EUROPEAN COMMISSION
Does capacity utilisation help estimating the TFP cycle?
Christophe Planas, Werner Roeger and Alessandro Rossi
Economic Papers 410| May 2010
EUROPEAN ECONOMY
Economic Papers are written by the Staff of the Directorate-General for Economic and Financial Affairs, or by experts working in association with them. The Papers are intended to increase awareness of the technical work being done by staff and to seek comments and suggestions for further analysis. The views expressed are the author’s alone and do not necessarily correspond to those of the European Commission. Comments and enquiries should be addressed to: European Commission Directorate-General for Economic and Financial Affairs Publications B-1049 Brussels Belgium E-mail: [email protected] This paper exists in English only and can be downloaded from the website ec.europa.eu/economy_finance/publications A great deal of additional information is available on the Internet. It can be accessed through the Europa server (ec.europa.eu) KC-AI-10-410-EN-N ISSN 1725-3187 ISBN 978-92-79-14896-5 doi 10.2765/41165 © European Union, 2010 Reproduction is authorised provided the source is acknowledged.
Does capacity utilization help estimating
the TFP cycle? ∗
C.Planas(a), W.Roeger(b) and A.Rossi(a)
(a) European Commission, Joint Research Centre
(b) European Commission, Economic and Financial Affairs
December 2009
Abstract
In the production function approach, accurate output gap assessment requiresa careful evaluation of the TFP cycle. In this paper we propose a bivariate modelthat links TFP to capacity utilization and we show that this model improves theTFP trend-cycle decomposition upon univariate and Hodrick-Prescott filtering. Inparticular, we show that estimates of the TFP cycle that load information aboutcapacity utilization are less revised than univariate and HP estimates, both with2009 and real-time TFP data vintages. We obtain this evidence for twelve pre-enlargement EU countries.
Keywords: Cobb-Douglas production function, Hodrick-Prescott filter, output gap,
revisions.
∗The views expressed in this paper are those of the authors and should not be attributed to theEuropean Commission.E-Mails: [email protected], [email protected], [email protected].
Contents
1 Introduction 3
2 A model for capacity utilization and TFP 4
3 Methodology for empirical validation 6
4 Country results 10
4.1 Belgium . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
4.2 Germany . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
4.3 Denmark . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
4.4 Greece . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
4.5 Spain . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
4.6 France . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
4.7 Ireland . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
4.8 Italy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
4.9 Luxembourg . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
4.10 Netherlands . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
4.11 Portugal . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
4.12 United Kingdom . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
5 Conclusion 72
6 Appendix 73
6.1 IG-priors for variance parameters . . . . . . . . . . . . . . . . . . . . . . 73
6.2 Correlations between CU series. . . . . . . . . . . . . . . . . . . . . . . . 74
7 References 75
2
1 Introduction
Output gap is the key variable of the cyclical adjustment of EU Member States budget
balance (see European Commission, 2005). Following a 2002 ECOFIN decision, the
European Commission (EC) measures output gap through a Cobb-Douglas production
function (see Denis et al., 2002) that relates the gap to the cyclical components of
labour and of total factor productivity (TFP). While the labour cycle is estimated using
unemployment and wage inflation in a Phillips curve relationship (see Denis et al., 2006),
so far the EC procedure extracts the TFP cycle with the Hodrick-Prescott filter (HP;
Hodrick and Prescott, 1997).
Output gap measures have been criticized for their real time performance; for Eu-
rope, see for example Runstler (2002), Planas and Rossi (2004), and Marcellino and
Musso (2008) who exploit a comprehensive real time data set for comparing several
methodologies as well as the various approaches adopted by international organisations.
Although overlooked in these studies, a most striking feature is that all methods failed
to identify a positive output gap in early 2000, towards the end of the IT boom. Output
gaps for this period have been substantially revised upward when information about
the 2002 economic downturn became available. Extending series with forecasts before
HP-detrending could not alleviate the problem, mainly because of the forecasts impre-
cision close to turning points. One possible strategy is to use economic indicators which
go along with the business cycle but are not revised. Capacity utilisation (CU) mea-
sures have been previously suggested in the literature (see e.g. Ruenstler, 2002, Proietti,
Musso and Wastermann, 2007, and European Commission, 2008, pp.94-105), but so far
no model-based justification have been given. Here we introduce CU within the pro-
duction function framework by explicitly allowing for variations in the use of the capital
stock. A strong correlation between CU and the cyclical component of TFP naturally
arises. As an alternative to HP detrending, we thus propose a model that links the
cyclical component of TFP to CU. We show that CU series do bring a gain in precision
to the TFP trend-cycle decomposition, and that they help overcoming the 2000 output
gap underestimation problem.
In Section 2 we discuss the link between TFP and CU in the Cobb-Douglas production
function framework. The bivariate system that we obtain has similarities with Kuttner’s
(1994) model for measuring potential output. For model estimation we resort to Bayesian
analysis. The Bayesian framework is convenient for imposing a strong prior about the
3
inertia of the productivity potential growth. It has also the advantage of eliminating
the pile-up effect, i.e. the occurence of 0-coefficient estimates for the unobserved shocks
variances (see Stock and Watson, 1988) that yields deterministic components. In real-
time, obtaining a trend that is sometimes deterministic and sometimes stochastic is
unacceptable because the decomposition results excessively instable over time.
To verify the relevance of CU to TFP cycle estimation, we compare the bivariate
estimates to those returned by a univariate decomposition and by the HP filter. The
comparison is made in terms of revisions in TFP cycle estimates recorded over the years
2000-2009 that cover two important boom bust episodes for which large revisions are
expected. If CU contains valuable information for TFP decomposition, its use should
limit the revisions in preliminary estimates. Details about the empirical methodology
are given in Section 3.
Section 4 reports results for each Member States. In order to give some actual rele-
vance to our investigation, we consider both 2009 data and real-time TFP vintages. For
CU, two types of series are used that mainly differ about the coverage of the service
sector. The exercise is carried out for twelve pre-enlargement countries, namely BE,
DK, DE, EL, ES, FR, IE, IT, LU, NL, PT, and UK. The other three pre-enlargement
countries AT, FI, and SW are left out for missing CU data. Overall we find that CU
has informative content for TFP trend-cycle decomposition in the twelve countries con-
sidered, and for both 2009 and real-time TFP vintages. The results are summarized in
Section 5.
2 A model for capacity utilization and TFP
According to the Cobb-Douglas production function, output Y is obtained from the
combination of capital stock K and labour L, both employed at the available total factor
productivity TFP:
Y = TFP K1−α Lα
The constant α represents the labour share of income. Because capital K cumulates
past investment at some depreciation rate, output gap only depends on labour gap and
on the TFP cycle, say C. These short-term fluctuations C are related to variations in
the capacity utilization of capital and labour inputs that we denote CUK and CUL,
respectively. TFP also contains persistent efficiency improvements P, so TFP = P ×C.
4
Writing the production function as:
Y = P (CUK × K)1−α (CUL × L)α
suggests that the link between TFP gap and capacity utilization is such that:
C = CU1−αK CUα
L
No capacity utilization measure however discriminates between the different factors.
Only aggregate capacity utilization series are available. They are usually built from
surveys, so by construction we expect CU and CUK to be significantly correlated. Given
that average hours worked per employee already contains some cyclical movements, the
link with labour utilization should be somewhat looser. But if there are fluctuations
in the degree of labour hoarding that are not captured by hours, a correlation between
labour and capital utilization should nevertheless be present. We thus assume:
cuL = γcuK + ε 0 < γ < 1
where small letters denote logarithms and ε is a random shock which can be itself au-
tocorrelated in case of movements in cuL that are not exactly synchronised with cuK .
Hence TFP is related to capacity utilization through:
tfp = p+ (1− α + αγ)cu+ αε
This link can be exploited for estimating the TFP trend in a bivariate model such as:
tfpt = pt + ct
cut = μcu + βct + ecut β = (1− α + αγ)−1 (2.1)
where the sub-index t = 1, · · · , T introduces time. The cyclical component ct is a
stationary factor that is common to both TFP and CU series. Given standard values
for the output elasticity of labour α and plausible values for γ, the loading coefficient β
should be greater than one. The dynamic behaviour of the unobserved components pt
and ct remains to specify. We consider:
Δpt = μt−1
μt = w(1− ρ) + ρ μt−1 + aμt V (aμt) = Vμ
ct = 2A cos(2π/τ) ct−1 − A2 ct−2 + act V (act) = Vc (2.2)
5
where aμt and act are white noises. Equation (2.2) describes the TFP long-term path
through a damped trend model with a coefficient w that catches the series average growth
rate. The cyclical movements are reproduced using an AR(2) model with complex roots
that are parameterized in terms of amplitude A and periodicity τ . For the stochastic
term ecut, we will consider either:
ecut = acut or ecut = δ ecut−1 + acut V (acut) = VCU (2.3)
where acut is a white noise. The insertion of the autoregressive lag will depend on the
statistical properties of the CU series. The bivariate system (2.1)-(2.3) is similar to the
Phillips-curve augmented unobserved component model proposed by Kuttner (1994) for
estimating potential output and output gap in the US.
3 Methodology for empirical validation
Model (2.1)-(2.3) describes a possible link between the cyclical movements of TFP and
CU series. If such a link exists, making use of CU information should yield a gain in
accuracy in TFP cycle estimates. We check this conjecture for twelve pre-enlargement EU
Member States. Three estimation methods are considered: HP filtering, the univariate
trend plus cycle model (2.2) and the bivariate system (2.1)-(2.3). The estimators are
compared in terms of revisions recorded in TFP cycle latest estimates, both with 2009
data sets and with real-time data vintages.
The data are annual series for BE, DK, DE, EL, ES, FR, IE, IT, LU, NL, PT, and UK,
all taken from AMECO database. The other three pre-enlargement countries AT, FI, and
SW are left out for data unavailability. The TFP time span covers 1965-2009 with ten
vintages available over the period 2000-2009. To capture cyclical fluctuations in capacity
utilization, we use two different indicators: the Capacity Utilization Indicator (CUI)
which is available for manufacturing only, and the EC Business Survey indicator (BS)
that is for both manufacturing and services. CUI has the advantage that it is available
since 1985 for most countries and since 1987 for few ones. Also BS is available for all
countries but surveys for services only start in the years 1995-1998, at the exception
of FR for which the starting date is 1988. The missing years before 1995 have been
completed by merging with CUI after proper re-scaling, so BS and CUI are identical
until the actual start of BS for services. A third CU measure called the Purchasing
Managers Indicator (PMI) exists for some countries only. Because exhaustive country
6
coverage is essential for any practical application, PMI have been excluded from this
exercise. For information, Table A2 in Appendix gives the cross-correlations between
CUI, BS, and PMI when available.
The exercise is performed using Bayesian techniques. Maximum likelihood estimation
is of course feasible and less computationally intensive, but in recursive analysis occa-
sional occurrences of 0-coefficient estimate for the unobserved component shock variances
cause some instability in the trend-cycle decomposition. In the Bayesian framework, this
pattern can be excluded by specifying an informative prior. Another advantage of the
Bayesian approach is that the information brought by macroeconomic knowledge can be
inserted into the analysis. In our context, we have a strong prior about the inertia of
the potential growth of productivity. Our model implies a β-coefficient in (2.1) that is
greater than one. And we also have some knowledge about the periodicity and amplitude
of the business cycle.
All computations are made using Program Bayesian GAP downloadable at eemc.jrc.ec
.europa.eu. Details about the procedures implemented can be read in Planas, Rossi and
Fiorentini (2008). For the parameters in (2.1)-(2.3), we consider the following priors:
• Cycle amplitude A Beta-distributed with mean 0.4 and standard deviation 0.2;
• Cycle periodicity τ Beta-distributed with mean of 8 and standard deviation 3.5;
• Average growth w normally distributed with mean 0.015 and standard deviation
0.005; for ES, mean at 0.003 and standard deviation 0.002. The average growth w
is always constrained to be positive.
• ρ normally distributed with mean 0.8 and standard deviation 0.3 restricted to the
stationary (0, 1) region;
• β given VCU normally distributed with mean of 1.4 and standard deviation 0.3 ×VCU ;
• for BE, FR, NL and UK, δ given VCU normally distributed with mean 0.5 and
standard deviation 0.5VCU with a restriction to the stationary region (0, 1);
• Inverted-gamma (IG) prior distributions are used for all variance parameters. As
a tuning by country has been necessary, we report in Table A1 of the Appendix
the hyper-parameters of the prior distribution of Vc, Vμ, and VCU for each country.
7
The estimators are compared in terms of revisions in TFP cycle estimates. Let xt
denote the set of observations available at time t, i.e. xt = (x1, · · · , xt). For univariate
analysis, xt represents TFPt while in bivariate xt contains both TFPt and CUt. The
cycle estimates for period t based on observations until period t + k is obtained as the
expectation of ct given observations xt+k: i.e. ct|t+k = E(ct|xt+k). Hence the cycle
estimates for a given point in time depend on the information available. A revision can
be defined as the correction of preliminary estimates due to incoming observations. For
instance, the difference ct|t+1 minus ct|t measures the revision in the concurrent estimate
ct|t due to the availability of one further observation. For each country, we show the
path taken by cycle estimates for the years 2000 to 2008 when observations are ending
in 2000, 2001, · · · , until 2009.Revisions in real-time TFP gap estimates come from two different sources: forecast
errors and parameter update - the signal extraction error or statistical uncertainty, and
the use of real-time data sets that are corrected every year, i.e. data vintages. In order
to shed light on the relative contribution of these two sources of revisions, we report
results obtained first using the 2009 data vintage and then using real-time data sets.
We summarize these revisions by computing their variances. For instance, averaging the
squared values of the revisions obtained with one more observation from t = 2001 to
t = 2009 approximates the variance of the first revision in concurrent estimates, i.e:
V (ct|t+1 − ct|t) � (1/9)2008∑
t=2000
(ct|t+1 − ct|t)2
The same computations can be done for evaluating empirically the variance of the second
revision in concurrent estimates, i.e. V (ct|t+2 − ct|t+1). As revisions are independent, we
can cumulate them to obtain the variance of the revisions with k more observations,
V (ct|t+k − ct|t). We obtain the variance of revisions when one more observation is avail-
able, when two more observations are available, and so on. For each country we report√V (ct|t+k − ct|t) for k = 1 to 4. This 1 to 4-period-ahead revision variance can be used
to build a confidence interval around concurrent estimates for the estimates that will be
obtained with k-more observations like for instance ct|t + / − 2√
V ar(revision). More
details can be found in Planas and Rossi (2004). The theoretical analysis of revisions
has been developed by Pierce (1980).
The model parameters are re-estimated every time the data set is updated. Notice
that when the data previously observed are not updated, revisions on the trend and on
8
the cycle sum to zero so they are equivalent in absolute value. This equivalence however
breaks down when past data are revised. Because most TFP vintages show level shifts,
real-time trend estimates do not converge over the different vintages. Here we focus
on the behaviour of revisions in the cycle estimates obtained with 2009 series and with
real-time vintages.
For each country, we report:
1. Fig.1: the TFP 2000-2009 vintages plus CU series;
2. Fig.2: the cycle and the trend growth estimated with the 2009 vintage;
3. Fig.3: for the 2009 data vintage, the prior and posterior distributions for a selection
of parameters.
4. Table 1: the bivariate model fitted using the 2009 TFP vintage and the two CU
series;
5. Fig.4: paths followed by the 2000-2008 cycle estimates over the years 2000-2009
with vintage 2009;
6. Fig.5: revisions standard deviation with up to 4 years of additional data, vintage
2009;
7. Fig.6: paths followed by the 2000-2008 cycle estimates over the years 2000-2009
with real-time vintages;
8. Fig.7: revisions standard deviation with up to 4 years of additional data, real-time
vintages.
For all figures, HP is in black, the univariate model is in red, the bivariate one with
CUI series is in blue, and the bivariate one with BS series is in green. A detailed
explanation of the figures is given for the BE case in pages 10-15.
For HP, the filter is run on series extended with four forecasts. The forecasting models
are I(1) for DK, IE, PT, BE, EL, IT, LU, UK; ARIMA(1,1,0) for DE; and ARIMA(0,1,1)
for ES, FR, and NL. A constant drift is always included. The inverse signal to noise ratio
is set equal to 100. Holding this ratio constant for all vintages gives a slight advantage
to HP.
9
4 Country results
4.1 Belgium
Figure 1
TFP vintages plus CU series
0.55
0.6
0.65
0.7
0.75
0.8
85 88 91 94 97 00 03 06 09
−0.06
−0.04
−0.02
0
0.02
0.04
The upper plot shows the different vintages of the TFP series: the 2000 vintage ends
in 2000, the 2001 one in 2001 and so on. Time labels are visible on the lower plot.
Two capacity utilization series are displayed: the continuous line represents the capacity
utilization indicator CUI and the dotted line is the EC business survey indicator BS. Both
series are displayed after a mean removal. The CUI and BS series are used alternatively.
10
Figure 2 below shows the posterior mean of the TFP trend growth for the years 1985-
2009 together with the TFP series growth in dots and the series cycle. The HP estimate
is in black, the univariate one is in red, the bivariate one with CUI series is in blue,
and the bivariate estimate with BS series is in green.
Figure 2 BE Vintage 2009
Trend growth and cycle (×100)
−1
0
1
2
3
Trend growth
85 88 91 94 97 00 03 06 09
−2
−1
0
1
2
Cycle
As can be seen, the trend growth is quite smooth. The cycle estimates obtained with the
three estimation methods differ mostly in the second sample half: loading CU information
increases the TFP gap for the years 1997-2009.
11
For the 2009 vintage, Figure 3 below shows the prior distribution (−−) together withthe posterior distributions obtained with CUI series (in blue) and with the BS series (in
green) for a selection of parameters.
Figure 3 BE Vintage 2009
Prior and posterior distributions
5 10 15 20 250
0.02
0.04
0.06
0.08
0.1
0.12
τ
0.005 0.01 0.015 0.020
50
100
150
ω
0.4 0.6 0.80
1
2
3
4
5
6
7
ρ
0.5 1 1.5 20
0.2
0.4
0.6
0.8
1
1.2
β
0 1 2x 10−4
0
2000
4000
6000
8000
10000
Vc
0 0.5 1x 10−5
0
1
2
3
4
x 105 Vμ
As can be seen, the data contain information about all parameters but Vμ. In particular,
the β coefficient that relates the TFP cycle to capacity utilization is sharply estimated,
with mode value above one as expected. The prior on Vμ has strongly imposed the view
that the TFP growth should evolve quite slowly, so the posterior could not depart from
this hypothesis.
12
Table 1 below summarizes the parameter posterior distributions obtained with the
2009 vintage in terms of modes and standard deviations. As can be seen, the posterior
modes are stable with respect to the use of CUI or BS series.
Table 1 BE Full sample estimation, 2009 vintage
Posterior modes and standard deviations
TFP: Δpt = μt (1− ρL)μt = (1− ρ)w + aμt (1− 2Acos(2π/τ)L+ A2L2)ct = act
CU: CUt = μCU + βct + aCUt/(1− δL)
w ρ Vμ A τ Vc μCU δ β VCU
CUI0.011 0.96 1.8×10−6 0.33 5.56 10.8×10−5 -0.001 0.4 1.28 40.5×10−5
( 0.003) ( 0.09 ) ( 0.13 ) ( 3.3 ) ( 0.01 ) ( 0.23 ) ( 0.37 )
BS0.011 0.97 1.7×10−6 0.34 5.75 10.8×10−5 -0.001 0.55 1.28 32.7×10−5
( 0.003) ( 0.09 ) ( 0.13 ) ( 3.26 ) ( 0.01 ) ( 0.2 ) ( 0.33 )
Figures 4 and 6 in the next pages show the behavior of the cycle estimate for the periods
2000, 2002, ..., 2008 obtained assuming that the data are ending in 2000, ..., until 2009.
Vintage 2009 means that past TFP data are assumed to not be updated. Figure 6 is
like Figure 4 but using real-time data. The x-axis displayed in the graph bottom line
refers to the last point of the dataset used. The first estimates displayed is always the
concurrent one, i.e. ct|t. For instance, in Figure 4 the first small plot in the upper
right corner shows the path taken by the estimate of the cycle for the year 2000 using
datasets ending successively in 2000, in 2001, and so on until 2009: i.e. c2000|2000+k for
k = 0, 1, · · · , 8.
13
Figure 4 BE Vintage 2009
Paths followed by the 2000-2008 cycle estimates
−0.474
0.907
2.287
−1.831
−1.097
−0.363
−1.714
−1.016
−0.318
−2.407
−1.521
−0.635
−0.598
0.614
1.825
−1.283
−0.311
0.66
00 01 02 03 04 05 06 07 08 09−1.267
−0.225
0.818
00 01 02 03 04 05 06 07 08 09−1.124
0.082
1.288
00 01 02 03 04 05 06 07 08 09−1.934
−1.11
−0.286
Figure 5 BE Vintage 2009
Revisions standard deviation (×100)
1 2 3 40.5
0.6
0.7
0.8
0.9
1
1.1
1.2
1.3
Additional observations
14
Figure 6 BE Real-time vintages
Paths followed by the 2000-2008 cycle estimates
−0.507
0.89
2.287
−1.164
−0.562
0.039
−1.937
−1.128
−0.318
−1.898
−1.266
−0.635
−0.695
0.565
1.825
−0.941
−0.141
0.66
00 01 02 03 04 05 06 07 08 09−0.882
−0.032
0.818
00 01 02 03 04 05 06 07 08 09−1.077
0.106
1.288
00 01 02 03 04 05 06 07 08 09−1.626
−0.956
−0.286
Figure 7 BE Real-time vintages
Revisions standard deviation (×100)
1 2 3 40.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
1.1
1.2
Additional observations
15
Figures 5 and 7 show the average squared first-four revisions due to the the incoming
of new observations from 2001 to 2009. These averages estimate the variance of the
revisions with k more observations, V (ct|t+k − ct|t) - see Section 3. The lower the revision
variance, the more reliable are the TFP cycle estimates. As can be seen, the estimates
that load CU data are performing better both in real-time and with the 2009 vintage.
Summary for BE:
• TFP data The TFP series level shifts downward after 2000 for all post-2000
vintages.
CU data The two series are quite similar.
Link TFP-CU The β-coefficient is significantly different from 0 and above than
1 as expected.
Revisions The bivariate model yields less revisions than univariate and HP de-
compositions both with 2009 data and real-time vintages, with both CUI and BS
series.
CUI vs. BS The BS series yields slightly less revisions than CUI.
As Figure 4 and 6 show, the HP filter fails to capture the cyclical nature of the strong
TFP growth at the end of the 90s and indicates a negative TFP gap in 2000. The
bivariate estimates give a positive TFP gap because of cyclical indicators pointing to
above average capacity utilisation in 2000. The revisions are not confined to the year
2000 but the HP filter is revised heavily in the direction of the bivariate estimates in
subsequent years. A similar phenomenom seems to be taking place for the years 2007-
2008 where the HP preliminary estimates are heavily revised upward.
For the other countries, Figures 1-7 and Table 1 are reported together with a short
comment and the summary.
16
4.2 Germany
Figure 1
TFP vintages plus CU series
0.45
0.5
0.55
0.6
0.65
0.7
0.75
85 88 91 94 97 00 03 06 09
−0.1
−0.05
0
0.05
Figure 2 DE Vintage 2009 (×100)Trend growth and cycle
−3
−2
−1
0
1
2
3
Trend growth
85 88 91 94 97 00 03 06 09
−3
−2
−1
0
1
2
Cycle
17
Figure 3 DE Vintage 2009
Prior and posterior distributions
5 10 15 20 250
0.05
0.1
0.15
τ
0.005 0.01 0.015 0.020
50
100
150
ω
0.5 0.6 0.7 0.8 0.90
1
2
3
4
5
6
7
ρ
0.5 1 1.5 2 2.50
0.2
0.4
0.6
0.8
1
β
0 1 2x 10−4
0
5000
10000
15000
Vc
0 1 2x 10−5
0
1
2
3
4
x 105 Vμ
Table 1 DE Full sample estimation, 2009 vintage
Posterior modes and standard deviations
TFP: Δpt = μt (1− ρL)μt = (1− ρ)w + aμt (1− 2Acos(2π/τ)L+ A2L2)ct = act
CU: CUt = μCU + βct + aCUt
w ρ Vμ A τ Vc μCU β VCU
CUI0.013 0.96 2.9×10−6 0.62 7.41 8.2×10−5 -0.001 1.66 63.1×10−5
( 0.003) ( 0.08 ) ( 0.14 ) ( 2.87 ) ( 0.01 ) ( 0.38 )
BS0.013 0.96 2.3×10−6 0.65 8.97 8.9×10−5 -0.003 1.21 74.9×10−5
( 0.003) ( 0.1 ) ( 0.14 ) ( 3.25 ) ( 0.01 ) ( 0.42 )
18
Figure 4 DE Vintage 2009
Paths followed by the 2000-2008 cycle estimates
−0.936
0.043
1.022
−0.74
0.147
1.034
−1.198
−0.266
0.665
−1.802
−0.827
0.148
−2.036
−1.092
−0.148
−2.261
−1.144
−0.026
00 01 02 03 04 05 06 07 08 09−1.057
0.347
1.752
00 01 02 03 04 05 06 07 08 09−0.562
0.785
2.131
00 01 02 03 04 05 06 07 08 09−1.135
0.276
1.687
Figure 5 DE Vintage 2009
Revisions standard deviation (×100)
1 2 3 40.4
0.6
0.8
1
1.2
1.4
1.6
1.8
Additional observations
19
Figure 6 DE Real-time vintages
Paths followed by the 2000-2008 cycle estimates
−1.02
0.001
1.022
−0.928
0.053
1.034
−1.263
−0.299
0.665
−1.704
−0.778
0.148
−1.727
−0.937
−0.148
−1.983
−1.005
−0.026
00 01 02 03 04 05 06 07 08 09−1.382
0.185
1.752
00 01 02 03 04 05 06 07 08 09−0.664
0.733
2.131
00 01 02 03 04 05 06 07 08 09−1.102
0.292
1.687
Figure 7 DE Real-time vintages
Revisions standard deviation (×100)
1 2 3 40.2
0.4
0.6
0.8
1
1.2
1.4
1.6
Additional observations
20
Summary for DE:
• TFP data The TFP vintages are systematically shifted upward.
CU data The CUI and BS series are similar. The largest difference occurs in
2008, CUI standing at one-percentage point higher than BS. Both series have a dip
in 2009.
Link TFP-CU The β-coefficient is significantly different from 0, with posterior
mode above one as expected. It takes larger values with the CUI series.
Revisions The bivariate model yields less revisions than univariate and HP de-
compositions both with 2009 data and real-time vintages, and with both CUI and
BS series.
CUI vs. BS The posterior distributions of model parameters seem robust to the
use of CUI vs. BS. Less TFP gap revisions are obtained with the CUI series.
The cyclical information for the year 2000, indicating above average capacity utilisation,
avoids a strong negative TFP gap for 2000 in the bivariate case. In contrast, the HP filter
fails to capture the cyclical nature of high TFP growth resulting in a strongly negative
TFP gap. In 2001-2006, the bivariate estimates did not outperform HP in terms of
revisions. In 2007-2008, HP goes through large positive revisions with a sign switch.
The bivariate estimates that use CUI are the most stable for these years.
21
4.3 Denmark
Figure 1
TFP vintages plus CU series
0.45
0.5
0.55
0.6
0.65
0.7
0.75
88 91 94 97 00 03 06 09
−0.06
−0.04
−0.02
0
0.02
0.04
0.06
Figure 2 DK Vintage 2009
Trend growth and cycle (×100)
−2−1
012345
Trend growth
88 91 94 97 00 03 06 09
−3
−2
−1
0
1
2
3
Cycle
22
Figure 3 DK Vintage 2009
Prior and posterior distributions, 2009 vintage
5 10 15 200
0.05
0.1
0.15
τ
0.005 0.01 0.015 0.020
20
40
60
80
100
120ω
0.4 0.6 0.80
0.5
1
1.5
2
2.5
3
ρ
0.5 1 1.5 2 2.50
0.5
1
1.5
β
0 1 2 3x 10−4
0
2000
4000
6000
8000
10000
Vc
0 0.5 1 1.5x 10−4
0
1
2
3
4
5
x 104 Vμ
Table 1 DK Full sample estimation, 2009 vintage
Posterior modes and standard deviations
TFP: Δpt = μt (1− ρL)μt = (1− ρ)w + aμt (1− 2Acos(2π/τ)L+ A2L2)ct = act
CU: CUt = μCU + βct + aCUt/(1− δL)
w ρ Vμ A τ Vc μCU δ β VCU
CUI0.012 0.88 19.3×10−6 0.53 8.71 11.1×10−5 0 0 1.29 23.6×10−5
( 0.004) ( 0.13 ) ( 0.13 ) ( 2.84 ) ( 0.01 ) ( 0 ) ( 0.4 )
BS0.011 0.85 18×10−6 0.56 8.98 12.4×10−5 -0.001 0 1.19 9.5×10−5
( 0.004) ( 0.12 ) ( 0.12 ) ( 2.81 ) ( 0.01 ) ( 0 ) ( 0.26 )
23
Figure 4 DK Vintage 2009
Paths followed by the 2000-2008 cycle estimates
−0.939
0.441
1.822
−1.811
−0.736
0.34
−2.01
−1.053
−0.096
−2.025
−0.961
0.103
−1.136
0.226
1.589
−0.965
0.652
2.268
00 01 02 03 04 05 06 07 08 09−0.966
1.026
3.018
00 01 02 03 04 05 06 07 08 09−2.027
0.353
2.734
00 01 02 03 04 05 06 07 08 09−3.311
−1.137
1.037
Figure 5 DK Vintage 2009
Revisions standard deviation (×100)
1 2 3 40.4
0.6
0.8
1
1.2
1.4
1.6
1.8
Additional observations
24
Figure 6 DK Real-time vintages
Paths followed by the 2000-2008 cycle estimates
−1.341
0.241
1.822
−1.178
−0.419
0.34
−1.595
−0.585
0.425
−1.812
−0.854
0.103
−0.981
0.304
1.589
−0.844
0.712
2.268
00 01 02 03 04 05 06 07 08 09−0.886
1.036
2.957
00 01 02 03 04 05 06 07 08 09−1.642
0.222
2.086
00 01 02 03 04 05 06 07 08 09−1.939
−0.32
1.3
Figure 7 DK Real-time vintages
Revisions standard deviation (×100)
1 2 3 40.2
0.4
0.6
0.8
1
1.2
1.4
1.6
Additional observations
25
Summary for DK:
• TFP data There is large level shift in the TFP vintages available after 2005.
CU data There is a large positive outlier in the CUI series in the year 2007. Both
series record a large dip in 2009.
Link TFP-CU The β-coefficient is significantly different from 0. The posterior
mode is above one as expected and slightly larger with CUI than with BS.
Revisions The bivariate model yields less revisions than univariate and HP de-
compositions both with the 2009 data and the real-time vintages, with both CUI
and BS series.
CUI vs. BS Less revisions are obtained with the BS series.
In 2000, the HP estimate miss the information about capacity utilization above average
and returns a negative cycle. In contrast, the bivariate measures point to a positive gap,
with agreement with the CU series. The HP preliminary estimate undergoes mostly
positive revisions that bring it close to the bivariate estimates. Again the HP estimates
for the years 2006-2007 are heavily revised.
26
4.4 Greece
Figure 1
TFP vintages plus CU
0.4
0.45
0.5
0.55
0.6
0.65
0.7
0.75
85 88 91 94 97 00 03 06 09−0.05
−0.04
−0.03
−0.02
−0.01
0
0.01
0.02
Figure 2 EL Vintage 2009
Trend growth and cycle (×100)
−3
−2
−1
0
1
2
3
4
Trend growth
85 88 91 94 97 00 03 06 09−3
−2
−1
0
1
2
3
Cycle
27
Figure 3 EL Vintage 2009
Prior and posterior distributions, 2009 vintage
5 10 15 20 250
0.02
0.04
0.06
0.08
0.1
0.12
τ
0.005 0.01 0.015 0.020
20
40
60
80
100
ω
0.4 0.6 0.80
0.5
1
1.5
2
2.5
3
3.5
ρ
0.5 1 1.50
0.5
1
1.5
β
0 2 4x 10−4
0
1000
2000
3000
4000
5000
6000
7000
Vc
0 1 2x 10−4
0
1
2
3
4
5
x 104 Vμ
Table 1 EL Full sample estimation, 2009 vintage
Posterior modes and standard deviations
TFP: Δpt = μt (1− ρL)μt = (1− ρ)w + aμt (1− 2Acos(2π/τ)L+ A2L2)ct = act
CU: CUt = μCU + βct + aCUt
w ρ Vμ A τ Vc μCU β VCU
CUI0.014 0.85 28.9×10−6 0.38 6.15 14.4×10−5 0 0.61 23.1×10−5
( 0.004) ( 0.11 ) ( 0.13 ) ( 3.33 ) ( 0 ) ( 0.31 )
BS0.013 0.86 34.3×10−6 0.38 6.16 14.7×10−5 0 0.69 16.7×10−5
( 0.004) ( 0.11 ) ( 0.13 ) ( 3.23 ) ( 0 ) ( 0.27 )
28
Figure 4 EL Vintage 2009
Paths followed by the 2000-2008 cycle estimates
0.615
1.424
2.233
0.561
1.531
2.502
−1.27
0.071
1.411
−0.835
0.648
2.131
−0.307
1.663
3.634
−1.884
0.02
1.924
00 01 02 03 04 05 06 07 08 09−1.79
−0.83
0.129
00 01 02 03 04 05 06 07 08 09−0.384
0.364
1.111
00 01 02 03 04 05 06 07 08 09−0.998
−0.29
0.418
Figure 5 EL Vintage 2009
Revisions standard deviation (×100)
1 2 3 40.6
0.65
0.7
0.75
0.8
0.85
0.9
0.95
1
1.05
Additional observations
29
Figure 6 EL Real-time vintages
Paths followed by the 2000-2008 cycle estimates
−1.035
0.1
1.235
0
0.911
1.822
−1.041
0.612
2.266
−0.225
0.953
2.131
0.025
1.829
3.634
−1.105
0.444
1.993
00 01 02 03 04 05 06 07 08 09−1.176
−0.265
0.647
00 01 02 03 04 05 06 07 08 09−0.723
0.194
1.111
00 01 02 03 04 05 06 07 08 09−0.757
−0.17
0.418
Figure 7 EL Real-time vintages
Revisions standard deviation (×100)
1 2 3 40.4
0.5
0.6
0.7
0.8
0.9
1
1.1
1.2
1.3
1.4
Additional observations
30
Summary for EL:
• TFP data The 2008 vintage has a large positive level shift. The 2009 vintage is
close to the 2008 series but has a greater growth during 1997-2000.
CU data Both CU series show a large dip in 2009. There is another dip in the
CUI data for 2005.
Link TFP-CU The link is slightly more pronounced with the BS series. With
both CUI and BS series the β-posterior mode is below one.
Revisions The bivariate estimates that uses the BS series dominates with both
the 2009 data and the real-time vintages.
CUI vs. BS: less revisions are obtained with the BS series.
For the year 2000 all methods have a similar revision history. For the recent boom bust
cycle, especially for the years 2007-2008 the HP TFP gap is more heavily revised.
31
4.5 Spain
Figure 1
TFP vintages plus CU series
0.56
0.58
0.6
0.62
0.64
0.66
0.68
0.7
88 91 94 97 00 03 06 09−0.08
−0.06
−0.04
−0.02
0
0.02
Figure 2 ES Vintage 2009
Trend growth and cycle (×100)
−0.5
0
0.5
1
1.5
Trend growth
88 91 94 97 00 03 06 09−1.5
−1
−0.5
0
0.5
1
1.5
Cycle
32
Figure 3 ES Vintage 2009
Prior and posterior distributions, 2009 vintage
5 10 15 200
0.02
0.04
0.06
0.08
0.1
0.12
τ
2 4 6x 10−3
0
50
100
150
200
250
300
350
ω
0.2 0.4 0.6 0.80
0.5
1
1.5
2
2.5
3
3.5
ρ
1 2 30
0.2
0.4
0.6
0.8
β
0 0.5 1x 10−4
0
0.5
1
1.5
2
2.5
3
x 104 Vc
0 2 4 6x 10−6
0
2
4
6
8
x 105 Vμ
Table 1 ES Full sample estimation, 2009 vintage
Posterior modes and standard deviations
TFP: Δpt = μt (1− ρL)μt = (1− ρ)w + aμt (1− 2Acos(2π/τ)L+ A2L2)ct = act
CU: CUt = μCU + βct + aCUt
w ρ Vμ A τ Vc μCU β VCU
CUI0.004 0.9 0.8×10−6 0.45 8.29 4.4×10−5 0 2.06 41.6×10−5
( 0.001) ( 0.13 ) ( 0.14 ) ( 3.14 ) ( 0.01 ) ( 0.72 )
BS0.004 0.87 1×10−6 0.52 9.43 4.9×10−5 -0.001 2.4 21.9×10−5
( 0.001) ( 0.13 ) ( 0.13 ) ( 2.94 ) ( 0.01 ) ( 0.47 )
33
Figure 4 ES Vintage 2009
Paths followed by the 2000-2008 cycle estimates
−1.764
−0.44
0.883
−1.735
−0.442
0.85
−1.834
−0.685
0.463
−1.781
−0.738
0.306
−1.747
−0.787
0.174
−1.643
−0.678
0.287
00 01 02 03 04 05 06 07 08 09−1.525
−0.547
0.43
00 01 02 03 04 05 06 07 08 09−1.253
−0.154
0.946
00 01 02 03 04 05 06 07 08 09−1.547
−0.644
0.259
Figure 5 ES Vintage 2009
Revisions standard deviation (×100)
1 2 3 40.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
Additional observations
34
Figure 6 ES Real-time vintages
Paths followed by the 2000-2008 cycle estimates
−0.962
0.229
1.42
−1.321
−0.074
1.173
−1.39
−0.17
1.051
−0.446
0.522
1.491
−1.166
−0.31
0.546
−1.903
−0.734
0.435
00 01 02 03 04 05 06 07 08 09−1.695
−0.566
0.563
00 01 02 03 04 05 06 07 08 09−1.509
−0.237
1.035
00 01 02 03 04 05 06 07 08 09−1.606
−0.702
0.203
Figure 7 ES Real-time vintages
Revisions standard deviation (×100)
1 2 3 40.5
0.6
0.7
0.8
0.9
1
1.1
1.2
1.3
Additional observations
35
Summary for ES:
• TFP data The post-1995 series growth seems to have almost-always been revised
downward.
CU data Both CU series show a large dip in 2009.
Link TFP-CU The β-coefficient is significantly different from 0. It takes larger
values when BS is used instead of CUI, and the posterior distribution is more
concentrated around the mode.
Revisions The bivariate model with BS data yields the less revisions with both
2009 data and real-time vintages.
CUI vs. BS Less revisions are obtained with the BS series.
There is the well known tendency for the HP gap to be revised upward over time, i.e. the
HP trend seems to be excessively optimistic in real time. With an initial HP estimate
for 2000 at -0.3, subsquently revised positively towards the bivariate estimates, the 2000
phenomenom appears also for ES. The bivariate estimates that use the BS series are
remarkably robust in the years 2006-2009. Again, HP is heavily revised in these years.
36
4.6 France
Figure 1
TFP vintages plus CU series
0.5
0.55
0.6
0.65
0.7
0.75
0.8
85 88 91 94 97 00 03 06 09
−0.1
−0.08
−0.06
−0.04
−0.02
0
0.02
0.04
Figure 2 FR Vintage 2009 (×100)Trend growth and cycle
−1
0
1
2
Trend growth
85 88 91 94 97 00 03 06 09
−4
−3
−2
−1
0
1
2
Cycle
37
Figure 3 FR Vintage 2009
Prior and posterior distributions, 2009 vintage
5 10 15 20 250
0.02
0.04
0.06
0.08
0.1
0.12
0.14τ
0.005 0.01 0.0150
50
100
150
200
250
300
350
ω
0.2 0.4 0.6 0.80
0.5
1
1.5
2
2.5
3
3.5
ρ
1 1.5 2 2.50
0.2
0.4
0.6
0.8
1
1.2
β
0 1 2x 10−4
0
2000
4000
6000
8000
10000
12000
14000
Vc
0 2 4x 10−6
0
2
4
6
8
10
12
14x 105 Vμ
Table 1 FR Full sample estimation, 2009 vintage
Posterior modes and standard deviations
TFP: Δpt = μt (1− ρL)μt = (1− ρ)w + aμt (1− 2Acos(2π/τ)L+ A2L2)ct = act
CU: CUt = μCU + βct + aCUt/(1− δL)
w ρ Vμ A τ Vc μCU δ β VCU
CUI0.012 0.93 0.5×10−6 0.58 10.47 9.8×10−5 -0.003 0.2 1.77 29.5×10−5
( 0.002) ( 0.16 ) ( 0.12 ) ( 2.85 ) ( 0.01 ) ( 0.28 ) ( 0.32 )
BS0.011 0.94 0.5×10−6 0.6 10.5 9.7×10−5 -0.004 0.59 1.68 31.3×10−5
( 0.002) ( 0.16 ) ( 0.12 ) ( 2.86 ) ( 0.01 ) ( 0.18 ) ( 0.37 )
38
Figure 4 FR Vintage 2009
Paths followed by the 2000-2008 cycle estimates
−0.026
1.318
2.662
−0.586
0.693
1.973
−0.515
1.006
2.526
−0.945
0.509
1.964
−1.29
0.035
1.359
−1.351
−0.078
1.195
00 01 02 03 04 05 06 07 08 09−1.018
0.432
1.882
00 01 02 03 04 05 06 07 08 09−1.543
−0.42
0.703
00 01 02 03 04 05 06 07 08 09−2.397
−1.545
−0.694
Figure 5 FR Vintage 2009
Revisions standard deviation (×100)
1 2 3 40.3
0.4
0.5
0.6
0.7
0.8
0.9
1
1.1
1.2
1.3
Additional observations
39
Figure 6 FR Real-time vintages
Paths followed by the 2000-2008 cycle estimates
0.057
1.36
2.662
−0.441
0.766
1.973
−0.38
1.073
2.526
−1.021
0.471
1.964
−0.586
0.387
1.359
−1.129
0.033
1.195
00 01 02 03 04 05 06 07 08 09−1.38
0.251
1.882
00 01 02 03 04 05 06 07 08 09−1.374
−0.335
0.703
00 01 02 03 04 05 06 07 08 09−3.303
−1.998
−0.694
Figure 7 FR Real-time vintages
Revisions standard deviation (×100)
1 2 3 40.4
0.6
0.8
1
1.2
1.4
1.6
1.8
Additional observations
40
Summary for FR:
• TFP data There is a noticeable level-shift in the 2008 and 2009 vintages.
CU data CUI and BS have similar variability, the BS series is leading one period
in the first sample half. Both have a large dip in 2009.
Link TFP-CU The β-coefficient is significantly different from 0. The β-posterior
distributions obtained with BS and with CU series are quite similar.
Revisions Bivariate decompositions are performing better than HP for both 2009
and real-time datasets.
CUI vs. BS: CUI yields less revisions.
For 2000, the HP TFP gap must be revised up strongly while the bivariate model cor-
rectly indicates a positive gap that is consistent with the cyclical indicators. We observe
strong upward TFP gap revisions until 2004. The HP estimate show large instability
in 2007-2008. The 2000 phenomenom that hits the HP estimates seems to take place
again in 2009 in the opposite direction, the 2009 HP trend in Figure 2 being excessively
pessimistic. On the contrary, estimates that exploit the CU information assign the last
TFP decline to the gap so the 2009 potential growth has more inertia.
41
4.7 Ireland
Figure 1
TFP vintages plus CU series
0.6
0.7
0.8
0.9
1
1.1
1.2
88 91 94 97 00 03 06 09
−0.02
0
0.02
0.04
Figure 2 IE Vintage 2009
Trend growth and cycle
−2
0
2
4
6
Trend growth
88 91 94 97 00 03 06 09
−4
−2
0
2
Cycle
42
Figure 5
Prior and posterior distributions, 2009 vintage
5 10 15 200
0.02
0.04
0.06
0.08
0.1
0.12
τ
0.01 0.02 0.03 0.040
10
20
30
40
50
ω
0.4 0.6 0.80
1
2
3
4
5
ρ
0.5 1 1.50
0.2
0.4
0.6
0.8
1
1.2
β
0 2 4 6 8x 10−4
0
500
1000
1500
2000
2500
3000
3500
Vc
0 1 2x 10−4
0
0.5
1
1.5
2
2.5
x 104 Vμ
Table 1 IE Full sample estimation, 2009 vintage
Posterior modes and standard deviations
TFP: Δpt = μt (1− ρL)μt = (1− ρ)w + aμt (1− 2Acos(2π/τ)L+ A2L2)ct = act
CU: CUt = μCU + βct + aCUt
w ρ Vμ A τ Vc μCU β VCU
CUI0.021 0.93 55.5×10−6 0.43 8.75 32.9×10−5 0.004 0.84 40.1×10−5
( 0.008) ( 0.09 ) ( 0.15 ) ( 3.31 ) ( 0.01 ) ( 0.35 )
BS0.022 0.93 55.7×10−6 0.48 9.09 34.7×10−5 0.003 1.1 44.2×10−5
( 0.008) ( 0.08 ) ( 0.14 ) ( 3.16 ) ( 0.01 ) ( 0.33 )
43
Figure 4 IE Vintage 2009
Paths followed by the 2000-2008 cycle estimates
1.428
2.372
3.315
−0.13
1.074
2.278
−0.073
1.788
3.65
−1.82
0.748
3.316
−2.915
−0.28
2.354
−3.17
−0.597
1.975
00 01 02 03 04 05 06 07 08 09−2.779
−0.598
1.583
00 01 02 03 04 05 06 07 08 09−2.732
−0.072
2.588
00 01 02 03 04 05 06 07 08 09−4.743
−2.719
−0.696
Figure 5 IE Vintage 2009
Revisions standard deviation (×100)
1 2 3 41.2
1.4
1.6
1.8
2
2.2
2.4
2.6
Additional observations
44
Figure 6 IE Real-time vintages
Paths followed by the 2000-2008 cycle estimates
1.682
3.083
4.485
0.485
1.932
3.378
−1.078
1.306
3.691
−2.543
0.387
3.316
−2.486
−0.066
2.354
−3.345
−0.685
1.975
00 01 02 03 04 05 06 07 08 09−2.96
−0.688
1.583
00 01 02 03 04 05 06 07 08 09−2.582
0.003
2.588
00 01 02 03 04 05 06 07 08 09−4.038
−2.367
−0.696
Figure 7 IE Real-time vintages
Revisions standard deviation (×100)
1 2 3 41.4
1.6
1.8
2
2.2
2.4
2.6
2.8
Additional observations
45
Summary for IE:
• TFP data Vintages are stable enough.
CU data The BS series seems slightly more variable than the CUI series. BS
and CUI have opposite growth directions over 2006-2008. There is no CU data for
2009.
Link TFP-CU The β-coefficient is significantly different from 0. It takes larger
values when BS is used instead of CUI.
Revisions The bivariate approach is best performing with both 2009 and real-
time datasets.
CUI vs. BS BS yields less revisions.
The HP filter generates a relatively smooth trend that has problems to capture the abrupt
change in the trend growth rate, in particular the increase in TFP growth towards the
end of the 90s and the subsequent decline. This can clearly be seen for the 2000 gap
estimate: in this year, the smoothness assumption makes it difficult to the HP filter to
properly account for the increase in trend TFP growth so the HP filter suggests a large
positive output gap. As time passes, trend growth is slowly adjusted upwards for 2000
and the 2009 estimate for the year 2000 approaches, though it does not reach it, the
bivariate. This last remains remarkably stable across the 2005-2009 vintages.
46
4.8 Italy
Figure 1
TFP vintages plus CU series
0.5
0.55
0.6
0.65
0.7
85 88 91 94 97 00 03 06 09
−0.08
−0.06
−0.04
−0.02
0
0.02
Figure 2 IT Vintage 2009
Trend growth and cycle (×100)
−3
−2
−1
0
1
2
Trend growth
85 88 91 94 97 00 03 06 09−4
−3
−2
−1
0
1
2
Cycle
47
Figure 3 IT Vintage 2009
Prior and posterior distributions, 2009 vintage
5 10 15 20 250
0.05
0.1
0.15
τ
0.005 0.01 0.015 0.020
20
40
60
80
100
ω
0.6 0.7 0.8 0.90
2
4
6
8
10
ρ
1.4 1.6 1.8 2 2.2 2.40
0.5
1
1.5
2
β
0 1 2x 10−4
0
2000
4000
6000
8000
10000
12000
14000
Vc
0 0.5 1 1.5x 10−5
0
1
2
3
4
x 105 Vμ
Table 1 IT Full sample estimation, 2009 vintage
Posterior modes and standard deviations
TFP: Δpt = μt (1− ρL)μt = (1− ρ)w + aμt (1− 2Acos(2π/τ)L+ A2L2)ct = act
CU: CUt = μCU + βct + aCUt
w ρ Vμ A τ Vc μCU β VCU
CUI0.008 0.97 3×10−6 0.57 7.2 10.1×10−5 -0.001 1.88 9.6×10−5
( 0.004) ( 0.06 ) ( 0.14 ) ( 2.76 ) ( 0.01 ) ( 0.2 )
BS0.009 0.98 2.9×10−6 0.55 7.09 10×10−5 0 1.6 12.9×10−5
( 0.004) ( 0.06 ) ( 0.14 ) ( 2.77 ) ( 0.01 ) ( 0.23 )
48
Figure 4 IT Vintage 2009
Paths followed by the 2000-2008 cycle estimates
−1.238
0.246
1.729
−1.064
0.376
1.817
−1.661
−0.47
0.722
−2.534
−1.527
−0.52
−2.34
−1.053
0.233
−2.022
−0.844
0.333
00 01 02 03 04 05 06 07 08 09−1.975
−0.605
0.764
00 01 02 03 04 05 06 07 08 09−1.766
−0.265
1.236
00 01 02 03 04 05 06 07 08 09−2.008
−0.746
0.517
Figure 5 IT Vintage 2009
Revisions standard deviation (×100)
1 2 3 40.2
0.4
0.6
0.8
1
1.2
1.4
1.6
Additional observations
49
Figure 6 IT Real-time vintages
Paths followed by the 2000-2008 cycle estimates
−0.954
0.567
2.087
−1.067
0.451
1.969
−1.596
−0.398
0.8
−2.099
−1.125
−0.151
−2.247
−1.007
0.233
−2.22
−0.943
0.333
00 01 02 03 04 05 06 07 08 09−1.999
−0.617
0.764
00 01 02 03 04 05 06 07 08 09−1.399
−0.081
1.236
00 01 02 03 04 05 06 07 08 09−2.001
−0.742
0.517
Figure 7 IT Real-time vintages
Revisions standard deviation (×100)
1 2 3 40.2
0.4
0.6
0.8
1
1.2
1.4
1.6
Additional observations
50
Summary for IT:
• TFP data The 2000-vintage has a low level and the 2009 one shows a positive
level shift.
CU data The two series are similar. Both end on a large dip in 2009.
Link TFP-CU: the β-coefficient is far away from 0. It takes slightly larger values
with CUI.
Revisions: the bivariate model yields less revisions than univariate and HP de-
compositions both with the 2009 data and the real-time vintages, with both CUI
and BS series.
CUI vs. BS: less revisions are obtained with the CUI series.
The HP TFP gap estimate fails to capture the cyclical boom in 2000 and in 2006-2007.
Also, revisions are slow to occur and affect all points between 2000 and 2007. There
are substantially less revisions when the information on capacity utilisation is used,
particularly for the years which exhibit peaks in capacity utilisation. Large revisions are
recorded with HP for the gap in the years 2007-2008.
51
4.9 Luxembourg
Figure 1
TFP vintage plus CU series
0.5
0.6
0.7
0.8
0.9
85 88 91 94 97 00 03 06 09
−0.1
−0.05
0
0.05
The EC business surveys series are not available for LU.
Figure 2 LU Vintage 2009
Trend growth and cycle (×100)
−4
−2
0
2
4
6
Trend growth
85 88 91 94 97 00 03 06 09−6
−4
−2
0
2
4
Cycle
52
Figure 3 LU Vintage 2009
Prior and posterior distributions, 2009 vintage
5 10 15 20 250
0.02
0.04
0.06
0.08
0.1
0.12
0.14
τ
0.005 0.01 0.015 0.02 0.0250
20
40
60
80
ω
0.4 0.6 0.80
1
2
3
4
5ρ
0.5 1 1.5 20
0.2
0.4
0.6
0.8
1
1.2
1.4
β
0 2 4 6 8x 10−4
0
500
1000
1500
2000
2500
3000
Vc
0 0.5 1 1.5x 10−4
0
1
2
3
4
5
x 104 Vμ
Table 1 LU Full sample estimation, 2009 vintage
Posterior modes and standard deviations
TFP: Δpt = μt (1− ρL)μt = (1− ρ)w + aμt (1− 2Acos(2π/τ)L+ A2L2)ct = act
CU: CUt = μCU + βct + aCUt
w ρ Vμ A τ Vc μCU β VCU
CUI0.015 0.93 23.2×10−6 0.48 8.08 36.8×10−5 -0.003 1.22 56.1×10−5
( 0.004) ( 0.1 ) ( 0.13 ) ( 3.05 ) ( 0.01 ) ( 0.29 )
53
Figure 4 LU Vintage 2009
Paths followed by the 2000-2008 cycle estimates
0.014
2.421
4.829
−1.676
0.127
1.931
−1.42
0.463
2.346
−2.154
−0.713
0.728
−2.125
−0.616
0.893
−1.147
0.331
1.809
00 01 02 03 04 05 06 07 08 09−0.86
0.972
2.804
00 01 02 03 04 05 06 07 08 09−0.865
1.43
3.725
00 01 02 03 04 05 06 07 08 09−3.305
−2.058
−0.811
Figure 5 LU Vintage 2009
Revisions standard deviation (×100)
1 2 3 41.1
1.2
1.3
1.4
1.5
1.6
1.7
1.8
1.9
2
2.1
Additional observations
54
Figure 6 LU Real-time vintages
Paths followed by the 2000-2008 cycle estimates
−0.813
2.725
6.263
−1.743
0.447
2.637
−4.725
−1.189
2.346
−3.781
−1.588
0.604
−1.688
0.075
1.837
−1.727
0.041
1.809
00 01 02 03 04 05 06 07 08 09−1.139
0.833
2.804
00 01 02 03 04 05 06 07 08 09−0.998
1.364
3.725
00 01 02 03 04 05 06 07 08 09−1.883
−0.769
0.344
Figure 7 LU Real-time vintages
Revisions standard deviation (×100)
1 2 3 41
1.2
1.4
1.6
1.8
2
2.2
2.4
2.6
2.8
Additional observations
55
Summary for LU:
• TFP data Real-time data show large enough revisions.
CU data There is a level shift in 1998 in the CUI series. No BS series available.
Link TFP-CU The β-coefficient is significantly different from 0 with a posterior
mode above one as expected.
Revisions The bivariate model yields less revisions than univariate and HP de-
compositions both with the 2009 data and the real-time vintages.
The bivariate estimates uses the capacity utilisation series to identify a cyclical peak in
2000-2001. The HP filter estimates instead a negative TFP gap, particularly in 2001
which is subsequently revised to become positive. The 2008 gap estimated with the CUI
series has a large revision in 2009. Most often BS series yield better results but it is not
available for LU.
56
4.10 Netherlands
Figure 1
TFP vintages plus CU series
0.45
0.5
0.55
0.6
0.65
0.7
0.75
85 88 91 94 97 00 03 06 09
−0.06
−0.04
−0.02
0
0.02
Figure 2 NL Vintage 2009
Trend growth and cycle (×100)
−4
−3
−2
−1
0
1
2
Trend growth
85 88 91 94 97 00 03 06 09−6
−4
−2
0
2
Cycle
57
Figure 3 NL Vintage 2009
Prior and posterior distributions, 2009 vintage
5 10 15 20 250
0.02
0.04
0.06
0.08
0.1
τ
0.005 0.01 0.015 0.020
50
100
150
200
250
ω
0.2 0.4 0.6 0.80
1
2
3
4
ρ
0.8 1 1.2 1.40
0.5
1
1.5
2
2.5
β
0 2 4x 10−4
0
1000
2000
3000
4000
5000
6000
7000
Vc
0 2 4 6x 10−6
0
2
4
6
8
10
x 105 Vμ
Table 1 NL Full sample estimation, 2009 vintage
Posterior modes and standard deviations
TFP: Δpt = μt (1− ρL)μt = (1− ρ)w + aμt (1− 2Acos(2π/τ)L+ A2L2)ct = act
CU: CUt = μCU + βct + aCUt/(1− δL)
w ρ Vμ A τ Vc μCU δ β VCU
CUI0.014 0.95 0.6×10−6 0.57 9.65 16.6×10−5 -0.001 0.78 1.11 8.9×10−5
( 0.002) ( 0.15 ) ( 0.15 ) ( 3.32 ) ( 0.01 ) ( 0.18 ) ( 0.16 )
BS0.014 0.95 0.7×10−6 0.57 8.99 17.2×10−5 -0.003 0.66 0.93 10.9×10−5
( 0.002) ( 0.15 ) ( 0.15 ) ( 3.34 ) ( 0.01 ) ( 0.22 ) ( 0.18 )
58
Figure 4 NL Vintage 2009
Paths followed by the 2000-2008 cycle estimates
−0.125
1.29
2.705
−0.859
0.452
1.763
−1.847
−0.725
0.398
−2.386
−1.351
−0.316
−1.477
−0.278
0.921
−1.126
0.071
1.269
00 01 02 03 04 05 06 07 08 09−0.993
0.438
1.868
00 01 02 03 04 05 06 07 08 09−0.584
1.232
3.049
00 01 02 03 04 05 06 07 08 09−0.969
1.149
3.267
Figure 5 NL Vintage 2009
Revisions standard deviation (×100)
1 2 3 40.5
1
1.5
2
Additional observations
59
Figure 6 NL Real-time vintages
Paths followed by the 2000-2008 cycle estimates
−0.494
1.204
2.902
−1.874
−0.055
1.763
−2.362
−0.982
0.398
−3.165
−1.741
−0.316
−2.355
−0.717
0.921
−1.745
−0.238
1.269
00 01 02 03 04 05 06 07 08 09−0.649
0.609
1.868
00 01 02 03 04 05 06 07 08 09−1.149
0.95
3.049
00 01 02 03 04 05 06 07 08 09−1.256
1.005
3.267
Figure 7 NL Real-time vintages
Revisions standard deviation (×100)
1 2 3 40.8
1
1.2
1.4
1.6
1.8
2
2.2
Additional observations
60
Summary for NL:
• TFP data The vintages are stable until 2005 included. The next vintages show
both a level shift and a change in the series growth between 1995 and 2000.
CU data The two series are quite similar. Both have a large dip in 2009.
Link TFP-CU The β-coefficient is significantly different from 0. When CUI is
used, the β-posterior mode is larger than 1.0 and greater than when using BS.
Revisions The bivariate model yields less revisions both with the 2009 data and
the real-time vintages, with both CUI and BS series.
CUI vs. BS: The results are very close.
The largest difference beween HP and bivariate estimates appear in the 2006-2009 years
for which HP gaps are heavily revised with sign switch.
61
4.11 Portugal
Figure 1
TFP vintages plus CU series
0.7
0.75
0.8
0.85
0.9
0.95
88 91 94 97 00 03 06 09
−0.06
−0.04
−0.02
0
0.02
Figure 2 PT Vintage 2009
Trend growth and cycle (×100)
−1
0
1
2
3
4
5
Trend growth
88 91 94 97 00 03 06 09−4
−2
0
2
4
Cycle
62
Figure 3 PT Vintage 2009
Prior and posterior distributions, 2009 vintage
5 10 15 200
0.02
0.04
0.06
0.08
0.1
0.12
0.14τ
0.005 0.01 0.015 0.02 0.0250
20
40
60
80
100
ω
0.4 0.6 0.80
1
2
3
4ρ
0.5 1 1.5 20
0.2
0.4
0.6
0.8
β
0 1 2 3x 10−4
0
2000
4000
6000
8000
Vc
0 1 2x 10−4
0
1
2
3
4
5
x 104 Vμ
Table 1 PT Full sample estimation, 2009 vintage
Posterior modes and standard deviations
TFP: Δpt = μt (1− ρL)μt = (1− ρ)w + aμt (1− 2Acos(2π/τ)L+ A2L2)ct = act
CU: CUt = μCU + βct + aCUt
w ρ Vμ A τ Vc μCU β VCU
CUI0.014 0.91 19.7×10−6 0.43 7.38 12.5×10−5 0.001 0.84 50.6×10−5
( 0.004) ( 0.11 ) ( 0.13 ) ( 3.09 ) ( 0.01 ) ( 0.43 )
BS0.015 0.91 19.9×10−6 0.45 7.59 12.5×10−5 0.001 0.96 48.2×10−5
( 0.004) ( 0.12 ) ( 0.13 ) ( 3.01 ) ( 0.01 ) ( 0.42 )
63
Figure 4 PT Vintage 2009
Paths followed by the 2000-2008 cycle estimates
−1.09
0.839
2.768
−2.105
−0.428
1.25
−2.817
−1.362
0.093
−3.128
−1.745
−0.363
−3.143
−1.841
−0.539
−2.728
−1.409
−0.091
00 01 02 03 04 05 06 07 08 09−2.828
−1.606
−0.384
00 01 02 03 04 05 06 07 08 09−1.616
−0.072
1.472
00 01 02 03 04 05 06 07 08 09−2.459
−1.356
−0.253
Figure 5 PT Vintage 2009
Revisions standard deviation (×100)
1 2 3 40.6
0.7
0.8
0.9
1
1.1
1.2
1.3
1.4
1.5
Additional observations
64
Figure 6 PT Real-time vintages
Paths followed by the 2000-2008 cycle estimates
−1.325
1.538
4.4
−2.334
−0.288
1.758
−3.04
−1.286
0.467
−3.007
−1.521
−0.035
−3.376
−1.943
−0.509
−3.115
−1.595
−0.076
00 01 02 03 04 05 06 07 08 09−2.903
−1.633
−0.363
00 01 02 03 04 05 06 07 08 09−2.46
−0.494
1.472
00 01 02 03 04 05 06 07 08 09−2.403
−1.328
−0.253
Figure 7 PT Real-time vintages
Revisions standard deviation (×100)
1 2 3 40.6
0.8
1
1.2
1.4
1.6
1.8
2
Additional observations
65
Summary for PT:
• TFP data: large revisions in the vintages.
CU data : There is a large dip in the series in 1993. The BS and CUI series are
similar.
Link TFP-CU The β-coefficient is significantly different from 0.
Revisions: the bivariate model yields less revisions than univariate and HP de-
compositions both with the 2009 data and the real-time vintages, with both CUI
and BS series.
CUI vs. BS: slightly less revisions are obtained with the BS series.
The largest differences between HP and the bivariate estimates occur in the last five
years. 2007 is particularly bad for HP.
66
4.12 United Kingdom
Figure 1
TFP vintages plus CU series
0.4
0.5
0.6
0.7
85 88 91 94 97 00 03 06 09−0.08
−0.06
−0.04
−0.02
0
0.02
0.04
Figure 2 UK Vintage 2009
Trend growth and cycle (×100)
−3
−2
−1
0
1
2
Trend growth
85 88 91 94 97 00 03 06 09
−3
−2
−1
0
1
2
Cycle
67
Figure 3 UK Vintage 2009
Prior and posterior distributions, 2009 vintage
5 10 15 20 250
0.02
0.04
0.06
0.08
0.1
0.12
0.14τ
0.005 0.01 0.015 0.020
50
100
150
ω
0.2 0.4 0.6 0.80
0.5
1
1.5
2
2.5
3
ρ
0.5 1 1.5 20
0.2
0.4
0.6
0.8
1
1.2
1.4
β
0 1 2x 10−4
0
2000
4000
6000
8000
10000
12000
Vc
0 2 4x 10−5
0
2
4
6
8
10
x 104 Vμ
Table 1 UK Full sample estimation, 2009 vintage
Posterior modes and standard deviations
TFP: Δpt = μt (1− ρL)μt = (1− ρ)w + aμt (1− 2Acos(2π/τ)L+ A2L2)ct = act
CU: CUt = μCU + βct + aCUt/(1− δL)
w ρ Vμ A τ Vc μCU δ β VCU
CUI0.015 0.87 6.9×10−6 0.65 9.3 9.5×10−5 -0.003 0.78 1.4 25.3×10−5
( 0.003) ( 0.14 ) ( 0.13 ) ( 2.95 ) ( 0.01 ) ( 0.21 ) ( 0.32 )
BS0.015 0.84 5.6×10−6 0.69 10.48 10.3×10−5 -0.004 0.77 1.2 27.7×10−5
( 0.003) ( 0.15 ) ( 0.12 ) ( 2.9 ) ( 0.01 ) ( 0.22 ) ( 0.29 )
68
Figure 4 UK Vintage 2009
Paths followed by the 2000-2008 cycle estimates
0.416
0.865
1.313
−0.3
0.189
0.679
−0.76
−0.062
0.637
−0.646
0.323
1.291
−0.39
0.632
1.653
−0.838
0.149
1.136
00 01 02 03 04 05 06 07 08 09−0.743
0.541
1.826
00 01 02 03 04 05 06 07 08 09−0.617
0.759
2.135
00 01 02 03 04 05 06 07 08 09−1.698
−0.13
1.438
Figure 5 UK Vintage 2009
Revisions standard deviation (×100)
1 2 3 40.5
0.6
0.7
0.8
0.9
1
1.1
1.2
1.3
1.4
Additional observations
69
Figure 6 UK Real-time vintages
Paths followed by the 2000-2008 cycle estimates
−0.897
0.209
1.315
−1.476
−0.398
0.679
−1.736
−0.55
0.637
−1.304
−0.006
1.291
−1.054
0.3
1.653
−1.315
−0.089
1.136
00 01 02 03 04 05 06 07 08 09−0.687
0.57
1.826
00 01 02 03 04 05 06 07 08 09−0.274
0.93
2.135
00 01 02 03 04 05 06 07 08 09−1.346
0.046
1.438
Figure 7 UK Real-time vintages
Revisions standard deviation (×100)
1 2 3 40.5
0.6
0.7
0.8
0.9
1
1.1
1.2
1.3
1.4
1.5
Additional observations
70
Summary for UK:
• TFP data The 2006-2009 vintages show large differences in the first sample-half
with respect to the other vintages. All series show cyclical fluctuations in the 1985-
1993 years that disappear afterwards.
CU data The BS series is slightly more variable in the last years.
Link TFP-CU The β-coefficient is significantly different from 0. It takes larger
values when CUI is used instead of BS.
Revisions Bivariate model using the CUI and BS series dominate with both 2009
data and real-time dataset.
CUI vs. BS The two series yield similar results.
For all years 2000-2008, the real-time concurrent estimates obtained with the three meth-
ods go through large positive revisions with a sign switch. Concurrent TFP gap estimates
are systematically excessively pessimistic.
71
5 Conclusion
The figure below puts together the revisions standard deviation recorded in real-time
for all countries. It can be seen that in all cases the bivariate method improve over the
univariate approach. This shows that CU series do have informative content for TFP.
The CU improvement makes the bivariate approach more reliable in real-time than HP
for all cases. CU series give a better outcome than BS in three cases, ie. DE, IT and FR,
BS dominates for DK, ES, IE, and the two series give equivalent results for the other
countries. The β-coefficient is significant for all countries, with posterior mode higher
than one except for EL.
Real-time vintages
Revisions standard deviation (×100)
0.25
0.56
0.89
be0.37
0.87
1.38
dk0.39
1
1.62
de
0.43
0.88
1.34
el0.54
0.81
1.08
es0.48
1
1.52
fr
1.42
1.95
2.49
ie0.31
0.91
1.52
it1.18
1.96
2.74
lu
1 2 3 40.87
1.42
1.98
nl1 2 3 4
0.74
1.17
1.61
pt1 2 3 4
0.57
1.02
1.48
uk
The analysis thus shows that especially around boom bust episodes, the use of cyclical
indicators leads to less revisions and sign changes of the TFP gap.
72
6 Appendix
6.1 IG-priors for variance parameters
Table A1
Mean and standard deviation of IG-variance priors
BE DE DK EL ES FR IE IT LU NL PT UK
Vc(×10−4) 1.6 1.6 1.6 1.7 2.0 1.6 14 1.6 3.8 2.6 2.4 1.6
Vμ(×10−6) 2.4 2.4 22 20 1.2 8.0 40 2.0 20 1.0 20 10
VCU(×10−4) 4.8 16.3 9.3 2.8 3.1 3.6 12.8 1.8 6.4 1.8 9.3 5.1
The prior distribution of variance parameters are tuned so that mean and standard deviations are equal.
73
6.2 Correlations between CU series.
Table A2 Correlations between CU series
Corr(CUI,BS) Corr(CUI,PMI) Corr(BS,PMI)
BE .917 — —
DE .862 .906 .863
DK .824 — —
EL .660 — —
ES .733 .868 .766
FR .647 .792 .742
IE .881 .837 .862
IT .927 .883 .920
NL .827 — —
PT .918 — —
UK .929 .784 .828
74
7 References
Denis C., Mc Morrow K. and Roeger W. (2002), “Production function approach
to calculating potential growth and output gaps - estimates for the EU Member States
and the US”, European Economy, Economic Papers nr. 176, European Commission.
Denis C., Grenouilleau D., Mc Morrow K. and Roeger W. (2006), “Calculat-
ing potential growth rates and output gaps - A revised production function approach”,
European Economy, Economic Papers nr. 246, European Commission.
European Commission (2005), “New and updated budgetary sensitivities for the EU
budgetary surveillance”, ec.europa.eu/economy finance/sg pact fiscal policy/sg progra-
mmes9085 en.htm.
European Commission (2008), Public finance in EMU - 2008, European Economy,
4-2008, Brussels: European Commission.
Hodrick R.J. and E.C. Prescott (1997), ”Post-war U.S. business cycle: an empir-
ical investigation”, Journal of Money, Credit and Banking, 29, 1, 1-16.
Kuttner K. (1994), “Estimating potential output as a latent variable”, Journal of
Business and Economic Statistics, 12, 3, 361-368.
Marcellino, M. and A. Musso (2008), “Real time estimates of the euro area output
gap”. European University Institute, mimeo.
Pierce, D.A. (1980), “Data Revisions in Moving Average Seasonal Adjustment proce-
dures”, Journal of Econometrics, 14, 1, 95-114.
Planas C. and Rossi A. (2004), “Can inflation improve the real-time reliability of
output gap estimates?”, Journal of Applied Econometrics, 121-133.
Planas C., Rossi A., and Fiorentini G. (2008), “Bayesian analysis of the output
gap”, Journal of Business & Economic Statistics, 19, 1, 18-32.
Proietti T., Musso T., and Wastermann T. (2007), “Estimating potential out-
put and output gap for the euro area: a production function approach ”, Empirical
Economics, 33, 85-115.
Runstler, G. (2002), “The information content of real-time output gap estimates: an
application to the euro area”. European Central Bank Working Paper 182.
Stock J. and Watson M. (1998), “Median unbiased estimation of coefficient variance
in a time-varying parameter model”, Journal of the American Statistical Association,
349-358.
75