Does capacity utilisation help estimating the TFP cycleec.europa.eu/economy_finance/publications/economic_paper/2010/pdf/... · Economic Papers are written by the Staff of the Directorate-General

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.

http://ec.europa.eu/economy_finance/publications/economic/index_en.htm

http://europa.eu/index_en.htm

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.


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


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


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


−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


1 2 3 40.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

1.1

1.2


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


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



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



−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



1 2 3 40.4

0.6

0.8

1

1.2

1.4

1.6

1.8


19

Figure 6 DE Real-time vintages


−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


1 2 3 40.2

0.4

0.6

0.8

1

1.2

1.4

1.6


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.


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


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


−2−1

012345

Trend growth

88 91 94 97 00 03 06 09

−3

−2

−1

0

1

2

3

Cycle

22


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





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



−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



1 2 3 40.4

0.6

0.8

1

1.2

1.4

1.6

1.8


24

Figure 6 DK Real-time vintages


−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


1 2 3 40.2

0.4

0.6

0.8

1

1.2

1.4

1.6


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.


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


−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



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





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



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



1 2 3 40.6

0.65

0.7

0.75

0.8

0.85

0.9

0.95

1

1.05


29

Figure 6 EL Real-time vintages


−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


1 2 3 40.4

0.5

0.6

0.7

0.8

0.9

1

1.1

1.2

1.3

1.4


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


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


−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



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





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



−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



1 2 3 40.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9


34

Figure 6 ES Real-time vintages


−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


1 2 3 40.5

0.6

0.7

0.8

0.9

1

1.1

1.2

1.3


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


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


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





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



−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



1 2 3 40.3

0.4

0.5

0.6

0.7

0.8

0.9

1

1.1

1.2

1.3


39

Figure 6 FR Real-time vintages


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


1 2 3 40.4

0.6

0.8

1

1.2

1.4

1.6

1.8


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


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


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





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



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



1 2 3 41.2

1.4

1.6

1.8

2

2.2

2.4

2.6


44

Figure 6 IE Real-time vintages


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


1 2 3 41.4

1.6

1.8

2

2.2

2.4

2.6

2.8


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.


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


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


−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



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





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



−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



1 2 3 40.2

0.4

0.6

0.8

1

1.2

1.4

1.6


49

Figure 6 IT Real-time vintages


−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


1 2 3 40.2

0.4

0.6

0.8

1

1.2

1.4

1.6


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-


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


−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



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





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



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



1 2 3 41.1

1.2

1.3

1.4

1.5

1.6

1.7

1.8

1.9

2

2.1


54

Figure 6 LU Real-time vintages


−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


1 2 3 41

1.2

1.4

1.6

1.8

2

2.2

2.4

2.6

2.8


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.


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


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


−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



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





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



−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



1 2 3 40.5

1

1.5

2


59

Figure 6 NL Real-time vintages


−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


1 2 3 40.8

1

1.2

1.4

1.6

1.8

2

2.2


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


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


−1

0

1

2

3

4

5

Trend growth

88 91 94 97 00 03 06 09−4

−2

0

2

4

Cycle

62



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





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



−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



1 2 3 40.6

0.7

0.8

0.9

1

1.1

1.2

1.3

1.4

1.5


64

Figure 6 PT Real-time vintages


−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


1 2 3 40.6

0.8

1

1.2

1.4

1.6

1.8

2


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-


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


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


−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



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





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



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



1 2 3 40.5

0.6

0.7

0.8

0.9

1

1.1

1.2

1.3

1.4


69

Figure 6 UK Real-time vintages


−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


1 2 3 40.5

0.6

0.7

0.8

0.9

1

1.1

1.2

1.3

1.4

1.5


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.


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


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

Does capacity utilisation help estimating the TFP cycleec.europa.eu/economy_finance/publications/economic_paper/2010/pdf/... · Economic Papers are written by the Staff of the Directorate-General

Documents