See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/24075766

Consequences of Hodrick–Prescott filtering for parameter estimation in a

structural model of inventory behavior

Article  in  Applied Economics · February 2006

DOI: 10.1080/00036840500427254 · Source: RePEc

CITATIONS READS
10 2,288

1 author:

David John Doorn

West Chester University
8 PUBLICATIONS   45 CITATIONS   

SEE PROFILE

All content following this page was uploaded by David John Doorn on 21 May 2014.

The user has requested enhancement of the downloaded file.

Applied Economics, 2006, 38, 1863–1875

Consequences of Hodrick–Prescott
filtering for parameter estimation
in a structural model of
inventory behaviour
D. Doorn
Department of Economics, University of Minnesota Duluth,
112 Library Drive, 165 SBE, Duluth, MN 55812, USA
E-mail: ddoorn@d.umn.edu

Much work in macroeconomics relies on detrending a time series prior to

analysis. A popular method of detrending has been the Hodrick–Prescott
(HP) filter. This filter has been widely applied in the Real Business Cycle
literature to isolate the behaviour of economic variables at business cycle
frequencies and to look at comovements between series over the business
cycle. Prior work has shown that the use of this filter can have serious
consequences for such analysis, such as inducing spurious correlations, and
that a researcher should proceed with caution when applying the filter.
Another use of HP filtering has been to achieve stationarity prior to
estimation of structural econometric models. Little work has been done
concerning the possible effects this method of detrending may have on
parameter estimation from such models. Given the problems with the filter
noted in the literature, it is likely these effects may be of some consequence
to estimation results. Using a common model of inventory behaviour,
a simulation study is conducted to assess the impact of using the HP filter
for detrending prior to estimation. A comparison will be made to other
methods of handling trend to gauge relative performance.

I. Introduction interest in and of itself for describing how a series

Since the important work of Nelson and Plosser
(1982) there has been much interest in determining
the time-series properties of macroeconomic data
series. Recent examples in this publication alone
include Gil-Alana (2004) and Sen (2004). Of partic-
ular interest has been the determination of whether a
time series follows a deterministic trend, possibly
piecewise linear,1 or a stochastic trend containing one
or more unit roots. While this determination is of
evolves over the long-run and establishing whether
macroeconomic shocks are permanent or transitory
in their effects, it is also useful in deciding upon the
best means of removing a long-run trend from a time
series prior to its further use.
In recent years the Hodrick–Prescott (HP) filter,
developed in Hodrick and Prescott (1997), has
become a popular method for removing the growth
component from data exhibiting long-run trends over
time. While application of the filter typically does not

1
Sen (2004) investigates this possibility for updated versions of the series’ examined in Nelson and Plosser (1982) and Doorn
(2003) finds trend breaks in aggregate inventories and sales at the two-digit SIC industry level.
Applied Economics ISSN 0003–6846 print/ISSN 1466–4283 online ß 2006 Taylor & Francis 1863
http://www.tandf.co.uk/journals
DOI: 10.1080/00036840500427254
1864
rely upon pre-determination of the time-series prop-
erties of the data, it is capable of removing both linear
and stochastic trends. The HP filter has been much
applied in the Real Business Cycle (RBC) literature as
a method of isolating business cycle frequencies from
the long-run trend component in both actual and
simulated data. The goal in that literature has been to
study the business cycle behaviour of different
macroeconomic variables and to determine the
degree of any comovements between them over
the course of the cycle. This generally involves a
correlation study of the filtered data.2 Unfortunately
for this work, studies of the impact HP filtering can
have on the second moments of a data series have
shown it to have the potential to introduce spurious
correlations in the filtered data and spurious cross-
correlations between series. This may cause the
stylized facts gained through the study of HP filtered
data to be questionable at best. There have been a
number of papers showing this potential for the HP
filter to distort results from RBC models and
empirical studies, most notably Harvey and Jaeger
(1993), King and Rebelo (1993), and Cogley and
Nason (1995).
Another application of the HP filter has been as an
ad hoc method of detrending data series which are
subsequently used in estimation of parameters from
structural econometric models. Most estimation
methods require the data used to be covariance
stationary in order for the underlying asymptotic
theory to hold. While seemingly useful as a tool for
detrending and achieving stationarity, the above-
mentioned problems have the potential to affect
estimation using HP filtered data as well. If the cross-
correlations between variables are affected by the
filtering process, then estimation of a model posited
as governing the relationship between those variables
will likely be affected as well, along with any
inferences based on the estimated parameters. It is
the goal of this paper to assess whether there is a
significant impact on parameter estimation when
using HP filtered data as opposed to other methods
of detrending.
To effect this study of the impact HP filtering may
have on estimation a simulation study is conducted
based on a widely-studied model of inventory
behaviour. The simulated data generated by the
model contain simple linear trends. The simulated
data is then used to estimate the population
parameters of the model using the Generalized
Method of Moments (GMM) estimator of Hansen
(1982). The model is estimated several times using
2
A recent example of such work in this journal is Sensier (2003), in which wide use of the HP filter was made in the study of
asymmetry in the business cycle behaviour of inventories in the UK.
D. Doorn
different criteria for detrending the data prior to
estimation, including removing a linear trend and HP
filtering with different values for the smoothing
parameter, discussed below. For comparison, the
model is also estimated without prior detrending of
the data but instead estimating deterministic compo-
nents along with the other parameters of the model.
The paper proceeds as follows: First the Hodrick–
Prescott filter is presented and some of the previous
results mentioned above discussed. Next a structural
model of inventory behaviour is derived with which
to simulate data. The model is then calibrated to
actual data on inventories and sales of non-durable
goods. Once this is done, simulated data is used to
estimate the parameters of the model several times
using different methods for dealing with the trend.
A comparison will be made of the distributional
characteristics across the different detrending meth-
ods. In addition, the effect of detrending method on
different test statistics resulting from the estimation
will be considered.
II. The Hodrick–Prescott Filter
This derivation of the Hodrick–Prescott filter closely
follows that of King and Rebelo (1993). The HP filter
is based on the assumption that a time-series process
can be modelled as the sum of a growth component
and a cyclical component, that is yt ¼ ygt þ yct . In the
real business cycle literature it is desirable to remove
the growth component, whether it be a stochastic or
deterministic trend, in order to study the behaviour of
the cyclical component and to compare that behav-
iour between different series. The HP filter achieves
this by defining the cyclical component as
yct ¼ yt
ygt ð1Þ
where yct becomes the desired detrended series. The
HP filter minimizes the variance of Equation 1 while
penalizing for excessive changes in the growth
component. This can be expressed in terms of the
following minimization problem
T n

X 2 
2 o
min yt
ygt þ
ygtþ1
ygt Þ
ðygt
ygt
1
g
fyt gTt¼1 t¼1
ð2Þ
where
is termed the ‘smoothing parameter’, since its
value determines the penalty applied to excessive
changes in the growth rate of the series. The value
Impact of Hodrick–Prescott filtering on parameter estimation 1865
chosen for
will differ depending on the sampling 1.0
frequency of the data and represents the trade-off
between smoothness and goodness of fit in the 0.8
resulting series.3
0.6
Making use of lag operators, the first-order
condition from the minimization problem results in 0.4

yt
ygt 2
¼
ð1
LÞ ð1
L
1 Þ2 ygt 0.2

or 0.0
2
yt ¼ ½1 þ
ð1
LÞ ð1
L Þ
1 2
ygt – 0.1 0.1 0.3 0.5 0.7 0.9 1.1
Frequency
By inversion this can be put in terms of a low
frequency filter applied to yt, which yields the growth Fig. 1. Frequency response of HP cyclical filter: ( )
lambda ^ 1600, ( ) lambda ^ 14 400, ( ) lambda ^
component ygt ¼ Hg ðLÞyt , where 86 400, ( ) lambda ¼ 129 600, ( ) Ideal-monthly, ( )
 
1 Ideal-Quarterly. This plot depicts the squared gain of the
Hg ðLÞ ¼ 1 þ
ð1
LÞ2 ð1
L
1 Þ2 ð3Þ filter HP(u) for j ^ 1600, j ^ 14 400, j ^ 86 400, and
j ^ 129 600. Also shown are the ideal highpass filters for
The cyclical component of the series is then given by monthly and quarterly data. See text for details
yct ¼ yt
ygt ¼ ð1
Hg ðLÞÞyt ¼ Hc ðLÞyt , where

½1
L2 ½1
L
1 2
L
2 ð1
LÞ4
Hc ðLÞ ¼ ¼
1 þ
½1
L2 ½1
L
1 2 1 þ
L
2 ð1
LÞ4
differencing the data with (1
L)4 and then applying
L
2 ð1
LÞ4 the filter L
2/(1/
þ L
2)(1
L)4 to the result.
¼ ð4Þ
ð1=
Þ þ L
2 ð1
LÞ4 Figure 1 plots the frequency response of the filter
In the frequency domain, the squared gain, or for different values of
, chosen to reflect values used
frequency response, of this filter is in the literature for quarterly data (
¼ 1600) and
monthly data (
¼ 14 400, 86 400, 129 600). Also
 

½1
e
i! 2 ½1
ei! 2 2 depicted are the ideal filters for both monthly and
 
HPð!Þ ¼   quarterly data, which only pass frequencies above a
1 þ
½1
e
i! 2 ½1
ei! 2 
  period of eight years, often chosen to be the cutoff
 4
½1
cosð!Þ2 2 period for business cycle research. On the vertical axis
 
¼  are the weights applied by the filter to different
1 þ 4
½1
cosð!Þ2 
 2 frequencies in the data. Notice that very low
 4½1
cosð!Þ2  frequencies are completely attenuated by the filter
 
¼  ð5Þ while high frequencies are passed without alteration.
ð1=
Þ þ 4½1
cosð!Þ2 
Also note that regardless of the value of
, more low
where ! is frequency measured in radians. This frequency content is passed by the HP filter than may
representation allows the effect the filter has on cycles be ideal if the goal is to study behaviour of a series at
of different frequencies in the data to be seen. business cycle frequencies. In addition, some higher
Note that HP(0) ¼ 0, indicating attenuation of low frequency content is attenuated to the right of the
frequencies, and also that HP () ¼ (16/(1/
þ 16)), cutoff points of the ideal filters. This attenuation
which is approximately equal to one for large
.4 This tends to decrease as
is increased, although more
implies close to unit gain on high frequencies. In fact, undesirable low-frequency content is then left as well.
as
goes to infinity, the filter passes all but the zero Regardless of the value of
, the HP filter always
frequency, meaning it removes a deterministic linear removes a linear trend, but operates further on
trend only. In addition, the four differences in the deviations from this trend. Singleton (1988) shows
numerator of Equation 4 indicate it is capable of that when applied to stationary data, the HP filter is a
rendering stationary integrated processes of up to very good approximation to an ideal high pass filter,
fourth order.5 This would be the equivalent of which only allows cycles of greater frequency than
3
The smoothing parameter can be viewed as the ratio of the variance of the cyclical component to that of the change in the
growth component, although alternative interpretations are available. See, for example, Harvey and Jaeger (1993), Hodrick
and Prescott (1997), and Ravn and Uhlig (1997).
4

is typically set to 1600 for quarterly data and 14 400 for monthly. See Hodrick and Prescott (1997).
5
See King and Rebelo (1993).
1866
some prespecified value to pass. This implies that if
the data is linear trend stationary, the HP filter acts as
a high pass filter on the residuals from the
deterministic trend.6 This property makes the filter
potentially useful in business cycle research by
allowing the isolation of business cycle frequencies
from the data. However, if the goal of filtering is
primarily to render data stationary prior to use in
estimation of econometric models, without regard to
business cycle behaviour, applying such a high pass
filter may be removing more low-frequency content
than is perhaps desirable. This may be especially
harmful if most of the power of the series lies in the
low frequencies, as tends to be the case with many
macroeconomic series. An example of this is given in
Fig. 2, which plots the spectral density of monthly
aggregate inventories of non-durable goods in the top
panel, and then the same series detrended using the
HP-filter with different values for
. Also included is
the spectral density of the linearly detrended series.
The dashed line in the figure indicates a period of
eight years, which is generally considered to be the
maximum length of the business cycle. The HP-filter
is designed to attenuate any frequencies with period
above eight years, which the figure indicates it does
fairly well. Again, this seems to be a good procedure
if the goal is to study the behaviour of a series at
business cycle frequencies and to compare that
behaviour across series. However, as a general
detrending method it is clear that HP filtering
removes much of the information in the data
contained in the lower frequencies. In addition,
even at frequencies above eight years, the filter
removes more of the variability from the series than
does linearly detrending, which is due to the
frequency response of the filter as depicted in
Fig. 1. Note, however, that as
is increased more
of the low frequency information is retained, which
makes sense since the filter becomes equivalent to
linear detrending as
goes to infinity.
Application of the HP filter to a time series results
in a trend that very closely follows the time path of
the data. While visually appealing, this may not
necessarily be desirable.7 Figure 3 indicates the
resulting differences in the residuals between remov-
ing a linear trend and applying the HP filter to the
aggregate inventory data. Clearly the HP filter results
in a very different detrended series.
6
See Cogley and Nason (1995).
7
As an example, McCallum (2000) contains an application of the HP filter to US real GDP over the 1920s and 1930s. The
resulting detrended series gives little evidence of the Great Depression.
8
See Harvey and Jaeger (1993) and King and Rebelo (1993).
9
Harvey and Jaeger (1993) discuss estimating
using maximum likelihood methods. Pedersen (1999) proposes other methods
to determine the optimal value of
.
10
They found that such detrending caused cycles to appear in the detrended data that were not present in the original series.
D. Doorn
Under certain conditions, the HP filter can be
shown to be the optimal linear filter, minimizing the
mean square error between the actual cyclical
component and its estimate.8 However, these condi-
tions are fairly restrictive, with the strongest being
that the growth component and the cyclical compo-
nent are uncorrelated. This would indicate that these
are generated by completely separate economic
events. In addition, the particular trend specification
required for optimality causes the overall process to
be I(2).For many economic time series this is unlikely
to be the case. Finally, optimality requires a priori
knowledge of the variances of the disturbance terms
for the separate cyclical and growth components in
order to set the smoothness parameter,
, at its
optimal value. Since the purpose of applying the filter
to begin with is to isolate the trend and cyclical
components, it is unlikely that these will be known
ahead of time.9 In practice
is generally set based
upon common practice and the sampling frequency of
the data. These conditions make it unlikely that the
HP filter is an optimal filter in most applications.
However, even when optimal the filter tends to cause
distortions in the auto-correlations of filtered series.
Ehlgen (1998) finds that the autocorrelations and the
variances of HP filtered data tend to be too small.
Further, he states that this can be a normal
characteristic of optimal signal extraction filters and
not just a problem with the HP filter.
Cogley and Nason (1995) look at the effects of HP
filtering on difference stationary data and the
implications of this for business cycle research.
They posit that, if the series is difference stationary
then applying the HP filter will likely result in
problems akin to linearly detrending a random
walk, which has been studied by Nelson and Kang
(1981).10 This implies that application of the HP filter
to difference stationary data will likely cause pro-
blems of filter induced cyclicality in the resulting
filtered series. Application of the HP filter to an I(1)
process is equivalent to applying the filter
L
2 ð1
LÞ3
SðLÞ ¼ ð6Þ
ð1=
Þ þ L
2 ð1
LÞ4
to the stationary differenced series. In this case, S(L)
is not a high pass filter. In investigating its properties
Cogley and Nason (1995) find that the filter S(L)
Impact of Hodrick–Prescott filtering on parameter estimation 1867

35000000000.00

20000000000.00

5000000000.00

473 157.67 94.6 67.57 52.56 43 36.38 31.53 27.82 24.89 22.52 20.57 18.92 17.52 16.31 15.26 14.33 13.51 12.78 12.13
236.5 118.25 78.83 59.13 47.3 39.42 33.79 29.56 26.28 23.65 21.5 19.71 18.19 16.89 15.77 14.78 13.91 13.14 12.45

800000000

500000000

200000000

473 157.67 94.6 67.57 52.56 43 36.38 31.53 27.82 24.89 22.52 20.57 18.92 17.52 16.31 15.26 14.33 13.51 12.78 12.13
236.5 118.25 78.83 59.13 47.3 39.42 33.79 29.56 26.28 23.65 21.5 19.71 18.19 16.89 15.77 14.78 13.91 13.14 12.45

Fig. 2. Spectral density of total aggregate inventories of non-durables: ( ) LDT, ( ) HP1, ( ) HP2, ( ) HP3. The
top panel is the spectral density of the series in levels. The bottom panel shows the spectral densities of the detrended series from
the different methods, with LDT indicating linear detrending and HP indicating HP filtering with j ^ 14 400 for HP1, j ^ 86
400 for HP2, and j ^ 129 600 for HP3. The dashed line indicates the cutoff period of eight years, which is generally considered to
be the maximum length of the business cycle

will result in spurious cyclicality when applied to
stationary data. Furthermore, HP filtering of two
independent random walk processes was found to
induce cross-correlations between the series that were
previously non-existent. Harvey and Jaeger (1993)
also show that the HP filter can induce spurious
cycles that are non-existent in the original series when
applied to an ARIMA(p, d, q) process. They found
that there is a non-zero probability of inducing
spurious cross-correlations between two independent
series that have been HP filtered.
The implication of these findings is that the cyclical
properties of HP filtered data cannot be relied upon
for inference about the business cycle behaviour of
1868
4000
2000
0
– 2000
– 4000
– 6000
0.00 100.00 200.00 300.00 400.00 500.00
Date
Fig. 3. Residuals: ( ) Linear trend ( ) HP1. A
comparison of residuals from detrending monthly aggregate
inventories of non-turables with a linear trend as well as with
the HP filter using j ^ 14 400
real economic series, particularly if the series are
likely to be integrated of order one or higher. To the
best of the author’s knowledge there have been no
studies to date on the possible impact these problems
may have on estimation of econometric models using
HP-filtered data. Although the filter always removes
a linear trend from the data, it was chosen to study
the effects of applying the filter to data simulated to
contain a linear trend, as this seems the simplest case
with which to start. As mentioned above, the HP
filter will remove the trend but also act as a high-pass
filter on what is left. The result, as seen in Fig. 2, is
that the filter removes more of the variation in the
data than would just simply detrending, which in turn
implies that there is less information left with which
to effect estimation. This suggests that the results
from estimation performed with filtered data are
likely to be impacted. It is the purpose of the rest of
this paper to give evidence of this.
III. The Model
To get an idea of how the Hodrick–Prescott filter
affects estimation results a simulation is used based
on the linear-quadratic model of inventory behaviour
that has been widely studied in the macroeconomics
literature. The simulation model used is that of West
and Wilcox (1996). In this paper, the model is
calibrated to monthly Bureau of Economic Analysis
data on inventories and sales of non-durable goods
spanning January 1959–May 1998.
Consider a representative firm that maximizes the
present discounted value of profits subject to a cost
11
See Sensier (2003) for a review and discussion of these properties of the inventory holding cost function.
D. Doorn
function that accounts for production costs, costs of
changing production, inventory holding costs, and
backlog costs that arise when sales cannot be met out
of inventories.11 This particular model also includes
an unobservable cost shock, that may be observable
to the firm but not the econometrician. The
optimization problem becomes
X
T
max lim Et bj ðptþj Stþj
Ctþj Þ
ðHtþj Þ T!1
j¼0
s.t.
a0 a1 a2
Ctþj ¼
Q2tþj þ Q2tþj þ ðHtþj
1
a3 Stþj Þ2
2 2 2
þ Htþj utþj þ deterministic terms ð7Þ
Qtþj ¼ Stþj þ Htþj
Htþj
1 ð8Þ
where pt þ j is price, St þ j is real sales, Ht þ j real end of
period inventories, Ct þ j real costs, and Qt þ j real
output. The a terms represent the marginal costs
faced by the firm, with a0 the marginal cost of
changing production, a1 the marginal cost of
production, a2 the marginal cost of holding inven-
tories, and a3 the cost associated with stock outs.
Unobserved cost shocks are represented by ut þ j, with
ut assumed to be I.I.D. Normal with zero mean.
Deterministic terms would account for any linear and
trend terms that may exist in the data. The final term,
b, is the discount factor and is usually set a priori in
estimation of the model. The parameters of the cost
function are generally what is estimated, and there-
fore these will need to be calibrated in the simulation.
The first order condition of the model yields:

Et a0 ð
Qt
2b
Qtþ1 þ b2
Qtþ2 Þ þ a1 ðQt
bQtþ1 Þ

þ ba2 ðHt
a3 Stþ1 Þ þ dðtÞ þ ut ¼ 0 ð9Þ
where d(t) represents deterministic terms. Leaving out
deterministic terms and substituting in for production
from Equation 8 gives

Et a0 b2 Htþ2
½a0 ð2b2 þ 2bÞ þ a1 bHtþ1
þ ½a0 ð1 þ 4b þ b2 Þ þ a1 ð1 þ bÞ þ ba2 Ht

½a0 ð2b þ 2Þ þ a1 Ht
1 þ a0 Ht
2
þ a0 b2 Stþ2
½a0 ðb2 þ 2bÞ þ a1 b þ a2 a3 bStþ1

þ ½a1 þ a0 ð1 þ 2bÞSt
a0 St
1 þ ut ¼ 0 ð10Þ
In order for there to be an optimal solution, the
Legendre–Clebsch condition states that the second
Impact of Hodrick–Prescott filtering on parameter estimation 1869
Table 1. Specification of simulation parameters innovations. Since the parameters in Equation 13
are functions of the ai’s they can be determined given
Sales process and cost shock
some choice for the cost parameters. West and
KS  1 2 "2S u2 "S , u Wilcox (1996) use several sets of values for these
531 0.001211 0.75 0.20 0.11625 3.5
0.5 based on other work. Here it is chosen to instead
Implied inventory process determine values for the cost parameters that fit the
particular data set being calibrated to. In order to do
KH  1 2 1 2
1050 0.002685 1.06
0.12 0.006
0.065 this it was necessary to jointly estimate the exogenous
sales process, Equation 12, and the reduced form for
inventories, Equation 13. From the parameter
estimates for Equation 13 estimates can be obtained
for combinations of the cost parameters using13
derivative of the objective function must be non- a1 =a0 ¼ 1 ðb

1
2 Þ
2b
2 ð14Þ
negative.12 This yields the requirement that
 2
a0 ð1 þ 4b þ b2 Þ þ a1 ð1 þ bÞ þ ba2  0 ð11Þ a2 =a0 ¼
b
1 
1
2 ð1 þ b1 Þ

In order to use the model to simulate data it is þ b2 2 þ ð1 þ bÞa1 =a0 þ ð1 þ 4b þ b2 Þ ð15Þ
necessary to solve for a reduced form equation in
inventories. This requires a specification of the sales and from estimates of the sales process a3 can also be
process. Here it is assumed for sales to be exogenous derived.
and described by an AR(2) process: In calibrating the model, estimates from BEA data
on total aggregated inventories and sales of non-
St ¼ KS þ t þ 1 St
1 þ 2 St
2 þ "St ð12Þ durable goods are roughly matched. In fixing the ai’s
where KS and t are constant plus trend terms and "St for the simulation Blanchard (1983) is followed
is assumed I.I.D. Normal. and a0 is normalized to one. Based on the estimates,
With the sales process specified and assuming this results in a0 ¼ 1, a1 ¼ 5.898, a2 ¼ 0.032, and
uncorrelated cost shocks the model can be solved for a3 ¼ 7.191.
a reduced form equation for inventories. The solution With estimates for the cost parameters in hand
methods used are described in Blanchard (1983) and data for inventories and sales can be simulated by
West (1992) and also are available in the form of a specifying the parameters of the exogenous sales
technical appendix to this paper, which may be process along with the distributions of the sales and
obtained from the author upon request. The derived cost innovations. The autoregressive parameters of
equation for inventories is the sales process are set to approximately match those
of the present data. These are given in Table 1.
Ht ¼ 1 Ht
1 þ 2 Ht
2 þ 1 St
1 Following West and Wilcox (1996), the sales innova-
tion variance is set to normalize the unconditional
þ 2 St
2 þ t þ KH þ "Ht ð13Þ
variance of the sales process to one. In addition
where the same cost shock variance and correlation with
2 2 sales are used given in that paper, as these values
"Ht ¼ ut þ "St conveniently result in the variance ratio and the
a0 2
correlation between inventories and sales closely
with 1and 2 being functions of the marginal cost matching that of the data. The sales innovation
parameters, 1 and 2 functions of 1, 2 and the variance and the variance of the cost shock also
parameters from the sales process, and t þ KH being are given in Table 1, along with the correlation
trend plus constant terms. The error term, "Ht, between the two. With the ai’s fixed and the sales
indicates the correlation between the inventory and process specified implied values for 1, 2, 1, and
sales processes. 2 can be determined.14 These also are reported in
In order to simulate the data it is necessary to Table 1.
specify values for the parameters of the exogenous To determine coefficients on trend terms West and
sales process, the reduced form inventory equation, Wilcox (1996) is again followed and these are set so
and also the distribution of the sales and cost that the implied coefficients of variation for
St
12
See Ramey (1991) for a discussion of this optimality condition.
13
See Blanchard (1983).
14
See technical appendix, available upon request.
1870
and
Ht matches that of the present data, which is
found to be approximately 14.6 for sales and 9.7 for
inventories. The coefficient of variation is  1/2/ 
where  is the variance of the process and the mean,
which will involve the trend terms for the process in
levels. The exact solution for the trend coefficient, as
well as constant terms, is available in the technical
appendix and yields the values given in Table 1.
The sales and inventory processes specified in
Equations 12 and 13 are used to generate simulated
data for sales and inventories. One thousand samples
were generated by first obtaining starting values for
St
1, St
2, Ht
1 and Ht
1 by generating series of
length T ¼ 1000 from each process while omitting the
trend terms. The last two observations from each of
these series are then used as the starting values in
generating an additional 1000 observations from both
the inventory and sales processes, this time including
the trend terms. The final 500 observations from each
series is then chosen as the sample.15
IV. Estimation Technique
In estimating the model using the simulated data the
goal is to retrieve estimates of the population values
of the cost parameters. However, it may be noted
from Equation 9 that these are not fully identified by
the Euler equation, as any simultaneous scaling of a0,
a1, or a2 would have no effect on the Euler condition.
Because of this the process is reduced to estimating
combinations of parameters rather than the individ-
uals. This requires some sort of normalization in
order to estimate the Euler equation, which amounts
to choosing a left-hand-side variable in Equation 9
and solving for that variable. Choice of which
variable to solve for has often varied in previous
studies, and although it should not affect the Euler
equation identities, it does often result in different
estimates for the parameter values.16
However, the goal here is to construct a valid
model with which to analyse the effects of filtering
techniques, rather than to make any particular
observations on the empirical behaviour of inven-
tories. With this in mind, normalization issues are
ignored, and it is continued to follow West and
Wilcox (1996) in choosing to solve for current period
inventories. This is a convenient choice in that it
amounts to dividing Equation 10 through by the
15
A sample size of 500 is chosen because this is close to the number of observations available in Bureau of Economic Analysis
data for monthly inventories and sales at this point.
16
See Fuhrer et al. (1995) for a discussion of different normalizations for the linear-quadratic inventory model.
D. Doorn
Legendre–Clebsch condition of Equation 11. The
first-order condition can now be expressed as

Et Ht
X1tþ2
2 X2tþ1
3 Stþ1
ut ¼ 0 ð16Þ
where
X1tþ2 ¼
b2 Htþ2 þ ð2b2 þ 2bÞHtþ1
þ ð2b þ 2ÞHt
1
Ht
2

b2 Stþ2 þ ðb2 þ 2bÞStþ1

ð2b þ 1ÞSt þ St
1
X2tþ1 ¼ bHtþ1 þ Ht
1 þ bStþ1
St
a0
1 ¼
c
a1
2 ¼
c
ba2 a3
3 ¼
c
c ¼ a0 ð1 þ 4b þ b2 Þ þ a1 ð1 þ bÞ þ ba2
Given the values of the cost parameters above, the
implied population values of the regression coeffi-
cients are 1 ¼ 0.056, 2 ¼ 0.332, and 3 ¼ 0.013.
Replacing expectations with realized values in
Equation 16 and solving for current period inven-
tories gives a linear equation which can then be used
in estimation.
Ht ¼ 1 X1tþ2 þ 2 X2tþ1 þ 3 Stþ1 þ vtþ2 ð17Þ
where vt þ 2 includes the unobservable cost shock
along with expectational errors, or
vtþ2 ¼ ut
1 ðX1tþ2
Et X1tþ2 Þ
2 ðX2tþ1
Et X2tþ1 Þ

3 ðStþ1
Et Stþ1 Þ ð18Þ
Although Equation 17 is linear in the parameters,
the correlation structure of the regressors rules out
ordinary least squares estimation as inconsistent.
Instead, an instrumental variables procedure is used
to obtain consistent estimates of the regression
coefficients. The procedure is implemented using an
iterated version of the Generalized Method of
Moments (GMM) estimator of Hansen (1982),
which is equivalent to classical instrumental variables
estimation when the equation is linear in the
parameters. Instruments used are a constant and
two lags each of inventories and sales. Because vtþ2 in
Equation 18 is MA(2), the GMM weighting matrix
used in estimation is the inverse of the Newey and
Impact of Hodrick–Prescott filtering on parameter estimation 1871
West (1987) heteroscedasticity and autocorrelation 10
consistent covariance matrix. Additional details of
the estimation procedure are given in the technical 8
appendix, available upon request.
The above model is estimated five times for each of 6
the 1000 samples, using a different procedure for
handling the trend each time. The first method, 4
labelled LDT, just removes a linear trend from the
data by regressing the series on a constant and time 2
and retaining the residuals. The next three methods
use the HP filter for detrending, but with different 0
– 0.10 – 0.05 0.00 0.05 0.10 0.15 0.20 0.25
values for the smoothing parameter
, the choices of
which have been used or suggested in various studies Fig. 4. Density of b1 estimates: ( ) LDT, ( ) HP1,
( ) HP2, ( )HP3, ( ) NDT. The vertical line indicates
with monthly data. These will be denoted HP1, HP2, the population value of 0.056
and HP3. HP1 uses the standard value of
¼ 14 400
generally applied in the literature to monthly data.
HP2 uses a value of
¼ 86 400, used by Bils and
Kahn (2000) in their study of inventory behaviour.
HP3 uses a value of
¼ 129 600, suggested as optimal 3
for monthly data in Ravn and Uhlig (1997). The final
procedure, denoted NDT, does not detrend the data
prior to estimation, but instead estimates constant 2
and trend terms along with Equation 17 and adds
time to the instrument set.
1

0
V. Simulation Results – 0.3 – 0.1 0.1 0.3 0.5 0.7 0.9

Fig. 5. Density of b2 estimates: ( ) LDT, ( ) HP1,

To assess the impact that HP filtering may have on
( ) HP2, ( )HP3, ( ) NDT. The vertical line indicates
estimation and test statistics the distributional char- the population value of 0.332
acteristics of the parameter estimates are looked at
first resulting from the different detrending methods
when estimation is done using simulated data
calibrated to inventories and sales of total non-
30
durables. In this section the impact on another
parameter not directly estimated from the model,
but still of interest to researchers, is also considered.
20
An estimate of scaled marginal cost of holding
inventories can be constructed from estimates of 1
and 2 as
10
a2  
4 ¼ ¼ b
1 1
1 ð1 þ 4b þ b2 Þ
2 ð1 þ bÞ
c
0
From the calibrated values used in the simulation, – 0.1 0.0 0.1 0.2 0.3
4 ¼ 0.0018. In what follows results for this parameter
Fig. 6. Density of b3 estimates: ( ) LDT, ( ) HP1,
are reported as well.
( ) HP2, ( ) HP3, ( ) NDT. The vertical line indicates
Figures 4–7 plot the densities of the parameter the population value of 0.013
estimates for each detrending method used. Table 2
indicates distributional characteristics. From these it
is clear that not detrending the data prior to
estimation, but instead estimating deterministic detrended data. The use of linearly detrended data
terms along with the other parameters, results in results in fairly good performance in terms of bias as
estimates that contain less mean bias away from the well, which is to be expected given that the data was
population values than do the estimates using generated to contain a linear trend. Although the HP
1872 D. Doorn
Table 2. RMSE, mean, and interquartile range of parameter distributions

LDT HP1 HP2 HP3 NDT

RMSE 1 0.0469 0.0473 0.0459 0.0458 0.0459
2 0.1419 0.1462 0.1395 0.1386 0.1318
3 0.0152 0.0654 0.0395 0.0359 0.0149
4 0.0090 0.0352 0.0223 0.0203 0.0090
Mean 1 ¼ 0.0560 0.0594 0.0712 0.0675 0.0666 0.0570
[0.029, 0.092] [0.045, 0.101] [0.040, 0.098] [0.039, 0.098] [0.031, 0.098]
2 ¼ 0.3320 0.3213 0.2718 0.2891 0.2927 0.3285
[0.220, 0.413] [0.184, 0.346] [0.196, 0.369] [0.198, 0.374] [0.200, 0.407]
3 ¼ 0.0130 0.0163 0.0595 0.0392 0.0361 0.0167
[0.006, 0.023] [0.029, 0.083] [0.019, 0.055] [0.017, 0.051] [0.003, 0.024]
4 ¼ 0.0018 0.0042 0.0330 0.0206 0.0187 0.0039
[0.001, 0.009] [0.024, 0.042] [0.014, 0.027] [0.013, 0.025] [0.000, 0.009]

reasonably-sized confidence intervals for these

50
40
30
20
10
0
–0.005 –0.030 –0.005 0.020 0.045 0.070 0.095 0.120
Fig. 7. Density of b4 estimates: ( ) LDT, ( ) HP1,
( ) HP2, ( ) HP3, ( ) NDT. The vertical line indicates
the population value of 0.0018
filter also removes a linear trend, its further action on
the data induces a bias away from the population
values of the parameters. This is most pronounced
when
¼ 14 400, the most commonly used value
when applying the filter to monthly data. As
is
increased to 129 600, the bias shrinks toward that of
the linearly detrended estimator, but is still consider-
able, especially for 3 and 4.
The root mean squared error (RMSE), measuring
variability of the estimates about the population
parameters, indicates similar performance across
methods for 1 and 2. For 3 and 4 both NDT
and LDT result in significantly less volatility in the
estimates than does HP filtering.
While the point estimates made using data that is
HP filtered tend to be biased, it may be that most
parameters contain the population values. To check
this, the first and third quartiles of each distribution
are looked at, given in Table 2. These give the
parameter values between which 50% of the estimates
lie, and therefore can be seen as a type of 50%
confidence interval.17
For 1 and 2 all methods result in the population
values being within the interquartile range. However,
for 3 and 4 the population values lie outside of the
range for all of the estimates obtained from HP
filtered data, regardless of the value of
used in
filtering, although the range gets closer as
increases
and the mean bias diminishes.
In addition to bias in the mean, all of the
distributions tend to be skewed and have positive
kurtosis. For 1 and 2 these are only slight
deviations from the expected normal distribution
when detrending the data prior to estimation,
regardless of method. For both 3 and 4 the
deviation from normality is more pronounced for
all methods, with the positive kurtosis particularly
extreme for NDT and LDT.
Another area in which the detrending method
might have an impact is in performing inference
about the parameters of the model based on the
estimates. Table 3 indicates the rejection rates for two
test statisics related to the estimation. The nominal
rejection rates for both of these tests is 0.05. The first
test is the t-statistic testing the hypothesis that the
estimated parameter value is equal to the population
value. This is equivalent to testing whether the true
parameter value lies within a 95% confidence interval

17
This is due to West and Wilcox (1996), who devise such an interval in their simulation study of instrumental variables
estimators.
Impact of Hodrick–Prescott filtering on parameter estimation
Table 3. Simulation results
LDT HP1
t 1 0.1300 0.1407
2 0.1310 0.1711
3 0.0962 0.1022
J 0.0491 0.0445
Note: The t-statistic and J-test entries denote the rejection rate of the statistics, as described in the text.
Both have nominal rejection rates of 5%.
constructed about the parameter estimate, so com-
plements the above results using the interquartile
range. However, the t-statistics are constructed using
asymptotic approximations to the standard errors as
generated in the estimation procedure, rather than the
actual standard deviations of the 1000 estimates in
each case, and so these results are not likely to be
consistent with those above. From the graphs of
the distributions of the estimates, it is clear that
t-statistics constructed using the actual standard
deviations would reject far more frequently in most
cases. This implies that the approximated standard
errors are much larger than the actual standard
deviations of our estimates.
In all cases but one the t-test over-rejects, with the
exception being the test of 3 from NDT. The degree
of over-rejection is similar between LDT and the
three HP specifications for 1 and 2, with slight
improvement as
is increased. For 3, HP2 and
HP3 outperform LDT, approaching the nominal
rejection rate as
is increased. NDT outperforms
the other methods in all cases, with near equality to
the nominal rate for 3. The poor performance
of the t-test when using detrended data is likely due
in part to the poor small sample performance of
GMM estimation in general, which has been well
documented in the econometrics literature. To test
this some of the experiments were repeated using a
sample size of 10 000. This did improve the perfor-
mance in all cases.
The second test statistic is the commonly used
J-test of instrument-residual orthogonality, which is a
test of model specification.18 This test is distributed
2 with degrees of freedom q
p, where p is the
number of parameters being estimated and q the
number of instruments used. For LDT, HP1, HP2,
and HP3, q
p ¼ 2, and for NDT, q
p ¼ 1. For this
test, all of the methods result in actual size very close
to the desired nominal size. There is clearly little
impact on the size of this statistic when using HP
filtered data.
18
See Hansen (1982).
1873
HP2 HP3 NDT
0.1365 0.1339 0.0800
0.1466 0.1440 0.0800
0.0754 0.0669 0.0460
0.0468 0.0528 0.0580
VI. Conclusions
The simulation results above provide evidence that
the use of the HP filter as a general detrending device
is potentially dangerous to estimation results when
applied to trend stationary data subsequently used in
that estimation. In particular, point estimation was
significantly affected. For at least one of the directly
estimated parameters and also for 4, constructed
from the other two, use of the HP filter induced
substantial bias in the parameter estimates away from
their true values. While the bias tends to shrink as the
smoothing parameter of the filter increases, it is still
significant even at the highest value we tested. For the
lowest value,
¼ 14 400, the one most frequently used
in practice, the bias is severe in all of the parameter
estimates. Also increased variability in the parameter
estimates was found, as measured by RMSE, when
using HP filtered data, particularly for 3 and 4.
This also diminished with increases in the smoothing
parameter of the filter.
In addition, the 50% confidence intervals given
above only contain the population values within them
for two of the parameters when using HP filtered
data, regardless of
. For 3 and 4 these confidence
intervals were quite distant from the true values. This
is again due to the significant bias of these distribu-
tions to the right of the true parameter values even for
large values of
. However, this also indicates
inference based on estimation using HP filtered data
may be questionable. The other two methods of
dealing with the trend were clearly superior to HP
filtering, in that there was much less bias and also less
variability about the population values in most cases.
This is, of course, due to the fact that the data were
generated with simple linear trends, for which those
methods are well suited. An interesting next step,
then, would be to repeat the above study using data
simulated to contain a unit root or even possibly a
piecewise linear trend. The performance of estimators
making use of HP filtered data in these instances may
be very different from those in this study.
1874
There was little impact on the test statistics looked
at. It was found that the use of HP filtered data did
not result in the t-test performing significantly worse
than with linearly detrended data, particularly for the
higher values of
. However, the size properties of
this test were found to be significantly better than
either of these when the data was not detrended prior
to estimation. The size of the J-test of model
specification was virtually identical across detrending
methods. Despite this result, it may be interesting in
future research to consider the possible impact HP
filtering might have on the power of the J-test in
detecting misspecification.
The implication of these results is that research-
ers should take great care in choosing a detrending
method to apply to nonstationary data prior to
estimation of structural models. Clearly the ad hoc
use of the HP filter may adversely affect results
gained from estimating such models. It is noted,
however, that the HP filter performs much better
when applied to trend stationary data if the value
of the smoothing parameter is set at
¼ 129 600,
which was suggested as the ideal for monthly data
in Ravn and Uhlig (1997). This suggests that
making use of unit root testing procedures to
determine the stationarity properties of a set of
data may help in guiding the choice of
prior to
detrending. This also bears further investigation. In
regards to the current study, previous research
using the BEA data set used in calibrating the
simulation have tended to remove a linear trend
prior to using the data in estimation, although
stationarity testing has generated somewhat ambig-
uous results in some cases.19
References
Bils, M. and Kahn, J. A. (2000) What inventory behaviour
tells us about business cycles, The American Economic
Review, 90, 458–81.
Blanchard, O. J. (1983) The production and inventory
behaviour of the american automobile industry,
Journal of Political Economy, 91, 365–400.
Cogley, T. and Nason, J. M. (1995) Effects of the
Hodrick–Prescott filter on trend and difference
stationary time series: implications for business cycle
research, Journal of Economic Dynamics and Control,
19, 253–78.
19
See, for example, Eichenbaum (1989) and Doorn (2003).
D. Doorn
Doorn, D. J. (2003) Trend breaks in finished goods
inventories and sales of non-durables, Unpublished
Manuscript, North Carolina State University.
Ehlgen, J. (1998) Distortionary effects of the
optimal Hodrick–Prescott filter, Economics Letters,
61, 345–49.
Eichenbaum, M. (1989) Some empirical evidence on the
production level and production cost smoothing
models of inventory investment, The American
Economic Review, 79, 853–64.
Fuhrer, J., Moore, G. and Schuh, S. (1995) Estimating the
linear-quadratic inventory model: maximum likelihood
versus generalized method of moments, Journal of
Monetary Economics, 35, 115–57.
Gil-Alana, L. A. (2004) Seasonal fractional components in
macroeconomic time series, Applied Economics, 36,
1265–80.
Hansen, L. P. (1982) Large sample properties of generalized
method of moments estimators, Econometrica, 50,
1029–54.
Harvey, A.C. and Jaeger, A. (1993) Detrending, stylized
facts and the business cycle, Journal of Applied
Econometrics, 8, 231–47.
Hodrick, R. J. and Prescott, E. C. (1997) Post-war US
business cycles: an empirical investigation, Journal of
Money, Credit, and Banking, 29, 1–16.
King, R. G. and Rebelo, S. T. (1993) Low frequency
filtering and real business cycles, Journal of Economic
Dynamics and Control, 17, 207–31.
McCallum, B. T. (2000) Alternative monetary policy rules:
a comparison with historical settings for the united
states, the United Kingdom, and Japan, Economic
Quarterly of the Federal Reserve Bank of Richmond, 86,
49–79.
Nelson, C. R. and Kang, H. (1981) Spurious periodicity in
inappropriately detrended time series, Econometrica,
49, 741–51.
Nelson, C. R. and Plosser, C. (1982) Trends and random
walks in macroeconomic time series: some evidence
and implications, Journal of Monetary Economics, 10,
139–62.
Newey, W. K. and West, K. D. (1987) A simple
positive semi-definite, heteroskedasticity and autocor-
relation consistent covariance matrix, Econometrica,
55, 703–8.
Pedersen, T. M. (1999) Spectral analysis, business cycles,
and filtering of economic time series: a Survey,
Unpublished PhD dissertation, Institute of
Economics, University of Copenhagen.
Ramey, V. A. (1991) Nonconvex costs and the behaviour
of inventories, Journal of Political Economy, 99,
306–34.
Ravn, M. O. and Uhlig, H. (1997) On adjusting the hp-filter
for the frequency of observations, Unpublished
Manuscript, Center for Economic Research, Tilburg
University, The Netherlands, May.
Sen, A. (2004) Are us macroeconomic series difference
stationary or trend-break stationary?, Applied
Economics, 36, 2025–30.
 Impact of Hodrick–Prescott filtering on parameter estimation 1875
Sensier, M. (2003) Inventories and asymmetric business West, K. D. (1992) Inventory models, Unpublished
cycle fluctuations in the UK: a structural approach, Manuscript, University of Wisconsin.
Applied Economics, 35, 387–403. West, K. D. and Wilcox, D. W. (1996) A comparison of
Singleton, K. (1988) Econometric issues in the analysis of alternative instrumental variables estimators of a
equilibrium business cycle models, Journal of dynamic linear model, Journal of Business and
Monetary Economics, 21, 361–86. Economic Statistics, 14, 281–93.

View publication stats