Conclusiones del tercer capítulo

My quarterly dataset covers the period between 1980q1 and 2015q4, meaning that all the major steps in the European integration process in the 1990s and 2000s are included. This dataset consists of twelve old EU countries (including two members

that entered in the 1980s and three members in mid 1990s, see footnote5) forming

a monetary union and three stand-alone old EU member states – Denmark, Sweden and the United Kingdom. This composition of EU countries will enable analysis of the effects and reactions of fiscal policies across different types (utilizations) of fiscal policy, and as a result, levels of indebtedness.

For the estimation of country-specific fiscal rules over time, both fiscal time series, and other macro-economic variables are needed. This study utilizes several sources of macroeconomic series (see below and in details in appendix). Despite the fact that quarterly fiscal time series have been published by Eurostat for EU countries for some time, their length and coverage vary substantially across coun-

tries.55 For the purpose of this study, it was necessary to reconstruct quarterly fiscal

series. My approach to reconstruction (Kalman filter based) is briefly described be- low, and in greater detail in appendix. Few alternative attempts have been made to reconstruct quarterly fiscal series from individual revenue and expenditure items. These attempts used macroeconomic aggregates based on the System of National Accounts (SNA) or its European version (European System of National Accounts,

ESA), such asParedes et al.(2014) for the Euro area (currently, EA-15 covering the

period 1980q1–2013q4), and a similar methodology as in the case of Spain (recently

53

Baumeister and Benati(2013) propose as an alternative, so-called inefficiency factors (IF) that

are calculated as the inverse of relative numerical efficiency. Series of IF should be below twenty to indicate convergence.

54

There was a ‘light’ problem with autocorrelation for some countries (slightly above the 95% confidence intervals), however, coefficients returned fast to confidence intervals. All countries passed the CD test (|CD|<2) as commonly required, even though, I had to increase the burn-in number of draws (to 50000 or 100000 draws for countries like Germany, Sweden or the UK), while keeping the stored sample size unchanged at 10000 draws. Values of the RNE varied across countries across specifications.

55_{Otherwise, one would have to start as late as around the first quarter of 2002 for most EA}

countries with few exceptions such as France or Finland. The case of stand-alone countries presents a somewhat better situation. This is because of the ‘obligatory’ beginning of publishing fiscal quarterly series that goes back to the first quarter of 1999.

CHAPTER 3. INSTITUTIONAL CHANGES AND FISCAL POLICY BEHAVIOUR

updated to 1970q1–2015q4), seeDe Castro et al.(2014). Even that reconstruction is

not without problems because of several changes in methodological concepts (ESA 1979, ESA 1995 and ESA 2010) over last decades. As illustrated for Spain in ap- pendix, my generated fiscal series (primary balance) matches the SNA-reconstructed series very closely, providing a visual check of the method (robustness) and support- ing the Bayesian approach. The same does hold for reconstructed public debt series

that is not shown because of space considerations.56

3.4.1 Reconstruction of quarterly series

This section summarises the main steps used to reconstruct quarterly fiscal and economic time series (for more details on data treatments see the data section in the appendix). The main source of quarterly series for my dataset is Eurostat for fiscal time series (Government finance statistics, ESA2010 and Quarterly national accounts, ESA2010) and OECD (Quarterly National Accounts). Yearly fiscal policy variables are primarily taken from the database of the European Commission (An- nual macro-economic database, AMECO) that is compatible with Eurostat, OECD (Economic Outlook database) and from a historical dataset of fiscal variables pre-

pared by the IMF (Public Finances in Modern History Database, seeMauro et al.,

2013). In order to have comparable series, the same database is used for coun-

try/year observations, that is debt and primary balance, in line with suggestions inBerti et al. (2016). Furthermore, to eliminate ‘spurious’ responses coming from during the Sovereign Debt Crisis conducted (one-off) interventions into financial in- stitutions’ balance sheets in a few EA countries (so-called government support to fi- nancial institutions – GAFS, for further details see the data subsection in appendix), these transfers are excluded from primary balance series. Since these measures did not directly affected public debt series (but changed so-called contingent liabilities), no adjustment of public debt series was carried out. Moreover, there were various one-off operations realised in the past (before 2007). However, there is no consistent and systematic evidence of these items, even for EA countries. Some authors have tried to adjust series for these effects; one approach uses differences in dynamics of

net capital transfers see Joumard et al.(2008). However, I decided not to distort

the dynamics of the fiscal series with imprecise corrections since their implications

56

I thank Javier P´erez (Banco de Espa˜na) for providing me with the latest version of their fiscal dataset accompanyingDe Castro et al.(2014).

Furthermore, an empirical illustration for GIIPS countries and the Netherlands (with and without GAFS series) is shown in the panels of figures (B.1) in the appendix. Details on one-off items and their treatment can be found inEC(2015), and a brief summary for GAFS series is presented in

CHAPTER 3. INSTITUTIONAL CHANGES AND FISCAL POLICY BEHAVIOUR

were also related to countries’ debt series.

Since quarterly fiscal time series are rather short or missing for primary bal-

ance and debt, and the output gap (see below),57 for a majority of EU countries

are available from 1999q1 onwards at best, they are extended for the whole sample period, with quarterly series created with help of the Kalman filter technique and a

Bayesian approach for decomposition of low frequency series.58 _{One of the biggest}

advantages of this approach compared to commonly used mechanical techniques for

temporal disaggregation59 is that quarterly series are constructed with the help of

the information provided by using other (macro) quarterly series that are highly correlated with to-be-reconstructed fiscal series (primary balance and debt). The set of variables employed for the reconstruction consists of the following series: un- employment rate, CPI index, short-term and long-term interest rate, real GDP and

government consumption.60 A similar procedure was employed when reconstruct-

ing other quarterly series (government current expenditures). Further details are provided in the data section in the appendix.

Output gap and the cyclical component of total current expenditures for individual quarters are calculated with the help of the Baxter and King (Band-pass)

filter. The calculation uses commonly used parameters (BK12(6,32) covering main

business cycle frequencies in the range 11_{/2−8 years) that provides better estimates,}

as compared to the Hodrick-Prescott filter (with λ= 1600) on quarterly frequency

(HP filtered series are utilized in robustness section).61 Since both filters have been

shown to have problems in the beginning and end of a time series (‘end-points’), and a few initial period are lost in the BK filter because of the filter construction),

57

For example even the most recent OECD publication on output gaps for individual OECD countries is for yearly frequency only, seeTurner et al.(2016); the same holds true for the ECB; however,Jarocinski and Lenza(2016) discuss methods of estimating output gap at quarterly fre- quency for the Euro area as a whole.

58_{This way of reconstructing quarterly series draws upon the contributions ofGiannone et al.}

(2015) andBa`nbura et al. (2015). I thank Giovanni Ricco for sharing an earlier version of their Matlab code used inCaruso et al.(2015).

59

The most commonly used are Chow-Lin, Fernandez or Litterman; for overview and details on available methods with references see for exampleQuilis(2004).

60_{I also tried to recalculate quarterly series utilizing both techniques, and the results were broadly}

similar in terms of trends and turning points.

61_{Other high values of smoothing parameters for the HP filter were utilized, such as those}

recommended byPerron and Wada (2009) or Market and Ravn(2007) for GDP. However, their gains compared to the standard HP filter were given by the length of available time series.Market

and Ravn(2007) argue that setting the BK filter equal to BK12(6,32) works well for quarterly

series; the closest counterpart of the HP filter for quarterly data would beBK12(2,32) according

toBaxter and King(1999).

I will treat the HP filtered series as a robustness check following a recent paper –Hamilton(2016) – arguing that one should use different filtering techniques other than the HP filter in empirical applications.

CHAPTER 3. INSTITUTIONAL CHANGES AND FISCAL POLICY BEHAVIOUR

series are extended with three or four years (12 or 16 quarters of observations) using

forecasting and backcasting in a bivariate VAR(p) model.62 For these extended

series, both filtering techniques were applied, fitted values were stored, and the extensions of series were dropped. All series were seasonally adjusted (either directly when accessed in particular databases or before any calculations using the ARIMA

X-13 method).63

Owing to data revisions, several studies have shown the importance of data

vintages’ effects on fiscal series (such as Golinelli and Momigliano, 2009), mainly

government balances (in particular on cyclically adjusted fiscal series). Unfortu- nately, real-time analysis cannot be carried out in the case of quarterly time series that are published by Eurostat. Even in the case of yearly series, the AMECO database that has been running since 2002, comparable series available since 2008, and the OECD Economic Outlook database, also provides yearly series.