Professional Documents
Culture Documents
Andrew Harvey
To cite this article: Andrew Harvey (2011) Modelling the Phillips curve with unobserved
components, Applied Financial Economics, 21:1-2, 7-17, DOI: 10.1080/09603107.2011.523169
Download by: [Universidad Nacional Colombia] Date: 24 September 2017, At: 18:16
Applied Financial Economics, 2011, 21, 717
The relationship between inflation and the output gap can be modelled
simply and effectively by including an unobserved random walk compo-
nent in the model. The dynamic properties match the stylized facts and the
random walk component satisfies the properties normally required for core
inflation. The model may be generalized so as to include a term for the
expectation of next periods output, but it is shown that this is difficult to
distinguish from the original specification. The model is fitted as a single
equation and as part of a bivariate model that includes an equation for
Gross Domestic Product (GDP). Fitting the bivariate model highlights
some new aspects of Unobserved Components (UC) modelling. Single
equation and bivariate models tell a similar story: an output gap 2% above
trend is associated with an annual inflation rate that is 1% above core
inflation.
1
An early draft of this article was written while I was visiting the Economics Department at the University of Canterbury,
Christchurch in late 2007. Clive Granger was a visitor at the same time and so it seems fitting to pay tribute to him by
publishing the article in this volume.
Applied Financial Economics ISSN 09603107 print/ISSN 14664305 online 2011 Taylor & Francis 7
http://www.informaworld.com
DOI: 10.1080/09603107.2011.523169
8 A. Harvey
depends only on current and (possibly) lagged values dogmatic statement of the way in which expectations
of the output gap, the model essentially reverts to its enter the model. Instead, we use UC models to yield
original specification. In these circumstances, it is a decomposition into persistent and transitory move-
difficult to produce convincing estimates of the ments, the interpretation of which is informed by
coefficient of the forward-looking term. economic theory.
The hybrid NKPC cannot adequately deal with The model is developed with US data. Output is
nonstationarity2 and is sometimes estimated after measured by the logarithm of quarterly real US GDP,
detrending, usually with a HodrickPrescott denoted yt, while the (annualized) rate of inflation,
(HP; 1997) filter. The use of detrended inflation t, is measured as the first differences of the quarterly
begs the question of what explains core inflation Consumer Price Index (CPI) multiplied by four. We
(see also Fukac and Pagan, 2010). Gali and Gertler have data from January 1947 to February 2007,
(1999, p. 203) write Oddly, enough, however, the obtained for GDP from Department of Commerce
hybrid Phillips curve has met with rather limited (website: www.bea.gov) and, for CPI, US Bureau of
success. Their response is to use marginal cost. But Labor Statistics (website: www.bls.gov). There is a
the problem is that the dynamic specification of the case for starting estimation in January 1952, since the
Downloaded by [Universidad Nacional Colombia] at 18:16 24 September 2017
hybrid NKPC is flawed and hence there is nothing early observations show quite extreme movements
odd about its failure. The proposed specification because of the Korean war. Models were also fitted
encompasses the hybrid NKPC and other models and using the GDP deflator to measure inflation.
in doing so exposes their shortcomings. Although there are some differences, the overall
The papers by Kuttner (1994), Planas and Rossi message remains the same and the results are not
(2004), Domenech and Gomez (2006) and Planas reported.
et al. (2008) are related in that they use Unobserved
Components (UC) in models linking inflation with
the output gap. The inflation equation proposed here Output gap
differs significantly from the ones in Kuttner (1994) A trend-cycle model can be set up as
and Planas and Rossi (2004) in that the lagged
growth rate of GDP is dropped and a stochastic trend y t t t "t, t 1, . . . , T 1
is included. Note that the main motivation in these
papers is to use the information in inflation to obtain where t is an integrated random walk,
better estimates of the output gap. This point is worth t t1 t1,
bearing in mind when bivariate estimation of our t 1, . . . , T 2
t t1 t,
model is considered. The statistical case for the
proposed model is made in Section II where univar- t is a stochastic cycle and "t is white noise. The
iate UC methodology is used to examine the output stochastic cycle
and inflation gaps and the relationship between them.
t cos c sin c t1 t
The model is set out in Section III and extended to ,
include a forward-looking term. Estimation is con- t sin c cos c t1 t
sidered in Section IV. Bivariate estimation makes use t 1, . . . , T 3
of some new capabilities in the Structural Time Series
Analyser Modeler and Predictor (STAMP) package where c is frequency in radians, is a damping
of Koopman et al. (2007). factor, with 0 1, and t and t are two mutually
independent white noise disturbances with zero
means and common variance 2 . The reduced form
is an Autoregressive Moving Averge (ARMA(2,1))
II. Preliminary Modelling and process in which the autoregressive part has complex
Stylized Facts roots. The disturbances, "t , t , t and t are serially
and mutually uncorrelated with variances "2 and 2
It is useful to begin with an exploration of the for the irregular and slope. The model is assumed to
relationship between inflation and output based on be Gaussian and is estimated by Maximum
fitting univariate UC models. We do not want to Likelihood (ML). The smoothed estimates of the
impose a tight theoretical specification at the outset, cycle can serve as a measure of the output gap. When
particularly one involving a specific lag structure or a the model is fitted to US GDP, "2 is estimated to be
2
Furthermore, GMM attempts to estimate the model with nonstationary series run into difficulties (Pesaran, 1987; Stock
et al., 2002; Mavroeidis, 2005).
Modelling the Phillips curve with unobserved components 9
zero and so the cycle is the same as the detrended seasonal, as specified in Koopman et al. (2007), have
series. A smoother cycle can be produced by using the been added to Equation 4, that is
higher-order models proposed in Harvey and
t t t t " t , t 1, . . . , T
Trimbur (2003).
The HP filter is widely used for detrending time If the inflation gap is estimated by the cycle, it is
series in macroeconomics. The detrended series can somewhat smoother than the detrended (and season-
be computed as the smoothed estimates of the ally adjusted) series because the irregular has been
irregular component in a model in which yt t filtered out.
"t , where t is an integrated random walk as in The above formulation in terms of additive com-
Equation 2 and signalnoise ratio, q 2 = "2 , is fixed ponents is preferred to one in which the dynamics of
at 1/1600 for quarterly data. The output gap series the inflation gap are picked up by a lagged dependent
produced by HP detrending is close to the cycle series variable, that is
obtained by fitting the trend-cycle model. However,
t t t1 "t , "t NID 0, "2 , t 1, . . . , T
while the HP filter may be relatively efficient in the
middle of a series, it is much less efficient at the end 6
Downloaded by [Universidad Nacional Colombia] at 18:16 24 September 2017
(Mise et al., 2005). Furthermore, a model is needed to The disadvantage of Equation 6 is that the dynamics
produce forecasts. of the inflation gap are imposed on the underlying
level (core inflation), since
0.05 0.00
0.00
0.01
0.05
1960 1980 2000 1960 1980 2000
0.025
Downloaded by [Universidad Nacional Colombia] at 18:16 24 September 2017
0.00
0.000
0.05 0.025
Fig. 1. Inflation and its decomposition into stochastic level, cycle, seasonal and irregular
0.08
0.04
0.02
0.00
0.02
0.04
0.06
1950 1955 1960 1965 1970 1975 1980 1985 1990 1995 2000 2005
Fig. 2. Smoothed estimates of the cycles obtained from univariate models for inflation and (log) GDP
the lagged dependent variable was replaced by a the discounted sequence of future output gaps. If xt is
future expectation to give assumed to be a stationary first-order Autoregression
(AR(1)) with coefficient
, the model reverts to (7)
t Et t1 xt "t , t 1, . . . , T 8
with =1
. If xt were assumed to be
where 0 1. However, when expectations are AR(p), for p 2 it would introduce p 1 lags of xt
based on this model, inflation depends exclusively on into the equation and allow to be identified
Modelling the Phillips curve with unobserved components 11
(Pesaran, 1987, propositions 6.1 and 6.2; Nason and steady state3 this is an Exponentially Weighted
Smith, 2008). More generally, identifying information Moving Average (EWMA), that is
is only available when p is greater than the number of X
1
lags in the original equation. Et1 t 1 j t1j xt1j 12
The hybrid NKPC is j0
p
t t1 Et t1 xt "t 9 and q q2 4q=2. This is very different
from just including a single lagged dependent vari-
Solving for Et t1 , as in Nason and Smith (2008),
able, t1:
and substituting in the original equation yields
The model can also be regarded as forward-
X1 j looking. The conditions under which a model with
1
t t1 Et xt1j an expectational term reverts to Equation 10 are
1 1 1 1 2 j0 2
derived below. The fact that the forecast function for
1 a random walk component is constant simplifies
xt "t
1 1 1 1 matters considerably.
Downloaded by [Universidad Nacional Colombia] at 18:16 24 September 2017
where 1 and 2 are, respectively, the stable and The reduced form of Equation 10 is an ARMA
unstable roots of L1 1 1 L: Identifiability with Exogenous terms (ARMAX) model in which
issues are much the same as with the NKPC. Thus, if Dt is equal to Dxt plus a moving average distur-
xt is AR(1), the model has the form of the original, bance term. The autoregressive distributed lag
Equation 9. The rate of inflation is stationary if reduced form
j=1 1 j 5 1. Setting 1 , implies 1 1 so X
1 X
1
3
In practice, the Kalman filter is initiated with a diffuse prior and it only approaches the steady-state asymptotically (Harvey,
1989). However, it simplifies matters to assume a steady-state at the outset.
12 A. Harvey
showing that inflation decomposes into a linear the trend and cycle are (negatively) correlated. They
combination of the expectation of core inflation, the estimate to be 0.29.
output gap, the discounted sum of expectations of the As a final point, note that the model proposed by
output gap and a residual. Domencech and Gomez (2006) in their Equation 8 is
The future expectations can be removed as before
t 1 t t1 xt "t , t 1, . . . , T
by making an assumption about the process followed
by xt . Thus if xt is AR(1) and j
j < 1, Equation 25 16
becomes where the notation has been adapted so as to be
similar to that used here and there is only one lag on
t yt xt xt "yt yt xt "yt t to simplify the discussion. Setting 0 gives
1
1
Equation 10, although if Equation 10 is generalized,
This cannot be distinguished from Equation 10, as is done later, so that t t xt is a stationary
but if xt is a stationary AR(2), which corresponds process, rather than white noise, the two formulations
approximately to the stochastic cycle used in are no longer nested; the reasons for preferring an
Equation 1, Et xt1j , j 0, 1, 2, . . . , depends on xt additive formulation were discussed at the end of the
Downloaded by [Universidad Nacional Colombia] at 18:16 24 September 2017
0.1 0.1
0.0 0.0
Inflation Level Inflation Level+reg
0.1 0.000
0.0
0.025 Inflation-Regression
0.0 0.00
Fig. 3. Components (excluding seasonal) in model relating inflation to (lagged) output gap
subsequent sharp fall in the early 1980s, this is not results with four lags are as shown in Table 1. There
surprising. However, if we start in January 1986, the is no clear interpretation to the pattern. Using the
diagnostics are much better, even without including a Autometrics option in the PC Give module4 in
cycle. There is no evidence for lags beyond one and Oxmetrics gives lags at one and three only and the
the contemporaneous and lagged one estimates of negative coefficient in lag 3 of the output gap is
output gap coefficients, 0 and 1, are 0.11 (0.34) and puzzling, to say the least. Note that the seasonally
0.39 (1.19). In fact, a contemporaneous output gap adjusted CPI series was used to create the inflation
provides a good it, the estimate of being 0.49 with a series since using unadjusted data with fixed seasonal
t-statistic of 3.56. Using the output gap series dummies would have resulted in the changing sea-
estimated from 1947 gives an estimate of 0.42. sonal pattern becoming incorporated into the lag
Multi-step forecasts from the end of 1997 are structure. The estimates from the full data also show
shown in Fig. 4. The movements, which are condi- no clear pattern.
tional on the output gap, are not big but the higher
inflation around 2000 is picked up. The volatility of
the series in recent years has made accurate forecast- Bivariate model
ing of any one quarter difficult, a point made by
Rather than first estimating the output gap from a
Stock and Watson (2007).
If filtered estimates of the output gap are used in univariate model for GDP, inflation and GDP may
the equation estimated from 1986, the diagnostics are be modelled jointly as
still satisfactory, but the coefficients on the output t t t "t
gap are small and insignificant. Using the predictive y 18
yt yt t "yt
(one-step ahead) filter, obtained starting in 1947, the
coefficient on the current output gap is 0.16(0.38). where t is a random walk, as in Equation 5, and yt
With the contemporaneous filter it is 0.09(0.26). is an integrated random walk, as in Equation 2. These
Autoregressive distributed lag models, Equation two stochastic trends are independent of each other.
13, are often used to make forecasts based on the The irregular disturbances may be correlated and
Phillips curve (see, e.g. Orphanides and van Norden, are assumed to have a covariance matrix, ".
2002). However, the estimates are erratic and difficult A seasonal component can be added to the model
to interpret. For a data set beginning in 1986, the and in the estimates reported seasonal effects were
4
Doornik and Hendry (2007).
14 A. Harvey
0.06
0.04
0.02
0.00
Downloaded by [Universidad Nacional Colombia] at 18:16 24 September 2017
0.02
1998 1999 2000 2001 2002 2003 2004 2005 2006 2007
Fig. 4. Multi-step predictions made from the end of 1997, conditional on the output gap estimated from the full sample
included in the equation for inflation. The stochastic univariate models, the trend, cycle, seasonal and
cycles are modelled as similar cycles, as in Harvey irregular components are assumed to be mutually
et al. (2007), so that if wt t , yt 0 then independent.
A simple transformation of the bivariate similar
wt cos c sin c wt1 jt cycle model allows the cycle in inflation to be broken
I2 ,
wt sin c cos c
wt1 jt down into two independent parts, one of which
depends on the GDP cycle,
that is t yt y
t ,
t 1, . . . , T y y
where Cov t , t =Var t Cov t ,yt =Var yt
19 and yt is a cyclical component specific to inflation.
where jt and jt are 2 1 vectors of the disturbances Substituting in the inflation equation in Equation 18
0
such that Ejt j0 t Ejt jt , where is a gives
0
2 2 covariance matrix, and Ejt jt 0: Because y y
t t t t "t
the damping factor and the frequency, and c , are
the same in all series, the cycles in the different series If the cycle disturbances and t are yt perfectly
have the same autocorrelation functions. As in the correlated, the above expression corresponds to
Modelling the Phillips curve with unobserved components 15
INFL ATION Level log_GDP Level
0.075
9.2
0.050
0.025
9.0
0.000
0.025 8.8
1985 1990 1995 2000 2005 1985 1990 1995 2000 2005
0.03 log_GDP-Cycle
INFL ATION Level+Cycle
0.075
0.02
0.050
0.01
0.025
0.00
Downloaded by [Universidad Nacional Colombia] at 18:16 24 September 2017
0.000 0.01
0.025 0.02
1985 1990 1995 2000 2005 1985 1990 1995 2000 2005
Fig. 5. Smoothed components from a bivariate model for GDP and inflation