RJTA Vol. 12 No.

3 2008

56
A Web-Based System for Fashion Sales Forecasting

Yong Yu, Tsan-Ming Choi, Kin-Fan Au* and Chui-Yan Kwan
Institute of Textiles & Clothing, The Hong Kong Polytechnic University,
Hung Hom, Kowloon, Hong Kong


ABSTRACT

The fashion industry faces more challenges from demand uncertainty than many other
businesses because it creates products that are highly seasonal with short lifetimes and
demand is inherently volatile. Such a situation introduces difficulties in fashion supply chain
management. The solution to the agile supply chain involves setting up seamless or
boundary-less connections between supply chain members. These connections can minimize
buffers between the different stages in the chain. In an agile network, such connection is
critical and can be enabled by web-software, allowing different actors to be connected
without needing to have the same computer system. With integrated web systems,
businesses in different geographical locations can behave as if they belong to the same
enterprise. The forecasting of future sales is one of the key constituents of these solutions.
Todays enterprises in fashion often employ various IT services in supply chain operations
and a forecasting system is often expected to be accessible by independent users and
systems through a standardized interface. Therefore, web-software is the right solution in
this circumstance. There are many methods in the sales forecasting field. In this paper, we
focus on exploring the implementation of a web-based forecasting system in which various
forecasting methods can be utilized. This web-based forecasting system is expected to bring
great flexibility into fashion enterprise operations and enhance their supply chain
management. A case analysis is presented in the paper in which a neural network is utilized
as the forecasting method. We believe that the implementation mechanism is highly
applicable to help fashion companies in improving their operations.

Keywords: Fashion business, forecasting, supply chain management, demand
variability, management information systems, web-based system


1. Introduction

Customer demand in the fashion industry is highly
variable and volatile (Brannon, 2005).

Fashion companies face more challenges from
demand uncertainty than most other industries
because of the inherent properties of the products:
short lifecycles, ever-changing design trends and
unstable demand (To & Leung, 1999-1, 1999-2,
2000). For example, the renowned fast-fashion
company Zara, designs, manufactures, distributes
and sells merchandise at the speed of fashion fads.
It ships new merchandises every few days and has
established the 2 weeks magic by creating a new
design which is ready to sell in the retail sales
floor in just 15 days (Tagliabue, 2003).
In order to achieve fast fashion, the supply chain
needs to be highly flexible and agile (Christopher,
2004).

The solution to the agile supply chain involves
setting up seamless or boundary-less connections
between supply chain channel members. These
connections can minimize buffers between the
different stages in the chain (Levy, 2003).

The connections and integrations can be created
both internally among departments and externally
with other companies. The popular ERP systems
(Laudon & Laudon, 2006) provide a way for
achieving a high degree of integration which
makes this idea possible. When ERP is utilized in
the fashion business, it relies highly on effective
forecasting.


* Corresponding author. Tel.: (852) 2766-6428; Fax: (852) 2773-1432
E-mail address: tcaukf@inet.polyu.edu.hk
RJTA Vol. 12 No. 3 2008

57
Information systems for fashion and apparel
enterprises have undergone a path of great
evolution: from material requirements planning
(MRP), manufacturing resources planning
(MRPII), just-in-time (J IT) production and
flexible manufacturing systems (FMS) to
enterprise resources planning (ERP) (Zhu, 2002).

In the 1990s, the internet and business process
re-engineering (BPR) was used in conjunction
with each other. They have emerged as important
tools which give many fashion companies the
leading competitive edge. Today, more and more
enterprises in fashion and apparel are integrating
their legacy systems with internet-enabled
applications and changing their business processes
to achieve better integration. One of the most
essential elements of the enterprise system is the
ability to forecast the future situation based on
previous knowledge, such as predicting future
sales and demand based on past historical data. An
effective sales forecasting scheme can help
increase productivity and lower inventory. This
will result in lower inventory levels, quicker
response to changes (Choi & Chow, 2008), better
pricing schemes (Choi, 2007) and achieve the
objective of J IT delivery (Au & Chan, 2002; Choi,
2006).

As information technology is widely adopted, the
fashion industry is now in an information-rich
environment (Brannon, 2005).

The marketers can enjoy up-to-the-minute
continuous flow of sales information by installing
point-of-sale (POS) scanners in retail stores
(Fisher, 1994).

The availability of POS data makes it possible to
identify smaller market changes; flexible
manufacturing makes it possible to produce
smaller quantities of product efficiently enough to
earn a profit. Owing to global competition, fashion
companies need to have faster product
development and create more products/extended
brands to react to market competition. Here, more
new products could mean that each one of them
tends to require a shorter lifetime. This result
implies that the sales history of many products is
short and forecasts have to be carried out based on
POS data in very short time. This is a great
challenge to the traditional forecasting system.
Moreover, as there are a huge number of products
and SKUs, the forecasting process (including
inputs and outputs) is no longer centralized work
by one or two individuals, but distributed in many
places. For example, the POS data acquisition may
occur globally and forecasting may be carried out
frequently by markers of different levels of the
companies in different places. As a consequence,
to come up with a solution, a web-based
forecasting system which can aggregate data from
POS systems across the world and distribute
forecasting to people anywhere would be a
promising choice.

In this paper, we explore an implementation
scheme of a web-based forecasting system, which
can be used to produce forecasting from POS data
for fashion companies. A case analysis of neural
network forecasting is discussed, while other
forecasting methods can easily be incorporated
into our proposed implementation mechanism.
Section 2 presents a brief introduction on the
neural networks as forecasting method. Section 3
discusses the case and results. An elaboration of
the system structure is given in Section 4. We
conclude in Section 5.

2. Forecasting Methods

Forecasting is a crucial step in effective and
efficient planning for fashion operations because it
can help ensure that an effective use of resources
is made. Nowadays, fashion companies are in
need of new forecasting procedures, which will
put them in a position to produce more accurate
forecasts. It is well-known that poor forecasts lead
to inefficient capital management (Fry, 2005).

In particular, the opportunities created by the use
of a new and more accurate forecasting method
are plenty and at the same time, substantial for
enhancing the efficiency of many operations in a
fashion company. For instance, a more accurate
forecast on the companys monthly demand will
lead to a better stocking policy, high inventory
fill-rate (and hence, lower chances of stockout),
more efficient warehouse management and better
product allocation to the companys stores.
Planning for the ordering decisions requires
precision to reduce deviations from the final sales
numbers. Accurate forecasting can help with more
precise ordering which helps improve business
profitability.

RJTA Vol. 12 No. 3 2008

58
It is clear that managers can improve resource
planning by understanding the limitations of
forecasts. These limitations are exemplified
through several strategic forecasting paradoxes
that managers should recognize (see Wacker and
Lummus (2002) for some more discussions). For
example, the linear forecasting methods are
accurate when the demand is somewhat linear, but
of course, not so accurate when the demand is
nonlinear. In recognizing these paradoxes,
managers can devote their attention to improving
the use and implementation of the forecast for
better resource decisions. There are a number of
forecasting methods or models that are available
to decision makers and the appropriate selection of
the optimal forecasting scheme requires a number
of considerations. For example, if a manager
believes that the future facing his company is
predictable or fairly predictable, then statistical
forecasting is a useful tool. On the other hand, if a
company faces a very turbulent environment
where the future demand is mostly unpredictable
and highly volatile, then there is very little reason
in attempting to utilize pure statistical techniques
to forecast the future demand. In dealing with the
demand uncertainties, there is the need to
recognize the importance of forecasting activities
in the configuration and effective utilization of a
companys supply chain. Modeling and
forecasting seasonal data is one of the major
research efforts in the past few decades. Although
traditional methodologies of sales forecasting,
such as time series, auto-regression (AR) and
integrated moving average models have been
proven effective under different conditions, they
have many shortcomings (such as performing
badly when nonlinear features appear in the
demand patterns), and fail in an environment in
which the sales amounts are highly dependent on
exogenous variables, such as size, price, colour,
climatic data, effect of media and price changes.

It is well-known that well-developed artificial
intelligence (AI) models have more flexibility and
can be used to estimate various non-linear
relationships. Therefore, a lot of researchers tend
to apply AI models to study all kinds of
forecasting problems. For example, Frank and
Garg (2003) proposed the use of an artificial
neural networks (ANN) approach to forecast
womens apparel sales and the experimental
results show that the ANN approach featured sales
forecasting models that outperform the other two
conventional models; namely, the single seasonal
exponential smoothing model and Winters three
parameter model. Chang et al. (2004) proposed a
hybrid model for stock price forecasting by
integrating multiple regression, Back propagation
neural network and an autoregressive integrated
moving average model. Versace et al. (2004)
evaluated the performance of a heterogeneous
mixture of neural network algorithms for
predicting the exchange-traded fund DIA. A
genetic algorithm was utilized to find the best
mixture of neural networks, the topology of
individual networks in the ensemble and to
determine the features set. Chang et al. (2005)
developed an evolving neural network (ENN)
forecasting model by integrating genetic
algorithms (GA) and ANN for sales forecasting.
Their experimental result shows that the
performance of ENN is superior to traditional
statistical models and back propagation (BP)
networks. Chang and Lai (2005) proposed a
hybrid system to combine the self-organizing map
(SOM) of neural networks with the case-based
reasoning (CBR) method for forecasting the sales
of some newly released books. Other approaches,
such as the extreme learning machines, have also
been studied (see Sun et al. 2007 for more details).

The Evolutionary Neural Network (ENN) is a
hybrid combination of evolutionary computation
and neural networks. ENN is a suitable candidate
for topology design and widely adopted (see Falco
et al., 1998; Chen et al., 1999 for more
discussions). Unlike most of the ENN approaches
which use indirect parametric representations (and
encode factors, such as the number of input and
hidden nodes, and input nodes separately), an
ENN approach with a direct binary representation
to every single neural network connection is
proposed in this paper for time series forecasting
(Cortez et al., 2006). With the automatic pruning
of the hidden and input nodes with no connection,
the number of input and hidden nodes, and input
nodes is not encoded directly in this approach. The
Bayesian Information Criterion (BIC) is used as
the fitness function to speed-up the searching
process of ENN as in Cortez et al. (2006), and a
pre-searching mechanism is also proposed to
facilitate searching.

In order to perform time series modeling with a
multilayer perception, a set of training cases is
created from the data by adopting a sliding time
RJTA Vol. 12 No. 3 2008

59
window. In the literature, this combination is
called time lagged feed-forward network
(Sandberg 2001). A sliding window is defined by
the sequence <k
1
, k
2
, , k
i
>, for a network with i
inputs and time lags. In this paper, a basic fully
connected topology will be adopted, with one
hidden layer, bias and direct shortcut connections,
from input to output nodes. To enhance
nonlinearity, the sigmoid activation function is
applied on the hidden nodes, while in the output
node, the linear function is used instead to scale
the range of the outputs (since the sigmoid
function has a [0, 1] output range). This
architecture, often adopted by the forecasting
community, avoids the need of filtering
procedures, which may give rise to information
loss. Therefore, the general model is given in the
following form (see Cortez et al., 2006 for more
details):

1
,0 , , ,0 ,
1 1 1
( )
i i
I Out I
t
Out t k Out i t k i j j Out j
i j I i
x w x w f x w w w

= = + =
= + + +


(1)

where
j i
w
,
denotes the weight of the connection
from node j to i (if j = 0 then it is a bias
connection), Out denotes the output node, f is
the sigmoid function ( ) 1 /( 1
x
e

+ ), Ki denotes the
lag of the input i, and i is the number of input
neurons. Figure 2.1 gives an example of a fully
connected multilayer network.

1
x
2
x
, i j
w
,0 i
w

Fig. 2.1 A fully connected multilayer network with 2 inputs, 2 hidden neurons, 1 output, bias and shortcut
connections

The genetic algorithm was first presented by
Holland (1973) and since then, the concept has
evolved into several computational variants that
share a set of common features. There are a
number of potential solutions (individuals) to a
problem, evolving simultaneously (a population).
Each individual is coded by a string (chromosome)
of symbols (genes), taken from a well-defined
alphabet. Each individual is assigned a numeric
value (fitness) that indicates the solutions
appropriateness. In every generation, a fraction of
the individuals is replaced by the offsprings,
which are generated by the application of genetic
operators, such as crossover and mutation, in order
to create new solutions (reproduction). The whole
process is evolutionary, where the fittest
individuals have greater chances of survival.
Cortez et al., (2006) employed an evolutionary
neural network approach with a direct binary
representation in time series forecasting. Since
statistical analysis has shown that many time
series are modeled by small networks, this
encoding can prevent the topologies from growing
too large. The connection between the hidden
nodes and the output ones always exists, since this
strategy enhances the creation of valid networks.

Apparently, the maximum neuron number H in the
neural network is a crucial parameter in this
evolution approach. When H is large, this
approach cannot always successfully identify a
simple structure network for the forecasting even
with the guide of BIC. Concerning the
computational performance, the evolution process
RJTA Vol. 12 No. 3 2008

60
from a maximum 7 neurons network to a 2
neurons network can be very time consuming. To
overcome these limitations, we propose an
improved forecasting scheme. It is obvious that
when the neuron number in a network is too large,
it will overfit a time series, the network will then
fail to forecast the future of the sales time series
correctly. In our approach, a pre-search
mechanism is devised to find an appropriate
maximum hidden neuron number H to start the
evolution. By this approach, the evolutions
process performance is expected to be improved.

3. Empirical Results

In this work, the time series of real sales data of a
fashion product from a fashion company has been
extracted for analysis. The analysis follows our
previous work in Au et al. (2007). Figure 3.1
illustrates the time series, which consists of the
number of daily goods sold in units for the apparel
item. The time series covers the whole life cycle in
one season of the product. Its life cycle lasts for
around 70 days. As we all know, fashion sales
figures are, by its nature, hard to forecast, since
there are too many factors which can influence
them. The situation is especially hard for this data
set since the sales data are collected from no more
than 10 shops. Owing to the small sample size, the
fluctuation of sales is large which we can observe
from Figure 3.1 and the random factors play a
more influential role than the cases when we have
a bigger sample size.

0 10 20 30 40 50 60 70
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Day
S
a
l
e
s

(
n
o
r
m
a
l
i
z
e
d
)

Fig. 3.1 Time series of apparel item As sales in a
season

In the forecasting process, as four-week data
points in the time series are needed in the training
and validation for the neural network, so
forecasting starts from the fifth week. Table 3.1
(Yu et al., 2007) lists the results of the forecasting
of weeks 5 8. Figure 3.2 illustrates the curves of
forecasted and actual time series during the
training, testing and forecasting period for the 7th
week.

Table 3.1 The forecasting results
Forecasted
Week
MSE of ENN
Max. 14 inputs
Hidden
Neuron
Number
Input
Window
(Dates)
5 0.0109 1 <2 3 6 7
10 13 14>

6 0.0126 1 <2 3 5 6 7
9 10 12>

7 0.0062 1 <1 2 4 5 6
7 13 14>

8 0.0039 1 <1 2 3 4 5
7 13>



Fig. 3.2 Forecasted (Fitted) and actual curve for
the 7
th
week

As shown in Table 3.1, the forecasting error
generally becomes smaller for the weeks which
are late in the temporal sequence. This is because
an increasingly later forecasting target results in
more data being used in the training of the neural
network and the forecasting can be more accurate.
For all the results, the best structure is simple,
with only 1 hidden layer neuron. The number of
input varies for these results. Generally, the input
lags from day 1 to 7 (one week earlier) are used
more frequently in the structures and the input lags
from day 8 to 14 (two weeks earlier) are used less
frequently, while days 13 and 14 appear more
frequently than others. Figure 3.2 depicts the
RJTA Vol. 12 No. 3 2008

61
situation of the forecasting for week 7. The first 3
weeks (week 3-5) are used for the training, week 6
for testing and week 7 for forecasting. Since we
need at least 2 weeks of data for the input, no
forecasting can be carried out for weeks 1 & 2.
Figure 3.2 illustrates that the best testing result is
reached after the evolution. As we can expect, the
fitted curves of the testing and forecasting periods
are close to the actual ones, while the curve for the
training period is not close. This is a common case
in forecasting, in which a well fitted curve to the
actual one during the training period may imply
overfitting and bad forecasting often occurs in that
situation.

Forecasting using ENN can be time consuming,
such as in our tests. It often costs tens of minutes
for the whole process of ENN model identification
and forecasting to be completed. Thus, the users
can first submit the forecasting request and input
the time series, and leave it to the system for
performing the forecasting. They can observe the
progress of the forecasting at any time and request
for the result after the forecasting is finished.

4. System Structure

A web-based forecasting system is proposed based
on the above methodology for forecasting. Figure
4.1 illustrates the structure for implementing the
system. The core of the forecasting system is a
web service object and support standard SOAP
interface both to the users using web client such as
IE, and to other enterprise services, such as ERP
(enterprise resources planning) and SCM (supply
chain management) services, through which these
direct users or services can send forecasting
requests and retrieve forecasting results from the
proposed forecasting system.

Internet or Intranet
Web Client Interface (ASP, JSP etc.)
Time Series Input Request for the forecasting result Forecasting Results Output
Web Sever Objects (SOAP interface)
ENN model
identification
procedure
(COM object)
Forecasting
using the NN
model
(COM object)
The selected
NN model
Request & response database
Other enterprise
services (ERP, SCM etc)
Internet or Intranet
Forecasting Results Output
Time Series Input
Request for the forecasting result

Fig. 4.1 The system structure of the proposed web-based forecasting scheme
RJTA Vol. 12 No. 3 2008

62
The SOAP (Mendelsohn et al., 2000) protocol is
adopted in our web-based forecasting system as it
offers an extremely flexible interface to other
applications, such as the end-user can view and
interact with web pages which are composed of
almost any web page script languages, such as
ASP or J SP. All these script languages support the
SOAP interface and can exchange information
with our forecasting server object. Not only is
SOAP flexible for web applications, it is also
flexible when working with other traditional
applications, such as the enterprise applications or
customized applications which are composed by
COBOL, C or C++. All such contemporary
programming languages support SOAP and can
invoke SOAP objects easily. By applying the
SOAP protocol, our web service objects can easily
be used either by end-users before a web client
(such as IE), or by other enterprise applications
which have knowledge on the specifications of our
SOAP services.

SOAP originally stood for Simple Object Access
Protocol and lately, also Service Oriented
Architecture Protocol, which was originally
designed by Dave Winer, Don Box, Bob Atkinson,
and Mohsen Al-Ghosein in 1998 with backing
from Microsoft (where Atkinson and Al-Ghosein
worked at the time) as an object-access protocol.
SOAP makes use of an internet application layer
protocol as a transport protocol. HTTP has gained
wider acceptance as it works well with today's
internet infrastructure. SOAP works well with
network firewalls. SOAP may also be used over
HTTPS (since it is the same protocol as HTTP at
the application level, but using an encrypted
transport protocol underneath) in either simple or
mutual authentication. This is a major advantage
over other distributed protocols, such as
GIOP/IIOP or DCOM, which are normally filtered
by firewalls. XML is chosen as the standard
message format because of its widespread use by
major corporations and open source development
efforts. Additionally, a wide variety of freely
available tools significantly ease the transition to a
SOAP-based implementation. By employing the
SOAP interface in the forecasting service, the
service can be utilized by applications across the
globe provided that web access is available,
regardless if the service is built on Windows or
Macintosh, in C programming language or J ava
etc.

While SOAP is used for the outer interface of the
forecasting service, COM is used for the interface
of the service since the SOAP interface is
considerably slower than COM and unsuitable for
internal forecasting computation where the
performance of the program is crucial.

Component Object Model (COM) is a platform for
the software component introduced by Microsoft.
It is used to enable inter-process communication
and dynamic object creation in any programming
language that supports the technology. The
essence of COM is a language-neutral way of
implementing objects such that they can be used in
environments different from the ones they were
created in, even across machine boundaries. For
well-authored components, COM allows reuse of
objects with no knowledge of their internal
implementation because it forces component
implementers to provide well-defined interfaces
that are separated from the implementation. The
different allocation semantics of languages are
accommodated by making objects responsible for
their own creation and destruction through
reference-counting. Casting between different
interfaces of an object is achieved through the
QueryInterface() function. The preferred method
of inheritance within COM is the creation of
sub-objects (called aggregation) to which method
calls are delegated.

In our forecasting system, the ENN algorithm is
implemented based on the Matlab NN toolbox, so
the algorithm is complied into a COM object from
the Matlab language. The SOAP interface is
implemented using C++programming language.
Upon receiving the request from the SOAP
interface, the C++ program invokes the related
COM objects and manages the running logic of
them, the forecasting result from the COM object
is returned to the C++ program and the C++
program translates the result and outputs to the
SOAP interface or puts the result in a database for
a delayed result request. The original time series,
and the request and response can all be stored in
the database for later use. Since the forecasting
process can be lengthy, users can submit their
forecasting requests while retrieving the results at
a later time by sending a delayed result request.
The final result, as the output, will be shown via
the standardized SOAP interface. Not only the
direct users from the web, but also other
applications (ERP, SCM) can use this service.
RJTA Vol. 12 No. 3 2008

63
5. Conclusion

Forecasting is an essential tool for planning and
decision making in the fashion business.
Understanding the role of forecasting in its
organizational and managerial contexts is very
important for a company in todays business. An
effective way of implementing forecasting is to
establish a web-based decision support system.
Web-based forecasting facilities the forecasting
process as inputs can be obtained via the network
from the sources in real time. This paper has
demonstrated the way that the proposed ENN
forecasting model can be successfully
implemented as a web-based forecasting system.
The user can upload the required data sets into the
system or auto-feed the POS data into the system
and get the forecasted time series. In addition, the
proposed forecasting system can be an important
module of an enterprise system which supports
various other vital functions of a fashion company.
The marketing of the fashion apparel industry has
been leading the trends of the marketing field,
with the help of technologies, such as web based
forecasting systems in the information-rich
environment. The goal of real-time marketing
will be realized in the future.

REFERENCES

[1] Au, K.F. & Chan, N.Y. 2002, Quick
response for Hong Kong clothing suppliers:
a total system approach, Proceedings of the
13
th
Annual Conference of the Production
and Operations Management Society, San
Francisco, USA.
[2] Au, K.F., Choi, T.M. & Yu, Y. 2007, Fashion
retail forecasting by evolutionary neural
networks, International Journal of
Production Economics, Special Issue.
[3] Brannon, E.L. 2005, Fashion Forecasting,
Second Edition.
[4] Chang, P.C. & Lai, C.Y. 2005, A hybrid
system combining self-organizing maps with
case-based reasoning in wholesaler's
new-release book forecasting, Expert
Systems With Applications, vol. 29, no. 1, pp.
183-192.
[5] Chang, P.C., Wang, Y.W. & Tsai, C.Y. 2005,
Evolving neural network for printed circuit
board sales forecasting, Expert Systems With
Applications, vol. 29, no. 1, pp. 83-92.
[6] Chang, P.C., Wang, Y.W. & Yang, W.N. 2004,
An investigation of the hybrid forecasting
models for stock price variation in Taiwan,
Journal of the Chinese Institute of Industrial
Engineering.
[7] Choi, T.M. & Chow, P.S. 2008,
Mean-variance analysis of quick response
program, Forthcoming in International
Journal of Production Economics, available
at sciencedirect.com http://dx.doi.org/
10.1016/j.ijpe.2007.06.009.
[8] Choi, T.M. 2006, Quick response in fashion
supply chains with dual information
updating, Journal of Industrial and
Management Optimization, vol. 2, pp.
255-268.
[9] Choi, T.M. 2007, Pre-Season stocking and
pricing decisions for fashion retailers with
multiple information updating, International
Journal of Production Economics, vol. 106,
pp. 146-170.
[10] Christopher, M., Lowson, R. & Peck, H.
2004, Creating agile supply chains in the
fashion industry, International Journal of
Retail & Distribution Management, vol. 32,
no. 8, pp. 367-376.
[11] Cortez, P., Rio, M., Rocha, M. & Sousa, P.
2006, Internet traffic forecasting using
neural networks, International Joint
Conference on Neural Networks, pp. 2635-
2642.
[12] Falco, I.D., Cioppa, A.D., Iazzetta, A. &
Natale, P. 1998, Optimizing neural networks
for time series prediction, Third World
Conference on Soft Computing (WSC3).
[13] Fisher, M. L., Hammond, J . H., Obermeyer,
W. R. & Raman, A. 1994, Making supply
meet demand in an uncertain world,
Harvard Business Review, vol. 72, no. 3, pp.
83-89.
[14] Fry, E. 2005, Big Ideas, CFO Magazine, pp.
1-5.
[15] Holland, J . 1973, Genetic algorithms and
the optimal allocation of trials, SIAM
Journal on Computation.
[16] Laudon, K.C. & Laudon, J .L. 2006,
Management Information Systems, 9
th

Edition, Prentice Hall.
[17] Levy, S. 2003, The connected company,
Newsweek, vol. 41.
[18] Mendelsohn, N., Nielsen, H.F., Thatte S &
Winer D 2000, Simple Object Access
Protocol (SOAP) 1.1, World Wide Web
Consortium (W3C) http://www.
RJTA Vol. 12 No. 3 2008

64
w3.org/TR/SOAP.
[19] Nikolopoulos, K., Metaxiotis, K. &
Assimakopoulos, V. 2003, A first approach
to e-forecasting: a survey of forecasting Web
services, Information Management &
Computer Security, vol. 11, no. 3, pp.
146-152.
[20] Sandberg, I.W. 2001, Nonlinear Dynamical
Systems: Feedforward Neural Network
Perspectives, Wiley-IEEE.
[21] Sun, Z.L., Choi T.M. & Au, K.F. 2007, A
hybrid neuron-fuzzy inference system
through integration of fuzzy logic and
difference evolutionary extreme learning
machines, IEEE Transactions on Systems,
Man, and Cybernetics, Part B Cybernetics,
37, 1321 1331.
[22] Tagliabue, J . 2003, Spanish clothing chain
Zara grows by being fast and flexible, New
York Times, [Web], pp. 1-2. Available:
www.nytimes.com (2003, J uly 7).
[23] To, K.M.C. & Leung C.S. 1999-1,
Monitoring Service Gaps in Hong Kong
Apparel Merchandising Business Part I:
Examining Customer's Expectations of
Adequacy and Superiority in Hong Kong
Apparel Merchandising Business, Research
Journal of Textile and Apparel, vol. 3, no. 1,
pp. 65-79.
[24] To, K.M.C. & Leung C.S. 1999-2,
Monitoring Service Gaps in Hong Kong
Apparel Merchandising Business Part II: A
study of Holistic Quality-oriented Service,
Research Journal of Textile and Apparel, vol.
3, no. 2, pp. 72-78.
[25] To, K.M.C. & Leung C.S. 2000, Monitoring
Service Gaps in Hong Kong Apparel
Merchandising Business Part III: An
Antecedents study of Prganizational Service
Quality Incongruence, Research Journal of
Textile and Apparel, vol. 4, no. 2, pp. 52-59.
[26] Versace, M., Bhatt, R., Hinds, O. & Shiffer,
M. 2004, Predicting the exchange traded
fund DIA with a combination of genetic
algorithms and neural networks, Expert
Systems With Applications, vol. 27, no. 3, pp.
417-425.
[27] Wacker, J . & Lummus, R. 2002, Sales
forecasting for strategic resource planning,
International Journal of Operations &
Production Management, vol. 22, no. 9, pp.
1014-31.
[28] Zhu, S. & Li , W. 2002, Research on the
System Framework of ERP, System
Engineering.