Integrating Modelling Technique

INFORMATION TO USERS
This manuscript has been reproduced from the microfilm master. UMI films the
text directly from the original or copy submitted. Thus, some thesis and
dissertation copies are in typewriter face, while others may be from any type of
computer printer.
The quality of this reproduction is dependent upon the quality of th e copy

submitted. Broken or indistinct print, colored or poor quality illustrations and
photographs, print bieedthrough, substandard margins, and improper alignment
can adversely affect reproduction.
In the unlikely event that the author did not send UMI a complete manuscript and
there are missing pages, these will be noted. Also, if unauthorized copyright
material had to be removed, a note will indicate the deletion.
Oversize materials (e.g., maps, drawings, charts) are reproduced by sectioning

the original, beginning at the upper left-hand comer and continuing from left to
right in equal sections with small overlaps.
Photographs included in the original manuscript have been reproduced

xerographically in this copy. Higher quality 6” x 9” black and white photographic
prints are available for any photographs or illustrations appearing in this copy for
an additional charge. Contact UMI directly to order.
Bell & Howell Information and Learning

300 North Zeeb Road. Ann Arbor, Ml 48106-1346 USA
®
UMI 800-521-0600
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
R eproduced w ith perm ission o f the copyright owner. Further reproduction prohibited w itho ut perm ission.
INTEG RATING MODELLING TECHNIQUES FO R FINANCIAL TIM E SERIES
A Thesis
Presented to
The Faculty of G raduate Studies
of
The University of Guelph
by
VICTOR KHOON-LEE TAN
In partial fulfilment of requirem ents
for th e degree of
Doctor of Philosophy
August, 1999
© V ictor Khoon-Lee Tan, 1999
National Library Bibliotheque nationale
1+1 of Canada du Canada
Acquisitions and Acquisitions et
Bibliographic Services sen/ices bibliographiques
395 Wellington Street 395, rue Wellington
Ottawa ON K1A0N4 Ottawa ON K1A0N4
Canada Canada
Your 1Ue Votre reference
OurSie Notre reference
The author has granted a non L’auteur a accorde une licence non
exclusive licence allowing the exclusive permettant a la
National Library o f Canada to Bibliotheque nationale du Canada de
reproduce, loan, distribute or sell reproduire, preter, distribuer ou
copies o f this thesis in microform, vendre des copies de cette these sous
paper or electronic formats. la forme de microfiche/film, de
reproduction sur papier ou sur format
electronique.
The author retains ownership o f the L’auteur conserve la propriete du

copyright in this thesis. Neither the droit d’auteur qui protege cette these.
thesis nor substantial extracts from it Ni la these ni des extraits substantiels
may be printed or otherwise de celle-ci ne doivent etre imprimes
reproduced without the author’s ou autrement reproduits sans son
permission. autorisation.
0-612-43275-0
Canada
R eproduced w ith perm ission o f the copyright owner. F urther reproduction prohibited w itho ut perm ission.
TAN
UNIVERSITY
W iUELPH Victor Khoon Lee 80302 (928 009 570)
GRADUATE PROGRAM SERVICES

Dept, o f Mathematics and Statistics PhD
The Examination. Committee has concluded that the thesis presented by the above-named candidate in partial
fulfilment o f the requirements for the degree
Doctor o f Philosophy
is worthy of acceptance and may now be formally submitted to the Dean o f Graduate Studies.
Title: In tegratin g Modelling Techniques for Financial Time S e r ie s
1.
Chair, Dqcjoral Examination Committee J . Mokariski
n.
External Examiner R* S^QC
in.
Graduate Faculty Member G. Hines
IV .
Advisory Committee p £-jm
v. Date:
Advisory Committee T. Stengos
Received by: __________________ Date: AU0 2 3 1999

for Dean o f Graduate Studies
rev. iv/96
ABSTRACT
INTEG RATING MODELLING TECHNIQUES FO R FINANCIAL TIM E SERIES
Victor Khoon-Lee Tan Supervisors:

University of Guelph, 1999 Professor P. T . Kim and A. F. Desmond
T he use of a related series of historical values as evidence or a guide line to predict

and to draw conclusions about future events is a practice common to reseaxchers of
m any disciplines. In order to achieve a better outcome, researchers are motivated to
seaxch for rules th a t can explain relationships between the past and future. However,
any forecast is affected by the unpredictable nature of future events as well as by the
lim itations of past data. Moreover, a forecasting model used by one field may not
necessarily be used by another, because there are problems which axe unique to each
field.
In standard statistical treatm ent of tim e series, tim e domain and frequency domain
are the classical techniques used as a basis for characterising an observed system and
forecasting its future behaviour. However, due to complicated d a ta structures of
tim e series, recent research developments-along with the availability of high-speed
com puters-have facilitated the development of m odem nonparam etric methods such
as the moving blocks bootstrap and the neural networks.
T he inadequacy of the efficient market hypothesis and the instability of d ata

structures found in financial market create serious challenges to the implementation
of modelling strategies for financial tim e series. However, a combination of both
classical techniques and m odem nonparam etric m ethods may offer a m ore effective
solution to financial forecasting problems.
This thesis will im plem ent a financial tim e series m odel which utilizes b o th clas
sical techniques and nonparam etric m ethods for estim ation and inference. It will
go on to propose a modified bivariate transfer function model. B oth tim e dom ain
and frequency dom ain techniques will be used in order to determ ine a relationship
between bivariate tim e series. After generating the less model de-pendent sam ples
using the m oving blocks bootstrap, a neural network will be applied to achieve b e tte r
point estim ation. Integrating these techniques provides th e best alternative solution
to financial forecasting problems. T he bivariate tim e series in question is interest
rate spreads and the spot Canadian dollar. Simulation results showed undercoverage
overall, but th e integrated modelling technique appears robust to th e choice of noise
distribution as well as th e sample size.
To: My Family
and
The Memory of My G randparents and Parents
ACKNOW LEDGEM ENTS
I would like to express my sincere appreciation to:
First and foremost, I would like to thank those professors on m y supervisory

com m ittee for their tim e and dedication. I am grateful to my supervisor Dr. Peter
Kim for his patience guidance, helpful suggestions and all the help since I came to
Guelph. I would also like to thank Dr. Tony Desmond for his constructive criticism
and suggestions. My appreciation is also extended to Dr. Thanasis Stengos and
Dr. Qi Li for the honour of having both of them on m y supervisory com m ittee. In
addition, I would like to thank the other professors who took tim e to answer any
questions I m ay have had.
I would like to thank th e D epartm ent of M athem atics and Statistics for all the
physical and financial supports. My thanks are also extended to Linda Allen, John
McCormick, departm ental stuff and m y fellow graduate students for th eir help and
friendship.
During my stay at Guelph, I am fortunate to have P eter Yong, Siew-Hong Tan,

Gary Rusland, Gary Cho, Dianqin Wang, and all others who axe associated with me
over the years as my friends. I would like to thank them for their friendship and for
making m y life more enjoyable.
A special thank to m y lovely girlfriend, Sandy Pui-Yee Li, and h er family for
encouragem ent, support and love. Finally, I would like to thank my fam ily for their
endless support in everything!
Contents
1 In tr o d u c tio n 1
1.1 In tro d u c tio n ...................................................................................................... 1
1.2 Scope of T he Thesis ...................................................................................... 5
2 T im e S eries M eth o d o lo g ies 7
2.1 Classical Tim e Series M e th o d s............................................................... 7
2.1.1 Spectral A n a ly s is ................................................................................. 8
2.1.2 Cross Spectra C oherency .................................................................... 12
2.1.3 Bivariate Transfer Function M o d e l .................................................. 16
2.2 N onparam etric M e t h o d s ................................................................................ 16
2.2.1 Moving Blocks Bootstrap S a m p le .................................................. 17
2.2.2 Modified Moving Blocks B o o t s t r a p ................................................. 18
ii
2.2.3 Neural N e t w o r k s ................................................................................. 20
2.2.4 Feed-forward Networks : M ultilayer P e rc e p tro n s ......................... 21
3 In te g r a te d M o d e llin g 28
3.1 Theoretical Foundations for Integrated Modelling..................................... 28
3.2 Modelling P ro c ed u re s....................................................................................... 34
3.2.1 Algorithm for Choosing the Block S i z e ......................................... 34
3.2.2 G enerating Bootstrap Bivariate Tim e Series.................................. 36
3.2.3 Neural Network E stim a tio n s................. 36
4 A p p lic a tio n to F in an cial T im e S eries 38
4.1 T he Interest R ate and Spot Canadian Dollar D a t a ................................... 38
4.2 Em pirical R e s u lts ............................................................................................. 47
5 S im u la tio n 50
5.1 Sim ulated Series I and Simulation R e s u l t s ............................................... 51
5.2 Sim ulated Series II and Simulation R e s u lts ............................................... 54
5.3 Sim ulated Series III and Simulation Results ............................................ 56
5.4 Sim ulated Series IV and Simulation Results ............................................ 58
iii
5.5 Sim ulated Series V and Simulation R e s u lts ............................................... 60
5.6 Sim ulated Series VT and Simulation Results ............................................ 62
5.7 Sim ulated Series VTI and Sim ulation R e s u lts ............................................ 64
5.8 D isc u ssio n .......................................................................................................... 66
6 C on clu sion s 69
6.1 F uture Research................................................................................................. 71
R eferen ces 73
A p p e n d ix 80
A. T he Back-Propagation A l g o r i th m ........................................................ 80
iv
List of Figures
2.1 A schem atic diagram of the moving blocks bootstrap for tim e series
(Modified from Efron and Tibshirani [21]). . ........................................ 19
2.2 Three-Layer Feedforward Network................................................................. 22
2.3 Threshold Functions for Neural Networks.................................................... 26
4.1 T im e Series Plots for Spot Canadian Dollar and Interest R ate Spread. 40
4.2 Changes in Spot Canadian Dollar (Yt) and Interest R ate Spread (A t). 41
4.3 Histograms of the Changes in Spot Canadian Dollar and Interest R ate
Spread D ata....................................................................................................... 43
4.4 Spectral Density Plots of Yt and X t Series................................................... 44
4.5 Cum ulative Spectrum Plots of Yt and X t Series......................................... 46
4.6 Plot of Squared Coherence Statistic.............................................................. 48
4.7 T he plot of standard error estim ates for all chosen block sizes. T he
true standard error is 0.02605 ..................................................................... 48
List of Tables
5.1 Results for Forecast Intervals on Sim ulated Series I ....................... 53
5.2 Results for Forecast Intervals on Sim ulated Series I I ............................... 56
5.3 Results for Forecast Intervals on Sim ulated Series I I I ............................ 58
5.4 Results for Forecast Intervals on Sim ulated Series I V .................... 60
5.5 Results for Forecast Intervals on Sim ulated Series V ....................... 62
5.6 Results for Forecast Intervals on Sim ulated Series V T .................... 64
5.7 Results for Forecast Intervals on Sim ulated Series V I I .................... 66
5.8 Results for 95% Forecast Intervals on Sim ulated Series ........................ 67
vi
Chapter 1
Introduction
1.1 Introduction
The use of a related series of historical values (or past d ata collected) as evidence
or a guide line to predict and to draw conclusions about future events, is a practice
common to reseaxchers of m any disciplines. In order to achieve a ‘b e tte r’ outcome,
researchers are m otivated to search for rules that can explain relationships between
the past and future. However, any forecast is affected by the unpredictable nature of
future events as well as by th e lim itations of past data. Moreover, a forecasting m odel
used by one field may not necessarily be used by another, because th ere are problems
which are unique to each field. However, an awareness of the characteristics of th e
field specific m odels m ay allow the application of a translational m odel to obtain an
adequate level of forecast accuracy.
In financial tim e series, the following two problems always exist. They axe:
1. T he inadequacy of the efficient market hypothesis; th a t is, market prices always
reflect the available information set. However, in the actual m arket, deviations
exist between true prices and market prices because not all traders have all
the existing information, and may not always act rationally. Moreover, “the
verification of th e efficient m arket hypothesis (EMH) is impossible” Granger
[27, p.12].
2. Instability of th e d ata structure. The system m ay be influenced by m ajo r factors
such as m onetary policies, economic indicators, and interest rates. In addition,
certain inform ation cannot be modelled statistically, which leads to an unstable
structure of past d ata sets.
These problems create serious challenges to th e im plem entation of modelling strate
gies for financial tim e series.
In standard statistical treatm ent of tim e series, the first step is to ‘look’ a t the
data, and then select suitable m athem atical models for the data. Before we can do
this, we must first study and understand the techniques which have been developed
for forecasting and drawing inferences from the series. Classical tim e series techniques
are used as a basis for characterising an observed system and forecasting its future
behaviour and, therefore, is the starting point for conventional statistical tim e series
m odel building.
In the time series literature, classical tim e series analysis is based on the theory of
stationary processes, and it has been dem onstrated that m ost stationary processes can
Reproduced with permission o f the copyright owner. Further reproduction prohibited without permission.
be approxim ated by a model drawn from the class of autoregressive moving average
(ARMA) models. In addition, some non-stationary processes m ay be represented by
an ARMA process after differencing. This is known as the autoregressive integrated
moving average (ARIMA) class of models, for which a model selection strateg y has
been developed by Box and Jenkins [8]. The significant contribution of Box and
Jenkins to tim e series analysis is based on the invertible ARMA class of processes and
th e four phases of: identification, estim ation, diagnostic checking, and forecasting.
P articularly in the identification phase, plots of the sample autocorrelation function
and the sam ple partial autocorrelation function can yield im portant inform ation in
identifying th e orders p and q of an autoregression moving average (A R M A (p,q))
m odel. Because of th e developments of Box and Jenkins [8], the problem of estim at
ing the ARMA (p, q) m odel has attracted considerable attention in the tim e series
literature.
The frequency dom ain analysis of tim e series has its own share in the tim e series
literature. Frequency domain analysis is m athem atically equivalent to tim e dom ain
analysis, which is based on the autocovariance function, but provides an alternative
way of viewing an observed system. In frequency domain analysis, “the spectral
density function describes how the variation in a tim e series m ay be accounted for by
cyclic components at different frequencies” Chatfield [10, p.7]. As a result, frequency
dom ain analysis provides many im portant insights which would not be apparent in
th e tim e dom ain analysis.
“T he spectral point of view is particularly advantageous in the analysis of mul
tivariate stationary processes and in th e analysis of very large d a ta sets, for which
num erical calculations can be perform ed rapidly using the fast Fourier transform” ,
Brockwell and Davis [9, p .112]. By combining the frequency dom ain technique and
the transfer function m odel-as proposed by Box and Jenkins [8] in the tim e d o m ain -
we are able to exploit th e relationship between any two financial tim e series.
Due to the com plicated d ata structures of financial markets such as stock returns,
exchange rates, and interest rates, there is a need to improve th e perform ance of
param eter estim ates in tim e series. In addition, th e behaviour of these financial
tim e series could perhaps be described more accurately by non-linear m athem atical
models. Such m ajor developments such as bootstrap techniques have been applied to
problems from time series analysis. In the tim e series literature th e use of bootstrap
techniques can be found in Efron and Tibshirani [19], Freedman [23], Kunsch [43] and
Liu and Singh [46].
Recently, artificial neural networks (A N N s) have formed a basis for an entirely dif
ferent non-linear approach to the analysis of tim e series. Together w ith the widespread
availability of powerful com puter hardw are and efficient training algorithms, neural
networks are now a subject of interest to professionals in almost every quantitative
field and there is increasing focus on th e potential benefits that neural networks can
offer. N eural networks have been applied to a wide variety of problems, such as p a tte rn
recognition, classification, process control, and prediction. Recent reviews from a
statistical perspective include: Cheng and T itterington [11], Ripley [51] and W hite
[55] [56]. Weigend an d Gershenfeld [54] provide a com puter scientist’s perspective of
neural networks in tim e series. In addition, Cybenko [13], Funanhashi [24], G allant
and W hite [25] [26], Hecht-Nielsen [31], and Hom ik, Stinchcombe and W hite [32]
[33] provide some im portant references on the existence properties of the feedforward
networks.
1.2 Scope of The Thesis
T he m ain objective of this thesis is to model financial tim e series. After a review
of classical tim e series approaches to tim e domain and frequency dom ain approaches,
the thesis will examine the moving blocks bootstrap and neural networks before going
on to integrate these techniques to form an integrated modelling technique for th e
analysis of financial tim e series.
C hapter 2 will discuss classical and m odem nonparam etric methodologies to tim e
series. In Section 2.1, tim e and frequency domain techniques are reviewed as these
techniques are used (in a later section of the thesis) to determ ine the existence
of a relationship between financial tim e series, and to develop a bivariate transfer
function model. In Section 2.2, nonparam etric m ethods-the moving blocks bootstrap
and neural netw orks-are discussed in detail. These methods, together with classical
techniques, provide the theoretical basis for the derivation of an integrated technique.
In C hapter 3, an integrated modelling technique will be developed to analyse
financial tim e series. T he proposed m ethod offers an improved solution to financial
forecasting problems.
Chapter 4 presents an application of the integrated modelling technique to fi
nancial time series, interest ra te spreads and spot Canadian dollar d a ta sets. The
empirical results support the use of the integrated modelling technique to financial
tim e series.
T he integrated modelling technique will be applied to sim ulated tim e series data
in Chapter 5, which will also summarise and present the results of the sim ulation.
T he simulation results show undercoverage overall, but the integrated modelling
technique appears robust to th e choice of noise distribution as well as th e sample
size. Conclusions and suggestions for future work will be presented in C h ap ter 6.
Chapter 2
Time Series Methodologies
T im e series analysis is an im portant technique in a wide variety of quantitative
fields. Over th e years, a num ber of quantitative techniques and nonparam etric
m ethods have been developed for analysing and predicting tim e series values (essential
references can be found in Chatfield [10]).
2.1 Classical Time Series Methods
To simplify the review of classical m ethods, we will not review tim e dom ain
techniques here. A dequate references about the autoregressive integrated m oving
average (ARIM A) models, known as the Box and Jenkins approach, can be found
in studies by Brockwell and Davis [9], Chatfield [10], and Priestley [48], which pro
vide readable introductions to tim e series analysis and cover both tim e domain an d
frequency domain techniques. In this section, a brief review of frequency dom ain
techniques is presented.
2.1.1 Spectral Analysis
“Spectral analysis is essentially a modification of Fourier analysis so as to m ake
it suitable for stochastic rather th an determ inistic functions of tim e” Chatfield [10,
p .105]. Hence, spectral analysis of stationary tim e series involves decomposing th e
given tim e series into a linear com bination of sinusoids.
For a given tim e series, Z t, t = 0, ± 1, ± 2 , . . . , w ith E (Z t ) — 0, the frequencies
k (n ~~ 1)
Vk .......... [?]■
are called th e Fourier frequencies of Z t and [•] denotes the greatest integer function.
Notice th a t k contains n integers. For example, a t frequency k, a waveform is th e
sinusoidal function
cos(2xi/fct) +- sin(27rif^t) , t = 1 ,... , n (2-2)
which constitutes cycles over t = 1 , . . . ,n , where k = —[^n^ ], ■■• ,[y]. Since th e
length of tim e required to complete one cycle is called the period, the frequency i/*
measures the num ber of cycles per tim e point t. Note that, by knowing th e Fourier
frequencies, we can capture the tim e p ath of Zt as a weighted sum of the sinusoids in
(2.2).
In order to estim ate the contribution of each sinusoid (2.2), we estim ate th e
autocovariances; y z (k ) = C<rv(Zt+k, Z t ), by
1
= - E (z ‘+w - ^ - £) <2-3 )
n «=i
for A: = —rc + I , . . . , n — 1 and Z = %t)- Note th at cz{0) is an estim ate of th e
variance of Zt . Following this we define the periodogram
n —1
In.{vk) = cz(0) + 2 5 2 cz(&) cos(2iri/kk) ■ (2.4)
fc=i
T he periodogram appears to be a natu ral way of estim ating the spectral density
function (or the power spectrum)
fz { y ) = 7z(0) + 2 ^ 7z(fc) cos(2irku) (2.5)

fc=i
where
lz (k ) = I 2 f z ( v ) cos{2rkv)du .
Moreover, the to ta l area under the periodogram is an estim ate of the variance of Z t ]
th a t is,
t~]
£ (Z t - Z f = £ / n( ^ ) . (2.6)
Although, th e periodogram is asym ptotically unbiased; th a t is
E [-^ M ] fz { v ) , asn-+oo (2.7)
the variance of I n(vk) is not zero as n —►oo. Hence I n(i/k) is not a consistent estim ator
for f z {u). Thus, in order to restore consistency, we will consider an alternative
procedure of estim ating a power spectrum -that is, the m ethod of sm oothing the
periodogram. T his procedure, which provides a consistent estim ate of th e spectral
density function, is called a lag window estimator, defined by
fz{u ) = cz {0) + 2 Y ^ w ( j ^ J cz(k)cos(2-irki/) , v € [ - |, , (2.8)
where the lag window, iu(-), is a weight function on [0,1] w ith 0 < w (x) < io(0) = 1,
and a sequence, m , is chosen such th a t ^ - » o o a s n - + oo. This m ethod elim inates
the inconsistency of the periodogram w ithout destroying its asym ptotic unbiasedness.
Moreover, it allows us to examine the sources of variability in the total variation of a
given tim e series. However, the choice of m is im portant because “the larger th e value
of m the smaller will be the variance of the resulting estim ate but the larger will be
the bias, and if m is too large th en interesting features of f z { y ) , such as peaks, may
be sm oothed out” Chatfield [10, p .117]. Since there is no effective m ethod for m aking
the choice of m , we can try several values in the region of ^ according to Chatfield
[10]. The most basic lag window estim ator is th e uniformly weighted average. In this
thesis, the modified Daniell sm oothing function, which is defined below, will be used
10
for d a ta analysis
X [I- + l i E ^ l ) _ 1I + | . x [ i + f c l i l , } , (2.9)
where span ( = 2 m -+-1) is an integer giving the length of th e smoothing window for
the given series. Note th a t u;(x,-) in (2.9) is the com putational expression of w ( ^ ) in
(2.20). Moreover, th e sm oothing window or average only considers the m neighboring
values an d x,- in sm oothing x,-; the sm oothing window actually includes span points
with half-weights on th e ends.
If Z t is a generic white noise series which is defined to be an uncorrelated series,
then we have j z ( 0 ) = <7 ^, and 7 z (k ) = 0 Vfc = ± 1 , ± 2 , Moreover, we have a flat
spectral density because f z ( v ) = 0 %for v € [— j ] . In addition, we have F z{v) = v
for all v E [0 , j] hence define the normalised spectral measure by
(2.10)
fo fz { v )d v
By defining the sam ple version of equation (2.10) in the following way
E
^ ) =& r ‘'it
"€N ' (2 -u )
we can use Un(is) to test whether or not a tim e series is w hite noise. T h at is, we
plot th e Un{y) function and check its com patibility with the diagonal line (uniform
11
distribution, function; F (x) = x , 0 < x < 1) using the Kolmogorov Smirnov test.
Moreover, for n > 62 (see Brockwell and Davis [9]), a good approxim ation to a level
a Kolmogorov Sm irnov test is to reject th e null hypothesis if the statistic falls outside
of the bounds.
2.1.2 Cross Spectra Coherency
Thus fax, we have been concerned with analysing a univariate tim e series. We will
now exam ine th e relationship between two tim e series. For two given non-white noise
tim e series, Yt and X t , we use the cross spectrum, to establish whether a stru ctu ral
relationship exists between them . The cross spectral density is defined as
fv x M = £ lY x W e -* ™ * , „ € f - i |] , (2.12)
fc = —oo
where
7 r x { k ) = Cov(Yt+k, X t ) , k = 0 , ± 1, ± 2 , . . . (2.13)
is called the cross covariance and its inverse representation is defined as
TrxW = f \ / r e W e " , fc = 0 , ± l , ± 2 ......... (2.14)
12
Note th a t f r x ( v ) is a complex valued function. The reason is th at y y x (k ) (2.13) is
not an even function; th at is, ~fYx(k) ^ ~fYX (—k).
T he cross-coherence function between two series is defined by
A 'O ) = frxM r _ 1 11 ( 2.15)

[/x M A 'O ')]’ L 2 ’ 2J ’
and the Cauchy - Schwarz inequality ensures that squared coherence, K 2(v), is botinded
as follows:
0 < A 'V ) < 1 , *e [-i i]. (2.16)
Here K 2(u) can be interpreted as the spectral equivalent of the ordinary linear
regression R 2 statistic. Note th at (2.16) measures the extent to which Yt a n d X t
series axe correlated at each frequency. Two situations of particular interest arise:
C a se I: W hen Yt series is an exact linear function of X t series, we have
fy (i/) = \A{u)\2f x {v) and fx v { v ) = A { v )fx {v), v € , (2.17)
where
00 r 1 11
A (u) = £ (2.18)
OO W
is called the frequency response function.
13
C a s e I I : W hen no correlation between Yt and X t series exists, we have
fx v M = 0 v „ 6 [ _ i, i] . (2.19)
Therefore, by th e definition of squared coherence, we have
\K {v)\2 = 1, v € i |J , for case I
and
\K (u )\2 = 0 , v € J-^ , , for case I I .
These axe the two extrem es th a t can happen to th e Yt and X t series. However, in
case II, no further analysis is required, as no relationship between Yt and X t series
exists.
To test whether or not the Yt and X t series are correlated, we use
f z ( r ) = cz ( 0 ) + 2 E w y Y j c z ( k ) cos(27rfci/) , v e [“ , 511 , (2-20)
as well as
fr x M = E crX ( * : ) u . ^ e - :,”i- ‘ , v 6 [ - i , |] , ( 2 .21 )
where w(-) is a weight function supported on th e interval [0 , 1] w ith 0 < w (x ) <
14
tu( 0 ) = 1 , and m is chosen to be a sequence such th a t - — > oo as n - t o o w ith
n —k
£ £ ( Yt+k- Y ) ( X t - X ) for 0 < fc< n -l,

t=i
J- E (Yt+k- Y ) ( X t - X ) for —re -f- 1 < k < 0 ,

t= -k + 1
and
K 2 ( 2 . 22)
- [ 4 -3 -
Asymptotically, we can test the hypothesis : K 2(i/) = 0 , v £ , and
th e alternative hypothesis H a : K 2{v) > 0 using the statistic
S - * ! * * .™ . (2.23)
l-I^ J I2 L 2 ’ 2j
Since K 2(v) is distributed as the square of a m ultiple correlation coefficient, so th at
S ~ F2Am under the hypothesis that K 2{u) = 0, we can therefore reject the hypothesis
K 2{v) = 0 if S > i 7].-a, 2, 4m, where F i_at 2, 4m is the (1 — a) quantile of th e F
distribution w ith 2 and 4m degrees of freedom. As we m entioned earlier, th e choice
of m can be im portant; it can also affect the F test as dem onstrated here.
15
2.1.3 Bivariate Transfer Function Model
After establishing the existence of a structural relationship between the X t and
Yt series, we have the following m o d el:
1712
Yt = £ a sX t. s + N t , (2.24)
5=771 X
where m i < m 2 and N t is a random noise which is assum ed to be uncorrelated with
X t . Here, m i and m 2 have to be estim ated. Note th a t if m i > 0, we assume the past
X t will influence future Yt , but not vice versa. Thus, in the former case the model
given in (2.24) would be called causal.
Having considered the model, we follow the procedures in Brockwell and Davis
[9, pp.455-457] to estim ate a a, s = m i , . . . , m 2 and possibly the Nt process in model
(2.24). The procedures will be presented in C hapter 3.
2.2 Nonparametric Methods
In this section, nonparam etric m ethods will be discussed. Bootstrap techniques
and neural networks are computer based (nonparam etric) techniques. Although
the two techniques use different com puter algorithms, they provide an alternative
approach to m odern d a ta analysis and are particularly useful in the analysis of
complicated d a ta structures.
16
2.2.1 Moving Blocks Bootstrap Sample
The bootstrap is a com puter based resampling technique (as initiated by Efron
[15]) used to gain information on the distribution of an estim ator. A m ore general
resam pling technique, called th e moving blocks bootstrap (see Kunsch [43], Liu and
Singh [46], and Leger, Politis and Romano [44], and Efron and Tibshirani [20]), is used
for m odel dependent data. T he procedure for obtaining th e moving blocks bootstrap
tim e series samples will be discussed. In the case where n ^ k ■I (where k is th e
num ber of blocks and I is th e block size), it is proposed th a t a modified moving
blocks bootstrap be used to generate the bootstrap tim e series.
Simple bootstrap resam pling involves fitting a model and then sampling from th e
residuals. An alternative m ethod is the moving blocks bootstrap, for bootstrapping
tim e series: For any given tim e series sample, we can generate a bootstrap tim e series.
In order to do so we m ust (1) choose a block length; (2) form the tim e points to all
possible contiguous blocks of chosen length; (3) draw th e samples with replacem ent
from these blocks; and (4) join all th e samples together to form a bootstrap tim e
series. These steps can be applied to generate a bivariate bootstrap tim e series-that
is, we pair-up the original bivariate tim e series and then apply the four steps to th e
series. For exam ple, if we have samples of size n, and the block length is chosen to be
/, then we can have k blocks to form the bootstrap tim e series. The num ber of blocks,
k , is chosen so th a t n « k - I. Figure 2.1 shows th e bootstrap samples for a bivariate
17
tim e series with, block length I = 3. ‘‘T he idea for the moving blocks bootstrap is to
choose a block size I large enough that observations m ore th an I time u n its apart will
be nearly independent” Efron and Tibshirani [20, p.102]. However, Kunsch [43] and
Liu and Singh [46] show th a t B{ = (X{, X i+ i, . . . ,X,-+r_i), i = 1,2,— , k, are i.i.d .
for fixed I under some conditions. Note th a t the moving blocks bootstrap coincides
w ith the classical i.i.d. bootstrap of Efron [15] when 1 = 1. The advantage of using
th e moving blocks bootstrap is th at we will not destroy the correlation th a t we axe
try ing to capture since we axe resampling from the blocks of observations. Next, we
propose a modified moving blocks bootstrap when k * I ^ re.
2.2.2 Modified Moving Blocks Bootstrap
We note th a t if k • I ^ n ; th a t is, when n is not exactly divisible by /, then Efron
and Tibshirani [20] suggest th a t we have to m ultiply the bootstrap stan d ard errors
by adjust for the difference in lengths of the series. We propose an additional
step to the moving blocks bootstrap procedure; th at is, the addition of a random
block w ith th e block size V = n — (k • I) to the bootstrap tim e series, in order to
resolve this adjustm ent problem .
18
Original Tim e Series Sam ple
Figure 2.1: A schematic diagram of the moving blocks bootstrap for tim e
(Modified from Efron and Tibshirani [21]).
19
2.2.3 Neural Networks
Artificial neural networks (ANNs) axe m athem atical m odels for hum an brain
activities. In general, neural networks are often applied to applications such as
classification, pattern learning, and prediction.
Neural networks are considered a nonparam etric m ethod for drawing statistical
inferences. The im portant aspect of neural networks is their capability to construct
nonlinear relationship between th e given input and output d ata. W hen used together
w ith powerful com puter hardware and training algorithms, ANNs allow statistical
inference to be m ade without any structural formula. This implies th at nonlinear
tim e series are more suitable for analysis by neural networks. Because tim e series are
statistically more volatile in financial markets th an in other applications, the financial
industry has become a prime application area for neural networks.
Although there are a numerous of networks of which m any are recurrent neural
networks (e.g. Kuan and Liu [42]), the multilayer perceptrons is the sim plest neural
network representation, and it can well approxim ate a large class of functions. Hence,
in this thesis, the feed-forward networks are used as the only nonparam etric model.
Authors such as Cybenko [13], Funanhashi [24], Gallant and W hite [25] [26], Hecht-
Nielsen [31], and Homik, Stinchcombe and W hite [32] [33] provide some im portant
references on the existence properties of the feedforward networks. Recent reviews
from a statistical perspective include Cheng and T itterington [11], Ripley [51], and
W hite [55] [56]. Weigend and Gershenfeld [54] present a com puter scientist’s perspec
tive of Neural networks in tim e series.
20
2.2.4 Feed-forward Networks : Multilayer Perceptrons
Feed-forward networks are an im portant class of neural networks. Typically, the
network consists of a set of source nodes that constitute the input layer of one or
m ore hidden layers of com putation nodes, and an output layer of com putation nodes.
T he input signal propagates through th e network in a forward direction, on a Iayer-
by-layer basis. These feed-forward networks are com m only referred to as mulitlayer
perceptrons (M LP) when there is a t least one hidden layer. From th e statistical point
of view, the m ultilayer perceptron networks provide nonlinear regression functions
th a t are estim ated by optimising some measure of fit to the training data. Figure 2.2
shows a typical m ultilayer perceptron with only one hidden layer, flowing from the
bottom up.
T he m ultilayer perceptron netw ork is a supervised neural network since, during
training, each input d a ta vector is paired with a corresponding desired taxget for
th e network o u tp u t. T he actual outputs in the o u tp u t layer axe now compaxed with
th eir desired targets, the difference being the o u tp u t errors. T he errors axe th en
fed back to th e network to improve the performance. This process is known as the
backpropagation algorithm. This algorithm is based on the mean squared error (MSE)
as an error-correction learning rule. Basically, th e back-propagation process consists
of two passes: th e forward pass a n d backward pass. In the forwaxd pass, the input
d a ta vectors axe applied to the source nodes and propagate, layer by layer, through
21
O utput layer
H idden layer
Input layer
Figure 2.2: Three-Layer Feedforward Network.
th e network. Once one forward pass is completed, a set of outputs is obtained as the
actual o u tp u t of the network. Note that the synaptic weights (which are th e inter
node connection weights of the network) are fixed during the forward pass. During
th e backward pass, the synaptic weights of th e network are all adjusted according
to the error-correction rule. A n error signal is then obtained by subtracting the
actual output of the network from the desired target. This updated error signal is
th en propagated backward through the network; th at is, in the opposite direction of
synaptic connections. The synaptic weights are adjusted to improve the perform ance
until a zero, or m inim um, error signal is obtained. T he learning process perform ed
by this algorithm is called backpropagation learning. A com plete presentation cycle,
or iteration, comprises the forward pass and backward pass cycles. One com plete
presentation of th e entire training set is called an epoch. T he learning process is
22
m aintained on an epoch by epoch basis, until the synaptic weights and the threshold
levels of the network stabilise and the E m se over th e entire training set converges to
some m inim um value. The E m se is defined by
r, U s Ho{Target - O utput )2 fet

£m” = JW , ’ ( ’
where N s is the num ber of training vectors in one complete presentation of the entire
training set during th e learning process (an epoch), and N0 is th e num ber of output
neurons.
It is advisable to randomise the order of presentation of training sets from one
epoch to th e next in order to make the search in weight space stochastic over the
learning cycles, and avoid the possibility of limit cycles in the evolution of the synaptic
weight vectors. Moreover, the randomisation of th e training sets allows for a testing
technique th a t reflects the practical utilisation of the network in actual forecasting
problems.
T he normalised mean squared error (NMSE) is used to evaluate the model of the
network, and is defined by
Y K (Y — Y ) 2
N M SE = * vT - •1 , (2.26)
where K is the num ber of d a ta in the validation set, Ys is the desired target, and Ys
is the actual output. Note th a t the quantity ( 1— NMSE) measures the percentage
variance in the d a ta which is explained by the model. Moreover, if we plot the learning
23
curve (the NMSE versus the epoch) for th e function estim ated from the train in g set,
we can easily see th a t the learning curve is m onotonicaily decreasing.
T he essence of backpropagation learning is to encode an input-output relation. A
m ultilayer perceptron will be considered well train ed if it learns enough abo u t th e past
to be able to generalize about the future. However, if we allow the learning process
to proceed indefinitely, we will overfit the d a ta and begin modelling only noise-that
is, the NMSE reduces to zero. T he cause o f this problem is understandable if we
consider the linear regression case: if we constantly add param eters to its full rank,
the R-square becomes 1. Hence, we need a stopping criteria in order to end the
learning process when the generalization perform ance is adequate.
Kim, M artin and Staley [39] identify the technique most commonly used to end
th e learning process: First, the available d a ta set (or the moving blocks bootstrap
tim e series samples) is randomly partitioned into a training set and a validation set
(there are no fixed rules for selecting the am ount of data for either the train in g set or
th e validation set). We propose to select 75% o f the data to be used for th e training
set, and 25% of the d a ta to be used for the validation set. T he m otivation here is to
validate the model on a data set which is different from the one used for param eter
estim ation. In this way, we can com pute th e NMSE (2.26) based on a validation
set using the resulting network used on the training set. Consequently, we plot the
learning curve of th e validation set, and look for th e epoch a t th e point on th e curve
where th e direction starts to change. Note th a t the learning curve of th e validation
set will not necessarily monotonicaily decrease. This number of epochs will then be
used to determ ine th e stopping point for the learning process. It should be n oted th at
24
this situation does not always occur, b u t if it does happen, th e network should be
re-trained on the validation set. Finally, th e configuration of the resulting m ultilayer
perceptron network on th e given training set and validation set is used for prediction
or forecasting on the given test s e t
In neural network architecture, Figure 2.3 shows the commonly used threshold
functions. The threshold function is used in the forward pass com putation. Using
th e architecture of Figure 2.2 as an exam ple, we process input X ’s to calculate new
hidden layer itfl neuron using
= + (2-27)
h=1
where / ( - ) is the threshold function, i/hi is the weight between the input layer an d hid
den layer, and is the hidden layer threshold values. Note th at th e backpropagation
algorithm is presented in Appendix A.
R e m a r k s : A m ultilayer perceptron has the following distinctive characteristics :
1. “T he model of each neuron in the network includes a nonlinearity at th e o u tp u t
end. The im portant point of em phasis here is th a t the nonlinearity is sm ooth;
th a t is, differentiable everywhere” Haykin [30, p .138]. In the prediction context,
a nonlinear function th at satisfies this requirement is a sym m etric sigmoidal
25
(c)topta
Figure 2.3: Threshold Functions for Neural Networks.
26
function defined by the logistic function :
Y = , ■, (2.28)
1 + e x p ( —x)
f = (l - Y ) Y , (2.29)
where f is the derivative of the logistic function. Note th a t the logistic function
is a commonly used function which transform s the input value in the range x 6
[—oo, + oo] into the output range O € [0, 1]. “The presence of nonlinearities is
im portant because, otherwise, the input - output relation of the netw ork could
be reduced to th a t of a single-layer perceptron” Haykin [30, p.139].
2. For the choice of the maximum number of layers in th e network, Lippm ann [45]
recommends th a t three layer networks, w ith just one hidden layer, are sufficient
to solve arbitrarily complex input-output mappings.
3. The hidden neurons which are not part of the input or output of th e network
enable the network to learn complicated d a ta by extracting more information
or features from the input vectors.
4. The network is connected between the layers by the synapses of th e network.
Hence, a change in the connectivity of the network requires a change in synaptic
weights of the network.
27
Chapter 3
Integrated Modelling
In this chapter, a strategy called integrated modelling will be im plem ented in
order to analyse tim e series based on financial d a ta . Since th e m odelling tool is
a com bination of th e methodologies described in th e previous chapters, the term
integrated modelling is used. The objective for implementing this technique is to
achieve a more reliable solution to th e modelling and forecasting problems which
occur in the financial world. The technique utilizes both classical techniques and
nonparam etric m ethods for estim ation and inferences. This chapter will begin w ith
an exam ination of th e theoretical foundations for integrated modelling.
3.1 Theoretical Foundations for Integrated Modelling.
In C hapter 2, th e theoretical background for th e derivation of classical techniques
an d nonparam etric m ethods is discussed. These techniques and m ethods form th e
basis of im plem enting techniques in modelling tim e series. In this section, with th e
28
support of these techniques and m ethods, we present the theoretical foundations in
detail for the derivation of the integrated m odelling technique.
An initial approach is to follow th e m ethod of Kim and Maxtin [40], using spectral
analysis to uncover certain statistical details. For a given bivariate tim e series, the
sta tistic Un(v) (see C hapter 2 ) is used to determ ine whether the two tim e series
axe w hite noise. T hat is, the Kolmogorov Sm irnov test is used to reject th e null
hypothesis w ith level a. if the function Un(u) falls outside of th e bounds for each tim e
series. Once we reject the null hypothesis and conclude that the two given tim e series
are not w hite noise, the cross spectrum is applied to determ ine the existence of the
relationship between the two series.
A fter establishing the existence of a structural relationship between the two series,
th e proposed m odel (which is modified from th e bivariate transfer function model) is
applied to furth er analyse the tim e series. Before implementing th e proposed model,
it is helpful to provide the rationale for m odifying the bivariate transfer function
m odel. T he bivariate transfer function model is restated below:
7Jl2
y t = £ a *x > - + N ‘ • (3-i)
Recall th a t (3.1) is valid only if it is causal, th a t is, the past X t will influence future
Yt , b u t not vice versa. In particular, the m odel requires th at m x < m 2 and m i > 0 .
N ext, we follow the procedures listed in Brockwell and Davis [9, pp.455-457] in order
to estim ate m x, m 2, and a s, s = m x, . . . , m 2- T he following steps are applied to
o btain m x and m 2 '.
29
S te p 1. We calculate the cross correlation function (C C F ); f>wtY‘ {h) of the two
filtered series; nam ely W t and Y ' . The filtered series Wt and Y ' are obtained
by applying the prewhitening techniques to X t and Yt series, respectively. This
is achieved by transforming the X t series to a white noise series by
i(B )0 ~ l (B )X , = Wt , (3.2)
and applying the same transformation; th at is, d>(B)0~l ( B ), to th e Yt series to
obtain
Yt" = 4>{B)d~\B)Yt . (3.3)
S te p 2. We perform ed an asymptotic test of no cross correlation at the 5% level
using the bounds ±1.96n~?. T he comparison of pwtYc'(h ) with the bounds
±1.96n- 2 gives an indication of the lags h a t which pwtYtm(h) is significantly
different from zero.
S te p 3. From the previous step, we estim ate rhi and m 2 to be the m inim um and
m aximum of lags h at which pwcYc'(h ) ^ 0, respectively.
30
To continue the process of estim ating ocs, s = m l7 — , m 2 for th e bivariate transfer
function model, we apply the filter <j>(B)0~l (B ) to th e model (3.1), which yields
7712
V,- = £ a ,W t- , + n ; . (3.4)
s=mi
T his implies th a t the transfer function between X t and Yt series isthe sam e as
th a t between W tand Y ’ . By m ultiplying both sidesof (3.4) by W t-h , and taking
expectation, we obtain
E {W t. k Y?) = {a 0E (W t. kWt) + <xxE (W t- kWt- i ) + • ••} + E {W t. kN ;) . (3.5)
Since
E (W t- kY~) = iw trc<k)
E (W t- kWt-,-) = 0 for j / 1
an d E (W t- kN ') = 0 , V k (because X t and N t axe uncorrelated). Therefore,
IW tY'iE ) = a k&wt ( 3 -6 )
31
and
lw tYt-
CCk = ------ -
C\V?
_ gyt‘ l w tY?{k)
<*Wt G Y ’ GWt
= - ^ - p W tY 'ik ) , fc = 0,1, 2, . . .
crwt
The estim ates of a for lag h at which pwtYfi}1) is found to be significantly different
from zero axe
O'h = PWtYS - ' 5 h — ™h ••• T7712» (3*7)

* <?wt
Now th at we understand th e frame work of th e bivariate transfer function m odel, we
present our proposed m odel for bivariate tim e series.
The first step in integrated modelling is a modification of (3.1) as follows:
p
E
Yt = £ w s( X t- s +
s=0
Pt-s) + Nt , (3.8)
where N t is random noise assumed to be uncorrelated w ith X t, and Pt is defined by
K if X t- 1 > X t-2
0 if X t—i = X t —2
X ifX t-i< X t-2
where P ’s, k, and A axe unknown param eters. Note th a t we fix th e term p to b e 4
in ( 3 .8 ) for daily closing d ata and, for simplicity we let X t — X t + Pt- In addition,
the coefficient w3 will be estim ated by using the maximum likelihood estimate (M LE).
32
Furtherm ore, we introduce two constant term s, namely k and A, in (3.8). The reason
for this is th a t we believe th a t the increase (decrease) in previous tim e points X t- i
will result in an increase (decrease) in present tim e point X t. The estim ates of k and
A are defined as follow:
k = average of —x t- \\X t > ^ t - i ) } ,
and
A = average of \ ^ 2 ( X t — X t- i \ X t < X t- i)
U=2
So far, we have only used the classical techniques w ith modifications. The next
step in integrated modelling is to incorporate the nonparam etric m ethods. We gen
erate bootstrap tim e series samples using th e moving blocks bootstrap m ethod for
drawing samples. It is im portant to note th a t we do not apply the moving blocks
bootstrap m ethod to the residuals. The reason for this is th at the bootstrapping of
residuals m ethod is “model dependent”, th a t is, the model, for exam ple an A R ( 1)
model, is fit to the original tim e series.
Next, we apply the multilayer perceptron network to the original bivariate tim e
series and th e bootstrap generated bivariate tim e series samples for estim ation and in
ference. These applications complete the derivation of the theoretical foundations for
integrated m odelling where the procedures of the nonparam etric methods introduced
in C hapter 2 are followed closely.
33
3.2 Modelling Procedures
After deriving the theoretical foundations for integrating a model, th e modelling
procedure is used to analyse bivariate financial tim e series.
T he correct approach for analysing bivariate financial tim e series is to use spectral
analysis in order to uncover certain statistical details. In particular, we need to
determ ine the existence of the relationship between the two given tim e series. Lf it is
determ ined th at a relationship exists, we continue to apply th e integrated modelling
technique to fu rth er the analysis of the bivariate tim e series. Having defined the
m odel structure, the moving blocks bootstrap is applied to th e given tim e series.
However, we need to know the block size I in order to generate the bootstrap tim e
series samples. Since there is no effective way of choosing the block size a t this
m om ent, we suggest an algorithm for choosing an appropriate block size.
3.2.1 Algorithm for Choosing the Block Size
S te p 1. We fit an ARMA(p,qr) model to the stationary tim e series. We use th e first
coefficient and its associated standard error as tru e values. For exam ple, we
have the AR(1) model for the given series zt , t = 1, 2 , . . . , n;
zt = f i z t - i + e .
We now choose $ and ap as our true values.
34
S te p 2 . Here we fix the block size to be I = 1,5,10, - -. , 100, and select B = 200
(W here I < ^ ) . For each I, we generate th e independent bootstrap tim e series
•?j6, b = 1 ,2 ,... , B , each consisting of n d a ta values draw n with replacement
from zt.
S te p 3. For each I, we fit the ARMA(p,^) m odel to each bootstrap tim e series
= 1 ,2 ,... , B . In th e exam ple given, we fit an AR(1) m odel to each
generated series an d obtain th e bootstrap coefficients of /?”(&), b = 1 , 2 , . . . , B .
S te p 4. For each I, we estim ate th e standard error <tf(P) by the sample standard
deviation of th e B replications
1_
= (e - £ '(-)]2 /[JB —1]}’ .
where
6=1
S te p 5. Here we plot the standard errors obtained from Step (4) versus th e block
size I. Finally, we choose th e block size I such th at the standard erro r of I is
closest to the tru e estim ate obtained in Step ( 1 ).
35
3.2.2 Generating Bootstrap Bivariate Time Series.
Once the block size is chosen, th e bootstrap tim e series samples axe generated by-
using th e following steps:
S te p 1 . Use the algorithm described above to select an appropriate block size L
S te p 2 . Select B independent bootstrap samples (X*1, l^*1), (X*2, Y ' 2), - - * , ( X “B, YtmB),
each consisting of n data values drawn w ith replacement from ( X t, Yt)- Here we
use the moving blocks bootstrap procedures to generate th e bootstrap sample
pairs. Note th a t we need to have B > 1000 for constructing the intervals (See
Tibshirani [53]).
3.2.3 Neural Network Estimations
A fter obtaining th e moving blocks bootstrap tim e series for each set of bootstrap
sam ple pairs (X "b, Ytmb), b = 1 , 2 ,... , B , we construct the m ultilayer perceptron net
work w ith 4 input layers, n hidden units for th e hidden layer (where n is determ ined
by trial and error in th e region of | of input vectors), and 1 o utput layer. The three-
layer perceptron network with th e backpropagation algorithm is now used to train on
the input-output pair; th a t is, ( { X t- 4 , X t- 3 , X t - 2 , X t- i) , Yt ) ,t = 1 ,2 ,... ,n , where
X - 3 , X - 2 , X - i and X 0 axe set to zero. Since the underlying (2.28) activation function
of th e network is logistic, it would be better to choose the d a ta th at fits a range
36
th a t does not “saturate” the network neurons. Thus, we have th e range of th e inputs
X t, t = 1 ,2 ,... , n to be X t € [—1, -hi], and th e range of the o utput Yt,t = 1 , 2 ,... ,n
to be Yt 6 [0 , 1] (i.e., Sigmoid range).
Finally, we can construct the bootstrap forecast intervals for the financial tim e
series. For example, if we want to have a forecast interval for one step ahead future
value, Yn+i, then the procedure is as follows:
S te p 1 . Use the neural network to obtain th e forecasted values; Y ~ ^ , b = 1 , 2 ,... , B .
S te p 2. We now construct the forecast intervals for Yn+x. Letting
D‘ = Y n+ i - 6 = 1, 2, . . . ,2?,
where Yn+i is the forecasted future value obtained from th e original tim e series,
we approxim ate the bootstrap distribution of D ' by its empirical distribution.
S te p 3. Let u'a be the a quantile of the bootstrap distribution of D~. Then a
bootstrap forecast interval is given by
[^ n + l Wl —a> ^n+1 U a •
N ote th at this is the bootstrap percentiles m ethod applied to prediction. Other
m ethods of bootstrap confidence intervals can be found in Efron and Tibshirani
[20].
37
Chapter 4
Application to Financial Time Series
Based on. the theoretical foundations discussed in C hapter 2, we derive th e inte
grated m odel presented in C hapter 3. This chapter will apply th e integrated model
to a given set of data, nam ely the interest rate spread and th e spot C anadian dollar
series, in order to exam ine how the integrated m odel will perform .
4.1 The Interest Rate and Spot Canadian Dollar Data
T h e interest rate spread d a ta (I S t) and the spot Canadian dollar d ata (C t) are
taken from Kim and M artin [40]. A brief description of the d a ta set is given below.
A to ta l of 1476 daily trading data, commencing January 1991 through to August
1996, axe used in this study. T he interest rate spread IS t (in percent) is defined as
the difference between 90 day Canadian and 90 day US treasury bill rates, and the
38
spot Canadian, dollar Ct is defined as th e closing price of th e exchange rate measured
in US cents. T he tim e plots of Ct and I S t axe presented in Figure 4.1.
N ext, we define the changes in the interest rate spread I S t by
X£ = IS t ~ I S t-i , (4.1)
and define th e changes in log Ct (or R etu rn R t ) by
Yt = log Ct - log Ct_ i . (4.2)
Note th a t th e tim e unit for Yt is one day. T he time plots of Yt and X t series axe
presented in Figure 4.2A and 4.2C. It is apparent that th e two series appear to be
random quantities over time.
In Figure 4.2C however, the X t series exhibits extrem e movements, while the Yt
series (Figure 4.2A) is bounded within a sm aller range over th e tim e period. Moreover,
since we axe trying to identify the “stable” model for the long term link between th e
X t and Yt series, we will remove these extrem e points and replace them as follows:
±ct if X t > ±<z

—a if X t < —a
+b if r £ > + 6
—b if Yt < —b ,
where a — 4 • crxt and 6 = 4 - oyt.
In particular, we will bound the m ovem ent in X t between ±38 ( = ± 4 • &xt) basis
39
Figure 4.1: Tim e Series Plots for Spot Canadian Dollar and Interest R ate Spread.
St
St
5
tN t tWZ 1M tm * (MS IIM
TkM
A ) Time S e r ie s P lo t for S p o t C anadian Dollar.
IM t IMS IM IN* IMS t*M
T im *
B ) Time S e r ie s P lo t for Interest R ate S p rea d .
40
Figure 4.2: Changes in Spot Canadian Dollar (Y t) and Interest R ate Spread (X t).
A ) T h e C h a n g e s in S p o t C a n a d ia n D ollar. B ) T h e C h a n g e s in S p o t C a n a d ia n D o lla r wit

A djustm ent.
C ) T h e C h a n g e s in In terest R a te S p r e a d . D ) T h e C h a n g e s in In te re s t R a te S p r e a d will
A djustm ent.
41
points, and bound the movement in Yt between ± 1.1 (= ± 4 * <5yt) basis points. This
m eans th a t we adjust the extrem e points and replace them as follow:
_ f 0.38 if X t > 0.38

4 “ [ -0 .3 8 if X t < -0.38
_ f 0.011 if Yt > 0.011
i -0.011 if Yt < -0.011 .
In Figure 4.2B and 4.2D, wereplot the tim e plots of Yt and X twith these changes.
T h ere axe a to ta l of 12extrem e points in X t an d a total of 5extrem e points in Yt
ad justed using the 4 *<x rule.
T he histogram of Yt and X t series axe presented in Figure 4.3. Notice th a t all the
histogram s axe sym m etrical a t zero.
Now th a t we have th e d a ta in hand, an initial approach is to use spectral analysis
to determ ine the existence of the relationship between the Yt and X t series. From
C h ap ter 2, we have a consistent estim ator of th e spectral density function, which we
apply to the Yt and X t series using the modified Daniell sm oothing function in Splus
w ith a span of 71 points (m = 35). In Figure 4.4, we plot th e spectrum s of th e two
series.
From the figures, we can see th a t there is some power in th e high frequencies in
Yt series and low frequencies in X t series. Although we do not know where th e power
is com ing from, we can relate some of the power to the short noisy appearance in the
tim e plots of Figure 4.2B and 4.2D.
N ext, Figure 4.5 shows the cumulative periodogram and th e 95% confidence limits
42
Figure 4.3: Histograms of the Changes in Spot Canadian Dollar and Interest R ate
Spread D ata.
C ) Histogram of th s C h anges In In terest R a ts O ) H istogram of the C h a n g es In terest R ate

S p re ad . S p re a d with Adjustment.
43
Figure 4.4: Spectral Density Plots of Yt and X t Series.
e
£
H U M
F req u en cy
A ) Spectral Density of y _ t Series.
••
FftqtMncy
B ) Spectral Density of X_t Series.
44
(Kolmogorov Smirnov boundaries) for both, the X t and Yt series. It is apparent th at
th e Yt series falls w ithin the confidence lim its, whereas the X t series fall outside of the
confidence limits. T he Kolmogorov Smirnov test (see C hapter 2) suggests th a t Yt is
w hite noise, although it is interesting to note th a t the spectral density of the Yt series
in Figure 4.4A is not “flat” enough; that is, this “visual” criterion disagrees w ith the
Kolmogorov Smirnov test. Thus, we reject th a t the two tim e series are white noise,
and we can carry on our analysis.
Since the bivariate tim e series is not w hite noise, we carry out the hypothesis
testing defined in C hapter 2 in order to test w hether these series are correlated at
each frequency. To continue our example, using the modified Daniell sm oothing
function with m = 35, and a = 0.1, we reject H 0 if
s = 1 — K*{y) > F °*>-2-« » = 2-3408 • C4-3)
In Figure 4.6, a plot of the statistic S is presented with the accompanying F statistic
( i ' 0.90, 2,140)• It is apparent th a t the hypothesis K 2(v) = 0 is rejected for all u € [o, |]
and th a t the Yt and the X t series have correlation over the entire frequency range.
Therefore, there is strong statistical evidence th a t the two series are correlated and
we can proceed w ith our analysis.
45
Figure 4.5: Cum ulative Spectrum Plots of Yt and X t Series.
F re q u e n c y
A ) C u m u la tiv e S p e c t r u m o f Y _ t S e r i e s .
F re q u e n c y
B ) C u m u la tiv e S p e c t r u m o f X _ t S e r i e s .
46
To proceed to our d a ta analysis, th a t is, to generate bootstrap tim e series samples,
we first apply the algorithm for choosing appropriate block size. Note th at here we
apply the algorithm to the Yt series (spot Canadian dollar d ata). We plot the standard
errors versus the block size in Figure 4.7. Note th a t the horizontal line is th e tru e
value of crp obtained from the Yt series. It is evident th at the plot suggests th a t 1 = 60
is preferred here. After fixing the block size I = 60, we follow the steps described in
C hapter 2 in order to generate B = 1500 bootstrap tim e series samples. N ext, we
can use the integrated model to obtain a point estim ate a n d /o r inference for spot
C anadian dollar series.
4.2 Empirical Results
In this section, the bootstrap forecast interval for the spot Canadian dollar is
presented. T he intervals are constructed with B = 1500, moving block size I = 60,
and the m ultilayer perceptron network with 4 input layers, 3 hidden units for the
hidden layer, and one output layer. Note th a t th e three-layer m ultilayer perceptron
network w ith the backpropagation algorithm is used to train on the input-output
pair; th a t is, Yt) ,t = 1 , 2 , . . . ,ra, where A'_3,X _ 2, X _1 and
X Q axe set to zero.
Before we provide the point estim ate and th e forecast interval for spot Canadian
dollar series, we need to carry out th e following transform ation:
47
Figure 4.6: P lot of Squared Coherence Statistic.
§ -
st -
I
05
Bloctc SlM 0)
Figure 4.7: The plot of standard error estim ates for all chosen block sizes. T he
standard error is 0.02605
48
Recall that Yt in (4.2) is defined as
Yt = log Ct - log Ct- u
hence th e forecast value for th e spot Canadian dollar Cn+1 is :
c n+1 =
= e*"+‘ ■C„ . (4.4)
As an example, the forecasted one step ahead future value forspot C anadian
dollar, at n = 1476, is given by
C n+1 = 73.0678 ,
and th e 95% forecast interval is given by
(72.0251 < C n+1 < 74.0755) .
Note th a t, the a ctu al value of Cn+i is 73.0033. These results justify the use of the
integrated m odelling technique to financial tim e series.
49
Chapter 5
Simulation
A sim ulation study is carried out to evaluate the performance of the model
described in C hapter 3. Note th a t the integrated modelling technique utilizes both
th e classical techniques and nonparam etric m ethods for this d ata analysis. For this
study, we generate n + 1 data with sample sizes n = 500, and n = 1000, for th e X t and
Yt series and sim ulate X t as Interest rate spreads d a ta and Yt as spot Canadian dollars
d ata. For each sam ple size, the sim ulated tim e series are generated w ith Gaussian
and non-Gaussian noises. In addition, the X t series are generated as follows: For
sim ulation studies I - V, we have M A = 0 order and the orders for autoregressive
(A R ) axe assigned to be 2, 3, 5,7, and 10; For sim ulation study VI, we have A R = 2
and M A = 2; and for simulation study VII, we have A R = 5 and M A = 5.
We will only use th e first n samples of (X t, Yt ) for model building. T he following
steps axe applied to each sim ulated series: (1) The extrem e points axe adjusted
according to th e “adjustm ent rule” described in C hapter 4; (2) spectral analysis is
50
used to determ ine the existence of the relationship between th e two series; (3) if
the two series axe not white noise and are correlated over th e frequency range, we
proceed to the next step, otherwise, we have to staxt on a new sim ulated series; (4) the
algorithm for choosing the block size is applied; (5) the bootstrap tim e series samples
are generated with B = 1000; ( 6 ) a three-layer m ultilayer perceptron network w ith the
backpropagation algorithm is used to train on the input-output pair; (7) th e future
value for i ^ +i, and the forecast interval for are forecasted; a n d ( 8 ) after obtaining
the 95% forecast intervals for the Yn+i, we then com pare them to th e actual Yn+i- By
repeating this procedure 200 times, with only the complete procedure counted, we can
then calculate the percentage of the intervals th at cover the a c tu a l or targeted Yn+i
for sam ple sizes n = 500 and n = 1000, respectively. Note th a t we use a bootstrap
resample size of B = 1000 to construct the bootstrap forecast intervals. Next, we
present th e sim ulated tim e series and the simulation results.
5.1 Simulated Series I and Simulation Results
S im u la ted Series 1(a): W ith G aussian N oise
W e sim ulate tim e series X t as A R I M A [ 2 , 1, 0 ), as shown below:
( l - ' E < f . i B i ) ( l - B ) X , = U, , V , ~ N ( 0 ,1 ) (5.1)

i=i
51
where th e coefficients <f>j, j = 1,2 axe random ly chosen to be 0.65 an d —0.2.
We sim ulate tim e series Yt as follows:
Zt = —0J2Xt -f v /F —ffi2^ - H t
Yt = \trunc { m in ( Z t)}\ + 1 + Zt , (5.2)
where the function m m (-) returns a num ber th a t is the minimum of th e inputs,
trunc(-) is the function th a t creates integers from floating point num bers by
going to th e next integer closer to zero, and H t is defined by
(1 - 4>B)(l - B )H t = Vt , K ~ W ( 0 ,1 ) (5.3)
and a random ly chosen coefficient, <f>= 0.32.
S im u la te d S eries 1(b): W ith N o n -G a u ssia n N o ise
We sim ulate tim e series X t as A R I M A ( 2 ,1 ,0 ), as shown below:
i= i
Ut ~ (80% of N ( 0,1) -I- 20% of C auchy{0 , 1)) (5.4)
where the coefficients <py, j = 1,2 axe random ly chosen to be 0.58 and —0.12.
Zt = -0 .2 X t + V l - 0.22 - H t
Yt = |trunc {m i n ( Z t )}\ + 1 + Z t , (5.5)
52
where Ht is defined by
(1 - 4>B){ 1 - B )H t = Vt , Vt ~ N ( 0 , 1) (5.6)
and a random ly chosen coefficient, <f>= 0.45.
S im u la tio n R e su lt I
T he sim ulation results are sum m arised in Table 5.1. Note th a t th e coverage
probability for sim ulation results axe lower than the nom inal level of 95%.
Moreover, th e results show indifferently for both sam ple sizes as well as the
noise distributions. However, for both samples sizes, th e average error is higher
in non-Gaussian noise series th an the Gaussian noise series. T he average error
Table 5.1: Results for Forecast Intervals on Sim ulated Series I
A = 500 N = 1000
A R = 2 and M A = 0 A R = 2 and M A = 0
Noise Nominal Coverage Average Nominal Coverage Average

Level Proportion Error Level Proportion Error
Gaussian 95 % 92.0 % 0.801 95 % 92.5 % 0.763
Non-Gaussian 95 % 92.0 % 0.922 95 % 92.0 % 0.920
is defined by
Average Error = ^ ■
t^ 1 ^ — — . (5.7)
53
5.2 Sim ulated Series H and Sim ulation R esu lts
S im u la ted Series H (a ): W ith G au ssian N o ise
We sim ulate tim e series X t as A R I M A ( 3,1,0), as shown below:
(1 - £ - B ) X t = Ut , Ut ~ N ( 0 , 1) . (5.8)
i=i
Note th at th e series is generated using the Splus function. We sim ulate tim e
series Yt as follows:
Zt = -Q .2 X t + V l - 0 .22 - H t
Yt = \trunc { m in ( Z t)}\ + 1 + Z t , (5.9)
where H t is defined by
(1 - 4>B){1 - B )H t = Vt , K ~ i V ( 0 ,l ) - (5-10)
Note th at th e series Ht is also generated using Splus function.
54
S im u la ted S eries 11(b): W ith N on -G au ssian N o ise
We sim ulate tim e series X t as A R I M A (3 , 1,0), as shown below:
- B ) X t = Ut ,
i= i
Ut ~ (80% of 1V(0 , 1 ) + 20% of C a u ch y{0 , 1)) . (5-11)
Zt = —0 .2X t + V l - 0 .22 - H t
Yt = \trunc { m in ( Z t)}| + 1 -f- Zt , (5.12)
(1 - <f>B){1 - B )H t = Vt , Vt ~ N ( 0 , 1 ) . (5.13)
S im u la tio n R e su lt H
The sim ulation results are su m m arised in Table 5.2 . Note th at the autoregres
sive order is 3 in this simulation study. Although the coverage probability
for sim ulation result II show slight improvement in coverage probability as
com pared w ith the sim ulation result I, the coverage probability is still under the
nominal level of 95%. The average error for non-Gaussian noise series rem ained
high at about 1 basis point (which is about 4-oyt in actual spot C anadian dollar
da ta in C hapter 4).
55
Table 5.2: R esults for Forecast Intervals on Sim ulated Series II
N = 500 N = 1000
A R = 3 and M A = 0 A R = 3 and M A = 0
N oise N om inal Coverage Average N om inal Coverage Average

Gaussian 95 % 93.0 % 0.873 95 % 93.0 % 0.861
N on-G aussian 95 % 92.0 % 0.902 95 % 92.5 % 0.911
5.3 Simulated Series III and Simulation Results
S im u la te d S eries 111(a): W ith G au ssian N oise
We sim ulate tim e series X t as A R I M A { 5,1,0), as shown below:
(1 - j r ^ ) ( l - B ) X t = Ut , Ut ~ iV(0 , 1 ) . (5.14)
i= i
Zt = -Q .2 X t + V T - O J P • H t
Yt = \trunc {rran(Z£)}| + 1 -+- Z t , (5.15)
( l - 4 > B ) ( l - B ) H t = Vt , Vt ~ N ( 0 , 1 ) . (5.16)
56
S im u la te d S eries 111(b): W ith N o n -G a u ssia n N o ise
We sim ulate tim e series X t as A R I M A { 5 , 1,0), as shown below:
i=i
Ut ~ (80% of 1V(0,1) + 20% of Cauchy( 0 , 1)) . (5-17)
Zt = - 0 .2X* + V l - 0 . 22 - H t
Yt = |£nm c { m in (Z t)}\ + 1 4- Zt , (5.18)
( l - < f > B ) ( l - B ) H t = Vt , Vt ~ N ( 0 , l ) . (5.19)
S im u la tio n R e su lt II I
T he sim ulation results are su m m arised in Table 5.3. In this study, the autore
gressive order isincrease from 3 to 5. The coverage probability for sim ulation
results show decreases as compared with sim ulation results I and II. However,
the average error for non-Gaussian noises axeappears closer to Gaussian noise
series for both sample sizes. T he coverage probability for these sim ulation
results are still lower than the n om in al level of 95%.
57
Table 5.3: Results for Forecast Intervals on Sim ulated Series III
N = 500 AT = 1000
A R = 5 and M A = 0 A R = 5 and M A = 0
Noise N om in al Coverage Average Nominal Coverage A verage

Gaussian 95 % 92.0 % 0.882 95 % 92.0 % 0 .863
Non-Gaussian 95 % 91.5 % 0.900 95 % 92.0 % 0 .8 7 4
5.4 Simulated Series IV and Simulation Results
S im u la ted S eries IV (a): W ith G aussian N o ise
We sim ulate tim e series X t as A R I M A ( 7 ,1 ,0 ), as shown below:
(1 - £ ^ ) ( 1 - B ) X t = Ut , Ut ~ W(0 , 1 ) . (5.20)
i=i
We sim ulate tim e series K as follows:
Z t = — 0. 2X t + V l - 0 . 22 • Ht
Yt = |tru n c (m in(Zt)}[ + 1 + , (5.21)
(1 - 1 - B )H t = Vt , Vt ~ W (0, 1) . (5.22)
58
S im u la te d S eries IV (b ): W it h N on -G au ssian N o ise
{ l - Y f 4>i B i ) { l - B ) X t = Ut ,
i=i
Ut ~ (80% of i\T(0,1) + 20% of C auchy(0,1)) . (5.23)
Zt = - 0 .2 X * + Vl - 0 .2 2 - Ht
Yt = \trunc {rm’n(Z t)}[ + 1 + Zt , (5-24)
( l - < f > B ) ( l - B ) H t = Vt , Vt ~ N ( 0 ,1 ) . (5.25)
S im u la tio n R e su lt IV
T he sim ulation results axe sum m arised in Table 5.4. Note th a t th e autore
gressive order is now 7. T h e coverage probability for the sim ulation results
show no improvement; they also show decreases as compared w ith th e previous
sim ulation results. As usual, th e coverage probability for the sim ulation results
are lowerth an th e nominal level. But the average errors are alm ost th e sam e
for b o th noise distributions w ithin the same sample size.
59
Table 5.4: Results for Forecast Intervals on Sim ulated Series IV
lV = 500 N = 1000
Aft = 7 and M A = 0 AR = 7 and M A = 0

Gaussian 95 % 91.5 % 0.781 95 % 92.0 % 0.832
Non-Gaussian 95% 90.5% 0.779 95 % 91.0 % 0.869
5.5 Simulated Series V and Simulation Results
S im u la te d S eries V (a ): W ith G a u ssia n N o ise
10
( 1 - 5 3 0 y ^ ') ( l - B ) X t = Ut , Ut ~ N ( 0,1) . (5.26)

y=i
We sim ulate tim e series K as follows:
Zt — —0.2X t 4" \ / l — 0.22 - S t
Yt — \tru n c {m in (Z t )}\ + 1 + Z t , (5-27)
(1 - 4>B){ 1 - B ) S t = Vt , Vt ~ iV (0,1) . (5.28)
60
S im u la te d S eries V (b ): W ith N o u -G a u ssia n N o ise
We sim ulate tim e series X t as A R I M A { 10,1,0), as shown, below:
10
(1 - £ f c # ) ( l - B ) X t = U t,
i= 1
Ut ~ (80% of N { 0,1) + 20% of C auchy{0,1)) . (5.29)
Zt = - 0 .2 X t + Vl - 0 .2 2 - Ht
Yt = |tru n c {m iu(Z t)}| + 1 + Z t , (5.30)
(1 - 4>B)(l - B ) H t = Vt , Vt ~ W (0,1) . (5.31)
S im u la tio n R e su lt V
T he sim ulation results axe sum m arised in Table 5.5. Notice th a t the coverage
probability for sim ulation results are decreasing as the autoregressive order
increases to 10. In fact, the average errors are higher than usual. In this
sim ulation study, we still have th e coverage probability lower than nom inal
level.
61
Table 5.5: Results for Forecast Intervals on Sim ulated Series V
N = 500 JV = 1000
A R = 10 and M A = 0 A R = 10 and M A = 0

Gaussian. 95 % 90.0 % 1.080 95 % 90.5 % 0.897
Non-Gaussian 95 % 89.0 % 1.110 95 % 90.0 % 0.903
5.6 Simulated Series VT and Simulation Results
S im u la te d Series VT(a): W ith G aussian N o is e
We simulate tim e series X t as A R IM A ( 2 ,1 ,2 ), as shown below:
(1 - f > / j S ' ) ( l - B ) X t = (1 + j 2 8 iB i) U t , Ut ~ t f ( 0 ,1) - (5.32)

3=1 3= 1
We simulate tim e series Yt as follows:
Zt = -Q .2X t + ■Ht
Yt = |£runc {rnin(Z t)}\ + 1 + Z t , (5.33)
( l - 4 > B ) ( l - B ) H t = Vt , Vt ~ AT(0,1) . (5.34)
62
S im u la te d Series VT(b): W ith N o n -G a u ssia n N oise
We sim ulate tim e series X t as A R IM A { 2 ,1 ,2 ), as shown below:
(1 - £ t i & X 1 - B )X , = (1 + £ » iB s)Ut ,
j =1 j —1
~ (80% of iV(0,1) + 20% of C auchy{0,1)) . (5.35)
Zt = -0 .2 X £ + V l - 0.22 - Ht
Yt = \trunc { m in (Z t )} \ + 1 + Z t , (5.36)
(1 - 4>B){1 - B )H t = Vt , Vt ~ AT(0,1) . (5.37)
S im u la tio n R e su lt VT
T he sim ulation results are sum m arised in Table 5.6. Note that we have the
M A = 2 and A R = 2 in this sim ulation study. T he coverage probability
for sim ulation results axe very sim ilar to sim ulation results I and II. However,
for respective sam ple size, the average errors are almost indifferent for both
Gaussian and non-Gaussian noises. Again, th e coverage probability results are
still lower than th e nominal level.
63
Table 5.6: Results for Forecast Intervals on Simulated Series VT
IV = 500 IV = 1000
A R = 2 and M A = 2 AJ2 = 2 and MA = 2

Gaussian 95 % 92.0 % 0.887 95 % 92.0 % 0.916
Non-Gaussian 95 % 92.0 % 0.879 95 % 92.0 % 0.920
5.7 Simulated Series VII and Simulation Results
S im u la te d Series V H (a ): W ith G au ssian N o ise
We sim ulate tim e series X t as A R I M A ( 5 ,1,5), as shown below:
(1 - £ 1 - B ) X t = (1 + £ OjBi)Ut , Ut ~ N (0,1) . (5.38)

j= i j= i
We simulate tim e series Yt as follows:
Zt = - 0 .2 X t + Vl - 0 .2 2 - Ht
Yt = |trtm c { m in (Z t)}j + 1 + Zt , (5.39)
where is defined by
(1 - 4>B){ 1 - R)IT£ = Vt , VJ ~ W (0,1) . (5.40)
64
S im u la ted S eries VTI(b): W ith N on -G au ssian N o ise
We sim ulate tim e series X t as A R I M A { 5 ,1 ,5 ), as shown below:
u-E -
b )x , = (i+e
h& W t ,
i= i i= i
Ut ~ (80% of N ( 0, 1) + 20% of C aucky(0,1)) . (5.41)
Zt = - 0 .2 X t + V l - 0.22 - Ht
Yt = \trunc {m in(Zt )}| + 1 + Zt , (5-42)
where Ht is defined, by
(1 — <f>B)(l — B ) H t = Vt , K ~ W(0,1) - (5-43)
S im u la tio n R e su lt V H
The sim ulation results axe sum m arised in Table 5.7. As we increase b o th the
order for M A and A R to 5, the coverage probability for sim ulation results axe
considered the lower as compaxed to previous simulation results. Also, the
average errors show no improvement.
65
Table 5.7: Results for Forecast Intervals on Sim ulated Series V II
IV = 500 N = 1000
A R = 5 and M A = 5 A R = 5 and M A = 5

Gaussian 95 % 90.0 % 1.053 95 % 89.5 % 0.976
Non-Gaussian 95 % 89.5 % 1.092 95 % 89.5 % 1.022
5.8 Discussion
In Table 5.8, it is apparent th a t the coverage probability for all sim ulation results
are lower than th e nom inal level of 95%. T here are five possible reasons for this
low coverage probability: (1) The estim ates taken to be the true values used for
constructing the bootstrap forecast intervals axe inaccurate. (2) T he confidence
intervals based on bootstrap percentiles m ethod m ay not have performed well due to
th e consequence of nonparam etric inferences. It may have underestim ated the tails
of the distribution of prediction. (3) T he num ber of repetitions for th e simulation
procedure in this study was too low (in this study the sim ulation procedure was
only repeated 200 tim es). (4) T he adjustm ent rule for extrem e points m ight affect
th e results for coverage probability. However, because we are trying to model the
long term relationship for bivariate financial series, it seems reasonable to apply
th e adjustm ent rule to those extrem e points for better m odel analysis. (5) As the
sim ulation results show, the coverage probability decreases as we increase the order
66
for M A a n d /o r A R . This m ay be due to com putation errors occurring as th e model
becomes m ore complicated.
Table 5.8: Results for 95% Forecast Intervals on Simulated Series
N = 500 N = 1000
Simulated Gaussian Non-Gaussian Gaussian Non-Gaussian

Series
I (A R = 2, M A = 0) 92.0% 92.0 % 92.5 % 92.0 %
H (AR = 3, M A = 0) 93.0 % 92.0 % 93.0 % 92.5 %
HI (AR = 5, M A = 0) 92.0 % 91.5 % 92.0 % 92.0 %
IV (A R = 7, M A = 0) 91.5 % 90.5 % 92.0 % 91.0 %
V (AR = 10, M A = 0) 90.0 % 90.0 % 90.5 % 90.0 %
VI (AR = 2, M A = 2) 92.0 % 92.0 % 92.0 % 92.0 %
VII (AR = 5, M A = 5) 90.0 % 89.5 % 89.5 % 89.5 %
T he results obtained when n = 500 do not differ greatly as compared w ith those
obtained when n = 1000. In addition, the sim ulation results show that the integrated
modelling technique is insensitive to the choice of distribution for the noise. Since we
m ust clean up th e extrem e points for each of th e sim ulated series, this m ay explain
why this technique is insensitive to the Gaussian or non-Gaussian noise. However,
there is a slight decrease in coverage probability when we increase the order for A R
67
and M A . This also suggests that we m ay want to keep the model simple in order
to save com putation tim e and obtain b e tter results. Again, the integrated modelling
technique is not sensitive to the com plicated series.
A lthough th e results seem reasonable when using the integrated modelling tech
nique, the operation costs axe high in term s of CPU .
To conclude, the integrated m odelling technique is worthwhile when the complex
ity of the d a ta requires a more versatile approach th an allowed by classical m ethods.
As to th e reason for under coverage in probability, this remains an issue to be resolved
in future studies.
68
Chapter 6
Conclusions
This thesis has demonstrated, the usefulness of both classical m ethods and non-
param etric m ethods in analysing financial tim e series. T he approach developed in
C hapter 2 give the basic ideas for im plem enting techniques in modelling financial
tim e series.
As discussed in C hapter 2, commonly used methodologies (such as AREMA mod
els) provide us with estim ation and forecast interval. However, in financial m arkets,
tim e series are statistically more volatile as compared to other applications. Thus,
simple models alone are unable to solve difficult problems such as exchange rate
prediction. T he bivariate transfer function model is introduced in order to improve
the forecast estim ation as compared with th e univariate case. However, th e transfer
function m odel requires the value of X n+ i, which is not available and m ust itself be
forecast in order to forecast Yn+1. The accuracy of the forecasting interval for 3^+i can
be improved by not requiring the estim ation of X n+i, thus reducing th e probability
of error in estim ation of Yn+1, as shown by th e integrated model.
69
T he frequency dom ain methods provide us (w ith the aid of high speed com puter
hardw are), a more efficient way to uncover certain statistical details. Particularly for
the bivariate tim e series, we need to determ ine th e existence of a relationship between
two tim e series in order to further th e data analysis. Together with the classical
techniques, m odem nonparam etric m ethods such as th e moving blocks bootstrap and
neural netw ork provide us with a theoretical basis for the derivation of an integrated
m odelling technique.
T h e integrated modelling technique described in C hapter 3 provides us w ith
the tools to further analyse the bivariate tim e series. The technique utilizes b o th
classical techniques and nonparam etric methods for b etter estim ation and inference.
In C h ap ter 4, an illustration of integrated modelling techniques is applied to th e
financial tim e series. T he empirical result support th e use of th e proposed m odelling
technique.
In C hapter 5, although the simulation study was only carried out for n=500 and
n=1000, the sim ulated tim e series were generated w ith Gaussian and non-Gaussian
noises. T he sim ulation results showed undercoverage overall, but the integrated
modelling technique appears robust to the choice of noise distribution as well as
the sam ple size.
70
6.1 Future Research
This research could be extended by adding more variables to th e model given in
chapter 3. For the spot Canadian dollar example, we can add oth er variables such
as the dem and an d /o r supply of Canadian dollars, dem and a n d /o r supply of U.S.
dollars, and th e stock indexes to the m odel. The dem and and supply of currency will
help us predict w hether or not future values will soon rise or fall. As for the stock
indices, increase or decrease, this would help us to determ ine w hether the m arket is
perform ing well. O ther possible variables to add axe those pertaining to available
information from the derivative m arkets.
By adding more variables to the model, two approaches to analysing the tim e
series are possible. The first approach would be to utilize m ultiple linear regression
techniques. Note th at economists describe a regression m odel as an econometric
model. Although m ultiple linear regression models are widely used due to th eir
com putational simplicity, there are problems inherent to this approach which need to
be pointed out. One problem is th at explanatory variables can be highly correlated,
leading to singularity problems. Hence, it is advisable to look at the correlation
m atrix of the explanatory variables before carrying out a m ultiple linear regression
analysis. A nother problem involves th e structure of th e error term s. It is always
assumed th a t th e errors are i.i.d.N (0, a 2), but this assum ption m ay not be correct.
If the residuals are correlated, one can try fitting a m ultiple linear regression m odel
71
with, autocorrelated errors (See Kendall et al. [37]).
T he second approach is to follow the approach, developed in Chapter 3. but to
pair up the spot Canadian dollar with other variables. However, how do we combine
the information from different pairings of Yt (spot Canadian dollar series) and other
variables, without producing too many forecast intervals?
Since the confidence intervals based on th e bootstrap percentiles m ethod have
less satisfactory probability coverage, this research could be improved by replacing
the confidence intervals w ith m ore advanced bootstrap intervals methods such as bias-
corrected and accelerated (B C a) method, and the approximate bootstrap confidence
intervals (A B C ) m ethod. These advance bootstrap intervals m ethods could partially
correct the undercoverage problems in our sim ulation studies.
In dynamic representation, the recurrent networks m ay be more in line with
nonlinear tim e series. Hence, by replacing th e feed-forward network with recurrent
network, we m ay achieve a b etter point estim ation and satisfactory probability cov
erage.
72
Bibliography
[1] A braham , B. and Ledolter, J . [1983]. Statistical methods fo r forecasting. John
W iley Sons, New York.
[2] Anderson, T .W . [1971]. The statistical analysis o f tim e series. John Wiley &
Sons, New York.
[3] Beran, R. [1982]. Estim ating sam pling distributions : The bootstrap and
com petitors. The Annals o f Statistics 10, 212-225.
[4] Beran, R. [1988a]. Balanced sim ultaneous confidence sets. Journal o f the
Am erican Statistical Association 83, 679-686.
[5] Beran, R. and Ducharme, G .R. [1991]. A sym ptotic theory fo r bootstrap methods
in statistics. Centre de Recherches M athem atiques : Universite de Montreal.
[6] Bickel, P .J. and Freedman, D.A. [1981]. Some asym ptotic theory for th e
bootstrap. The Annals o f Statistics 9, 1196-1217.
[7] Bloomfield, P. [1976]. Fourier analysis o f time series : An introduction. John
W iley & Sons, New York.
[8] Box, G.E.P. an d Jenkins, G.M . [1976]. Tim e series analysis : Forecasting and
control. Holden-Day, San Francisco.
[9] Brockwell, P.J. and Davis R .A . [1987]. Tim e series : Theory and methods.
Springer-Verlag, New York.
[10] Chatfield, C. [1996]. The Analysis o f Tim e Series : An introduction. 5 th Ed.,
C hapm an and H all, London.
[11] Cheng, B. and T itterington, D.M . [1994]. Neural networks : A review from a
statistical perspective. Statistical Science 9, 2-54.
[12] Christensen, R. [1991]. Linear models fo r multivariate, time series, and spatial
data. Springer-Verlag, New York.
[13] Cybenko, G. [1989]. Approxim ation by superposition of a sigmoidal function.
M athematics o f Control, Signals and System s 2, 303-314.
[14] Diaconis, P. an d Efron, B. [1983]. C om puter intensive m ethods in statistics.
Scientific American 248(5), 116-130.
[15] Efron, B. [1979a]. Bootstrap m ethods: A nother look at the jackknife. The Annals
o f Statistics 7 , 1-26.
[16] Efron, B. [1979b]. Computers and the theory of statistics : Thinking the
unthinkable. S IA M Review 21, 460-480.
[17] Efron, B. [1981]. Nonparam etric estim ates of standard error : T he jackknife, th e
bootstrap and th e other m ethods. Biom etrics 68, 589-599.
74
[18] Efron, B. [1982]. The jackknife, the bootstrap and other resampling plans. Society
for Industrial and Applied M athem atics. Philadelphia, Pennsylvania.
[19] Efron, B. and Tibshirani, R .J. [1986]. Bootstrap m ethods for standard errors,
confidence intervals and other measures of statistical accuracy. Statistical
Science 1, 54-77.
[20] Efron, B. and Tibshirani R .J. [1993]. A n introduction to the bootstrap. C hapm an
and Hall, New York.
[21] Franke, J. and Hardle, W . [1992]. O n bootstrapping kem al spectral estim ation.
The Annals o f Statistics 20, 121-145.
[22] Freedman, D.A. [1981]. Bootstrapping regression models. The Annals o f Statis
tics 9, 1218-1228.
[23] Freedman, D.A. [1984]. On bootstrapping two-stage least-squares estim ates in
stationary linear models. The Annals o f Statistics 12, 827-842.
[24] Funanhashi, K.I. [1989]. On the approxim ate realization of continuous mappings
by neural networks. Neural Networks 2, 183-192.
[25] Gallant, A.R. and W hite, H. [1988]. There exists a neural network th a t does not
make avoidable mistakes. Proceedings o f the second annual IE E E Conference on
Neural Networks, San Diego, CA, Vol 1 (IEEE Press, New York), 657-664.
[26] Gallant, A.R. and W hite, H. [1992]. On learning th e derivatives of an unknown
m apping with multilayer feedforward networks. Neural Networks 5, 129-138.
75
[27] Granger, C -W .J. [1992]. Forecasting stock market prices : Lessons for forecasters.
International Journal o f Forecasting 8, 3-13.
[28] Hall, P. [1986]. O n the num ber of bootstrap simulations required to construct a
confidence interval. The Annals o f Statistics 14, 1463-1562.
[29] Hall, P. [1988]. Theoretical comparison of bootstrap confidence intervals (w ith
discussion). The Annals o f Statistics 16, 927-985.
[30] Haykin, S. [1994]. Neural networks : A comprehensive foundation. M acmillan,
New York.
[31] Hecht-Nielsen, R. [1989]. Theory of the backpropagation neural networks. Pro
ceedings o f the International Joint Conference on Neural Networks, W ashington
D. C., Vol 1 (IEEE Press, New York), 593-605.
[32] Horaik, K., Stinchcombe, M. and W hite, H. [1989]. M ultilayer feedforward
networks are universal approxim ators. Neural Networks 2, 359-366.
[33] Homik, K., Stinchcombe, M. and W hite, H. [1990]. Universal approxim ation of
an unknown m apping and its derivati ves using m ultilayer feedforward networks.
Neural Networks 3, 551-560.
[34] Homik, K ., Stinchcombe, M., W hite, H. and Auer, P. [1994]. Degree o f approx
imation results fo r feedforward networks approximating unknown mappings and
their derivatives. UCSD discussion paper.
[35] Hwang, J .T . and Ding, A.A. [1997]. Prediction intervals for artificial neural
networks. Journal o f the American Statistical Association 92, 748-757.
76
[36] Jenkins, G.M. and W atts, D.G. [1969]. Spectral analysis and its applications.
Holden-Day, California.
[37] Kendall, M.G., S tuart, A. and Ord, J.K . [1983]. The advanced theory o f statistics.
vol. 3 (4th edn.), London, Griffin.
[38] K im , P.T ., M artin, P., and Kelly, D. [1992]. Unanticipated changes in interest
rate spreads and th eir effect on the Canadian spot exchange rate : A n update.
Fixed Income Research : Wood G undy Inc.
[39] Kim P.T ., M artin P. and Staley M. [1996]. A neural network approach to
forecasting the Canadian spot exchange rate. Fixed Income Research : Wood
G undy Inc.
[40] Kim P.T . and M artin P. [1996]. On the relationship between the interest rate
spread and the spot Canadian dollar. University of Guelph Lecture Notes.
[41] Kuan, C.M. and W hite, H. [1994]. Artificial neural networks: An econometric
perspective. Econometric Reviews 13, 1-91.
[42] Kuan, C.M . and Liu, T. [1995]. Forecasting exchange rates using feedforward
and recurrent neural networks. Journal o f Applied Econometrics 10, 347-364.
[43] Kunsch, H.R. [1989]. The jackknife and th e bootstrap for general stationary
observations. The Annals o f Statistics 17, 1217-1241.
[44] Leger, C., Politis, D.N. and Romano, J.P . [1992]. Bootstrap technology and
applications. Technometrics 34, 378-398.
77
[45] Lippm ann, R.P. [1987]. An introduction to computing with neural nets. IE E E
A S S P Magazine, April Issue, 4-23.
[46] Liu, R.Y. and Singh, K. [1992]. Moving blocks jackknife and bootstrap capture
weak dependence, In Exploring th e lim its of bootstrap, eds. LePage, R . and
Billard, L-, John Wiley, New York.
[47] Ljung, G.M. and Box, G.E.P. [1978]. On a measure of lack of fit in tim e series
models. Biometrika 65, 297-303.
[48] Priestly, M.B. [1981]: Spectral analysis and tim e Series. Academic Press, New
York.
[49] Quenouille, M.H. [1956]. Notes on bias in estim ation. Biometrika 43, 353-360.
[50] Rao, V.B. and Rao, H.V. [1993]. C++ neural networks and fuzzy logic. MIS:Press,
New York.
[51] Ripley, B.D. [1993]. Statistical aspects of neural networks. In Networks and
Chaos - Statistical and Probabilistic Aspects (eds O.E. Bamdorff-Nielsen et al.).
C hapm an & Hall, London, 40-123.
[52] Stacey, D. [1995] Lecture notes : COMP 27-6420 (Spring Semester). Depaxtment
of C om puting and Information Science, U. of Guelph.
[53] T ibshirani, R .J. [1985]. How many bootstraps ? (Tech. rep. no. 362). Stanford
University, D ept, of Statistics. Stanford, CA.
[54] Weigend, A.S. and Gershenfeld, N.A. (eds) [1994], Time series prediction,
Reading, MA : Addison-Wesley.
78
[55] W hite, H. [1989]. Learning in artificial neural networks : A statistical perspective.
Neural Computation 1, 425-464.
[56] W hite, H. [1992]. Artificial neural networks: approximation and learning theory.
Cambridge, MA: Blackwell.
[57] W u, C .F.J. [1986]. Jackknife, bootstrap and o th er resampling m ethods in
regression analysis. The Annals o f Statistics 14, 1261-1295.
79
A p p e n d ix A : T h e B ack -P rop agation A lg o rith m
Before we provide the back-propagation algorithm , we present the m ultilayer
perceptron network model used for financial d a ta analysis. For the given input
X (s ) = (X !(s),X 2( 5 ) , . . . ,Xfc(s))f, s = 1 ,2 ,... ,n , the o utput of the netw ork is
defined as
G e (X ) = r„ + £ m, f ( v ' , X + * .)
3=1
where th e activation function /(*) is the logistic function defined in (2.28), and 0 =
(iox,u72, . . . . . . , 1; * , ^ , $ ! , ^ , . . . , $ p) where u* = (ua l, v s2, . . . , u ifc), are
the weights and threshold values, and the superscript t is denoted as transpose.
T h e following backpropagation algorithm is modified from th e lecture notes by
Stacey [52].
1. N o ta tio n s :
Li :the input layer

L ff :the hidden layer
Lo :the output layer
Vih :the synaptic weight between the L i an d L h
Who '■the synaptic weight between the L h and Lo
$ h '■ TffProcessing Elements (PE) threshold values
T0 : To PE threshold values
N ote th at we assign random values in the range [ —1, + 1] to \h,Who-> and

IV
80
2. F o rw a rd P a s s C o m p u ta tio n : For each input-output pair ( X a, 0 S):
a. Process X / s values to calculate new L h P E activations using

k
hi = x hv hi + 4 \)
h= 1
where /(-) is th e sigmoid function defined in equation 2.28.

b. F ilter th e L h activations through the weights; W , to Lo using
p
Oj = fetDii + r i
t=i
c. Com pute the output differences by dj — ( 0 | — Oj)
3. B a c k w a rd P a s s C o m p u ta tio n :
I. Com pute th e output error at L o PE values using
ey = Oj{l - Oj)dj
II. Calculate the error of each L h P E relative to each dj with

p
t{ —
—6f(l bi) Wij&j
X=1
TIT. A djust th e L h to Lo connections; that is, th e am ount of change m ade to

the connection from the ith L h to the j th L o .
A w ij = a6,-ey
where a is a positive constant controlling th e learning rate.

IV . Adjust th e Lo threshold value APy = aey
V. Adjust th e L j to L h connections AVhi = where /3 is a positive
constant controlling the learning rate.
VT. A djust th e L h threshold value A$,- = /?ft-
4. I t e r a t i o n : Repeat Step 2 and Step 3 until ey’s are all either zero or at th e
m inim um value.
81
5. R e c a ll : T he recall mechanism consists of th e two feed-forward operations.
1. C reate Lh P E values
k
bi = f (5 2 x hVhi + $ f )
h=l
2. After all L h PE activations have been calculated, th ey axe used to compute

the output.
Oj = 52biwij + Tj
82

Integrating Modelling Technique

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Integrating Modelling Technique

Uploaded by

Copyright:

Available Formats

INFORMATION TO USERS

The quality of this reproduction is dependent upon the quality of th e copy

Oversize materials (e.g., maps, drawings, charts) are reproduced by sectioning

Photographs included in the original manuscript have been reproduced

Bell & Howell Information and Learning

The University of Guelph

In partial fulfilment of requirem ents

© V ictor Khoon-Lee Tan, 1999

OurSie Notre reference

The author retains ownership o f the L’auteur conserve la propriete du

GRADUATE PROGRAM SERVICES

Title: In tegratin g Modelling Techniques for Financial Time S e r ie s

Advisory Committee p £-jm

Received by: __________________ Date: AU0 2 3 1999

INTEG RATING MODELLING TECHNIQUES FO R FINANCIAL TIM E SERIES

Victor Khoon-Lee Tan Supervisors:

T he use of a related series of historical values as evidence or a guide line to predict

T he inadequacy of the efficient market hypothesis and the instability of d ata

I would like to express my sincere appreciation to:

First and foremost, I would like to thank those professors on m y supervisory

During my stay at Guelph, I am fortunate to have P eter Yong, Siew-Hong Tan,

1.1 In tro d u c tio n ...................................................................................................... 1

1.2 Scope of T he Thesis ...................................................................................... 5

2 T im e S eries M eth o d o lo g ies 7

2.1 Classical Tim e Series M e th o d s............................................................... 7

2.1.1 Spectral A n a ly s is ................................................................................. 8

2.1.2 Cross Spectra C oherency .................................................................... 12

2.1.3 Bivariate Transfer Function M o d e l .................................................. 16

2.2 N onparam etric M e t h o d s ................................................................................ 16

2.2.1 Moving Blocks Bootstrap S a m p le .................................................. 17

2.2.2 Modified Moving Blocks B o o t s t r a p ................................................. 18

2.2.4 Feed-forward Networks : M ultilayer P e rc e p tro n s ......................... 21

3.1 Theoretical Foundations for Integrated Modelling..................................... 28

3.2 Modelling P ro c ed u re s....................................................................................... 34

3.2.1 Algorithm for Choosing the Block S i z e ......................................... 34

3.2.2 G enerating Bootstrap Bivariate Tim e Series.................................. 36

3.2.3 Neural Network E stim a tio n s................. 36

4 A p p lic a tio n to F in an cial T im e S eries 38

4.1 T he Interest R ate and Spot Canadian Dollar D a t a ................................... 38

4.2 Em pirical R e s u lts ............................................................................................. 47

5.1 Sim ulated Series I and Simulation R e s u l t s ............................................... 51

5.2 Sim ulated Series II and Simulation R e s u lts ............................................... 54

5.3 Sim ulated Series III and Simulation Results ............................................ 56

5.4 Sim ulated Series IV and Simulation Results ............................................ 58

5.6 Sim ulated Series VT and Simulation Results ............................................ 62

5.8 D isc u ssio n .......................................................................................................... 66

6.1 F uture Research................................................................................................. 71

(Modified from Efron and Tibshirani [21]). . ........................................ 19

2.2 Three-Layer Feedforward Network................................................................. 22

2.3 Threshold Functions for Neural Networks.................................................... 26

4.4 Spectral Density Plots of Yt and X t Series................................................... 44

4.5 Cum ulative Spectrum Plots of Yt and X t Series......................................... 46

4.6 Plot of Squared Coherence Statistic.............................................................. 48

true standard error is 0.02605 ..................................................................... 48

5.1 Results for Forecast Intervals on Sim ulated Series I ....................... 53

5.2 Results for Forecast Intervals on Sim ulated Series I I ............................... 56

5.3 Results for Forecast Intervals on Sim ulated Series I I I ............................ 58

5.4 Results for Forecast Intervals on Sim ulated Series I V .................... 60

5.5 Results for Forecast Intervals on Sim ulated Series V ....................... 62

5.6 Results for Forecast Intervals on Sim ulated Series V T .................... 64

5.7 Results for Forecast Intervals on Sim ulated Series V I I .................... 66

common to reseaxchers of m any disciplines. In order to achieve a ‘b e tte r’ outcome,

which are unique to each field. However, an awareness of the characteristics of th e

These problems create serious challenges to th e im plem entation of modelling strate

“T he spectral point of view is particularly advantageous in the analysis of mul