Fourier Analysis For Demand Forecasting in A Fashion Company
Fourier Analysis For Demand Forecasting in A Fashion Company
Fourier Analysis For Demand Forecasting in A Fashion Company
): INFORMATIK 2015
Lecture Notes in Informatics (LNI), Gesellschaft für Informatik, Bonn 2015
Abstract: The fashion industry faces different challenges regarding accurate forecasts for future
fashion products. The consumer demand is volatile and sales periods of fashion products are short
due to production plants in Asia and target markets in Europe. Besides standard statistical
approaches based on historical data and advanced methods such as the application of artificial
neural networks or fuzzy logic, there are fashion experts, who use different information sources,
e.g. fairs, social media, fashion websites, to predict design-trends as well as sales volumes. In this
paper we follow this expert-driven approach by collecting data from fashion weblogs, news sites
and fashion magazines, in order to identify actual and future design- trends. For this aim, we
develop the TrendFashion Tool which collects data from these fashion sources and analyse them.
On a higher level, this tool successfully separates fashion related posts from non-fashion related
posts. And on a lower level, it identifies fashion related words and weights them according to an
index.
Keywords: Fashion Forecasting, Social Media Analytics, Trend identification
The fashion industry has to handle diverse idiosyncrasies on different levels in order to
avoid overstocked and stock-out inventories. External factors such as changing weather
conditions, holidays or (sports) events have an impact on short-term customers’
purchasing decisions and buying behaviours [Th10]. The availability of a product is
crucial for fashion items, since buying decisions are often made at the point of sale
[NGP13]. Moreover, different design trends are presented in fashion shows with a long-
term range of approximately six months. Besides, street trends are emerging
spontaneously with a mid-term range of approximately three months. Additionally, the
most fashion products are shorter lived and will stay only for some weeks at the stores.
1 Universität Bremen, International Graduate School for Dynamics in Logistics (IGS), Hochschulring 20,
28359 Bremen [email protected]
2 BIBA - Bremer Institut für Produktion und Logistik GmbH at the University of Bremen, Hochschulring 20,
28359 Bremen, Germany, [email protected]
4
Faculty of Production Engineering, University of Bremen, Germany, [email protected]
1195
Samaneh Beheshti-Kashi et al.
Product related factors such as the high variety in sizes and colours intensifies the
demand uncertainty [CLP04]. In this regard, we can distinguish between traditional and
fast fashion retailing. While traditional fashion retailing consists of producing one
collection for the spring/summer and one for the autumn/winter season by producing the
products in Asia and shipping them to Europe, the Spanish retailer Inditex is an excellent
example of fast fashion retailing. Zara, the major division of Inditex has become the
leader in rapid development of fast changing fashions by addressing latest street trends
with products made in Spain with lead times less than several weeks. On the other hand,
most production plants are placed in Asian countries where products can be produced
cheaply [MTD11], but the time-to-market is long. Therefore, the reproduction of good
selling products is rarely possible [Fs01]. One option for companies is to fly their goods
to Europe or install additional production plants in Turkey or North Africa in order to be
time efficient. However, the traditional as well as the fast fashion retailing strategy face
challenges, which are worthy considering new approaches of trend identification and
forecasting.
2 Related Work
The literature describes different approaches within the field of fashion forecasting. Due
to the described characteristics of the fashion industry and the lack of historical data,
standard forecasting methods such as simple statistical approaches often fail in providing
accurate forecasting results. In recent approaches more advanced methods such as fuzzy
logic or artificial neural networks are considered [SAC07] [Th10]. Further learning
algorithm such as the extreme learning machine [Sz08] or the evolutionary neural
network [ACY08] obtains accurate results. Though, in particular hybrid models achieve
reasonable accuracy [WG10] [Ct14] [Tm14]. However, these works do not consider
further influencing factors in their models. In contrast, Thomassey [Th10] provided a
complex approach including external information such as weather conditions or
information on promotions and events [Th10]. Considering these additional information,
the accuracy of forecast results will increase, since the consumer demand is impacted by
exactly these diverse factors. Another approach is followed by Mostard et al. [MTD11]
and Teucke et al. [Tm14]. Both focus on pre-order information and include them into
their forecasting models. Within the fashion industry the selection of the right colour is a
highly important decision [Ki12]. Accordingly, this problem is also addressed by
researchers. [Ct12] introduced a model for colour forecasting with very little data. In Gu
et al. [GL10] a computer assisted colour forecasting database is presented. Beheshti-
Kashi et al. [BT14] suggest integrating social media applications such as Twitter or
weblogs into the fashion forecasting process [BT14]. This paper follows this approach
and suggestion and introduces the so-called TrendFashion Framework.
119
TrendFashion
Within the last 5 years the fashion world has undergone a huge change. In particular,
within the last 2-3 years new concepts such as outfittery5, edited6, AboutYou7 or Zalon8,
Zalando’s curated shopping solution, have emerged. All these applications require more
consumer engagement, for instance, providing more personal data than for a traditional
online purchase. Consequently, the companies obtain far more information about their
customers than before. Information regarding each purchase, on colours, sizes, number
of purchases, preferences, which products have been purchased together, or just product
views are easily accessible. Consequently, a huge number of data is available about the
consumer for the companies. The availability of these data, paves the way for data driven
approaches within the fashion industry. In addition, to the available information
provided by the consumer, further data is available throughout the internet. In recent
times some start-ups9 10 have emerged who focus on these kinds of fashion related data.
Other approaches focus on a typical problem occurring in the fashion industry: sizes and
the perfect fit. Every brand has its own particular sizes. Therefore, each garment or shoe
ideally should be fitted before purchasing, which is not always possible, in particular in
the case of online shopping. Recently, applications such as virtual try-on technologies or
3D Scanners have emerged which can be used by the consumers also at home in order to
build their digital avatar. Both technologies are used for ensuring a more accurate fit for
the consumer. In addition, to these applications augmented reality environment have
obtain increased relevance. These environments allow the user for instance to browse a
catalogue and try on apparels [BA15]. Ashdown and Loker consider the virtual try-on as
a promising 3D technology for enhancing mass customization and online sales [AL10].
Google as the market leader in search engine technologies has conducted a study on
fashion trends considering search queries. In the “Google Fashion Trends Report (U.S.)”
they publish their results examining six billion apparel related queries in the United
States. They categorize the extracted trends based on user search behaviour into the
following six clusters applying Time Series Clustering: Sustained Growth, Seasonal
Growth, Rising Stars, Sustained Decline, Seasonal Decline and Falling Stars [ZH15].
This work only considers search queries and not free and unstructured text. Therefore,
no automated text processing and Natural Language Processing (NLP) methods are
applied. Whereas, considering weblogs and digital fashion magazines require conducting
NLP methods, in order to obtain valuable information.
5
http://www.outfittery.de/
6
http://www.edited.de/
7
http://www.aboutyou.de/
8
https://www.zalon.de/
9
http://www.fashiongps.com/
10
https://poshly.com/
119
Samaneh Beheshti-Kashi et al.
Text Mining methods and approaches have gained increased relevance within recent
years. One reason is the huge emergence of free and unstructured data on the web. In
particular, due to the numerous social media applications, which empowered every
individual user to publish content, Text Mining is required in order to process the data
and to extract valuable information. However, Text Mining methods are not limited to
web usages, any textual database can profit from Text Mining. Text Mining includes
several fields such as information retrieval, text analysis, information extraction,
clustering, categorization, visualization, database technology, machine learning and data
mining [TA99].
Granitzer [GR06] has identified the following four processes in Text Mining:
- Pre-processing: lexical analysis, part-of-speech tagging, stemming, stop word
removal
- Information extraction: machine learning and linguistic analysis
- Feature generation: statistical analysis such as frequency analysis, word co-
occurrences
- Operations on feature spaces: clustering, classification etc.
Within the last years the focus of Text Mining has gone towards Natural Language
Processing (NLP). NLP is an older research field and involves technologies such as word
stemming, lemmatization, multiword phrase grouping, synonym normalization and part-
of-speech tagging [KP04].
1198
TrendFashion
Within the data processing phase, the content is analysed by the mean of NLP methods,
and the paragraphs are separated into sentences. At this stage, we have included a
“Moderelevanz” Index which weights the fashion relevancy of words and sentences.
Relevant sentences are those which include for instance colours, brands or textile names.
These sentences are separated into words. These words are matched with the thesaurus
and a “Moderelevanz” Index is assigned to them. In the following sections, the
collecting, processing as well as the thesaurus functionalities are presented.
The tool is designed to handle different formats, namely RSS feeds and HTML-websites.
The first approach was to use RSS feeds exclusively, since they always provide a
complete time stamp. A detailed time reference to the written text is highly relevant for
the given objective. The significance of the time aspect is due to the fast paced and
rapidly changing fashion trends. Therefore, we have to focus on data from the same time
period in order to correctly aggregate the data and to conclude the right information.
Since a large number of the blogs do not provide RSS feeds, we added the second
approach, HTML formats, to the tool. Currently, the tool considers 100 German
1199
Samaneh Beheshti-Kashi et al.
After collection, the data has to be processed. For this task, we apply NLP methods such
as sentence segmentation, tokenization, named entity recognition and Part-of-Speech-
Tagging. The basis for further processing is the single posts which are stored as single
documents. In the first step, the posts are separated into single sentences looking based
punctuations. Then, the tokenization process will follow. After separating the sentences,
the syntactic analysis starts. The Part-of-Speech-Tagger (POS-Tagger) annotates of
every single word of sentence to a specific part-of-speech such as noun, verb or
adjective. For the Tagging process we selected the German version of Wiktionary.org,
which is a free online wiki-based dictionary, for identifying the correct part-of-speech.
The “Moderelevanz” Index, filters and weight the separated POS according to their
relevance to fashion topics and filter out stop words such as articles or pronouns at the
same time. The extracted and tagged words are stored in a separated table.
3.3 Thesaurus
In order to enhance the semantic functionality of the tool, we have included a thesaurus.
Such a thesaurus is often designed for a particular domain, and in contrast to a dictionary
it provides relations between words. These relations are required for automated text
analysis and are based on several taxonomies in which we classified different product
groups. So far, we have included taxonomies only on a high level. Detailed taxonomies
are in progress and will be integrated for enhancing the learning capability of the tool.
The thesaurus includes broader, narrower and related terms. In addition, to these
corresponding terms, the words are stored according to their ”Moderelevanz” Index.
Furthermore, the part of speech of each entry is provided. In the current version of the
tool, the “Moderelevanz” Index is adjusted and adapted manually.
4 Results
The TrendFashion Database handles different tasks. It collects and processes the data. It
filters the content from the noise of a website (navigation, HTML tags etc.) and stores
each post separately. Based on the posts, sentences are separated; from these sentences
words and word combinations are extracted and stored. Besides the collection and
storage of posts, sentences, and words, it provides lists of word occurrences. This is the
first step in order to extract future trends. Though, a further focus is on extracting words
belonging together, in order to increase the semantic ability of the tool. For identifying
fashion trends, adjectives such as colours are crucial. Therefore, it is important to find
adjective-noun or adjective-adjective-noun combinations belonging together. In the
current version, the tool is able to identify and extract such combinations in a reasonable
1
TrendFashion
way. Figure 4.1 illustrates matched word combinations in the area of evening robe,
glitter and sparkling clothing. The font size of the word combinations correspondents to
their frequency rates: the larger the font size, the higher is their occurrences within the
text. In the most cases, the tool matched adjective-noun combinations. However, in the
example of femininenSmokingLook, we can see that in a first step the two nouns Smoking
and Look are combined, and in the second step the adjective feminine is matched to the
new combination.
With the word cloud in Figure 4.1 a broader and more general view on the extracted
word combinations is presented. However, the tool can also identify more specific
combinations and relations. Since in fashion just a slight variation of a certain colour,
shape or cut will have an impact on the success of a product, we assume that for the
prediction of a future design trend more specific features on the colour, cut or style are
required and not just a general overview of a product. For instance, looking at the
product “blouse”, the following adjective noun combinations are extracted bodenlange
Bluse, rosafarbene Bluse, bieder Bluse, hochgeschlossene Bluse. In this example,
different features describing a “blouse” are provided: a length (bodenlange), a colour
(rosafarbene), a shape (hochgeschlossene) as well as information on the style (bieder).
Similarly, the tool extracts diverse adjective noun combinations for the product jeans.
We consider the presented tool as the basis for future predictions. Since the objective is
1 1
Samaneh Beheshti-Kashi et al.
to provide trend information also on a detail level, it is crucial to classify the extracted
information according to the different features. This will be also done by the help of
taxonomies. Figure 4.2 illustrates the extracted word combinations as a hierarchal tree.
In this case, we have taken the features colour, cut, style as well as additional describing
attributes and matched the corresponding adjectives to them.
P
J
F
F F F
A
C C S
W E L
R
W T C
H S E
B
L S E
H K
W
L B S
1
TrendFashion
1
Samaneh Beheshti-Kashi et al.
References
[ACY08] Au, K.-F., Choi, T.-M., & Yu, Y. : Fashion retail forecasting by evolutionary neural
networks. International Journal of Production Economics, 114(2), 615–630, 2008.
[AL10] Ashdown, S., Loker, S.; Mass-Customized Target Market Sizing: Extending the Sizing
Paradigm for Improved Apparel Fit. Fashion Practise 2/2, 147-174.
[BA15] Baytar, F., Ashdown, S.: An Exploratory Study of Interaction Patterns around the Use
of Virtual Apparel Design and Try-on Technology. Fashion Practise: The Journal of
Design, Creative Process and the Fashion Industry, 7/1, 31-52, 2015.
[BT14] Beheshti-Kashi, S., Thoben, K.-D.: The usage of social media text data for the demand
forecasting in the fashion industry fashion sales forecasting and the predictive power
of online. In Dynamics in logistics. Fourth international conference, LDIC 2014
Bremen, Germany, February 2014 Proceedings, 1–4, 2014.
[CLP04] Christopher, M., Lowson, R., Peck, H.: Creating agile supply chains in the fashion
industry. International Journal of Retail & Distribution Management, 32/8, 367–376,
2004.
[Ct14] Choi, T., Hui, C., Liu, N., Ng, S., & Yu, Y.: Fast fashion sales forecasting with
limited data and time. Decision Support Systems, 59, 84–92, 2014.
[Ct12] Choi, T., Hui, C., Ng, S., & Yu, Y.: Color Trend Forecasting of Fashionable Products
with Very Few Historical Data, 42/6, 1003–1010, 2012.
[Fs01] Fissahn, C.: Marktorientierte Beschaffung in der Bekleidungsindustrie - Schriften zur
Textilwirtschaft, Band 53,. FATM, Münster, 2001.
[Gr 06] Granitzer, M.: KnowMiner: Konzeption und Entwicklung eines generischen
Wissenserschliessungsframeworks, PhD Thesis TU Graz, Graz, 2006.
[GL10] Gu, W., & Liu, X.: Computer-Assisted Color Database For Trend Forecasting, 0–3,
2010.
[Ki12] King, J. A.: Colour trends and selection in fashion design. Colour design: Theories
and applications, 535–550, Woodhead Publishing Limited, 2012.
[KP04] Kao, A., Poteet, S.: Text Mining and Natural Language Processing –
Introduction for the Special Issue. SIGKDD 7/1, 2004.
[MTD11] Mostard, J., Teunter, R., & de Koster, R.: Forecasting demand for single-period
products: A case study in the apparel industry. European Journal of Operational
Research, 211/1, 139–147, 2011.
[NGP13] Nenni, E. M., Giustiniano, L., & Pirolo, L..: Demand Forecasting in the Fashion
Industry: A Review. International Journal of Engineering Business Management, 5,
2013.
[SAC07] Sun, Z.-L., Au, K.-F., & Choi, T.-M. : A neuro-fuzzy inference system through
integration of fuzzy logic and extreme learning machines. In: IEEE Transactions on
Systems, Man, and Cybernetics. Part B, Cybernetics : A Publication of the IEEE
Systems, Man, and Cybernetics Society, 37/5, 1321–31, 2007.
1 4
TrendFashion
[Sz08] Sun, Z.-L., Choi, T.-M., Au, K.-F., & Yu, Y.: Sales forecasting using extreme learning
machine with applications in fashion retailing. Decision Support Systems, 46/1, 411–
419, 2008.
[TA99] Tan, A.-H.:Text Mining: The state of the art and the challenges, PAKDD 1999
Workshop on Knowledge Discovery from Advanced Databases, Beijing, (1999), 65-
70, 1999.
[Th10] Thomassey, S.: Sales forecasts in clothing industry: The key success factor of the
supply chain management. International Journal of Production Economics, 128/2,
470–483, 2010.
[Th14] Thomassey, S.: Sales Forecasting in Apparel and Fashion Industry: A Review. In.
(Choi, T.M. et al.): Intelligent Fashion Forecasting Systems: Models and Applications.
Springer-Verlag Berlin Heidelberg 2014, 9-27, 2014.
[Tm14] Teucke, M., Ait-alla, A., El-berishy, N., Beheshti-Kashi, S., & Lütjen, M.: .
Forecasting of Seasonal Apparel Products. In (H. Kotzab, J. Pannek, & K.-D. Thoben):
Dynamics in Logistics. Fourth International Conference, LDIC 2014 Bremen,
Germany, February 2014 Proceedings. Springer, 2014.
[WG10] Wong, W. K., & Guo, Z. X.: A hybrid intelligent model for medium-term sales
forecasting in fashion retail supply chains using extreme learning machine and
harmony search algorithm. International Journal of Production Economics, 128/2,
614–624, 2010.
[ZH15] Zimmer, O., Horwitz, Y.: Google Fashion Trends Report (U.S.), 2015.
1 5