Chapter 1. Euro Area Macroeconomic Outlook and Forecasts. Annex B

Transcription

1 Chaper Euro Area Macroeconomic Oulook and Forecass Annex B

2 . Inroducion In his paper we compare alernaive procedures for forecasing fiscal variables for he larges counries in he Euro area. An imporan moivaion for his exercise comes from he recogniion ha fiscal forecass are playing an increasing role in macroeconomic policy decisions. This has been paricularly obvious in he European conex where, for example, he operaing procedures of he Sabiliy and Growh Pac involve reference o forecas values of he fiscal defici and deb a more han one poin. We consider five differen ypes of forecass. Firs, sandard ARMA model based forecass, which perform quie well for several European macroeconomic variables, boh on a counry by counry basis and a he euro area aggregae level, see e.g. Marcellino (00a, 00b) and Baneree, Marcellino and Masen (003). Also, Aris and Marcellino (00) found ha even simple random walk forecass someimes ouperform he leading inernaional organizaions such as he IMF, he EC or he OECD. Second, VAR based forecass, since VARs have been ofen used o model fiscal variables and heir ineracion wih oher macroeconomic variables, see e.g. Blanchard and Peroi (00?) for he US, Peroi (00?) for some OECD counries, and Marcellino (00c) for he larges counries in he Euro area. Third, forecass from small scale srucural models conaining hree ypes of variables: macroeconomic indicaors, fiscal policy indicaors and moneary policy indicaors. We consider boh naional models, along he lines of Favero (00) who used similar models o sudy he ineracion beween fiscal and moneary auhoriies, and a larger euro area model, where he naional models are linked up ogeher o ake ino accoun he implicaions of he convergence process sared by he adopion of he single currency, and in paricular he presence of a single moneary policy wih differen fiscal policies. Fourh, pooled forecass obained by aking eiher he mean or he median of he previous hree ypes of forecass. Since he pioneering work of Baes and Granger (96?) pooling has been found o be useful in improving he forecasing accuracy, see e.g. Clemens and Hendry (00) for possible reasons underlying his resul. Recen papers highlighing he good performance of pooling for forecasing macroeconomic variables include Sock and Wason (999) for he US and Marcellino (00d) for he Euro area. Sock and Wason (00) find ha he simple average or median of he single forecass perform well compared wih more sophisicaed pooling procedures. Finally, we consider he OECD forecass, as published in he World Economic Oulook. The forecass in quesion are no direcly derived from formal macroeconomeric models bu emerge from he ieraive inerplay beween parial formal modelling, commiee ieraion and udgmenal discreion. Besides four key fiscal variables, i.e. governmen expendiures and receips, he defici and he governmen deb, we also consider forecasing he oupu gap, inflaion and a shor erm ineres rae, since hese are imporan variables o deermine he evoluion of he fiscal aggregaes. All daa are semi-annual and are exraced from he OECD daase, wih deails provided below. We repor resuls for one-sep and wo-sep ahead forecass, ha can be used o derive curren year and year ahead forecass. We also summarize he findings for four-sep ahead forecass o evaluae wheher he gains from srucural models increase wih he forecas horizon. Longer horizons are no worh evaluaing because of he subsanial uncerainy surrounding he forecass and he large biases ha emerge.

3 We can anicipae some of he main resuls we obain. Firs, for he macroeconomic variables he ARMA forecass are ofen he bes, wih a slighly worse performance a he longer horizon. Second, for he fiscal variables he ime-series forecass in general are he mos accurae a he shorer horizon, while more mixed resuls are obained a he longer horizon. Third, he good performance of he random walk forecass menioned before emerges also from our analysis, hough in general i is possible o find a model ha ouperforms he random walk. Fourh, in general he srucural models do no yield any subsanial forecasing gains, and a similar resul holds for he OECD forecass a he shores horizon. This finding is likely due o he fac ha our models are no fine-uned for forecasing, bu i is ye anoher indicaion ha simple ime series models or pooling ofen yield he bes forecass. Finally, subsanial uncerainy surrounds he forecass, so ha he compeing forecass are seldom saisically differen, and he size of he average forecas error for he fiscal balance, perhaps he mos ineresing fiscal variable from he policy poin of view, is raher large. The srucure of he paper is he following. In secion we briefly describe he daase. In secion 3 we discuss he differen forecasing mehods we adop. In secion 4 we presen he resuls. In secion 5 we summarize and conclude.. Daa We focus on he four larges counries of he Euro area, namely, Germany, France, Ialy and Spain. For each counry we consider seven variables: he oupu gap; he CPI inflaion rae; a moneary policy indicaor (a money marke rae); primary governmen deficis, also decomposed ino revenues and expendiures; and oal governmen deb. The fiscal variables are expressed as raios o GDP. The daa source is he OECD and he frequency is half-yearly. This choice conrass wih he sandard adopion of monhly or quarerly daa for he analysis of macroeconomic variables. I is dicaed firs by daa availabiliy, and second by he fac ha in mos counries he maor fiscal decisions are aken once a year, and possibly revised once. Peroi (00) consrucs a quarerly daase, bu Germany is he only counry wihin he Euro area for which such daa are available. For all counries he sample under analysis is 98:-00:, as in Marcellino (00c). Though for some counries longer series are available, boh Favero (00) and Peroi (00) found a clear indicaion of differen behaviour of fiscal and moneary policy policy afer he 70s, which suggess o focus on he mos recen period. The variables are graphed in Figure. There is a subsanial co-movemen of he business cycles of France, Germany, Spain and Ialy, in line wih he more deailed analysis in Aris, Marcellino and Proiei (003). The convergence process in inflaion and ineres raes is also eviden. Boh feaures of he series should be aken ino consideraion in he model specificaion sage. Figure also shows he working of he Maasrich Treay in reducing he fiscal defici and he governmen deb in all he four counries, a reducion ha appears o be due more o expendiure cus han o ax increases. The figure does no highligh a non-saionary behaviour for he variables, possibly wih he excepion of he deb o GDP raio. Since here are srong economic reasons o assume ha all he seven variables are saionary, we will proceed under his assumpion even hough he oucome of

4 ADF uni roo ess is mixed, likely due o he low power of hese ess in samples as shor as our (4 observaions). 3. Forecasing models We now describe he four differen ypes of forecass we consruc, namely, ARMA, VAR, Simulaneous Equaion Model (SEM), and pooled, focusing more on he SEM since i is he mos original mehod. All models are specified using he full sample available, which is raher shor (4 observaions) so ha recursive modelling is no suied. For he specificaion of he ARMA models we sar wih an ARMA(,) for each variable and counry, and selec he combinaion of AR and MA lengh ha minimizes he BIC. The resuling models are summarized in Table. Overall he fi is good, hough his does no represen a reliable indicaion for forecasing, wih lower values in he case of Germany. I is also ineresing o poin ou he similariy of he models for Ialy and Spain, and he fac ha an MA componen is always included in he model for inflaion. In he subsequen analysis, following sandard pracice, we will also include a random walk based forecas. For he (seven variable) VAR models, we can only include one or wo lags because of he degrees of freedom consrain. Raher han selecing he lag lengh wih an informaion crierion, we compue forecass for boh cases, which also allows o compare he performance of he same model for each counry and variable. Abou he SEM, i is useful o disinguish beween naional models and he euro area model. The general specificaion of he naional models follows Favero (00) and is skeched below, wih indexing he counries, more deails are provided in he Appendix. π u () = c π c y NP, y US = c3 c4 y c5 c6i c7 g c8τ cy π u () NP, i = c π u (3) 0 ci c c3 y c4i M, 3 g τ c = c c g x ( x π ) c π y ( x π ) DY c y u g, 5 = c3 c4 c5 y c6 y c 8 τ x π DY u τ, 6 c ( x π ) c 7 ( x π ) avc avc * DY * DY (4) (5) The noaion is as follows. π is annual inflaion of he GDP deflaor; y is he oupu gap, i.e., he percenage difference beween oupu and poenial oupu as measured by he OECD, i is he moneary policy rae; avc is he average cos of deb, i.e., he raio of ineres paymen on

5 governmen deb o GDP; g is he raio of governmen expendiure o GDP; τ is he raio of governmen revenue o GDP; DY is he raio of governmen deb o GDP; and x is real annual GDP growh. Equaions () and () represen aggregae supply and demand. The specificaion is similar o he one adoped in he recen srand of he empirical macroeconomeric lieraure based on small scale models, see e.g. Rudebusch and Svensson (999), Clarida, Gali and Gerler (000). In he demand equaion we inroduce lagged governmen expendiures and revenues, o ake ino accoun he delays in he effecs of fiscal policy and allow for a differen elasiciy of oupu o he wo fiscal componens. Demand can be also influenced by he corresponding US variables, and by he ineres rae, possibly in real erms. From he esimaed models repored in he Appendix, in all counries he oupu gap eners wih he proper sign ino he specificaion of he aggregae demand (Phillips curve) equaion, bu i is significan only for France and Spain. Fiscal and moneary policy appear o have a limied effec on he evoluion of he oupu gap in all counries, wih ofen a negaive coefficien for public expendiures. Insead, in all counries he oupu gap reacs posiively and significanly o he US gap. Equaion (3) is a moneary reacion funcion, in line wih a Taylor-rule ype of specificaion. I can be derived as he soluion of he opimizaion problem of a cenral bank ha has a quadraic obecive funcion in he deviaion of inflaion from arge, he oupu gap, and volailiy in he policy raes, see e.g. Favero and Rovelli (00). The inclusion of he German ineres rae in he equaion for he oher counries capures he anchor role of Germany over his sample period, see e.g. Clarida, Gali and Gerler (998). From he Appendix, boh inflaion and he oupu gap have he proper sign and are significan for Germany, he oupu gap seems o maer less for he oher counries (likely due o he overall marked decline of inflaion over our sample period), while he German ineres rae exers an imporan role. To evaluae wheher he moneary auhoriy reacs o fiscal policy we have also included he governmen defici and/or deb in he specificaion, bu hey were never saisically significan. The evoluion of governmen expendiures and receips is deermined by equaions (4) and (5). The specificaion of hese equaions follows Bohn (988), who allows for a smooh reacion of primary deficis o he oupu gap and o he deb o GDP raio. Ye, we prefer o separaely model he componens of he primary balance since hey separaely ener he demand funcion. Moreover, our specificaion allows for a ime-varying reacion of he primary defici (and is componens) o he deb o GDP raio, which depends on he nominal rae of growh of oupu and he average cos of deb. From he Appendix, in all counries here is subsanial ineria in public expendiures, and hey also increase in he presence of negaive oupu gaps, bu virually wihou any long run effecs. Taxes are also persisen, he effecs of he oupu gap are minor (he oupu level maers more), while axes increase significanly wih he cos of public deb. The model includes an equaion for he evoluion of he average cos of deb, avc = c u (6) 5 avc c6i NP, 4

6 and for dynamic simulaion purposes i is closed by he wo equaions below, describing he evoluion of he deb o GDP raio and he relaionship beween real GDP growh and he oupu gap. DY avc x π = DY DY ( x π ) ( g τ ) (7) y (8) x = c9 c30 The parsimonious specificaion of he naional models reflecs he limied number of degrees of freedom. Though more complex dynamics or cross variable relaionships migh exis, hey can be hardly deeced and accuraely esimaed wih such a shor sample. On he oher hand, he esimaed equaions (using SUR), repored in he Appendix, in general provide a good fi and do no presen heavy signals of misspecificaion. Moreover, parsimony is usually a benefi when forecasing is he goal of he analysis, as in our case. Similarly, he use of dummy variables could furher improve he fi and diagnosic ess of he model, bu i could deeriorae he forecasing performance of he model by making is specificaion oo much sample dependen. Since forecasing is our goal, we are also no ineresed in invesigaing wheher he backward looking srucure of he model is genuine or wheher i is he reduced form of a forward looking model. Insead, i can be ineresing o evaluae he dynamic behaviour of he model in equaions ()-(8) when all shocks are se o zero. The shor run behaviour is of paricular relevance for our shor erm forecasing exercise, bu he long run behaviour is also imporan o evaluae he soundness of he economic hypohesis we made in specifying he model. The dynamic behaviour of he naional models is summarized in Figure, and overall i is quie saisfacory. The gap, inflaion and ineres rae end o move ogeher across counries. There are some differences in he long run values bu sochasic simulaions of he model have shown ha hese differences are no saisically significan. Acually, as expeced, he sandard errors around he poin esimaes end o be quie large a he long horizon. Abou he fiscal variables, he expendiure and receip raios do no show any marked dynamics, while he governmen balance flucuaes in he range [-.5%,0] and he deb raio converges o values below 60%. The laer is an imporan finding since i indicaes ha we do no need o impose any resricions on he model o guaranee ha he Maasrich crieria are saisfied. We can now discuss he euro area model. This model links he naional models ogeher bu also akes ino accoun he convergence process associaed wih he moneary union ha was already eviden from he graphs of he macroeconomic variables. The euro area variables are consruced as averages of heir naional counerpars using real 995 GDP weighs. The main characerisics of he model are he following. The naional inflaion raes can reac also o he lagged euro area inflaion and is change, and in general hey do. The naional oupu gap can reac o is pas difference wih respec o he area gap. This erm usually has a negaive sign (excep for Ialy where i is no significan) supporing real convergence. The German ineres rae can reac no only o naional bu also o area wide inflaion (posiive and significan) and oupu gap (posiive bu no significan). The equaions for expendiures and receips are similar o hose for he naional models, since fiscal policy is no coordinaed a he euro area level. A deailed descripion of he area model can be found in he Appendix. The dynamic simulaion of he model is repored in Figure 3. The resuls are similar o hose for he naional models, possibly

7 wih a closer convergence of he macro variables. The sandard errors around he poin esimaes remain quie large, in paricular a longer horizons. Finally, we consider wo forecas pooling procedures, he mean and he median of he forecass we discussed so far, which nowihsanding heir simpliciy have performed quie well in previous analyses. 4. Forecasing fiscal variables In his secion we briefly review he forecasing mehodology, which is raher sandard, presen he resuls, and finally discuss a comparison wih he OECD fiscal forecass. 4. Forecasing mehodology As we menioned in he previous secion, he specificaion of he forecasing models is based on he full sample. Ye, he chosen model is re-esimaed over he forecas period, eiher recursively wih he firs sample ending in 995:, or wih a 5 year rolling window, so ha he firs window ends again in 995:. The esimaed models are used o produce -, - and 4- semeser ahead forecass, where he laer are compued by forward ieraion of he model raher han by means of dynamic esimaion o avoid a furher specificaion search (see e.g. Marcellino, Sock and Wason (003) for deails on dynamic esimaion). The resuling forecass and he acual values are used o compue he mean square error (mse) and he mean absolue error (mae). Boh he mse and he mae of each model are expressed as a raio of he corresponding values for he random walk forecass, so ha raios smaller han one indicae a worse performance of he random walk forecass. Finally, we compue he Diebold and Mariano (995, DM) es saisic o evaluae he saisical significance of he loss differenials. Two commens are in order on his opic. Firs, even hough we apply he small sample correcions in Harvey, Leybourne and Newbold (99?), he very limied number of forecass cass some doubs on he reliabiliy of saisical esing in our conex. Second, since some models are nesed, he asympoic disribuion of he DM es could be differen from he sandard normal, see e.g. Clark and McCracken (00?). Ye, Giacomini (00) has shown ha he use of a fixed rolling window for esimaion resores he validiy of he DM resuls also in he case of nesed models. 4. Resuls Table presens he MSE of each forecasing mehod relaive o he random walk. ARMA models are clearly he bes a he shores horizon for mos variables and counries (7 ou of 8), wih pooled forecass ranked second (6 ou of 8). The performance of he ARMA models deerioraes wih he forecas horizon, ARMA produce he lowes MSE in ou of 8 cases for h= and 9 ou of 8 for h=4 (Table 6), while ha of he pooling mehods is basically unaffeced (7 ou of 8 bes forecass for h= and 8 ou of 8 for h=4). The srucural models do slighly beer a he longes horizon, hey are he bes in 6 ou of 8 cases for h=4 and only in 3 ou of 8 cases for h=, bu are sill beaen ofen by he ime series mehods.

8 These models perform bes for gap and expendiure forecass in Germany and for he ineres rae in France. Tables 3 and 6 repor he relaive MAE for h=, and 4, respecively. Basically, hey show ha he ranking above is robus o he use of he MAE o compare he forecass. Tables 4, 5 and 7 repea he MSE and MAE comparison using rolling esimaion. Also in his case here are no maor changes in he ranking of he forecass, while no clear cu comparison of rolling and recursive esimaion emerges. As we menioned before, because of he shor sample size he forecass are surrounded by a raher large uncerainy. As a consequence, he MSEs and MAEs are seldom saisically differen from hose of he random walk model, even hough he laer is sysemaically beaen by he bes forecas in erms of poin MSE and MAE values. Finally, comparing he MAEs wih he average value of he variables in he las column of Tables and 6, i emerges ha he forecass for he expendiure and receip raios and for he deb raio are much more accurae han hose for he fiscal balance. 4.3 Comparison wih OECD forecass The OECD publishes curren year forecass in June and year ahead forecass in December for some of he variables we consider. I is herefore ineresing o compare heir forecass wih ours, using he same mehodology as above, bu wih an accurae choice of he iming (o reflec he availabiliy of OECD forecass), and forecas definiion. Noice ha our models are slighly advanaged by he full sample specificaion. We also include pooled OECD srucural model forecass in he comparison. The resuls in Tables 8- indicae ha pooled (mean) forecass perform quie well for he curren year, wih he OECD being he bes for all counries only for he ineres rae. The OECD improves for he year ahead forecass, bu pooling or one of our models remains bes for gap and deb. Again he resuls are robus o he choice of loss funcion (MSE or MAE) and mehod of esimaion (recursive or rolling). The good performance of he random walk is confirmed also wih respec o he OECD, in paricular one sep ahead. 5. Conclusions The common finding of good performance of simple ime-series or pooled forecass for macroeconomic variables is confirmed also for fiscal variables for he larges counries in he Euro area. This finding can be due o several reasons, including he shor sample available ha makes he specificaion and esimaion of srucural models complicaed, he robusness of simple mehods o srucural breaks (his is paricularly so for random walk and pooled forecass), and he difficuly of modelling he oin behaviour of several variables in a period of subsanial insiuional and economic changes. References Aris, M. and Marcellino, M. (00). Fiscal forecasing: he rack record of IMF, OECD and EC, Economerics Journal, 4, s0-s36.

9 Aris, M., Marcellino, M., Proiei, T. (003), ) Daing he Euro area business cycle, CEPR WP Baneree, A., Marcellino, M., Masen I. (003), Leading indicaors for Euro area inflaion and GDP growh, CEPR WP Baes, J.M. and Granger, C.W.J. (969). The combinaion of forecass, Operaions Research Quarerly, 0, Blanchard, O. and Peroi, R. (999), An empirical characerizaion of he dynamic effecs of changes in governmen spending and axes on oupu, Quarerly Journal of Economics, forhcoming. Bohn 988 Clarck, T.E. McCracken, M.W. (00), Tess of equal forecas accuracy and encompassing fr nesed models, Journal of Economerics, 05, Clarida gali gerler 000 Clemens, M.P. and Hendry, D.F. (00). Pooling of forecass, Economerics Journal, (forhcoming). Diebold, F.X. and Mariano, R.S. (995), Comparing predicive accuracy, Journal of Business and Economic Saisics, 3, Favero, C.A: (00), How do European moneary and fiscal auhoriies behave?, CEPR WP 346. Favero rovelli 0 Giacomini, R. (00), Tess of condiional predicive abiliy, mimeo, UCSD Harvey, D., Leybourne, S., and Newbold, P. (997), Tesing he equaliy of predicion mean squared errors, Inernaional Journal of Forecasing, 3, 8-9. Marcellino, M. (00a). Insabiliy and non-lineariy in he EMU, CEPR WP 33. Marcellino, M. (00b). Forecasing EMU macroeconomic variables, CEPR WP 359 (forhcoming in he Inernaional Journal of Forecasing) Marcellino, M (00c), Some sylized facs on fiscal policy in he Euro area, CEPR WP Marcellino, M (00d), ) Forecas pooling for shor ime series of macroeconomic variables, Oxford Bullein of Economics and Saisics, forhcoming. Marcellino, M., Sock, J.H. and Wason, M.W. (003). Macroeconomic forecasing in he Euro area: counry specific versus Euro wide informaion, European Economic Review, 47, -8.

10 Peroi, R. (00), Esimaing he effecs of fiscal policy in OECD counries, mimeo, European Universiy Insiue. Rudebusch Svensson Sock, J.H. and Wason, M.W. (999). A comparison of linear and nonlinear univariae models for forecasing macroeconomic ime series, in Engle, R. and Whie, R. (eds), Coinegraion, causaliy, and forecasing: A fesschrif in honor of Clive W.J. Granger, Oxford: Oxford Universiy Press, -44. Sock, J.H. and Wason, M.W. (00). Combinaion forecass of oupu growh and he 00 recession, mimeo, Harvard Universiy and Princeon Universiy.

11 Figure : Macro and Fiscal variables 98: 00: GAP GAP GAP GAP INFL INFL INFL INFL MMR MMR MMR MMR RECY RECY RECY RECY TEXPY TEXPY TEXPY TEXPY BALY BALY BALY BALY DEBTY DEBTY DEBTY DEBTY

12 Figure : Simulaion Single Counry Models Esimaion sample 98: 00: MODF_GAP MODF_GAP MODF_GAP MODF_GAP MODF_INFL MODF_INFL MODF_INFL MODF_INFL MODF_MMR MODF_MMR MODF_MMR MODF_MMR MODF_RECY MODF_RECY MODF_RECY MODF_RECY MODF_TEXPY MODF_TEXPY MODF_TEXPY MODF_TEXPY MODF_BALY MODF_BALY MODF_BALY MODF_BALY MODF_DEBTY MODF_DEBTY MODF_DEBTY MODF_DEBTY Noes: he figures repor dynamically simulaed pahs of macroeconomic and fiscal variables over he sample 003: 03:

13 Figure 3: Simulaion Area Model Esimaion sample 98: 00: EF_GAP EF_GAP EF_GAP EF_GAP EF_INFL EF_INFL EF_INFL EF_INFL EF_MMR EF_MMR EF_MMR EF_MMR EF_RECY EF_RECY EF_RECY EF_RECY EF_TEXPY EF_TEXPY EF_TEXPY EF_TEXPY EF_BALY EF_BALY EF_BALY EF_BALY EF_DEBTY EF_DEBTY EF_DEBTY EF_DEBTY Noes: he figures repor dynamically simulaed pahs of macroeconomic and fiscal variables over he sample 003: 03:

14 Table : Selecion of ARMA models Germany France Ialy Spain R BIC BIC Gap ARMA(,) Infl ARMA(,) Inrae ARMA(,) Bal AR() Exp AR() Rec AR() Deb ARMA(,) Gap ARMA(,) Infl ARMA(,) Inrae ARMA(,) Bal AR() Exp AR() Rec AR() Deb AR() Gap AR() Infl ARMA(,) Inrae ARMA(,) Bal ARMA(,) Exp ARMA(,) Rec ARMA(,) Deb AR() Gap AR() Infl ARMA(,) Inrae AR() Bal ARMA(,) Exp ARMA(,) Rec ARMA(,) Deb AR() Noes: he able repors he min-bic ARMA specificaion for each variable, along wih is adused R-squared, BIC and he BIC of he ARMA(,) specificaion.

15 Table : Relaive RMSE Recursive esimaes -Sep ahead -Sep ahead ARMA VAR VAR S.C.M AW.M Mean Med RMSE ARMA VAR VAR S.C.M AW.M Mean Med Germany Gap Infl Inrae (*).49 (*) Bal (*) Exp Rec Deb France Gap 0.75 (*) (*) (*) (*) (*) Infl (*) (*) (*) Inrae Bal (*) (*) Exp Rec (*) (*) Deb (*).99 (*).636 (*) (*) Ialy Gap Infl (*) Inrae (*) (*) (*) Bal Exp Rec (*) Deb (*) (*) Spain Gap (*) (*) 0.59 (*) 0.54 (*) Infl (*) Inrae (*).33 (*) (*) (*) (*) Bal (*) (*).67 (*).93 (*) 0.8 (*) (*) (*).40 (*) Exp Rec (*).048 (*) (*).83 (*).585 (*) Deb RMSE Noes: The able enries are he RMSEs of differen models, relaive o he RMSE of a random walk model, for one and wo-sep ahead simulaed forecass. Esimaion sample is 98: 995:. Forecass are performed over he sample 996: 00:. Resuls are repored for ARMA models (ARMA see able xx for deails), one and wo-lag VARs (VAR and VAR), single-counry srucural models (S.C. M see ex for deails), an area-wide model (AW. M see ex for deails), and for pooled forecass (compued each period as he mean and he median of he forecass of all models - MEAN and MED respecively), along wih he RMSE of he random walk model ( RMSE). A es (see Diebold and Mariano (995) and Harvey a al. (997) is also performed on he significance of he mean of he difference beween he squared errors of he differen models and hose of he random walk. (*) denoes 5% significance.

16 Table 3: Relaive MAE Recursive esimaes -Sep ahead -Sep ahead ARMA VAR VAR S.C.M AW.M Mean Med MAE Germany ARMA VAR VAR S.C.M AW.M Mean Med Gap Infl Inrae (*).633 (*) (*) Bal 0.98 (*) Exp Rec Deb MAE Variable Average France Gap (*) (*) 0.64 (*) (*) (*) Infl (*) (*) Inrae Bal (*) (*) Exp Rec (*) Deb (*).078 (*).034 (*) (*).66 (*) Ialy Gap (*) Infl (*) Inrae (*) (*) Bal (*) Exp (*) (*) (*) Rec (*) Deb (*) (*) Spain Gap 0.48 (*) (*) 0.56 (*) Infl (*) (*) (*) Inrae (*).309 (*).0.67 (*) (*).63 (*) (*) Bal (*) (*) (*) Exp Rec (*).897 (*) (*) 3.9 (*) 3.04 (*) Deb Noes: The able enries are he MAEs of differen models, relaive o he MAE of a random walk model, for one and wo-sep ahead simulaed forecass. Esimaion sample is 98: 995:. Forecass are performed over he sample 996: 00:. Resuls are repored for ARMA models (ARMA see able xx for deails), one and wo-lag VARs (VAR and VAR), single-counry srucural models (S.C. M see ex for deails), an area-wide model (AW. M see ex for deails), and for pooled forecass (compued each period as he mean and he median of he forecass of all models - MEAN and MED respecively), along wih he MAE of he random walk model ( MAE) and he average value of he variables over he forecasing sample. A es (see Diebold and Mariano (995) and Harvey a al. (997) is also performed on he significance of he mean of he difference beween he absolue errors of he differen models and hose of he random walk. (*) denoes 5% significance.

17 Table 4: Relaive RMSE Rolling esimaes -Sep ahead -Sep ahead ARMA VAR VAR S.C.M AW.M Mean Med RMSE Germany ARMA VAR VAR S.C.M AW.M Mean Med Gap Infl Inrae Bal (*) Exp Rec Deb RMSE France Gap (*) 0.73 (*) (*) Infl Inrae Bal (*) (*) Exp (*) Rec (*) (*) Deb (*).947 (*).566 (*).473 (*) (*) Ialy Gap Infl Inrae (*) Bal (*).36 (*) (*).378 (*) Exp Rec (*).078 Deb (*) Spain Gap (*) (*) 0.54 (*) Infl (*) Inrae (*).343 (*) (*) (*) Bal (*).7 (*).395 (*) (*) 0.86 (*) (*).586 (*) Exp Rec (*) (*).80 (*).608 (*) Deb Noes: The able enries are he RMSEs of differen models, relaive o he RMSE of a random walk model, for one and wo-sep ahead simulaed forecass. Esimaion sample is 98: 995:. Forecass are performed over he sample 996: 00:. Resuls are repored for ARMA models (ARMA see able xx for deails), one and wo-lag VARs (VAR and VAR), single-counry srucural models (S.C. M see ex for deails), an area-wide model (AW. M see ex for deails), and for pooled forecass (compued each period as he mean and he median of he forecass of all models - MEAN and MED respecively), along wih he RMSE of he random walk model ( RMSE). A es (see Diebold and Mariano (995) and Harvey a al. (997) is also performed on he significance of he mean of he difference beween he squared errors of he differen models and hose of he random walk. (*) denoes 5% significance.

18 Table 5: Relaive MAE Rolling esimaes -Sep ahead -Sep ahead ARMA VAR VAR S.C.M AW.M Mean Med MAE Germany ARMA VAR VAR S.C.M AW.M Mean Med Gap (*) 4.85 (*) Infl Inrae Bal Exp Rec Deb MAE Variable Average France Gap (*) 0.53 (*) (*) (*) (*) 0.59 (*) 0.63 (*) 0.67 (*) Infl Inrae Bal (*) (*) Exp Rec (*) Deb (*).3 (*).939 (*).803 (*) (*) Ialy Gap (*) Infl (*) (*) Inrae Bal (*) (*) (*) (*).489 (*) Exp (*) (*) Rec Deb (*) (*) Spain Gap (*) Infl Inrae (*).95 (*) (*) (*) (*) Bal (*) (*) (*) Exp (*) Rec (*) (*) 3.00 (*).988 (*) Deb Noes: The able enries are he MAEs of differen models, relaive o he MAE of a random walk model, for one and wo-sep ahead simulaed forecass. Esimaion sample is 98: 995:. Forecass are performed over he sample 996: 00:. Resuls are repored for ARMA models (ARMA see able xx for deails), one and wo-lag VARs (VAR and VAR), single-counry srucural models (S.C. M see ex for deails), an area-wide model (AW. M see ex for deails), and for pooled forecass (compued each period as he mean and he median of he forecass of all models - MEAN and MED respecively), along wih he MAE of he random walk model ( MAE) and he average value of he variables over he forecasing sample. A es (see Diebold and Mariano (995) and Harvey a al. (997) is also performed on he significance of he mean of he difference beween he absolue errors of he differen models and hose of he random walk. (*) denoes 5% significance.

19 ARMA VAR VAR S.C.M AW.M Mean Med Table 6: 4-sep ahead forecass Recursive esimaes RMSE MAE RMSE Germany ARMA VAR VAR S.C.M AW.M Mean Med Gap Infl Inrae Bal Exp Rec Deb MAE Variable Average France Gap Infl Inrae Bal (*) (*) Exp 0.90 (*) Rec Deb Ialy Gap Infl Inrae Bal Exp Rec Deb Spain Gap Infl Inrae Bal Exp Rec Deb Noes: The able enries are he RMSEs and MAEs, respecively, of differen models, relaive o hose of a random walk model, for four-sep ahead simulaed forecass. Esimaion sample is 98: 995:. Forecass are performed over he sample 996: 00:. Resuls are repored for ARMA models (ARMA see able xx for deails), one and wo-lag VARs (VAR and VAR), single-counry srucural models (S.C. M see ex for deails), an area-wide model (AW. M see ex for deails), and for pooled forecass (compued each period as he mean and he median of he forecass of all models - MEAN and MED respecively), along wih he RMSE/ MAE of he random walk model ( RMSE/ MAE) and he average value of he variables over he forecasing sample. A es (see Diebold and Mariano (995) and Harvey a al. (997) is also performed on he significance of he mean of he difference beween he squared/absolue errors of he differen models and hose of he random walk. (*) denoes 5% significance.

20 ARMA VAR VAR S.C.M AW.M Mean Med Table 7: 4-sep ahead forecass Rolling esimaes RMSE MAE RMSE Germany ARMA VAR VAR S.C.M AW.M Mean Med Gap Infl Inrae Bal Exp Rec Deb France Gap Infl Inrae Bal Exp Rec Deb MAE Variable Average Ialy Gap Infl Inrae Bal Exp Rec Deb (*) (*) (*) (*) Spain Gap Infl Inrae Bal Exp Rec Deb (*) (*) (*) 0.46 (*) 0.43 (*) Noes: The able enries are he RMSEs and MAEs, respecively, of differen models, relaive o hose of a random walk model, for four-sep ahead simulaed forecass. Esimaion sample is 98: 995:. Forecass are performed over he sample 996: 00:. Resuls are repored for ARMA models (ARMA see able xx for deails), one and wo-lag VARs (VAR and VAR), single-counry srucural models (S.C. M see ex for deails), an area-wide model (AW. M see ex for deails), and for pooled forecass (compued each period as he mean and he median of he forecass of all models - MEAN and MED respecively), along wih he RMSE/ MAE of he random walk model ( RMSE/ MAE) and he average value of he variables over he forecasing sample. A es (see Diebold and Mariano (995) and Harvey a al. (997) is also performed on he significance of he mean of he difference beween he squared/absolue errors of he differen models and hose of he random walk. (*) denoes 5% significance.

21 Table 8: Relaive RMSE Recursive esimaes Comparison wih OECD forecass -Sep ahead -Sep ahead OECD S.C.M AW.M Mean Med RMSE OECD S.C.M AW.M Mean Med RMSE Germany Gap Infl Inrae Bal Deb France Gap Infl Inrae Bal Deb Ialy Gap Infl Inrae Bal Deb Spain Gap Infl Inrae Bal Deb Noes: The able enries are he RMSEs of differen models, along wih hose of OECD forecass (as repored in he OECD Economic Oulook), relaive o he RMSE of a random walk model, for one and wo-sep ahead simulaed forecass. Esimaion sample is 98: 995:. Forecasing sample is 996: 00:. Resuls are repored for single-counry srucural models (S.C. M see ex for deails), an area-wide model (AW. M see ex for deails), and for pooled forecass (compued each period as he mean and he median of he forecass of all models - MEAN and MED respecively), along wih he RMSE of he random walk model ( RMSE).

22 Table 9: Relaive MAE Recursive esimaes Comparison wih OECD forecass -Sep ahead -Sep ahead OECD S.C.M AW.M Mean Med MAE OECD S.C.M AW.M Mean Med MAE Variable Average Germany Gap Infl Enrae Bal Deb France Gap Infl Enrae Bal Deb Ialy Gap Infl Enrae Bal Deb Spain Gap Infl Enrae Bal Deb Noes: The able enries are he MAEs of differen models, along wih hose of OECD forecass (as repored in he OECD Economic Oulook), relaive o he MAE of a random walk model, for one and wo-sep ahead simulaed forecass. Esimaion sample is 98: 995:. Forecasing sample is 996: 00:. Resuls are repored for single-counry srucural models (S.C. M see ex for deails), an area-wide model (AW. M see ex for deails), and for pooled forecass (compued each period as he mean and he median of he forecass of all models - MEAN and MED respecively), along wih he RMSE of he random walk model ( RMSE) and he average values of he variables over he forecasing sample.