Forecast Evaluation of Economic Sentiment Indicator for the Korean Economy*

Transcription

1 Forecas Evaluaion of Economic Senimen Indicaor for he Korean Economy* Hyejung Moon and Jungick Lee 2 Absrac The economic senimen indicaor (ESI) for he Korean economy is recenly developed by combining he BSI and CSI componens ha have a high correlaion wih GDP growh and a leading feaure wih respec o GDP cycle. We evaluae he forecasing performance of he Korean ESI wih respec o GDP growh and cycle. For his purpose we use Granger causaliy ess o show ha he consruced ESI conains useful informaion in predicing GDP growh. Using a probi model, we also show ha he ESI is helpful in monioring and predicing he urning poins of GDP. Keywords: ESI, GDP, cycle, Granger causaliy, probi. Inroducion Survey daa measuring economic agens senimen provide useful informaion o assess he curren sae of he economy and forecas shor-erm economic developmen. Besides he informaion iself ha business and consumer surveys provide, he survey daa have many advanages. They have an informaional lead in ha he daa are available ahead of hard economic daa like GDP and indusrial producion ha are usually published wih delays of or 2 monhs. In addiion, he survey daa are generally available a monhly frequencies and hence suiable for reflecing volaile economic developmens. Therefore, he survey daa such as business survey index (BSI) and consumer survey index (CSI) are widely used as a key * Preliminary. This paper is prepared for he presenaion a he 6 h Irving Fisher Commiee Conference, BIS, Basel, Swizerland in Augus 202. Economic Saisics Deparmen, The Bank of Korea, 39 Namdaemun-ro, Jung-gu, Seoul , Korea, E- mail: hjmoon@bok.or.kr 2 Economic Research Insiue, The Bank of Korea, 39 Namdaemun-ro, Jung-gu, Seoul , Korea, jilee@bok.or.kr

2 complemen o quaniaive saisics. The BSI and CSI daa consis of muliple componen series ha concern diverse faces of economic aciviy in differen secors of he economy. The demand o incorporae mos of he informaion conained in muliple indicaors ino a single indicaor has led o he consrucion of a composie indicaor. The single composie indicaor is useful o reflec economic agens overall percepion of economic aciviy. The European Commission (EC) has calculaed an economic senimen indicaor (ESI) since 985 a he EU and he euro-area level as well as a he individual EU member sae level; see European Commission (2006) for a deailed descripion of he EC s ESI. Previous research on he consrucion and evaluaion of composie confidence indicaors include Sock and Wason (2002), Bruno and Malgarini (2002), Gayer (2005), Gayer and Gene (2006), Gelper and Croux (2007). Moon (20) develops he ESI for he Korean economy ha has officially published since June 202. We exend Moon (20) by adding he forecas evaluaion of he ESI for he Korean economy. To evaluae he predicive conen of he ESI wih respec o GDP, Grangercausaliy ess (Granger, 969) and a probi model are used. Relaed lieraures ha examine he forecas performance of he leading indicaors in idenifying urning poins include Esrella and Mishkin (998), Krysalogianni e al. (2004), Croce and Haurin (2009), and Coşar (202) among ohers. The res of he paper is organized as follows. Secion 2 describes he consrucion of he ESI for he Korean economy. In Secion 3, we evaluae he forecas performance of he ESI wih respec o GDP growh and cycle. Secion 4 concludes wih some remarks. 2. Consrucion of he Korean ESI 2. Daa To consruc he Korean ESI, we use he monhly BSI and CSI daa from 2003 o 20 published by he Bank of Korea (BOK). While he monhly series of he BSI daa are available from January 2003, hose of he CSI daa are only available from July 2008 as he BOK had conduced he survey of consumers on a quarerly basis before hen. The BOK has conduced he survey of consumers on a monhly basis since July 2008 when Saisics Korea 2

3 (SK), a naional saisics office ha had conduced a separae survey of consumers, ransferred is monhly compilaion of he CSI o he BOK. Accordingly, we esimae he monhly series of he CSI daa from January 2003 o June 2008 using emporal disaggregaion, a process of deriving high frequency daa from low frequency daa. For he CSI componens ha exis boh in he BOK and SK surveys, he monhly daa are esimaed so ha he disaggregaed series from he quarerly daa of he BOK keep rack of he movemens of he SK s monhly daa as a reference indicaor. For he CSI componens wihou a corresponding reference indicaor in he SK survey, he monhly daa are produced based on a smoohing mehod using ECOTRIM, he sofware released by Eurosa. The BSI daa consis of 30 componen series in he manufacuring secor and 0 componen series in he non-manufacuring secor. In each secor, half of he oal componen series provides he judgmen for he curren monh, while anoher half represens he oulook for he nex monh. For he CSI daa, 9 componen series are considered, seing aside several componen series ha have recenly been added. Hence here exiss a oal of 49 componen series from he business and consumer survey daa. Le us use hese componen series and heir variable names inerchangeably for convenience. The ype of hese 49 variables is he index moving around 00, ranging from 0 o 200. Each variable has 08 observaions from January 2003 o December 20. GDP is used as a reference variable o represen he enire economy. In paricular, he growh rae and cycle of GDP are used o consider shor-erm and long-erm characerisics of he economy. However, since he monhly GDP daa are no published, hey are esimaed using he emporal disaggregaion by sae space mehod. Then he GDP growh rae is measured by he year-on-year percenage change of monhly GDP series. The GDP cycle is exraced using he double Hodrick-Presco (HP) filer. The HP filer is applied wice o achieve a smoohed de-rended cycle; removing a long-erm rend from he seasonally adjused GDP and hen smoohing he de-rended GDP. The cycles of he 49 variables are exraced in he same fashion as he GDP cycle excep for he de-rending procedure because he 49 componen series have no rend. 2.2 Selecion of he ESI componens and weighs 3

4 The ESI needs o be consruced o rack GDP well so ha i can be used as a useful complemen o GDP. Should he ESI and GDP move differenly, i may cause confusion in assessing he curren sae of he economy. So he ESI mus be highly correlaed wih GDP. By he way, if he ESI racks GDP wih a lead of a few monhs, hen he ESI will also be useful for predicing fuure GDP developmens. Inherenly he survey daa relaed o respondens' expecaions have he poenial o have a leading propery. This is because enerprisers and consumers end o increase heir producion and consumpion if hey feel posiive abou he curren and fuure economic siuaion. Therefore, he screening procedure is aimed a selecing informaive componens ha are no only closely correlaed wih GDP bu also deec urning poins of economic movemens earlier han GDP. Cross-correlaion analysis and urning poin analysis are used here. Le z be he reference variable and zi be he i h variable o compare wih. Then he cross-correlaion beween he reference variable and he i h variable shifed m monhs is defined as ρ ( m) i Cov( z Var( z, z i+ m ) ) Var( z for i,,49. If m0, hen i is a conemporaneous correlaion beween he reference variable and he i h variable. The maximum cross-correlaion can be obained from differen choices of posiive or negaive ineger values of m. If he maximum is found for negaive m, hen i means ha he i h variable has he larges correlaion wih he reference variable when i is shifed m monhs ahead. Here he sample cross-correlaions beween he GDP growh rae and each of he 49 variables are calculaed. Denoe he sample conemporaneous correlaion by r 0, he maximum sample cross-correlaion by r max and he value of m wih r max by max. A variable having a large r max a he negaive max is considered o have leading behavior. The leading propery is also examined in erms of he cyclical movemen. The BUSY sofware based on he rouine by Bry and Boschan (97) can be used o deec he urning poins. I idenifies he urning poins of he reference variable and hen denoes he leading or lagging monhs of each of he 49 variables by negaive or posiive values a he reference urning poins. However, he urning poins produced by he BUSY sofware are no obvious in some periods, due o a relaively shor lengh of ime series. So only he urning poins which are obviously idenifiable even by he naked eye are considered. Variables wih a negaive sign a hese ime poins are considered o have he leading propery. i ) 4

5 The preliminary screening of he individual variables is carried ou in each of hree secors: manufacuring, non-manufacuring, and consumer. In each secor, he variables having high levels of cross-correlaion and leading characerisics are pre-seleced for furher invesigaion. Under hese crieria, he 9 variables in he manufacuring secor, 4 variables in he non-manufacuring secor and 5 variables in he consumer secor are pre-seleced respecively. In each secor, all possible combinaions of he pre-seleced variables are examined. There are 2 k -k- possible combinaions when here exis k variables wihin a secor. In each combinaion, he variables are aggregaed by a simple average of he sandardized series, no he original series. This prior sandardizaion is necessary o avoid he dominan effecs of highly volaile variables on he composie indicaor. The racking performance of he aggregaed series in relaion o GDP is esed based on he cross-correlaion and urning poin analyses. The previous wo crieria used in he preliminary screening are reapplied. Among all combinaions in each secor, a hree-variable se (oulook for expors, capaciy uilizaion and financial siuaion), a wo-variable se (oulook for business condiions and financial siuaion) and anoher wo-variable se (oulook for household income and spending decisions) are seleced as he bes combinaion respecively. Therefore, 7 variables among a oal of 49 variables are finally seleced o consruc he ESI. The crosscorrelaion analysis and he urning poin analysis o hese 7 variables are given in Table. Overall, he seleced variables show he leading propery, which is consisen in he fac ha hese variables reflec anicipaions. BSI CSI Manufacuring Nonmanufacuring Table. ESI Componens Componens Cross-correlaion Turning poin r 0 r max max Mean Median Oulook for Expors Oulook for Capaciy uilizaion Oulook for Financial siuaion Oulook for Business condiions Oulook for Financial siuaion Oulook for Household income Oulook for Spending decisions

6 To deermine he weighs of he seleced variables, principal componen analysis is used. The firs principal componen explains abou 82% of he oal variance of he 7 variables. This means ha he firs principal componen can replace he 7 variables wihou much loss of informaion. The coefficien of he firs principal componen measures he imporance of he each variable o he firs principal componen, irrespecive of he oher variables. In paricular, he relaive sizes of imporance are deermined based on he squared coefficiens which sum o. Based on he sum of he squared coefficiens wihin he manufacuring and non-manufacuring secors and he sum of hose in he consumer secor, he weighs of BSI and CSI pars are deermined by 0.75 and Wihin he BSI par, he weighs of he manufacuring and non-manufacuring secors are deermined based on he conribuions o GDP growh. The conribuion of he nonmanufacuring secor o GDP growh is compued by excluding he indusries for which he business survey is no conduced (agriculure, financial inermediaion, public adminisraion and defense, compulsory social securiy, educaion, healh and social work, and oher service aciviies). The average of he conribuions over 2003 o 20 is.62%p for he manufacuring secor and 0.98%p for he non-manufacuring secor, so he raio of heir relaive magniudes is almos 0.6 and 0.4. This raio is sable for oher ime periods. Thus he weighs wihin he manufacuring and non-manufacuring secors are deermined as 0.6 and 0.4. Since he BSI par has a weigh of 0.75 in oal, he weighs of he manufacuring and non-manufacuring secors are finally allocaed o 0.45 and To sum up, he weighs of he manufacuring, non-manufacuring and consumer secors are se by 0.45, 0.30 and Wihin each secor, he individual variables have equal weighs as shown in Table 2. Table 2. Weighs allocaed o he ESI Componens Componens Weighs Oulook for Expors 0.50 Manufacuring Oulook for Capaciy uilizaion BSI Oulook for Financial siuaion 0.50 Non-manufacuring Oulook for Business condiions Oulook for Financial siuaion CSI Oulook for Household income Oulook for Spending decisions

7 2.3 Calculaion of he ESI Afer deermining he 7 informaive variables (or componens series) and he corresponding weighs, he exac calculaion of he ESI is made as follows. Sep : Sandardize he original componen series where S T Y i, X i, - X X i, is he i h componen series observed a ime, - T ( X i X 2, - i ) for i,,7. Sep 2: Aggregae he 7 sandardized series using he weighs S i X T T i X i, and Z 7 i w Y i i, where w i is a weigh of he i h componen such ha w. Sep 3: Scale Z o have a mean of 00 and a sandard deviaion of 0 7 å i i ESI æ Z ö - Z ç è S z ø where T Z and SZ T å ( Z T - T Z - Z ) 2. The ESI value of 00 marks a long-erm average over he ime period from,,t. Values greaer han 00 indicae an above-average posiion, while values below 00 indicae a below-average posiion. The fixed sandard deviaion of 0 implies ha abou 68% of he ESI values fall wihin a range beween 90 and 0 assuming approximae normaliy. Unlike he ESI, a value of 00 in he BSI and CSI daa means he equal proporion of negaive and posiive opinions. In addiion, he BSI and CSI daa have ofen fallen below 00 due o he cauiousness of respondens, even when he economy is booming. The ESI solves his problem by rescaling in Sep 3. Moreover, he ESI is easy o inerpre because he longerm average of 00 plays a yardsick role for making judgmens. 7

8 Noe ha he sandardizaion in Seps and 3 is carried ou over he period from,...,t. The end poin T is exended every year o include up-o-dae informaion, bu does no change wihin a single year. For example, he ESI values from January o December in 202 are calculaed based on he sandardizaion period from January 2003 o December 20. Bu he ESI values in 203 are compued using a new sandardizaion period exended o December 202, and he ESI values before 203 are all revised a once a he beginning of 203. Tha is, he revision of he ESI daa is underaken every year. This revision may confuse users, bu i is ineviable in order o reflec he recen economic siuaion adequaely. The cyclical componen of he ESI is compiled o rack he cyclical paerns of economic senimen, and is calculaed by removing seasonal and irregular componens from he ESI. 3. Forecas Evaluaion 3. Tracking performance of he ESI Following he mehod described in Secion 2, he ESI for he Korean economy are compued for he period of January 2003 o May 202. Figure shows ha he ESI and GDP growh move closely ogeher. Noe ha in Table 3 he ESI has a maximum cross-correlaion of when i is one monh ahead of GDP. Figure. ESI and GDP growh 8

9 Table 3. Cross-correlaion of ESI and GDP growh Leading (-) or Lagging (+) Monhs The movemens of he cyclical componens of ESI and GDP are shown in Figure 2. Over he period of 9 years, he GDP cycle records wo peaks in February 2008 and June 200, and wo roughs in February 2005 and The cyclical componens of he ESI deecs urning poins 4 monh, 6 monhs, monh, and 2 monhs ahead of he corresponding reference dae, respecively, or abou 3.25 monhs early on average. The leading feaure of he cyclical movemen is also found in he cross-correlaion analysis of he cycles. The maximum cross-correlaion is when he cyclical componen of ESI is 3 monhs ahead of GDP as shown in Table 4. Overall, he ESI racks GDP well, being well correlaed and co-moving wih GDP wih leads of a few monhs. Figure 2. Cyclical componens of ESI and GDP Noe: The deceleraing phase saring in June 200 is subjec o change as more daa are available. Table 4. Cross-correlaion of he cycles of he ESI and GDP Leading (-) or Lagging (+) Monhs

10 3.2 Granger causaliy ess In his subsecion, he leading behavior of he ESI wih respec o GDP growh is furher examined using Granger-causaliy es. To see wheher movemens in he ESI precede movemens in he GDP growh or vice versa, Granger-causaliy ess are carried ou based on he lag lenghs of 4, 5, and 6. As can be seen in Table 5, he null hypohesis ha ESI does no Granger-cause GDP growh is rejeced, while he null hypohesis ha GDP growh does no Granger-cause ESI is no rejeced, a a significance level of 5% for all choices of lag lenghs, indicaing he ESI precedes GDP growh. Tha is, he ESI shows significan posiive conribuion o explain fuure GDP growh, implying ha he ESI is helpful in forecasing GDP growh. Table 5. Granger causaliy ess Lag H 0 : GDP growh does no Granger-cause ESI H 0 : ESI does no Granger-cause GDP growh F-Saisic Prob. F-Saisic Prob. Resuls ESI GDP ESI GDP ESI GDP 3.3 Forecas using probi model We use a probi model o furher examine he leading propery of he cyclical componen of ESI wih respec o ha of GDP and hen evaluae he forecas performance of he ESI in idenifying he urning poins of GDP. Suppose ha a binary dependen variable, Y, akes on only values of one and zero as follows. Y, if he economy is in deceleraion period 0, oherwise Then he esimaed probabiliy of being in he deceleraion period is of he form P Y x, β) F( β + β x + L β ( 0 k xk ) 0

11 where x,, x k are k explanaory variables, β,, βk are he corresponding regression coefficiens and F is he cdf of a sandard normal disribuion, i.e., F( a) Φ( a) 2π a / 2 - e 2 - d Here he deceleraion period is deermined based on he peak and rough poins of he cyclical componen of GDP. The conemporaneous and lagged values of he cyclical componen of ESI are considered as explanaory variables. The number of lags is deermined so as o minimize Akaike Informaion Crierion (AIC) and Schwarz s Bayesian Crierion (SBC). The cyclical componen of ESI shifed monh ahead (ESIC(-)) is included in our probi model as an explanaory variable along wih is conemporaneous value (ESIC). Table 6 shows ha he esimaed model has 53% of he explanaory power and all explanaory variables are saisically significan. Noe ha he sign of esimaed coefficien of ESIC(-) is posiive, implying ha ESIC(-) and he conemporaneous GDP cycle end o move in he same direcion. Table 6. Probi model esimaion Variable Coefficien Sd. Error z-saisic Prob. C ESIC ESIC(-) McFadden R From he esimaed probi model, he probabiliy of being in he deceleraion period can be compued for each observaion of he daa, which is called in-sample forecas. The probabiliies from he in-sample forecas are ploed ogeher wih he GDP cycle in Figure 3. The esimaed probabiliies are shown o be high in he shaded areas of he deceleraion. In paricular, he esimaed probabiliies are close o during he financial crisis from February 2008 o February Overall, our esimaed probi model seems o successfully idenify he deceleraion phase.

12 Figure 3. Esimaed probabiliies and he GDP cycle Noe: The deceleraing phase saring in June 200 is subjec o change as more daa are available. Recursive ou-of-sample forecass are made over he period from January 200 o May 202. In paricular, he probabiliies obained from he -sep and 3-sep ahead ou-ofsample forecass are presened in Figure 4. The predicive power of hese forecass is evaluaed by comparing a percenage of correc classificaion based on he cuoff value we specified. There are wo kinds of correc classificaions. One is ha he prediced probabiliy is greaer han he cuoff and he observed Y, and anoher is ha he prediced probabiliy is less han or equal o he cuoff and he observed Y 0. The fracion of Y observaions ha are correcly prediced is called sensiiviy, while he fracion of Y 0 observaions ha are correcly prediced is called specificiy. In his problem, he sensiiviy is compued for he deceleraion phase and he specificiy for he acceleraion phase. Moreover, a percenage of correc classificaion among oal observaions is compued. Table 7 presens he forecas powers for he -sep and 3-sep ahead ou-of-sample forecass based on hree cuoff values of 0.4, 0.5, and 0.6. All forecass correcly idenify he phase in he acceleraion period from January 200 o June 200 for all hree cuoff values. Bu he percenage of correc classificaion in he deceleraion period depends on he choice of he cuoff. I is highes when he cuoff value is 0.4 since he smaller cuoff value is easier o declare he deceleraion. Obviously he oal forecas power ends o decrease as he cuoff value ges bigger. Comparing he ou-of-sample forecas mehod, he -sep ahead ou-of- 2

13 sample forecas has a higher forecas power han he 3-sep ahead ou-of-sample when he cuoff value is 0.6, bu almos he same for he oher cuoff values. Figure 4. Probabiliies from ou-of-sample forecass Noe: The deceleraing phase saring in June 200 is subjec o change as more daa are available. Table 7. Ou-of-sample forecas evaluaion: probi model Cuoff value Phase Percenage of correc classificaion -sep ahead 3-sep ahead 0.4 Deceleraion Acceleraion Toal Deceleraion Acceleraion Toal Deceleraion Acceleraion Toal

14 4. Conclusions We consruc he ESI for he Korean economy in a similar fashion o he European Commission s; ha is, we aggregae he sandardized BSI and CSI componen series by a weighed average and hen rescale i o have a mean of 00 and a sandard deviaion of 0. However, we focus on selecing informaive componens of he BSI and CSI daa and deermining he weighs so ha he composie indicaor has a high correlaion wih GDP growh and a leading feaure wih respec o GDP cycle. We evaluae he forecasing performance of he Korean ESI wih respec o GDP growh and cycle. The ESI urns ou o have a good racking performance as a leading indicaor of GDP. Using he Granger causaliy ess we show ha he consruced ESI precedes GDP growh, implying he former conains useful informaion in predicing GDP growh. Also, using a probi model, we show ha he ESI is helpful in monioring and predicing he urning poins of GDP. The performance of our probi model could be furher improved by adding more relevan economic variables as explanaory variables o he model. Overall, he recenly developed ESI for he Korean economy is useful in forecasing shorerm economic developmens as well as in reflecing economic agens overall percepion of economic aciviy or condiions. References Bruno, G. and Malgarini, M. (2002), An indicaor of economic senimen for he Ialian economy, Working Paper No. 28/02, Insiue for Sudies and Economic Analysis, Rome. Coşar, E. (202), Analysis of cyclical behaviour of invesmen expendiures for he Turkish economy, Applied Economics Leers, 9, Croce, R. and Haurin, D. (2009), Predicion urning poins in he housing marke, Journal of Housing Economics, 8, Esrella, A. and Mishkin, F. (998), Predicing U.S. Recessions: Financial Variables as Leading Indicaors, Review of Economics and Saisics, 80, European Commission (2006), The join harmonized EU programme of business and consumer surveys, Special Repor No 5. 4

15 Gayer, C. (2005), Forecas Evaluaion of European Commission Survey Indicaors, Journal of Business Cycle Measuremen and Analysis, 2, Gayer, C. and Gene, J. (2006), Using facor models o consruc composie indicaors from BCS daa A comparison wih European Commission confidence indicaors, European Economy, Economic Papers 240, European Commission, Brussels. Gelper, S. and Croux, C. (2007), The predicive power of he European economic senimen indicaor, K.U.Leuven, FETEW Research Repor KBI 0707, -5. Granger, C. (969), Invesigaing Causal Relaions by Economeric Models and Cross- Specral Mehods, Economerica, 37, Krysalogianni, A., Maysiak, G. and Tsolacos, S. (2004), Forecasing UK commercial real esae cycle phases wih leading indicaors: a probi approach, Applied Economics, 36, Moon, H. (20), Consrucion of an Economic Senimen Indicaor for he Korean Economy, The Korean Journal of Applied Saisics, 24, Sock, J. and Wason, M. (2002), Forecasing using principal componens from a large number of predicors, The Journal of he American Saisical Associaion, 97,