WPS6303 Policy Research Working Paper 6303 Real-Time Macro Monitoring and Fiscal Policy Eduardo Ley Florian Misch The World Bank Poverty Reduction and Economic Management Network Economic Policy and Debt Department January 2013 Policy Research Working Paper 6303 Abstract This paper considers the effects of inaccurate real-time Monetary Fund’s World Economic Outlook, comprising output data on fiscal management, both with respect final and real-time output data for 175 countries, over to budgetary planning and fiscal surveillance. As newer a period of 17 years. The authors simulate the effects and better information becomes available, output data of output revisions on revisions of the overall balance, available in real time get revised and are likely to conflict the structural balance and debt accumulation. It finds with final figures that are only released some years later. that output revisions may have substantial effects on the Nevertheless, fiscal policy needs to be inevitably based on ability of governments to correctly estimate the overall real-time figures. The paper develops a simple modeling balance and the structural fiscal balance in real time, and framework to formalize these linkages and combines it that the effects may imply substantial debt accumulation. with a newly compiled dataset from the International This paper is a product of the Economic Policy and Debt Department, Poverty Reduction and Economic Management Network. It is part of a larger effort by the World Bank to provide open access to its research and make a contribution to development policy discussions around the world. Policy Research Working Papers are also posted on the Web at http:// econ.worldbank.org. The author may be contacted at eley@worldbank.org. The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about development issues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry the names of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank and its affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent. Produced by the Research Support Team Real-Time Macro Monitoring and Fiscal Policy † Eduardo Ley Florian Misch Keywords: Real-time Output Data, Fiscal Policy, Data Revisions, Public Debt JEL Codes: E01, E62, H68 We thank for their useful comments, Gerhard Kempkes, Jeffrey D. Lewis, Lukas Reiss, the participants of the 68th Congress of the International Institute of Public Finance (IIPF) held in Dresden, Germany, and partic- ipants of a workshop organized by the National Bank of Slovakia held in Bratislava, Slovakia. Financial support from a World Bank Research Support Budget grant is gratefully acknowledged. Poverty Reduction and Economic Management (PREM) Network, The World Bank, Washington DC, USA. † Centre for European Economic Research, Mannheim, Germany. “It does make economic sense to target cyclically adjusted rather than actual deï¬?cits. But the improvement in economics is at the cost of a reduction in precision. Nobody knows what a structural deï¬?cit is.â€? —Martin Wolf, Financial Times, March 6, 2012. “The most important product of knowledge is ignorance.â€? —David Gross, 2004 Physics Nobel laureate. 1. Introduction This paper assesses the implications of output data revisions on the overall and the struc- tural ï¬?scal balances, as well as for debt accumulation. The ability to correctly observe and predict the ï¬?scal balance is of central importance for budgetary planning and for ï¬?s- cal surveillance. The public-debt dynamics are driven by the overall ï¬?scal balance, which is therefore the main reference indicator in assessing ï¬?scal sustainability. The structural balance, in turn, which is calculated as the difference between the overall balance and the cyclical balance, relates to another key aspect of assessing the governments’s ï¬?scal policy stance and is a key ingredient of ï¬?scal surveillance as it separates the impact on the budget of changes in the economic environment from that of discretionary policies. However, budgetary planning and ï¬?scal surveillance may be impeded by fundamental data uncertainty, resulting in frequent output data revisions: as newer and better information becomes available, GDP ï¬?gures are revised so that real-time GDP ï¬?gures rarely correspond to ï¬?nal GDP ï¬?gures. Previous papers have documented revisions of real-time GDP estimates and their impli- cations for output gap estimates; see for instance Orphanides and van Norden (2002) using U.S. data, Marcellino and Musso (2011) using Euro area data, and Ley and Misch (2013) using IMF World Economic Outlook (WEO) data covering a large number of countries. Often, output revisions may be so large that in many cases, governments may be unaware whether a given country is in a recession or not, or what the sign of the growth rate is. Such revisions of output data are obviously a major concern for ï¬?scal policy, and con- ceptually, there are several channels through which such revisions translate into revisions of real-time estimates of the overall and structural balances prior to the ï¬?scal year. First, changes in output automatically induce changes in the volume of public spending and 1 revenue. Even if spending and revenue output elasticities were known with complete cer- tainty, inaccurate real-time information of the output growth therefore implies that it is not possible to correctly estimate ï¬?scal revenue and expenditure streams prior or even dur- ing the ï¬?scal year. Second, ï¬?scal indicators are frequently reported as shares of GDP. As output revisions occur, the denominator of these ratios change as well. Third, estimating the structural balance requires knowledge about the magnitude of the output gap which is also constantly revised, given revisions of output data. Depending on their magnitude, revisions of the overall and structural balances represent signiï¬?cant challenges. On the one hand, with revisions of the overall balance, governments inevitably miss deï¬?cit targets so that depending on the nature of output data revisions, ï¬?scal policy is either too tight or results in unplanned debt-to-GDP accumulation. On the other hand, ï¬?scal surveillance becomes considerably more difficult and less reliable. There is rapidly growing literature focusing mostly on developed countries that documents revisions of ï¬?scal data, i.e., the difference between ï¬?scal projections or forecasts, and outcomes in terms of actual spending and revenue. Recent evidence suggests that such revisions may in some cases be sizable; see for instance Beetsma et al. (2012) who reports that on average, the overall budget balance is downward revised by 0.5%. Hallett et al. (2011) ï¬?nd that real-time estimates of structural balances that are published by the OECD for 19 countries are a poor predictor of of episodes of signiï¬?cant ï¬?scal loosening. Much of the earlier literature, summarized in great detail by Leal et al. (2008), discusses the properties of ï¬?scal forecasts and in particular examines whether there is a systematic bias, whereas the more recent literature summarized by Cimadomo (2011) also includes an assessment of why such revisions occur. Obviously, there may be many reasons of why observed revisions of ï¬?scal policy data occur. In this paper, we solely focus on one particular factor that causes revisions of ï¬?scal indicators, namely output data revisions which is the central impediment of the ability of governments to correctly report ï¬?scal indicators. This factor contrasts with other political or institutional factors for revisions related to the willingness of government to avoid ï¬?scal data revisions, such as strategic and deliberate misreporting or mis projecting by gov- ernments. Auerbach (1995) distinguishes three types of factors that result in ï¬?scal data 2 revisions: economic errors which relate to forecast errors of macroeconomic variables used, policy errors which are mainly of political nature, and technical errors which relate to the forecasting model used including its underlying assumptions about behavioral responses. Most existing papers that examine the determinants of revisions to ï¬?scal data contribute these to the ï¬?rst two factors, namely both budgetary institutions and other political factors as well as to revisions of GDP growth projections; see for instance Strauch et al. (2004), Pina and Venes (2011), and de Castro et al. (2011) for evidence from developed countries o and Poplawski-Ribeiro (2011) and Frankel (2011) for rare examples that include and Lled´ both developed and developing countries. Easterly (2012) suggests that growth slowdowns coupled with overoptimistic growth projections are key to explain the rapid accumulation of debt in Europe and the US. alez-M´ Similarly to Kempkes (2012) and Gonz´ ınguez et al. (2003), we exclusively use real-time and output data to infer revisions of ï¬?scal indicators as a share of GDP.1 Using estimates of the bias of the real-time output gap of the EU-15 countries, Kempkes (2012) roughly estimates the implications for deï¬?cits in a hypothetical scenario in which a ï¬?scal rule prescribes a deï¬?cit target that is inversely related to the output gap (i.e., that allow a higher deï¬?cit in times of recession and vice versa). He multiplies the estimated bias with country-speciï¬?c cyclical elasticities of the budget balance and ï¬?nds that under these assumptions, the deï¬?cit projected using real-time data would on average exceed the deï¬?cit under ï¬?nal data by 0.5 to 0.6 percent of potential output in the EU-15 countries. Similarly, alez-M´ Gonz´ ınguez et al. (2003) discuss the effects of revisions to the output gap to revisions 2 of the structural balance. Our approach differs from these papers, and we make two contributions. The ï¬?rst one is an analytical one: we develop a simple but comprehensive modeling framework which 1 Given that we focus on revisions of ï¬?scal variables in terms of GDP, we ignore revisions of other macroe- conomic variables, such as inflation. 2 Other papers focus on different aspects that affect output gap estimates and that thereby have implica- tions for ï¬?scal surveillance. Larch and Langedijk (2007) for instance discusses the implications of using alternative smoothing parameters for output gap estimates in the context of EU budgetary surveillance. 3 we use to simultaneously study the implications of output data revisions for deviations from a given target for the overall balance, for revisions of the structural balance from original estimates, and for unwanted debt accumulation as a result of mistakes in budgetary planning. This framework reflects all three transmission channels through which output data revisions affect estimates of the overall and the structural balances prior to the ï¬?scal year, namely (i) revisions in growth which affect estimates about the magnitude of the cyclical component of revenues and expenditures, (ii) revisions of the level of GDP as the denominator which affects the shares of ï¬?scal variables in GDP, and (iii) revisions of the output gap, which, in addition to the two other factors, also affect the accuracy of estimates of structural balances. An important assumption of the model is that the government anchors its estimate of the a period’s ï¬?scal balance on the previous period’s ï¬?scal balance which limits the effects of output data revisions and thereby makes our results more credible. We essentially distinguish two ‘release’ dates of ï¬?scal data which differ with regard to the availability of ï¬?nal output data. We then compare the concurrent estimates of the overall and structural balances using output estimates available at the time of budget preparation with the ones computed when all ï¬?nal release output data is available and derive an expression about how revisions of the overall balance affect public debt over time. The theoretical framework will set the groundwork for the simulation exercise to quantify the effects of output revisions on the overall balance, the structural balance and debt accu- mulation using a novel dataset which is our second contribution. Our data come from the IMF’s World Economic Outlook (WEO), comprising real-time output growth and gap data for 1990–2007 and ï¬?nal output data from 2012 covering 169 countries.3 This allows us to obtain for every country and every year for 1990–2007 a real-time and ï¬?nal estimate of out- put and of the output gap, which together with few assumptions on structural parameters, enables us to compute revisions of ï¬?scal balances and to simulate debt accumulation. We essentially assume that the government is benevolent and mechanically uses output data 3 In a companion paper (Ley and Misch, 2013), we present the dataset in detail and include various more detailed descriptive statistics on output growth and output gap revisions. 4 available in real time to make their predictions. We ï¬?nd that while on average revisions of ï¬?scal balances are small, in a relatively high share of cases, these revisions are nevertheless likely to create signiï¬?cant challenges for ï¬?scal planning and ï¬?scal surveillance. The advantage of this type of simulation exercise is that we are able to evaluate the effects of one particular and potentially important cause of ï¬?scal revisions that determines government ability and ‘switch off’ all other factors, in particular strategic and political considerations of governments to wrongly estimate ï¬?scal indicators. For our purposes, WEO data is particularly suitable, even though governments of more advanced countries that may use higher frequency data from national sources for ï¬?scal policy. The WEO predictions are the result of a comprehensive and systematic procedure. The country desks, in consultation with country governments and other observers, submit their forecasts to the WEO division. The WEO division makes sure that ‘the pieces ï¬?t in’, checking the compatibility of the forecasts between countries that have signiï¬?cant trade, or share signiï¬?cant trade partners. Several iterations with individual desks may occur before it is settled on the published WEO (spring and fall). In addition, compared to government GDP forecasts which are predominantly used by papers analyzing the determinants of revisions of ï¬?scal indicators, WEO data is likely to be less affected by political interference.4 Together with the fact that release dates are identical across countries, this implies that the WEO data is therefore probably the best comparable source of real-time data available for a large number of countries. This feature of the data allows us to obtain credible measures of government ability, in contrast to willingness, of correctly projecting ï¬?scal balances. Given the large number of country-year observations, we are able to draw general conclusions for different country groups. The results have important policy implications. For instance, the results sug- gest that countries may deviate from ï¬?scal plans even if governments are benevolent and 4 There are no systematic differences between WEO data and other sources that are deemed reliable; Timmermann (2007) and Abreu (2012) for instance suggests that the quality of consensus forecasts and WEO forecasts is similar. Irrespective of the relatively high quality of WEO data, Aldenhoff (2007) and Dreher et al. (2008) still ï¬?nd evidence of political interference, and recently, Blanchard and Leigh (2013) show that during the global ï¬?nancial crisis, forecasters underestimated the magnitude of ï¬?scal multipliers. 5 want to stick to ï¬?scal targets which is important for whether to impose sanctions or not. In addition, our results provide guidance for ï¬?scal contingency plans in terms of safety margins. The paper is organized as follows. Section 2 contains a description of the data. Section 3 develops the conceptual framework. Section 4 presents the simulation results and Section 5 concludes. 2. A First Look at the Data As noted, the data comes from the IMF World Economic Outlook and consist of output data from 1966 to 2012 for 169 countries.5 The dataset contains output ï¬?gures released in spring and fall from 1990 to 2012—i.e., real-time output ï¬?gures from 24 different vintages where we consider the one from fall 2011 as the one that contains ï¬?nal ï¬?gures, with some exceptions. We clean the data for outliers to ensure that our simulation results are not driven by extreme output revisions or output gaps.6 We use the perspective of the previous-year fall WEO which is the latest data available for the preparation of the budget provided that the ï¬?scal and calendar years are congruent to appraise current year’s GDP. Thus, for example, for 1991, we focus on the fall 1990 forecast for 1991 which we shall call concurrent estimate. In a robustness test however, we use same-year spring estimates which should, in principle, be more accurate. The justiï¬?cation for using the same-year April WEO vintage would be that if the level of economic activity could be reliably assessed from this perspective of the ï¬?rst quarter, then corrective measures could perhaps be implemented to re-direct ï¬?scal policy for the 5 We discard 9 countries where output ï¬?gures appear to systematically not be revised for several years or where the output series from different vintages do not allow us to compute all variables required for the simulation in Section 4. 6 For instance, due to unusually large shocks such as wars or natural disasters where revisions or output gap estimate. In particular, we discard all observations where the absolute estimated and ï¬?nal output gap is larger than 25 percent, and where the absolute deviation of the ï¬?nal growth rate and the estimated concurrent growth rate from the median of the ï¬?nal growth rate is larger than 25 percentage points. In addition, we discard vintages with fewer than 30 observations which we consider as reasonable to compute trend output which amounts to less than 0.2% of observations in our dataset. 6 remaining of the year. We shall see that is not so. Final estimates are assumed to be available ï¬?ve years later, so the 1990–2012 data allows us to study the reliability of the 1990–2007 concurrent estimates. The left panel in Table 1 summarizes descriptive statistics of ï¬?nal growth rates by country groups. During the period considered, overall growth averages to an annual four percent, with signiï¬?cant smaller dispersion among high-income OECD countries, and larger disper- sion among LICs. The right panel in Table 1 summarizes the revisions of the growth rate in percentage points which is the difference between the ï¬?nal ï¬?gure and concurrent estimates. Given that positive and negative revisions tend to cancel each other out, the mean is close to zero in most instances, but in 50% of the cases, that is below the 25th and above the 75th percentiles, the absolute revision is above 1.75%. The median revision in absolute value is almost two percentage points of GDP (not shown). The standard deviation of the revisions for all country groups is about the median value of the ï¬?nal growth rates, almost the same order of magnitude as the standard deviations of the ï¬?nal growth ï¬?gures, and larger than the standard deviations of the original projections (not shown here). Table 1. Final growth rates and growth revisions (169 countries: 1990-2012; N = 2804) Final output gaps (in % of GDP) Revision to output gaps (in % of GDP) Quartiles Moments Quartiles Moments Country Group 25 50 75 Mean StDev 25 50 75 Mean StDev High income: OECD 1.53 2.97 4.25 2.97 2.52 -1.08 0.05 1.15 -0.08 2.21 High income: nonOECD 2.02 4.69 7.24 5.12 5.02 -1.62 0.61 2.60 0.76 4.36 Upper middle income 1.53 4.08 6.40 3.74 4.50 -2.33 0.20 2.41 -0.15 4.31 Lower middle income 2.30 4.48 6.61 4.51 4.52 -1.88 0.01 1.79 -0.07 4.15 Low income 1.70 4.51 6.59 3.88 5.26 -2.87 -0.53 1.27 -1.09 5.09 All countries 1.84 4.03 6.24 3.99 4.49 -1.91 -0.01 1.76 -0.21 4.18 Source: WEO data and own compilation We compute the output gap for each vintage. While there are alternative ï¬?ltering meth- ods, we use the Hodrick-Prescott (HP) ï¬?lter, which is the most common method used, to extract the trend from a time series. In addition, we show in Ley and Misch (2013) that differences across ï¬?lters tend to be fairly small. The left panel in Table 2 summarizes descriptive statistics of the ï¬?nal output gaps—i.e., the gaps calculated using data from 7 the latest vintage. Ignoring the sign, and just looking at their absolute value, the median values would be about one percentage point of GDP. However, there is some variation from 0.8 percentage points of GDP for high-income OECD countries to 1.5 percentage points of GDP for upper-middle income countries. Table 2. Final output gaps and output gap revisions (169 countries: 1990-2012; N = 2804) Final output gaps (in % of GDP) Revision to output gaps (in % of GDP) Quartiles Moments Quartiles Moments Country Group 25 50 75 Mean StDev 25 50 75 Mean StDev High income: OECD -0.81 -0.05 0.96 0.07 1.52 -0.48 0.17 1.02 0.28 1.32 High income: nonOECD -1.32 -0.02 1.48 0.04 2.81 -1.08 0.31 1.35 -0.10 2.83 Upper middle income -1.43 0.04 1.63 0.07 2.73 -1.01 0.21 1.65 0.29 2.41 Lower middle income -0.94 0.02 1.06 0.04 2.58 -0.74 0.20 1.14 0.18 2.35 Low income -0.91 0.03 1.26 0.02 2.96 -0.82 0.20 1.29 0.15 2.89 All countries -1.01 -0.00 1.27 0.05 2.57 -0.76 0.20 1.27 0.19 2.39 Source: WEO data and own compilation The right panel in Table 2 represents the summary statistics of the revisions to the output gap. Most of the statistics are of the same order of magnitude than those in the left panel. This is a conï¬?rmation of the ï¬?nding, for the U.S., by Orphanides and van Norden (2002); namely that the uncertainty in the real-time estimates of the output gap is about the magnitude of the gap itself. In absolute value, roughly for a gap around one percentage point of GDP, the typical revision will be ±1 percentage points of GDP. Figure 1 illustrates the uncertainty intrinsic in output-gap estimates.7 The concurrent output gaps correspond now to the previous-year’s fall WEO data while ‘ï¬?nal’ output gaps correspond to the most recent WEO in the sample. The correlation coefficient is a rather low (0.39). Note that the concurrent estimates are computed with the data that become available at the end of the ï¬?rst quarter of the same year. In more than one-third of the cases, the output gap changes its sign from the concurrent estimate to the ï¬?nal one (Figure 2). 7 Trend output is estimated here, in this example, with a Hodrick-Prescott ï¬?lter and a parameter value of λ = 6.25. The results are similar when other standard univariate ï¬?lters are used to estimate trend output. 8 q 20 q q q q q q q 10 qq q q q q q qq q q q q qq q qqqq q q q q q q q q q qq q q q q q q q qq qq q qqq q q q q q q q qq qqq qqq qqq q qqq q qq q qq qq q qq q qq qqq qqqqq qq q qqq qqq qq q q q q qq q qqqqq q qq qq q qq q q q q qq q q q qq q qq qq q q qq q q q qqq q qqq q q q qqq qqq q q Final gap q q qqq q qqq q q qq q q q qq q qq q qqq qq qqq q q qq q q q qq qq q q q q qq q q q qq q q q qq q q qq qq q qqq qqqq q q q q qqqq q qqq q q qq q q qq q q q q qq q q q qq q qq q qq qq qqq q q qqq qq q qq qq q qq q q qq q q q qq q q qq q q qq q q q qq q q q qq q qq q qq q qq qq qq q q q qq q qq q qq q qq q qq qq q q qq qq q q q q qq q q qq q q qq q q qq q q qqq qqq q q qq q qqq qq q qq q qq q qq qq q q qq q qq q qq q qq qq q qqqq q q q qq q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q qq q q qq q q q q q q qq q q q q q q q q q q qq qq q q q qq q q qqq qq q q qqq q q qq q qqqq 0 qq qq qq q qq qq qq q q q q qq q q qq q q q qqq q q q q qq q q q qq q qq q q qq q q q qq qq q q qqqq qqq qq q qq qq q q qqq qq q qq q q q qq q qq q q q qqq q q q qq q q qq q q q qq q q q qq q q qq qq q qq qq q q q q q qq qqqq q q q qq qq q q q q q q q q q q q q q q q q q q q q q q q q qq q q q q q q q q q q q q q q qq q q q q q q q qq q q q q q q q q q q q q q q q q q q q q q q qq q q q q q qq q qq qq q qq qqq q q qq q qq q qq q qq q q qq q q qq q q qq q q q qq q qq qqqqq qqqq qq q q qq qq q q q qq q q q q q q q q q q q q q q q q q q q q qq q q q q q q q q q q q q q q q q q q q q q q q qqq q q qq q qq q qq qq q q qq qq q q qq q qq qq qq qq qqq qqq q qqq qq qq q qq q q q qq q qq qqq qq qq q q q q q q q q q q q q q q q qq q qq q qq q qqq q q q q q q qq q q q q q q qq q q q q q q q qq q q qq q q qqq q qq q qqq q q q qq q qqq q q q q qq q q qq qq q q qq q q q q q −10 q q q q qqq q q q q q q q q q q q q q −20 q −10 −5 0 5 10 15 20 Predicted gap Fig. 1. Scatter Plot of the Output Gap: Concurrent vs Final (169 countries: 1990-2010) ⊕: 64% cases ⊕ 44% time; of which, the Final Gap is : 36% cases (changes sign) Predicted Output Gap ⊕: 39% cases (changes sign) 56% time; of which, the Final Gap is : 61% cases Fig. 2. Effect of GDP Revisions on the Output Gap: Concurrent vs. Final (169 countries 1990-2010) One important reason behind these revisions in the output-gap estimates is, of course, that GDP growth projections get substantially revised. This is shown on the right panel of Table 1, which presents statistics on growth revisions. Apart from the preliminary nature of output data available in real time which naturally result in output gap revision, output gap estimates are also affected by methodological difficulties. Virtually all methods for estimating trend or potential output at time t require future observations for a number of 9 periods beyond t (i.e., the methods are based on backward- and forward-looking symmetric ï¬?lters). In order to estimate the output gap in real time, the government therefore has to rely on truncated ï¬?lters which are suboptimal, and as observations become available, the past trend (and gap) gets revised, often substantially. As a result, and similarly to output, ˆt in real time where zt = z the government observes z ˆt so that the government may even misperceive whether actual output is above or below potential output. 3. Modeling Framework 3.1. Economic Fluctuations and Fiscal Policy In our model economy, aggregate output, yt , grows over time at rate γt , thus yt = (1 + γt )yt−1 , and is subject to short-run, exogenous shocks so that actual output fluctuates ¯t , giving rise to an output gap, zt : around potential output represented by its trend, y yt − y¯t zt ≡ × 100 ≈ [log(yt ) − log(¯ yt )] × 100 (1) ¯t y >0 economic activity is over potential = <0 economic activity is below potential The output gap, zt , deï¬?ned in (1), reflects the cyclical position of the economy: when it is positive then economic activity is above potential, whereas when it is negative then output is below potential, and there are unemployed resources in the economy. We assume that the government spends on public services and social transfers and raises taxes to ï¬?nance these expenditures. Let e represent government expenditure, and let r represent its revenue. The resulting budget balance b is simply bt = (rt − et ), which can be negative or positive, depending on whether the government runs a deï¬?cit or a surplus. Parts of expenditure and revenue automatically respond to changes in output: an increase in economic activity results in higher tax revenue and lower social expenditure, for instance due to a decrease in unemployment beneï¬?ts. The elasticity (strictly, buoyancy) of revenue, Ï?, and of expenditure, , can be written as ∆r r−1 ∆r 1 Ï?= ÷ = × (2) ∆y y−1 r−1 γ ∆e e−1 ∆e 1 = ÷ = × (3) ∆y y−1 e−1 γ 10 The change of the budget balance as a result of changes in output can be written as ∆b = ∆r − ∆e = (Ï? r−1 − e−1 )γ (4) 3.2. Information Available in Real Time Key of our modeling framework is the assumption that governments are imperfectly in- formed about the state of the economy at the time when they need to project future expenditure and revenue streams. In or shortly prior to period t, available output data referring to period t is inaccurate, and ï¬?nal ï¬?gures become only available ex-post, i.e., a ˆt , γ couple of periods after t. As a result, estimates of yt , γt and z denoted by y ˆt , ˆt and z respectively, made in real time need to be revised later on as more information becomes available. We assume that revisions occur at least within four years after period t implying that the ï¬?nal ï¬?gures yt , γt and zt are only available in t + 5 (i.e., ï¬?ve years in the future). Obviously, the magnitude of the revisions depends on in which period the output data was released, i.e., when the estimates of yt , γt and zt were made. For the purpose of this paper, we focus to what we refer to as concurrent or real-time estimates and assume ˆt , γ that y ˆt are released approximately at the end of the third quarter of period ˆt and z t − 1 for two reasons, although we slightly relax this assumption as part of the robustness checks discussed below. On the one hand, supposing that the ï¬?scal year is congruent with the calendar year, this is most likely to be the newest information that governments may use to prepare the annual budget. On the other hand, in our data, these dates roughly correspond to the release dates of the fall estimates of the World Economic Outlook. Revisions to these estimates are typically non-negligible as for instance Figures 1 and 2 suggest, and result in ï¬?nal ï¬?gures to differ signiï¬?cantly from concurrent estimates. We ˆt gets revised by φt × 100 percent express revisions of concurrent output in relative terms: y ˆt (1+ φ). By contrast, and in line with the literature, we express revisions so that yt , is yt = y ˆt ) and (zt − z of the growth rate and of the gap in absolute terms, namely as (γt − γ ˆt ), respectively. 11 As a result of preliminary and possibly inaccurate estimates of γt , the government is unable to correctly project revenue and expenditure for t in real time. In particular, it ˆt , as actual estimates revenue, rt (γ ), and expenditure, et (γ ), using growth estimates, γ expenditure and revenue flows occurring in t are not yet observed. In order to make these estimates more reliable, the government uses revenue and expenditure ï¬?gures from (t − 1) as an anchor which we assume is observed with reasonable accuracy towards to the end of this period.8 While this assumptions may seem optimistic as growth in t − 1 and thereby revenue and expenditure streams are still uncertain, it lowers the effects of output data revisions implying that our simulation results below are not overly pessimistic. Using equations (2)–(3), real-time estimates of revenue and expenditure may therefore be written as r ˆt = r ˆt (ˆ ˆ r = rt−1 + Ï? γ γt ) = rt−1 + ∆ ˆ r−1 (5) and e ˆt = e ˆt (ˆ ˆ e = et−1 + γ γt ) = et−1 + ∆ ˆ e−1 (6) so that ˆ rt − ∆ ˆb = ∆ ∆ ˆ et = (Ï?rt−1 − et−1 )ˆ γt (7) By contrast and for simplicity, we assume that there is complete certainty with respect to the magnitude of the revenue and expenditure elasticities. 3.3. Revisions of the Overall Balance This sub-section derives the effects of output level and growth revisions on the ability of the government to project correctly the overall balance which coincides with the ability of governments to meet a given target for the overall balance. The magnitude of the revision of the overall budget balance in terms of GDP is the variable of interest and can be written 8 We implicitly assume that the government uses cash accounting. Under accrual accounting, this assump- tions would not be reasonable. 12 as as the difference between the actual balance, (b/y ), and the projected / targeted balance, (ˆ ˆ), in t: b/y bt (γt ) ˆbt (ˆ γt ) yt − ˆ bt (γt )ˆ γt )yt bt (ˆ − = (8) yt ˆt y ˆt yt y Equation (8) shows that we take into account another transmission channel through which inaccurate concurrent output estimates affect revisions of the budget balance, in contrast to existing papers. In addition to automatic changes in the budget, as a result of changes in output which the government estimates based on preliminary growth ï¬?gures, we also consider changes in the level of output. The underlying assumption here is that targets of the overall balance are typically expressed in terms of GDP. There is also another difference to Kempkes (2012): he assumes that the target of the overall balance itself is wrongly speciï¬?ed as a result of the forecast bias of output gap estimates, and then implicitly considers the revisions of the original target. By contrast, we consider the deviations from a given (and implicitly correctly set target) using essentially differences between predicted growth rates available at the time of budget preparation and ï¬?nal growth rates released ex-post. Using b = b−1 + ∆b and yt = (1 + γt )yt−1 , (8) can be re-written as bt ˆ bt yt − (bt−1 + ∆ˆ (bt−1 + ∆bt )ˆ bt )yt − = (9) yt ˆt y ˆt (1 + γt )yt−1 y ∆bt + (1 + φt )∆ˆ bt φt bt−1 = − · (10) (1 + γt )yt−1 1 + γt yt−1 yt . Substituting for ∆b and ∆ˆ where yt = (1 + φt )ˆ b using (4) and (7) yields bt ˆ bt ˆt − γ γt − γ ˆt φt rt−1 et−1 φt bt−1 − = · Ï? − − · (11) yt ˆt y 1 + γt yt−1 yt−1 1 + γt yt−1 This expression represents the effects of output level and output growth revisions on revisions of the overall budget balance when expenditure and revenue streams of t − 1 are certain and those of t are not observed. These revisions may either be positive implying that the ï¬?nal deï¬?cit is smaller than projected, or negative, implying that the ï¬?nal deï¬?cit is larger than projected. 13 3.4. Revisions of the Structural Balance This sub-section derives the effects of output level, growth and gap revisions on the ability of the government to correctly report the structural balance in real time. We therefore further split the overall balance into the cyclical balance, bc (z ), that depends on the cyclical position of the economy, and the structural balance, bs , that depends on discretionary ï¬?scal policy and coincides with the overall balance when z = 0. The structural balance cannot be observed and therefore must be calculated as a residual: bs c t (γt , zt ) = bt (γt ) − bt (zt ) (12) Note that we assume that structural balance is both driven by output growth which affects the overall balance and by the output gap which affects the cyclical balance. Using (2) and (3), the cyclical balance can be expressed as bc t (zt ) = zt · [Ï? rt (γt ) − et (γt )] (13) so that the structural balance is a function of γ and z . Changes in the growth rate γ induce automatic changes to public spending and revenue and thereby the overall balance, and changes in the gap z affect the cyclical balance. The difference, as percentage of GDP, between the actual and the estimated structural balance can be written as bs ˆ bs bt ˆ bt bc ˆ bc t − t = − − t − t (14) yt yˆt yt y ˆt yt yˆt where the ï¬?rst difference on the RHS can be obtained from (11). The second difference on the RHS may be written as9 bc ˆ bc Ï? rt − et ˆt − e Ï?r ˆt t − t = zt −z ˆt (15) yt yˆt yt ˆt y 9 From (15) to (16), we substitute for r, e, ˆr, and ˆ e using (2), (3), (5) and (6). From (16) to (17), we ˆ = yt /(1 + φt ). substitute for y = (1 + γt )yt−1 and y 14 Ï? rt−1 (1 + Ï? γ ) − et−1 (1 + γt ) = zt yt ˆt ) − et−1 (1 + γ Ï? rt−1 (1 + Ï? γ ˆt ) −z ˆt (16) yˆt Ï? rt−1 (1 + Ï? γt ) − et−1 (1 + γt ) = zt (1 + γt )yt−1 ˆt ) − et−1 (1 + γ Ï? rt−1 (1 + Ï? γ ˆt ) −z ˆt (1 + φt ) (17) (1 + γt )yt−1 Equation (14), together with (11) and (17), represents the magnitude of the revision of the structural balance that is required due to output data revisions. Again, these equations are based on the simplifying assumption that expenditure and revenue flows in t − 1 are observed towards the end of t − 1 when the projections are made. As a consequence towards the end of period t, bt is known (as expenditure and revenue streams are known) so that bt = ˆ bt . This implies that revisions of of estimates of bs t made towards the end of t are only driven by revisions of zt and yt . Based on (14) and (15), revisions of estimates of the structural balance made towards the end of period t can therefore be written as bs ˆ bs 1 1 zt ˆt z t − t = bt − − (Ï? rt − et ) − (18) yt yˆt yt y ˆt yt ˆt y 3.5. Debt Accumulation as a Result of Output Data Revisions This sub-section examines the implications of output data revisions for debt accumulation over time. Mistakes in budgetary planning due to output data revisions have consequences for the level of public debt: unplanned debt accumulation (reduction) occurs if the planned balance exceeds (is below) the actual balance. In contrast to revisions of the overall balance, actual debt accumulation is not driven by the unit of measurement, i.e., revisions to the level of GDP do not matter. By contrast, for the purpose of budgetary planning and ï¬?scal surveillance, ï¬?scal balances are mostly expressed in terms of GDP. The change in debt due to output data revisions is simply (ˆ bt (γt ) − bt (ˆ γt )). If the latter expression is positive, the stock of debt increases; in the opposite case, the level of debt 15 falls. The change in the level of debt resulting from n of these revisions after period t0 , ∆Dt0 +n , expressed for convenience in terms of ï¬?nal GDP of the nth period, is n ∆Dt0 +n 1 = (ˆ bt0 +s − bt0 +s ) (19) yt0 +n yt0 +n s=1 ˆt , and e Using (2), (3), (5), (6) to substitute for rt , et , r ˆt , respectively, yields n ∆Dt0 +n 1 = γt0 +s − γt0 +s )] [(Ï?rt0 +s−1 − et0 +s−1 )(ˆ (20) yt0 +n yt0 +n s=1 Using yt0 +n = yt0 +n−1 (1 + γt0 +n ) to express the latter expression in terms of rt−1 /yt−1 and et−1 /yt−1 where t = 1 . . . n to facilitate the numerical simulation in the next section yields n ∆Dt0 +n rt0 +s−1 et0 +s−1 ˆt0 +s − γt0 +s γ = Ï? − · n (21) yt0 +n s=1 yt0 +s−1 yt0 +s−1 k=1+s (1 + γt0 +k ) Whether over time debt increases or not is an empirical matter and driven by the nature of output data revisions, in particular the revisions of growth rates, which we turn to below. 4. Simulation 4.1. Parameters This section assesses numerically the effects of output data revisions on revisions of the overall balance, revisions of the structural balance based and on unplanned debt accu- mulation or reduction. From equations (11), (14) and (21), this type of exercise requires assumptions about few structural parameters including the magnitude of the overall bal- ance, cyclical elasticities of revenue and expenditure, and revenue and expenditure shares in GDP. We obtain country-speciï¬?c values on output level, output growth and output gap revisions for which we obtain country-speciï¬?c values from our dataset (we generate output level revisions using the data on output growth revisions). Unfortunately, reliable estimates of revenue and expenditure elasticities are scarce, no- tably for developing countries. In order to address this problem, we proceed as follows. For each of the four parameters and for each country group, we set lower and upper bounds. 16 We then assume that the parameters are equally distributed within this range (although later, we relax this assumption and we draw from a triangular distribution as a robustness check), and then draw values for each parameter and each country group 1,000 times. Note that for each draw, the resulting parameter values are identical across all countries within each country group contrary to the information on output data revisions which we have at the country level. With respect to the elasticities, expenditure elasticities are proportionally related to the level of income, given that social transfers are low or non-existent in many developing countries (Berg et al. , 2009). By contrast, the differences in terms of revenue elasticities between country income groups can be expected to be not that large because direct and indirect taxes always respond to changes in income, although administrative difficulties to exploit the full potential tax base in developing countries weaken the link between income and tax revenue. e (2005) is There are few papers available that present estimates. Girourard and Andr´ the standard reference for OECD countries. For revenue elasticities in high-income OECD countries, their estimates range from 1.10 to 1.42. From this group of countries, we exclude Slovakia, Czech Republic, and Greece which are now classiï¬?ed as high-income countries but were likely to share critical features of middle-income economies in past at the time from which most of the data comes that was used to estimate elasticities. In the latter e (2005) range from 1.01 to 1.29. group of countries, the estimates of Girourard and Andr´ For expenditure elasticities in high-income OECD countries (other OECD countries), their ucker (2012) who uses a estimates range from -0.23 to -0.02 (from -0.06 to -0.02). In Br¨ novel instrument to estimate revenue elasticities in sub-Saharan countries the range is from 0.48 to 2.7. Berg et al. (2009) estimate that the elasticity of the overall budget is 0.2 for sub-Saharan countries. Finally, Martner (2006) estimates revenue elasticities for a number of Latin American countries which are probably somewhat representative for middle-income countries. The range is from 0.31 to 1.95. IMF (2009) makes the assumption that the revenue and expenditure elasticities for middle-income G-20 countries e (2005) are one and zero, respectively, which provides not covered by Girourard and Andr´ additional guidance for our purposes. 17 We use these ranges as guidance for our simulation. However, the evidence base is sketchy, and the literature does not provide representative estimate ranges for the lower and upper bounds for each income group. We therefore have to make in part our own assumptions. We also adjust the ranges in a way that excludes the possibility to draw extreme values which would potentially inflate the size of ï¬?scal revisions that we obtain. For instance, we ucker (2012) as unrealistically high. These ranges are consider the upper estimate by Br¨ summarized in Table 3. Table 3. Ranges of parameter values for the simulations r/y b/y Ï? | | Country Group min max min max min max min max High income: OECD 0.25 0.55 -0.10 0.05 0.90 1.20 0.30 0.02 High income: nonOECD 0.20 0.50 -0.10 0.05 0.80 1.10 0.25 0.01 Upper middle income 0.15 0.40 -0.10 0.05 0.70 1.00 0.10 0.05 Lower middle income 0.10 0.30 -0.10 0.05 0.60 0.90 0.05 0.00 Low income 0.10 0.20 -0.10 0.05 0.60 0.80 0.00 0.00 All countries 0.10 0.55 -0.10 0.05 0.60 1.20 0.00 0.30 Source: based on own assumptions We apply a similar procedure to obtain revenue and expenditure shares which are likewise not available for many country-year observations included in our dataset. We therefore as- sume a lower and upper bound for each country group and assume a uniform distribution. However, in contrast to elasticities, we do not draw revenue and expenditure shares inde- pendently which may result in unrealistic budget balances. Rather, we draw the revenue share and the overall balance and then compute the expenditure share. Table 3 summa- rizes the assumptions with respect to the range of each parameter and each country group. As a robustness check, we assume a triangular distribution which puts less weight on the extreme parameter values. 4.2. Results What are the implications of inaccurate real-time output data for ï¬?scal management? Assume that the government is targeting a speciï¬?c deï¬?cit/surplus, but that the information on the level of economic activity is uncertain and subject to revisions such as the ones 18 reported in Section 2, for each country-year observation in our sample. Assume further a distribution of the values of the revenue and expenditure elasticities as well as revenue and expenditure shares as described in the previous subsection. Drawing country group-speciï¬?c values of these parameters for each country-year observation 1,000 times (so that values are always identical per draw within one country group) and calculating the revisions of the overall balance and the structural balance in each case according to equations (11) and (14) results in the data underlying Tables 4 and 5. Table 4 presents descriptive statistics of the revisions of the overall balance. Overall, as positive and negative revisions tend to cancel each other out, the mean across all country groups is near zero. By contrast, the distribution of these revisions is widely dispersed: in 20% of the cases, that is below the 10th and above the 90th percentiles, the revision in absolute terms is at least 1.18% of GDP. The results show that if a government targets a deï¬?cit of 3% of GDP for instance, the chances are 10% that the deï¬?cit is indeed above 4.26% and 10% that the deï¬?cit is indeed below 1.82%. Both cases are signiï¬?cant deviations from the original target and are likely to represent signiï¬?cant challenges for governments. These results appear to be consistent with existing evidence: using data from the WEO from 2002 to 2007, Cebotari et al. (2009) report that the 10th percentile of revisions of the overall balance is 1.7% across all countries. This implies that the simulated revisions of the overall balance are smaller than observed revisions which makes sense, given that in practice, other factors contribute to ï¬?scal data revisions as well whereas we only consider output data revisions. The revisions of the overall balance are a result of the nature of the output data revisions and the nature of elasticities where both larger output data revisions and larger elasticities imply larger revisions of the overall balance. As a result, there is also heterogeneity across country groups. Revisions of the overall balance are largest in high-income countries which are not members of the OECD. One underlying reason is that in these countries, revisions of growth rates are fairly large as a number of countries in this group are major exporters of oil and other natural resources whose prices are difficult to predict. In addition, given the stage of development, revenue and expenditure elasticities are likewise relatively large. 19 In upper-middle income countries, the dispersion as measured by the 10th and 90th percentiles is likewise relatively large, which is due to relatively large elasticities and sig- niï¬?cant output data revisions. By contrast, while real-time output data is also inaccurate in low-income countries, revenue and expenditure elasticities are likewise low implying that output data revisions do not translate into large revisions of the overall budget. Table 4. Revisions of the overall balance, % of GDP (169 countries: 1990-2012; N = 2805000) Percentiles Moments Country Group 10 25 50 75 90 Mean StDev High income: OECD -1.29 -0.55 0.03 0.60 1.17 -0.03 1.21 High income: nonOECD -1.77 -0.69 0.17 1.01 2.19 0.18 2.04 Upper middle income -1.54 -0.61 0.04 0.66 1.35 -0.05 1.39 Lower middle income -0.95 -0.36 0.01 0.39 1.03 0.07 1.36 Low income -1.12 -0.46 -0.06 0.25 0.74 -0.16 0.90 All countries -1.27 -0.49 0.00 0.51 1.18 -0.01 1.35 Source: WEO data and own compilation Table 5 presents descriptive statistics of the revisions of the structural balance by country group in percent of GDP. Here, the picture is similar compared to Table 4. While revi- sions of the structural balance are also driven by revisions of the output gap, the mean and the dispersion of these revisions are not systematically larger than revisions of the overall balance. Indeed, the dispersion of revisions as measured by the 10th and the 90th per- centiles of the structural balance across all countries is slightly smaller than the dispersion of revisions of the overall balance. This suggests that revisions of the output gap and of output growth may in some cases have different signs so that they cancel each other out, at least to some extent, and that growth revisions appear to be relatively more important for ï¬?scal projections. There is a similar pattern of heterogeneity across country groups with high-income countries that are not members of the OECD and upper-middle incomes showing the largest dispersion revisions of the structural balance, whereas low-income and lower-middle income countries showing the smallest dispersion. As a ï¬?nal step, Table 6 translates the revisions of the overall balance into debt accumula- tion over a 10-year period. We proceed as follows: based on the dataset with 1,000 country group-speciï¬?c parameter draws for each country-year observation, we draw all parameters 20 Table 5. Revisions of the structural balance, % of GDP (169 countries: 1990-2012; N = 2805000) Percentiles Moments Country Group 10 25 50 75 90 Mean StDev High income: OECD -1.45 -0.70 -0.08 0.47 1.04 -0.15 1.11 High income: nonOECD -1.61 -0.71 0.17 1.03 2.20 0.25 1.95 Upper middle income -1.50 -0.69 -0.04 0.56 1.24 -0.10 1.24 Lower middle income -0.89 -0.39 -0.04 0.36 0.99 0.04 1.33 Low income -1.06 -0.49 -0.09 0.21 0.69 -0.17 0.84 All countries -1.26 -0.54 -0.05 0.44 1.12 -0.05 1.27 Source: WEO data and own compilation and output data revisions 10 times for each country group to calculate debt accumulation over a 10-year period using equation (21). We repeat this procedure 1,000 times. Table 6 presents the change of government debt in percent of GDP. Compared to Table 4, the dispersion and the mean across all countries is signiï¬?cantly larger (note that whereas in Table 4, a positive sign denotes a surplus, in Table 6, a positive sign refers to increases in the stock of public debt). In 10% of the cases (i.e., above the 90th percentile), the increase of the stock of debt is at least 3.37% of GDP, whereas in 10% of the case (i.e., below the 10th percentile), the decrease of the stock of debt is at least 3.91% of GDP. While these changes of debt may appear to be small, they are nevertheless signiï¬?cant given that they are solely due to missing deï¬?cit or surplus targets as a result of inaccurate real-time output growth ï¬?gures. There is again heterogeneity across country groups. This time, the 90th percentile in high- income countries and upper-middle income countries is largest across all country groups suggesting that here, the risk of unwanted debt accumulation is largest. By contrast, relatively large surprise debt reductions may also occur: in high-income countries that are not members of the OECD, in 10% of the cases, a surprise debt reduction is at least 8.16% of GDP. The possibility of a surprise debt reduction tends to be lowest in low-income countries where in only 10% of the cases, such a debt reduction is above 0.94% of GDP. 4.3. Robustness Checks We perform two robustness checks to test how vulnerable the magnitude and the dispersion 21 Table 6. Debt accumulation over 10 years, in % of the 10th period’s GDP (169 countries: 1990-2012; N = 50000) Percentiles Moments Country Group 10 25 50 75 90 Mean StDev High income: OECD -3.33 -1.44 0.21 2.23 4.33 0.41 3.17 High income: nonOECD -8.32 -5.28 -2.14 0.86 3.28 -2.34 4.65 Upper middle income -3.48 -2.02 -0.18 1.79 4.26 0.11 3.16 Lower middle income -2.06 -1.15 0.05 1.23 2.19 0.09 1.79 Low income -0.94 -0.06 0.86 1.79 2.81 0.91 1.50 All countries -3.91 -1.66 0.09 1.62 3.37 -0.16 3.27 Source: WEO data and own compilation of the revisions of the overall and structural balances are to our assumptions. First, in Tables 7 and 8, instead of using previous-year fall output data, we use same-year spring estimates to compute the revisions of the overall budget and the structural balance. Here, the rationale for the virtual experiment is that instead of using output data released in fall of the previous year, the government could make use of more recent and hence more up-to-date output information released during the course of the current ï¬?scal year; for simplicity, we still maintain the assumptions that revenue and expenditure streams of t − 1 are observed whereas those of t are not observed. However, our results suggest that even with same-year spring-WEO output data, the dispersion of the revisions of the overall budget balance and the structural balance remain large: in particular, in 20% of the cases, the revisions are in both cases above 1% of GDP. With respect to revisions of the structural balance, some measures of dispersion, such as the 90th percentile across all country groups, appears to even increase. While the use of more accurate output data lowers the magnitude of both output growth and gap revisions, the magnitude of these effects may differ for output growth and output gap revisions, and both types of revisions may have opposing signs thereby exerting opposite effects on the revisions of the structural balance. This may imply that the revisions of the structural balance actually increase. Output gap and output growth revisions may have opposing signs as the output gap is determined by both trend output and actual output which in turn may be affected differentially by output revisions. Second, and similarily, we use same-year fall-WEO output data to compute revisions of the structural balance according to (18) in Table 9. According to our (simplifying) 22 Table 7. Revisions of the overall balance, in % of GDP, same-year spring output data Percentiles Moments Country Group 10 25 50 75 90 Mean StDev High income: OECD -0.78 -0.31 0.19 0.65 1.13 0.14 1.02 High income: nonOECD -1.66 -0.55 0.25 1.07 2.13 0.20 2.27 Upper middle income -1.32 -0.49 0.09 0.65 1.37 0.05 1.29 Lower middle income -0.87 -0.31 0.03 0.40 1.02 0.10 1.27 Low income -1.01 -0.41 -0.04 0.26 0.73 -0.11 0.90 All countries -1.06 -0.38 0.05 0.53 1.19 0.06 1.31 Source: WEO data and own compilation Table 8. Revisions of the structural balance, in % of GDP, same-year spring output data Percentiles Moments Country Group 10 25 50 75 90 Mean StDev High income: OECD -1.08 -0.48 0.06 0.56 1.09 0.00 1.03 High income: nonOECD -1.51 -0.62 0.24 1.07 2.21 0.30 2.07 Upper middle income -1.35 -0.58 0.00 0.58 1.27 -0.04 1.23 Lower middle income -0.88 -0.35 -0.02 0.36 0.96 0.06 1.25 Low income -1.00 -0.47 -0.09 0.21 0.67 -0.16 0.85 All countries -1.12 -0.46 -0.01 0.47 1.13 0.01 1.25 Source: WEO data and own compilation assumptions, towards the end of any period t, expenditure and revenue streams of that period are known so that at this point in time, ˆ bt = bt . Revisions of the structural balance fall, but they still remain large. In more than 20% of the cases, revisions exceed 0.72% of GDP. Table 9. Revisions of the structural balance, in % of GDP, same-year spring output data Percentiles Moments Country Group 10 25 50 75 90 Mean StDev High income: OECD -0.92 -0.48 -0.08 0.27 0.61 -0.12 0.72 High income: nonOECD -1.29 -0.62 -0.08 0.42 1.08 -0.11 1.46 Upper middle income -1.00 -0.47 -0.05 0.32 0.81 -0.07 0.88 Lower middle income -0.63 -0.26 -0.03 0.23 0.67 0.04 1.18 Low income -0.72 -0.32 -0.04 0.20 0.60 -0.07 0.65 All countries -0.87 -0.38 -0.05 0.27 0.72 -0.05 0.99 Source: WEO data and own compilation As a third robustness check, we test whether our results are driven by extreme values of the revenue and expenditure elasticities and shares in GDP. Instead of assuming a uniform parameter distribution, we assume a triangular distribution function which places less 23 emphasis on the extreme parameter values, i.e., the lower and upper bounds. Compared to Tables 4 and 5, the results of Tables 10 and 11 hardly change suggesting that our main results are not driven by extreme draws from the tails. Table 10. Revisions of the overall balance, in % of GDP, triangular distribution Percentiles Moments Country Group 10 25 50 75 90 Mean StDev High income: OECD -1.29 -0.56 0.04 0.60 1.13 -0.04 1.17 High income: nonOECD -1.89 -0.73 0.19 1.05 2.16 0.13 2.17 Upper middle income -1.54 -0.63 0.06 0.70 1.37 -0.04 1.38 Lower middle income -0.97 -0.38 0.02 0.42 1.01 0.05 1.20 Low income -1.14 -0.47 -0.06 0.25 0.68 -0.17 0.87 All countries -1.28 -0.51 0.01 0.53 1.17 -0.02 1.32 Source: WEO data and own compilation Table 11. Revisions of the structural balance, in % of GDP, triangular distribution Percentiles Moments Country Group 10 25 50 75 90 Mean StDev High income: OECD -1.45 -0.74 -0.08 0.46 0.99 -0.17 1.08 High income: nonOECD -1.72 -0.75 0.19 1.07 2.17 0.23 1.99 Upper middle income -1.51 -0.71 -0.04 0.60 1.26 -0.10 1.23 Lower middle income -0.90 -0.43 -0.05 0.38 0.97 0.02 1.17 Low income -1.06 -0.51 -0.10 0.18 0.62 -0.19 0.80 All countries -1.27 -0.57 -0.06 0.45 1.10 -0.06 1.22 Source: WEO data and own compilation As a ï¬?nal exercise, we empirically analyze more in depth the drivers of our results, i.e., the role that different factors play for the magnitude and nature of the overall balance and structural balance revisions, using basic OLS regressions. To this end, we standardize the absolute value of the revisions of the overall and the structural balance and the variables and parameters that affect them, including the output growth, gap and level revisions, ˆ ), θ, and (z − z (γ − γ ˆ), respectively, and of revenue and expenditure elasticities and shares in GDP, as a means, although an imperfect one, to linearize equations (11) and (14). Table 12 presents the results. It shows that revisions of output growth, the level of output have a relatively large effect on the overall balance. With respect to the structural balance, these revisions together with revisions of the output gap, are important relatively to the 24 remaining coefficients. The results further suggest that the expenditure elasticity and the revenue elasticity and the share of expenditure in GDP hardly play any role for our results. The share of revenue in GDP lies somewhere in between these extremes, but the revenue elasticity appears to be also less important. Table 12. Determinants of the revisions of the overall and the structural balances log of standardized absolute value of: [b/y − ˆ ˆ] b/y [bs /y − ˆ bs /y ˆ] (γ − γ ˆ) 0.499* 0.347* (0.000) (0.000) θ 0.481* 0.501* (0.000) (0.000) 0.036* 0.026* (0.001) (0.001) Ï? 0.056* 0.054* (0.001) (0.001) r/y 0.185* 0.184* (0.001) (0.001) e/y 0.029* 0.028* (0.001) (0.001) (z − z ˆ) 0.046* (0.000) Standard errors in parentheses. *p < 0.01 All variables in standardized absolute values. 25 5. Conclusions The difference between same-year GDP projections and the ï¬?nal ï¬?gures released a few years later impairs the ability of policy makers to project correctly the ï¬?scal revenues and expenditures in real time. This ability is driven by revisions to the output data which, in turn, result in revisions of the overall and the structural ï¬?scal balances. We develop a comprehensive theoretical framework that considers three transmission channels through which output data revisions matter for ï¬?scal policy in practice, namely through (i) revisions in original estimates of GDP growth, (ii) the output gap, and (iii) the level of GDP. We consider both revisions to the overall and to the structural balance. Revisions to growth matter given that tax revenue and public spending automatically respond to changes in economic activity. In addition, structural balances are also affected by output gap revisions. Finally, revisions to the level of GDP matter given that ï¬?scal balances are typically reported as shares of GDP. Our simulation results with respect to the magnitude of the revisions of the overall balance and the structural balance may be regarded as a lower bound because we eliminated outliers from the WEO data, and we excluded the recent period of global economic turmoil where output data revisions may have been particularly large. We also took care not to consider extreme values of revenue and expenditure elasticities, and we assumed that the previous- year overall budget balance is observable at the time when the budget is prepared, or when the structural balance is computed. Nevertheless, even under these assumptions, our simulation results suggest that revisions of the overall and the structural balance may be substantial and above 1 percent of GDP in absolute terms in more than one-ï¬?fth of the cases. These results are robust to various changes in the underlying assumptions. Chances are in the same order of magnitude that as a result of these revisions of the overall balance, unplanned debt accumulation or a surprise debt reduction over a period of 10 years of above 3.4 percent of GDP occurs. We have also explored differences between country income groups and found that they are signiï¬?cant. Our results with respect to the simulated overall balance appear to be consistent with evidence on actual revisions of the overall balance provided by Cebotari et al. (2009). 26 Future research could address various immediate extension of this paper: on the one hand, our model predictions with respect to the revisions of the overall balance as a func- tion of output data revisions could be compared with actual revisions of ï¬?scal indicators. Alternatively, our modeling framework could be extended to consider uncertainty about the exact magnitude of revenue and expenditure elasticities. For instance, when predicting the budget balance, governments could wrongly estimating these elasticities in addition to relying on possibly inaccurate output data. This model extensions would be likely to in- flate simulated values of balance revisions, but at the same time, it would make our model more realistic. The results presented here have important policy implications. On the one hand, they caution about taking real-time estimates about the structural balance for the purpose of ï¬?scal surveillance at face value and suggest that they may be wrong by signiï¬?cant margins. On the other, the results suggest that in the context of budgetary planning, governments should take into account that they miss ï¬?scal targets due to output data revisions. This in turn implies that governments may set targets in a way that encompass safety margins in case for instance growth estimates are signiï¬?cantly revised. The bottom line of the paper is that, in real time, the overall and structural balances should be better considered a known unknown instead of wishfully being treated as a known known ï¬?scal indicator.10 What you do not know that you do not know may be sometimes more important that what you certainly know. 10 U.S. Secretary of Defense, Donald Rumsfeld, insightfully remarked: “There are known knowns : there are things we know that we know. There are known unknowns : that is to say there are things that, we now know we don’t know. But there are also unknown unknowns : there are things we do not know we do not’t know.â€? 27 6. References Aldenhoff, F.-O., 2007, “Are Economic Forecasts of the International Monetary Fund Po- litically Biased? A Public Choice Analysis.â€? The Review of International Organizations, vol. 2, no. 3, pp. 239-260. Abreu, I., 2011, “International organisations vs. private analysts growth forecasts: an evaluationâ€? Economic Bulletin, vol. 17, no. 2, pp. 2344. Auerbach, A., 1995, “Tax Projections and the Budget: Lessons from the 1980’s.â€? American Economic Review, vol. 85, no. 2: 165-69. Beetsma, R., B. Bluhm, M. Giuliodori, and P. Wierts, 2012, “From budgetary forecasts to ex-post ï¬?scal data: exploring the evolution of ï¬?scal forecast errors in the EU.â€? Contemporary Economic Policy, forthcoming. Berg, A., N. Funke, and A. Hajdenberg, 2009, “Fiscal policy in sub-Saharan Africa in response to the impact of the global crisis.â€? IMF Staff Position Paper, 2009. Blanchard, O., and D. Leigh, 2013, “Growth Forecasts and Fiscal Multipliers.â€? IMF Working Paper, no. WP/13/1. ucker, M., 2012, “An instrumental variables approach to estimating tax revenue elastic- Br¨ ities: Evidence from sub-Saharan Africa.â€? Journal of Development Economics, vol. 98, no. 22. Cebotari, A, J. M. Davis, L. Lusinyan, A. Mati, P. Mauro, M. Petrie, and R. Velloso, 2009, “Fiscal Risks. Sources, Disclosure, and Management.â€? Fiscal Affairs Department, IMF. Cimadomo, J., 2011, “Real-time Data and Fiscal Policy Analysis: a Survey of the Litera- ture.â€? Federal Reserve Bank of Philadelphia Working Papers, no. 11-25. Croushore, D., 2011, “Frontiers of Real-Time Data Analysis.â€? Journal of Economic Liter- ature, vol. 49, no. 1, p. 29. erez, and M. R. Vives, 2011, “Fiscal data revisions in Europe.â€? ECB de Castro, F., J. P´ Working Paper, no. 1342. 28 Dreher, A., S. Marchesi, and J. Vreeland, 2008, “The politics of IMF forecasts.â€? Public Choice, vol. 137, pp. 145-171. Easterly, W., 2012. “The Role of Growth Slowdowns and Forecast Errors in Public Debt Crises.â€? in: “Fiscal Policy After the Financial Crisis.â€?, eds. A. Alesina and F. Giavazzi, forthcoming. Frankel, J. 2011. “Over-optimism in Forecasts by Official Budget Agencies and Its Impli- cations.â€? Oxford Review of Economic Policy., vol. 27, no. 4. pp. 536-562. e, 2005, “Measuring cyclically-adjusted budget balances for Girouard, N., and C. Andr´ OECD countries.â€? OECD Economics Department Working Papers, no. 434. alez-M´ Gonz´ ınguez, J., and P. de Cos, 2003, “An analysis of the impact of GDP revisions on cyclically adjusted budget balances (CABS).â€? Documentos Ocasionales - Banco de na, no. 9. EspaËœ Hallett, A. H., R. Kattai and J. Lewis, 2012, “How Reliable are Cyclically Adjusted Budget Balances in Real Time?â€? Contemporary Economic Policy, vol. 30, no. 1, pp. 75-92. International Monetary Fund, 2009, “Companion PaperThe State of Public Finances: Out- look and Medium-Term Policies After the 2008 Crisis.â€?, available at www.imf.org. Kempkes, G., 2012, “Cyclical adjustment in ï¬?scal rules: Some evidence on real-time bias for EU-15 countries.â€? Discussion Paper, Deutsche Bundesbank, no. 15/2012. Larch, M., and S. Langedijk, 2007, “Testing the EU Fiscal Surveillance: How Sensitive is it to Variations in Output Gap Estimates?â€?, European Economy, no.285. erez, M. Tujula, and J.-P. Vidal, 2008, “Fiscal Forecasting: Lessons from Leal, T., J. J. P´ the Literature and Challenges.â€? Fiscal Studies, vol. 29, no. 3, pp. 347-386. Ley, E., and F. Misch, 2013, “Cross-Country Real-Time Estimates of the Output Gap.â€? unpublished manuscript. o, V., and M. Poplawski-Ribeiro, 2011, “Fiscal Policy Implementation in sub-Saharan Lled´ Africa.â€? IMF Working Papers, no. WP/11/172. 29 Marcellino, M., and A. Musso, 2011, “The reliability of real-time estimates of the euro area output gap.â€? Economic Modeling, vol. 28, no. 4, pp. 1842-1856. Martner, R., 2006, “Cyclical indicators of ï¬?scal policy in Latin American Countries (with special reference to Chile).â€? unpublished working paper; available at http://papers.ssrn.com. Orphanides, A., 2001, “Monetary policy rules based on real-time data.â€? American Eco- nomic Review, vol. 91, no. 4, pp. 964-985. Orphanides, A., and S. van Norden, 2002, “The Unreliability of Output-Gap Estimates in Real Time.â€? Review of Economics and Statistics, vol. 84, no. 4, pp. 569-583. ´ Pina, A.M., and N.M. Venes, 2011, “The political economy of EDP ï¬?scal forecasts: An empirical assessment.â€? European Journal of Political Economy, vol. 27, no. 3, pp. 534-546. Strauch, R., M. Hallerberg, and J. von Hagen, 2004, “Budgetary Forecasts in Europe — the Track Record of Stability and Convergence Programmes.â€? ECB Working Paper Series, no. 307. Timmermann, A., 2007, “An evaluation of the World Economic Outlook forecasts.â€? IMF Staff Papers, vol. 54, pp. 1-33. 30