CHAPTER 5. results generated from the selected methodology in the previous chapter. The chapter is

Transcription

1 CHAPTER FINDINGS This chaper presens he findings and furher discusses he analysis of he resuls generaed from he seleced mehodology in he previous chaper. The chaper is segregaed ino hree secions consis he preliminary analysis, model selecion and finally, he srucural break effec in he consisency of hedging performance measuremen. In he firs sub secion, he research discusses he saisical characerisic of boh CPO and FCPO series. Addiionally, in he second secion, he chaper proceeds o seek he bes dynamic model ha will give he bes hedging performance resul. The analysis compares nine dynamic models encompass he GARCH framework wihin he mean-variance and minimum variance measuremen. Using he dynamic model seleced his second sub secion, he nex secion will inroduce he srucural break effec which possibiliy could affec he hedging performance resuls. Finally, he hird secion will hen relaes he srucural break effec wih he consisency of hedging performance resuls wihin he whole research period. 07

2 5. PRELIMINARY ANALYSIS SECTION I 5.. Descripive Saisic Resuls Table 5.: Saisical Properies for CPO and FCPO reurns Mean Median Max Min Sd. Dev. Skewness Kurosis CPO Jarque-Bera Q(9) Q(5) Q 2 (9) Q 2 (5) (0.0000) (0.02) (0.0000) (0.0000) (0.0000) Mean Median Max Min Sd. Dev. Skewness Kurosis FCPO Jarque-Bera Q(9) Q(5) Q 2 (9) Q 2 (5) (0.0000) (0.00) (0.0000) (0.0000) (0.0000) P-value are provided in parenheses for Jarque-Bera, Q(9), Q(5), Q 2 (9) and Q 2 (5). Table 5. summarises he saisical properies of CPO and FCPO reurns for he period of January 996 o Augus The CPO has a wider range of reurns ranging beween -4.07% and 5.29% compared wih he FCPO, which ranges beween - 0.0% and 3.84%. The FCPO reurns, on he oher hand, exhibi a slighly higher sandard deviaion han he CPO reurns. Boh reurns show a non-symmeric disribuion, wih he FCPO (CPO) reurns disribuion being posiively (negaively) 08

3 skewed. Boh reurns series exhibi an excess kurosis. The non-normaliy feaure in boh reurns series is found in he Jarque-Bera es resuls. Such a feaure is consisen wih many oher financial and commodiy reurns (see Baillie and Myers, 99; Kroner and Sulan, 993 and Ford, Pok and Poshakwale, 2005). The Q saisic for he residuals and squared residuals were done a lagged 9 and lagged 5. Similar resuls were repored in he CPO and FCPO reurns (Liew and Brooks, 998 and Azizan e al., 2007) wih boh being lagged, inferring he presence of serial correlaion, and an auoregressive and heeroscedasiciy problem. These resuls also ally wih he evidence documened in oher markes including Mili and Abid (2004) in he Canadian Bankers Accepance reurns, Yang and Allen (2004) in he Ausralian Sock Index reurns and Ford, Pok and Poshakwale (2005) in he Malaysian Sock Index reurns. In conras, Bailie and Myers (99) failed o find he presence of any serial correlaion problem in Beef, Coffee, Corn, Coon, Gold and Soybeans commodiy prices. In sum, he non-normaliy feaures and he presence of boh serial correlaion and ARCH effec in boh series unanimously prove he imporance of considering he surrounding informaion in modelling boh reurns Uni Roo Tes Resuls The usage of he uni roo es confirming he evidence of saionariy in imes series is said o be less adequae. Consequenly, he KPSS saionary es is ailored o complemen he oher uni roo saisical es (DF es or PP es) in deermining he 09

4 saionariy of he series. Thus, he research adoped boh he ADF es and he PP es o es boh CPO and FCPO series uni roo, while he KPSS es is o idenify he series saionariy. Tesing boh uni roo and saionariy hypohesis may give a srong conformaion wheher he series are saionary or inegraed (Kwiakowski e al, 992). Many recommend including boh ess o srenghen he conclusion of he presence of saionariy in he esed ime series. Table 5.2: Uni Roo Tess Resuls ADF Tes PP Tes KPSS Tes Level s Diff Level s Diff Level s Diff CPO (0.7498) (0.000) (0.6249) (0.000) (0.0000) (0.4809) FCPO (0.7920) (0.000) (0.6503) (0.000) (0.0000) (0.4872) P-values are provided in parenheses for hese hree ess. Table 5.2 illusraes he resuls for he uni roo es employed for CPO and FCPO. Previous researchers employed he ADF and PP ess for boh spo and fuures reurns, and hey suppored he non-saionary characerisic in he esed series (Kroner & Sulan, 993; Tong, 996; Bera, Gracia & Poh, 997; Liew and Brooks, 998; Brails e al., 2002; Brooks, Hendry & Persand, 2002; Chen e al, 2002; Tunara & Tan, 2002; Floros and Vougas, 2004, Ford, Pok and Poshakwale, 2005; Norden, 2006; and Azizan e al., 2007). The ADF and PP es resuls indicae ha hese wo series virually unanimously fail o rejec he null hypohesis where he series have a uni roo a level [inegraed a 0 or I(0)]. Such resuls validaed a non-saionary characerisic in hese variables a any 0%, 5% and % significan level. Conrary resuls were porrayed 0

5 when he series was aken a is firs differen or inegraed a [I()]. In addiion, he KPSS es repors a rejecion of is null hypohesis where he series is saionary. The rejecions drive he same conclusion for boh he ADF and PP ess. On he oher hand, when examining he series a is firs difference, we end o accep he KPSS null hypohesis. Overall, hese hree ess infer a srong non-saionariy of series a is level bu which urned o be saionary a is firs differen. As he uni roo ess suppor ha he series are saionary a is firs differen (I()), we can safely conjecure he possibiliy of boh series being coinegraed in he long run. As such, we es he coinegraion relaionship wihin boh series via he Johansen Coinegraion Tes and he resuls are repored in Table 5.3. The resuls obviously exhibi a long run relaionship beween boh series a he 5% level of significance. Generally, hese uni roo es resuls indicae he need of series ransformaion for he hedging effeciveness modelling process. Consequenly, hese CPO and FCPO selemen prices are ransformed ino reurns and hese compued reurns are saionary a is level. Table 5.3: Johansen Coinegraion Tes Resuls Unresriced Coinegraion Rank Tes (Trace) Hypohesized Trace 0.05 No. of CE(s) Eigenvalue Saisic Criical Value Prob.** None * A mos Trace es indicaes coinegraing eqn(s) a he 0.05 level

6 * denoes rejecion of he hypohesis a he 0.05 level **MacKinnon-Haug-Michelis (999) p-values Unresriced Coinegraion Rank Tes (Maximum Eigenvalue) Hypohesized Max-Eigen 0.05 No. of CE(s) Eigenvalue Saisic Criical Value Prob.** None * A mos Max-eigenvalue es indicaes coinegraing eqn(s) a he 0.05 level * denoes rejecion of he hypohesis a he 0.05 level **MacKinnon-Haug-Michelis (999) p-values 2

7 SECTION II 5.2 DIVERSE MEAN AND VARIANCE SPECIFICATIONS versus HEDGING PERFORMANCES Conclusive evidence indicaes ha he hedging decision is a dynamic process raher han a saic process. The hedging conrac effeciveness, which ofen refers o he degree of variance (risk proxy), can be furher reduced when marke paricipans include hedging in heir invesmen decision. Empirical evidence ess a diverse range of dynamic models ha aim o esimae he variance and covariance srucure and furher deermine which model ends o produce he larges risk reducion. However, he evidence is mainly from he examinaion of developed fuures markes. As par of risk reducion he invesor uiliy maximizaion is acually equally imporan as i covers he risk and reurn, however, only a few researchers invesigae boh. Since Lien (2004) finds ha he omission of long run equilibrium in mean specificaion may give a downward bias hedging raio, we believe ha a differen mean specificaion will generae various hedging performance resuls. To address his issue, his research moves he aenion of variance-covariance srucure specificaion o vary mean reurns modelling specificaion and invesigaes is implicaion on hedging performance. The research adoped he inercep, vecor auoregressive and vecor error correcion erm mean specificaion for he BEKK, Consan Correlaion and Dynamic Condiional Correlaion models. These nine models were used o esimae he variance and covariance in he Malaysian Palm Oil Commodiy markes. Subsequenly, he 3

8 research analysed he hedging performance using boh he risk reducion and uiliy maximizaion funcion. The performances are esed wihin in-sample and ou-sample muliple forecasing periods (, 5, 0, 5 and 20 days) Maximum Likelihood Esimaion Resuls Table 5.4: Maximum Likelihood esimaion resuls for he BEKK, CCC and DCC models. BEKK CCC DCC Inercep VAR VECM Inercep VAR VECM Inercep VAR VECM Mean Specificaion α 0s α 0f α s *** *** *** -0.56*** -0.80*** -0.94*** α s *** -0.28*** *** *** *** *** α s *** * *** * *** * α s ** * α s ** ** * α s α f *** *** *** 0.302*** 0.38*** 0.385*** α f-2 0.9*** 0.592*** 0.205*** 0.422*** *** 0.523*** α f *** *** 0.065*** 0.048** *** ** α f * ** ** α f * *** *** 0.00 α f * *** *** e s *** *** *** α 2f *** ** *** ** *** * α 2f *** *** *** α 2f *** *** ** *** *** α 2f *** *** *** -0.46*** *** α 2f *** *** * *** α 2f *** *** *** α 2s *** 0.385*** *** 0.572*** *** 0.503*** α 2s *** ** 0.69*** *** 0.79*** 0.095*** α 2s * 0.067** 0.454*** *** 0.478*** *** 4

9 α 2s *** 0.054** 0.395*** *** 0.465*** *** α 2s *** *** *** α 2s *** 0.074*** *** e f 7.254*** *** *** BEKK CCC DCC Inercep VAR VECM Inercep VAR VECM Inercep VAR VECM Variance-Covariance Specificaion C s *** 0.943*** *** 0.065*** *** *** *** *** 0.036*** C sf -0.46*** 0.392*** *** 0.534*** 0.597*** *** C f E E E *** 0.048*** *** *** *** *** A s -0.3*** ** *** A sf A fs *** ** A f *** * G s *** -0.39*** G sf *** *** ** G fs *** *** *** G f -.543*** *** ** A s *** *** *** *** *** *** A f 0.933*** 0.987*** 0.953*** 0.925*** 0.907*** *** B s 0.299*** *** *** 0.353*** 0.*** 0.075*** B f *** *** *** *** 0.076*** *** a 0.077*** 0.52*** 0.096*** b *** *** *** *** represens % level of significance ** represens 5 % level of significance * represens 0 % level of significance This able repors join maximum likelihood esimaes of he condiional means and he covariance marix of he reurns of CPO and FCPO for he following specificaion: 5

10 6 Mean Specificaion Consan s s s r ε α + = ; ; Ω s ε ~N(0,H ) f f f r ε α + = ;; Ω f ε ~N(0,H ) VAR = = = k i s i f f i s k i s s s r r r,, ε α α α = = = k i f i s s i f k i f f f r r r, 2, 2 ε α α α VECM = = = k i s s i f f i s k i s s s Z e r r r,, ε α α α = = = k i f f i s s i f k i f f f Z e r r r, 2, 2 ε α α α Variance Specificaion (BEKK) * *' * ' *' * *' k K k k k K k k G H G A A C C H = = + + = ε ε Variance Specificaion (CCC) = = f s f s ff sf sf ss h h h h h h h h H,,,,,,,, ρ ρ

11 7 2,, * * = = + + = s q j s p j s s B h A h ε γ 2, 22, 22 * * = = + + = f q j f p j f f B h A h ε γ f s sf h h h ρ = Variance Specificaion (DCC) = = f s f s ff fs sf ss h h Q Q h h h h h h H,,,,,,,, ,, * * = = + + = s q j s p j s s B h A h ε γ 2, 22, 22 * * = = + + = f q j f p j f f B h A h ε γ ` ( ) + + = = 2 2 f f s f s s f sf sf s B A r r r r B A Q ρ ρ ε ε ε ε ε ε ρ ρ Table 5.4 presens he hree mean specificaions of he BEKK model, Consan Correlaion model and Dynamic Condiional Correlaion. The maximum likelihood resuls are caegorized according o he hree mean models including Inercep, VAR and VECM specificaion, 8 respecively. 8 The number of lags in VAR and VECM models are deermined by lag lengh crieria procedure.

12 5.2.. Mean specificaion Resuls Based on he mean specificaion resuls, he inercep mean model does no provide a good model in posuring boh CPO and FCPO reurns in all hree variancecovariance specificaion models. Addiionally, he VAR models deec significan evidence of boh esed reurns depending on is own and is counerpar lagged erm. The movemen of he CPO reurn ends o be driven by is own lagged erms ( o 3 lagged erms for he BEKK model and CCC model bu o 5 lagged erms for he DCC model) and FCPO lagged erms ( o 4 lagged erm for BEKK model and o 6 lagged erms for he oher wo models). However, he FCPO reurns movemen was deermined by CPO and FCPO reurns lagged erms. 9 The CPO and FCPO reurns are likely o have an inverse movemen wih heir own lagged erms bu no wih heir counerpars. The Johensen coinegraion es resul exhibied he exisence of a coinegraion relaionship beween boh series. Therefore, he VECM mean specificaion was employed and he significance of boh error erms confirmed ha boh series are highly coinegraed in he long-run. The resuls are consisen wih he exising lieraure for when series are saionary a I();he series will possibly coinegrae in he long-erm (Kroner and Sulan, 993; Gagnon and Lypny, 995; Wilkinson, Rose and Young, 999;Floros and Vougas, 2004; and Azizan e al., 2007). In conras, however, Bailie and Myers (99) found no evidence of coinegraion in beef, corn, coon, gold and soybean prices. 9 o 6 lagged erms 8

13 Variance and Covariance Specificaion Resul The Inercep-BEKK has shown he presence of a significan opposie relaionship in he CPO variances wih is previous volailiy shocks (refer o G parameers) and is own shocks (refer o A parameers). Meanwhile, he FCPO variance was influenced by he movemen of is own previous volailiy bu no evidence of is squared residual erms. The Inercep-BEKK and VAR-BEKK model repored greaer inverse coefficiens in boh series volailiy shocks compared o he effec of is squared residuals. This resul infers ha he volailiy of boh series was highly influenced by is previous volailiy movemen raher han is own shocks. Meanwhile, he VECM-BEKK model exhibied conrary resuls for he FCPO volailiy modelling (Inercep and VAR- BEKK). Addiionally, he VECM-BEKK model has proven ha here is no evidence o confirm ha curren CPO volailiy is affeced by is own volailiy shocks. The covariance resuls have esablished a posiive significan movemen wih is previous covariance in he VECM-BEKK model. However, he resuls are mixed in he VAR-BEKK and Inercep-BEKK models, where he covariance is eiher posiively or negaively affeced by is previous series covariance. The covariance esimaion resuls conclude ha he series covariance was highly affeced by is own lagged covariance erm. Subsequenly, he previous erm for boh CPO and FCPO residuals did no asser any influence on he movemen of he covariance. The Consan Correlaion and Dynamic Condiional Correlaion model exhibi srong evidence ha CPO and FCPO volailiy are deermined by heir own volailiy 9

14 shock (refers o A parameers) in all hree mean specificaion models. Much higher coefficiens were found in he FCPO compared o he CPO volailiy clusering model. Such resuls lead one o conclude ha he FCPO reurns end o be influenced by heir own shocks (refer o B parameers) in a higher parallel magniude han he CPO reurn volailiy. Boh series squared residuals demonsraed a posiive relaionship ha may significanly influence he movemen of boh CPO and FCPO volailiy for all hree mean models in he CCC and DCC models. However, here is no evidence o infer he possibiliy ha he FCPO squared residual effecs he volailiy of he curren FCPO reurns in he Inercep-CCC model. The DCC model suppors he evidence ha he correlaion beween CPO and FCPO is affeced by boh residual erms and is previous period correlaion. In summary, boh he VAR and VECM model appear o provide a good represenaion model in posuring he CPO and FCPO reurn. As for BEKK variance clusering models, he CPO variance is likely o be influenced by is own volailiy shocks bu a a higher magniude han is own squared residuals (in he VAR and VECM mean model). Consisen resuls for he BEKK, CCC and DCC models suppor he evidence ha FCPO variance ends o be driven more by is own volailiy shocks han is own squared residual for all hree mean models. However, he negaive coefficien in variance and residuals simply means ha a negaive shock in he FCPO and CPO reurns will increase boh marke reurns volailiy or vice versa. While, a posiive coefficien for FCPO own volailiy shock in he VECM-BEKK model describes a posiive movemen in is own pas variance, which will lead o an increase 20

15 of FCPO marke reurns volailiy. Similar resuls were demonsraed in he FCPO variance movemen, which were repored in all hree mean models for he CCC and DCC models. However, he resuls failed o prove he exisence of a relaionship beween he FCPO squared residuals wih he FCPO variance in he Inercep-BEKK and Inercep-CCC models. Neverheless, he covariance oulined by he BEKK model was more influenced by is own covariance pas movemen han beween he residuals for CPO and FCPO (insignifican H sf in hree mean models). Based on he above coefficien esimaion resuls, we can draw a conclusion here is a srong persisency in volailiy in he Malaysian CPO and FCPO markes. The residual and squared residual diagnosic resuls for all hree mean models for BEKK, CCC and DCC models are shown in Appendixes C, D and E. The Q saisic for boh residuals and squared residuals are presened using he, 4 and 0 order serial correlaions. There was evidence of serial correlaion in all Inercep mean models for he BEKK, CCC and DCC models ha presence in he CPO and FCPO residual series. However, he same mean models are able o accoun he minor presence of he auoregressive and heeroscedasiciy problem in he CPO and FCPO sandardized residuals. The Inercep-CCC model was able o ackle he ARCH problem where he resuls infer no evidence of ARCH effec for eiher he CPO or FCPO squared residuals. Furhermore, he Inercep-BEKK and Inercep-DCC were able o suppor he absence of ARCH effec in FCPO squared residuals in lagged and lagged 0 for Inercep- DCC only. 2

16 All hree VAR mean models gave a beer performance in addressing he serial correlaion presence in he CPO and FCPO residuals. Only a minor serial correlaion was deeced in he VAR-DCC FCPO residual a lagged 4. Furher, he findings exhibied an average ARCH effec in all hree VAR models. The resuls suppor he presence of he ARCH effec in VAR-BEKKCPO residuals a a higher lagged order. However, an idenical conclusion was found in lagged 4 and 0 for he VAR-CCC squared residuals series. Only a minor ARCH problem was found in he lagged 0 VAR-DCC FCPO squared residual. Overall, he VAR-BEKK ends o be he mos successful model o overcome boh he serial correlaion and ARCH effec ha is presen in he FCPO residual series. Meanwhile, he VAR-DCC model appears o be he bes as i fully addressed boh problems ha presence in he CPO residual series. The VECM mean model for BEKK, CCC and DCC models parly overcomes he serial correlaion and ARCH problem in residuals and squared residuals in he esed series. There is no evidence of serial correlaion for VECM-BEKKCPO and FCPO series a lagged 4 and lagged for he oher wo variance models. The VECM-BEKK ouperform oher models as no ARCH evidence was found in he CPO squared residuals. The resuls furher sugges ha he ARCH effec sill exiss in he FCPO squared residual a lagged 0 for all VECM mean models. However, he CPO squared residuals a order porrayed he non-absence of ARCH problems in he CCC and DCC models. In ARCH effec resuls, he CCC and DCC gave consisen resuls for he VECM mean models. 22

17 In conclusion, he VAR-BEKK and VAR-DCC appear o provide he bes model for CPO and FCPO, respecively, for solving he serial correlaion and ARCH issues ha are presen in boh residual series. Yang and Allen (2004) have repored ha he VECM-DVECH model failed o encouner he ARCH effec bu usefully solved he serial correlaion in he residual. They recommended a more dynamic GARCH model o couner hese ARCH effecs in he squared residual. Based on our resuls, all hree VECM models are considered o be equal second bes as hey are able o ackle boh residual serial correlaion and ARCH effec albei marginally. Finally, compared o he oher models, he Inercep model appears o be he wors Hedging Performance Minimum variance resuls (Risk reducion) Table 5.5: Hedging Raio esimaion resuls wihin minimum variance framework No Of Days Forecas BEKK CCC DCC Inercep() VAR(2) VECM(3) Inercep(4) VAR(5) VECM(6) Inercep(7) VAR(8) VECM(9) In-sample Day Days Days Days Days No Of Days Forecas Ou-sample Day Days Days Days Days Noe: The hedging raio is calculaed based on h Ω - = cov sf Ω - /σ f 2 Ω -. 23

18 Table 5.5 repors he hedging performances hrough he percenage of risk reducion achieved by all hree mean models for he BEKK, CCC and DCC models. The ables are segregaed according o ou-sample and in-sample daa for each model. The resuls include each, 5, 0, 5 and 20 days forecased period ahead, which are caegorized according o he hree mean models. The Inercep model has posulaed a wider range of hedging raio wihin 0.2 o.43 (refer o Column ) compared o he ou of sample raio, which is from 0.38 o 0.6 (refer o Column ). However, a sable esimaion was posured by boh VECM-BEKK and VAR-BEKK models wihin he ou-sample daa, beween 0.48 o 0.6 (refer o Column 2 and 3). In addiion, he Inercep-BEKK demonsraes ha a higher ime horizon will lead o a higher hedging raio esimaion. Subsequenly, a conrasing finding was repored in he VAR-BEKK models, which porrayed an inverse relaionship beween he hedging raio and forecasing period ahead. However, wihin he in-sample period, VAR and VECM- BEKK exhibi a similar finding generaed in he Inercep-BEKK model. The evidence suppors a posiive movemen beween he hedging raio and he percenage of risk reducion where he lower he raio, he risk reducion ends o be low and vice versa. The hedging raio esimaion from he CCC model exhibied a larger range beween 0.7 o.84 (in Column 4) for he in-sample Inercep-CCC model and 0.6 o 2.08 (in Column 4) for he same ou-sample forecased period. This evidence indicaes ha wihin he in-sample analysis, he CPO marke paricipan ends o hedge from he range of 7% o 84% of is spo posiion and 6% o 208% of is spo posiion wihin he ou-sample resuls. The highes raio was generaed from he 0-day forecasing 24

19 period, which was 2.08 for he ou-sample and.84 for he in-sample Inercep model. However, a consisen range of hedging raio was repored for boh he VAR and VECM-CCC models for boh he in-sample and ou-sample daa (beween 0.46 o 0.67 Column 5 and 6). Similarly, he VAR-BEKK, VAR and VECM-CCC models suppor he evidence ha a lesser ime horizon ends o give a higher hedging raio. However, he risk reducion findings gave a monoonic hedging performance a any forecased period. In he heoreical framework chaper, he risk reducion measuremen similarly refers o he squared correlaion beween he FCPO and CPO reurns. Therefore, he consan risk reducion esimaed in he CCC model is no surprising since he model conjecures a consan correlaion beween boh series. In conras o he CCC model, he DCC model assumes a dynamic process for series correlaion. Wihin he hree variance models, he DCC model esimaed he mos similar hedging raio resuls eiher in he in-sample or ou-sample analysis. The model esimaed ha beween 39% and 60% spo proporion (refer o Column 7, 8 and 9) needs o be hedged and ha hese proporions or raios shared he same cycle hroughou he muliple forecasing horizons. In addiion, he longer he period forecased ahead, he higher he risk reducion can be achieved eiher in he in-sample or ou-sample esimaion resuls. Overall, he above evidence ends o suppor ha he ime facor exiss in hedging raio esimaion and coness he saic hedging raio. 25

20 Table 5.6: Hedging performance resuls wihin minimum variance framework BEKK CCC DCC No Of Days Forecas Inercep VAR VECM Inercep VAR VECM Inercep VAR VECM Insample Day 6.8% 4.70% 4.3% 28.53% 35.66% 36.8% 30.69% 27.63% 27.42% 5 Days 44.36% 7.45% 9.80% 28.53% 35.66% 36.8% 30.60% 26.22% 3.53% 0 Days 8.06% 26.35% 22.92% 28.53% 35.66% 36.8% 3.0% 34.02% 38.02% 5 Days 65.84% 3.80% 30.95% 28.53% 35.66% 36.8% 32.06% 43.9% 44.88% 20 Days 69.8% 43.24% 46.08% 28.53% 35.66% 36.8% 42.05% 54.99% 54.22% No Of Days Forecas Ousample Day 27.30% 40.2% 8.8% 29.37% 36.66% 37.86% 29.69% 26.45% 26.0% 5 Days 32.88% 35.94% 40.57% 29.37% 36.66% 37.86% 30.47% 25.52% 29.7% 0 Days 30.46% 38.93% 35.8% 29.37% 36.66% 37.86% 30.57% 33.3% 36.89% 5 Days 34.65% 34.9% 45.94% 29.37% 36.66% 37.86% 3.87% 45.49% 46.94% 20 Days 68.68% 32.77% 45.36% 29.37% 36.66% 37.86% 42.79% 55.69% 55.66% Noes The variance of unhedged porfolio is generaed from he variance of CPO (Var (UnHE) = σ s 2 ). The variance of Hedged porfolio is compued based on Var (HE) = σ s 2 + h 2 σ f 2 2 h σ sf. The hedging effeciveness or risk reducion is calculaed based on HE = [ Var(HE)*/Var (UnHE)] = ρ 2. Table 5.6 describes he risk reducion achieved by he hedger. I is segregaed ino, 5, 0, 5 and 20 days forecasing period ahead for all he esed models. 20 The hedging performance resuls show ha he Inercep-BEKK model is likely o give he highes variance reducion of 60% for in-sample daa (5 and 20 days forecasing period) and ou-sample daa (20 days forecasing period). However, he Inercep- BEKK model gives he wors performance for he day forecased period. During ha day, he hedgers were only able o minimize 4% from heir oal price risk exposure 20 Please refer o Appendices F, G and H for he deailed resuls for minimum variance measuremen. 26

21 (wihin he in-sample period). More similariy wih he hedging performance was found in boh he CCC and DCC models for he in-sample and ou-sample period. An average range of 30% o 38% variance reducion was achieved in he CCC esimaion while DCC indicaes ha an average of 26% o 55.6% of variance reducion can be aained in all hree mean models (in all forecased horizon). Alhough boh he CCC and DCC models have porrayed a consisen risk reducion he magniude is modes. In addiion, he BEKK model (especially in is Inercep model) ends o have a wider ranger of risk reducion, beween 4% and 66%. The BEKK model demonsraes he same movemen beween he forecased period ahead and is risk reducion; a higher forecased horizon will give a beer risk reducion resul. In conclusion, here is no definie answer as o which model is considered bes in erms of he risk reducion achieved in hedging porfolio agains he non-hedging posiion. Alhough he evidence is mixed, based on he findings, he BEKK-Inercep ends o oucas he oher models for he 5, 5 and 20 period for in-sample and he 20 forecasing period for ou-sample esimaion. 2 These resuls do no fully suppor ha he VECM mean model is superior in erms of variance comparison o all he dynamic models, similar o Kroner and Sulan (993), Yang and Allen (2004) and Ford, Pok and Poshakwale (2005). However, i is in conras o Lien, Tse and Tsui (2002) where he CCC model ends o be less superior albei hey compared he hedging performance o 2 Similar evidence repored in: Lee and Yoder (2007) he ousanding performance from he BEKK model. Baillie and Myers (99), Bera e al. (997), Haigh and Hol (2002), Kumar e al. (2008) he dynamic model gives beer performance han he saic model. 27

22 he saic model (no o oher dynamic models). In he hedging raio conex, Lien (2004) concludes ha he exclusion of error erm (ECM) in means specificaion esimaes a lesser hedging raio. Obviously, he above findings do no suppor Lien (2004), since he VECM models, generally, do no esimae he highes hedging raio as compared o he Inercep and VAR mean models. For example, he Inercep mean model of he BEKK, CCC and DCC models provides a higher hedging raio han he VECM mean models. Furher, Wilkinson, Rose and Young (999), and Floros and Vougas (2004) documen evidence ha he EC model was no he bes model o esimae higher hedging raio han he OLS model Uiliy maximizaion Funcion Table 5.7: Hedging Performance in he Uiliy Maximizaion Funcion for he BEKK model Φ Inercep-BEKK VAR-BEKK VECM-BEKK In-sample Comparison

23 Ou-sample Comparison Noe: Uiliy Maximizaion funcion for hedging porfolio and unhedged porfolio are compued based on 20 days forecasing period ahead. The uiliy quadraic funcion is generaed from equaion 56 and he Φ denoes he degree of risk aversion for invesors ranging from 0.5 o 3.0. Table 5.8: Hedging Performance in he Uiliy Maximizaion Funcion for he CCC model Φ Inercep-CCOOR VAR-CCOOR VECM-CCOOR In-sample Comparison

24 Ou-sample Comparison Noe: Uiliy Maximizaion funcion for hedging porfolio and unhedged porfolio are compued based on 20 days forecasing period ahead. The uiliy quadraic funcion is generaed from equaion 56 and he Φ denoes he degree of risk aversion for invesors ranging from 0.5 o 3.0. Table 5.9: Hedging Performance in he Uiliy Maximizaion Funcion for he DCC model Φ Inercep-DCC VAR-DCC VECM-DCC In-sample Comparison

25 Ou-sample Comparison Noe: Uiliy Maximizaion funcion for hedging porfolio and unhedged porfolio are compued based on 20 days forecasing period ahead. The uiliy quadraic funcion is generaed from equaion 56 and heφ denoes he degree of risk aversion for invesors ranging from 0.5 o 3.0. Thus far, he previous secion discussed he performance of hedging sraegy using he minimum variance framework; his secion expands he hedging performance in he uiliy maximizaion framework in various dynamic models. The resuls for all he esimaion models are presened in Tables 5.7, 5.8 and 5.9, respecively. The framework measures he performance of such sraegy considering mean reurn, risk aversion and variance aained in he hedging sraegy (refer o equaion 9). Previous evidence has compared he saic and he non-saic model and inferred ha he larges uiliy maximizaion is achieved by he non-saic model (Kroner and Sulan, 993; Gagnon and Lypny, 995; and Yang and Allen, 2004). However, his research was more focused 3

26 on comparing he uiliy maximizaion among he dynamic models (BEKK, CCC and DCC model). All hree ables (refer o Tables 5.7, 5.8 and 5.9) ouline he hedging performance in he uiliy maximizaion funcion wihin 0.5 o 3.0 risk aversion level wihin he in-sample and ou-sample for 20 days ahead forecasing period. The resuls suppor ha he Inercep model ouperforms in boh in-sample and ou-sample daa for all hree GARCH models. However, overall, he Inercep BEKK model gives he larges uiliy maximizaion wihin he in-sample and ou-sample period. In conras, he VECM model exhibis he wors hedgers uiliy maximizaion performance among all esed models. The resuls furher suppor ha he higher level of hedger s aversion, he less he uiliy maximizaion funcion is achieved. In addiion, empirical evidence suppors a lower mean reurn posure in he dynamic models compared o he saic models (Yang and Allen, 2004). Inuiively, when invesors have a higher risk aversion i porrays a lesser olerance owards he addiional risk exposed by hem. Furher, a φ RH. higher level of risk aversion (Φ) will lead o a larger variance { / 2 VAR( )} Ω Ulimaely, he imbalance beween he mean reurn and he variance will resul in a larger negaive uiliy maximizaion achieved by hedgers, especially when he reurn porion{ E ( RH Ω ) } is small. 32

27 5.2.3 Conclusion Iniially, he research invesigaed wheher various mean specificaions have a significan effec on he hedging effeciveness in he Malaysian Crude Palm Oil markes. The sudy focuses on he Inercep, VAR and VECM mean modelling for he BEKK, Consan Correlaion model and Dynamic Condiional Correlaion model (refers o nine models). Apar from an evaluaion of common muli-variance specificaion models, our research aemps o prove he imporance of various mean specificaions ha may give differen hedging performance resuls. The diagnosic es resuls provide he exisence of non-normaliy feaures in boh he CPO and FCPO series. Serial correlaion and auoregressive and heeroscedasiciy problems were esablished in boh residuals and squared residuals, respecively. Therefore, dynamic models are more appropriae o model he ime varying second momen of he CPO spo and fuures reurns. Boh he VAR-BEKK and VAR-DCC were found o fi wih he CPO and FCPO, respecively. The models were able o couner boh he serial correlaion and ARCH effec ha were presen in boh he residual and squared residual series, however, in he VECM models i is likely o parly overcome he issues. I is no surprising ha he Inercep models were acknowledged o be he leas saisfacory among all he models in overcoming he serial correlaion and ARCH issues, since he means were only ran agains is inercep. Therefore, i is undersandable ha his inercep model has a less saisfacory resul han he oher models. 33

28 In respec of he hedging raios esimaions resuls, he esimaion proved ha he hedging raio ends o be in a non-monoonic process, which is consisen wih prior empirical evidence. Addiionally, he variance reducion resuls were mixed and he Inercep-BEKK models appear o be he bes in all BEKK, CCC and DCC models wihin he in-sample forecased periods (20 days). Meanwhile, wihin he ou-sample analysis, similar resuls were repored for he BEKK and DCC models. However, he CCC model has proven ha he VAR model ouperforms he oher mean models. In addiion, when he uiliy maximizaion funcion is considered, he Inercep-BEKK and VECM-BEKK models end o be superior for he in-sample and ou-sample periods. I was also revealed ha when hedgers are willing o olerae a risky posiion, i will elevae he hedger s uiliy level. Overall, he findings acknowledge ha he error erm mean specificaion may influence he degree of risk minimizaion, however, he magniude is merely low. Neverheless, ineresingly, he inercep model urned ou o be superior when judged agains he invesor s uiliy maximizaion funcion. The conclusion is inuiively appealing where differen mean and variance specificaion models end o affec boh he degree of risk minimizaion and he hedgers uiliy maximizaion, albei marginally. Addiionally, he evidence suppors he superior performance of he Inercep-BEKK models, which gave he bes hedging performance measuremen resuls. 34

29 SECTION III 5.3 EFFECT OF STRUCTURAL BREAKS ON VOLATILITY CLUSTERING BEHAVIOUR AND HEDGING PERFORMANCE ESTIMATION Modelling he srucural breaks has aken he cenre sage in empirical macroeconomics and finance. This is eviden from he ever-increasing number of publicaions ha have discussed his issue in recen decades. The implicaions of failing o accoun for srucural breaks in economeric modelling are many. Two of he wellknown implicaions are: () he endencies o erroneously suppor ha he ime series behave as a non saionary process raher han a saionary process in he preliminary uni-roo diagnosic es (Zivo and Andrew, 992) and (2) a misspecified model, which could lead o an error in esimaion and forecass. Bai and Perron (998) developed a comprehensive es ha allows for muliple break idenificaion, which may exis in series means. The noion of a srucural break is no sricly resriced o he mean specificaion of a series. In fac, more recenly, he regime shif idenificaion has exended ino he series second momen specificaion (see Inclan and Tiao, 994). They inroduced a similar es based on he Ieraed Cumulaive Sums of Squares [ICSS] algorihm bu caered for a series variance. These procedures have been applied o many macroeconomic variables such as exchange rae (Rapach and Srauss, 2008), US ineres rae (Bai and Perron, 2003), growh naional produc (Fang, Miller and Lee, 2008) and in he securiies markes (Aggrawal, Inclan and Leal, 999). By and large, he empirical evidence poins o he imporance of 35

30 idenifying and modelling hese breaks in boh he mean and volailiy specificaions in order o generae correc esimaes of he model and is forecass. On ha basis, i is rudimenary o check for he exisence of srucural breaks in he series and o accoun for hem in he modelling exercise. In he hedging conex, pracically, he hedging decision is likely o change over ime. By definiion, he hedging decision synonymy refers o he hedging raio ha shows he proporion of he fuures conracsagains he spo marke. Also, he hedging decision is believed o be in a non-monoonic fashion since hedgers someimes ener ino he marke o hedge less and someimes more (Karp, 987). Empirical evidence confirms he raionaliy of he non-monoonic characerisic of hedgers decisions because hey hen change he hedging percenage in consideraion of he informaion available in he marke. Fung e al. (2006) infer ha fund managers end o have a nonsaic hedging decision. They end o change heir hedging sraegies o correspond o heir risk facor concerning he environmenal changes. A similar resul was found in Meligkosidou and Vronos (2008). The source of environmenal changes can be demarcaed wihin he inernal conex (local aspec) and he exernal conex (refers o he inernaional aspec). Empirically, many researchers have deermined he presence of a regime shif in various macroeconomic series and while mos concenrae on he inernaional conex (see Fang, Miller and Lee, 2008; Fang and Miller, 2008; Rapach and Srauss, 2008; Andreou and Ghysels, 2002) some combine hese wo aspecs (refer o Aggrawal, Inclan and Leal, 999 and Zhang, Jeffrey and Rusell, 200). However, 36

31 very few sudies have explored he significance of srucural breaks wihin spo and fuures prices. In addiion, Lien (2005) specifies hree elemens ha may poenially make he hedging raio esimaion less accurae:i) a smaller sample size (in esimaion and es sample), which makes he esimaion of hedging raio less accurae and, furher, will fail o prove he effeciveness of hedging correcly;ii) he presence of a regime shif in he esed series and, finally, iii) inconsisen crierion specified in he esimaed and esed sample. His paper concepually proved ha he ECM model is able o ouperform he OLS model when a srucural break is considered in he esimaion model. Lien furher highlighs he omission of a srucural break ha may spuriously esimae he hedging raio and, herefore, we believe ha he hedging performance esimaions will also be affeced. However, limied research invesigaes he implicaions of srucural breaks in he spo and fuures reurn on hedging decisions and is performance (see Lee and Yoder, 2007). They sugges ha a regime shif is an imporan elemen ha may give a superior resul for he ECM esimaion model compared o he convenional model. Based on his evidence, we can safely assume ha a regime shif may influence he hedging performance resul. As such, his research aemps o invesigae he effec of a srucural break on he hedging decision process wihin he BEKK esimaion model in he crude palm oil 37

32 marke. 22 To idenify he srucural break number and daes, he sudy applies he Bai and Perron (998, 2004) procedure for mean, wih boh he Inclan and Tiao (994) and he Modified ICSS procedure (Sanso e al., 2004) for series variance. Furher, he research will posulae he seriousness of he non-inclusion of he srucural break in hedging performance analysis vis-à-vis he srucural break model. The research will consider boh he minimum variance and he uiliy funcion measuremen for hedging performance analysis. This research exends he exising lieraure in a number of ways. Firs, i complemens previous research on he issue of srucural breaks wih applicaions on financial and macroeconomic series from developed markes, by considering he applicaion of srucural break ess on commodiy reurns series from an emerging marke. Second, here has been considerable invesigaion of he issue of srucural changes in macroeconomic variables while very lile aenion has been given o agriculural commodiy reurns. As he agriculural secor is inerwined wih oher secors and consiues a major conribuion o economic aciviy, economy-wide changes in he levels of economic aciviy would have a direc impac on he agriculural secor. Furhermore, concerning changes in he economic srucure, agriculure is perhaps more prone o shocks caused by weaher, which can have susained and lasing effecs. In addiion, echnological changes can aler produciviy levels and can shif he way resources are allocaed, hus, having a permanen effec on he agriculural secor. Moreover, major policy reforms, boh a he naional and 22 Refer o he hird objecive of his sudy in chaper. The BEKK inercep model was seleced in his srucural break invesigaion since he model appears o be superior in boh hedging performance resuls presened in Secion II (DIVERSE MEAN AND VARIANCE SPECIFICATIONS versus HEDGING PERFORMANCES)- pg

33 inernaional levels, can induce some srucural changes in prices. Third, idenificaion of breaks in he mean and variance of he reurns series, as well as in he coinegraing relaionship beween he spo and fuure reurns suggess ha hese breaks need o be incorporaed in he model specificaion o provide more precise persisency inference. The research aemps o associae he srucural break effec on he hedging performance esimaion resul. Ulimaely, we analyse he consisency of hedging performance achieved across he esed period Time Series Volailiy Analysis Figure 5.: Plo for CPO and FCPO prices Figure 5. (a): Plo for CPO prices (RM) CPO 4,500 4,000 3,500 3,000 2,500 2,000,500, Year 39

34 Figure 5. (b) FCPO prices (RM) FCPO 4,500 4,000 3,500 3,000 2,500 2,000,500, Year Figure 5. (c): Plos for CPO vs FCPO Prices (RM) 5000 CPO vs FCPO Prices CPO FCPO 0 /2/996 7/2/996 /2/997 7/2/997 /2/998 7/2/998 /2/999 7/2/999 /2/2000 7/2/2000 /2/200 7/2/200 /2/2002 7/2/2002 /2/2003 7/2/2003 /2/2004 7/2/2004 /2/2005 7/2/2005 /2/2006 7/2/2006 /2/2007 7/2/2007 /2/2008 7/2/2008 Year 40

35 Figure 5.2: Plo for CPO and FCPO reurns Figure 5.2 (a): Plo for CPO reurns (%) RETCPO Year (%) Figure 5.2 (b): Plo for FCPO reurns 5 RETFCPO Year 4

36 The srucural break effec analysis begins wih he basic ploing of esed series prices and reurns hroughou he sampling period. Figure 5. presens boh he CPO and FCPO prices for he period beween January 996 and Augus These series end o esablish a homogeneous paern, and we can assume ha hese series are likely o move ogeher (see Figure 5. (c)). During he Asian Financial Crisis, crude palm oil prices recorded a seep hike from RM,300 per meric onne in mid 997 o RM2,500 per meric onne in early 998. During he same period, he CPO producion increased due o a good biological yield cycle during ha paricular year. Laer, a ranquil movemen in CPO and FCPO prices was regisered afer he Asian financial crisis, from 999 up uno Anoher dramaic price movemen was exhibied in early 2007 and hereafer he CPO price reached is peak a almos RM4,000 per meric onne in early In spie of he global recession saring in early 2007, hese prices were pushed up by a slower CPO producion. The shor supply was due o he weaher condiions ha affeced he level of CPO producion during ha specific period. Furhermore, he lower producion was also due o he seasonal downcycle during ha year. The supply shorage coninued ino 2008, suppored by he robus global demand for such commodiies, boosing he price o he highes amoun during ha period. Referring o volailiy ploing in Figure 5.2, he CPO reurns experienced he mos urbulen period in 200 and again in However, in 2007 a less volaile movemen was repored for FCPO reurns. A more uncerain movemen was coninuously exhibied in FCPO from 997 up o is peak in year 200. During he 42

37 Asian financial crisis, boh series experienced a highly volaile paern, similarly, during he curren global recession and commodiy producion pressure. Based on he above evidence, we can conclude ha boh series, prices or reurns, presened cerain regime changes hroughou 999 o These changes were conribued from he local and exernal facors ha direcly or indirecly affec hose series movemens. The CPO prices and reurns volailiy were more affeced by he sudden marke condiion from he Unied Saes han is own surrounding urbulence (such as he Asian Financial Crisis). For inernal facors, he supply or producion forces play a criical role in making his vegeable oil s price and reurns more variable. While he demand funcion is less likely o make he reurns more uncerain because of he srong consisen global demand for he vegeable oil. The curren jump in crude oil prices made many counries search for an alernaive source of energy such as biofuels and spurred he CPO demand curve. In addiion, he CPO is also largely used in he food processing indusry, especially in China and India, which srenghened he CPO demand over ime Srucural Break Tes Resuls Referring o he CPO and FCPO ploing price and reurn series in he previous secion, we can generally expec some regime shif hroughou he sampling period. However, we do no know he exac dae and number of regime shifs wihin he sampling period. Since he lieraure has proven he imporance of deecing he correc 43

38 number of breaks in capuring more precise volailiy esimaion, as such, we furher proceed o perform he hree break idenificaion procedures wihin for mean, variance and coinegraion conex Srucural Breaks in Variance The sudy adoped he IT ICSS and adjused ICSS (k and k 2 ) ess and he resuls are presened in Appendix I. The IT ICSS resuls idenified 4 breaks presen in he CPO variance, while only 2 breaks were confirmed by he k ICSS es resuls. The similar variance break daes deeced by boh he IT and k ICSS ess were 4/04/997, 9/09/997, 02/0/998, 0/03/998, 08/09/999, 0/0/200, 05/08/2002, 25/03/2005 and 24/0/2005. In addiion, he k es idenified addiional breaks 07/0/996, 0/05/999 and 02/08/999. However, no srucural breaks were repored under he k 2 es. By referring o he FCPO, he IT ICSS proved fewer breaks han posied in he CPO variance while similar break numbers were shown in he k es. There were 22 breaks for IT ICSS and breaks in he variance series. However, only four breaks were recognized by he k 2 es. All hree ess have one similar srucural dae, which was 3/0/996. In addiion, boh he IT and k consisenly idenified breaks on 3/0/996, 20/0/997, 26/2/997, 2/02/998, 08/09/998, 0/04/2005, and 3/03/2006. However, he k furher idenified new srucural changes in variance presen on 02/07/200 and /0/2008. Finally, one new srucural change was deeced from k 2 's four breaks, which was locaed on 8/03/

39 Based on he findings, obviously mos of he breaks occurred during he Asian Financial Crisis (wihin 997 and 998) and pos Asian Financial Crisis in 999 where he Malaysian governmen placed a capial conrol and pegged he Malaysian Ringgi agains he US Dollar. Locally, he volailiy during his period migh be conribued from a generous growh in producion ha was due o a good biological cycle for he commodiy. Anoher break was regisered in Ocober 200, a few weeks afer he erroris aacks in he US. However, he srucural change in he variance CPO reurn is no direcly hur by his even. The aack had an almos insananeous affec on he US sock markes and conagiously owards he global sock markes volailiy. This series variance changes end o be influenced by domesic forces, which consis of CPO producion shorage (lower biological cycle) and more markes owards he sock markes movemen. In addiion, he inense compeiion wih oher vegeable oil (soy oil, rapeseed oil and sunflower oil) producers who increased heir producion may have made he CPO marke more volaile ha year. The lower producion coninued o be experienced in 2002, parly due o he low biological cycle of he CPO rees, and he unsable weaher, which migh have conribued o he supply shorage. These facors furher pressured he CPO markes and explicily ranslaed ino volailiy in CPO reurns. The oil price shocks in 2005 and 2006 furher benefied he CPO producers. The populariy of he biofuel as anoher energy opion increased he world CPO demand curve and pushed he price furher. The downward supply in oher vegeable oils and fas in 2005 made he CPO prices flucuae more. 45