PUBLiC dEBT GRoWTH , GREAT RECESSioN ANd FiSCAL CoNSoLidATioN-THE SERBiAN EPiSodE

This paper analyses the dynamics of public debt growth in Serbia between the fourth quarter of 2004 (2004Q4) and the same quarter of 2017 (2017Q4). The empirical estimates capture the upward shift in public debt growth from the onset of the Great Recession. The results also show how policy makers have managed to curb the growth of public indebtedness in Serbia between 2014Q4-2017Q4. The government should, however, put greater fiscal efforts to reduce the overall share of public debt in gross domestic product in accordance with the fiscal rules of the Republic of Serbia and the Maastricht convergence criteria.

of GDP defined in fiscal rules.In 2014Q1, public debt breached yet another thresholdthe upper limit of 60% of GDP defined in Maastricht convergence criteria (Andric et al., 2016a). 1  Faced with aforementioned public finance trends, the Serbian government launched a fiscal consolidation programme at the end of 2014.The government managed to curb the growth of public indebtedness which led to the stabilization of public debt around 70% of GDP in the first 2 years of fiscal consolidation programme, and its subsequent decline to around 65% of GDP at the end of 2017.The analysis presented in this paper search, hence, for potential structural breaks in the dynamics of public debt growth between 2004Q4 and 2017Q4.
This study adds two contributions to the existing empirical literature on fiscal sustainability.First, this paper is one of the first to analyse the growth of public indebtedness in the case of small open transition economy from Southeastern Europe.Uctum et al. (2006) assess, for example, fiscal sustainability in the case of G7 and some Asian and Latin American economies between 1970 and 2002.Afonso and Jalles (2012) extend fiscal sustainability analysis to OECD economies for the period 1970-2010. Yildiz and Yildirim (2014) analyze fiscal sustainability in Eurozone, i.e., in EMU 12 countries, using annual panel data set between 1995and 2011. Despotovic and Durkalic (2017) compare budget deficits of 34 European economies with respect to numerous statistical and public finance criteria.Lovre et al. (2017) analyze public sector efficiency in 19 developed economies focusing on the relation between government expenditures and economic growth during the global financial crisis.Martins and Duarte (2017) stress the importance of public debt sustainability from historical perspective in the case of The First Portugese Republic.Finally, Jawadi and Sousa (2013) analyse public debt dynamics in the United Kingdom (UK) and the United States (US) for the period 1962Q4-2009Q2 and 1970Q1-2009Q2, respectively.Second, the empirical model presented in this paper captures the dynamics of public debt growth both before and after the Great Recession.The approach outlined in this paper is, hence, similar to the one of Krajewski et al. (2016) who consider a panel of 10 emerging economies from Central and Eastern Europe between 1990and 2012. Contrary to Krajewski et al. (2016), the focus of this paper is on a single economy which enables the analysis of fiscal developments in a greater detail, with a particular emphasis on the Great Recession and the fiscal consolidation package launched at the end of 2014.
The econometric estimates quantify the jump of public debt growth in Serbia from the onset of the Great Recession.In particular, the growth of public debt in Serbia trended around its mean value of -1.75% of GDP before the global financial crisis.After the crisis hit Serbia in the second half of 2008, the growth of public debt has fluctuated around its mean value of 1.82% of GDP.The results, hence, capture the sharp decrease of public debt growth in Serbia before the crisis, as well as its abrupt increase after the crisis.In addition, the findings do not change if the fiscal consolidation package between 2014Q4-2017Q4 is taken into account.Policy makers managed to curb the growth of government debt in Serbia between 2014Q4-2017Q4, but the government should put further fiscal efforts to reduce its overall indebtedness in accordance with the fiscal rules of the Republic of Serbia and Maastricht convergence criteria.
The rest of the paper is organized as follows.The second section provides theoretical background for subsequent empirical estimates.The third section presents baseline results, as well as the results of sensitivity analysis.The fourth section concludes.

ECoNoMETRiC METHodoLoGY
Cafiso (2012) defines the growth of public debt ∆B t as: In equation ( 1), B t stands for the stock of public debt at time t, B t-1 is the stock of public debt from the previous time period t-1, PB t denotes primary fiscal balance, defined as the difference between overall government revenues and primary government expenditures, while i t represents implicit nominal interest rate in time period t on B t-1 .Finally, SFA t measures stock-flow adjustments which are equal to the difference between public debt growth and the officially reported overall fiscal deficit.In particular, these adjustments include all the operations which influence the debt, but do not generate the deficit, and vice versa.Von Hagen and Wolff (2006) provide empirical evidence for the case of advanced European economies that engaged in creative accounting practices by covering large fiscal deficits with stock-flow adjustments after the Stability and Growth Pact was agreed upon in 1998.Izák (2009) also reports how the analysis of stock-flow adjustments has become more important as the EU budgetary surveillance may have provided incentives for shifting items from the officially reported fiscal deficit to the stock-flow adjustments.Finally, Bornhorst et al. (2011) stress the importance of one-off operations such as the re-evaluation of financial assets and liabilities due to exchange rate changes.Since the Serbian government has issued around 80% of its debt in foreign currency, and given the sensitivity of Serbian real effective exchange rate to net capital flows, the change in public debt is chosen as the primary measure of public indebtedness growth (Andric et al., 2016a).Bai (1997) and Bai and Perron (1998, 2003a, 2003b) consider a multiple linear regression model with T periods and m potential structural breaks, i.e., m+1 regimes.In particular, for the observations in the regime j, Bai (1997) and Bai and Perron (1998, 2003a, 2003b) estimate the following least squares regression: for the regimes j=0,1,2,…m, and white noise process ε t .The model in (2) is presented in its most general form, since variables corresponding to vector X t do not vary across regimes, while variables corresponding to vector Z t are allowed to vary across regimes.For a specific set of m breaks, Bai (1997) and Bai and Perron (1998, 2003a, 2003b) minimize the following sum of squared residuals: using standard least squares regression to obtain estimates . The global m-break optimizers are the set of breaks and corresponding coefficient estimates that (3) minimize sum of squared residuals across all possible sets of m-break partitions (Bai and Perron 1998, 2003a, 2003b).Following Jawadi and Sousa (2013), who apply described Bai-Perron testing procedure in the cases of U.S. and the UK, the meanshift model with m potential structural breaks (T 1 ,T 2 ,…T m ) is estimated: in which j=1,2,…m+1, T 0 =0 and T m+1 =T, μ j refers to the regression coefficients with property μ i ≠μ i+1 (1≤i≤m) and ε t corresponds to the error term.Bai (1997) was the first to derive the consistency, rate of convergence and asymptotic distribution of the ordinary least squares (OLS) estimates of a change point model from equation ( 2).The OLS estimates are consistent even if the disturbances are dependent and heteroscedastic.In addition, Bai (1997) and Bai and Perron (2003a) construct confidence intervals for the OLS estimated break dates (T 1 ,T 2 ,…T m ).
Building on the paper of Bai (1997), Bai and Perron (1998) have introduced several structural break tests.First, Bai and Perron (1998) propose two double maximum tests of the null hypothesis of no structural break against an unknown number of breaks given some upper bound M. 2 The upper bound M for the number of breaks is inversely proportional to the size of trimming percentage ϵ, ϵ=h⁄T, in which h represents the minimal length of each sub-regime.As Bai and Perron (2003b) show, larger values of trimming percentage ϵ are needed to achieve tests with correct size in finite samples, especially if one allows for heterogeneity across segments and/or serial correlation in the errors.Second, Bai and Perron (1998) construct a test of ℓ versus ℓ+1 breaks, labelled supF T (ℓ+1|ℓ ) test.The test assumes the application of (ℓ+1) tests of the null hypothesis of no structural change vs. the alternative hypothesis of a single change.
Finally, Bai andPerron (2003a, 2003b) propose the following algorithm for determining the overall number of structural breaks: 1) prespecify the upper bound for the number of breaks M by setting the value of trimming percentage ϵ; 2) test the null hypothesis of no structural break against the alternative of a prespecified number of breaks defined in step 1) by using double maximum tests of Bai and Perron (1998); 3) if double maximum tests indicate the presence of at least one structural break, proceed with the application of supF T (l+1|ℓ) test by selecting M such that supF T (l+1|ℓ) is insignificant for ℓ ≥ M.
An alternative approach to the one proposed in Bai (1997) would be to use unit root tests with endogenously determined breakpoints, such as those developed in Zivot and Andrews (1992), Vogelsang and Perron (1998), Lumsdaine and Papell (1997) and Lee and Strazicich (2003).The results presented in Bai (1997), and later developed in Bai and Perron (1998, 2003a, 2003b), hold, however, for both nonstationary and trending regressors.In addition, the results of Bai (1997) support more general forms of non-linearity, especially with respect to the number of breakpoints and with respect to the statistical properties of disturbance terms.3

EMPiRiCAL EVidENCE
This section consists of three subsections.Subsection 3.1 discusses major stylised facts regarding the dynamics of public debt growth.Subsection 3.2 presents a mean-shift model which captures stylized facts from subsection 3.1.Finally, subsection 3.3 provides the findings of sensitivity analysis.

Stylized Facts
The empirical estimates are based on quarterly data set between 2004Q4 and 2017Q4.More precisely, the analysis presented in this paper search for potential structural breaks in the dynamics of public debt growth between 2004Q4 and 2017Q4.The availability of official quarterly data from the Ministry of Finance of the Republic of Serbia determines both the beginning and the end of the sample span.The baseline estimates refer to the period 2004Q4-2014Q4, while the robustness checks refer to the period 2004Q4-2017Q4, taking into account, hence, the period of fiscal consolidation package between 2014Q4-2017Q4.The growth of public debt is measured as % of GDP, following the leads of Hakkio andRush (1991) andBohn (2005).Hakkio and Rush (1991) argue how shares of GDP represent the most appropriate measure for growing economies.This measure is, hence, suitable in the case of Serbia, given its high GDP growth rates in the pre-crisis period.Bohn (2005) supports the claims of Hakkio and Rush (1991) by documenting how standard deviations of GDP ratios are much less volatile with respect to the standard deviations of nominal and real variables.
Figure 1 displays three distinctive subperiods in the dynamics of public debt growth: i) the first sub-period spans from 2004Q3 to 2008Q2, and captures the period before the crisis; ii) the second sub-period spans from 2008Q3 to 2014Q3, and captures the period after the crisis up to the beginning of fiscal consolidation programme in 2014Q4; iii) the third sub-period spans from The most important discretionary factors encompassed the extraordinary increase in public wages and pensions in the wake of the Great Recession, accompanied with the reduction in custom rates on imported goods from the European Union (EU).Although the government implemented some ad-hoc fiscal consolidation measures between 2008Q3 and 2014Q3, it failed to curb the growth of public indebtedness in Serbia.Adhoc fiscal consolidation measures on the revenue side encompassed the hikes in standard VAT rate, corporate income tax rate and excise taxes on tobacco, oil, and oil derivatives.Ad-hoc fiscal consolidation measures on the expenditure side encompassed slower than inflation indexation scheme for public wages and pensions, public sector hiring freezes and reductions in monthly governmental salaries which exceeded 60.000 RSD during 2014. 4 Finally, in 2014Q4, at the beginning of the third sub-period, the Serbian government launched a 3-year fiscal consolidation package.The package centred on the reduction of public sector wages and pensions.In particular, government reduced nominal pensions by 5% and public sector wages by 10%.As a result of introduced fiscal austerity measures, the government curbed the growth of public debt in Serbia in the first two years of fiscal consolidation programme, and managed to stabilize public debt-to-GDP ratio around 70% of GDP.Finally, in the last year of fiscal consolidation programme, public debt declined to around 65% of GDP.

Baseline Estimates
The empirical estimates build on the previous contributions of Arsic et al. (2013) and Andric et al. (2016aAndric et al. ( , 2016b) ) which are concerned with the sustainability of fiscal policy in Serbia after the year 2000.The baseline estimates, which are obtained for the period 2004Q4-2014Q4, consist of two sets of results.First, unit root tests are implemented to determine the persistence of shocks to public debt growth.If the shocks to public debt growth are transitory, i.e., if public debt growth is stationary stochastic process, then the dynamics of public debt growth is predictable, and policy makers can control movements in public debt with discretionary fiscal policy measures, and vice versa.Second, the equation ( 4) is estimated in accordance with the methodology developed in Bai (1997) and Bai and Perron (1998, 2003a, 2003b).
Table 1 presents the results of unit root tests.In particular, table 1 shows the results of point-optimal invariant unit root tests proposed in Elliott et al. (1996), as well as the results of M-unit root tests proposed in Ng and Perron (2001).The results from table 1 support the stationarity hypothesis, i.e., the hypothesis about the transitory nature of public debt growth innovations.This finding is consistent with the sample autocorrelation function of public debt growth with the first lag autocorrelation coefficient of only 0.46.These particular unit root tests are chosen since they are robust to small sample bias.The data are first detrended by generalized least squares (GLS), since Elliott et al. (1996) show how GLS detrending yields power gains for unit root tests and allows for a more precise autoregressive spectral density estimate, especially in the case of an unknown mean.In addition, the number of lags in unit root testing regressions is determined in accordance with the modified Akaike criterion (MAIC) in which the maximum number of lags is set to 4, since the analysis is based on quarterly data.Ng and Perron (2001) show how MAIC leads to substantial size improvements over standard information criteria.
Since shocks to public debt growth are transitory, then policy makers can use discretionary fiscal policy to control the growth of overall public indebtedness.To investigate the nature of these measures in a greater detail, the mean-shift model from equation ( 4) is estimated.
Table 2 presents the estimates of the mean-shift model for the period 2004Q4-2014Q4.The estimated econometric specification explains 53% of variations of public debt growth.The structural break in the dynamics of public debt growth occurred in 2008Q4 which coincides with the arrival of the global financial crisis to Serbia. 5 The results from table 2

Table 1. Unit Root Tests, 2004Q4-2014Q4
5 The 95% confidence interval for the break date is .The confidence interval is calculated according to the formula in which is the estimated break date, is the integer part of , is the 97.5th quintile from the symmetric case CDF for which and !"# $ $ % &'( )* + is a scale factor with , and * + defined as the public debt growth coefficient before the break point, public debt growth coefficient after the break point and the estimated variance of , $ from (2.4), respectively.The use of symmetric CDF is appropriate, since the model's residuals are stationary on the whole sample.The results of ADF test, which are available from the authors upon request, confirm this finding.
with the findings of Andric et al. (2016aAndric et al. ( , 2016b) ) who detect the sharp decline of primary fiscal balance response to public debt accumulation, interest payments and implied effective interest rate on Serbian government bonds after the Great Recession hit Serbian economy.
The estimated coefficient before the crisis equals -1.75% of GDP, while the estimated coefficient after the crisis equals +1.82% of GDP.Reported estimates capture, hence, the abrupt shift in public debt dynamics after the global financial crisis.In obtaining the estimates from table 2 above, the upper bound for potential number of breaks is first specified, i.e., the minimal regime length is determined by setting the value of trimming percentage ϵ.Given the short sample span, the value of ϵ is set to 25%, which corresponds to a minimal regime length of approximately 13 quarters, i.e., to a maximum of two structural breaks M=2.Second, double maximum tests UD max and WD max from Bai and Perron (1998) are implemented in order to see if at least one break is present, given the ϵ-prespecified upper bound for the potential number of breaks.Both test statistics are statistically significant and equal 49.03 with an estimated structural break in 2008Q4.Finally, given the presence of a break in 2008Q4, the sequential test of ℓ+1 versus ℓ globally determined breaks is applied, as in Bai and Perron (1998, 2003a, 2003b).The sequential procedure reaffirms 2008Q4 as sole structural break in data generating process for public debt growth.

Sensitivity Analysis
To check for the robustness of the baseline estimates, the mean-shift model is modified to allow for heterogeneous errors across breaks, and the equation ( 4) is subsequently estimated for the period 2004Q4-2017Q4.Heterogeneous error distributions across breaks are justifiable due to regime changes induced by the Great Recession and fiscal consolidation.The inclusion of a period between 2014Q4 and 2017Q4 should evaluate achieved fiscal accomplishments of the 3-year consolidation programme. 6able 3 shows the results of sensitivity analysis.The results are almost identical to those from table 2. Structural breaks occur, however, both in 2008Q4 and in 2014Q4, with the 95% confidence intervals between [2008Q2-2009Q2]   consolidation package has been initiated.After two years of fiscal consolidation, the government has managed, hence, to stop further growth of public indebtedness in Serbia.Public debt has stabilized around 70% of GDP at the end of 2016, and declined to 65% of GDP at the end of 2017.Further fiscal efforts are needed, therefore, to reduce the share of public debt in GDP in accordance with the upper debt limit defined in fiscal rules (45% of GDP) and Maastricht convergence criteria (60% of GDP).

CoNCLUSioN
This paper modelled the dynamics of public debt growth in Serbia between 2004Q4 and 2017Q4.The presented empirical estimates capture the upward shift in public debt growth from the onset of the Great Recession.The results also show how policy makers have managed to curb the growth of public indebtedness in Serbia between 2014Q4-2017Q4.The government should, however, put greater fiscal efforts to reduce the overall share of public debt in gross domestic product in accordance with the fiscal rules of the Republic of Serbia and Maastricht convergence criteria.Consequently, the results are relevant for public debt management, since lower public indebtedness implies lower borrowing costs, which is in accordance with the official public debt management strategy of the Republic of Serbia that stresses the provision of regular budgetary servicing at the lowest possible cost and at the acceptable level of risk.
This study is one of the first to analyse the growth of public indebtedness in the case of small open transition economy from Southeastern Europe.Other contributions from the literature predominantly focus on advanced OECD economies.The explicit focus on a single economy enables the analysis of fiscal developments in a greater detail, with a particular emphasis on the Great Recession and the fiscal consolidation package between 2014Q4-2017Q4.
The focus on a single national economy, however, has certain limitations.First and foremost, the short sample span puts certain constraints on making more general statements about the historical behaviour of public debt in Serbia.In addition, single case study means that the results cannot be easily extended to other emerging European economies.An important avenue for further research concerns, hence, the extension of the analysis presented in this paper to other economies from Central and Eastern Europe.