um Specification II: adds inflation indicator An, = aiCtM -ac^1) + a2(X rf-x^ 2) + YACln(M2)-ln(GNP))M +n? + ^fn*i FRB CHICAGO Working Paper June 1991, WP-1991-14 unable to rejectthe null hypothesis of a unitroot in the unemployment rate over the postwar sample. One possible solution isto model the naturalrate as an additional non-stationary latentvariable, and specify the measured unemployment rateas a function of the output gap and the time-varying natural rate. However, because the focus of this paper ispotentialGNP ratherthan thenaturalrateofunemployment, an appealing alternative approach issimply toremove the unitrootby firstdifferencing. This would be a perfect solution to the time-varying natural rateproblem only if U* were a pure random walk, U*t= lft_x+v,, whose errorswere uncoirelatedwith the stochastic components of the output gap. In this case, firstdifferencing (4) yields: 8“(L)At/r= <KL)A(jc;- xt)+ V,+ A$f; or,combining thev, and 8 intoa single MA(1) composite error term X “ ( L ) t i “ , “(L)Af/, = « - x,)+ X“(L)nr. (5) whereT}*isuncorrelatedwithqxand eatalllags. Inpractice,theseassumptionsare ratherdubious. While serialcorrelation inv(could be taken care ofby expanding the XU(L)polynomial, the literatureon sectoral shifts suggests sources ofnonzero correlationbetween thestructuralunemployment innovations,v, and dieotherdis turbances,Ty*and e,violatingthedynamic factorassumptions. At best,one would hope that first-differencing would remove most low-frequency structural unem ployment fluctuations. Adding the unemployment rate as a third indicator of the outputgap, therefore, may introducea tradeoffbetween bias and precision. While the additional indicator reduces the signal extraction error associated with esti mated potential GNP, italso introduces the possibility thatestimated fluctuations in the output gap may be contaminated by changes in structuralunemployment. Including population growth A second extension involves including population as a determinant of potential output. The simplestway todo thisistosubtractthe logarithm ofpopulation from thelog ofpotentialoutput,x*-nt= x*,and expressper capitapotentialoutput,X* as a stochastic trend: Ax( *= p* + e„ (6) asbeforeinequation 1. The threeindicatorequations 2,3, and 5,remain thesame, exceptthesum ofthelatentvariableX* and the (exogenous) log ofpopulation now appear in place ofx*. FRB CHICAGO Working Paper June 1991, WP -1991-14 9 Table 2 Extended Estimation Results SpecificationIII Estimate StdError Specification TV Estimate Std Error n* Oe 0.733 0.505 0.052 0.134 0.349 0.545 0.060 0.164 e 0.084 0.411 0.209 -0.132 0.948 0.048 0.356 0.160 0.364 0.078 0.076 0.494 0.038 0.273 0.162 0.285 0.076 0 .0 2 1 1 0% 0.096 -0.075 -0.549 0.051 0.273 <>0 2 1 .1 1 ♦l 8 ? X? 6.62 0.338 -0.283 0.170 % «5 Ox (Xl a2 Ou LLF SEa, SEa* SEaii 0 .2 2 2 -0.188 0.954 0 .0 2 1 0.089 -0.072 -0.505 0.044 0.279 0 .0 2 1 2.17 3.66 0.106 0.198 0.026 21.27 7.71 0.300 -0.351 0.157 2.13 3.49 0.103 0.293 0.039 0.019 0.095 0.026 10.648 0.846 0.291 0.207 0 .0 2 2 0 .0 2 1 0.082 0.025 10.637 0.894 0.296 0.204 Specification III:adds unemployment indicator A U, - 4>oA(j:*- x ,) + <>!A(jc( *_!- xM )+ 6?AC/M + rtf+ X^JLi Specification IV: potentialoutput inper capitaterms x* = n,+5* where Ax* = p* + e, FRB CHICAGO Working Paper June 1991,WP-1991-14 The results for these two models, designated specifications m and IV, appear in table2. Adding the unemployment rateindicatorhas a relatively small impact on most of the parameter estimates. The most noticeable change isin the size of the shocks topotentialoutput; specificationm delivers an estimatedaeofonly 0.S0S, compared with 0.728 in specification II. The specificationIV resultsaresimilartothosefrom III.The main differenceisthat p* now represents thegrowth rate ofper captiapotentialoutput, whose estimated annual growth rate isroughly 1.4%. Surprisingly, as measured by the standard er rorof theone-step-aheadforecasts, theoutput and inflationequations’fitdeclines slightlyrelativetospecification in. Estim ates o f potential GNP Perhaps themost usefulresultsofthesemodels aretheestimatedpathsofpotential GNP and theirstandard errors. Extracting an estimate of the latentx* isstraight forward; conditionalon themaximum-likelihoodparameterestimates,theKalman filtercan compute one-sided estimates ofpotentialoutputand itssignalextraction error,denoted% and Applying the Kalman smoothing algorithm yieldsanal ogous two-sided estimates,x^t and Q 4 7 . However, theseestimatesofthesignalextractionerrorneglecttheuncertaintyfrom the parameter estimates’variance. To compute thisvariance, we adopt theproce dure proposed by Hamilton (1986). This involves decomposing the totalvariance of state vector element of interest, say xjjr,into the signal extraction or “filter” uncertainty, £ {tf-*3r.fc)2IZr.po}, and theparameter uncertainty, { (xqr.fL ~ x3r,Po)21Zr} , where Po istheestimatedparametervector,andZj denotes thedatathrough period T. The Q ( | 7 from the Kalman smoother approximates the filteruncertainty. The parameter uncertainty isestimated via Monte-Carlo: an artificialsample of Ps is drawn from a multivariate normal population, generating a sample of xjjr series, which is then used to compute the sample variance of each observation in the series. The same procedure applies to the one-sided estimates, Figure 1 shows thetwo-sided estimateof thepotentialGNP seriesfrom specifica tion H (the two-indicator model with inflation), along with the 5th and 95th per centiles of itsempirical distribution (analogous to the 90% confidence interval). FRB CHICAGO Working Paper June 1991, WP-1991-14 11 Figure 1 Figure 2 FRB CHICAGO Working Paper June 1991, W P1991-14 Figure 3 FRB CHICAGO Working Paper June 1991, WP-1991-14 Table 3 Sources of uncertainty in potential GNP Specification I II III IV Parameter variance 7.68 9.54 2.07 1.74 2.08 1.02 0.64 1.17 Overall standard error 4.07 1.75 1.54 1.80 One-sided Filter variance T\vo-sided Filter variance 4.44 1.03 0.95 1.12 Parameter variance 6.80 0.65 0.40 0.87 Overall standard error 3.30 1.29 1.16 1.41 The estimated parameter variances are based on 200 draws. The reported averages exclude the first 8 quarters o f the sample. The solid line plots log real GNP. The local trend in potential output varies con siderably over the sample; from a relatively rapid growth rate in the early 1960s, it begins to slow late in the decade, a tendency which continues throughout the 1970s. In the early 1980s, however, its growth rate picks up once again. As measured by the two-indicator specification n , the output gap is statistically significant during only four episodes: the 19 6 5 -9 ,1 9 7 3 -4 and 1978-9 expansions and the 1981-2 recession. The 1974-5 recession and the late-80s expansion are marginal episodes. Here, the deviations barely exceed the two-sided 90% confi dence bounds, but not the one-sided bounds shown in figure 2. Figure 3 shows the one-sided estimate from specification III. Including the unemployment rate as a third indicator makes it possible to distinguish the 7 4-75 recession, although the late 80s remain within the 90% bounds. Except for the fact that the parameters are estimated over the entire sample, the one-sided estimates correspond to the estimates that would have been available to policymakers in “real tim e.” The fact that som e o f the deviations o f output from potential are discernible in the two-sided but not in the one-sided estim ates illustrates the point that discerning the appropriate course o f monetary policy is frequently easier with the benefit o f hindsight. FRB CHICAGO Working Paper June 1991, WP-1991-14 14 Table 3 reports the average uncertainty associated with the estimates o f potential GNP, including the separate contributions o f filter and parameter uncertainty. The table demonstrates the significant reduction in the standard error obtained in going from a one- to a two-sided estimate; in the case o f specification U, it shrinks by 30%. The table also shows the beneficial effect o f adding additional indicators o f the output gap. The reduction in uncertainty is particularly spectacular in going from specification I to II, in which the overall standard errors fall by over 60%. Including the unemployment rate as an additional indicator yields an additional improvement in the standard errors. M oving to the per capita specification, how ever, increases the standard errors, perhaps due to the deterioration in the m odel’s fit in the output equation. Except in the case o f the (somewhat overparameterized) specification I, signal extraction contributes more uncertainty to the estim ates than the sampling variance o f the parameter estimates. Spectral properties o f the estim ated m odels Another interesting way to view the results is from the frequency domain, exam ining the spectral density function o f Ax im plicit in the specification. The contri butions o f the two shocks to the spectrum o f the change in log real GNP can be computed from their transfer functions: / at(co) = oj\ 6 8*(e’ to) ( l - t r to) + e + o? (l-e-^GT*0) because the e and t |x shocks are uncorrelated, the spectral density o f Ax is just the sum o f the two shocks’ individual contributions. The spectral density from the two-indicator specification II appears in figure 4. Except for sm all differences in the size o f the frequency-zero variance, the other specifications’ spectra are virtually identical. The shapes o f the shocks’ individual contributions reflect the error-correction structure o f equation 2. The transitory q* shocks contribute nothing to the spectrum at frequency zero, while the persistent e shocks contribute spectral mass only near the zero frequency; by co = 0.S (12.5 quarters), its contribution is virtually z o o . The decomposition o f the k-step-ahead forecast error variance, shown in table 4 , is an alternative way to view the information in the spectral density function. W hile almost none o f the one-quarter-ahead variance comes from the potential output shocks, that proportion rises sharply at horizons longer than 12 quarters. A t 16 quarters, the e shock accounts for more than three-quarters o f the variance. As mentioned earlier, the univariate representation o f Ax, im plied by these poten tial output specifications is equivalent to an ARM A(3,2). An interesting restriction on these specifications involves imposing 0 = 0 (which also renders o e irrelevant). FRB CHICAGO Working Paper June 1991, WP-1991-14 15 Figure 4 Spectrum of output growth, specification II Table 4 Decomposition of forecast variance ^-step-ahead variance, % k FRB CHICAGO Working Paper June 1991. WP-1991-14 T |x shocks e shocks 1 94 6 4 8 12 88 82 12 18 71 29 16 23 77 16 Figure 5 Spectrum of output growth, ARMA(2,1) Doing so produces redundant unit roots in both the autoregressive and m ovingaverage polynomials; the (1 - L) factors cancel, and the model collapses to an ARM A(2,1). Imposing this restriction and estimating the restricted ARMA yields the follow ing model: (1 + 0.33L - 0.32L2)Axt = (1 + 0.59L)e,. The comparison between the restricted and unrestricted models reveals a lot about their low-frequency behavior. The restricted ARM A(2,1) spectrum, which appears in figure 5, differs significantly from the unrestricted m odel’s — particularly near frequency zero, where the ARM A(2,1) spectral density function attains a peak o f roughly 2.0. By contrast, the unrestricted model with its ARM A(3,2) represen tation allows for a much richer characterization o f the low frequencies o f Ax, an observation consistent with the Christiano and Eichenbaum (1990) analysis o f the low-frequency properties o f alternative ARMA specifications. Examining the roots o f the m odels’ characteristic polynomials is another way to understand this behavior. The effect o f moving from zero to positive values o f 6 FRB CHICAGO Working Paper June 1991, WP-1991-14 17 and o e is m ove the redundant unit MA root away from the unit circle. Likew ise, a positive 6 moves the AR root away from unity. The result is to create two nearly redundant roots. In specification II, for example, the largest MA root is -0 .9 0 , and the modulus o f the largest AR root is 0.81. Some comparisons An U nobserved C om ponents m odel In his discussion o f univariate detrending methods, Watson (1986) proposes de com posing difference-stationary time series as the sum o f an 1(1) component and a stationary autoregressive com ponent For real GNP, Watson uses the follow ing Unobserved Components (UC) specification: x, = x, + ct where X, = p + t,_i + ej and ct = a ic t- 1 + a 2c t- 2 + ect , a random walk stochastic trend xf, and a stationary AR(2) cyclical com ponent ct. The estimated parameters o f this model for the 1960:1-1991:1 sample period are: A = 0.721, <h = 1.604, &2 = -0 .6 4 7 , d t = 0.702, and d c = 0.476. The standard error is 0.89% . As a decom position in which the stochastic trend is modeled as a latent variable, the UC model has much the same flavor o f the latent-variable specification pro posed in this paper. Indeed, the spectrum o f Ax it im plies, shown in figure 6, is alm ost identical to the that o f the potential GNP model. The difference between them lies in the nature o f the two shocks’ contributions to the spectrum: in the case o f the UC model, the shock to the trend x contributes to the spectrum across all frequencies. In contrast, by virtue o f the error-correction specification o f the potential GNP model, the shock to the potential output trend contributes only at very low frequencies. The result is a somewhat smoother estimated trend, free o f the higher-frequency movements that characterize the UC trend. The H odrick-Prescott filter Another interesting comparison is between the estimated potential output sa le s and the trend component from the Hodrick-Prescott (HP) filter, which is essen tially a sophisticated two-sided m oving-average.13 Figure 7 plots the two-sided specification n estimate o f potential along with the HP trend, with X set to 1600. For reference, figure 7 also includes a deterministic linear trend. In general ap pearance, the two series are comparable; the smoothness imposed by the stochas tic trend specification far x* is similar to the smoothness delivered by X. = 1600. FRB CHICAGO Working Paper June 1991, WP-1991-14 18 Figure 6 Spectrum of output growth, unobserved components model One obvious distinction between the two procedures is that without an underlying statistical model, the HP filter cannot deliver the standard errors computed by the Kalman filter. The series them selves also reflect a number o f important discrepancies. Two o f these occur in 1973 and 1979, when potential output growth takes a dip not re flected in the HP trend. The estimate o f potential apparently associates the in creasing rate o f inflation over those periods with a depressed level o f potential output. Or, to attach a more structural interpretation, the potential GNP estim ate successfully picks out the 1973 and 1979 OPEC-induced supply shocks. Another important discrepancy occurs in 1982-4, where the HP trend is depressed relative to estimated potential. Because the HP technique simply involves smooth ing the observed output series, som e portion o f the 1981-2 recession appears in the filtered trend component. By contrast, the potential GNP procedure attributes FRB CHICAGO Working Paper June 1991, WP-1991-14 19 Figure 7 Figure 8 FRB CHICAGO Working Paper June 1991, WP-1991-14 virtually all o f the 1981-2 decline in output to a widening output gap — in other words, a pure demand shock. M id-expansion trend GNP A final instructive comparison is between this estimate o f potential GNP and the more traditional mid-expansion measure constructed by the Bureau o f Economic Analysis, which, because it uses ex post business cycle dating information, can also be thought o f as a two-sided estimate. Figure 8 compares the “output gaps” implied by the two measures, where the gap is defined as actual GNP less potential. Again, the two series differ significantly during a number o f important episodes. Like the HP trend, the mid-expansion measure m isses the 1973 and 1979 dips in potential, implying smaller output gaps than those based on the specification II measure. The discrepancy goes the other way in the late 60s. During this period, the mid-expansion gap is about twice the size o f the one based on the specification II estimate. Only during the 1981-2 recession do the two measures coincide. Conclusions The goal o f this paper has been to apply new econometric techniques to the estima tion o f a key unobservable macroeconomic quantity, potential output. The paper proposed using multiple indicators— output growth, inflation, and unemployment — to extract information on the level o f potential real GNP, which was modeled as a latent stochastic trend. W hile consistent with Okun’s original conception o f potential output, these new estimates embody two important advantages relative to traditional measures. First, they distill disparate sources o f information into a single estimate o f potential out put, promising to improve its reliability and precision. Second, they explicitly recognize the uncertainty involved in tire estimation and provide a way o f gauging that uncertainty. Another way to interpret these results is as a decomposition o f real GNP fluctu ations into potential output shocks, and deviations from output around potential (traditionally, “ supply” and “demand” shocks). The results presented here are consistent with the idea that shocks to potential are a very small source o f output variance at horizons shorter than three years. However, over four-year and longer horizons, shocks to potential output are responsible for m ost o f the variance in real GNP. FRB CHICAGO Working Paper June 1991, WP-1991-14 21 One possible extension to the method outlined above is to expand the supply side o f the potential output specification, which was modeled as a stochatic trend tv as a sim ple function o f population. Augmenting that specification to include measures o f factor inputs is a promising area for future work. FRB CHICAGO Working Paper June 1991, WP-1991-14 22 Footnotes 1. Some econom ists would dissent from this characterization. On one hand, clas sical econom ists criticize the potential output concept for its lack o f a general equilibrium foundation. Plosser and Schwert (1979) make this case. On the other side, DeLong and Summers (1988) argue that potential output should correspond to the highest feasible level o f production. 2. Since the 1970s, latent-variable models have found extensive application. A partial list includes Aigner et al. (1986), Engle and Watson (1985), Hamilton (1985), andNorrbinandSchlagenhauf (1988). A comprehensive guide to their application to econometrics is Harvey (1989). 3. Okun (1970), pp. 132-3. 4. Although this series remains unpublished, it is widely used in Federal Reserve Board research, such as Hallman Porter and Small (1989). 5. As the e shocks produce persistent changes in the sustainable level o f out put, while the q* shocks produce only transitory fluctuations around that level, one might interpret them as “supply” and “demand” shocks respectively, as in Blanchard and Quah (1989). While appealing, this interpretation is not essen tial to the model. 6. See Nerlove, Grether and Carvalho (1979). 7. A useful taxonomy o f latent variable econometric models appears in Watson and Engle (1983). 8. These results use the BFGS and BHHH algorithms from the Gauss MAXLIK procedure. 9. In the single-indicator model, real GNP changes alone provide very little infor mation on the level o f potential output Therefore, when a very loose prior on is specified, the procedure manages to fit a trend-stationary model by rais ing the initial value o f log potential GNP upwards, to over 7.5. ly in g down a more plausible initial condition with only a 2% standard error on xj forces the procedure to fit a difference stationary model with nonzero e variance.10 10. In practice, results based on the CPI are very similar to those obtained with the fixed-weight deflator when oil price changes are appropriately controlled for. FRB CHICAGO Working Paper June 1991, WP-1991-14 23 11. Statistical releases from the Department o f Commerce caution against com puting an inflation rate from the im plicit deflator “[Because] the prices are weighted by the com position o f GNP in each period,.. .the Im plicit Price D e flator reflects not only changes in prices but also changes in the com position o f GNP, and its use as a measure o f price change should be avoided.” 12. See, for example, Lilien (1982) and Rissman (1986). 13. See King and Rebelo (1989). FRB CHICAGO Working Paper June 1991, WP-1991-14 24 References Aigner, D.J., F. Schneider, and D. Ghosh (1986), “Me and My Shadow: Estimat ing the Size o f the U .S. Hidden Economy from Time Series Data” in Bar nett, Bemdt, and White, (ed.), Dynamic Econometric Modeling. Cambridge: Cambridge University Press. Blanchard, O J . and D. Quah (1989), ‘T he Dynamic Effects o f Aggregate Demand and Supply Disturbances,” American Economic Review 79, September 6 5 5 73. Boschen, J. and L. M ills (1990), “Monetary Policy with a New View o f Potential GNP,” Federal Reserve Bank o f Philadelphia Business Review, July-August, 3-10. Bureau o f Economic Analysis, Commerce Department (1991), “GNP Report and Tables,” various issues. Christiano, L J. and M. Eichenbaum (1990), “Unit Roots in Real GNP: Do We Know, and Do We Care?” Federal Reserve Bank o f Chicago Working Paper #W P-90-2. Clark, P.K. (1979), “Potential GNP in the United States, 1948-80,” The Review of Income and Wealth 25, June 141-166. Clark, P.K. (1983), “Okun’s Law and Potential GNP,” Federal Reserve Board o f Governors, Washington, D.C.. DeLong, J.B. and L.H. Summers (1988), “How Does Macroeconomic Policy A f fect Output?” Brookings Papers on Economic Activity 1 ,4 3 3 -8 0 . Gall, J. (1989), “How Well does the IS-LM Model Fit Postwar U .S. Data?” Manuscript' Department o f Economics, Massachusetts Institute o f Technol ogy. Goldberger, A .S. (1974), “Unobservable Variables in Econometrics” inZarembka, (ed.), Frontiers in Econometrics. N ew York: Academic Press. Gordon, R J. (1975), “The Impact o f Aggregate Demand on Prices” Brookings Papers on Economic Activity 3 ,6 1 3 -6 6 2 . Gordon, R J. (1977), “Can the Inflation o f the 1970s Be Explained?” Brookings Papers on Economic Activity 1 ,2 5 3 -7 6 . FRB CHICAGO Working Paper June 1991, WP-1991-14 25 Gordon, R J. (1990a), "What Is New-Keynesian Econonomics?” Journal o f Eco nomic Literature 28, September 1115-1171. Gordon, R J. (1990b), Macroeconomics 5th edition. Glenview, IL: Scott Foresman. Hallman, J.J., R.D. Porter, and D.H. Small (1989), “M2 per Unit o f GNP as an Anchor for the Price Level,” Board o f Governors o f the Federal Reserve System: Staff Study #157. Hamilton, J.D. (1985), “Uncovering Financial Market Expectations o f Inflation,” Journal o f Political Economy 93, December 1224-1241. Hamilton, J.D. (1986), “A Standard Error for the Estimated State Vector o f a StateSpace M odel,” Journal o f Econometrics 33, December 387-398. Harvey, A.C. (1981), Time Series Methods. New York: W iley. Harvey, A.C. (1989), Forecasting, Structural Time Series Models, and the Kalman Filter. Cambridge: Cambridge University Press. King, R.G. and S.T. Rebelo (1989), “Low Frequency Filtering and Real Business C ycles,” Manuscript: Econom ics Department, Northwestern University. Lilien, D. (1982), “Sectoral Shifts and Sectoral Unemployment,” Journal o f P o litical Economy 90, August 7 77-93. Nerlove, M ., D. Grether, and J. Carvalho (1979), Analysis o f Economic Time Se ries: A Synthesis. New York: Academic Press. Norrbin, S.C. and D.E. Schlagenhauf (1988), “An Inquiry into the Sources o f Macroeconomic Fluctuations,” Journal o f Monetary Economics 2 2 ,4 3 -7 0 . Okun, A.M. (1970), The Political Economy o f Prosperity. Washington, D.C.: The Brookings Institution. Perloff, J.M. and M.L. Wachter (1979), “A Production Function - Nonaccelerating Inflation Approach to Potential Output: Is Measured Potential Output Tbo High?” Carnegie Rochester Coherence Series on Public Policy 10, 113164. Perry, G i . (1977), “Potential Output: Recent Issues and Present Trends,” in U.S. Productive Capacity: Estimating the Utilization Gap, Cento' for the Study o f American Business Working Paper #23, Washington University. FRB CHICAGO Working Paper June m i , WP.1991-14 26 Rissman, E.R. (1986), “What Is die Natural Rate o f Unemployment?” Federal Reserve Bank o f Chicago Economic Perspectives X, September-October 3-17. Plosser, C.I. and G.W. Schwert (1979), “Comment on Perloff and Wachter,” Carnegie Rochester Conference Series on Public Policy 1 0 ,1 7 9 -1 8 6 . Stock, J.H. and M. W. Watson (1988), “Variable Trends in Econom ic Time Series,” Journal o f Economic Perspectives 2, Summer 147-74. Watson, M.W. (1986), “Univariate Detrending Methods with Stochastic Trends,” Journal o f Monetary Economics 1 8 ,4 9 -7 5 . Watson, M.W. and R J7. Engle (1983), “Alternative Algorithms for the Estimation o f Dynamic Factor, MIMIC, and Varying Coefficient Regression M odels,” Journal o f Econometrics 23, 385-400. FRB CHICAGO Working Paper June 1991, WP-1991-14 27