He Huang1, 3,

Qiurui Liu2
1
College of Fashion and Design,
Donghua University,
Shanghai, China
E-mail: huanghedhu@gmail.com;
2
College of Public Administration,
Huazhong Agricultural University,
Wuhan, China
E-mail: qiurui.liu@hotmail.com
3
Faculty of Management,
McGill University,
Montreal, Canada
GENERAL PROBLEMS OF THE FIBRE AND TEXTILE INDUSTRIES
Intelligent Retail Forecasting System for New
Clothing Products Considering Stock-out
DOI: 10.5604/12303666.1227876
Abstract
Improving the accuracy of forecasting is crucial but complex in the clothing industry, espe-
cially for new products, with the lack of historical data and a wide range of factors affecting
demand. Previous studies more concentrate on sales forecasting rather than demand foreca-
sting, and the variables affecting demand remained to be optimized. In this study, a two-stage
intelligent retail forecasting system is designed for new clothing products. In the first stage,
demand is estimated with original sales data considering stock-out. The adaptive neuro fuzzy
inference system (ANFIS) is introduced into the second stage to forecast demand. Meanwhile
a data selection process is presented due to the limited data of new products. The empirical
data are from a Canadian fast-fashion company. The results reveal the relationship between
demand and sales, demonstrate the necessity of integrating the demand estimation process
into a forecasting system, and show that the ANFIS-based forecasting system outperforms
the traditional ANN technique.

Key words: intelligent forecasting system, demand estimation, stock out, adaptive neuro
fuzzy inference system, new clothing product.

Introduction
For the clothing industry, it is critical to
improve the accuracy of demand fore-
casting [1]. A good forecasting system
could avoid problems such as lost sales,
inadequate inventory level, inferior cus-
tomer service, etc. However, four main
factors increase the difficulty in cloth-
ing forecasting: (1) The clothing supply
chain is considered inflexible and com-
plex with many sectors and companies;
[2] it results in the high sensitivity of
the bullwhip effect; (2) a large number
of new products introduced to each col-
lection leads to extensive historical data
for past products, but little data for new
products; (3) The sales data of clothing
products is usually noisy and affected by
multiple factors, such as weather, pro-
motion, macroeconomics, etc. [3]; (4)
acquiring an accurate demand pattern is
not easy because of stock-out, thus using
historical sales data instead of demand
data to implement forecasting is a com-
mon way.
To explore the issue of clothing retail
forecasting, some papers have investigat-
ed statistical methods, while others have
studied artificial intelligence (AI) tech-
niques. In recent years, hybrid techniques
have aroused the interest of experts.
Among the various hybrid techniques,
the adaptive neuro fuzzy inference sys-
tem (ANFIS) is a competent alternative
that combines the advantages of the ar-
tificial neural network (ANN) and fuzzy
system [4]. In addition, previous studies
heavily focus on developing forecasting
models/techniques, but ignore dataset
processing. In fact, the purpose of retail
forecasting is to satisfy consumer de-
mand. Therefore integrating the demand
estimation process into the forecasting
system should be taken into account.
As a result, this study aims to design an
intelligent system for the demand fore-
casting of new clothing products. This
system comprises of a demand estima-
tion stage and demand forecasting stage.
The empirical data are from a Canadian
fast-fashion company.
Forecasting techniques
in the clothing industry
As time goes on and computer technolo-
gy develops, more and more forecasting
models/techniques have been created for
the clothing industry.
Traditional techniques
An inspection of previous studies reveals
that classical statistical methods are ex-
tensively used [5-7]. However, most
methods such as regression models are
only efficient for seasonal or cyclical
data [8], and limited to a linear struc-
ture. Moreover the demand for clothing
products is affected by numerous factors,
so that statistical methods may not be
capable of achieving a satisfactory re-
sult. Moreover a substantial amount of
data are often required to generate good
results, but the data available for new
clothing products are limited. Therefore
AI techniques have gained more atten-
tion [9-11]. Despite the fact that AI tech-
niques could obtain satisfactory results
in many studies, it is not easy to achieve
a balance between result accuracy, high
computational speed and system sta-
bility. Consequently to combine the ad-
vantages of different techniques, hybrid
techniques were investigated.
Hybrid techniques
A great quantity of studies have presented
that hybrid techniques lead to better per-
formance. As a hybrid technique, ANFIS
combines the knowledge representation
ability of the fuzzy system and the learn-
ing ability of ANN. Consequently ANFIS
has been widely used for forecasting in
different industries [12-14]. However,
there is not an adequate quantity of stud-
ies in the literature on retail forecasting
with ANFIS in the clothing industry.

10 Huang H, Liu Q. An Intelligent Retail Forecasting System for New Clothing Products Considering Stock-out.
FIBRES & TEXTILES in Eastern Europe 2017; 25, 1(121): 10-16. DOI: 10.5604/12303666.1227876
 Method and dataset
Proposed system and data selection
As an important information flow, de-
mand information greatly influences the
performance of clothing retail compa-
nies. However, unlike other areas where
the past is a great predictor of the future,
The data selection process is explained
as follows:
n Based on the concept of ANFIS and
data splitting theory, two product
groups are identified: group 1 has 56
items with full-year data from 2014,
and group 2 has 8 items with data
from several months in 2015.
it indicates that stock-out may happen.
We also realize that the demand on each
day of the week varies a lot. For instance,
sales at the weekend are usually stronger
than those on weekdays. Then each day
of the week has different sales weight.
We express the weekly demand of each
item D as below:
m
k

such as grocery or hardware retail, there n All products belong to the same prod-
D = ∑ D       (1)
m T m
is no amount of historical data or experi- uct department, and there are three k t =1 tk

ences that can truly and effectively pre- product classes in group 1. 7 7

D tk = ∑ ( S tkd ∗V tkd ) / ∑ (W d ∗V tkd )   (2)

m m m m
dict how customers would react to new n With the help of company experts, d =1 d =1
clothing products. This reality is further products in group 1 are selected as
made complex by the fact that demand comparable items to products in W d
= S 'd / ∑ d =1 S 'd     
7
(3)
is affected by numerous factors, and it is group 2. Indeed, given the limited
difficult to acquire the demand pattern be- data for new products, many fashion Where:
cause of stock-out. In addition, tradition- companies consider the performance D tk – the demand of item m in store t in
m

al techniques have certain disadvantages of comparable items, such as ZARA week k. m = (1,2,…,M); t = (1,2,…,T);
such as an inability to address a complex [15]. k = (1,2,…,K).
non-linear structure, a need for more his- n For each item, daily data from each D – the demand of item m in week k
m

torical data, the production of relatively store are collected rather than aggre- which aggregates the demand of all
worse results, etc. In this section, a two- gated data. stores.
stage intelligent retail forecasting system S tkd – sales quantity of item m in store t
m

that could remedy the shortcomings of
traditional techniques is proposed, see
Figure 1. In the first stage, demand is es-
timated with original sales data consider-
ing stock-out. In the demand forecasting
stage, ANFIS is employed and the sig-
nificant variables affecting demand are
selected as the ANFIS inputs. Given the
data of new clothing products is limited,
a data selection process is integrated in
this system.
To validate the effectiveness of this
system, we applied it to a Canadian
fast-fashion company. The company has
hundreds of stores across North Amer-
ica and a large Asian sourcing base.
In addition, the life cycle of fast-fashion
products is short, usually 6-10 weeks
[16]. In our database, the products in
group 2 include a complete product life
cycle of 10 weeks. Therefore, in this
study, the forecasting time period is 10
weeks with a weekly horizon.
Demand estimation
The occurrence of stock-out results in in-
equality between sales and demand [17].
In other words, if there is no stock-out,
sales quantity can be regarded as de-
mand quantity. Therefore the basic idea
of stage one is using the sales data with-
out stock-out to deduce demand. In our
database, if the daily inventory is zero,
on day d of week k.
V tkd – the binary variable equals 0 when
m
stock-out may exist (daily inventory
= 0), otherwise 1 (daily inventory > 0).
W d – salces weight of day d.
d – index of the day, d = (1,2,…,7),
Monday is 1 and Sunday – 7.
S’d – aggregated sales quantity on day d,
used to compute Wd.
The basic idea behind this approach is
consistent with one of the methods dis-
cussed in [18], although our approach is
slightly different. To generate the sales
weights, the daily sales of group 1 are
aggregated by each day of the week, then
the aggregated sales are divided by the

Stage 1: Demand Estimation

Demand
Original Demand
Estimation
Sales Data Data
Model
Actual
Data Selection

Value
Variable 1
ANFIS

Variable 2 Fuzzy Predicted

Artificial Value
Factors affecting Inference
Neural Compare Demand
demand System
…… Network
(Sugeno)

Variable N
Parameters adjustment

Stage 2: Demand Forecasting

Figure 1. Schematic of the intelligent retail forecasting system.

FIBRES & TEXTILES in Eastern Europe 2017, Vol. 25, 1(121) 11

weights, the daily sales of group 1 are aggregated by each day of the week, then the

weekly
ggregated sales are divided by the total sales of the whole weekly demand
demand
week, see obtained
obtained
Equation instead of
3. instead of weekly
weeklysales
saleswas
wasused
usedasasthe
theinputs
inputsofofthe
thesecon
seco

Figure 2 shows the results. stage.

stage.
25% 25%

23.51% 20%
20%
19.66%

Sales weight
15%
15%
Sales weight

15.33%
10%
10% 12.44%
10.14% 10.65% 5%
8.27%
5%
0%
MON. TUES. WED. THURS. FRI. SAT. SUN.
0% Day
MON. TUES. WED. THURS. FRI. SAT. SUN.
Day Class A Class B Class C

Fig.
Figure 2. Sales 2 Salesfor
weights weights for each
each day of theday of the week
week. Fig.33Sales
Figure 3. Sales Fig. Sales
weights weightsfor
for weights
different for differentclasses
different
classes. classes

The three days with the highest sales are Saturday, Friday
3.3.and
3.3. Thursday,
Demand
Demand with
forecasting
forecasting
total sales of the whole week, see Equa- Equation 1. The weekly demand obtained ployed, which has become common prac-
ales weights of 23.51%, 19.66%, and 15.33%, respectively. Surprisingly
ItIt is
is sales
well sales on
well was
known
known that AIin-
techniques, suchas
asimplementation
thefuzzy
fuzzysystem
system and ANN,have
havetht
tion 3. Figure 2 shows the results. instead of weekly usedthat AI
as the techniques,
tice in such
ANFIS the and
[12]. ANN,
All
puts of the second stage. functions were carried out in MATLAB.
Sunday are relatively
The three low.
days The
with main reasonsales
the highest might
are be that retail stores of
capability
capability have
of shorterhuman
imitating
imitating human reasoning.
reasoning.ANN
ANNhas
hasaastrong
strongcapability
capabilityofoflearning
learningwit
w
Saturday, Friday and Thursday, with sales Demand forecasting The structure of a typical ANFIS is
usiness hoursweights of 23.51%,
on Sunday 19.66%,
in North and 15.33%,
America. It is well
Then another known
question that
is
parallel data. AI techniques,
whether
data. However, the
However, given such shown in is
given knowledge
knowledge Figure 4. For into
is integrated
integrated ease of
into thetheillustra-
wholenetwork
networkana
parallel whole
respectively. Surprisingly sales on Sunday as the fuzzy system and ANN, have the tion, it is assumed that each variable is
are relatively low. The main reason might capability of imitating human reasoning. fuzzed by two fuzzy sets. The node func-
ales weights observed would vary depending on the products cannotin different
be broken
broken classes,
up into
intolearn-
individual parts; the network isis essentially
essentially
be that retail stores have shorter business ANN has cannot be
a strong capabilityup of individual tionsparts;
of eachthe network
layer are explained below: aablack-box
black-bo
hours on Sunday in North America. Then ing with parallel data. However, given
ven if they belong
anothertoquestion
the sameis department.
whether theWe then examined
sales knowledgeThethe is sales weights
integrated
fuzzy system intoisfor
the
good whole
at reasoning
reasoning with the linguistic information i obtainedfrom
fro
fuzzy system is good at ANFISwithinputs thearelinguistic
represented as neuron xobtained
information
weights observed would vary depending network and cannot be broken up into in- (i = 1, 2, …, n). Here, they are the factors
on the
he three classes products
of group 1. in
Asdifferent
expected,classes, even quantity
the sales dividual parts;class
ofexpert
each the network is essentially affecting demand. Each neuron simply
is different,
expert knowledge.
knowledge. However, the
However, the learning
learningcapability
capabilityisisinferior inferiorandandcannot
cannotadapt
adaptitse
its
if they belong to the same department. a black-box. The fuzzy system is good at passes external crisp signals to Layer 1.
ut the sales We then examined
weights the salesconsistent,
are remarkable weights forsee Figure
reasoning 3. with the linguistic that
It demonstrates information
the three classes of group 1. As expected, obtained to from
a newnew environment
environment
expert knowledge.[19]. [19].
How- As aaLayer
As hybrid
hybrid technique, first
technique,
1 (Fuzzification first proposed
proposed
Layer): Neuronby Fby [20],ANFI
[20], ANF
ij
the sales quantity of each class is differ- ever, the learning capability is inferior (i = 1, 2, …, n; j = 1, 2) is the linguistic
roduct class ent,
has but
no significant impact
the sales weights areonremarkable
sales weights,
andand hence
cannot theitself
adapt salestoweights
a new environ-
combines
combines the
the learning
learning ability
ability of oflabel
ANN
ANN corresponding
with
with the to
the semanticthe fuzzytransparency
semantic set asso-
transparency ofof fuzz
fuz
consistent, see Figure 3. It demonstrates ment [19]. As a hybrid technique, first ciated with the input. The neuron func-
n Figure 2 can
thatbeproduct
used for each
class hasitem. After obtaining
no significant the salesby
im- proposed weights, the weekly
[20], ANFIS combines the tion can be modelled by the fuzzy mem-
systems. Meanwhile
Meanwhile the the domain
domain knowledge
knowledge represented
representedasasfuzzy fuzzyrulesrulesandandlinguisti
pact on sales weights, and hence the sales learningsystems. ability of ANN with the seman- bership function (MF).
linguis
emand of eachweights
item ininFigure 2 canwas
each store be used for each
computed tic transparency
by Equation 2, and thenof fuzzy systems. Mean-
the weekly
item. After obtaining the sales weights, while the variables
variables is
is integrated
domain knowledgeintegrated into
into ANFIS.
represented ANFIS. In In this
this1ijstudy,
Output = µ Fij (the
study, i ) =Sugeno-type
xthe membership fuzzy
Sugeno-type fuzzyinferenc
inferen
the weekly demand of each item in each as fuzzy rules and linguistic variables is function format, i = 1, 2, … , n, j = 1, 2.
emand of eachstoreitem
was that aggregates
computed all stores
by Equation was generated
2, and integrated by Equation
into ANFIS. In1.thisThe study,thethe Mamdani-type
system
system (FIS)
(FIS) rather
rather than
than the Mamdani-type FIS FIS isis employed,
employed, which which has has becom
becom
then the weekly demand of each item that Sugeno-type fuzzy inference system (FIS) Layer 2 (Rule Layer): Each neuron wl
aggregates all stores was generated by rather than the Mamdani-type FIS 7 is em- (l = 1, 2, …… , L) corresponds to a fuzzy
common
common practice
practice in
in ANFIS
ANFIS implementation
implementation [12].
[12]. All
All functions
functionswere
werecarried
carriedout
outi

MATLAB.
MATLAB.
Layer 1 Layer 2 Layer 3 Layer 4 Layer 5

F 11 w1 x1 xn
w1
x1 F 12 Π N y1
w1 * y1
• • • • •
Input • • • • • ∑ y Output
• • • • •

xn F n1 Π N yL wL * yL
wL
F n2 wL x1 xn

Figure 4. Structure of ANFIS.

12 FIBRES & TEXTILES in Eastern Europe 2017, Vol. 25, 1(121)


Original Price Level
Promotion Level
Size Preference
Allocation Status
Calendar Factor
Figure 5. Sugeno-type fuzzy inference systemm.
rule and functionally computes the firing
strength. The neuron function should be
a certain T-norm function, such as,
n
Output 2 l = wl = ∏ µ F i ( xi ) ,
l
i =1
l = 1, 2, … , L.
Layer 3 (Normalization Layer): Neuron
wl (l = 1, 2, … , L) determines the nor-
malised firing strength. For a given rule,
the ratio between the firing strength and
the sum of the firing strengths of all rules
represents the normalised firing strength.
L which is associated with three fuzzy
Output 3l = wl = wl ∑ w, l = 1, 2, … , L.
l sets: high, medium and low.
l =1
Layer 4 (Defuzzification Layer): Neurons
yl (l = 1, 2, … , L) compute the conse-
quent of the rules. The neuron function in
this layer is a linear function.
Output 4 l = wy l l l
l l = wl ( p 0 + p 1 x1 +…+ p nxn )
l = 1, 2, … , L.
Layer 5 (Summation Neuron Layer):
Neuron y determines the final output of
ANFIS by computing the sum of the re-
sults of all defuzzification neurons.
L L
Output 5 = y = ∑ wl ∗ yl = ∑ wl ∗ yl
l =1 l =1
As discussed earlier, neuron xi are the
inputs of ANFIS, which are the factors
affecting demand. Therefore selecting
the appropriate factor is important. Fur-
thermore it is usually difficult to analyse
qualitative variables with conventional
methods. However, ANFIS is a powerful
technique to cope with both quantitative
and qualitative variables. By interview-
ing company experts and reviewing pre-
vious literature, five factors were deter-
mined.
n Original Price Level: This variable
not only affects the demand pattern,
FIBRES & TEXTILES in Eastern Europe 2017, Vol. 25, 1(121)
Demand
Forecasting f(u)
(sugeno)
Demand Quantity
but also determines the target custom-
ers. It has three fuzzy sets: high, medi-
um and low.
n Promotion Level: Some previous
studies used “promotion” and “no
promotion” as the fuzzy sets. In fact,
for a given product, it sets the final
sales price after more than one mark-
down, and the promotion level varies
between normal weekends and big
holidays. Meanwhile different promo-
tional levels have different impacts on
customer consumption. Therefore the
“promotion level” was introduced,
n Size Preference: Unlike colour pref-
erence, which changes every year or
even every season, size preference is
relatively stable. The main reason is
that the target customers in a specific
city are stable. Three fuzzy sets were
determined: high, medium and low.
n Allocation Status: In the planning
stage, not every item is allocated to
each store. Some items are allocated
to all stores, while others are not. The
allocation strategy also affects the de-
mand pattern. Two fuzzy sets were
L
identified: normal and special.
∑w l n Calendar Factor: It is easy to observe
l =1 that holidays or big events influence
demand. For example, the sales on
Black Friday are extremely high,
whereas those on normal workdays
are low. This variable has two fuzzy
sets: normal and special.
It should be noted that although colour
preference also has a great influence on
demand, the large number of existing
colours and new ones introduced every
season make it impractical to identify the
fuzzy sets. After determining the five var-
iables, four main steps were implemented
to build a robust ANFIS.
Input Output
trapMF
triMF Constant
gaussMF Linear
gbellMF
Figure 6. Different MF combinations for
ANFIS structure.
In the first step, determining the train-
ing and validation datasets is important.
Given comparable products have similar
demand patterns, it is reasonable to use
such data rather than random product
data for forecasting. Consequently group
1 data were used as training data to build
the networks. Group 2 data were used as
validation data to validate the reliability
of the networks. Furthermore given the
variables were fuzzed by several fuzzy
sets, with the help of company experts,
numerical scales were used to convert
linguistic labels to numerical data.
In the second step, a Sugeno-type FIS
(see Figure 5) is constructed for ANFIS.
Determining the MF type of input vari-
ables is an important but complex task.
To obtain the best performance, differ-
ent experiments were prepared for four
commonly used MFs: the triangular MF
(triMF), trapezoidal MF (trapMF), gen-
eralised bell MF (gbellMF) and Gaussi-
an MF (gaussMF). The output function
has two types, constant and linear, and
experiments prepared for both functions
were conducted as well. The 8 different
MF combination experiments used to
identify the best forecasting capabili-
ty of the ANFIS structure are shown in
Figure 6.
In the training step, the ANFIS learning
algorithm is used to optimise FIS param-
eters. The hybrid algorithm, consisting
of the least squares approach and back-
propagation gradient descent approach,
was utilised. The ANFIS structure with
5 inputs and 1 output was trained for
100 epochs with a 0.01 error tolerance.
Considering the overfitting problem, the
model was run ten times for each MF
combination with varying the parame-
ter. Group 1 data were used for this pro-
cess. With different experiments, Gauss
13
Table 1. Example rules of Sugeno-type FIS. put is the value of forecasted demand.
rate coefficient η was 0.5 and the momentum μ - 0.7. The same datasets of ANFIS were used.
Input
The model with 5 inputs and 1 output was
Rule Original Promotion Size Allocation Calendar Output trained for 1000 epochs with a 0.01 error
NO.
In addition,
price level
(x1)
level
(x2) to compare
preference
(x3) the
status
(x4) forecasting
factor
(x5) capability of the system
tolerance. proposedruleand
The backpropagation was
1 Low High Medium Special Normal F1(x1, x2, x3, x4, x5) used for the training process, and all da-
2 Medium Medium Low Normal Special F2(x1, x2, x3, x4, x5) tasets were normalised between 0 and 1.
ANN,
3 two widely
High Low used accuracy
Medium Specialmeasures
Normal were
F3(x1, chosen:
x2, x3, x4, x5) theThemean
networkabsolute
was also runpercentage
ten times for
4 High Medium Low Normal Special F4(x1, x2, x3, x4, x5) each network architecture. Three hidden
5 Medium Low High Special Normal F5(x1, x2, x3, x4, x5) layers with 10-10-11 neurons were iden-
tified. The learning rate coefficient η was
error (MAPE) and mean square error (MSE). 0.5 and the momentum μ – 0.7.
Table 2. Demand and sales patterns of a specific product (unit). In addition, to compare the forecasting
4.2. Results
Week
comparison
1 2 3 4 5 6 7 8 9 10 capability of the system proposed and
Sales quantity 285 569 708 766 743 607 432 354 331 215 ANN, two widely used accuracy meas-
Demand quantity 391 657 912 941 891 888 634 553 433 355 ures were chosen: the mean absolute per-
Difference For most106 clothing
88 204 companies,
175 148 281 past
202 sales
199 data140are
102 usually
centage error used
(MAPE) for product
and mean square
error (MSE).

planning, allocation
Table 3. Forecasting performancesdecision, inventory management, etc.
of ANFIS and ANN. If comparison
Results products are always
For most clothing companies, past sales
MAPE MAPE MSE MSE data are usually used for product plan-
Product No. Difference Difference
available
1
for sales29.05%
(ANFIS)
7.45%
and without
(ANN)
21.60%
stock-out,
(ANFIS)
4.450.55
then the
(ANN)
35,608.07 31,157.52
demand
ning, and salesdecision,
allocation information
inventory
management, etc. If products are always
2 9.02% 31.70% 22.68% 6,098.67 38,900.83 32,802.16 available for sales and without stock-out,
3 17.27% 37.80% 20.53% 27,468.70 84,127.40 56,658.70 then the demand and sales information
would 4 be identical.
6.37% However,
28.42% 22.05%due 3,291.09
to the demand
32,169.09 for clothing
28,878.00 wouldproducts
be identical. being affected
However, due to
5 10.72% 35.69% 24.97% 6,765.71 76,841.22 70,075.51 the demand for clothing products being
6 8.59% 33.41% 24.82% 5,642.73 58,466.45 52,823.72 affected by multiple factors and many
by multiple
7 factors40.88%
11.42% and many 29.46%factors are 104,590.77
7,654.59 unpredictable96,936.18 or factors
uncontrollable,
are unpredictable orstock-out
uncontrolla-
8 14.43% 32.90% 18.47% 9,851.00 42,346.01 32,495.01 ble, stock-out is rarely avoided, causing
Average 10.66% 33.73% 23.07% 8,902.88 59,131.23 50,228.35 the amount sold to be less than the ac-
is rarely avoided, causing
Data source: calculated using MATLAB the amount sold to be less than
tualthe
demand actual demand
[17]. Therefore one [17].
of the
main purposes in this study is to consider
stock-out when implement forecasting.
MF and Constant MF were chosen for
Therefore one of the main purposes
inputs and outputs, respectively. Since
Results and this study is toThe
in discussion first stage is estimating demand data
consider stock-out when
based on original sales. Aggregated de-
each rule in ANFIS is represented as “If Technique comparison mand and sales data of the training and
input, Then output”, 108 Sugeno–type The traditional ANN technique was in- validation datasets are shown in Fig-
implement forecasting.
FIS rules were generated. The first
Table 1 shows troducedstage is estimating
to facilitate a comparison withdemandure 7, data based
and Table ondata
2 presents original
of a spe-
example rules. the system proposed. To construct the cific product.
ANN network, the forecasted demand
sales. Aggregated
The last step is measuring thedemand
forecasting and sales
value data byofthethe
is generated training
vectors of the and validation
As expected, the demanddatasets
quantitiesare
are
capacity of the system. Group 2 data were five variables (original price level, pro- significantly higher than those for sales,
used, which include 8 items with a com- motion level, size preference, allocation which means that the sales data cannot
shown in life
plete product Figure
cycle of7,
10and
weeks.Table 2 presents
status datafactor),
and calendar of aand specific
the out- product.
accurately reflect real customer demand.

Training data set Validation data set

60,000 8,000

50,000 7,000
Quantity (unit)
Quantity (unit)

6,000
40,000
5,000
30,000 4,000
3,000
20,000
2,000
10,000
1,000
0 0
W1 W2 W3 W4 W5 W6 W7 W8 W9 W10 W1 W2 W3 W4 W5 W6 W7 W8 W9 W10

Week Week
Sales Demand Sales Demand

Figure 7. Demand and sales patterns of Fig. 7 Demand

datasets. and sales patterns of datasets
14 FIBRES & TEXTILES in Eastern Europe 2017, Vol. 25, 1(121)

Table 2 Demand and sales patterns of a specific product (unit)

Week 1 2 3 4 5 6 7 8 9 10
 clearly demonstrate that the system proposed outperforms the traditional ANN

technique.

Product 3 Product 8
1,800 1,200
1,600
1,000
1,400

Quantity (unit)
1,200 800
Demand

1,000
600
800
600 400
400
200
200
0 0
 W1 W2 W3 W4 W5 W6 W7 W8 W9 W10  W1 W2 W3 W4 W5 W6 W7 W8 W9 W10

Week Week
Actual ANFIS ANN Actual ANFIS ANN

Figure 8. Weekly data of products 3 and Fig.

8. 8 Weekly data of products 3 and 8

Furthermore
In other the just
words, if the company forecasting
uses different capability wasbutalso
impacts on demand, compared
they are with the
rate and unreliable, company’s
and this task would
sales data as the input of its forecasting too numerous to establish an exhaustive be extremely tedious if the number of
system, it would lead to a certain amount list, and some are consistently unavail- new products is large. More importantly,
current approach.
of lost sales. Eventually, theCommercial
total revenue able, forecasting
such as weather,software is integrated
political factors the companyinto
failedtheir enterprise
to consider the issue
must be affected. As a result, it is essen- and competitors’ strategies [21]. Nev- of demand estimation, and thus their re-
tial to integrate the demand estimation ertheless, apart from the special points, sults are only sales quantities. With such
resource
stage into the planning (ERP) system.
forecasting system. However,
the ANFIS-based it generally
system can is well a consideration,
known that it is notcommercial
appropriate to
achieve values that are close to the real compare the demand quantities forecast-
In addition, forecasting performanc- demand. The comparison results clearly ed by our system and the sales quantities
es were compared between the AN- demonstrate that the system proposed the company computed by using MSE
software
FIS-based systemis usually designed
proposed (ANFIS, for universal
for outperforms companies
the traditional ANN tech- and is notHowever,
and MAPE. a good fitbetween
the gap for a
simplification)
curves strongly indicatesand ANN, see Table
the lost sales3.ignored nique.
by the company, see Figure 9. Such a curves strongly indicates the lost
the two
sales ignored by the company, see Fig-
specific
According to company.
the MAPE andMoreover
MSE values, this software
Furthermore is inadequate
the forecasting capability for
ure new
9. Suchproduct
a situation forecasting,
would result in
situation would result in system
the ANFIS-based enormous losses inwasthe
significantly alsolong run. with
compared In fact, some clothing
the company’s enormous losses in the long run. In fact,
increases the forecasting accuracy com- current approach. Commercial forecast- some clothing companies have attempted
pared with ANN. However, we noticed ing software is integrated into their en- to develop their individual customised
companiesthushave
that
company
attempted
sometimes
experts
the to develop
forecasting
needtheirto
results
make adjustments
individual customised
terprise resource
with
forecasting
planning (ERP)
theirforecasting
subjective
system. system
knowledge to a
system [21] based on the
are not very close to the real demand, see However, it is well known that com- company’s actual situation with the con-
Figure 8. The forecasted values of prod- mercial software is usually designed for sideration of more factors. If the demand
[21] based on the company’s
baseline computed actual
bysituation with the The
thefarsoftware. consideration
results of more factors.inaccurate
If
uct 3 in weeks 2 and 4 are relatively universal companies and is are usually
not a good estimation process isand unreliable,
integrated into their
from the real demand compared with the fit for a specific company. Moreover this customised system, we believe lost sales
the demandperformance
estimation process
of other weeks.isSuch
integrated
a situ- into theiris customised
software inadequate forsystem, we believe
new product and end-of-season inventories could be
and
ationthis
is alsotask would
observed be extremely
with product tedious
8 in forecasting, if the number
thus company experts needof new
reduced,products
meanwhile is the large. More
retail efficiency
week 7. A possible reason is that the var- to make adjustments with their subjective and total revenue could be improved.
lost sales iables
and may end-of-season inventories
not fully encapsulate demand could be reduced,
knowledge to a baselinemeanwhile
computed by the the retail
behaviour. Indeed different factors have software. The results are usually inaccu-
importantly, the company failed to consider the issue of demand estimation, and thus
Conclusions
efficiency and total revenue could be improved. To sum up, by first observing the sales
and demand patterns of datasets, the
their results 12,000 are only sales quantities. With such a consideration, it is not appropriate
system proposed could significantly re-
10,000 duce lost sales. In addition, the forecast-
ing capabilities of the system proposed
to compare 8,000
the demand quantities forecasted by our system and were
and ANN the compared.
sales quantities
The results
Quantity (unit)

6,000
demonstrate that the system proposed
outperforms ANN. Moreover the fore-
casting performance of the system pro-
the company 4,000
computed by using MSE and MAPE. However, posedthe
was gap between
compared the two
with the company’s
2,000
current approach. Given the company’s16
0 approach forecasted only with respect to
 P1 P2 P3 P4 P5 P6 P7 P8 sales, lost sales were observed from the
Product gap between the demand quantities our
system forecasted and the sales quanti-
Intelligent System (Demand) Company Approach (Sales) ties the company computed. Neverthe-
less more data is required to obtain robust
Fig. 9.9 Forecasting
Figure Forecastingresultsresults of intelligent
of the the intelligent
systemsystem and company
and company approach. approach sales weights, and the variables affecting

FIBRES & TEXTILES in Eastern Europe 2017, Vol. 25, 1(121) 15

5. Conclusions
 demand should be optimised continuous- AFIS fibre properties by using linear re-
ly. However, given the limitation of our gression models. Fibres and Textiles in
database, these improvements should be Eastern Europe 2007; 15, 4(63): p. 63.
 8. Kuo RJ, Wu P and Wang C. An intelli-
taken into account in future work.
gent sales forecasting system through
India (25%)
integration of artificial neural networks
Summary and fuzzy neural networks
Pakisthan with fuzzy
(25.1%)
weight elimination. Neural networks
China (23.3%)
This study aims to forecast the demand 2002; 15(7): p. 909-925.
Thialand (1.6 %)
for new clothing products. A two-stage  9. Frank C, et al. Forecasting women’s
intelligent retail forecasting system was apparel sales using Others mathematical
(25% ) mod-
eling. International Journal of Clothing
designed and applied to a Canadian fash-
Science and Technology 2003; 15(2):
ion company. In the first stage, demand p. 107-125.
is estimated based on original sales. AN- 10. Hui CL et al. Learning–based fuzzy col-
FIS is employed in stage two to forecast our prediction system for more effective
weekly demand. Meanwhile a data se- apparel design. International Journal of
Figure 10. Importers of raw jute in the World. Source:Clothing IJMA. Science and Technology 2005;
lection process is presented due to the
limited data for new products. According 17(5): p. 335-348.
the system References
n The totalresults
production of conventional 11. Au K-F, Choi T-M and. Yu Y. Fashion
to the and discussion,
retail forecasting by evolutionary neural
jute goods is
proposed is also
well seeing
suited for an the
increas-
clothing 1. Ahmad networks.
DM and International
Parrey A H. Journal of Pro-
July. Socio
ingindustry.
trend and sacking cloth production ductionPotential
Economics 2008; 114(2): p. 615-
economic of handicraft Industry
covers more than 75% of the total jute 630.
in Jammu and Kashmir; Opportunities
products in India. B twill bags con- 12.
andWang F-K, Chang
Challenges. K-K and Tzeng
International Monthly C-W.
tributes 85% of jute bags production Using adaptive network-based fuzzy in-
Referred Journal of Research in
in India. ference system to forecast automobile
Management and Technology 2013; 2.
sales. Expert Systems with Applications
Acknowledgments
n Jute is now facing challenges from the 2. Ghimire
2011;TB andp.Thakur
38(8): N C. Constraint
10587-10593.
cheaper
The authorspriceswould
of synthetic substitutes.
like to thank the compa- and Opportunity
13. Bektas Ekici ofB Raw Jute Production:
and Aksoy UT. Predic-
Hence
ny formore development/diversifica-
providing data and expert knowledge. A case study
tion of of Eastern
building energy Terai,
needs Nepal.
in early
tion
Thisofresearch
jute products are required
is supported for
by the Chinese stage ofJournal
Agronomy design by (Agron Expert
using ANFIS.
of Nepal JN),
theScholarship
viability ofCouncil (NO. 201406630046
the industry. 2013;Systems with Applications 2011; 38(5):
3: 117-122
and NO. 201406240017). 3. Islam p. 5352-5358.
M Sl and Alauddin M.. World
n The cooperation of government, the 14. Aksoy A, of
Ozturk N and Sucky E. A de-
owners of organizations, employees Production Jute: A Comparative
cision support system for demand fore-
Analysis of Bangladesh. International
etc. is required to make a concerted casting in the clothing industry. Interna-
Journal of Management and Business
References
effort for the future sustainability of tional Journal of Clothing Science and
2012;
this industry.
Studies
Technology 2012;224(4): (1):p. 221-236.
014-022,
 1. Lenort R and Besta P. Hierarchical Sales International
15. Gallien J, Scholars Journals
et al. Initial Shipment (http://
Deci-
Forecasting System for Apparel Compa- internationalscholarsjournals.org),
sions for New Products at Zara. Opera-
nies and Supply Chains. Fibres and Tex- Received December
tions Research 17, 63(2):
2015; 2011; Accepted
p. 269-286.
tiles In Eastern Europe 2013; 21, 6(102):
Editorialp. note7-11.
16. Bergvall-Forsberg
(January 3, 2012). J and Towers N. Cre-
atingGagile
4. Karthik and supply
Ramyanetworks
N..‘SwotinAnalysis
the fash-
1) ‘A 2. Şen A. The US fashion industry: A sup-
twill bag’ refers a special type of ion industry:
of Indian A pilot study
Jute industry. of the
Indian Europe-
Streams
ply chain review. International Journal an textile and clothing industry. Journal
100%ofjute bag which is used as
Production Economics 2008; 114(2):packag- Research Journal 2012; 2, 10.
of the Textile Institute 2007; 98(4):
ing ofp. 571-593.
food grains of 100 kg capacity in 5. Khondake G M, Rahman T and Sahan
p. 377-386.
India, Bangladesh.
 3. Wong WK andWheat,Guo ZX.paddy, rice,
A hybrid intelli- A.. Wecker
17. Jute WE. Manufacturing sector from
Predicting demand in
sugar,gent
spices model for medium-term
etc are packing in this sales
bagfore- Bangladesh,
sales data Opportunities,
in the presenceChallenges
of stockouts.
casting in fashion
for transportation. The termretail ‘Twill
supply ‘chains
is andManagement
Policy options. Occasional
Science 1978; paper24(10):
using extreme learning machine and 78, Centre for
p. 1043-1054. Policy Dialogue (CPD),
a textile design pattern which
harmony search algorithm. International
imparts
18. Garro A. New product demand
Dhaka-1209,Bangladesh, 2009. forecast-
duringJournal
weaving of cloth atEconomics
of Production loom stage. 2010;
6. RayingKamal.
and distribution
Economics optimization:
of Summer a case
The bag itselfp.has
128(2): weight of 1120 g as
614-624. Massachusetts Institute
Paddystudy– atJute
Zara.substitution: A profile of of
per 4.
Bureau
Babuška of Indian
R and Standards.
Verbruggen H. Neu- Technology, 2011.
ro-fuzzy Environmental Cost. Asian Journal
2) ‘B twill bag’ methods
refers to for 100%nonlinear system
jute bag 19. Negnevitsky M. Artificial intelligence:
of Research in Social Science &
identification. Annual Reviews in Control a guide to intelligent systems. Pearson
used for packaging
2003; 27(1): p. 73-85.
of rice of capacity Humanities 2012; 2, 3.
Education, 2005.
50 5.
kg.Mostard
The bagJ, itself
Teunter hasRweight
and de 650 g R. 7. Singh
Koster R A. J-SR.
Golden ANFIS:
fibre shines brighter.
20. Jang adaptive-net-
as per Forecasting
Bureau of Indian demand Standards.Twill
for single-period Journal of Business line, 2004.
work-based fuzzy inference system.
is a design
products:which imparts
A case studyon in cloth dur- in- 8. Monthly
the apparel summary
Systems, Man and of Jute and Gunny
Cybernetics, IEEE
dustry. European
ing weaving at loom stage. Journal of Operational Statistics, IJMA,on,
Transactions SL1993.
no -836,
23(3):November
p. 665-685.
Research 2011; 211(1): p. 139-147. 21. Thomassey S. Sales forecasts in cloth-
2014.
 6. Ozbek A, et al. Prediction of Turkey’s ing industry: The key success factor of
9. www.worldjute.com
Denim Trousers Export Using Artificial the supply chain management. Interna-
Neural Networks and the Autoregressive 10. www.ijma.com
Acknowledgement tional Journal of Production Economics
Integrated Moving Average Model. Fi- 11. www.jutecorporation.org
2010; 128(2): p. 470-483.
We wouldbres like&toTextiles
acknowledge
in Eastern andEurope
express 2011; 12. www.ijira.org
19, 3,
our gratitude (86):
to all thep.executives
86. of the Bally
 7. UreyenLtd
Jute Company MEfor and Kadoglu
their H. Theand
full support predic-
tion of cotton ring yarn properties from
cooperation. Received 
Received 13.04.2016 Reviewed
13.04.2015 Reviewed  23.06.2016
22.06.2015

FIBRES
16& TEXTILES in Eastern Europe 2016, Vol. 24, 6(120) FIBRES & TEXTILES in Eastern Europe 2017, Vol. 25, 1(121)
25