Sei sulla pagina 1di 133

London South Bank University

Faculty of Business, Computing & Information Management

Efficient Market Hypothesis: An empirical analysis of


forecasting share returns and volatility on Chineae
stock market
I3y

U i w m t i o n submitted in partial fulfillment ofthe requirement for


the d e g m of Msc Financc &Accounting

Contents
CHAPTER 1
Intmduction
1,

CHAPTER 2

Ch.ineae Stmk Market

2.1, ,H,iswricnlbackground and development

2.2 Structural,m d 1:nstitutional Characteristics


2.3 The 'llmfilcs

C,HaPTER 3

Literature 'Review

9 9
13
15

3. I,, :Emerging M,arkr?&

3.2 Random NMks


3.3 ,E:Eficicnt Market Hypothesis

3.4 Volatility

18

22 22
24

25

C W E H5
5.1 Objective

Time-Serieo ,EmnometrimMethaddolop

26 26

5.2 %sting for Stationary


5.2.'1 Stochastic Fmess and :Distri,buti,on

26
26

5.2.2 Inteegrated Processes end 'Differencing


5.2.3 Tcsti,ngfor Autocorrelation

29
31
31
I

5.2.3.1 The Cormlogram

5.2.3.2 T h e Box-,PierceT W & the :Iju,ng-,Box %st


5.2.3.3 T h e Lagrangc Multiplier ' k s t & the P %st
5.2.4 Unit Root 'bst
5.2.4.'1, Introd uctim

34

36 38 38 39
42
45

5.2.4.2 The Dkkcg-FullerQ6t


5.2.4.3 Thc Augmented Dickey-Fuller lkst
5.2.44' The AIC & the SBC

5.3 'Forecastingwi,th A R I A U Models

48

CHAPTER 6

Empirical bsu:lta

51
51

6.1, ,EmpiricalDistribution of lbhms


6.2 Carrelatian Coefficient of Returns
6.3 1:nregratcd:Pra~sses and Differencing For

55

Shnre Price Indims & Retwns


6.4' The Co,rrclogramfor Shnre Price I'ndices & Returns
6.5 The Q & the Q ' %st ,fm Returns

6.5 The IAfl & the F lkst for Return6


6.7 'Unit Root Ikst

6.7.1 Unit Root Tcst for Share Pri,ce 1,ndices


6.7.2 Unit b o t %st. for 'Returns

6.8 Formasting Returns with ARMA Model


0.8.1, Identification d t h c value ofp &. q

6.8.2 Modeli,ngA , W Spcci.fimti,ons

6.8.3 Formasting '&turns 83 83


85

37

89
11

93
9 ' 7
1,01,

CHAPTER 8

Cancluaion

104

107
111

116

Acknowledgements
I would like to cxpress my sincere gratitude to ,my supervisor,

Dr ,Howard Grifiths, for his guidance nnd encouragement


towards my rompletion of this thes,is. This thesis would noL

hare been possible without: his help.

I would HISO like to thank those who have kindly given me


rn1,unble'hdp ,i.nthc development [and completion o f my thesis.

Finnlly, 'I am parti,cularly grateful to my pamnts and rng sistc,r

for their support and encouragement throughout my studies.

Synopsis
This thesis examims whcthcr or not: the behaviour of the Chinese
s w , k rnn,rkcts (CShfld appear to be efici,entbased on econometrics
time-series rnethod,ol,ogics.

T h e first part o:fthis study shows. by nmalysing iheir ermr serial


correlations based on the SACIP and, the Uni,tRoot tests, the Cshls'

rcturns a,re not stationaxy tirne.sePies duri.ng the period October

1992 to September 2003,.


In the semnd l s ~ r of t this thcsis, acmrdi.ng t a the appemance o:fthe stationary time-series, we model the return processes vin ARWlA,
speci,fi,cation,s with :Fmed cm,Eficien,tsestimated from p w t d a h , to
forecnst the CSAls' Future return processes.

The last: pnrt of this thesis is mnclerned with volatility measurement, si,nce the classicnl assumption o f errors having a constant variance cannot he indicttied. We examine rdntility using models of thc G,A.R-C,HI :family nnd more spccifically the

G'A.RC,H(I,,I,) models mgctrhcr with comparison t n the ARhtAl'I,.'l)


tXIodcls.

Our main :6,ndi,ng ,isthat the historical rcturns can he used tro
forecast: stock 'returns in short,horizon in the CSAls, especially ,i.n

the Shengwn-A L5ZAI stock market. The existing potential pred'ictability implies that the CSMs are Pn'r from the weak form of
the emcient rnerk,ec hypothesi,s(EMHI.

,Furthermom, we a h a,nnlysed the correlation hetween the CSMs


themselves and the Nikkci 225. W e dkmvered that the
combination of the emerging CSMs and the well developed Nikkei
225 would be a good strategic play t u avoi,d risk,

v
L

Notation
AkHike lnformnbion Crilxrion
Augmented 'Dickey Fuller

AIC
ADF

China Securities R e p l a t m y Commission

CSRC
CSMs
DF

Chinese Stock Markets

Dickey Fuller

DOW Jones Chi.na 88


Efficient Ma.rket Hypothesis

D588

EM H
11'0s

Initial Public Offerings


Jarquc-Bera

JB

Lagrange Multiplier bIa.xirnurn Ci keli hood Renminbi Rmt Mcan Squa.re Error Sample Autocorrelation Function

LM
A4 L
R R m

RiiKE

SACF S BG
SHSE
SZSE

Schwa,rx Bayesian Criteri,on

Shmghai Stock Exchmgc


Shenzhen Stock Exchangc Shanghai-A Stwk Exchange

SHA
SHB
SZA

Shanghai-B Stock 'Exchange


Shenzhen-A Stwk Exchmgc Shenzhen-B Stwk Exchange

SZB
WTO

World Trade Orpniztitiun

v1

List of Tables
Teble 5.2.3. I Shtiona,ry or Non-stationary:

i,decision r d e

Table 5.2.3.2 Stationary or Non-stetionmy: Q,p' decision rulc


Table 5.2.4.1, Stationa,ry or Non-stationary: Unit mt d,ecision rule

Ta,ble 5.2.4.2.a The 'DF unit mot ksting p:rmdurc

Tabl,e 5.2.4.2.b Stntionary or 'Non-stationary: Thc IlF test: decision rule Table 5.2.4.3.aT h e ADF unit mot ksting pmcdure
'11~1Iilc5.2.4'.3.b Stationnry or Non-stationary: The ,ADFk s t decision rule

m i l e 6.1,

,Uistci,butionof returns stntistics Estimated cmrelstion coefficients of the retu,rnS


Surnmarkd Q & Q' statistiics of :returns

Table 6.2
Tnhle 6.5

Tabh 6.6 Table 6.7. I,,

Serin1 correlation of return$ tcsidual,~


H:U,F(:1,2) test For 'the sham price ind'ices

Table 6.7.2~A I ' F (1,2) ksc for che S:HA& the S%Areturns -1, Table G.7.2.b A.DP k s t Tor the SHA & the SZA retu,rns -2 Table S.7.2.c I W test for the SH'H & the SZ:H returns
'lhblc 6.8.1,
Estimating the A.RnM(l,q) models

Table 6.8.2.a li'he 0 1 1 2 3 esti,mations of the partmetcrs

Tnble 6.8.2.h Estirnnted A,'KMAmodels

Tabk 6.8.3
Tablc 7.2

&SUI,& of forecasts

ARCH: effect decision rules


h s u l i s of the a:KC:n effect

Table '7.3 Table 7.4 TAble -7.5 Table 7.5.a

ARCH effect decision rules


h s u l t s of Wf!%ent6 of GA:RCH(1,1,)models
Assumkg n t-distribution

Table -7.5.b

Assumi.ng a normal distribution


Forccasring vo~luti.l.ity with lGARCHh,'l) model fur

Table 7.6.a Table 7.G.b


7'nblc 7.6.c

SHA

Cornpari,n#estimarcd errors for SH.A

Forecasting volatil.ity for S U


h~:ca,surcs of forecast performances

'klble 7.7

List of Figures
Figure 6.'l,,.b The mflximum & rnin'imumstatistics of returns
F,igure 6.3 :Figure G.4,.a
Patterns oFshnre prim ind.ices And returns

Corrclogram o l share ppicc indices and returns Sample automrrclation,cocficienu of retu:rns

:Figure 6.4.b

d forccnsts 'Figure 6.8.3 'Plot of ~ t u and

,Figure 7.7

I!lot of actual nnd Forccas&

CHAPTER 1

Introduction

The purpose of the thesis


Mfe seek to build upon knowledgt? gsined from the MSc Finance Br
Accounti.ng ~
U P S Cfor

grcat~rundershnding, both empirically

and a.nalyticnlly of the Chi.nese smk ma.rkets (CSMs) pricing eEciency.

The Chi.nesc stock, marke& are emerging markets and bad not

been establ.ishcd, until htte 1,990. h behaviour has not heen


ex,arnincd in detail up until temntly. Themfore, our pu.rposes H.W:

1. 7b i.ndicate the comlations between the CSMs themelves and

the other biggest sbck market: in Asia, i.n pa:rticula.r the


Japancse stock market. Nikkei 225.
2. To form a diversi:fied pottiol.io to i.mprove risk-return trade-om.

3.

To exarni.ne whether Che behnviour o f thc CSMs appears to be


efficient.

4. lb test whether the CSMs wru,rns can

be forccast and thei.r

volntilitics (risk) ca.n be mcasurcd based on econometrics methodologies. 5. To learn mote about: emerging mnrkcts.
1

6. ' I b provide useful information for investors to understand the

bchtlviour of the CSMs

' b achieve the above purpows, some spec&


addressed BW followe,

questions

will be

>'

Can the investment risk be diversified away or reduced


si.grdicantly by H portfolio?

4 !f the

CSMs do

nQt appear

t a be eficient, how i,neEcient

are they'?

>

How to fQreCfi6tshare prim4 returns and measwe risk, hy


usi.ngeconometrics mebhodoloigies?

3. What

k.ind of models can be used IO do so, and ifany! which

measurement will perfo,rm better?

Structure of the thesis


This thesis is organized into the following chapkrs:

C h ~ p t e2 r Presents the background, development, structural and


institutional characteristics, the prof,leG, and lists some of the shortcomings of the CSMs.

Chapter 3 Reviews the most 'i:mpnrta:nt 1,iterature regarding the


'random walks, the eficicnb market hypothesis and Ibrecasting
volatility ns well 8s the major a.rbicles regarding the emerging

market$.
2

Chapter 4 Describes the appl.ication o f o r i g h d dAta, log data, log


mtutns nnd emnomettic sofiware pnckagcs.

Cha,pter 6 Intmd,uces the the-series emnometrics methodologies

used in this thesis. for testi.ng sbtionnry prowsses in the CSMs.


Thesc approaches n.m

integrated processes and di,ffercncing


the correlogrm thc ;I~OK-I?~EPZC t e s t & the ,Lju,ng-Rox&et, the hgrmgc Mul~iplier t e s t & the ,F &st
uni,troot bst including the IIF test i % the M3,F * ~ t

the ATC & the SRC selection crikrin

Forecasting with ARI,MA models

Chapter b A.nalgse the empirical results. Outcomes i.nclude 'the


tnhlcs and Ggurcs.

Chapter 7 Exa.mi,ncs volatility usi.ng mod&

o t thc GA:RCH,

hmily tcgethe'rwith mmpsrimn b the A:WAmodels.

Cha,pter 8 Concludes the m,ai.nti.ndi,ngsof this thesis and point


out the extension,of stud,ies.

CHAPTER 2

Chinese Stock Market

2.1 Historical background and development


In the enrly 'halfofthe 2 0 t h mntury, Shanghai was regarded, as the
fa,r.cn$d fi.nn,ncislmnwr with n glory stock msrket. :However,nftcr
R
I

short period of rehlgcncc, this ever financial, center had been in

drcariness u,ntil the middle 198Os, Chi.nesc government decided t r ~

rebuild stock rna.rkc& which was part o f the new policy o:f
'eoonornic reform a,nd open-ouwide'. At the ti.me, thc issues and
trades were nll owr-the-cmners. Thc situation had not heen chnngcd until late 1,440 when Shanghai S t o c k 'Exdmnge (S:RSE)

nnd Shenahen Stack Ewhmge (S7SK)were established.

b,t tho very bcgi.nni.ng of its establishment,Chi.n,csesiock market


WAS H

very small scale OF market;: 13 listed companies, several,

bill.ion RMR of ma.rkct capitdization, thousands i.nrresbors. tens o:f

brokerage :fi.rms and werage one million :IZklB of da:ily trading


volume. Along with the wide npplicntion of high-tech and h u g

demand o f market, the following one decode has b x n seen tz rapid


progress i.n the C h e w stlock market.

'I

The growth ofthc stock mar,kcte in thc last: 1,3year5 hns bcen most
impressive by any stnndmd. The h6al rnN,rkctcapitalization ~t thc

end of the year 20)02 was up to ovcr $4.4' trillion (RMB38 trillion).

This made Chinese mwket thc second largest in Asia &er Japan.
In the market there are 1.238 lisbed companies, and over 69
million investors acmu.nts 12'1.

2.2 Structural and Institutional Characbrhtics

Options available to n company for a stlock exchange listing

include:
2 Domestic hti.ng ofA:sha.ms;

>
>

Domestic Listing of B-shares;

Overseas listi.ng of Foreign 'Investment Shares;

4 ,&d-chip 1,isti.ngwhich can take various forms.

A company could list itfi domestic investment shnres (in the form of

Asbares) or foreign i.nvostment share3 (in the form of B-shares) in Shanghai or Shenzhen or its foreign invesLnient shares on
exchanges which have signed
H

Memorandum of Understanding

with the CSRC or underttake H red-chip listing.

A Shams A-shares wrc diffcscnt fmm other categories ufdomcstic


investment shares such HY etate-owned shwrm. A-Aharcs are

domestic investment shares ,issued by Chinese companies which


5
I

g.rc liskd on the S,HSE:and SZSE.A-shares are subscribed by nnd

traded nmong Chinese c i , t , h n sand/or entities.

,B Shares Foreign investment shn.res are h t e d as B.she:reson


the S:HSE or S E X T h e t ~ r mB-shn.res.d,ornesticn.lly 1,ist;cd
foreign investment shares and spccin.1 ,re,nminbi

C m t R )-

denominated shA.resall refer f a the same thing-ordi.nn.ry shares of Chi.nee shnreholdi,ng compfinies that n.m d,enorni.natd i.n RMB but: traded, i.n loreign currencies, such
8s

US dallsrs, on a Chinese

3ecu.rities e,xchn,nge. B-sha.rcscan nnlr be subscri,bcd by and,


traded n.mong foreignlcgd and naruralperwns m d ,other entities,
legal nnd natural persons from Hang ,Kong, Macau and Taiwan,

nnd Chi.ncse citimn who a.* resident: abroad.

0verSeasLis.fing of Foreign Shems All Ibrci.gn l.isting must be


approved by rhe CSRC nnd the Foreign stock exchange (and
regulahry nuthmity+such
EISthe

SEC i.n the case of n US 1isti.ng).

:Foreign sha.rcs (such

HS

H:-sha.res)must be issued in rcgktered

form and denomi,nated i.n RMR e v m though they are traded, in


Foreign currencies. :IIcpository rewipw issued ovcr H-shares arc

nlso treated RS roreign shn.ri?s.

Red Chip fikting Ked-chip compnnies a.rc those i,ncorpora&d i.n


:Ho,ng Kong

and listed on the Hong KoongStock Exchange but with

contml1,ing shnreho1,dersfrom mai.nh.nd Chi.nesc entities.


6

2 . 3 The Pmfdes
The stock mn.rkets have b e n d.ramaticn,lly gt%wi,n,g,c o r n p a d
with its short history. The high emcient; tradi.ng system has m d e

iis most i.mpressive profile. i.n berms OF central cleari.ng nnd hmk.entry, market capitoliaacion, volume of rnising money, enln.rgement of i.nrrcstorbase, number of listed Eompanics, 8s well,
a$ daily trnd'i.ngvolume cnm,pn.red with any other markem in the

world,.

I,n fiddi,tion, since its establiihrncnt, Chinesc stock rnnr,kct has been on the trip of glubnliaation, considering the Fnct that o lot of

Chi,nesc cornpnnies issued, 'Bsha:resor I.iskd in H:ong ,Kong Stock

Exchangc, Singnpore Stock %:xchAnge, Tbkyo Stock Exchange,


k n d o n Stm'k,:E:xchang!NASI?AQ and, New York, Sin& Exchange.
Meanwhile, ,B shares n,ndo v ~ r w a s I.isting he1,ped'the market start:

t o norrndkc its practicc and systcrn of Inw and a,munti.ng in li.ne


wiLh international, standa.rds. Even though, Chinew stock mn,rket

has been still suffering a series of shortmmi.ng6 AS followi.n@, due


lo i,ts i:mmaturi,ty and emnornic bacckgcound under Chi.nese

transitionnl s h p from planed economy i.nto mn.rk,eremnomy.

3hoztcomi.nm

>

hfarket mechanism is far fmm CQrnpleteness.

3 The government sometimes plays H role by changing pdicies.

CHAPTER 3

Likrature Review

3 . 1 Emerging Markets
An merging mark,& is one t h s t has publicly traded secu.rities in
developed ma.rket. Broadly
a.n effort to
EL

less

de

cd. an emerging rnarkct is H country making

change

clnd

i.mprovc its economy with the p a l of raising its

performance lo that of tbe worlds mow advanccd countries. k I ~ . u y these ,

rnarkeb ogeratc i n countries with IOW pcr-capital i.ncome. The cmergence of

capital markets in developi.n,g muniirirics is. pa.rt of a general process of globalization of capital rnarkee. Research on emerging markets has suggeskd th.rce market features: high avera,gc retu.rns (cl+al

of high

p w t h economies), high, volAtility and low correlations both across thc


emerging rnn.rkets and with developed markets.

These emerging rn~.rkets attract international i n v e s h ~ s based in developed


muntries primwily because of the high potrntial for high rctu.ms in
B

relatively short pcriod of ti.me. These high return5 exist mai.nly because the

countries, newly i:ndustrisli-lk.d and developing countries, arc experiencing


phenomenally high growth :rakes. This trend of h.igh ~ o w t h is expected t o

continue hence the high raws of return, BR w d l ns the diversification benefits it offers, ~ P , D are expected to perfiist. Rekacrt and Urias (1999) studied the

rewards t o be gained from ho1dingema:rgingmarket Rtockv i n

global equity

portfolio. They used the mesn varitrnce a,nalgsic;: whereby covariance

estimate6 w e r e used to construct expectcd exccss ,returns that corresponded

to a hypothetical efiicicnt portfo1:io. In most

CAWS,

the resultsshowed that

the expected return for the emerging 'market smcka surpassed that of the developed world equity market: i:nde,xfor opti.mal portfolios, with at least 10%
invested in emerging markets.

Emerging market, returns are found to be mow pmdict,tlble tha.n developed market returns. The sources of this predictability could be time varying risk
exposures, timc-va.ryi.rig risk premiums a.nd it could d s o be

induced by

fundamental inefficiencies. Beknert, Erb, Harvey and Viskanta (1996)

establ.ished that as a mR.rk,ec hccomes more inwgrated into world capital


markets, it is more hkely that w ~ l dinformation , will have a great impacbon

the time varyi.ng mean rcturns. Most of the predictability is strong12


influenced by local information, which is consistent a i & the fact that some ofthese countries arc segmented from world capital markeh. On the other

band, these high returns me also acwm,panied by high volatility. There is a


p a t deal of risk involvcd in investments in the emerging markets, because
emerging rnarkcts a . ~ by , d e h i t i o n , ,in H state of transition and a . subject ~
t u unexpected political and economic upheavals. The values of their stocks,

bonds and currency can chmgc d.casticnl.ly and without notice.

Divecha, Drnch and Stefck (1992)give some o f t h e reasons for this volatility
in H papcr. A few of the explnnationb

this volfitility are political instability,


10

cumncp risks, i:ll.iq,uid'ity risks and high trwnsaction

COS~S.

Market

concentration is suggcsted to be a n i.mpottant: factor, w'hereby i.n these markets, largcr stodk,s represent a higher proportion r r l the overall market capitn1,izntion. This leads to a situation wherc them are fcwer opporcuni,ties

for diversi.rIcation because the rcturns to thcsc large stocks dominate the

orernli rna,rkct return. Another k e y fncbr i s that emerging markets tend to


have H strong ma.rket-relakd forcc that afleects a'I:I sbcks withi.n a rnawket,

thu8 accentuating volatility. Thss is very much unlike the developed markets,

where forces afkfectthe various s c c t ~ rof s the economy in difkrent ways.

Bekaert and, 'Wtias (1999) a b examined the risks involved 'from holdkg

emerging rnsrket stock&

AS

part o f a n i.nternational portfoho. Their resulb

demonstrated t h a t emerging market stmks, when 8,dded to H rnean-skandn.td


deviation dhgra.m. an eficient frontier, the frontier either stayEd, thc same

or shifted 'tothe lek! i.e. the investor wi.ll &her be i,n the same, or h t k r
position as regn.rds risk. Also, tests to hnd out whether thc 1efLwa.rd shiR i.n the hontier was signi:hcnnt showed that there wcre statistically signi:ficant diversification benefits du.ringthe test period.

Furthermore, even though, ernergkg rnn.rkets are risk,)?i.ndividunU,y, :low

oorrelations between thcm and with developed, ma.rkets lead to risk,


reduction Tor i.nvesmr8.'Divccha, Drach a.od Sbfek (1,9921,k k a e r t and Urias
(1,999) and Enker: Grant and, Wmdard (2000) all demonstrate that as a

general rule, em,etging rnn.rkew are much less oorrelakd wi,th one another
than n.re the developed mn.rkets, with the exceptions M n g rnmkets in n

specific region, for example Sou,th E ~ s A t s h . A, possible reason for thew low
mrrclations that exist: is thnt the emerging market economics are unrelated.

This is as a resuk of the few economic and trade links that emerging mnrkets
have with ench other. Another 1.i.koly PCHSOII
that many of the emerging

market econmnies have, or have had,, scvere restrictions on outsiders

pa,rticipsting i.n thc:i:rrnarkek This results in the emerging markets bei,ng


insulated . F r o m the worldwide, or even regional,, pa,tkrns i.n stock market

mtu.ms. Thus. contmry fa popular belief, R modest i.nvestmenr i.n emcrgi.ng

mrukek will lead to a . d u c t i o n , rnther tha.n a.n increase, in portfolio risk for
an i:nnvt.e.hx.

'Even a , b r the two scvere fi.nancia1,and economic crises that have afn.icted
the

emerging markets, che collnpse of the Mex.imn peso :in 1994 and the

coUnpss orthe 'rhsi bath i,n 1,9912//97, there a.rc still diversi.ficatrion bCnefik6 t a

be rcaped. 'Ibk,aertn,nd Urias (1,999) have identi.fied that: there is still a 'free
lunch' i.n the emerging ma.rket equities. They have shown that dircct
exposure t n emerging markeL indices d m o s t always gives benefits at least 8s

Strong ns those From manngcd Funds. They dso note that Iacbrs common

CO

emerging m a r k e h such

8,s

poor l.iquidity amd currencyz as well

HS

macroeeconomic i,nstabil.ity,which sffecb its performance, a.re not dlcctm! by

benchmark i.ndices. Thus, the performance of i,nvcstment in emerging


ma.rke& represented by benchrna:rk i.ndices may not: nlnnys be nchicvable.

12

3.2 Random Walke


The concept o:f market eficiency arose acccid,entally. Eu.rly

resewchers were concerned about the t h e series of changes i.n


stock prices for prediction purposes. The first ,rcoord of studies i.n
the change in stock, prices ca,n bc traced back to 1,,990 when

Bachei.ier(1,901)) concluded that: the ptims i.n the llrench Hou,rsc


chmgeed in randomly fashion. His results w e l r ignored for many
years and only i n 1934, Working (1,,934) did
8

shi1a.r study on the

US s b k ma.rket.

The idea ofeKciency WM fi.rstmnccived a h r the work, ofKendnll

(19531,who invcstigared, the chn.ngcs

:i.n the prices

o f the UK

rnn.rket. W e studied the wcekly chnngcs i.n the price indices for 23
weeks .and, a.rrived at the conclusion that they did not ,follow H cycle

or trend, and that the s u w s s i v e changes were i.ndepend,ent ofcnch

other, in other words, that they moved randomly and that the

historical path provided no uscful i.nforma,tion. He added bhnt


investors made profits not: by lwrk,iag at past prices hut by luck,

inside informa,tion, speed, of reaction a.nd by the scale of their


operations. H i s resulk! puzzled many fi.nancin1emnomists, 86 they

serried to imply that the market: was d.riven by animnl spi.rits

who followed no tulcs. Many anal,ysk were led to behew that i.f prices rnovwl randomly, without much rationalle, the systcm went
ugai.n6t rill the pri.nci,p!esof classical economics.
13

dlcxa,nder (J,,961,1, however. showed that all that the rsndom,ness


in price changes i.mplied WEIS that prices adju,sbd i.mmed,iakly t.0
new in:forrnation,which i.n Facc represented thc

bl.issofcconomists.
II

H'e found that ,fi.Iter kchniques did, not suoceed in beAting

buyhold strategy considering, becnuse of transaction and tn.xntion


ws&. Athough the filter techniques he used wcrc not exactly
tl

random ptoccss, he was one

OF rhe first

to nc,k,nowldge that a,

random walk is not needed t~ ensure et%?iiciencg;this concept of =fair game" rnther than 'random wnlk WHS t o devdoped hy Osborne

(1,959 and Fnma (119'13)years later.

Back in the US, 'Rober& I19SS'I mrnpn.red the movements of the


Dow index with H model b n s d on random chnngcs a.nd Found thn,t
the patkerns were si.mi.lnr,concludi,ng that past i.nformstion was of

no prcdicLive value. H'e wns also the first t o classify the eficiency
ofthc stock exchanges i.nto wcak and semi-strong a.nd srrong.

'Fnrnn, ,Eugene (1,,9ti5)studied the serial correhtion o f the Dow 30

'Industr,ial cornpan,ies for

space of :five years. :He cxmsidered the

natuml logarithm of prices for up t~ skLeen

lam %.ndFound no

evidcnce ofsuhstanti~l l.i.ncnrdependence. :He admitted thnt there

might: be non-linenr dependence which mighb hnve i.mplied profitnble tradi.ngsystem.Pn,ma,also studied, the behnviot o,fp,rices
i.n &rms of direction. H:e concluded that large daily price changes
14

werc followed by si.m'ilarly lmgc changes with 'rmdom sign. This albeit not proper of A random w d k p m s s , WAS ta'ken as
R

si.gn of

market eEcicncy. Hk explsined that these fluctu,~tion were rhe


resulr o:foverand under Adjustment.ta 'new i,nforrnstian.

Niede,rhoEer and Usborne (I,,Ff66)studied the sign of variations for

the W S E mncludi.ng how more I.i.kely it is to observc reversals


(changes of sign> than mntinuations, and that theso wcm more
hequent a:&r sirniht continuations. H,e exphi.ned this behavior

through the study of huy/scll orders to stockb,rokcrsand rhe likely


profitabi1,ityo f unexecuted orders.

The work of British, resetircher, Dtgden (1 970) yielded similar resuks of uncorlrlation for the 'LondonStock 'Exchnngc.

3.3 Efficient Market Hypothesis


The

%MH WAS developed, as B result o f rescatxhes conducted by


WRS

Renddl ('1953)and, Cwtner (1,964) who bund tho,t it:

excess,ively diEcult to 6:nd any predictn,ble patterns i,nthe market.

In pmticular 'Kcnd,aI.I, writes: "The series looks ,like B wonde,ring

one, aImust HS ,ifonceR wmk the ,Uc,moriof Chnncc drew H rnndom


number fmrn a s,ymmetricd papthtion of fixed dispersion and
addcdit tu the current price f a Clekrminc the nmt week2 price.."ln cssenm whnt they had dcveloped w a s the ,EMH where nn emcient
1,s

m~.rket is one which Al.rcndy reflects all available i.nformation.


Ever since many academics have cmne up with mn,ny definitions of

market eficiency.

Th.ree levels of eEfciency are usually considered, after the works of

Roberts WX91 and Fa.ma (1,965); thew are wen.k form eficiency,
=mi-strong form eficiency and strong 'Perm e,Eficicncy.These th:rce

forms d i . k i.n t w m s of thc types of inforrnntion which arc used in


dewloping invcstrnent strategies.

>

&rni:strong 'FormEmciency was the cumnt priw i.ncorporatcs


a11 publicly available 'information including thRt o f thc we8.k

fom.

>

Strong II?ootm Efliciency was the current price i.ncludes a'I:I

information i,ncludi,ngthat available to insiders.

On the topic o f Chinesc stock ma.rkct efficiency, the literatu.rc is quite cxmf)icting. Laurence, Cai and

Sun (1,997) i.ndbtlt.c that LhC

ex.irtence of

tl

weak-furm eEciency i.n thc Chi.nese stock mn.rker

(for A-sha:res), testi.ng the sample period 11,991-'1996. Lang, pnync


and Ferrg (1,,999)oxami.ned meuk is thnt Chinese s b k rnsrkot

follows

random willk, over the sample period of 1,992to 1994, on

the wcekl,y bnse, by mcms of n wmianm-rntio methodology. Su,

Dongwei (20031 teskd Chi,nese B-share mntkcts a.m wcak-form

efkient: fmm 1,4411,, ia 1996.

Darrar and, Zhong (20001 provide a strong evidence for rejecting


the random-wdk hyprpothcsis i.n Chi.nesc smk m,a.tkets by U 6 h g
two difhrent nppmnches. They uti.l.hfid, daily date of the A-shn.ms

closi.ng hdex prices of the SHEE from its i.nwptrion on 20th

:December 1990 t a 1,9lh October '1,,498; nnd of the SZSE fmm i,ts
i.nmption on
4,Lh April

1,990 in '1,WOctober 11,998. They used the

stwndnrd, varinnce-rho k c s t of Lo and M~ncKi.nlny(1,988) And a


1,7

model-compariwn test thnt compa.rcs the ex post fomcnsb From

N A h G (random wa1.k ' m d e U model w,ith thow obta,i,ned b r n

several,alternative mode'ls OI.RIMA, GARCH and Arti:Eicid Neuml


Net.work-A:NNN). To ~ v d u a k ex post forecasts, they ma.ke use of
scviynl prmdu'ms including RMSE, 'Wi, Theik Inequity
b f i c i r n t , and encompassing tlcsts. Both approaches rcjccbd the

hypothesis thac Chi:nese stock ma.rket is weak-form eEciency.

Furthermom, compa.rcd with the va.riancc.rntio &st, ,resulkj Emm


the model-comparison approach a.re qu,itc decisive i.n rejecti.ng the

random-walk hypopochesis in hoth Chi.nese $rock markets.

3.4 Volatility

Mnndelbmt (1,9631 stated thnt h,rgc changes (of either sign) to


follow lmgc changes a,nd,small changes (of either sign) to Follow

sma'll changes. This is a very fmnous description o f the volrl,tility


cluster:i.ng.

Other i.mpoctant empi.tical regdariticrs are pmvidcd, by the pmsence of fat tails i,n the retu.rns, see among ochers the same Mnndelbrot (1,963) as well a s Fama (1,9911, by the existence of
forecastable events such 8s thC presence o f a high volatility before centrnl.bank decision, studied by lH,arvcyand h a n g (1,991)nnd by

the presence of non-trading period that: lad to volatility pressmc,


studied by A.ndcrsen and Rollerslev (1997).
1,8

Thew relevant tesults state the i.mpotta,nce of modding not only


the 6rst but also the second moment, while esplnini.ng the
bchnriou.r of hnncial series. HIisimicdly, ir is only since the en.rly
'80s that an etTective econometrically analysis ot it bas been

developed. The first groundhreaki.ng paper must be attributed to

Engle (1,982). Starti,ng from the

bilineA:r model developed by

Granger a,nd Andersen (1978),'he tried M capture the ch,ange in


the ~ c o n d moment , o:f the Mnditianal density function th.mugh

ti.me by an autmcgressire process of the past-obwrved c ~ n d i t i o n d


moment values. The modcl cnn bc esti.mnted eficiently by

maximum l.i.keIlibaod as showed by the m m e 'Engle who defi.ned


also the cond'itions for the process

t n be weakly stationa.ry and

meani.ngful.

All the peper. which have followed, can be seen


i.mprove or
CO ada,pt:the

HS a

WRY

original one. The fi.rstand most i.mportant

improvement 'is given by the GA:RC,H, model devcloped by Bollerslev (1986). The wnditiond variance in now c s p r e s d as

the sum of

moving average n.nd an nubregressi,rre prwess. This

permits

H mort

pa.tsimonious reprcscntation, and a consequent

esti.mntion, of the mnditionlz'l wrin,nct?process in exnctly the sa.me


WHY

as an AHMA p'rocess encompasses R IMA or a n A.R ones. The

m m e Bollerslcv defined the basic cond,itions for a menn,i.ngful


model 8s well
RS

the ones for

wcakly stationary process. The


19

Istkr is just given by the need for the cquivihent ARCH

representation to be ststionary. A deeper ~t.nalysis of them can tie

found i.n Nelson and CHO (1,992).

Rcga.rdi,nga u n i v a r h k approacb, we review

GA,KCH(1,1,) model.

The hrst mai.n extension of the hasic o.pproach was fa move from
a n assumption of normality o:fthe wnditional difitrihution of the
tmr, c1ea:rIy noc renlistic enough, t o H diffcrent distribution. The
sn,me Bollerdev (1,987) applicd
R

tdistribution, Jorion (1,988) H


(1889) n normal-lognormal

norrnd-Poisson distribution, &eh

distribution, Nelson (1992) an exponentinl distribution and, :En& Gonzales-Rivera (1,991)


H

semi-pantmetric estimation o f the

distribution.

Othor exkentiow h w e mod,i:Cedthe type o f relationship between

the dependent v

able a.nd the explanatory ones t o co,trcctspecific

drawbacks of the origi.nnl approach. Tnylor (1,9861 and Schwerr

(1,9891 tried M modcl the conditional stnndard devin,tion with a


oonseguent rcduction of the i.mportance of hrgc changes. Nelson
(1,991)devcloped the %:GA:RC,H {cxwnential GARCHI model where

the logarithm of the vn.riancc was explai.ncd by a n fiubregressive


part, a omstant and
H

non-symmetric newseffect based on the

non-squn,rederror term

OF the

previous :period,.Nion.symmetry

present also in the GJ,R model developcd by Closten, daga,nnathnn


and Run.kle (1,993) AS well as i.n the QCkRC,H (Quad.raticCA:RCHj
20

proposed, by Engle and Ng (1,993). The former obtains ic h o u g h a


dum.my variable with n value of I i.n correspondence with negative
past error terms whereas Lhe latter subtrmt fl constnnt value &om

the past error terms before 8qun.ring them. Higgins and Bern
t199!!) d e h e d the N

CH' (non-1.i.nea.rARCH). which c m be seen


85

as a mcipe for the th.me pmvious models, as well

the GA.RCH

itsclf. The Inst. covered extension of the GA.RCH wirh A univariak


appro~chis given by the FIG'ARCH (Frncrionn.lly Inkgrated

GARCHI developed by ,Baillie. Bollerslev and Mikkekon

Iisssj.

Th.is approach tries to solve the grnernl negative feature present

in,a.U the previous models of a fast &cay of persistence or, i,n other
words, H long temporal dependence ofvolati.l.ity.

The above su.rvey list i s publ.ished by BolIersIev and Kroner ('14921

in Journal dEconamclrics; ARCH modeling in Pi.nnnce.

21

CHAPTER 4

Applied Data

4.1 Data Description

Meanwhile, the Nh'ikhi, 225 index i s Used for oompmison. 'I'he


reawn why wc choose Ui.k,,kci 225, i s thet it is one o l Asia's, and,
even thc

wmld's major m,a,rketsand ids economic fundamentals are

some of the strongest: i n the Asian rcgion. BIowcwr, benri.ng i.n


mind that there a.- suspicions about: the bubbly dapanew stmk

market. Thus data i s downloaded from h t . e . ~ ~ : / / ~ t ~ t ~ ~ . e h ~ n ~ ~ a t i . c o n ~

In addition. there are severd reasons supporting why we choose

SASE &. S7S,Eindices. rather than

an official nahnwide stock

market index or Dow Jones China B8 tDJ88) i.ndex. ,Firstly.

recalhg Chapter '1, when China establ.ished n nationwide equity


ma.rket with two stock e x c h m g m locnkd, in Shanghai and Shenahcn. There
WHS

no oficid

stock market index ,ee,flectingthe


22

total, Chinese stock rnmket p e r f o r m s n ~At ~ . pwsent this rem,ai.ns


the snme.

Secondly, although the cnmpositc s b k a pkked up by Dd88 am

a,lrnost those large-cap companies with rclntivc good pcrfo,rmance of operation, they do not f i f l ~ c t the truth OF tlhei.r smock, m,a.rket
performance. Furthermom! these 88 stocks' i.ncl,uded, represent
only 33.22% of the total mnwkct cnpitdkation and, about 30% of

the tradable mn.rket capitalization.

On

the whole, we mrrns,ider that

SHA, SH& S w \ & 52,B indices


be certain that our

prescnt t,he whole set of stock,s, and we can

resulb obtai.ned arc unbimscd.

1.

DJ88 io hnwd on thc following Fnctors: the mn,rketcapihliz~tion o f public

tradable smks, the cowmgc of the industrial scctoar~a,nd the liquidity

or

stocks. According fa the Irtest ediiion of the components, t h e n are 61


mmpmics from SASE R p r c s e n h g 6i.1'996 of E w x ~ ~ A marker ~ capitdimtion in SHSE. While 27 from %LSE. rcprc5cnt 32.51% of free-fiont

market cnpitalizntion in

.%SE.

Them figures are obtaincd

from

f n ~ - i n x-ml'u.tin htm

4.2

Application of Log Data & Log &turn3

LQEdata
Firstly, we take the logarithms of the original data as applied
share indices, which is calculated
a6

follow

P, = In X ,

Where X , denotes the ocigind share indices at time t and

I n denotes the natural logarithm.

h a returas
Then, we calculate returns from a series of

4.

There a.re two

b,inds of farmat.ions of retu.rn8, wh.ich are s h p l e returns, also


known as change returns and continuously compounded returns,

also known as log returns. :In th.is thesis, we employ the log
returns, which are shown aa follows,

Where Y( denotes the mntinuously compounded return at time t.

24'

4.3 Emonometric Software Package8

We use Microfit

4.1 for: our data processing, graphic display,

estimation, hypothesis testing, etc.

25

CHAPTER 5

Time-Series Econometrics Methodology

5.1 Objective
In this chapter. we introduce some of the time-series cwnornecrics

methodologies for th,is thesis. Section 5.2 describes how to test for
stationary p r m s s . Foremsting with A:RII:M:A, mod,el,si s i,ncluded,in

Section 5.3.

Be8.r i.n mind,althvugh we are goi.ng to &et. the characteristics of

both the share price indices.

e +and thcir rcmarns.

V,,

we only

represent d l equations on 1:. Obviously, thc testing processes of

p,

a.re the snme as I", .

5.2 Testi.ngfor Stationary 52.1 SbchaBtic Pmcess and Distribution


Stmhstic tIrom&8
Gujarnti (199,Sl states that
B

stochnstic ppoccss is said

be

stationary if i t s maan and variance ape constant over time, and the
value of covariance between two t h e periods depends only on the

distnncc or lag between the two time periods nnd not on the actual
26

the

a t which the mvaria,nmis com,puted.

'In the time series likmturc such H stO(;hRStic prowss i s known

HS

a wctikly or wvfiri,am* stationary stochastic p m s s . Note that H

white noise p,tocess hns constant mean and variance, and


covminnm, except R E h g mw.

7 ~ w

The properties of a stochflstic c m be described by the moments,


which am as follows.

Whcre cr'

,isB mensure of the random variahle a,round i t s mean.


the values o f
trnd,

T,~ is H meaeure of the covarianm between

Ke,

n t 1,ag ,k. The first tr'nd,second rnomcnt equations Ill,u,strak

that n stationary p:rwessshould have B Mnstn,,nt mean, H constant


~arinnce, and n constant covn,rin.nwstructu.rc.

Djetribution
The price changes and returns are nor,mnllydistrs,bukd ove,t time. If transactions m e 'in sufic,iont number and evenly spread, across

time, prim changes should follow B norrnnl distribution.


2i

The third and fourth morncnL9 are used to measure prohability


d.istdbution,its skewness h., lack of symmetry) and kurtosis (i.e.?

~nllness or flatness), re~pecciv<ly. The measures of skewness and,


kurtosie. can be combined to determi.ne whether the returns follow
a normal distribution. Measure of skewness (SI and kurtusis

( K1

are dehned as ToUou~

The skewness i s a symmetric distribution obtained when it is

equal t o 0.While the kurtosis controls the height of the density,


the parameters kurtosis must

be 0. In other words, for a normal

distribution the value of skewness ia zero and thevalue of kurtosk

i s also zero, furthemore, tho C~ils of the density should be positive.


Thus, a normal distribution displays symmetric and mevoku.rtic
with higher tail$, higher menn. nnd lowcr midpoint.

In addition, we conduct a Jarque-Beera (JB)test for normdity. Thr

JB test i s defined as foUows,

Where n is the number of observations, S is skewness and K is


28

Under the null hypothesis khat:

is n o r m d y distribubd, the

deckion role is that if the calcuhted v d u e of JB is p a t e r than


the critical 5% level in
A

chi-equarc

(x')

distribution with 2

degrees of freedom, which is 5.99, the null, hypothesis of 8. normal

distribution is rejected.

5.2.2

Integrated Proceases and Differencing

IC i s important to hour for the time series whether ur not the


underlying stochastic proccss, that generated the series, c a n be
aEsumed to be invariant with respect to t h e .

If the characteristics of thc

stmhsstic process changed or;er time

G.e., non-stationwy), it will often be dimcult to represent the time


series over

past and future intervals of time by a simple algebraic

mdel. On the other hand, if the stochastic process is fixed in time


Le., stationary), h e n we uin model the process via an equation

with Gxed coefficients that C

he ~ estimated, from past data.

Morwver, the statmnary data knds to return to its mean and

fluctuation around it. On the other hand. H non-stationary process will have different mean a t different points in time.

In addition, running regre.ssiun on nun-stationan; data

can give

24

rise to spurious v a h e s ol: K ' , Durbin-Watson (DW) test, and t


statistics causing ecunatnists wriously misleading to conclude,

that a metdngful relntionship o,xit;btr among the rcgression


variables. Becausc nun-sta'cionriry variwldea have nn infini,te variance, inferenw: using O M is ,invalid.

Although it is difficult. to model non-strltiona.rypruceases, they can


often be transformed into stationary or approximately stationa.ry

processes by integrated processes and differencing. For example,

#ere

iiy

is the first-di.ffemnceof

Y, and

K,

is assu.med to be

I I

white noise p m e s s . The flhovc equation describes chat- the series


is mid tn be integrated of order one, denokd l r l j , because

taking a first dimerence prod,uces

stationaq process. A

non-stationary series i s integrated of order d, denoted I(d! i.f it


becomes stationary after being first d.Secenced dtlme6.

30

5-2-3 Testing for Autocorrelation

5 . 2 . 3 . 1 . The Comelogram
The oorrclogram is a plot of 'the autocosrcletion function (ACF).
The ACP provides t3 partinl description of the ptocess for modeli.ng purposes. I t tells u s if the true disturbances are nutommelated

and how much mrrelation there i,s between n,eighboring datn


points in 'thc wries

5, W e define 'the nutoconeletion with lag k as

Where p r is the autocorrelation coe:t%cientas a function of the lag k. ( p a is the wvarianrn betweea
and

<,,,

normdized by

dividi.ng it by the variance cif 1' .I

We assume that

is distributed as B s t a n d d normal, then the

sample autcmrrclat.ion coefkients are also a p p r d m a k l y normal


in their distribution.

Where n is the number of observations, p1 dcnotes the


aukorrelation coefficient nt h g k estimated from H sample, called

the sample autocorrelation function (SACF).


31

lficients

up to the 12th lag, Far the share price index and its rcturns,

respectivelj?

Secondly, we will plot the mrrelogrnms in order Eo determine


whether th,eg depict the
CO

cients hI.l.ing off

nurntiers

insigni.ficantIg difXeren,tfrom 7xr0 and ff uctunting around it, if it is,

the series nppears to be stationary. Conversely, i.f they dccrease

slowly toward mro. the se,ries illustrates non-stationary


Consequentlyt for n strlti,onnr)l variable the correlogrnm should

show autooorrelations h a t dic out f d y quickly


large.

a5

k becomes

Thirdly, we will conduct significance tests for the autocorrelation

cwfficients, by constructi,ngB rnnfidence interval for the estimated


autmorrela.tian coeficients t u determi,ne whether they are
signifi~nntly different from am. A rough 95% confidence htervnl

will be drawn at * I . S S / m . A t H gl,ance,i.f bt falls outside this


region, then we reject; the nu,U hypothesis that the true value of the

mffrcient: tit that lag k is ze,m. On the other hand, if we cannot


reject the null hypothesis, it implies that the ACF 'ie
R

zero
32

mvarimce white noise process.

By recnlhng

the definition of the

weakly or covariance stntionaq stochastic pnxess born Section 52.1, we can conclude that the mries Y , expresses to be a

stationruy p m o s s .

FindlJ: the SACF docisiu rulo is

rmmlirimd i n Table 5.2.3.1.

Non-stationary

Gm?:
=0
ij, # 0
dfnotes the value at the 95% confidence intend
denotes the value at the 5% sigoificance level

It is worth noting that thc test: is subjective, because given the fib
of up to the 12th lag is only for sampling. Therefore, one may

condude that the series iu $latAioaa,ry, while others may consider

that it is a non-stationary eeonom,ic time series.

However, to test the joint, hypothesis that all the autceorrelation

coefficients are zero, we use ihe Q statistic intruduced by B o x


and P i e m and the mod.ified statistic developed by Ljung-Box in
the following section.

5.2.3.2 T h e Box-Pierce Teat


The BQX - fierce 'Cest

the Ljung-Box Test

Box and Pierce (1,!170) test the join,t hypathmis that 811 lag k o f the

autacorrelation coefficients arc eimu,ltaneously equal to zcm. The

Q test is carried out

H S follow^,

Where n is the number of observations nnd k is the maximum lag

lea@. Q statistic is distributed


reedom,

RS

ch.i square with k degrees of

xi,under the null hypothesis that all k autxmrrelation

mficients are ' e m .

The Lima - BQX Teat


Avarimt of the Box . Pierce test, has been developed by Ljung and
Box (1378). Si.nce the Box . Piercc test has poor small sample

properties, implyi,ng that it mriy lead to the wrong decision fmquently for small sa,mples. :Ljung a.nd Box show that the Q' test is as follows,

However,we forecast the

two tests am equivalent. Because the


34

sample size is relatively large in this thesis, the (n+2) and (n-k) terms in the Ljung - Box formulation will not be significant. Therefore, the
Q'

test would be close to the

test.

Consequently, this test examines Y, ' s correlations between the residuals, E , , and k lagged values of the residuals.

Table 5.2.3.2 Stationary or Non-stationzy.. Q, Q' decision ru1e


e

Stationary

1
I

Nonxtationary

1
I

Table 5.2.3.2 describes that if the calculated value of Q & Q' are greater than the critical 5% level in a x,' table, rewritten' as
Q&Q' > x i , we can conclude that the null hypothesis of

non-autocorrelation in

E,

is rejected. Note that rejecting the null

hypothesis, means accepting a n alternative that at least one autocorrelation is not zero. In other words, we reject the null hypothesis that the time series is generated by a white noise process (in which all autocorrelations should be zero), namely, rejecting stationary process.

35

5.2.3.3 The Lagrange Multiplier Test & the F Test


The Lagrange Multiplier test
An alternative approach to test if the true disturbances are serial
correlation, is the Lagrange Multiplier Breusch and Godfrey (1978).

(LM)test

introduced by

The test is the one in which the square of the OLS residual is regressed on an intercept and its lagged values, with the number of observations, R,times the RZ.It is shown as follow

The LM statistic is distributed a s chi square with k degrees of freedom, x i . Therefore, the decision role is that if LM > x i , we reject the null hypothesis of serial uncorrelation in we do not reject it.
E,,

if otherwise

The F test
Testing for serial correlation of errors can also be done by using an

F test to examine the coefficients of lagged OLS residuals, in a


regression of the OLS residuals on the lagged OLS residuals and the original regressors.

The F test statistic is operated as follows,

36

test stutistic =

RSS l(n - k )

F(k-1,n-k)

Where ESS is explained sum of square, RSS is residual sum of squares, n is the number of observations and k is the number of regressors in unrestricted regression.

The decision rule is that if F > F , , ( k - 1 , n - k ) , hypothesis of serial uncorrelation in


E,,

we reject the null

and vice versa.

On the whole, the LM and the F statistics have the same


distribution asymptotically. However, in small samples the F version is generally preferable to the LM version. As our samples are relatively large, we employ both of them in the thesis.

37

5.2.4 Unit Root Test 5.2.4.1 Introduction


The most common test of stationary is known as the unit root test, introduced by David Dickey and Wayne Fuller (Fuller, 1976; Dickey and Fuller, 1979). Consider the simple model;

y, = q-, + E ,

Where

E,

is a stationary error term with zero mean, constant

variance and non-autocorrelated. Recall integrated processes and differencing from Section 5.2.2,
E,

is I(0). Thus,

can be

denoted I(l), since A T = E , . The above model can be expressed in more general form as the first-order autoregressive process, AR(1)

Y,

= ax-, + E,

Table 5.2.4.1 Stationary or Non -stationary Unit root decision rule

c
Non-stationary

&
Stationary

38

Table 5.2.4.1 carries out that if la(<l, then Y, is I ( O ) , namely stationary, however, if a = l , then non-stationary. Thus, we will test if root, to investigate
U

I')

is I(l), meaning

=1,

in other words a unit

is stationary or non-stationary time series.

We use the t statistic to test whether a is significantly different from 1. Note that if
E,

is I(O), we will use Dickey-Fuller (DF) test


E,

as the t statistic. On the other hand, if

is not I @ ) , (i.e., the

error term is serial correlation,) the augmented Dickey-Fuller (ADF) test will be performed. To investigate the properties of the error term, we will mainly depend on the result of the LM & the F

tests.

5.2.4.2 The Dic-ey-Fuller test


Recall in the AR(1) model

Let the above equation subtracts

Z-l

on both sides, we have

Where

= a- 1 . For theoretical and practical reasons, drift and


39

However, the DF t-test statistics do not follow the usual tdistribution under the null hypothesis, since the null is one of non-stationary. The statistics follow a non-standard distribution, and critical values have been tabulated by Dickey and Fuller on the basis of Monte Carlo simulations (1979). The DF test decision rules are described in Table 5.2.4.2.b.

Table 5.2.4.2.b

Stationary or Non-stationary The DF test decision rule


DF > critical value
- - I ;

DF < critical value

Table 5.2.4.2.b exhibits that when using the computed DF t-statistic whether it will be greater or not than the relevant

critical value, and to prove if the null hypothesis of non-stationary


will be rejected or not.

For example, if DF > critical value, then y

= 0 , hence

containing a

unit root can be identified. This implies that after taking a first
41

difference procedure, the series

Y,

appears to be a stationary

process. Hence, we can arrive a t the conclusion that Y, expresses non-stationary time series.

Conversely, if DF < critical value, the null hypothesis of non-stationary time-series will be rejected.

5.2.4.3

The Augmented Dickey-Fuller Test


E,

The DF tests are valid only if

is white noise. However, if

E,

is

serial correlation, the ADF test will be employed. The ADF test is proposed to accommodate error autocorrelation by adding lagged differences of

Y, .

Recall the case (iii) model in Table 5.2.4.2.a, the DF unit root testing procedure,

Which can be rewritten by adding lagged changes in U, on the right-hand side of the equation as follows,

The lags of A& cancel out the serial correlation, to ensure that

E,

42

,,.,le

5.2.4.3.b

Stationary or Non-stationary The ADF test decision rule


ADF > critical value

+
Non-stationary

I
I

f
Stationary

1
I

The above table shows the ADF test decision rule is that if ADF < critical value, we reject the null hypothesis of non-stationary time-series, otherwise we do not reject it.

It is important to note that how to determine the appropriate lag length for the augmented regression, since including too few will

not remove all of the serial correlation and using too many will
increase the coefficient standard errors.

Concerning this issue, we will firstly observe a rough idea from the graphs of the correlogram, and the LM & the F test statistics. Then we will posit a preliminary number o f lag length for the

augmented regression to compute the ADF tests.

44

Secondly, the Akaike information criterion (AIC) and Schwartz Bayesian criterion (SBC) will be employed to measure criterion at each lag, in order to identify the correct lag length for each stock market. The AIC and SBC will be explained respectively in the following section.

Lastly, we will use the correct lag length to reexamine the ADF unit root test to further prove the characteristics of the series Y, .

5.2.4.4

TheAIC & the SBC

The test criteria to determine appropriate lag lengths are the AIC proposed by Akaike (1973) and the SBC proposed by Schwarz
(1978).

The AIC
The AIC is defined as follows,

AIC

= In(s)- k

Where In(s) is the maximized value of the log-likelihood function, and

is the maximum likelihood (ML) estimator based on size n,

k i s the number of dependent variables. While, h(g) is calculated as below

45

I"(@

= --[1+

n
2

log(2r52)I

Where S2 is the ML estimator of the error term,


E:

E,,

given by

l n . Thus, the equation of the AIC can be rewritten as

AZC=--[l+l0g(2~5*)]-k

= 3

AZC =--(I n 2

+ 10g2z)--10g52 n
2

-k

The decision rule is to select the highest value of AIC to determine the appropriate lag lengths.

The SBC
The SBC is defined by

Where

B"

is computed the same a s in AIC.Thus, the SBC can also

be presented as

SBC

= --(I

n 2

+ log 2 r ) --log Z2 --log


2

k 2

46

The decision rule is that choosing the highest value of SBC to determine the appropriate lag lengths.

It is worth noting that since logn will be greater than 2 (nis 131
in our samples), the SBC will always appear to be less than the

AIC in this thesis. Moreover, in practice the use of the SBC is


likely to result in selecting a lower order of lag length than when using the AIC.

47

5.3 Forecasting with AFtIMA Models

Assuming the return series will be demonstrated stationary in previous section. The next matter is how we can model a stationary time series, in order to describe its behavior. Moreover, how we can use the fitted econometric model to forecasting returns.

Autoregressive Integrated Moving Average, ARIMA (p,d,q), model associated with Box and Jenkins(l976) is common method of modeling stationary time series and forecast. It is a sophisticated extrapolation method, using only past values of the variable being forecast to generate forecasts.

Autoregressive processes
We have already referred to AR(1) model in the unit root test section. Generally, an autoregressive model of order p, denoted an AR(p), is expressed as follows,

=p +

CaiK-; +E,
i=l

Where

E,

is a white noise error term. An AR model describes that

Y) depends on the values in the previous time period and an error


term. In other words, we could forecast U, by using AR(p)
48

processes.

Moving average processes


A moving average model of order q, denoted a MA(q), is defined as follows,

A MA(q) model is simply a linear combination of white noise


processes. This means depends on the current and previous

values of a white noise error term.

ARMA processes
An ARMA(p,q) model is obtained from combining the AR(p) and MA(q) models, showing as follows,

The above model states that

depends linearly on its own

previous values plus a combination of current and previous values


of a white noise error term. In short, the characteristic of a n

ARMA process is a combination of those from the AR and MA


parts.
49

To identify the value of p and q, the correlogram is the common


method, known as the Box-Jenkins (BJ) methodology. However, by just a visual inspection of the estimated correlogram, this can be a rather casual way. Thus, we are going to select models specification with the highest value of the AIC and/or the SBC to determine the appropriate values of p and q.

Integrated processes
We have introduced an integrated process of order d in Section
5.2.2. Recall from the section, a n integrated process of order d, denoted an I(d), means after differencing d times, Y , appears to be stationary.

Forecastina urocesses

>

Identifying the values of p, d and q for each stock market from October 1992 through June 2003.

>

ModelingARIm (p,d,q) equations for the above sample period.

9 Forecasting three month stock monthly returns based on the


estimated ARIMA model from July 2003 to September 2003.

9 Plotting the forecasted and actual series for the monthly

returns in the four Chinese stock markets.

50

largest one month fall in the four Chinese stock markets is about double compared to the Nikkei 225.

From the above evidence, we understand that the Chinese stock returns seem to exhibit much higher volatility than the well-developed Nikkei 225 during the test period. The certainly high volatility implies that the risk is absolutely great in the Chinese stock markets than in the Japanese stock market. Referring to the relationship between risk and returns, the high risk may reflect that there are considerably great returns in the Chinese stock markets than in Japan. Nevertheless, the risks are also obviously high.

Consequently, it can reasonably be concluded that normality is rejected for any of the four Chinese stock returns. Conversely, the Nikkei 225 reflects a normal distribution in the test period.

54

coefficients may reflect that there is no correlation between the two Shanghai shares and the Nikkei 225.

These correlation coefficients may provide information on how investors can deal with risk. It has been widely argued in finance literature that the standard deviation of returns on an individual asset is not itself an appropriate measure of risk since the investor, by holding a portfolio, can diversify some risk away. The ability to do this will largely depend on the correlation between returns on different shares. This suggests that investors could form a diversified portfolio to improve their risk-return trade-off.

By recalling from Chapter 2, A shares are subscribed by and traded among Chinese citizens and/or entities. While, B shares can only be subscribed by and traded among foreign legal and natural persons and other entities, legal and natural persons from Hong Kong, Macau and Taiwan, and Chinese citizens who are resident abroad.

Hence, an individual investor could only form a diversified portfolio from either A shares or B shares. Although a combination of an A share and a B share can diversify some risk away, this would not be applicable in the Chinese stock market. However, for the international investors, the combination of the Chinese and the Japanese shares would be a good case to avoid risk.
56

After considering the generalities of the empirical distribution and correlation in the Chinese and the Japanese stock markets, we turn to concentrate on testing the stationary in the four Chinese stock markets. We first investigate whether their share price indices exhibit stationary processes, and then consider their first differences of share price indices stochastic processes, namely returns.

Figure 6.3 plots their share price indices and returns during the period considered. By visual inspection of the graphs, we can easily arrive at the conclusion that the share price indices stochastic properties cannot be assumed to be invariant with respect to time, since none of them having the tendency to return
to its mean and fluctuate around it. Therefore, we consider that

the four stock price indices stochastic processes are clearly non-stationary.

However, the graphs of their returns in Figure 6.3 illustrate conversely that their characteristics of the stochastic processes seem to be approximately constant and fluctuating around their means. In other words, all of the returns seem to be stationary time-series.

Nevertheless, it is important to note that there are some higher volatilities which cannot be considered to be constant, especially
58

for SZA. The variation in the series appears to be fluctuating with several clusters of large movements. Although the variance becomes unusually high, we still consider that returns seem to be stationary time-series in the stock returns.

Recall our discussion of the integrated processes and differencing in Section 5.2.2, we can summarize that the four stock share price indices imply IO), since taking the first difference produces a stationary process. While, their returns appear to be I ( O ) , namely stationary time-series.

6.4

The Correlogram for Share Price Indices g, Returns


Another method to test the stationary is the correlogram, which we analysed in Section 5.2.3.1. The correlograms for the four share price indices and for their returns are shown in Figure 6.4.a. The complete version of the results is included in Appendix.

At a glance, we observe that the share price indices decrease


slowly toward zero as the number of lags increase. While, the correlograms for the returns illustrate that the coefficients fall quickly to near zero a t lag 1 and fluctuate around it. These phenomena perhaps suggest that the series of the four stock returns plot white noise processes. Conversely, their original share
59

It is worth noting that the phenomena of the nearly zero SACF at the early lags, indicates that there is no strong evidence of autocorrelation in the stock returns, and the successive price returns exists independently.

The above finding illustrates that all of the SACF are not statistically different from zero for the four stock returns. This confirms the suggestion observed from the correlogram, that returns plot white noise processes. In other words, they appear to be stationary processes.

Consequently, the empirical results of the correlograms consist with the outcome observed from integrated processes and differencing, that original share price indices indicate

non-stationary time-series, but their returns show stationary time-series.

62

lower than the critical value a t the 5% level. These results suggest that none of the four stock returns is significantly autocorrelated
for the sample autocorrelation series. This supports the outcome

from SACF.

Hence,

we

can

conclude
E,

that

the

null

hypothesis

of

non-autocorrelation in

is not rejected. In other words, the four

stock returns appear to be stationary time-series.

64

As we can see from Table 6.6, the two B shares show that there is
certainly no serial correlation of residuals. Even at the lower lags, they still reflect significantly higher probability of falsely rejecting the null hypothesis. This result is consistent with our prior findings of stationary time-series from integrated processes and differencing, the correlogram and the Q & Q' tests.

In contrast, both the LM and the F tests provide a serial correlation of residuals a t the lower lags for the two A shares. However, the remaining statistics do not appear to be serial correlated, apart from
9th

lag for SHA. Hence, it is difficult to

conclude whether the returns of the SHA and the SZA exhibit stationary time-series, even major the LM and F statistics < x i . We
will employ the unit root test to illustrate this matter in the

following section.

66

We have detected from Figure 6.4.a, that a certain serial correlation exists of up to 12 lags in the four share price indices for the samples. While, Figure 6.3 reveals the patterns of all of the indices which are random walk with clearly drifting and no strong trend. Hence, we conduct the ADF unit root test to demonstrate the stationary for the share price indices by using case (id, namely random walk with drift, referring to Table 5.2.4.3.a.

Table 6.7.1 reports the ADF test up to 12 lags for all of the log of original share price indices. It provides all the computed ADF test statistics that are greater than the critical value of -2.8857 at 5% significant level. Thus, we cannot reject the null hypothesis. This means all of the four share price indices series have a unit root and are non-stationary time-series.

This finding consists with the outcomes observed from integrated processes & differencing method and the correlogram method.

68

6.7.2 Unit Root Test for Returns


In previous Section 6.6, we have detected there is serial correlation of residuals in the SHA and the SZA a t lower lags. Therefore, we employ the ADF test for the A shares. On the other hand, the DF test will be selected for the B shares, since their residuals appear to be independent.

Moreover, reviewing the Figure 6.3 again, the patterns of the four stock returns show white noise process visibly has no trending behaviour. Hence, the random walk with drift model can be specified for the four stock returns.

SHA&SZA
Before carrying out the unit root test, the order of the ADF regression needs to be selected. Looking a t the Figure 6.4.b again,

it exhibits that the autocorrelation coefficients of SHA and SZA a t


9th lag appear to be the highest one in the samples. Furthermore,

the LM & the F test statistics in Table 6.6 illustrate that there are serial correlations at lag 1, 2 and 9 in the SHA, and a t up to lag 3 in SZA, respectively.

Thus, we posit the appropriate lag length for the augmented regression is 9. However, we choose 12 lags, just in case the lag
i

length of 9 cannot minimize the value of an information criterion.


69

Table 6.7.2.c gives the computed D F test statistics they are considerably smaller than the critical value of -2.8838 at 5% significant level. Thus, we can conclude to reject the null hypothesis. This means that the SHB & the SZB returns appear to be stationary time-series.

Furthermore, we can conclude that both of the SHB and SZB express AR(1) process in the same sample period.

Consequently, the above findings are consistent with our


previous results of non-stationary share price indices and stationary returns, examined from Section 6.3 and 6.4, which are the integrated processes & differencing and the correlogram methods.

Conclusion
In conclusion, the characteristics of the four Chinese stock share price indices can be assumed to change over time, in other words, they appear to be random walk in the sample period.

In contrast, the Chinese stock returns illustrate invariant with respect to the same period. This suggests that we can model the return process via a n ARIMA(p,d,q) model with fixed coefficients estimated from past data, to forecast future return process.
73

Note that since all the sample stock returns appear to be I(O), thus we employ the ARIMA(p,O,q) or simply the ARMA(p,q)

specification to forecast returns.

74

In the previous section, we have identified all of the four stock returns appear to be AR(1) process. Thus, the ARMA(p,q) models will simply be the ARMA(1,q) specifications for our samples. Table 6.8.1 gives the details of estimating the ARMA(1,q) models.for q=0,1,2,3,respectively. Comparing the values of the AIC andlor the SBC, we select the model specification with the highest value. Thus, we observe that the two A share returns show an ARMA(1,l) model, while the SHB returns appears to be simply an AR(1) model.

Note that there is a problem of the specification of the MA(q) in the SZB, in which when we using the AIC, we have ARMA(1,2) is selected, on the other hand, ARMA(1,O) is chosen by the SBC. To deal with this task, we adopt the Box-Jenkins approach using the correlogram provided in Figure 6.4.a. Reviewing the correlogram again, it can be seen that the SZB returns does not seem to be MA(2) process. Therefore, we conclude that anARMA(1,O) model is selected for the SZB, namely an AR(1) specification.

76

almost capture the trends, and the predicted values are very close to the actual values in the tested three months, especially for the first two month forecasts.

SZB This

is the only one forecasted values fluctuating the

actual values. The AR(1) model appears to capture the trend in July, but fails to predict the turn occurred from August to September.

On the whole, the regression models seem to be efficient for the first one month stock return prediction in a l l of the Chinese stock markets during the sampling period. Moreover, the ARMA(1,l) model shows a powerful tool in forecasting monthly returns in the

SZA. In addition, the AR(1) model predicted successfully in the


first and the third month returns in the SHB.

Conclusion
We can arrive at the conclusion that the historical returns can be used to forecast stock returns in short horizon in all of the Chinese stock markets, especially in the SZA. This evidence of the statistical studies reveals that share prices appear not to be random walks and exist significantly potential predictability in the Chinese stock markets. This implies that the weak form of the

EMH seems to be unsatisfied over the testing period from October


81

1992 to June 2003. The inefficient Chinese stock markets indicate that investors are able to identify share price movements using past sequence of share prices, and may can statistically or economically beat the markets.

82

CHAPTER 7

Modelling Volatility

7 . 1 Introduction
In Section 5.2.1, we introduced variance is assumed to be constant over time in a stationary time series, known as homoscedasticity. Then, in the following chapter, we determined that all the Chinese stock returns series appeared to be stationary in the sample period.
Thus, we may consider that the variance of the error terms,
(r2

(second moment), is constant over that time.

However, &om the Figure 6.3, it can be easily seen that there are evidences of some clumps of large and small errors, known as volatility clustering originally described by Mandelbort (1963), who stated that large changes (of either sign) to follow large changes and small changes (of either sign) to follow small changes. This kind of volatility clustering implies that volatility at time t tends to be positively correlated with its level during the immediately preceding periods, such as at time t-1.Hence, we could consider that volatility is autocorrelated. Furthermore, the Figure 6.3 also expresses that periods of high volatility follow by periods of low volatility.

These phenomena would suggest that the classical assumption of

errors having a constant variance, both unconditionally and conditionally, is not valid. Conversely, o 2 appears to change from period to period. In other words, o 2 is heteroscedastic in a l l the Chinese stocks returns and, this implies that o2 depends on the volatility of the errors in the recent past.

Under such heteroscedasticity, we are able to model volatility by using autoregressive conditional heteroscedastic (ARCH) models introduced by Robert Engle(1982) and generalised ARCH (GARCH) models developed by Tim Bollerslev (1986).

84

7 . 2 ARCH Models
ARCH models spec& how to model and forecast conditional

variances of the error terms, o,? .a : is explained as follows,

, ! is equal to the conditional The above equation describes that a

expected value of

E , ? . In

other words, a : depends on how large the

errors were in the past. An ARCH(q) model is given as

a, = a,

+Ea,,-,
2 ,=I

In the ARCH(q) model, a , ?depends on q lags of squared errors.

E,

is normally distributed with zero mean and variance a,?. In short,


a : has two components:

c,!= a constant + the ARCH term

Where a constant is a, and the ARCH term means the last period's squared errors, namely, last period's news about volatility. The ARCH effect decision rules are given in Table 7.2.

85

Table 7.2 ARCH effectdecision rdes


H, :a, #O,a, # O;..,a,
f

No AFK!H,effect

ARCH effect

Table 7.2 suggests that if there is no autocorrelation in the error variance, we have the null hypothesis of no ARCH effect, otherwise,
we have the alternative hypothesis of ARCH effect presented.

Referring to our samples, we have already demonstrated in last chapter that all the Chinese stock returns appear to be autoregressive processes of order 1. Thus, ARCH(q) models would be

ARCH(1) models in this thesis. The ARCH(1) test involves running a


regression of squared OLS errors estimated from the AR(1) model. The key point is to examine whether a, = 0 , we will employ the LM test and F test to ensure this matter. Recall from Section 5.2.3.3,we have introduced the LM & the F tests. The decision rules for ARCH effects are if LM > x2 and F > F,,(I,n-2) , we reject the null hypothesis of no ARCH effect and vice versa.

86

based on stationary time series do seem to be valid.

It is important to note that although we reject the null hypothesis of


no ARCH effect in the SHA and the SZA returns, this does not certainly imply that the conditional variance of the error terms is not constant. The reason for this is that the serially correlated error term at the lower lags, reported in Table 6.6, may cause t.he conditional variance to be variable. Consequently, the ARCH effects may be less significant in this case.

However, we still consider the ARCH effects are present in the A shares returns series. By recalling the volatility literature review, the ARCH effects suggest that the two A share returns series appear to be stationary process. This fmding is consistent with our previous results from last chapter.

The further discussion would be that whether or not there are generalized GARCH effects presented in the SHA & the SZA returns over the same sample period.

88

7.4

GARCH Models

Tim Bollerslev (1986) extended Rober Engle (1982)'s work from ARCH(q) models to GARCH(p,q) models. GARCH(p,q) models are shown as follows,

In the GARCH(p,q) model, the conditional variance depends on q lags of squared errors and p lags of the conditional variance. In other
words, the conditional variance of the error terms has three

components:

0; = a

constant +

the ARCH term

the GARCH term

0 :

= a constant + last periods volatility

+ last period's variance

The above expression means the variance today depends on all past volatilities and all past variance and a constant term.

Table 7.4 GARCH effectdecision rules


H,

:PIf 0,P2 + O,...,pp f 0


GARCH effect

No GARCH effect

89

Table 7.4 shows the decision rules of GARCH effects. It can be easily seen that if the coefficients of the conditional variances are equal to zero, namely no autocorrelation, we cannot reject the null hypothesis
of no GARCH effect. Conversely, we reject the no GARCH effect

hypothesis.

whether GARCH affects are present in the Tests for determining . errors in our samples, we conduct using the following steps;

F To detect the order of p in GARCH(p,q) models for the SHA & the

SZA returns, we refer to our previous result of order 1gained from


MACS) progress approach in Section 6.8.1. The reason why, depending on the moving average of order q, is that GARCH(p,q) models are effectively ARMA(p,q) models for the conditional variance. Since, the equation of GARCH(p,q) models look very much like an ARMA(p,q) process [Walter Enders (1995) pp 1461. Therefore, we consider that the order of q in an ARMA(p,q) model is equal to the order of p in a GARCH(p,q) model.

Recall from Section 7.3, we have already examined the order of q


is 1for ARCH. Hence, GARCH(p,q) models would be GARCH(1,l)

models for the SHA & SZA returns from October 1992 to June 2003. Therefore, we have the following GARCH(1,l) model for the

two A share returns.

90

D To estimate the coefficients of p, , with the aid of Microfit we use


the maximum likelihood (ML) estimation.

D To find out if p, = 0 , we use t-tests shown as follows,

t=

coefficiert Std. Error

The t-test of significance decision rules is that if the computed t statistic is greater than 1.645 a t the 5% level, then the coefficient of the conditional variances is significant. Hence, we reject the null hypothesis of no GARCH effect. On the other hand, if t < 1.645, we have the null hypothesis of no GARCH effect.

It is important to note that although the conditional variance is changing over time, the unconditional variance of shown as follows.
E,

is constant and

If Ip,I 4 , a, >O and a, +p, (1, then we have a positive unconditional variance, this means the, GARCH(1,l) model is a stable process.

91

If ai +pi >1, variance of


E,

Vur(&,) becomes negative, the unconditional

is not defined, and this would be termed

non-stationary in variance. Moreover, as the forecasting horizon increases, the conditional variance of the error terms will tend towards infinity.

If a, +PI =1, it is known as a unit root in variance, called


integrated GARCH(1GARCH). IGARCH models are not stationary GARCH models, they become less and less useful as the prediction horizon increases.

In addition, GARCH estimations can be carried out by using either a


normal or a t-distribution for the conditional distribution of the errors. We will firstly estimate the coefficients of a,, a, and

pi under the

two assuming distributions respectively; then compare which

distribution is more suitable in our samples, based on the AIC & the SBC model selection criteria. The selection procedure involves choosing the appropriate distribution with the highest value of the AIC and/or the SBC.

92

* Red color figure denotes significance at the 5% level in t-ratio column. * Red color figure denotes the highest criterion in the A l C and SBC columns.
Under the assumptions of a tdistribution and a normal distribution for conditional errors, we have estimated the above results of coefficients of GARCH(1,l) models for the SHA and the SZA returns in the sample period from October 1992 to June 2003.

The values reported in Table 7.5 based on the AIC & the SBC selection criteria, suggest that the conditional distribution of the SHA return errors appear to be normal. Thus, the results in Table 7.5.b are taken into account. The table b shows all of the t-statistics are significantly higher than the critical value of 1.645. This seems to reflect GARCH effect presented in the SHAreturns.

Nevertheless, since at +PI = 1.0015 > 1, then we have the unconditional variance of the errors, Vur(~,), given by

a, / l - ( a l +PI) is negative. This means that V u r ( ~ , ) cannot be

defined and is non-stationary. It will lead to conditional variance of the error terms tend t o infinity, a s the forecasting horizon increases.

Moreover, a, +/3, = 1.0015 is very close to unity. This implies that shocks to the conditional variance will be highly persistent. In fact,
we have observed an IGARCH(1,l) model for the SHA returns. As we

introduced early, an IGARCH model is not a stable GARCH model


94

and becomes weaker and weaker when the prediction horizon increases. Therefore, the evidence of GARCH effect observed from tstatistics would be less important.

Consequently, we consider that there is a GARCH effect in a very short forecasting horizon. Under this consideration, we will forecast the volatility month by month for July, August, and September in
2003 respectively in the next section. The IGARCH(1,l) model is

estimated as follows.

SZA
As we can see from Table 7.5, the AIC & the SBC selection criteria choose the t-distribution for the conditional distribution of the SZA returns errors.

In Table 7.5.a, the calculated tstatistics of

p, , the coefficient of the

GARCH term, is at 1.410, which is slightly less than the critical value of 1.645 at the 5% signifcant level. Thus, we have p,=O. This means
no GARCH effect presented. However, if we consider the significance

at the 10% level, we have p, is greater than the relevant critical value of 1.282, then we have effect presented.

PI z 0 . This would suggest the GARCH

To determine whether or not the GARCH effect is present, we refer to


95

the assumption of a normal distribution, ignoring the AIC & the SBC selection. The result is that GARCH effect is strongly presented, since
t-statistic of

p,

is absolutely higher than the critical value of 1.645

at the 5% significant level, and also greater than the critical value of
2.326 at even 1% significant level. Furthermore, since the restrictions

of

IP,1<1, a,>O and a, +/3,<1 are all tested satisfactorily, then we

have Vur(&,) is positive. This suggests that the conditional variance of the error terms satisfies a stable GAFXH(1,l) model under the assumption of a normal distribution.

Therefore, we can reasonably consider that GARCH effect is present in the SZA over the sample period. Hence, we will attempt to forecast volatility based on the GARCH(1,l) model observed from Table 7.5.b shown as follows,

It is important to note that the sum of the coefficients on the lagged


squared error and lagged conditional variance, a, +p, , are approximately 0.83 and 0.81, respectively, in both cases. The relatively large figures imply that a large positive or a large negative return will lead future forecasts of the variance to be high for a long period. Furthermore,, the constant terms, a,, are all very small, while a, and

p, are similar in proportion and around 0.4. This

implies that a : depends on both its last periods volatility and variance appropriately equally.
96

7.6 Forecasting Volatility with GARCH(1,l) Models


Volatility (risk) is one of the most important concepts in the whole of finance. GARCH models are the most popular non-liner financial models for modeling and forecasting volatility, which allows the behaviour of a series to follow different processes at different points in time. In this thesis, we forecast volatility using the GARCH(1,l) models estimated from last section for the two A sharesreturns.

SHA
Recall the results of GARCH(1,l) effects observed &om Section 7.5, we have found the presence of IGARCH effect in the SHA returns, namely an IGARCH(1,l) model. Thus, we forecast its volatility in only one-months time. To do this we follow three steps;

1. Using IGARCH(1,l) model estimated, from October 1992 to June 2003 in Section 7.5, to forecast volatility for July 2003.

2. %-estimate a new GARCH(1,l) model using the period from

October 1992 to July 2003, by adding the actual volatility in July,


to forecast volatility for August 2003.

3. Repeat the procedure again for 6; in September 2003.

The reason for forecasting only on one month each time is that, in the theory, an IGARCH model is not a stable GARCH model and becomes
97

* Red c o l o r figure denotes the smallest absolute error i n the same month.
It is interesting to note that the forecasts observed based on repeating
the IGARCH'(1,l) models do not show better performances than the IGARCH(1,l) method
as

expected.

Conversely,

using

the

IGARCH(1,l) model gives us some slightly smaller forecasting errors, compared to using the IGARCH' (1,l)model. This result may be due

to the different testing horizons, between the one month period and
the three month period, cannot be seen horizon increases significantly. Therefore, to discover whether or not the IGARCH(1,l) models become less and less useful as horizon increases, as the theory stated, will remain for our further work.

More interestingly, from the inspection of the forecasting based on the ARMA(1,l) model, we can see that the ARMA model predicted a slightly better result than the most popular GARCH model in a short horizon. However, the GARCH model does show that its predictabilities are stronger than the increases, in the case of the SHA.

ARMA

model as horizon

SZA
Under the assumption of normal distribution for conditional errors, we forecast volatility based on the GARCH(1,l) model estimated from Table 7.5.b for the SZA returns, and summarised in Table 7.6.c together with the forecasts observed from the ARMA(1,l) model in Section 6.8.3. 99

Conversely, the ARMA(1,l) models show more accurate forecasting volatility for the SHA, rather than for the SZA.

However, when we measure the forecast performances based on the two models in either the SHA or the SZA, we can conclude that the GARCH(1,l) models express their forecasts having a smaller degree of errors than the ARCH(1,l) models in our samples.

In addition, it is important to note that all of the values reported in


Table 7.7 are considerably small. This means that the GARCH(1,l) models and the ARMA(1,l) models determine that they are appropriate for predicting future return volatility in the two A share stock markets. In other words, return volatility can be forecasted well by A R M and GARCH models. This implies that the SHA & the SZA appear to be inefficient markets in the test period.

103

CHAPTER 8

Summary and Conclusion

In general, the EMH is concerned with whether stock prices fully


reflect all the information available at that point in time. Weak form tests of the EMH model focus on the information subset of historical price or return consequences. This thesis examines the behaviour of the CSMs, with a view to determine whether or not they are consistent with the weak form of the EMH.

In this thesis, we have examined the CSMs do not appear to be


consistent with EMH, by using some econometrics time-series methodologies.

We firstly carried out by analysing the empirical distribution of returns, and found out that the CSMs' returns diverge from the normal distribution, compared to the Nikkei 225, and identified that there were certainly high volatility occurring in the CSMs.

In the following section, we estimated the correlation between the


CSMs themselves and the Nikkei 225, and discovered that the combination of the emerging CSMs and the well developed Nikkei 225 would be a good strategic play to avoid risk.

104

After considering the above issues, we turned to concentrate on


testing the stationary time-series in the CSMs share price indices and

their returns by testing their error serial correlations. To do so, several approaches were employed, including the integrated processes & differencing, the SACF, the Q & Q' tests, the LM test, the F test, the Unit Root test (the DF test & the ADF test) with aid of the AIC & the SBC selection criteria. Through these tests, we identified that the characteristics of the CSMs share price indices appeared to be random walk, conversely their returns illustrated invariant with respect to the test period.

According to the stationary returns time-series, we could carry out forecasting returns with ARIMA(p,d,q) models using fixed coefficients estimated from the in-sample returns, to forecast the out-of-sample returns, and showed the forecasted versus the actual returns for three month period. We arrived at the conclusion that the historical returns can be used to forecast stock returns in short horizon in all of the CSMs, especially in the SZA.

Furthermore, we examined volatility using models of the GARCH family and more specifically the GARCH(1,l) models together with comparison to theARMA(1,l) models. We found that the GARCH(1,l) models are appropriate for predicting future return volatility in the two A share stock markets. Meanwhile, the GARCH(1,l) models express their forecasts having a smaller degree of errors than the ARCH(1,l) models by using the RMSE models.
105

Consequently, we can arrive at the conclusion that the historical returns can be used to forecast stock returns in the CSMs, especially in the SZA. This reveals that share prices appear not to be random walks and exist significantly potential predictability. This implies that the weak form of the EMH seems to be unsatisfied over the test period. The inefficient CSMs indicate that investors are able to identlfy share price movements using past sequence of share prices, and may can statistically or economically beat the markets.

Finally, concerning our extension of studies, we intend to apply and test whether the Neural Network Regression (NNR) and Recurrent Neural Network (RNN) models can produce a substantial

improvement in the out-of-sample performance of our volatility forecasts.

106

Bibliography

1. www.cia.gov/cia/publications/factbook/index.html

2. www.csrc.gov.cn

3. Alexander, Sidney S. (1961) Price movements in speculative market: trends or random walks Industrial Management

Reviey 2 (May), 7-26.

4. Bachelier, Louis (1900) Theorie de la speculation Paris:


Gauthier-Villars, and reprinted in English in
-

Paul Cootner

(1964 ed.). The random character of stock market prices


Cambridge: M.I.T.

5. Bekaert, G and Urias, MS (1999) Is there a Free Lunch in


Emerging Market Equities?

The journal

of Portfolio

Management, Spring 1999, 83-95.

6. Bekaert, G, Erb, CB, Harvey, CR and Viskanta, TE (1996) The


Behaviour of Emerging Market Returns Working Paper, Duke University and Stanford University.

7. Cootner, Paul. (1964) The Random Character of Stock Market


Prices Cambridge, Mass.: MIT Press.
107

8. Divecha, AB, Drach, J and Stefek, D (1992) A Quantitative Perspective !/?he journal of Portfolio Management, Fall 1992,

41-50.

9. Dryden, M. H. (1970) A statistical study of UK share prices

Scottish Journal of Political EconomJs 17, 369-389.

lO.Eaker, M, Grant, D and Woodard, N (2000) Realised Rates of Return in Emerging Equity Markets The journal of Portfolio

Management, Spring 2000,41-49.

ll.Elton, Edwin J., Gruber, Martin J., Brown, Stephen J. and Goetzmann, William N. (2003) Modern Portfolio Theory and Investment Analysis ( Sixth Edition), John Wile & Sons, Inc.

12. Fama, Eugene. (1965) The Behavior of Stock Market Prices

Journal of Business, 38, January, 34-105.

13. Fama, Eugene. (1970) Efficient Capital Markets: A Review of


Theory and Empirical work Journal of finance, XXV, No.2 May, 383-417.

14. Fama, Eugene, and MacBech, James. (1973) Risk, Return


and Equilibrium: Empirical Tests Journal of Political EconomJs

81, No.3 May/June, 607-636.


108

15. Gandhi, Devinder K., Saunders Anthony and Woodward

Richards. (1980)

Thin capital market: a case study of the

Kuwaiti stock market Applied Economics, 12, 341-349.

16. Jarque, C. M. and Bera, A. K. (1987) A Test for Normality of

Observations

and

Regression

Residuals

International

Statistical RevieF Vo1.55, 163-172.

17. Kendall, Maurice G. (1953) The analysis of economic time series, part I: prices Journal of the RoyalStatistical Societs: 96, 11-25.

18. Laurence, M., Cai, F., Qian, Sun (1997) Weak-form efficiency and causality tests in Chinese stock markets Multinational

finance Journal, Vol.1, No.4, 291-307.

19. Long D.Michael, Payne Janet D. and Feng, Chenyang (1999) Information transmission in the Shanghai equity market The

Journal of financial Research, Vol. XXII. No. 1,29-45.

20.M. Mandelbrot (1963) The Variation of Certain Speculative Prices Journal of Business, Vol. 36, 394-419.

21. Neiderhoffer, Victor, and Osborne, M. F. M. (1966)

Market

Making and Reversal on the Stock Exchange Journal of the


109

Americin Statistical Associa tion, 6 1(December), 897- 9 16.

22. Roberts, Harry V. (1959) Stock market Patterns and financial analysis: Methodological Suggestions Journal of

Finance, 14 (March), 1-10.

23. Su Dongwei (2003) Chinese stock market World Scientific.

24. Working Holbrook (1934) A random difference series for use in the analysis of time series Journal of the American

Association, 29 (March), 11-24.

25. Zhou, G (1999) Security factors as linear combinations of economic variables Journal ofFinanciaIMarkets, 21, 402-432.

110

Reference
A: JOURNALS

1. Alexandros E. Milionis and Demetrios Moschos (2000) On the

validity of the weak-form efficient markets hypothesis applied to the London stock exchange: comment Applied Economics

Letters, 7, 419-421.

2. Chris Fawson, Terry F. Glover, Wenshwo Fang and Tsangyao


Chang (1996) The weak-form efficiency of the Taiwan share market Applied Economics Iatters, 3, 663-667.

3. D.A. Dickey and W.A. Fuller, Distribution of the Estimators


for Autoregressive Time Series with a Unit Root Journal of the

American StatisticalAssociation, Vol. 7, 1979, 427-431.

4 . Epaminondas E. Panas (1990) The behaviour of Athens stock


prices Applied Economics, 22, 1715-1727.

5. Gulnur Muradoglu, Nese Akkaya and Jamel Chafra (1998)


The effect of the establishment of an organized exchange on weak form efficiency: the case of Istanbul Gold Exchange The

European Journal of Finance, 4 85-92.


111

6. Musta

1. Gultekin,

Bulent Gultekin. (1983) Stock market

seasonality, Journal of financial Economics, 12, 469-481. North-Holland.

7. Richard D. F. Harris and C. Coskun Kucukozmen (2001) The Empirical Distribution of UK and US Stock Returns Journal

of Business f i a n c e &Accounting, Vol. 28 No 5 & 6 JuneIJuly.

8. Robert

Engle

(1982)

Autoregressive

Conditional

Heteroskedasticity with Estimates of the Variance of U.K. Inflation Econometrica, Vol. 50, pp. 987-1008.

9. Robert L. Brown and Stephen A. Easton (1988) Weak-form Efficiency in the Nineteenth Century: A Study of Daily Prices in the London Market for 3 per cent Consols, 1821-1860

Economica, 56, 61-70.

10. Sarath P. Abeysekera (2001) Efficient Markets Hypothesis and the Emerging Capital Market in Sri Lanka: Evidence from the Colombo Stock Exchange
-

A Note Journal of Business

finance & Accounting,


0306-686X.

28 (1) & (21, JanuaryMarch,

11.Simon Keane Emerging Markets - The Relevance of Efficient


112

Market Theory ACCA Research: Technical & Research

Committee Occasional Research Paper, No.15.

12.Tim Bollerslev (1986) Generalized Autogressive Conditional Heteroscedasticity Journal of Econometrics, Vol. 31, pp. 307-327.

B: TEXTS

1. A. Koutsoyiannis (1977) Theorv of Econometrics 2nd Macmillan

2. Christopher Dougherty (2002) Introduction to Econometrics 2nd Oxford University Press

3. Damodar

N.

Gujarati.

(1995) Basic

Econometrics

3Id

McGraw-Hill, Inc

4. David Madsen (1992) Successful Dissertations and Theses San

Francisco: Jossey-Bass Publishers

5. Edward J . Kane (1969) Economic Statistics & Econometrics;


An introduction to auantitative economics Harper & Row, Publishers, Incorporated

6. George G. Judge, R. Carter Hill, William E. Griffiths, Helmut


113

Lutkepohl and Tsoung-Chao Lee (1988) Introduction To The Theorv And Practice Of Econometrics 2nd John Wiley & Sons

7. G. S. Maddala & In-Moo Kim (1998) Unit Roots. Cointegration,


and Structural Change Cambridge University Press

8. John Sloman (2001) Essentials of Economics 2nd Financial Times Prentice Hall

9. John Y. Campbell, Andrew W. Lo and A. Craig Mackinlay. Econometrics of Financial Markets New Jersey: Princeton University Press, Princeton

10.M. Hashem Pesaran & Bahram Pesaran (1997) Working with Microfit
4.0:

Interactive

Econometric

Analysis

Oxford

University Press

11.Peter Kennedy (2003) A Guide To Econometrics 5th Blackwell, Simon Fraser University

12.R. F. Engle & C. W. J. Granger (2000) Long-Run Economic Relationships; Readings in Cointeeration: Advanced Texts in Econometrics Oxford University Press

13.Robert S. Pindyck & Daniel L. Rubinfeld (1995) Econometric


114

Models & Economic Forecasts 4th McGraw-Hill

14.Walter Enders (1995) Applied Econometric Time Series John Wiley & Sons, Iowa State University

15.William H. Greene (2003) Econometric Analysis 5th Prentice Hall, New York University

C: INTERNET RESOURCES

115

Appendix

116

Page 1 5/3/2004 15:31:32

...............................................................................
118 observations used in the estimation Of all ADF regressions. Sample period from 19931112 to 2003M9
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Unit root tests for variable DLSHA The Dickey-Fuller regressions include an intercept but not a trend

DF ADFl1) ADFl2) ADFl3)


~~~~

Test Statistic -11.2144 -9.3326 -6.7332 -6.4319


-h.2RSR
~~~~~~

AnF141 , ~ ,

...............................................................................
95% critical value for the augmented Dickey-Fuller Statistic = -2.8859 LL = Maximized log-likelihood AIC = Akaike Information Criterion SBC = Schwarz Bayesian Criterion HQC = Hannan-Quinn Criterion Unit root tests far variable DLSHA The Dickey-Fuller regressions include an intercept and a linear trend

ADFl5) ADFl6) ADFl7) ADFIB) ADFl91 ADFllO) ADFl11) ADFI12)

-5.7205 -4.6316 -4.9674 -3.4940 -3.8012 -3.0532 -2.8294 -2.5009

LL 87.5132 89.5833 89.6755 90.6085 91.7227 91.9459 92.4677 94.0400 98.1694 99.3952 100.7443 100.7986 101.2446
~~ ~

AIC 85.5132 86.5833 85.6755 85.6085 85.7227 84.9459 84.4677 85.0400 88.1694 88.3952 88.7443 87.7986 87.2446

SBC 82.7426 82.4273 80.1341 78.6818 77.4106 75.2485 73.3850 72.5719 74.3160 73.1565 72.1201 69.7891 67.8498

HQC 84.3883 84.8959 83.4255 82.7961 82.3477 81.0084 79.9678 79.9776 82.5445 82.2078 81.9944 80.4862 79.3697

............................................................................... ...............................................................................
Test Statistic DF -11.1637 ADF(11 -9.2914 ADFIZI -6.7039 -6.4038 ADFl31 ADFl4) -6.2579 ADFI51 -5.6948 AUFl6) -4.6064 ADFII) -4.9397 ADFIB) r-3.4645 ADF191 -3.7690 ADFI10) -3.0238 -2.8047 ADFl11) -2.4819 ADFl12) LL 87.5148 89.5856 89.6777 90.6115 91.7240 91.9469 92.4735 94.0425 98.1908 99.4047 100.7577 100.8108 101.2524 AIC 84.5148 85.5856 84.6777 84.6115 84.7240 83.9469 83.4735 84.0425 87.1908 87.4047 87.7577 86.8108 86.2524 SBC 80.3588 80.0442 77.7510 76.2994 75.0266 72.8642 71.0054 70.1891 71.9520 70.7806 69.7483 67.4160 65.4723 HQC 82.8274 83.3356 81.8652 81.2366 80.7866 79.4470 78.4111 78.4176 81.0034 80.6548 80.4453 78.9359 77.8151 118 ohservations used in the estimation of a l l ADF regressions. Sample period from 1993M12 to 2003M9

...............................................................................
95% critical value for the augmented Dickey-Fuller Statistic = -3.4481 LL = Maximized log-likelihood AIC = Akaike Information Criterion SBC = Schwarz Bayesian Criterion HQC = Hannan-Quinn Criterion

Page 1 5/3/2004 15:32:11

Unit root tests for variable DLSHB The Dickey-Fuller regressions include an intercept but not a trend
******************~*****fftff*ff*ff*f**************************~,***********************

...............................................................................
DF ADFi1) Test Statistic -9.6221 -7.5454
LL

129 observations used in the estimation of all ADF regressions. Sample period from 1993M1 to 2003M9 77.9462 78.0242 AIC 75.9462 75.0242 SBC 73.0864 70.7345

...............................................................................
95% critical value for the augmented Dickey-Fuller Statistic = -2.8838 LL = Maximized log-likelihood AIC = Akaike Information Criterion SBC = Schwarz Bayesian Criterion HQC = Hannan-Quinn Criterion Unit root tests for variable DLSHB The Dickey-Fuller regressions include an intercept and a linear trend

HQC 74.7842 73.2812

...............................................................................
...............................................................................
DF ADFI11 Test Statistic -9.5922 -7.5247
LL 77.9923 78.0737

129 observations used in the estimation of all ADF regressions. Sample period from 1993111 to 2003M9 AIC 74.9923 74.0737 SBC 70.7026 68.3541

HQC 73.2493 71.7497

**+**+*****I******************t*ttttt+tt***************************************

95% critical value for the augmented Dickey-Fuller Statistic = -3.4450 LL = Maximized log-likelihood AIC = Akaike Information Criterion SBC = Schwarz Bayesian Criterion HQC = Hannan-Quinn Criterion

Page 1 5/3/2004 15:31:6

Unit root tests for variable DLSZA The Dickey-Fuller regressions include an intercept but not a trend
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

118 observations used in the estimation of all ADF regressions. Sample period from 1993M12 to 2003M9
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

DF ADEll) ADF(2)
ADPITI . . . .,. ,

Test Statistic -10.0042 -7.9052 -5.6219


-4.91154
~ ~~~~

...............................................................................
95% Critical value for the augmented Dickey-Fuller Statistic = -2.8859 LL = Maximized log-likelihood AIC = Akaike Information Criterion SBC = Schwarz Bayesian criterion HQC = Hannan-Quinn Criterion Unit root tests f o r variable DLSZA The Dickey-Fuller regressions include an intercept and a linear trend
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

AUFl4) ADF151 ADFi61 ADFi71 ADFi81 ADF191 ADF(10) ADFl11) ADF(12)

-4.7431 -4.1580 -3.5779 -3.4453 -2.6461 -2.9396 -2.4942 -2.3255 -2.1642

LL 79.8726 80.2551 81.0973 81.1017 81.3692 81.4457 81.8068 81.8304 84.6193 85.8072 86.6707 86.7566 86.8576
~~ ~~~

AIC 77.8726 77.2551 77.0973 76.1017 75.3692 74.4457 73.8068 72.8304 74.6193 74.8072 74.6707 73.7566 72.8576

SBC 75.1019 73.0991 71.5560 69.1749 67.0571 64.7483 62.7241 60.3623 60.7659 59.5684 58.0466 55.7472 53.4628

HQC 76.7476 75.5676 74.8474 73.2892 71.9942 70.5082 69.3069 67.7680 68.9944 68.6198 67.9209 66.4442 64.9827

118 observations used in the estimation of all ADF regressions. Sample period from 1993M12 to 2003M9
...............................................................................

Test Statistic OF -9.9658 ADFlll -7.8762 . . -5.6024 ADFi2) ADFi3) -4.8891 ADF(4) -4.7273 ADFi5) -4.1448 -3.5666 ADF(6) -3.4334 ADF(71 ADF181 -2.6372 ADFi9) -2.9296 ADF(10) -2.4866 ADF(11) -2.3201 ADF(12) -2.1595

LL 79.8994 80.2841 81.1252 81.1293 81.3974 81.4729 81.8375 81.8604 84.6598 85.8506 86.7094 86.7907 86.8914

AIC 76.8994 76.2841 76.1252 75.1293 74.3974 73.4729 72.8375 71.8604 73.6598 73.8506 73.7094 72.7907 71.8914

SBC 72.1434 70.7428 69.1985 66.8172 64.7000 62.3901 60.3694 58.0070 58.4211 57.2265 55.6999 53.3959 51.1112

HQC 75.2120 74.0342 73.3128 71.7544 70.4600 68.9730 67.7751 66.2355 67.4725 67.1007 66.3970 64.9158 63.4540

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

95% critical value for the augmented Dickey-Fuller statistic = -3.4481 LL = Maximized log-likelihood AIC = Akaike Information Criterion SBC = Schwarz Bayesian Criterion HQC = Hannan-Quinn Criterion

Page 1 5/3/2004 15:32:29

...............................................................................
129 observations used in the estimation of all ADF regressions. Sample period from 1993M1 to 2003M9
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Unit root tests for variable DLSZB The Dickey-Fuller regressions include an intercept but not a trend

DF ADFI1)

Test Statistic -11.0473 -6.7912

LL 64.8640 66.2586

AIC 62.8640 63.2586

SBC 60.0042 58.9689

HQC 61.7020 61.5156

...............................................................................

95% critical value for the augmented Dickey-Fuller Statistic = -2.8838 LL = Maximized lag-likelihood AIC = Akaike Information Criterion SBC = Schwarz Bayesian Criterion HOC = Hannan-Quinn Criterion Unit root tests for variable DLSZB The D1Ckq-Fuller regressions include an intercept and a linear trend
**+**+************************tfftifttf*********************************~+,****

...............................................................................
DF ADF(1) Test Statistic -11.1093 -6.8504 LL 65.4710 66.7011 AIC 62.4710 62.7011 SBC 58.1813 56.9814

129 observations used in the estimation of a l l ADF regressions. Sample period from 1993M1 to 2003119

...............................................................................
95% critical value for the augmented Dickey-Fuller Statistic = -3.4450 LL = Maximized log-likelihood AIC = Akaike Information Criterion SBC = Schwarz Bayesian Criterion HOC = Hannan-Quinn Criterion

HOC 60.7280 60.3771

Page 1 5/3/2004 15:37:57

GARCHI1,lj assuming a t distribution converged after 37 iterations


. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Dependent variable is DLSHA 127 observations used for estimation from 1992M12 to 2003M6
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

...............................................................................
R-Squared S.E. of Regression Mean Of Dependent Variable Residual Sum of Squares Akaike Info. Criterion DW-statistic
.0041514

Regressor ONE DLSHA (-11

Coefficient .0037798 -.076680

Standard Error .0071960 .lo813

T-Ratio[Prob] .52527[.6001 -.70914 L.4801

...............................................................................
Parameters of the Conditional Hetemscedastic Model Explaining H-SQ, the Conditional Variance Of the Error Term
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

.13797 .0079073 2.3414 103.8520 2.0072

R-Bar-Squared -.020138 F-stat. F ( 3, 123) .17092[.916] S.D. of Dependent Variable ,13660 Equation Log-likelihood 108.8520 Schwarz Bayesian Criterion 96.7415

...............................................................................
H-SQ stands for the conditional variance of the error term. E-SQ stands for the square of the error term.

Constant E-SQI- 1) H-SQI- 1 1 D.F. of t-Dist.

Coefficient .0017566 ,47825 .44197 7.9051

AsmDtotic Standard Error .0012218 .32477 .27864 7.4899

Page 1 5/3/2004 15:35:48

...............................................................................
Dependent variable is DLSHA 127 observations used for estimation from 1992M12 to 2003M6 ............................................................................... Standard Error T-Ratio[Probl Coefficient Regressor .0068328 .53772[.592] ,0036741 ONE .096826 -.99981[.3191 -.096807 DLSHA 1-1
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

GARCHI1,l) assuming a normal distribution converged after 28 iterations

...............................................................................
Parameters of the Conditional Heteroscedastic Model Explaining H-SQ, the Conditional Variance Of the Error Term Coefficient .0019623 .lo969 .29146

R-Squared S.E. of Regression Mean of Dependent Variable Residual Sum of Squares Akaike Info. Criterion DW-Statistic

.0034333 ,13802 ,0079073 2.3431 104.2561 1.9678

R-Bar-Squared -. 020873 F-stat. FI 3, 1231 ,14125L.9351 S.D. of Dependent Variable ,13660 Equation Log-likelihood 108.2561 Schwarz Bayesian Criterion 98.5677

...............................................................................
Constant E-SQC- 1 1 H-Sal- 1) Asymptotic Standard Error .8622E-3 ,24364 .12328

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

H-SQ stands for the conditional variance Of the error term. E-SQ Stands for the square of the error term. *WARNING* The unconditional variance of the above GARCH processis undefined!

Page 1 5/3/2004 15:37:17

...............................................................................
Dependent variable is DLSZA 127 observations used for estimation from 1992M12 to 2003M6
................................................................................

GARCH(1.11 assuming a t distribution converged after 31 iterations

Regressor
ONE

DLSZA (-11

Coefficient -.011222 -.012075 -.Ole773 .13014 .0055453 2.0831 99.8709 1.7258

Standard E r r o r ,0079244 ,10099

T-Ratio[Prob] -1.4162 [ .I591 11957 [ .go51

-.

...............................................................................

...............................................................................
Parameters of the Conditional Heteroscedastic Model Explaining H-SQ, the Conditional Variai.ce of the Error Term Coefficient ,0036313 ,46629 .36446 4.2673

R-Squared S.E. of Regression Mean of Dependent Variable Residual Sum of Squares Akaike Info. Criterion DW-statistic

R-Bar-Squared F-stat. Fl 3 , 1231 S.D. of Dependent Variable Equation Log-likelihood Schwarz Bayesian Criterion

-.043621
*NONE*

.12739 104.8709 92.7604

...............................................................................
constant E-SQ(- 1) H-SQ(- 11 D.F. Of t-Dist. ASympt.OtiC Standard EIIOC .0026930 .26352 ,25855 1.6274

...............................................................................
H-SQ stands f o r the conditional variance of the error term, E-SQ stands for the square Of the error term.

Page I 5/3/2004 15:36:42

GARCH(1.11 assuming a normal distribution converged after 40 iterations


. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Dependent variable is DLSZA 127 observations used for estimation from 1992M12 to 2003M6
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Regressor ONE DLSZA (-1)

Coefficient -.0077389 ,0016952 -.010804 .12963 ,0055453 2.0668 92.6830 1.7616

Standard Error .0096004 .lo693

T-Ratio[Probl -.80610[.4221 ,015854I.9871 -.035457 *NONE* ,12739 96.6830 86.9946

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

...............................................................................
Parameters of the Conditional Heteroscedastic Model Explaining H-SQ, the Conditional Variance of the Error Term
...........................................................................

R-Squared S.E. of Regression Mean Of Dependent Variable Residual Sum o f Squares Akaike I n f o . Criterion DW-statistic

R-Bar-Squared F l 3, 1231 F-stat. S.D. of Dependent Variable Equation Lag-likelihood Schwarz BayeSian Criterion

...............................................................................
H-SQ stands for the conditional variance of the error term. E-SQ Stands for the square of the error term.

Constant E-SQI- 1) H-SQ(- 11

Coefficient .0035234 ,40389 .40191

Asymptotic Standard Error .0014804 ,18650 .15136

Potrebbero piacerti anche