Can we rely upon fiscal policy estimates in countries with a tax evasion of 15 per cent (or more) of GDP?

Transcription

1 (or more) of GDP? 1 December 2010 Raffaella Basile Ministry of Economy and Finance, Department of the Treasury Bruno Chiarini University of Naples Parthenope Elisabetta Marzano University of Naples Parthenope

2 Data Two novelties The dataset used in this paper is characterized by two statistical novelties consisting in: quarterly government national account estimates for direct spending and net revenues over the period , computed starting from official yearly government data issued by ISTAT; quarterly estimates for the undeclared VAT base over the period , computed starting from yearly time series recently elaborated by the Italian Revenue Agency (Marigliani and Pisani, 2007). 2

3 Quarterly general government estimates Main characteristics of data Availability of alternative data Methodology (disaggregation model, definition of input data) Results and comparison with official time series

4 Quarterly general government estimates Two-side disaggregation of government account Note: euro millions at 2000 prices. Government direct spending includes the expenditure for consumption (wages and current purchases of goods and services) and investments. 4

5 Quarterly general government estimates Two-side disaggregation of government account Note: percentage rate over total GDP. Net revenues are found as total revenues minus current and capital transfers to households and firms. Interest payments are not considered, as out of control of policy makers. In the VECM model net revenues are taken as a ratio over total GDP. 5

6 Quarterly general government estimates Main characteristics Both estimates for direct government spending and net revenues: are computed by applying the temporal disaggregation method suggested in Proietti (2006) to yearly observations from the Conto Consolidato delle Amministrazioni Pubbliche issued by ISTAT; refer to the General Government sector, including Central administrations, Local bodies and Social Security funds; are coherent with national accounts rules (only economic transactions are considered and evaluated on the base of the economic accrual principle); are fully comparable with government statistics issued by other NSIs as their institutional coverage and economic definition are coherent with ESA95 rules. 6

7 Quarterly general government estimates Availability of data from several sources Account/ Source Institutional sector Frequency Since Accountancy Principle Estimation method ISTAT EUROSTAT Quarterly GG account CON-ISTAT database Yearly GG account New Cronos database GG Q 1999:I Accrual Indirect: CL (1971) or FE (1981) NM Q 1980:I Accrual Indirect: CL (1971) or FE (1981) when no direct observation is available GG Y 1980 Accrual Direct obs. GG Q 1991:I Accrual Indirect: CL (1971) or FE (1981) OECD Economic Outlook, Stan database GG Q 1965:I Accrual GG Y 1960 Accrual Indirect: Boot, Feibes and Lisman (1971) Direct obs. from national sources BANK OF ITALY Statistic bulletinpublic finance SB M 1967:I Cash Direct obs. MEF-RGS Quarterly Cash Report SS Q 1980:I Cash Direct obs. PS Q 1994:I Cash Direct obs. Legend: GG = General Gov. sector; NM = Non Market sector; SB = State Budget; SS = State sector; PS = Public sector; Y= yearly; Q = quarterly; M = monthly. 7

8 Quarterly general government estimates Time coherence between yearly and quarterly data from ISTAT CETAP CECAP Differences QI QII QIII QIV Year (sum) Year Year Current taxes on income and wealth 29,487 50,388 39,381 66, , ,378 0 Gov. compensation of employees 34,362 35,741 32,583 47, , ,866 1 Gov. fixed capital formation 6,831 8,540 8,818 9,237 33,426 33,426 0 Social contributions receivable 38,141 40,821 43,761 53, , ,968-1 Taxes on production and imports 47,793 48,596 46,098 52, , ,455 0 Total revenues 126, , , , , , , Current taxes on income and wealth 31,356 48,862 41,120 68, , ,815 0 Gov. compensation of employees 34,635 34,871 33,861 53, , ,542 0 Gov. fixed capital formation 6,869 8,578 8,891 9,372 33,710 33,711-1 Social contributions receivable 39,776 42,581 44,940 56, , ,445 0 Taxes on production and imports 47,954 51,230 49,227 54, , ,736 0 Total revenues 130, , , , , , , Current taxes on income and wealth 34,662 58,937 45,512 74, , ,308-1 Gov. compensation of employees 36,373 39,696 36,884 49, , ,889 0 Gov. fixed capital formation 7,821 8,158 8,286 10,527 34,792 34,792 0 Social contributions receivable 40,780 44,020 46,583 58, , ,683 0 Taxes on production and imports 52,875 55,908 52,693 58, , ,181 0 Total revenues 140, , , , , , , Current taxes on income and wealth 36,785 61,378 51,379 84, , ,660 0 Gov. compensation of employees 36,965 36,833 36,920 53, , ,645 0 Gov. fixed capital formation 7,393 8,357 8,701 11,683 36,134 36,134 0 Social contributions receivable 42,067 48,978 49,503 64, , ,772 0 Taxes on production and imports 54,561 57,695 54,338 59, , ,928-1 Total revenues 145, , , , , , ,270 8

9 Quarterly general government estimates Main issues related to fiscal policy data from other sources VAR analysis on fiscal policy for Italy are very few and rely on either: OECD half-year/quarterly estimates (see Marcellino 2002); or cash-basis data from administrative sources (see Afonso and Sousa 2009 and Giordano, Momigliano et al. 2008). Two main kind of shortcomings arise from the use of these data. 9

10 Quarterly general government estimates OECD estimates are interpolated without a guide According to Perotti (2004), OECD quarterly estimates for the General Government are not more informative than yearly data. When official high-frequency estimates are not available from national sources, the OECD computes quarterly estimates by applying the Boot, Feibes and Lisman (1967) procedure to yearly observations. The procedure is a pure mathematical method consisting in minimizing a quadratic penalty function subject to the proper time constraint. 10

11 Quarterly general government estimates Cash-basis versus accrual data According to macroeconomic models with forward-looking agents, committed expenditures and assessed revenues might have a major effects on the behaviour of the economic agents than receipts and payments. The IMF and the EUROSTAT consider the accrual principle as the proper one when measuring fiscal quantities. Also, a relevant stream of literature on fiscal policy relies on genuine quarterly national account GG observations (see Blanchard and Perotti 2002, Perotti 2004, Perotti 2007). On the other hand, the impact of tax changes on consumer spending would occur when the policy change is implemented, not when it is enacted or credibly announced. The response of households consumption spending to current disposable income would be consistent with standard optimizing behavior with some households facing borrowing constraints (Auerbach 2009). In line with this view, a stream of empirical literature relies on cash-basis data (see Giordano, Momigliano et al 2008). 11

12 Quarterly general government estimates Cash-basis versus accrual data In the case of Italy: Monthly data from the monthly Bullettin of Bank of Italy refer to the State budget. Data is available for current spending and capital spending, not allowing for the focus on the direct public spending (consumption and investment spending). Consolidated data for the Public sector is only available since Before that year, quarterly consolidated data issued in the Quarterly Cash Report refer to the State sector. (the Public sector is not perfectly coherent with the corresponding one referring General Government defined by ESA95). Some authors tried to overcome the problem by performing the consolidation of quarterly cash-basis data in the period (see Giordano et al. 2008). Others relied on monthly cash data published by Bank of Italy and the Italian MEF referring to the Central Government (Afonso and Sousa 2009). 12

13 Quarterly estimates Methodology We use the approach based on a first order Autoregressive Distributed Lag (Adl) model proposed in Proietti (2006), assuming the existence of a regression model at high frequency (quarterly level in our case). The model is an extension of standard techniques widely used by NSIs to estimate high frequency (monthly or quarterly) series, such as: Chow-Lin (1971) stationary AR(1) process; Fernàndez (1981) unit root process without drift. The model is univariate; a model-based method based on high-frequency related series; optimal, providing BLU estimates. First, the use of the broad specification avoid the insurgence of spurious regressions between economic time series. Second, the statistical treatment of the model, based on its state space representation points out the role of: diagnostics based on the standardised kalman filter innovations; revision histories. 13

14 Quarterly estimates Methodology: the unifying specification The broader Autoregressive Distributed Lag model of first order nests both the Chow-Lin and the Fernandez methods: y = y x x t =1,..., n. By using the Lag operator, the equation can be expressed as in the following: 1 L y = x ' L y = x u u ' ' t t 1 t 0 t 1 1 t : t ' t t 0 t t NID /1 L t 2 (0, ) t t 0 1 t If the common factor condition ' 1 L y = x 1 L t t 0 t = 1 0 holds, the Adl model yields: By dividing both sides of equation for the common factor, the Adl yields the CL model: If we also impose =1, then the Adl model yields the Fe one. 14

15 Quarterly estimates Methodology: the statistical treatment Following Proietti (2006), the quarterly Adl model (1,1) model with constant term and linear trend: y y m gt x x NID ' ' 2 t = t 1 t 0 t 1 1 t t : (0, ) is expressed in a State Space representation, as the following: y t = t t = t 1 Wt t where and = [ mg,,, ] ' ' ' ' ' Wt = [1, t, xt, xt 1] 0 0 The quarterly SS model is aggregated at yearly level by using a cumulator variable, in order to estimate regression parameters: y y y c c t = t t 1 t, = 1, 0, t = s( 1) 1, =1,2,..., N otherwise That actually transforms the problem of disaggregation into a problem of unobserved variables. Parameters are estimated by the maximization of the LL function, by using the recursive Kalman Filter (augmented version due to de Jong 1991), which provides standardized innovations used to check stochastic properties of estimates. 15

16 Quarterly estimates Methodology: model selection Three models are estimated: the broader Adl model with constant term and deterministic trend; the Cl model with constant term and deterministic trend; the Fe model with constant term. The selection of the suitable model is based on the ability of the models to minimize the revision errors associated to quarterly estimates, in line with Eurostat guidelines. Starting from 1992 we perform a rolling forecast experiment consisting in re-estimating the three models with the addition of one more observation at a time. At the beginning of each year we made dynamic projections for the four subsequent quarters by only using the observations up to the previous year. We then revise the estimates of the last year by re-estimating the model using the new yearly observation. We compute the revision error for each quarter as the difference between projections and revised estimates. We summarize results over the whole forecast period by computing the mean square root revision error for each quarter. 16

17 Quarterly general government estimates Definition of input data YEARLY SERIES 1 QUARTERLY INDICATORS 2 SPENDING SIDE Gov. compensation of employees Gov. consumption expenditure Gov. fixed capital formation Gov. Fixed capital formation excluding divestments REVENUE SIDE Current taxes on corporate income Current taxes on household income Current taxes on income and wealth Social contributions receivable Compensations of employees of non-market sector Consumption expenditure of non-market sector Fixed capital formation of total economy Fixed capital formation of total economy GDP at market prices less compensation of employees of total economy Gross wages and salaries of total economy The sum of estimates for Current taxes on corporate incomes and Current taxes on households incomes Compensations of employees less gross wages and salaries of total economy Taxes on production and imports VAT and other taxes on imports The sum of estimates for the items of Current taxes on income, Taxes on Total revenues production and imports and Social contributions The sum of estimates for the items of Current taxes on income, Taxes on Total revenues net of transfers to household and firms production and imports and Social contributions 1) Source: ISTAT, Conto Consolidato delle Amministrazioni Pubbliche. 2) Source: CON-ISTAT database. Seasonally-adjusted with trading day correction at the source. The estimation period ranges from 1980:I to 2007:IV 17

18 euro million Quarterly general government estimates Definition of input data 40,000 35,000 Gross fixed capital formation Real estate divestments Gross fixed capital formation excluding divestments 30,000 25,000 20,000 15,000 10,000 5, Source of data: ISTAT, Conto Consolidato delle Amministrazioni Pubbliche and Relazione Previsionale e Programmatica, Sezione II, 18

19 Quarterly general government estimates Government consumption spending: results Original series y au c Indicators x Filtered Estimates y t t, filtered lwl ucl Smoothed Estimates y t T, smoothed lwl ucl Estimates for GG final consumption are very precise and lengthen the availability of quarterly data for the GG sector since 1980:I (Fe estimates in lower right panel). 19

20 Quarterly general government estimates Government consumption: comparison with Eurostat estimates 80,000 75,000 70,000 65,000 60,000 55,000 50, I 99 IV 00 III 01 II 02 I 02 IV 03 III 04 II 05 I 05 IV 06 III 07 II GC_Q_FE Government consumption spending is not provided NMCE by ISTAT in the quarterly account of the GG EU_GC_Q_SA sector. The item is published by Eurostat since 1999:I (thin blue line). The CON-Istat database provides the quarterly aggregate for the Non Market sector (red line), that we used as indicator for the estimation of our quarterly time series (bold line). Log differences from the time series by Eurostat range from -0,018 per cent (2007:IV) to 0,016 per cent (2006:QII). 20

21 Quarterly general government estimates Government investments: results Original series y au c Indicators x Filtered Estimates y t t, filtered lwl ucl Smoothed Estimates y t T, smoothed lwl ucl Quarterly estimates for Government investments clearly show an improvement in precision when we consider the time series excluding real-estate divestments earning (compare CL estimates in the lower right panel with the one in the next slide). 21

22 Quarterly general government estimates Government investments excluding divestmemts: results Original series y au c Indicators x Filtered Estimates y t t, filtered lwl ucl Smoothed Estimates y t T, smoothed lwl ucl Fe smoothed estimates in the lower right panel. 22

23 Quarterly general government estimates Gov. investments: comparison with quarterly Istat estimates 10,000 8,000 6,000 4,000 2, I 82 I 84 I 86 I 88 I 90 I 92 I 94 I 96 I 98 I 00 I 02 I 04 I 06 I GK_Q_FE Our quarterly estimates not considering divestment GK_Q_SA outturns (upper bold line) constitute a more meaningful variable compared with the one issued GKGD_Q_FE by ISTAT (red line), which is greatly affected by real-estate divestment earnings in most recent years. 23

24 Quarterly general government estimates Net revenues: results Original series y au c Indicators x Filtered Estimates y t t, filtered lwl ucl Smoothed Estimates y t T, smoothed lwl ucl From the revenues side we obtain precise estimates for several sub-items, especially for Social security contributions. In the case of net revenues we estimate a very relevant variable for economic literature which is not directly issued by ISTAT and EUROSTAT (CL estimates in the lower right panel). 24

25 Quarterly estimates for the undeclared VAT base Methodology (definition of input data) Results Comparison with official time series

26 Quarterly estimates for the evaded VAT base Definition of input data Yearly data on non-reported Value Added Tax base for the period has been recently estimated by the Revenue Agency. The yearly series has been computed by comparing actual values (VAT returns) and theoretical one, estimated on the base of macroeconomic data including: consumption spending of households and non-profit institutions; central government purchases; other expenditure which incurs non-refundable VAT. In order to obtain quarterly estimates for evaded VAT base, we first applied the temporal disaggregation procedure suggested in Proietti (2006) to the yearly series from the Revenue Agency referring to: Total VAT base Declared VAT base In the same fashion of Marigliani and Pisani (2007), we then found the quarterly series of undeclared VAT base as difference between the two. 26

27 Quarterly estimates for the evaded VAT base Definition of input data To disaggregate the yearly time series of Total VAT base we use the quarterly indicator computed as the sum of GDP and net imports. To disaggregate the yearly series of Declared VAT base we use the quarterly series of net indirect taxes (net of contributions on production), which includes actual VAT returns. For a robustness check we also considered the ratio of employees over selfemployed. The quarterly indicators are from the CON-Istat quarterly database, at nominal prices and seasonally adjusted with trading day correction at the source. 27

28 Quarterly estimates for the evaded VAT base Total VAT base: estimation results Total VAT base Main parameter estimates Method (Student's t in brackets) Adl (5.71) (-5.74) Cl (4.97) Fe (5.19) Note: Student s t in brackets. The indicator exerts significant effects on Total VAT base. Results also suggest the common root restriction in equation can be accepted, and therefore the use of the Chow-Lin or the Fernandez model would be advisable. 28

29 Quarterly estimates for the evaded VAT base Total VAT base: Fe estimates Original series y a u c Indicators x Filtered 5Estimates y t t, filtered lwl ucl Smoothed 5 Estimates y t T, smoothed lwl ucl Estimates obtained by the Fe model are very precise. 29

30 Quarterly estimates for the evaded VAT base Total VAT base: Fe estimates KF Innovations u t -2 s.e 2 s.e. 0 Density N(s=0.985) Correlogram ACF Standardised Innovations Also, the standardised KF innovations can be considered a white noise process; they show a symmetric distribution, without any problem of Kurtosis and very close to a Normal. Autocorrelation seems to be absent. 30

31 Quarterly estimates for the evaded VAT base Total VAT base: Fe estimates 1Q 2Q 3Q 4Q Mean percentage revision error Adl Cl Fe Mean revision Error Adl Cl Fe Mean absolute revision error Adl 1, , , , Cl 1, , , , Fe , , , Root mean square revision error Adl 2, , , , Cl 1, , , , Fe , , , The Fe model is the one providing the least root mean square revision error over quarters, therefore we consider the estimates provided by the method as the correct for computing the evaded VAT base. 31

32 Quarterly estimates for the evaded VAT base Declared VAT base: estimation results Declared VAT base Main parameter estimates Method (Student's t in brackets) Adl (0.13) (0.25) Cl (1.59) Fe (1.96) Note: Student s t in brackets. The quarterly indicator referring to net indirect taxes does not exert statistical significant effects on the variable to disaggregate, especially in the Adl model. The common root condition can be accepted and the CL or the Fe model can be considered the proper specifications instead of the broader Adl one. 32

33 Quarterly estimates for the evaded VAT base Declared VAT base: CL estimates Original series c y a u Indicators x Filtered 5 Estimates y t t, filtered lwl ucl Smoothed 5 Estimates y t T, smoothed lwl ucl Smoothed estimates for Declared VAT base are less precise compared with Total VAT base estimates. 33

34 Quarterly estimates for the evaded VAT base Declared VAT base: CL estimates KF Innovations u t -2 s.e 2 s.e Density N(s=0.974) Correlogram ACF Standardised Innovations However, standardized KF innovations present a symmetric distribution and no significant autocorrelation problems. 34

35 Quarterly estimates for the evaded VAT base Declared VAT base: CL estimates 1Q 2Q 3Q 4Q Mean percentage revision error Adl Cl Fe Mean revision Error Adl , Cl , Fe Mean absolute revision error Adl 5, , , , Cl 4, , , , Fe 1, , , , Root mean square revision error Adl 7, , , , Cl 5, , , , Fe 2, , , , Finally, in term of RMSE of revision errors, the CL model outperforms on average the Fe model in the second, third and fourth quarters. 35

36 Quarterly estimates for the evaded VAT base Declared VAT base: estimation results by using two indicators We disaggregate the series of Declared VAT by also considering the ratio of employees over self-employed. The inclusion of this second indicator strengthens regression effects exerted by net indirect taxes in the Fe model. The common root condition is accepted by the data. According to the rolling experiment, CL estimates still out-perform the ones deriving from the Adl and the Fe model. Anyway, no systematic difference emerges between the two CL series (with one and two indicators) and differences vary within a very small range. Therefore we considered estimates found by using the most parsimonious model. 36

37 Quarterly estimates for the evaded VAT base Declared VAT base: estimation results by using two indicators Declared VAT base Main parameter estimates x1 x2 x1 x2 Method (Student's t in brackets) (Student's t in brackets) Adl (0.71) (-0.34) (1.81) (-1.61) Cl (1.33) (1.64) Fe (2.26) (1.99) Note: x1= net indirect taxes; x2= ratio of employees over self-employed. Student s t in brackets. The inclusion of the second indicator strengthens regression effects exerted by net indirect taxes in the Fe model. The common root condition is accepted by the data. 37

38 Thanks for the attention!