National Bank of the Republic of Macedonia. Working Paper. GDP Data Revisions in Macedonia Is There Any Systematic Pattern?

Transcription

1 Naional Bank of he Republic of Macedonia Working Paper GDP Daa Revisions in Macedonia Is There Any Sysemaic Paern? Jane Bogoev 1 Gani Ramadani 2 Absrac: This paper invesigaes he exisence of any sysemaic relaionship beween preliminary esimaes and subsequen revisions of GDP growh raes in Macedonia. Accordingly, we use various saisical ools for esing he 'news' and 'noise' hypoheses and empirically assess if GDP daa revisions are unbiased and efficien and vice versa. The resuls based on differen empirical approaches in general provide mixed resuls abou he predicabiliy of GDP revisions, alhough he empirical evidence ends o incline ha he long-erm and final GDP revisions are efficien and conain new informaion ha is in line wih he 'news' hypohesis. We conclude ha each subsequen revision is unpredicable. Key words: preliminary daa, final daa, revision, GDP, Macedonia JEL classificaion: B22, C22, E60 Sepember Jane Bogoev, PhD, Research Deparmen, Naional Bank of he Republic of Macedonia, bogoevj@nbrm.mk 2 Gani Ramadani, MSc, Research Deparmen, Naional Bank of he Republic of Macedonia, ramadanig@nbrm.mk The views expressed in his paper are hose of he auhors and do no necessarily represen he views of he NBRM.

2 1. Inroducion Cenral banks and oher macroeconomic acors regularly face he problem of inaccurae iniial measuremens of macro variables such as Gross Domesic Produc (GDP). Naional accoun figures ofen undergo revisions and he firs esimae of a cerain quarer's GDP can differ significanly from he subsequen revised esimaes. Thus, using real-ime informaion abou cenral macro economic variables, such as GDP, can be problemaic, especially if real-ime daa ends o be biased and unreliable. GDP daa represens one of he key indicaors in he process of making macroeconomic decisions. Therefore, revisions make he process of decision making difficul because every curren decision relaed o he economic aciviy is based on he preliminary published daa which laer could be revised. Hence, he possibiliy o predic daa revisions, such as deerminaion of heir size and direcion, may be of grea imporance for opimal decision-making. Daa revisions are disinguished by wo polar characerizaions inroduced by Mankiw and Shapiro (1986), namely, he 'noise' and he 'news' characerizaion. Under he noise characerizaion, he revisions are biased, so ha hey are correlaed wih he preliminary esimaes. The preliminary esimaes conain informaion which could be useful in predicing subsequen GDP revisions. Revisions arise from measuremen error in preliminary esimaes. Such error could occur, for example, if he preliminary esimaes are based on unrepresenaive and insufficienly large sample, conain clerical misakes, and so forh. In conras wih he 'noise' characerizaion, revisions are unbiased in he 'news' characerizaion. For example, if GDP esimaes released afer he preliminary esimaes reflec news, here is no correlaion beween he preliminary esimae and he revision erm because he esimae conains all available informaion. Thus, under he 'news' characerizaion, revision is unpredicable. Consequenly, he major aim of his analysis is o examine if he GDP daa revisions are efficien and unbiased, i.e. wheher he 'news' or 'noise' hypohesis applies in he case of he Republic of Macedonia for which differen empirical approaches will be applied. The remainder of his paper is as follows: Secion 2 presens he lieraure review of daa revisions, mosly focusing on advanced economies. Consrucion of daase is described in Secion 3. Empirical mehods for esing he news and noise hypoheses are given in Secion 4. Secion 5 discusses baseline resuls and sensiiviy analysis, whereas he final secion provides he concluding remarks of his research. 2

3 2. Lieraure review The research on real GDP daa revision issues in Macedonia and in he region, according o auhors' knowledge, is limied as a resul of daa issues and difficulies o have access o daa documenaion. Our research in his field is one of he firs sudies in Macedonia; besides some earlier shor analyses by he Naional Bank of he Republic of Macedonia (NBRM) on he size and direcion of GDP daa revisions. 3 A review of earlier research focuses mainly on he sudies of developed counries because of he researchers' abiliy o consruc daabase and ge access o daa documenaion. Many sudies invesigaed he size of revision errors, e.g. Faus e al. (2004). They argue ha revisions o GDP announcemens are quie large in all G-7 counries and many revisions in quarerly GDP growh are over a full percenage poin a annualized rae. The mean revision o GDP growh is posiive for all counries, excep for Japan, indicaing a general endency oward 'pessimism' in iniial numbers. Moreover, Faus e al. (2004) examined he predicabiliy of G-7 GDP daa revisions and found ou ha he degree of predicabiliy varies hroughou he G-7. Thus, he evidence shows ha for he U.S., revisions are very slighly predicable, bu for Ialy, Japan and he UK, abou half of he variabiliy of subsequen revisions can be accouned for by he informaion available a he ime of he preliminary announcemen. For hese counries, hey showed ha revisions reflec, o a significan exen, removal of 'noise' from he preliminary numbers, raher han arrival of 'news'. Palis e al. (2004) exended Faus e al. (2004) analysis o sudy revisions o he Brazilian GDP esimaes. They documened ha revisions o he Brazilian GDP are largely relaive o hose of he G-7 counries. They also found some evidence ha he Brazilian GDP revisions are predicable, which is consisen wih he view ha GDP revisions correc errors in preliminary GDP raher han reflec 'news'. However, hey poin ou ha GDP revisions are far from being enirely predicable. Alhough GDP revisions were he larges only one year following he iniial GDP release, sill hose revisions are nearly unpredicable. Mankiw and Shapiro (1986) already did some earlier research and concluded ha he revisions for he US Gross Naional Produc (GNP) are more or less unpredicable. York and Akinson (1997) analyzed he behavior of revisions for he seven larges OECD counries and 3 See he box eniled "Revision of GDP (size, direcion and predicabiliy)", Quarerly Repor, NBRM, July

4 found ha revisions of GDP growh were large, bu no significanly differen from zero, so here was no sysemaic bias in he preliminary naional accouns figures. Roodenburg (2004) invesigaed he qualiy of he firs Duch GDP releases by using he same echnique as in he sudy of Faus e al. (2004). His findings suggesed ha Duch GDP revisions were also predicable o some exen. Those resuls were reinforced wih he applicaion of he more general sae-space esimaion procedure in he second par of his paper. Daniel Heller Sahlgren (2006), in he firs par of his sudy, invesigaed revisions of Saisics Sweden's firs announcemens of quarerly GDP and found weak evidence suggesing ha some of hese revisions could be prediced. In he second par, he followed Orphanides and van Norden (2002) who examined he accuracy and reliabiliy of alernaive oupu derending mehods in real-ime. He argued ha he main reason for revising he Swedish oupu gap were no he revisions in he naional accouns, bu he uncerainy abou he rend a he end of he sample. Revisions due o changes in he naional accouns were of lower magniude and persisence han he revisions ha occur because of new informaion abou he posiion of he economy in he business cycle. Aruoba (2006) documened he empirical properies of revisions o major macroeconomic variables in he Unied Saes 4. He found ou ha he revisions of variables of ineres do no saisfy simple desirable saisical properies. In paricular, he found ha he means of final revisions are differen from zero, which indicaes ha he iniial announcemens of saisical agencies are biased. Aruoba (2006) also found ha he magniudes of revisions are quie large compared o he original variables. Consequenly, he found evidence of predicabiliy of revisions using he informaion se a he ime of he iniial announcemen, which means ha he iniial announcemens of saisical agencies are no raional forecass. Overall, he assessed empirical sudies provide differen resuls abou he developed economies, whereas he empirical evidence abou he developing and ransiion economies is quie scarce. Also, auhors in heirs empirical sudies use several approaches o es for news and noise hypoheses of daa revisions. Hence, in he reminder of his paper we aim o es if Macedonian GDP daa revisions conain news or noise and wheher any sysemaic paern can be deermined. 4 The analysis focused on eigh variables: growh of real oupu, real final sales, nominal oupu, inflaion based on oupu deflaor, unemploymen rae, levels and growh raes of employmen, capaciy uilizaion and indusrial producion. 4

5 3. Daa In order o analyze year-on-year GDP growh revisions, we use daa from Sae Saisical Office (SSO) of Macedonia. Our daa sources include GDP press releases, beginning wih he press release for 2002Q1. The las recorded press release in his paper ends wih he period 2011Q3. The daa on he Macedonian GDP and is componens (invesmens, household and public consumpion, expors and impors) are subjec o many revisions before publishing he final daa. The firs daa on he quarerly GDP is published approximaely 75 days afer he end of he reference quarer, while he final daa are released wih longer ime lag. For insance, he provisional daa for annual naional accouns are being published approximaely 10 monhs afer he end of he reference year, while he full se of naional accouns wih final daa is being published 15 monhs afer he end of he reference year. Alhough he SSO has no official calendar for implemenaion of he revision process, he quarerly daa on GDP are revised almos in each new quarerly press release. We suspec ha hese esimaes may conain some measuremen error and may differ from he final daa. Throughou his paper for he GDP preliminary and revised esimaes and GDP revisions we use he following definiions and noaions: P is a preliminary GDP year-on-year growh esimae (release) a ime ; n P is a revised GDP year-on-year growh esimae (release) a ime +n, where n = 1,2,3. f P is a final GDP year-on-year growh esimae (release) a ime f; n R is a difference beween he GDP revised esimae a ime n and he preliminary GDP esimae (release) a ime called GDP revision. Consequenly, in his paper we define: a) shor-erm GDP revision ha is a difference beween he revised GDP esimae afer one quarer and he preliminary GDP esimae: 1 1 R P = P - ; b) long-erm GDP revision ha is a difference beween he revised GDP esimae afer seven quarers 5 and he preliminary GDP esimae: 7 7 R P = P - ; 5 This is he las announcemen of revised GDP esimae published in quarerly press releases. 5

6 c) final GDP revision ha is a difference beween he revised GDP esimae afer more han seven quarers (reaed as final esimae from SSO) and he preliminary GDP esimae: f R = f P - P 4. Tesing he 'news' and 'noise' hypoheses This secion presens he major saisical ools for esing he 'news' and 'noise' hypoheses. In oher words, his secion describes how we can empirically assess if GDP daa revisions are unbiased and efficien. In ha respec, we es for he following five hypoheses: 1. If he GDP revised esimaes are unbiased and conain new informaion, he regression line of he scaer plos beween preliminary and revised GDP esimaes will lie on he 45-degree line, i.e. on he so-called "line of perfec forecass" (see Mincer and Zarnowiz, 1969). Oherwise, if i inersecs wih he "line of perfec forecass", he revised GDP esimaes are biased and may conain measuremen errors. Moreover, addiional indicaion of bias in he GDP revisions may be observed in he regression line from he scaer plos beween GDP preliminary esimaes and GDP revisions. Any sysemaic paern of his line ha is differen from zero, could suggess bias in he revisions, hus supporing he 'noise' hypohesis. 2. If he GDP revisions are unbiased, he mean of he revisions (μ r ) is expeced o be equal o zero: E (μ r ) = 0. This is se according o he Mincer and Zarnowiz (1969) forecas efficiency hypohesis arguing ha when he mean of he revisions is zero, he revised values of he preliminary esimae of GDP lie on he so-called "line of perfec forecass", indicaing ha GDP daa revisions conain 'news'. Oherwise, if he mean of he revisions is differen from zero, he revisions are biased and may conain 'noise'. 3. If he revisions are efficien and unpredicable, he revisions and preliminary GDP esimaes should be uncorrelaed: corr ( P, n R = 0). This is derived from he Mincer and Zarnowiz (1969) and Mankiw and Shapiro (1986) argumens abou he 'news' hypohesis. More precisely, if he revisions of GDP conain new se of informaion ha was no available a he ime when he preliminary esimae was made, he revision should be uncorrelaed wih he preliminary esimae of GDP. In conras, according o he 'noise' hypohesis, if he revision is correlaed wih 6

7 he preliminary GDP esimae, i is an indicaor of 'noise' in he GDP daa revisions (Mankiw and Shapiro, 1986; Faus e al., 2004 and Aruoba, 2006). Moreover, according o Mankiw and Shapiro (1986) and Aruoba (2006), anoher way o es he 'news' hypohesis of efficien GDP forecass is ha he GDP revision and he revised GDP esimae should be correlaed. More precisely, if he revised esimae of GDP conains new informaion ha was no available when he previous GDP esimae was made, he revised esimae of GDP should hen be correlaed wih is laes revision. 4. Under he 'news' hypohesis, if he preliminary release of GDP is efficien forecas, ". he variance of he subsequen esimaes increase. Efficien forecass are necessarily smooher han he objec being forecas." (Mankiw and Shapiro, 1986, p. 22). This implies ha he variance of he subsequen revisions of he GDP preliminary esimaes should increase and vice versa, for he 'noise' hypohesis. 5. The predicabiliy of he GDP revisions can also be esed by using a regression analysis. In respec of 'news' hypohesis, if he GDP revisions conain 'news' and herefore, are efficien forecass, he revisions should be unpredicable (Mincer and Zarnowiz, 1969 and Mankiw and Shapiro, 1986). The raionale behind hese argumens can be explained wih he following equaions: P = f P + ε (1) where: preliminary GDP esimae ( P ) equals final GDP esimae ( P ) plus a disurbance erm ε (Faus e al., 2004). In he case of efficien forecas (he 'news' hypohesis), he disurbance erm should be whie noise process and orhogonal o he preliminary GDP esimae. In he case of 'noise' hypohesis, he opposie conclusion should hold: he disurbance erm should be orhogonal o he final GDP esimae while correlaed wih he preliminary GDP esimae. In he second case, when he disurbance is correlaed wih he preliminary GDP esimae, he laer may be used for predicing he final GDP esimae. Under his condiion, he 'noise' hypohesis holds, and in accordance o Mincer and Zarnowiz (1969) work, he equaion 1 can be rearranged as follows: f n R = β 0 + β 1 P + u (2) 7

8 where: R is he revision of he GDP forecas a ime + n (explained in Secion 3); P is preliminary GDP esimae, β 0 is he inercep erm, β 1 is a parameer o be esimaed, u are disurbances of he model and is a ime subscrip. In order o es for he efficiency of he GDP forecas and hus, he predicabiliy of he possible revisions, and wheher he 'news' hypohesis holds, we es for he following join hypohesis: H 0 : β 0 = β 1 = 0 (3) If we fail o rejec he join hypohesis H 0, i indicaes ha he inercep and he slope coefficien of equaion 2 are saisically no differen from zero. This implies ha he whole regression is meaningful and consequenly, he revisions may no be predicable. This means ha he GDP forecas may be efficien and he revisions are in line wih he 'news' hypohesis, i.e. he newly gahered informaion is incorporaed in he revised GDP esimae and he revisions are no prone o measuremen errors. As argued by Aruoba (2006), alhough he 'news' and 'noise' hypoheses are muually exclusive (he rejecion of he 'news' hypohesis from equaion 3 may implicily sugges non-rejecion of he alernaive 'noise' hypohesis) however, "..hey are no collecively exhausive, ha is, we can rejec boh hypoheses, especially when he uncondiional mean of revisions is no equal o zero." (p. 14). In order o check for he laer, according o Mankiw and Shapiro (1986) and Aruoba (2006), he 'noise' hypohesis can also be esed separaely. For example, if GDP esimaes are efficien, and wih each subsequen revision new informaion is incorporaed ha was no available previously, he revision should be correlaed wih he revised GDP esimae. In his ligh, in he syle of Auroba (2006), we can es he 'noise' hypohesis separaely wih he following equaion: n R n = α 0 + α 1 P + (4) where: same as in equaion 2, R is he revision of he GDP esimae a ime + n; n P is revised GDP esimae a ime + n, α 0 is he inercep erm, α 1 is a slope parameer o be esimaed, are disurbances of he model and is a ime subscrip. Under he 'noise' hypohesis, we es for he following join hypohesis: 8

9 H 0 : α 0 = α 1 = 0 (5) where non-rejecion of he null hypohesis implies ha boh coefficiens in equaion 4 are joinly saisically no differen from zero and herefore, we do no rejec he 'noise' hypohesis. Rejecion of he null hypohesis from equaion 5 suggess ha he revised GDP esimae is correlaed wih he revision, which is in line wih he 'news' hypohesis. Consequenly, in he nex secion, we saisically es he aforemenioned five hypoheses. 5. Discussion of he resuls In his secion, we discuss he resuls from esing he hypoheses presened in he previous secion. By esing hypohesis 1, he scaer plos beween he preliminary GDP esimaes and revised GDP esimaes are presened in Figure 1, whereas he scaer plos beween preliminary GDP esimaes and revisions (difference beween revised esimaes and preliminary esimaes) are presened in Figure 2. The scaer plo beween he GDP preliminary esimae P and he very shor-erm revised esimae 1 P (see Figure 1, panel a) indicae ha regression line has a slope of almos 45-degree line, and lies on he so-called "line of perfec forecass", implying ha he revised GDP esimaes afer one quarer are likely unbiased. The scaer plos beween he preliminary GDP esimae and he long-erm GDP revised esimae (afer seven quarers) and he final GDP esimae (Figure 1, panels b and c) imply ha he regression lines inersec he "line of perfec forecass", suggesing ha he long-erm and final GDP revised esimaes may be biased. 9

10 GDP_F GDP_P1 GDP_P7 Figure 1: Scaer plos beween GDP preliminary esimae (horizonal axis) and GDP revised esimaes (verical axis) 8 Panel a 10 Panel b GDP_P Panel c GDP_P GDP_P Panel a is a scaer plo beween GDP preliminary esimae and shor-erm revised esimae; Panel b is a scaer plo beween GDP preliminary esimae and long-erm revised esimae; Panel c is a scaer plo beween GDP preliminary esimae and final revised esimae. Source: Auhors' calculaions based upon daa from SSO performed in Eviews 7. By analyzing he scaer plos beween he GDP preliminary esimae and revisions (shor-erm, long-erm and final revision, see Figure 2), he regression line is downward sloping in all hree figures and is much seeper and sysemaically differen from zero beween he GDP preliminary esimae and he long-erm and final revisions (see Figure 2, panels b and c). This indicaes a possible exisence of a sysemaic relaionship beween he GDP preliminary esimaes and revisions. The downward sloping regression line in scaer plos in Figure 2 (panel b and c) generally relies on he posiive zone of x and y axis, which may imply ha high preliminary esimaed GDP growh raes are revised upward in a smaller size or magniude. The 10

11 GDP_RF GDP_R1 GDP_R7 opposie is rue when he preliminary esimae of GDP growh rae is small or low. Then he upward revision is of a bigger size. Figure 2: Scaer plos beween GDP preliminary esimae (horizonal axis) and GDP revisions (verical axis) Panel a Panel b GDP_P Panel c GDP_P GDP_P Panel a is a scaer plo beween GDP preliminary esimae and shor-erm revision; Panel b is a scaer plo beween GDP preliminary esimae and long-erm revision; Panel c is a scaer plo beween GDP preliminary esimae and final revision. Source: Auhors' calculaions based upon daa from SSO performed in Eviews 7. When assessing he mean of revisions (esing hypohesis 2 from previous secion), he daa presened in Table 1 indicae ha he mean is close o zero only for he shor-erm revisions, while for he long-erm and final revisions, i is increasing and is above zero especially for he final GDP revision. This implies rejecion of he null hypohesis of Mincer and Zarnowiz (1969) forecas efficiency es and hence, he revisions may be biased and conain 'noise'. These resuls 11

12 can also be suppored wih he -saisics, by which we es if he mean of he revisions is saisically no differen from zero. The resuls from he -saisics for he shor-erm revisions indicae rejecion of he null hypohesis ha he mean of he revisions is saisically no differen from zero a 10% level of significance, whereas he -saisics resuls for he long-erm and final revisions imply rejecion of he null hypohesis a 5% and 1% level of significance, respecively. Consequenly, he -saisics sugges ha he mean of revisions is no saisically equal o zero. Table 1: Mean of revisions, variance of GDP esimaes and correlaion coefficiens beween GDP preliminary esimae and revisions. Mean of he revisions: Correlaion coefficiens beween GDP preliminary esimae and revisions: Correlaion coefficiens beween GDP revised esimae and revisions: Variance of he GDP esimae: Preliminary / / / 4.84 Shor-erm 0.15* Long-erm 0.71** -0.49** Final 1.53*** ** 5.78 */**/*** indicaes rejecion of he null hypohesis ha he esimaed value is saisically differen from zero a 10%, 5% and 1% level of significance, respecively. Source: Auhors' calculaions based upon daa from SSO performed in Eviews 7. When esing hypohesis 3, explained in secion 4 abou he correlaion beween revisions and preliminary GDP esimaes, he resuls in Table 1 sugges a negaive correlaion beween preliminary GDP esimaes and revisions (shor-erm, long-erm and final). These correlaion coefficiens, apar from he ones beween GDP preliminary esimaes and long-erm revisions, are saisically no differen from zero a any level of significance. This is in line wih he Mincer and Zarnowiz (1969) and Mankiw and Shapiro (1986) argumens for he 'news' hypohesis implying a non-rejecion of he null hypohesis according o which, he correlaion beween he GDP preliminary esimaes and revisions should be saisically no differen from zero. Therefore, we may conclude ha revisions conain news raher han measuremen errors. In addiion, we also es he 'news' hypohesis by analyzing he correlaion coefficiens beween he GDP revisions and revised GDP esimae. The resuls in Table 1 sugges again ha he correlaion is saisically differen from zero a 5% level only for he correlaion coefficien beween GDP final esimae and final revision, poining ou ha he final GDP revisions conain 'news'. However, for he res of he correlaion coefficiens, his canno be suppored. Overall, he correlaion coefficiens for he long-erm revisions imply ha hey may conain 'noise' and 12

13 exhibi sysemaic paern, whereas for he final GDP revisions, he resuls are in line wih he 'news' hypohesis. Neverheless, he resuls for he shor-erm GDP revisions based on correlaion coefficiens are mixed. Regarding he fourh hypohesis in secion 4 ha addresses he variance of he GDP subsequen esimaes, he resuls in Table 1 show ha variances of each subsequen GDP esimae are increasing, which is in line wih he hypohesis of efficien forecass, i.e. he 'news' hypohesis. In order o empirically examine if here is a sysemaic paern in he revisions and if his is rue, wheher i can be prediced hrough ime (esing hypohesis 5 from secion 4), we run regressions se wih equaion 2 for he shor-erm, long-erm and final revisions as a dependen variable, respecively. Wih hese regressions, we es he null hypohesis from equaion 3, where non-rejecion of he null hypohesis implies ha revisions canno be prediced and conain 'news', and herefore, preliminary GDP esimae is an efficien forecas. In addiion, we use Newey-Wes (1987) sandard errors ha are robus o heeroskedasiciy and auocorrelaion. The resuls from he esimaed regressions are presened in Table 2. Table 2: Regression resuls for esing he 'news' hypohesis from equaions 2 and 3. Regression 1: Shor-erm revisions Regression 2: Long-erm revisions Regression 3: Final revisions Coefficien: β * S.E Coefficien: Β S.E F-es for join 'news' hypohesis from equaion 3 P-value Adjused R-squared Diagnosic ess (p-values): Serial correlaion Heeroskedasiciy Normaliy Funcional form */**/*** indicaes saisical significance a 10%, 5% and 1% level, respecively. Newey-Wes robus sandard errors are used. Source: Auhors' calculaions based upon daa from SSO performed in Eviews 7. 13

14 Before we sar wih he inerpreaion of he resuls, we firs examine if he residuals are whie noise. We conduc baery of diagnosic ess such as Breusch-Godfrey for serial correlaion, Engel's auo-regressive condiional es for heeroskedasiciy (ARCH), Jarque-Bera es for normal disribuion of residuals and Ramsey's rese es for funcional form. Examining wheher he residuals are whie noise is imporan for esing he 'news' hypohesis from equaion 2, especially he one for serial correlaion, because as argued by Sahlgren (2006), Roodenburg (2004) and Faus e al. (2006), residuals ha are no whie noise and in he case hey are serially correlaed hen i is an indicaor of 'noise' hypohesis. As saed by Sahlgren (2006) "Auo correlaed residuals sugges ha hey are no independen of pas informaion, which in effec means ha he 'news' hypohesis canno hold." (p. 12). The resuls regarding he afore-menioned diagnosic ess are saisfacory and hus, he residuals are whie noise apar from he normaliy es for Regressions 1 and 2 ha may be due o he ouliers in he sample for which a robusness check has been done in he nex secion. Analyzing he esimaed parameers from he regressions presened in Table 2 and heir join saisical significance, i.e. esing he 'news' hypohesis from equaion 3, we can noice ha in all regressions for he shor-erm, long-erm and final revisions a 10% level, we canno rejec he null hypohesis ha he inercep erm and he slope coefficien are joinly saisically no differen from zero. In addiion, he values of he adjused R-squared are very low, ranging from 0.01 o These resuls imply ha here is no any sysemaic paern beween he GDP preliminary esimaes and subsequen revisions and consequenly, revisions canno be prediced. This suppors he 'news' hypohesis ha revisions conain new informaion and are no prone o measuremen errors. According o auhors' knowledge, alhough here is no any exising empirical sudy ha analyzes he predicabiliy of GDP revisions for he Cenral and Souh Easern Europe, we compare he resuls for he exising empirical sudies for advanced economies. The esimaed resuls for he case of he Republic of Macedonia are consisen wih hose for he US economy presened in Faus e al. (2006) and Mankiw and Shapiro (1986), for Sweden esimaed by Sahlgren (2006) and for Brazil in he research of Palis e al. (2004). As explained in secion 4, for predicabiliy of he GDP revisions based on regression approach we also es he 'noise' hypohesis se in equaion 4, where we examine if here is any sysemaic relaionship beween he revised GDP esimae and revision. As menioned before, we firsly check he diagnosic ess and hen we inerpre he resuls from he esimaed regression. 14

15 As presened in Table 3, he diagnosic es poin ha residuals are whie noise, where again he only problem appears wih he normaliy es for Regressions 1 and 3 ha may be due o ouliers in he sample. Regarding he join saisical significance of he consan erm and he slope coefficien, apar from regression 1 from Table 3 beween he shor-erm revision and revised GDP esimae afer one quarer where we failed o rejec he null hypohesis a 10% level, for he oher wo regressions 2 and 3 we can rejec he null hypohesis a 5% level, suggesing efficien GDP forecass. Furhermore, he values of adjused R-squared are much greaer compared o he same ones from Table 3 and for regressions 2 and 3, hey range beween 0.14 and These resuls for he long-erm and final revisions are consisen wih he previously discussed resuls from Table 2 where we esed he 'news' hypohesis, as well as wih he saisical significance of he correlaion coefficiens. Table 3: Regression resuls for esing he 'noise' hypohesis from equaions 4 and 5. Regression 1: Shor-erm revisions Regression 2: Long-erm revisions Regression 3: Final revisions Coefficien: α S.E Coefficien: α *** 0.34** S.E F-es for join 'news' hypohesis from equaion 5 P-value Adjused R-squared Diagnosic ess (p-values): Serial correlaion Heeroskedasiciy Normaliy Funcional form */**/*** indicaes saisical significance a 10%, 5% and 1% level, respecively. Newey-Wes robus sandard errors are used. Source: Auhors' calculaions based upon daa from SSO performed in Eviews 7. 15

16 Overall, alhough he resuls discussed in his secion are ambiguous in some cases, sill, for he long-erm and final revisions hey end o incline ha here is no any sysemaic correlaion beween GDP preliminary esimaes and revisions, which is in line wih he 'news' hypohesis. In general, i can be concluded ha he preliminary GDP esimaes are raional forecass, and long-erm and final GDP revisions conain 'new' informaion which reduce he forecas errors done wih he preliminary GDP esimae. 5.1 Robusness check 6 In order o check he robusness of he resuls, following he approach of Sahlgren (2006) and Faus e al. (2005), we re-examined he same hypoheses discussed in secion 4 by adjusing he sample for he ouliers defined as a mean value of he series wih a deviaion of + / - 3 percenage poins. The resuls from he scaer plos are consisen wih he ones already discussed a he beginning of he previous secion. The scaer plos imply again ha he revisions may be biased o some exen because he regression lines in he scaer plos beween he preliminary and revised GDP esimaes inersec he 45-degree line, whereas he regression lines on he scaer plos beween GDP preliminary esimae and revisions is downward sloping. Regarding he mean of he revisions and he variance of he GDP esimaes, he resuls are also in line wih he resuls discussed in he previous secion. Namely, he mean of he revisions is no saisically equal o zero, while he variance of each subsequen esimae is increasing. The resuls from he correlaion coefficiens poin again ha correlaion coefficiens beween he GDP preliminary esimaes and shor-erm and final revisions are saisically no differen from zero a 10% level, while he correlaion coefficien beween GDP preliminary esimae and long-erm revision is saisically significan a 5% level. The resuls for he correlaion coefficiens beween revised GDP esimaes and revisions indicae once again ha only he correlaion coefficien beween final GDP esimae and final GDP revision is saisically significan a he 5% level. Checking he robusness of he resuls for he same regressions repored in Tables 2 and 3, he esimaed resuls are consisen again. The 'news' hypohesis could no be rejeced in all 6 The resuls discussed in his secion are available from he auhors upon reques. 16

17 regressions a 10% level, which is in line wih he resuls presened in Table 2. The 'noise' hypohesis could be rejeced a 10% level for he regressions beween long-erm and final revisions, bu no for he shor-erm revisions, which is consisen wih he regressions repored in Table Conclusions The aim of his analysis is o examine GDP daa revisions in a Souh Eas European economy such as he Republic of Macedonia. In ha respec, his paper examines if here is any sysemaic paern beween GDP esimaes and subsequen shor-erm, long-erm and final revisions, by which we may be able o predic he size and he direcion of he revisions. This may be of considerable imporance for opimal decision-making. We esed he 'news' and 'noise' hypoheses using various empirical approaches. The resuls based on hese empirical approaches in general provide mixed resuls for shor-erm GDP revisions alhough he empirical evidence ends o incline ha he long-erm and final GDP revisions are efficien and conain new informaion ha is in line wih he 'news' hypohesis. More precisely, he scaer plo diagrams and he mean of he revisions sugges ha hey are biased and conain measuremen errors, i.e. 'noise', whereas he variance of he revisions and he regression-based approaches imply ha long-erm and final GDP revisions are efficien and conain new informaion ha is in line wih he 'news' hypohesis. Overall, by carefully assessing he empirical resuls discussed in his paper, we may conclude ha he empirical evidence ends o incline ha he GDP daa revisions conain new informaion ha was no available when he preliminary GDP esimae was done. Therefore, one may conclude ha he preliminary GDP daa are efficien forecass and each subsequen revision incorporaes new informaion ha suppors he 'news' hypohesis. As open research issues are o explore if here is any asymmeric paern in he GDP daa revisions. For example, alhough his analysis assumes linear relaionship beween GDP esimaes and subsequen revisions, neverheless, his relaionship may be non-linear and may indicaes differen paern beween upward and downward revisions and preliminary GDP esimaes. 17

18 References Aruoba, S.B. (2006): Daa Revisions are no Well-Behaved, Journal of Money, Credi and Banking, 2008, 40(2-3), Faus, J., Rogers, J.H., Wrigh J.H. (2005), News and Noise in G-7 GDP Announcemens. The Journal of Money, Credi and Banking, Vol. 37, No. 3, Howrey, E.P. (1978): The Use of Preliminary Daa in Economeric Forecasing, Review of Economics and Saisics, 60, pp Howrey, E.P. (1984): Daa Revision, Reconsrucion and Predicion: An Applicaion o Invenory Invesmen, Review of Economics and Saisics, 66, pp Mankiw, N.G. and M.D. Shapiro (1986): News or Noise: An Analysis of GNP Revisions, Survey of Curren Business, May 1986, pp Mincer, J. and V. Zarnowiz (1969): The Evaluaion of Economic Forecass in J. Mincer (ed.), Economic Forecass and Expecaions, NBER, New York. Newey, W.K. and K.D.Wes (1987): A Simple Posiive Semidefinie, Heeroskedasiciy and Auocorrelaion Consisen Covariance Marix, Economerica, 55, pp Orphanides, A. (1998): Moneary Policy Rules Based on Real-Time Daa, Board of Governors of he Federal Reserve Sysem, Finance and Economics Discussion Paper, forhcoming American Economic Review. Orphanides, A. (2000): The Ques for Prosperiy Wihou Inflaion, European Cenral Bank Working Paper. Orphanides, A. and van Norden S. (2002): The Unreliabiliy of Oupu-Gap Esimaes in Real Time. The Review of Economics and Saisics, Vol. 84, No. 4, Palis, R., R. Ramos and P. Robiaille (2004): News or noise? An analysis of Brazilian GDP announcemens, Inernaional Finance Discussion Papers, No 776. Roodenburg, Olivier (2004), On he predicabiliy of GDP daa revisions in he Neherlands. Working Paper, No. 4, July. Neherlands Cenral Bank (DNB). Sahlgren, D. H. (2006): A Real-Time Daa Se for Swedish GDP - GDP Revisions and he Oupu Gap in Real-Time, Sockholm School of Economics. 18