No 1015 / February Inflation forecasting in. Member States. by Olga Arratibel, Christophe Kamps and Nadine Leiner-Killinger

Transcription

1 Working Paper Series No 1015 / Inflaion forecasing in he New EU Member Saes by Olga Arraibel, Chrisophe Kamps and Nadine Leiner-Killinger

2 WORKING PAPER SERIES NO 1015 / FEBRUARY 2009 INFLATION FORECASTING IN THE NEW EU MEMBER STATES 1 by Olga Arraibel 2, Chrisophe Kamps 3 and Nadine Leiner-Killinger 4 In 2009 all publicaions feaure a moif aken from he 200 banknoe. This paper can be downloaded wihou charge from hp:// or from he Social Science Research Nework elecronic library a hp://ssrn.com/absrac_id= We would like o hank, wihou implying, Hans-Joachim Klöckers, Ricardo Mesre, Philippe Mouo, Ad van Rie, Philipp Roher, Jean-Pierre Vidal and paricipans a he European Commission Workshop Wha drives inflaion in he new EU Member Saes? for heir valuable commens and suggesions. We also hank an anonymous referee for helpful commens. The views expressed are our own and do no necessarily reflec hose of he European Cenral Bank. 2 Corresponding auhor: European Cenral Bank, EU Counries Division, Kaisersrasse 29, D Frankfur am Main, Germany; olga.arraibel@ecb.europa.eu 3 European Cenral Bank, Moneary Policy Sraegy Division, Kaisersrasse 29, D Frankfur am Main, Germany; chrisophe.kamps@ecb.europa.eu 4 European Cenral Bank, Fiscal Policies Division, Kaisersrasse 29, D Frankfur am Main, Germany; nadine.leiner-killinger@ecb.europa.eu

3 European Cenral Bank, 2009 Address Kaisersrasse Frankfur am Main, Germany Posal address Posfach Frankfur am Main, Germany Telephone Websie hp:// Fax All righs reserved. Any reproducion, publicaion and reprin in he form of a differen publicaion, wheher prined or produced elecronically, in whole or in par, is permied only wih he explici wrien auhorisaion of he or he auhor(s). The views expressed in his paper do no necessarily reflec hose of he European Cenral Bank. The saemen of purpose for he Working Paper Series is available from he websie, hp:// eu/pub/scienific/wps/dae/hml/index. en.hml ISSN (online)

4 CONTENTS Absrac 4 Non-echnical summary 5 1 Inroducion 6 2 Inflaion developmens and inflaion forecass in he NMS: sylised facs Non-moneary deerminans of inflaion in he new EU Member Saes Moneary deerminans of inflaion in he new EU Member Saes 13 3 The empirical framework Daa Mehodology 16 4 Empirical resuls Shor-erm forecas of inflaion developmens (4 quarers ahead) Long-erm forecas of inflaion developmens (12 quarers ahead) 24 5 Conclusions 28 References 29 Appendix 31 European Cenral Bank Working Paper Series 37 3

5 Absrac To he bes of our knowledge, our paper is he firs sysemaic sudy of he predicive power of moneary aggregaes for fuure inflaion for he cross secion of New EU Member Saes. This paper provides sylized facs on moneary versus non-moneary (economic and fiscal) deerminans of inflaion in hese counries as well as formal economeric evidence on he forecas performance of a large se of moneary and nonmoneary indicaors. The forecas evaluaion resuls sugges ha, as has been found for oher counries before, i is difficul o find models ha significanly ouperform a simple benchmark, especially a shor forecas horizons. Neverheless, moneary indicaors are found o conain useful informaion for predicing inflaion a longer (3- year) horizons. Keywords: Inflaion forecasing, leading indicaors, moneary policy, informaion conen of money, fiscal policy, New EU Member Saes JEL Classificaion: C53, E31, E37, E51, E52, E62, P24 4

6 Non-echnical summary In recen years inflaion has acceleraed, someimes very sharply, in mos of he non-euro area EU Member Saes ha have joined he EU since 2004 (hereafer, he new EU Member Saes: NMS). The recen rise in inflaion has been largely aribued o a series of adverse supply shocks, such as energy and food price increases, as well as o ongoing changes in he economic srucures of hese caching-up economies. A leas in some counries, capaciy pressures also seem o have played a role. Accordingly, he rise in inflaion would appear by and large ineviable and, as regards he occurrence of adverse shocks, unpredicable. Indeed, over he pas few years, inflaion forecass by inernaional organisaions and oher professional forecasers appear o have sharply underesimaed he recen rise in inflaion in he NMS. The quesion arises wheher he recen rise in inflaion could have been prediced beer by accouning for he signals of moneary indicaors like rend money growh, which have no received much aenion in discussions of recen inflaion developmens in he NMS. Agains his background, his paper (1) analyses poenial economic and fiscal variables as explanaory variables for inflaion developmens and (2) akes a close look a he leading indicaor properies of moneary and credi aggregaes for fuure inflaion. I focuses on inflaion developmens in eigh Cenral and Eas European EU Member Saes, namely Bulgaria, he Czech Republic, Esonia, Lavia, Lihuania, Hungary, Poland and Romania. To he bes of our knowledge, his paper is he firs sysemaic sudy o analyse he performance of a large se of moneary, economic and fiscal indicaors in explaining inflaion dynamics in he NMS, including bivariae models and rivariae (wo-pillar Phillips curve) models. Three findings can be highlighed. Firs, we find ha he performance of alernaive forecasing models differs subsanially across counries in ha differen variables seem o be influencing o a differen exen inflaion developmens a differen forecas horizons. Second, a he shor-erm (one-year) horizon, wih he excepion of a few counries, i is difficul o find models wih forecass significanly ouperforming hose of a simple benchmark (he random walk model). This finding is in line wih empirical resuls for he U.S. and oher OECD counries suggesing ha inflaion has become increasingly hard o forecas. In he few counries for which we find forecasing models significanly ouperforming he benchmark a he one-year horizon, such models are in general based on economic and fiscal indicaors. Third, we find ha, over he longer erm (hree-year) forecasing horizon, moneary indicaors conain useful informaion for predicing inflaion in mos NMS counries. Ineresingly, he role played by moneary facors in driving inflaion dynamics does no differ subsanially beween he NMS wih a fixed exchange rae regime and hose wih a floaing exchange rae regime. Thus, moneary indicaors appear o become imporan variables o explain inflaion rends across he majoriy of he NMS over he long-erm forecasing horizon. 5

7 Inflaion is always and everywhere a moneary phenomenon (Milon Friedman (1963), Inflaion: Causes and Consequences, Asia Publishing House, New York.) 1. Inroducion In recen years inflaion has acceleraed, someimes very sharply, in mos of he non-euro area EU Member Saes ha have joined he EU since 2004 (hereafer, he new EU Member Saes: NMS). The recen rise in inflaion has been largely aribued o a series of adverse supply shocks, such as energy and food price increases, as well as o ongoing changes in he economic srucures of hese caching-up economies. A leas in some counries, capaciy pressures also seem o have played a role. Accordingly, he rise in inflaion would appear by and large ineviable and, as regards he occurrence of adverse shocks, unpredicable. Indeed, over he pas few years, inflaion forecass by inernaional organisaions and oher professional forecasers appear o have sharply underesimaed he recen rise in inflaion in he NMS. The quesion arises wheher he recen rise in inflaion could have been prediced beer by accouning for he signals of moneary indicaors like rend money growh, which have no ye received much aenion in discussions of recen inflaion developmens in he NMS. The exen of moneary accommodaion could be expeced o be a key deerminan of he ransmission of shocks o inflaion over medium-erm o long-erm horizons. Agains his background, his paper aemps o answer he following quesions: To wha exen do he above menioned non-moneary facors explain he rise in inflaion in hese counries? Why has he rise in inflaion over recen years been so heerogeneous across he NMS in he presence of common exernal shocks and similar levels of economic developmen? Wha is he role of money in explaining inflaionary developmens in he NMS? Could he rise in inflaion, and also is heerogeneiy across NMS, have been prediced? To address hese quesions his paper (1) analyses poenial economic and fiscal variables as explanaory variables for inflaion developmens and (2) akes a close look a he leading indicaor properies of moneary and credi aggregaes for fuure inflaion. I focuses on inflaion developmens in eigh Cenral and Eas European EU Member Saes, namely Bulgaria, he Czech Republic, Esonia, Lavia, Lihuania, Hungary, Poland and Romania. Depending on daa availabiliy, he ime period analysed in his paper is 1993Q1 o 2008Q2. To he bes of our knowledge, our paper is he firs sysemaic sudy of he predicive power of moneary aggregaes for fuure inflaion in he eigh NMS. Individual-counry sudies exis for a few NMS, see e.g. Dabusinskas (2005) and Kolowski (2005). These sudies in general conclude ha moneary aggregaes provide useful informaion for forecasing inflaion. Our forecas evaluaion exercise builds on he empirical framework firs applied by Sock and Wason (1999) o U.S. daa and subsequenly applied by Nicolei-Alimari (2001), 6

8 Carsensen (2007) and Hofmann (2008) o euro area daa. 1 In his exercise, we consider a large se of moneary and non-moneary (economic and fiscal) indicaors. We find ha he performance of alernaive forecasing models differs subsanially across counries in ha differen variables seem o be influencing o a differen exen inflaion developmens a differen forecas horizons. We also find ha a he 4-quarer horizon, wih he excepion of a few counries, i is difficul o find models wih forecass significanly ouperforming hose of a simple benchmark (he random walk model). This resul is in line wih he findings of Sock and Wason s (2008) lieraure review suggesing ha inflaion has become increasingly hard o forecas in he U.S. and oher OECD economies. However, our resuls also sugges ha, over he longer erm (3-year) forecasing horizon, moneary indicaors conain useful informaion for predicing inflaion in mos NMS counries. The paper is organised as follows. Secion 2 firs reviews pas inflaion developmens and assesses inflaion forecass for he eigh NMS. I hen provides sylised facs on moneary and non-moneary (economic and fiscal) facors ha have conribued o inflaion developmens in he NMS. Secion 3 discusses he daa and mehodology used in he forecas evaluaion exercise and Secion 4 presens he empirical resuls. Secion 5 concludes, highlighing he implicaions of moneary indicaors for fuure inflaion developmens in he NMS. 2. Inflaion developmens and inflaion forecass in he NMS: sylised facs Since he end of he 1990s unil he beginning of his decade inflaion raes in mos NMS followed a broadly downward pah from relaively high levels (see Figure 1). A differen poins in ime beween 2003 and 2005, his paern sared o change quie noiceably. Firs, he disinflaion process was inerruped in all counries bu Romania. Second, inflaion sared o increase again, in some counries o remarkably high levels. Third, while he process of disinflaion unil around 2003 had aken place irrespecive of he exchange rae sraegy chosen, he rise in inflaion over he period has been paricularly virulen in he counries ha have oped for hard-pegs (i.e. he Balic Saes and Bulgaria). In he oher counries, which follow an inflaion-argeing sraegy allowing for a higher degree of exchange rae flexibiliy (hereafer, he floaers ), he surge in inflaion has been generally more conained. Alhough he economic cycle seems o have peaked in 2008 in many cases, mos noably in he Balic Saes, srong price pressures remain (see for a survey of economic indicaors for he NMS Table A.1 in he Appendix). In such a changing environmen, rying o predic he behaviour of price dynamics has become paricularly demanding. This is illusraed in Figure 2 below, which shows for he 1 For oher mehodologies o forecas inflaion in he euro area see e.g. De Sanis e al. (2008), Hubrich (2005), Calza e al. (2001). These models end o require relaively large numbers of observaions and a sable environmen and are hus less suied for forecasing inflaion in he NMS. 7

9 eigh counries considered and for he years , he annual average inflaion rae and he respecive inflaion forecas ha was made 18-monhs ahead, prior o he release of he final oucome. For example, he forecas for he average annual inflaion in 2004 is he one made in May-June The forecass have been aken from he Easern Europe Consensus Forecas. Figure 1: 4-quarer average inflaion in NMS (1998Q2 2008Q2) 20 Esonia 20 Lavia 20 Lihuania 20 Bulgaria Czech Republic 20 Hungary 20 Poland 50 Romania Noe: 4-quarer average inflaion is defined as 100*ln(CPI /CPI -4 ). Ineresingly, unil 2003 inflaion oucomes surprised on average on he downside: wih he main excepion of he floaers in he year 2000, inflaion urned ou raher close o and, a imes, below he inflaion forecas expeced by he Consensus Forecas 18 monhs before. Thereafer, however, inflaion has surprised on he upside, paricularly in he las wo years. Moreover, developmens across counries have sared o diverge. Specifically, forecasing inflaion dynamics in he hard-peg counries seems o have become a paricularly challenging ask, wih he forecass having underprediced inflaion five years in a row. The quesion arises wheher his sysemaic underpredicion of fuure inflaion was unavoidable given a series of unpredicable unfavourable shocks or wheher he forecass missed imporan informaion ha would have allowed avoiding a downward bias. 8

10 Figure 2 Inflaion forecass and acual inflaion ( ) "HARD-PEGS" - Inflaion and 18-monhs ahead inflaion forecass* (annual average, CPI index, %) Source: Easern Europe Consensus Forecass and IMF (Inernaional Financial Saisics). * Unweighed average of annual inflaion raes in Bulgaria, Esonia, Lavia and Lihuania. The arrows heading up (down) denoe higher (lower) han expeced inflaion.the value for 2008 indicaes he laes daa available (Augus 2008). "FLOATERS" - Inflaion and 18-monhs ahead inflaion forecass* (annual average, CPI index, %) Source: Easern Europe Consensus Forecas and IMF (Inernaional Financial Saisics). * Unweighed average of annual inflaion raes in he Czech Republic, Hungary, Poland and Romania. The arrows heading up (down) denoe higher (lower) han expeced inflaion. The value for 2008 indicaes he laes daa available (Augus 2008). 9

11 The forecas evaluaion exercise presened in Secions 3 and 4 aemps o shed ligh on his quesion. In his exercise he naïve random walk model will serve as he benchmark o evaluae he performance of alernaive forecasing models. As has been shown in he forecasing lieraure, i has proven difficul o find model-based forecass or forecass by professional forecasers ha significanly ouperform his simple benchmark (see, e.g., Sock and Wason 2007). Visual inspecion of Figure 2 confirms ha for our se of counries he Consensus forecass in general did no ouperform he simple no-change forecas implied by he random walk model. Going beyond visual impression, we compued a simple meric of forecas performance, he roo mean squared forecas error, over he period boh for he Consensus forecass and for he random walk model (in he laer case using he average annual inflaion rae over he previous welve monhs in June of a given year as he forecas for average annual inflaion in he nex year). These compuaions confirm ha across our se of counries on average he random walk model performed equally well han he Consensus forecass, wih he laer performing slighly beer han he random walk model in he floaing counries bu worse in he hard-peg counries. 2 Given hese findings and also because only oo narrow a se of Consensus forecass is available, we use he random-walk model as he benchmark in our formal forecas evaluaion exercise presened below. Before urning o his exercise, several non-moneary economic and fiscal as well as moneary facors are analysed wih respec o heir poenial impac on inflaion developmens in he NMS. 2.1 Non-moneary deerminans of inflaion in he new EU Member Saes The rise in inflaion in he NMS has generally aken place agains he background of dynamic economic condiions, wih exernal facors also having played a role. On he domesic side, srong domesic demand, underpinned by robus growh in disposable income, large inflows of foreign direc invesmen, low real ineres raes and buoyan credi growh, have conribued o inflaionary pressures in mos NMS. 3 Moreover, he rapid ighening of he labour marke, exacerbaed by labour ouflows o oher EU counries, has conribued o rapid increases in wages, ofen significanly above labour produciviy growh, hence leading o high growh in uni labour cos, paricularly in he fases growing economies. Under such circumsances, capaciy consrains and signs of overheaing emerged in many counries, e.g he Balic Saes, Bulgaria, he Czech Republic and Romania. 4 2 Romania is a clear oulier, for which he random-walk model performs very poorly because i misses he srong downward rend in inflaion over he sample period, which a leas o some exen was predicable. 3 The main excepion would be Hungary, where oupu growh has remained relaively subdued. 4 In mos NMS, noably in hose ha have oped for a hard-peg o he euro, he process of caching-up in real income wih he euro area is likely o manifes iself in higher inflaion. However, he exac size of such effec is difficul o assess. According o mos sudies, i generally accouns for beween 0.4 and 2.4 percenage poins of inflaion in he NMS (see Darvas and Szapáry (2008)). 10

12 As for he exernal facors, among he mos imporan drivers of inflaion in he NMS are food prices, which had been declining in bu have since rebound sharply, and soaring energy prices. These wo componens have a relaively large weigh in he consumpion baske of hese counries, represening around 43% of he oal (61.2% in he case of Romania). In some counries wih flexible exchange raes, hese price increases have been parly dampened by an appreciaing currency (i.e. he Czech Republic and Romania). Moreover, he impac of globalisaion and srong compeiion, noiceably from emerging Asia, seem o have generally conribued o conaining growh raes in he price of radable goods (e.g. exiles, furniure, IT equipmen) in many NMS. Also fiscal policy, hrough he impac of, iner alia, public finance reform, he fiscal sance, changes in direc and indirec axaion, public wage-seing as well as he liberalisaion of adminisered prices in he wake of EU enry affeced inflaion performance in he NMS. In several of he eigh NMS considered here, direc axes and social securiy conribuions were lowered in an aemp o reduce ax wedges and hus increase labour supply in he presence of labour shorages and someimes large shadow economies in hese fas growing economies. To he exen ha capaciy consrains were lowered, hese measures conribued o conaining inflaion. A he same ime, o he exen ha hese measures raised domesic demand, hey ended o conribue o a rise in inflaion. Furhermore, in several NMS, public wages rose significanly. As Figure 3 indicaes, he increase in governmen spending on compensaion of public employees in he general governmen secor reached double digi values in he Balic counries, Hungary and Poland beween 1995 and In Bulgaria and he Balic counries, which faced a srong pick-up in inflaion since 2003, compensaion of public employees over his period coninued o rise a hese high raes. A he same, apar from Romania, i fell in hose NMS wih disinflaion, such as in Hungary and Poland, or coninued o grow a relaively lower growh raes, such as in he Czech Republic. 11

13 Figure 3: Compensaion of public employees, general governmen (annual average percenage changes) % Bulgaria Czech Republic Esonia Lavia Lihuania Hungary Poland Romania Noes: For Romania , for Hungary Source: ESCB, calculaions. Furhermore, he liberalisaion of adminisered prices in secors such as e.g. energy and ranspor conribued o price increases ha varied across he NMS. As regards indirec axes, over he las decade, hree of he eigh NMS considered lowered heir VAT raes, parly subsanially in view of EU ax harmonisaion, namely he Czech Republic, Hungary and Romania. 5 While in Romania hese measures ook place in a disinflaion environmen, hey were followed by a pick-up in inflaion in he ohers. Moreover, mos of he NMS receive parly subsanial EU Srucural Funds wihin he EU cohesion policy framework ha may affec inflaion. In 2007, he allocaion of EU Funds o he eigh NMS considered here amouned o abou 2½% of GDP on average. While hese funds are aimed a improving longerm growh and employmen prospecs in he suppored regions hey may in he shor run raise domesic demand significanly, hus raising inflaion paricularly a imes of capaciy consrains. 6 As regards fiscal policy, overall, fiscal effors o comply wih he requiremens of he Sabiliy and Growh Pac ended o foser fiscal consolidaion in he NMS ha should in 5 In Hungary, he VAT rae was lowered by 5 percenage poins o 20% in 2006, and in he Czech Republic by 3 percenage poins o 19% in 2003 (see Figure A.1 in he Appendix). In Romania he VAT rae was reduced by 3 percenage poins o 19% in See for deails Kamps e al. (2009). I is, however, difficul o gauge he impac of hese funds on inflaion as one has o disenangle allocaions from acual paymens, which do no coincide in general. Due o daa limiaions, we canno accoun for he role of EU funds in our forecasing exercise. 12

14 principle have worked in he direcion of inflaion conainmen. Neverheless, one may conjecure ha, a imes, a loosening fiscal sance, resuling from e.g. high capial spending in he presence of large infrasrucure needs and measures o raise labour supply, may have conribued o an increase in inflaion. In addiion, a some insances, he fiscal sance has no been igh enough o sufficienly conribue o inflaion conainmen (see Figure A.2 in he Appendix) Moneary deerminans of inflaion in he new EU Member Saes The economic and fiscal deerminans of inflaion developmens in he NMS share he following feaure: i is ofen unanicipaed shocks o hese variables and heir more or less sluggish propagaion wha drives inflaion developmens over he shor o medium run. Shocks o food and energy prices, for example, no only heavily affec relaive prices bu, in he shor o medium run, also srongly affec overall inflaion. However, in he long run such shocks need no impac on overall inflaion unless moneary policy chooses o a leas parly accommodae hem. This subsecion presens one indicaor of moneary-policy induced inflaionary pressures, namely he evoluion of rend money growh. The empirical lieraure has esablished a srong link beween money growh and inflaion for OECD counries, looking a daa ha span more han a cenury (see, e.g., Lucas 1996 and Benai 2008). This evidence suggess ha surges in money growh in general precede highinflaion episodes, making money growh a poenially useful indicaor for fuure inflaion. More recenly, Assenmacher-Wesche and Gerlach (2007, 2008) showed ha he empirical link beween money growh and inflaion is paricularly srong a low frequencies, while he relaionship is raher noisy a business cycle frequencies. Figure 4 shows he low-frequency componens of annualised quarer-on-quarer money growh and inflaion for he eigh NMS under review for he period 1995 o 2008, where he low-frequency componens of he wo series have been esimaed wih a Hodrick-Presco filer. The figure reveals a number of ineresing findings. Firs, here has been pronounced disinflaion in all eigh counries in he second half of he 1990s, accompanied by a srong deceleraion in rend money growh. Second, rend inflaion and rend money growh sared o accelerae in all counries excep for Hungary and Romania around he sar of he curren cenury. Third, in all hese counries he urning poin for rend money growh preceded he urning poin for rend inflaion. On average, rend money growh reached is low poin five quarers before rend inflaion (median: four quarers) did, wih he lead being longes in he case of Lavia (almos hree years) and shores in he case of he Czech Republic (half a year). Also, he pick-up in rend money growh suggess ha he recen surge in inflaion in he NMS is no exclusively due o he occurrence of unanicipaed shocks bu a leas parly 13

15 has moneary roos. 7 Fourh, in some counries (Esonia, Lavia and Lihuania) rend money growh peaked in 2005 and deceleraed hereafer, suggesing gradually declining inflaionary pressures in hese counries. All in all, he sylised evidence presened in his subsecion suggess ha rend money growh conains useful informaion for predicing fuure inflaion, a hypohesis ha will be esed in he remainder of his paper. Figure 4: Trend money growh and inflaion in NMS (1995Q1 2008Q2) Noe: Money growh based on broades available moneary aggregae (M3 for Bulgaria, Hungary and Lihuania, M2 for remaining counries). Trend esimaed wih Hodrick-Presco filer (uning parameer se equal o 1600). 3. The empirical framework This secion presens he empirical framework used o evaluae he performance of alernaive forecasing models. The firs subsecion discusses he daa, while he second subsecion presens he alernaive forecasing models used for his exercise. The empirical resuls are presened in Secion Daa We use daa for eigh NMS, namely Bulgaria, he Czech Republic, Esonia, Lavia, Lihuania, Hungary, Poland and Romania. Mos series were rerieved from he Haver Analyics DLX daabases Eurosa, EMERGECW and IFS (see Appendix A.2 for deails). Mos ime series are seasonally adjused by he source. Whenever his is no he case, we seasonally adjus he respecive series using he Census X11 muliplicaive procedure. Depending on daa availabiliy, we use quarerly daa for 1993Q1 o 2008Q2. However, daa availabiliy differs subsanially across counries and for some counries some series sar only in he end- 7 Noe ha even he recen surge in commodiy prices (one of he shocks ) migh have been parly driven by loose moneary policy, albei a he global level (Frankel 2006). 14

16 1990s. For he forecasing regressions, we only used daa ha were available a leas from 1995Q1 onwards. For Bulgaria and Romania, daa availabiliy is paricularly limied and also daa qualiy appears o be an issue, suggesing ha he resuls for hese counries should be aken wih a grain of sal. We group he explanaory variables ha we consider in he forecasing exercise ino hree groups, namely moneary, economic and fiscal indicaors. The moneary indicaors esed include he moneary aggregaes M1, M2 and M3 as well as loan growh. For hose counries for which M3 is no available we use M0, M1 and M2 insead. The moneary indicaor variables also include conceps such as rend broad money growh, he real money gap, he moneary overhang and he change in P-sar (see Appendix A.1 for deails). Economic indicaors include boh domesic and exernal facors. Domesic facors comprise he oupu gap, real GDP growh, employmen growh, a measure of he employmen gap, he change in he unemploymen rae, a measure of he unemploymen rae gap, growh in compensaion per employee, wage growh, indusrial producion growh and, finally, he level and change in producers selling price expecaions as well as consumers price expecaions based on surveys compiled by he European Commission. Exernal facors include he nominal effecive exchange rae, commodiy price inflaion as well as changes in oil prices. Commodiy prices and oil prices expressed in euro are aken from he Monhly Bullein daabase and convered ino naional currency using spo exchange raes. Fiscal variables comprise growh in general governmen revenue and expendiure, changes in direc and indirec axes, changes in VAT revenues, changes in governmen consumpion and growh in compensaion of governmen employees. Unforunaely, a long enough ime series of he laer variable is available for he Czech Republic only. However, given ha compensaion of governmen employees is by far he larges sub-iem in governmen consumpion, one can inerpre changes in governmen consumpion as a proxy for governmen-induced wage pressures. The rend and gap measures referred o above (e.g. rend money growh and he oupu gap) are consruced as follows. Firs, he rend componen of he respecive variable is esimaed wih a Hodrick-Presco filer (wih uning parameer se equal o 1600). We follow Sock and Wason (1999) and Hofmann (2008), using a one-sided filer o esimae he rend componens. This filer for each period uses only informaion up o ha period, which is imporan for he pseudo real-ime ou-of-sample forecasing exercise we carry ou. We esimae he one-sided Hodrick-Presco filer using he Kalman filer. Once he rend componen has been esimaed, he gap measure is obained as he (log) difference beween he raw series and he Hodrick-Presco rend. 15

17 3.2 Mehodology The forecas evaluaion exercise builds on he empirical framework firs applied by Sock and Wason (1999) o U.S. daa and subsequenly applied by Nicolei-Alimari (2001), Carsensen (2007) and Hofmann (2008) o euro area daa. 8 To he bes of our knowledge, our sudy is he firs aemp o apply his mehodology o sudy he performance of alernaive forecasing models for he NMS. The following descripion of he mehodology closely follows Hofmann (2008). As is common in his lieraure, we concenrae on single-equaion forecasing models wih few independen variables for he following reasons. Firs, single-equaion inflaionforecasing models have been found in he forecas evaluaion lieraure o ouperform large muliple-equaion models such as vecor auoregressions or radiional srucural macroeconomeric models (see, e.g., Sock and Wason 2008). Second, we concenrae on models wih few independen variables in order o accoun for our shor daa samples. Increasing he number of indicaor variables included in he models would quickly exhaus he degrees of freedom. Ye, in order o reflec he large se of informaion available o professional forecasers in pracice, we also presen resuls for combined forecass ha average over he large se of forecasing models (beween 40 and 60 per counry) we consider in he exercise. To begin wih, he variable o be forecased has o be chosen. I has become common pracice in he lieraure o focus on annualised average inflaion over he coming h quarers, which h can be formally expressed as 400/ h)ln( CPI / CPI ) and where denoes he h ( h inflaion rae and CPI denoes he consumer price index. In he empirical exercise we consider hree alernaive forecasing horizons, namely 4-quarer ahead forecass, 8-quarer ahead forecass and 12-quarer ahead forecass (h = 4, 8, 12). Benchmark model for forecas evaluaion In order o evaluae he forecasing performance of alernaive models, we compare heir forecas errors wih hose obained for a simple benchmark model. Two candidae benchmark models have been rouinely used in he forecasing model. Firs, he naïve random walk model (hereafer RW), which saes ha he forecas of h-quarer ahead inflaion a ime is simply he las observable value of h-quarer inflaion: (1) ˆ, h h h 1 8 See Fischer e al. (2006) for a discussion of he role hese forecasing models play in moneary analysis. 16

18 h where 400/ h)ln( CPI / CPI ). Here, we assume as is he case in pracice ha 1 ( 1 h 1 h he inflaion rae a ime,, is no par of he ime- informaion se. Second, a univariae auoregressive model for inflaion (hereafer AR), which can be expressed as: (2) h h 1( ) h 0 L u h, where 1( L ) is a finie polynomial of order p in he lag operaor L ( p L L L 1 ( ) p ). Again, we assume ha he inflaion rae a ime,, is no par of he ime- informaion se. We follow Carsensen (2007) and use a sepwise procedure o deermine which lagged values of inflaion ener he model, considering up o four lagged values of inflaion ( p 4 ). 9 The h- h quarer ahead forecas, afer model esimaion, is calculaed as ˆ ˆ ( L) ˆ h 0 1. The resuls presened in Secion 4 sugges ha, in he cross-secion of counries considered in his paper and across forecasing horizons, he random walk model on average ouperforms he auoregressive model, in he sense ha i produces lower roo mean squared errors (RMSE). Therefore, we pick he random walk model as he benchmark model when esing wheher he alernaive forecasing models presened below significanly ouperform he benchmark. For his purpose, we use he es for equal predicive abiliy developed by Diebold and Mariano (1995), which is applicable in he case of non-nesed models. 10 Bivariae forecasing models The firs se of forecasing models used in he forecasing exercise exends he univariae auoregressive model discussed above by including one indicaor variable, x, a a ime: (3) L) ( L) x u, h h 0 1( 2 h h where ( ) is a polynomial of order p in he lag operaor L. Again, we assume ha he 2 L respecive indicaor variable a ime, x, is no par of he ime- informaion se. A sepwise procedure is used o deermine which lagged values of inflaion and of he respecive indicaor variable ener he model, considering up o four lagged values of boh inflaion and he 9 Lagged values ener he model only if hey are significan a a pre-specified significance level (here he overall significance level is se o 0.2). Compared o lag-lengh selecion crieria he sepwise procedure has he advanage of addiional flexibiliy. For example, i migh be ha lags 1 and 4 urn ou o be highly significan bu lags 2 and 3 are insignifican. A convenional selecion crierion would probably pick a model wih one lag only as he bes model. However, his would miss he informaion conained in he fourh lag, which may srongly affec forecas accuracy. The sepwise procedure, insead, would include lags 1 and 4 in he model. 10 Noe ha some of he models are nesed, implying non-sandard asympoic disribuions under he null hypohesis, see Clark and McCracken (2005). Since forecass are muliple-sep, he proper size of he ess would need o be derived by Monecarlo simulaion. As doing his for all he models considered in his paper would be compuaionally prohibiive, we oped for he Diebold-Mariano es, noing is limiaions in some cases. 17

19 respecive indicaor variable ( p 4). The h-quarer ahead forecas, afer model esimaion, is h calculaed as ˆ ˆ ( L ) ˆ ( L) x ˆ h In he empirical exercise he bivariae models are esimaed for each counry and each indicaor variable in urn. The maximum number of bivariae forecasing models is 32 per counry, given ha we have up o 8 moneary indicaors, up o 17 economic indicaors and up o 7 fiscal indicaors. 11 Trivariae (wo-pillar Phillips curve) forecasing models Gerlach (2003, 2004) proposed a simple rivariae (wo-pillar Phillips curve) forecasing model inended o capure a he same ime he informaion of boh moneary and economic indicaors. Such rivariae (wo-pillar Phillips curve) forecasing models specify h-quarer ahead average inflaion as a funcion of is own lags, lags of rend broad money growh ( ), esimaed wih he one-sided Hodrick-Presco filer discussed above, as well as lags T m of a non-moneary indicaor variable ( x ): (4) ( L) m u, h h 0 1( L) 2 ( L) x 3 T h h where 3( L) is a polynomial of order p in he lag operaor L. Again, we assume ha rend T broad money growh a ime, m, as well as he respecive indicaor variable a ime, x, are no par of he ime- informaion se. A sepwise procedure is used o deermine which lagged values of inflaion, of rend money growh and of he respecive indicaor variable ener he model, considering up o four lagged values of boh inflaion and he respecive indicaor variable ( p 4). The h-quarer ahead forecas, afer model esimaion, is calculaed h T as ˆ ˆ ( L) ˆ ( L) x ˆ ( L m. ˆ h ) In he empirical exercise he rivariae (wo-pillar Phillips curve) models are esimaed for each counry and each indicaor variable in urn. The maximum number of rivariae (wopillar Phillips curve) forecasing models is 24 per counry, given ha we have up o 17 economic indicaors and up o 7 fiscal indicaors Given ha no all series are available for each counry, he effecive number of bivariae forecasing models ranges beween 20 (Romania) and 31 (Czech Republic, Esonia, Lavia). 12 Given ha no all series are available for each counry, he effecive number of rivariae forecasing models ranges beween 14 (Romania) and 24 (Czech Republic). 18

20 Forecas combinaions In addiion o he individual-indicaor bivariae and rivariae (wo-pillar Phillips curve) forecasing models we also provide evidence on he performance of simple forecas combinaions. We consider five alernaive forecas combinaions: combining he forecass of (i) he moneary-indicaor bivariae models, (ii) he economic-indicaor bivariae models, (iii) he fiscal-indicaor bivariae models, (iv) all bivariae models, and (v) all rivariae (wo-pillar Phillips curve) models. All hese forecas combinaions are compued as he simple average of he forecass produced by he models wih he individual indicaors, for each poin in ime, forecasing horizon and counry. The reason for including hese forecas combinaions is ha such combinaions have been shown o significanly improve forecas accuracy in previous sudies (see, e.g., Sock and Wason 1999 and Hofmann 2008). Also, forecas combinaions based on simple averages have been shown o perform equally well or even beer han combinaions assigning differen weighs o he forecass produced by differen models according o some opimaliy crierion. This resul suggess ha forecas combinaions based on simple averages can bes deal wih uncerain srucural breaks ha may heavily affec he forecas performance of individual models. In his conex, i may be wise no o assign oo much weigh o a paricular model. These consideraions are paricularly relevan for our se of counries, characerised by an ongoing convergence process and subsanial srucural change. Forecas evaluaion sample In he simulaed ou-of-sample forecasing exercise, he individual forecasing models are esimaed using daa prior o he forecasing period in such a way as o reflec he informaion se available o forecasers in real ime. Noe, however, ha since we lack a real-ime daabase we have o rely on revised daa from he mos recen vinage of official saisics. In his sense, we conduc a pseudo real-ime exercise. The sample period for he recursive forecasing regressions sars in 1998Q2 for all models, in order o allow for he long lags in he dependen and independen variables. Take he bivariae model wih h=12 as an example. The observaion of h-quarer inflaion for 1998Q2 (which is price level in 1998Q2 divided by price level in 1995Q2) is modelled as depending on values of he indicaor variable in 1995Q1 (lag 1), 1994Q4 (lag 2), 1994Q3 (lag 3), 1994Q2 (lag 4). This is he earlies observaion of h-quarer inflaion we can model given he daa consrains we face. The forecas evaluaion sample runs from 2003Q3 o 2008Q2 and is, hus, based on quarerly observaions for five years. In general, for each forecas horizon h he firs ou-of-sample 19

21 forecas is consruced in period 2003Q3-h in order o ensure ha for each forecasing horizon we obain a forecas evaluaion sample of equal lengh. Take he forecas for 2003Q3 and he bivariae model wih h=4 as an example. This forecas is made in 2002Q3 and is based on a forecasing regression using daa up o 2002Q2. Once he forecas for a given quarer in he forecas evaluaion sample has been compued, he procedure moves one quarer forward and uses one addiional daa poin per sep o esimae he forecasing regression and o consruc he forecas. The procedure sops afer he forecas for 2008Q2 has been consruced, he las period in our daa sample. 4. Empirical resuls This secion discusses he forecas evaluaion resuls for he above-menioned bivariae and rivariae (wo-pillar Phillips curve) models using alernaive moneary and non-moneary (economic and fiscal) variables over he hree alernaive forecasing horizons. I focuses on he shor-erm horizon (4-quarer ahead) and he long-erm horizon (12-quarer ahead); he resuls for he medium-erm horizon (8-quarer ahead) are displayed in he Appendix (see Table A.2). In presening he empirical resuls, counries have been grouped according o heir choice for exchange rae flexibiliy (i.e. Bulgaria, Esonia, Lavia and Lihuania as hardpegs and he Czech Republic, Hungary, Poland and Romania as floaers ) in an aemp o idenify possible differences in he abiliy o forecas inflaion rends among hese wo groups of counries. As menioned above, he performance of he bivariae and rivariae (wo-pillar Phillips curve) models is assessed by comparing heir respecive RMSE wih he RMSE obained from a simple random walk (RW). Three findings become immediaely apparen jus from a visual inspecion of he resuls. Firs, he resuls differ subsanially across counries in ha differen variables seem o be influencing o a differen exen inflaion developmens a differen forecas horizons. This is likely o derive from he fac ha, nowihsanding some similariies, economic srucures and insiuions do vary across he NMS. Moreover, moneary and fiscal policies and srucural reform in hese counries have addressed o a differen exen counry-specific policy requiremens reflecing, for example, differen saes of real convergence and/or differen policy opions. Second, a he 4-quarer horizon, wih he excepion of Bulgaria, Esonia and Hungary, none of he bivariae and rivariae (wo-pillar Phillips curve) forecasing models produce a RMSE ha is significanly smaller han ha of a RW model. This resul echoes he findings of Sock and Wason s (2008) recen lieraure review suggesing ha inflaion is hard o forecas in he U.S. and oher OECD economies and ha i is difficul o improve upon simple benchmark models. Third, a he 12-quarer horizon a number of bivariae models and rivariae (wo-pillar Phillips curve) models significanly improve upon he forecass of he 20

22 RW model. In paricular, moneary indicaors do appear o conain useful informaion for predicing inflaion a longer (3-year) horizons in hese counries. 4.1 Shor-erm forecas of inflaion developmens (4 quarers ahead) Following he mehodology described in Secion 3.2, Table 1 displays he absolue RMSE for he RW as well as he values of he RMSEs of he differen models relaive o he RW. Hence, relaive RMSEs smaller han uniy correspond o models ouperforming he RW. When marked bold, relaive RMSEs smaller han uniy correspond o models ha significanly ouperform he RW (by reference o a Diebold-Mariano es). Looking a Table 1, he resuls show ha a he 4-quarer horizon he bivariae and rivariae (wo-pillar Phillips curve) models perform worse han he RW benchmark in all counries bu Bulgaria, Esonia and Hungary. This confirms ha a he shor-erm horizon i is difficul o improve he forecas performance upon his simple RW benchmark. In he hree counries where eiher bivariae or rivariae (wo-pillar Philipps curve) models are able o significanly ouperform he RW, i appears ha differen facors can help explaining inflaion developmens as indicaed by he bes-performing models. In Bulgaria he bivariae model wih commodiy price inflaion performs bes, possibly reflecing he relaively large weigh of food and energy prices in he consumer s baske of his economy. The bivariae model wih growh of compensaion per employee and he rivariae (wo-pillar Phillips curve) model wih rend M3 growh and commodiy price inflaion also significanly ouperform he RW and o a lesser exen also improve upon he AR model. In Esonia he bivariae model wih ULC growh performs bes. The bivariae model wih producers' price expecaions also performs significanly beer han eiher he RW or he AR model (he laer performs worse han he RW for Esonia). In addiion, he rivariae (wo-pillar Phillips curve) models wih he oupu gap, he unemploymen gap and ULC growh perform beer han he RW. Wih he excepion of ULC growh, hese rivariae (wo-pillar Phillips curve) models perform beer han he bivariae models wih hese economic variables, which suggess ha combining a moneary indicaor (here: rend M2 growh) wih economic indicaors in some cases improves forecas accuracy. In Hungary he bivariae models wih consumers price expecaions and M1 growh perform bes. Boh models display considerably beer forecasing performance han he AR model, which iself performs significanly beer han he RW. The bivariae 21

23 models wih loan growh, he change in he unemploymen rae and governmen consumpion growh also perform significanly beer han he RW even hough he improvemen upon he AR model is no very sizeable. Ineresingly, Hungary wih is high and raher volaile general governmen defici is among he few counries for which we find ha fiscal policy has predicive power for inflaion in he shor run. As regards he forecas combinaions a he 4-quarer horizon, hey are no able o significanly ouperform he RW forecas in any NMS apar from Bulgaria and Hungary. In Hungary, he combined forecas of moneary indicaors performs he bes among hese combined forecass. Overall, we find ha when rying o forecas inflaion developmens a he 4-quarer horizon, in he few counries in which we find models performing significanly beer han he RW, he bes performing models are bivariae models wih economic indicaors. In Esonia and Bulgaria some labour marke indicaors seem o be paricularly capable o ouperform he RW. This may broadly reflec a labour marke ighening over he sample period in hese fas growing and a imes overheaing economies. 22

24 Table 1. Relaive roo mean squared error for alernaive forecasing models (forecas horizon = 4 quarers ahead) Hard-Pegs Floaers BG EE LV LT CZ HU PL RO Random walk model (RMSE) Auoregressive model Bivariae models (a) Moneary indicaors M3/M2 (change) Trend M3/M2 (change) P* (change) Real money gap Moneary overhang M1/M0 (change) M2/M1 (change) Loans (b) Economic indicaors Oupu gap Real GDP (change) Employmen (change) Employmen gap Unemploymen rae (change) Unemploymen rae gap Compensaion per employee (change) Wages (change) ULC (change) Indusrial producion (change) Nom. eff. exchange rae (change) Commodiy prices (change) Oil prices (change) Cons. price expecaions (level) Cons. price expecaions (change) Prod. price expecaions (level) Prod. price expecaions (change) (c) Fiscal indicaors Direc axes (change) Indirec ax (change) VAT (change) Social conribuions (change) Tax burden (change) Governmen consumpion (change) Governmen compensaion (change) Trivariae models (a) Economic indicaors Oupu gap Real GDP (change) Employmen (change) Employmen gap Unemploymen rae (change) Unemploymen rae gap Compensaion per employee (change) Wages (change) ULC (change) Indusrial producion (change) Nom. eff. exchange rae (change) Commodiy prices (change) Oil prices (change) Cons. price expecaions (level) Cons. price expecaions (change) Prod. price expecaions (level) Prod. price expecaions (change) (b) Fiscal indicaors Direc axes (change) Indirec ax (change) VAT (change) Social conribuions (change) Tax burden (change) Governmen consumpion (change) Governmen compensaion (change) Forecas combinaions (a) Bivariae models Moneary indicaors Economic indicaors Fiscal indicaors All indicaors (a) Trivariae models Economic and fiscal indicaors Noe: The firs line of he able displays he roo mean squared error (RMSE) for he random walk model. The remaining lines repor he RMSE of he respecive alernaive forecasing model divided by he RMSE of he random walk model. Figures in bold indicae ha he respecive forecasing model significanly ouperforms he random walk model (based on Diebold-Mariano es, a 10% significance level)