Forecasting Seasonal Footwear Demand Using Machine Learning

by
Majd Kharfan
Bachelor of Economics, Accounting, Damascus University, 2011
and
Vicky Wing Kei Chan
Bachelor of Business Administration, Global Supply Chain Management,
The Hong Kong Polytechnic University, 2011

SUBMITTED TO THE PROGRAM IN SUPPLY CHAIN MANAGEMENT

IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF
MASTER OF APPLIED SCIENCE IN SUPPLY CHAIN MANAGEMENT
AT THE
MASSACHUSETTS INSTITUTE OF TECHNOLOGY
JUNE 2018
© 2018 Majd Kharfan and Vicky Wing Kei Chan. All rights reserved.
The authors hereby grant to MIT permission to reproduce and to distribute publicly paper and
electronic copies of this capstone document in whole or in part in any medium now known or
hereafter created.

Signature of Author....................................................................................................................
Majd Kharfan
Department of Supply Chain Management
May 11, 2018
Signature of Author......................................................................................................................
Vicky Wing Kei Chan
Department of Supply Chain Management
May 11, 2018
Certified by..................................................................................................................................
Dr. Tugba Efendigil
Research Scientist, Center for Transportation and Logistics
Capstone Advisor
Accepted by..................................................................................................................................
Dr. Yossi Sheffi
Director, Center for Transportation and Logistics
Elisha Gray II Professor of Engineering Systems
Professor, Civil and Environmental Engineering
Forecasting Seasonal Footwear Demand Using Machine Learning
by
Majd Kharfan
and
Vicky Wing Kei Chan
Submitted to the Program in Supply Chain Management
on May 11, 2018 in Partial Fulfillment of the
Requirements for the Degree of Master of Applied Science in Supply Chain Management

ABSTRACT
The fashion industry has been facing many challenges when it comes to forecasting demand
for new products. The macroeconomic shifts in the industry have contributed to short product
lifecycles and the obsolescence of the retail calendar, and consequently an increase in demand
variability. This project tackles this problem from a demand forecasting perspective by
recommending two frameworks leveraging machine learning techniques that help fashion
retailers in forecasting demand for new products. The point-of-sale (POS) data of a leading
U.S.-based footwear retailer was analyzed to identify significant predictor variables
influencing demand for footwear products. These variables were then used to build two models,
a general model and a three-step model, utilizing product, calendar and price attributes for
predicting demand. Clustering and classification were used under the three-step model to
identify look-alike products. Regression trees, random forests, k-nearest neighbors, linear
regression and neural networks were used in building the prediction models. The results show
that the two forecasting models based on machine learning techniques achieve better forecast
accuracy compared to the company’s current performance. In addition, the proposed
methodology offers visibility into the underlying factors that impact demand, with insights into
the importance of the different predictor variables and their influence on forecast accuracy.
Finally, the project results demonstrate the value of forecast customization based on product
characteristics.

Capstone Advisor: Dr. Tugba Efendigil

Title: Research Scientist, Center for Transportation and Logistics

1
Acknowledgements
We would like to thank the MIT family of the Supply Chain Management program for giving

us the opportunity to challenge ourselves and to broaden our experience. Special thanks go to

our advisor, Dr. Tugba Efendigil, for her mentorship and guidance throughout this project. Also,

we would like to extend our sincerest gratitude to our sponsoring company. In particular: Shruti,

Monica, Stephen, Dan, Luis, Daniele and others who challenged us and were there to answer

our questions.

Vicky & Majd

I feel honored and blessed for everything I have learnt and experienced while working on this

project at MIT. I want to acknowledge the efforts of my wonderful partner, Vicky. Without her,

this work would not have been possible. Also, a huge thanks to my beautiful mom, Nisreen,

and my lovely wife, Afnan, for their patience and support throughout this journey.

Majd

I am grateful for the learning opportunities that I have been given while working on this project

at MIT. I especially want to thank my amazing project partner, Majd, for his dedication and

efforts to this project. I would also like to thank my family and friends who always support me

in everything I do.

Vicky

2
Table of Contents
ABSTRACT .............................................................................................................................. 1
Acknowledgements .................................................................................................................. 2
Table of Contents ..................................................................................................................... 3
List of Tables............................................................................................................................. 5
List of Figures ........................................................................................................................... 5
1. Introduction .......................................................................................................................... 6
1.1 Overview of the Retail Fashion Industry ..................................................................... 6
1.2 The Company and Motivation ...................................................................................... 7
2. Literature Review ................................................................................................................ 9
2.1 Demand Forecasting Methods ...................................................................................... 9
2.2 Predictor Variables in Demand Forecasting ................................................................ 9
2.3 Traditional Techniques vs. Machine Learning Techniques ...................................... 10
2.4 Application of Machine Learning Techniques in Industry ...................................... 12
3. Methodology ....................................................................................................................... 14
3.1 Machine Learning Techniques Used .......................................................................... 15
3.1.1 Supervised Learning Techniques ......................................................................... 15
3.1.2 Unsupervised Learning Techniques .................................................................... 16
3.2 Scope and Granularity of Data ................................................................................... 16
3.3 Feature Selection and Engineering ............................................................................. 18
3.4 Dataset Partitioning ................................................................................................ 20
3.5 Model Building ........................................................................................................ 21
3.5.1 General Model ....................................................................................................... 21
3.5.2 Three-Step Model .................................................................................................. 21
3.6 Performance Measurement .................................................................................... 26
4. Results ................................................................................................................................. 28
4.1 General Model .............................................................................................................. 28
4.2 Three-Step Model ......................................................................................................... 30

3
 4.2.1 Clustering and Classification ............................................................................... 30
4.2.2 Prediction ............................................................................................................... 33
5. Discussion............................................................................................................................ 37
5.1 Implications .................................................................................................................. 37
5.2 Limitations .................................................................................................................... 37
5.3 Future Research ........................................................................................................... 38
6. Conclusion .......................................................................................................................... 40
References ............................................................................................................................... 41

4
List of Tables
Table 1. Comparison between Traditional and Machine Learning Forecasting Approaches .. 12
Table 2. Applications of Machine Learning Techniques in Forecasting in Different Industries
.................................................................................................................................................. 13
Table 3. List of Attributes from the Aggregated Data by Month at the Style Level ................ 17
Table 4. List of Attributes for Feature Selection ...................................................................... 19
Table 5. Overview of Datasets Generated for the General Model ........................................... 20
Table 6. Overview of Datasets Generated for the Three-Step Model ...................................... 21
Table 7. List of the Variables Considered by Each Step of the Three-Step Model .................. 23
Table 8. List of Attributes Selected for Model Building .......................................................... 28
Table 9. Forecast Accuracy of the General Model ................................................................... 30
Table 10. Forecast Bias of the General Model ......................................................................... 30
Table 11. The Silhouette Score for Different k Number of Clusters ........................................ 31
Table 12. Comparison of the Classification Accuracy by Algorithm ...................................... 32
Table 13. Characteristics of Styles Distribution among Clusters ............................................ 33
Table 14. Best Performing Algorithm by Cluster (Validation Set) .......................................... 36
Table 15. Best Performing Algorithm by Cluster (Testing Set) ............................................... 36

List of Figures
Figure 1. The Proposed Methodology ..................................................................................... 14
Figure 2. The Four Sub-Steps Followed in Clustering ............................................................ 22
Figure 3. The Three Sub-Steps Followed in Classification ..................................................... 24
Figure 4. The Three Sub-Steps Followed in Prediction ........................................................... 25
Figure 5. Cross Validation Error by Number of Attributes ...................................................... 29
Figure 6. The Confusion Matrix Resulted from Using Five Clusters and SVM Algorithm .... 32
Figure 7. Forecast Accuracy and Bias of the Three-Step Model (Validation Set) ................... 34
Figure 8. Forecast Accuracy and Bias of the Three-Step Model (Testing Set) ........................ 35

5
1. Introduction
This section provides a high-level overview of the state of the retail fashion industry. It

discusses how agile supply chain strategies can enable fashion companies to adapt to current

trends. Finally, it highlights the essential role of demand forecasting in supporting agile supply

chain strategies and the optimization of other business functions.

1.1 Overview of the Retail Fashion Industry

The fashion industry has immensely evolved in the past few decades, especially after the

introduction of e-commerce. Consumers’ taste has become the major demand driver for fashion

products. It is generally influenced by internal and external factors including personalization,

omni-channel competition, social media influencers, political movements and others. This

continuous change in consumers’ behavior has led to shorter product lifecycles and more

volatile demand. In addition, consumer expectations have become greater as high quality,

guaranteed availability and fast delivery are no longer negotiable.

With all these challenges, fashion companies must develop overarching strategies that are

adaptable to the constant changes in the industry. Such strategies should embrace marketing as

a demand creation tool and digital capabilities like e-commerce and mobile apps as growth

enablers. Innovation and speed to market are other important features a strategy should focus

on. These features help companies stay competitive in today’s global market where brands like

Zara and H&M refresh their assortment every few weeks.

Once such an overall strategy is set, the role of an agile supply chain strategy that focuses on

responsiveness, competency, flexibility and quickness comes into play. An agile supply chain

will work as an enabler and executor through a number of aligned initiatives that collectively

6
work toward achieving the company’s objectives. Examples of such initiatives may include

manufacturing lead times, which can be fostered by applying ABC analysis to discover and

solve bottlenecks in the process system. Development of postponement strategies through

staging materials or semi-finished products at distribution centers (DCs) or factories is also

essential to provide flexibility and give the company extra time to see better market signals.

Moreover, inventory policies need to be visited to ensure safety stock and order quantity

parameters are set based on statistical analysis that considers the trade-offs between cost and

level of service.

Having an agile supply chain cannot be accomplished without optimizing demand planning

and especially demand forecasting, which will be the area of focus throughout this research

project. Demand forecasting can be defined as the art and science of predicting customers’

future demand for products. It serves as a major input for planning across different supply chain

and business functions, including raw materials planning, supply planning, inventory

management, sales and merchandising. Poor forecasting results can lead to stock outs and loss

in revenues and market share to competitors, or to excessive inventory, i.e., frozen capital and

high obsolescence. Therefore, having good demand forecasting capability is essential to

optimize other functions and to support the overall supply chain and company’s strategies.

1.2 The Company and Motivation

The sponsoring company for this research is a major footwear manufacturer and retailer based

in the US with operations across the globe. The company sells its products through its own

inline (full price) and outlet (discounted price) brick-and-mortar retail stores as well as through

its online website and wholesale partners.

The United States is the largest market for this company and the scope of this research. Like
7
other fashion retailers, the sponsoring company is at the crossroads of two key macro shifts:

the “Buy Now/Wear Now” consumer mentality influenced by social media and the love of

personalization, and the economic challenges facing the retail industry in the form of declining

mall traffic and the obsolescence of the traditional retail calendar. With that in mind, the

company is reworking its strategy to improve its position in the marketplace by becoming

closer to consumers and quicker in responding accurately to demand signals. This will

consequently bring to the company operational efficiencies in the form of minimized order

cancellation rates and healthier levels of inventory in the marketplace, which will be translated

into cost savings and additional revenues.

Through carrying out this research project we aim to recommend solutions to the sponsoring

company that will improve the demand forecasting capabilities and prediction accuracy.

Applying machine learning will maximize the utilization of the point-of-sale (POS) data and

help uncover new insights to be used in developing a demand forecasting framework that meets

the company’s strategic objectives.

8
2. Literature Review
This section explores the demand forecasting methods and common predictor variables that

have been used in industry, compares traditional and machine learning forecasting techniques,

and reviews the application of machine learning techniques in different industries. This

information sets the stage based on which we built our forecasting models through selecting

appropriate predictor variables and using suitable techniques.

2.1 Demand Forecasting Methods

Demand forecasting in the apparel and footwear industry is extremely challenging due to

volatile demand, strong seasonality, Stock-keeping-unit (SKU) intensity and for seasonal and

fashion items, short lifecycles and lack of historical data (Thomassey, 2010). Consumer

demand is the result of the interplay among a number of factors, which ideally should serve as

predictor variables in generating demand forecasts. However, in practice sometimes the effect

of these factors can be difficult to decouple. For example, price and seasonality are

interdependent on each other (Kaya, Yeşil, Dodurka & Sıradağ, 2014). Traditional forecasting

methods usually only take into account a single factor or at most a few factors, so part of the

variation remains unexplained in the forecasting model when in fact there may be patterns

undiscovered. In this research, different machine learning based forecasting techniques will be

explored to identify the most suitable approach for the sponsoring company.

2.2 Predictor Variables in Demand Forecasting

The most common type of data used in demand forecasting is POS data, or downstream data,

which is widely used in both traditional time-series forecasting and advanced machine learning

techniques. For retailers, POS data are usually readily available and relatively accessible, as

they are automatically captured at consumers’ checkout upon each purchase transaction.
9
Wholesalers and manufacturers depend on their downstream retail partners for visibility to POS

data.

In addition to POS data, there are many other types of data that are being used in industry or

proposed in academic research papers in demand forecasting. One important type of data is lost

sales. Demand that is not satisfied because of stock-outs is not captured in POS data and results

in potential lost sales. In such cases, true demand may be underestimated if sales are treated as

being equal to demand (Kaya et al., 2014). Therefore, lost sales need to be taken into account

during the forecasting process to reflect true historical demand. Other types of data include

price and promotion, consumer loyalty, calendar and holidays, weather, geographic location,

competition, item features, fashion trends, store count and mode of distribution, as well as

macro-economic trend data such as purchasing power and unemployment rate (Thomassey,

2010). These types of data lead to a large number of decision variables to be explored in

improving forecasting accuracy. Some factors are believed to have more impact compared to

the others. For example, in building a demand signal repository (DSR) for a fast-moving

consumer goods (FMCG) company, Rashad and Spraggon (2013) found year, month, weekday

and holidays to be the most significant factors in shaping demand out of the many variables

studied.

2.3 Traditional Techniques vs. Machine Learning Techniques

For the past few decades, traditional forecasting methods, including time series (extrapolatory)

and regression (explanatory) techniques, have been widely used in demand forecasting. Naïve,

moving average, trend, multiple linear regression, Holt-Winters, exponential smoothing and

ARIMA are among these traditional techniques. Recently, their performance has been used in

research to benchmark against those of advanced machine learning techniques, which have

10
gained attention and popularity in recent years due to the advancement in technology. For

example, Carbonneau, Laframboise & Vahidov (2008) performed studies on the application of

machine learning techniques such as support vector machine (SVM) and neural networks on

demand forecasting and compared the results with traditional methods including naïve, trend,

moving average and linear regression.

The emergence of big data, cloud computing and improved computing storage and processing

capabilities has led to increased availability and accessibility to large volumes of data, making

advanced machine learning techniques a viable option for demand forecasting in the industry.

Traditional and machine learning techniques differ in their capabilities and requirements.

Traditional time series and regression techniques normally consider either a single or a few

variables such as trend, seasonality and cycle. Machine learning-based techniques are able to

process an unlimited number of predictor variables, determining the ones that are significant.

The data source for traditional demand forecasting is mainly from demand history, while

machine learning-based techniques can make use of limitless data sources. However, this also

means that machine learning-based techniques are more reliant on the availability of data. The

more data there are, the better the learning will be. In traditional approaches, multiple single-

dimension algorithms are used separately for different product styles or categories based on

different data constraints. Thus, more manual data manipulation and cleansing work is required

and the algorithms are less generalizable. In machine learning, an array of general algorithms

is used to fit demand patterns across the entire product portfolio, creating a synchronized and

integrated forecast. In terms of technology requirements, machine learning is much more

dependent on computing power than traditional methods and may therefore be costlier to

implement.

11
Machine learning and predictive analytics provide an advantage over traditional forecasting

methods that use only limited demand factors to create more accurate demand forecasts.

Machine learning-based forecasting combines learning algorithms to identify underlying

demand drivers and uncover insights (Chase, 2017). Table 1 summarizes the comparison

between traditional and machine learning forecasting approaches.

Table 1. Comparison between Traditional and Machine Learning Forecasting Approaches

Traditional Forecasting Machine Learning Forecasting

Number of predictor variables Single or a few Unlimited
Data source Mainly demand history Multiple
Algorithms A number of single- An array of integrated algorithms
dimension algorithms
Manual data manipulation and High Low
cleansing need
Data requirements Low High
Technology requirements Low High

2.4 Application of Machine Learning Techniques in Industry

Machine learning techniques that have been applied in demand forecasting in research or

practice in the fashion apparel industry include neural networks, support vector machine

(SVM), fuzzy inference system (FIS), extreme learning machine (ELM), extended extreme

learning machine (EELM), harmony search (HS) algorithm and grey method (GM). In addition,

a hybrid combining different techniques tend to perform better than a single method. For

example, Wong & Guo (2010) proposed a model combining ELM and HS algorithm. The

proposed model performed much better than the traditional ARIMA model and certain other

neural networks models in making medium-term forecasts. Choi et al. (2014) also proposed a

hybrid model that produced satisfactory forecast accuracy results by utilizing a combination of

EELM and GM. Table 2 shows the industries that each technique has been applied to, the

12
preferred input variables and the forecasting horizon.

Table 2. Applications of Machine Learning Techniques in Forecasting in Different Industries

Machine Learning Industry Variables Horizon

Technique
Neural Networks Apparel Fashion, FMCG, POS, Order (Shipment), Short term
(NN) Medical Products Product attributes,
(Thomassey, 2010; Vhatkar & Consumer attributes
Dias, 2016)
Fuzzy Inference Apparel Fashion, Financial Price, Holidays, Period/ Long term
System (FIS) Forecast (Stocks), Technology Season, Financial time
Assessment series, Patent data,
(Thomassey, 2010; Kaya et al., Publication data and
2014) market research reports
Support Vector FMCG, Consumer Electronics Promotion, Number of Short term
Machine (SVM) (Pillo, Latorre, Lucidi, & opening hours, Price and
Procacci, 2016; Lu, 2014) number of daily receipts
(Forecast), Month,
Day of the month, Day of
the week, POS
Harmony Search (HS) Apparel Fashion POS Medium term
(Wong & Guo, 2010)
Grey Method (GM) Apparel Fast Fashion POS Short term
(Choi et al., 2014)
Decision Trees Apparel Fast Fashion Prototypes of sales, Long term
(Thomassey, 2010) Descriptive criteria of
historical items
Extreme Learning Apparel Fast Fashion, Apparel Long term forecasts, Last Short/
Machine (ELM) Fashion sales (At least 2 weeks) Medium term
(Wong & Guo, 2010; Choi et
al., 2014)
k-Means Clustering Apparel Fast Fashion Historical sales of a range Long term
(Thomassey, 2010) of products
Multivariate Adaptive Consumer Electronics POS (Sales amount,
Regression Splines (Lu, 2014) Trend, Growth ratios,
(MARS) Volatility)

13
3. Methodology
This section explains how we used the data collected to identify significant predictor variables

of sales and build the forecasting models. The objective is to find out how the data can be

leveraged to improve the demand forecasting capability, especially for seasonal products

without sales history. This section is structured as follows: We first describe the types of

machine learning methods used in feature selection and forecasting model building, and define

the scope and granularity of the data involved. We then move on to describe the process of

feature engineering and selection, and finally outline the steps in building two forecasting

models: the general model and the three-step model. The flow of the methodology is laid out

in Figure 1.

Figure 1. The Proposed Methodology

14
3.1 Machine Learning Techniques Used

This sub-section describes the types of machine learning techniques used in feature selection

and model building.

3.1.1 Supervised Learning Techniques

Supervised learning provides an algorithm with records that have a known output variable. The

algorithm “learns” how to predict this value with new records where the output is unknown.

The definition of each supervised learning technique used is listed below (Shmueli, Bruce,

Yahav, Patel, & Lichtendahl, 2018).

Regression and Classification Trees: Trees separate records into more homogeneous subgroups

in terms of the outcome variable by creating splits on predictors, thereby creating prediction or

classification rules. These splits create logical rules that are transparent and easily

understandable.

Random Forests: Random Forests combine the predictions or classifications from individual

trees by drawing random samples from the data and using a random subset of the predictors at

each run. The results are obtained either through voting for classification or averaging for

prediction.

Neural Networks: Neural networks mimic how human brain works and combine the predictor

information in a very flexible way that captures complex non-linear relationships among

variables. In neural networks, the user does not need to specify the correct form of relationship.

Instead, the network tries to learn about such relationships from the data. A feedforward neural

network consists of an input layer with nodes that accept predictor values, hidden layers that

15
receive inputs from previous layers and perform non-linear transformation, and finally an

output layer that classifies or predicts the outcome variable.

k-Nearest Neighbor (k-NN): k-NN classify or predict a new record by finding “similar” records

in the training data. k-NN identifies k records in the training data that are closest to the new

record in terms of predictor variables to derive a classification or prediction for the new record

by voting (for classification) or averaging (for prediction).

3.1.2 Unsupervised Learning Techniques

Unsupervised learning attempts to learn patterns in the data rather than predicting an output

value. In other words, there is no “correct answer” for the outcome. The definition of each

unsupervised learning technique used is described as follows (Shmueli et al., 2018).

k-Means Clustering: k-means clustering divides the data into a predetermined number k of non-

overlapping homogeneous clusters by minimizing a measure of dispersion within the clusters.

A common measure of within-cluster dispersion is the sum of distances (or sum of squared

Euclidean distances) of records from their cluster centroid.

t-distributed Stochastic Neighbor Embedding (t-SNE): This algorithm is one of the manifold

learning techniques. It is used to reduce the dimensionality of the data non-linearly, in a way

that helps visualizing the data points on a Euclidean space.

3.2 Scope and Granularity of Data

Two types of data were collected from the company: sell-in (shipment) and sell-through (POS)

data. The POS data collected were at the daily style-location level from 115 retail outlet stores

and include product attributes, calendar attributes, store attributes, price and promotion
16
attributes as well as the sales units. The total number of records in the sell-through data is

13,295,485, spanning a total of nine and a half seasons from July 2013 to March 2018. The

Spring/Summer season consists of January to June while the Fall/Holiday season consists of

July to December. Since the focus of this project is to support the decision of how much of

each style to order from the manufacturer for the whole season, the data were aggregated to the

level at which this decision is made; i.e., across all stores at the monthly level. The list of

attributes of the aggregated data is shown in Table 3.

Table 3. List of Attributes from the Aggregated Data by Month at the Style Level

Variable category Variable Description

Meta Data Style Unique style number of each product
Meta Data Style Description Description of the style
Calendar Year Fiscal year
Calendar Month Fiscal month
Product Attributes Color Color code
Product Attributes Basic Material Type of upper material
Product Attributes Gender Gender or age group description
Product Attributes Category Product family
Product Attributes Sub-Category Classic vs modern
Product Attributes Retail Outlet Sub-Department Basic vs. seasonal
Product Attributes Cut Ankle height
Product Attributes Pillar Product sub-brand
Product Attributes Product Class Product main feature
Price and Price Status Full-price vs mark-down
Promotion
Price and Manufacturer’s Suggested Retail Ticket price
Promotion Price (MSRP)
Price and Average Unit Retail (AUR) Actual selling price
Promotion
Sales Units Retail Sales Units (Target variable) Retail sales units

The products sold at outlet stores may either be discounted products from regular inline stores
17
or products made exclusively for launching at the outlet stores. In the context of demand

forecasting, we were only interested in the latter category. In addition, products with excess

inventory after the intended product lifecycle are discounted, and this distorts the demand.

Meanwhile, the sponsoring company is specifically interested in studying seasonal products

which typically have an intended lifecycle of 2 – 4 months. Therefore, records were removed

accordingly so that only records for outlet-exclusive products with full-price status and a

product lifecycle of 1 – 4 months were included in our analysis. In this case, product lifecycle

was estimated based on the POS data by counting the number of consecutive months with full

price sales records for a particular style. The data were pre-processed, filtered and aggregated

as described above using Alteryx software package.

3.3 Feature Selection and Engineering

Some features were modified or added in preparation for building the model. There are many

unique observations under the attribute color, some of which are very similar. In order to make

this attribute more meaningful, colors were aggregated into groups based on similarities.

Because it is commonly cited as one of the predictor variables in demand forecast, store count

was added as a candidate variable. It refers to the number of stores at which a style was sold,

which was estimated using sales record.

Pillar and Category are similar attributes with one-to-one relationship; i.e., they are completely

correlated with each other. Therefore, Pillar was dropped as Category already captured the

same information. The Retail Outlet Sub-department is the same across all seasonal styles and

was therefore dropped as well.

For building the forecasting model, three variables related to product lifecycle were added:

18
lifecycle, lifecycle month and lifecycle start month. As seasonal styles are launched at different

times of the year with short lifecycles, their sales are believed to be dependent on the lifecycle

attributes in addition to the calendar attributes; i.e., sales are not only related to which calendar

month the sale occurs in, but also to which month the product is launched. Lifecycle refers to

the total number of months in the lifecycle of a style. Lifecycle month refers to the number of

months since product launch. Lifecycle start month refers to the month the lifecycle started in.

The complete list of attributes subsequently being considered in the feature selection process

is shown in Table 4.
Table 4. List of Attributes for Feature Selection

Variable Category Variable Description

Meta Data Style Unique style number of each
product
Meta Data Style Description Description of the style
Calendar Year Fiscal year
Calendar Month Fiscal month
Product Attributes Color Group Color code
Product Attributes Basic Material Type of material
Product Attributes Gender Gender or age group description
Product Attributes Category Product family
Product Attributes Sub-Category Classic vs. modern
Product Attributes Cut Ankle height
Product Attributes Product Class Product main feature
Price and Promotion Manufacturer’s Suggested Retail Ticket price
Price (MSRP)
Price and Promotion Average Unit Retail (AUR) Actual selling price
Lifecycle Lifecycle The total number of months in the
lifecycle of a style
Lifecycle Lifecycle Month The number of months since
product launch
Lifecycle Lifecycle Start Month The month at which the lifecycle
started
Store Store Count Number of stores selling a style
Sales Units Retail Sales Units (Target variable) Retail sales units
19
Recursive feature elimination, a backward feature selection method, was used to eliminate

features based on their contribution to improving forecast accuracy. A random forests algorithm

was used on each iteration to evaluate the model with different subsets of the 14 input variables.

A 10-fold cross-validation on the training data was used. Random forests was selected in view

of its capability in handling multi-collinearity.

3.4 Dataset Partitioning

For building the general model, the data were partitioned into training and validation sets. The

data from the first six seasons (Fall/Holiday 2013 – Spring/Summer 2016) were used as training

set for building the model while the data for the next three seasons (Fall/Holiday 2016 –

Fall/Holiday 2017) were used as validation set for measuring the predictive performance of the

model. The number of styles and records in each data set is listed in Table 5.

Table 5. Overview of Datasets Generated for the General Model

Dataset Months of sales Number of Styles Number of records

Training 36 578 1796
Validation 18 195 560

For the three-step model, we split the database into three sets, a training set, a validation set,

and a testing set. For simplicity, the sponsoring company’s fiscal year (June – May) was the

factor used to split the data. The training set included all the sales records occurred before fiscal

year 2017, except for products with sales overlap in both fiscal years 2016 and 2017, which

were allocated to the validation set. For example, the records of a style that started selling in

April, fiscal year 2016 and continued selling through July, fiscal year 2017 was entirely moved

to the validation dataset to prevent data overlap. The validation set covered the sales records in

fiscal year 2017 and the overlap from 2016 plus seven months of records from fiscal year 2018.

20
The testing set included three months of records from fiscal year 2018. Table 6 gives an

overview of the three datasets generated for the three-step model.

Table 6. Overview of Datasets Generated for the Three-Step Model

Dataset Months of sales Number of Styles Number of records

Training 35 539 1558
Validation 19 201 591
Testing 3 58 155

3.5 Model Building

3.5.1 General Model

For seasonal styles without sales history, we built a general model utilizing product attributes,

calendar attributes, lifecycle attributes, store count and price attributes selected from the feature

selection process as described in Section 3.4. We explored using regression trees, random

forests, k-nearest neighbor (k-NN) and neural networks to build the model. In addition,

ensemble methods taking the median and average of the outputs from the four individual

methods were also considered. The prediction results from each method are compared in

Chapter 4.1.

3.5.2 Three-Step Model

The three-step model can be distinguished from the general model that it consists of three

separate stages: (i) clustering, (ii) classification, and (iii) prediction. The main objective behind

this model is to identify look-alike group of products from the training set. Once these products

are identified, their average sales can be used as a proxy to forecast the sales for brand-new

products in both the validation and testing sets.

21
In a similar fashion to the general model, the initial variables used in the three-step model were

those that resulted from the feature selection process. However, these variables were mixed

differently across the three stages. Additionally, two new variables were created from clustering

and then used in classification and prediction. Cluster number in the training set refers to the

cluster to which a style belongs. The average sales variable is calculated for a group of products

that belong to the same cluster and share similar lifecycle and calendar attributes. A complete

list of the variables considered for each stage is presented in Table 7.

3.5.2.1 Clustering
The main objective for the clustering stage was to partition and group all the seasonal styles in

clusters based on similarities across eight different attributes. The targeted data were a

combination of both the training and the validation sets. The only reason for including the

validation set was to later test the classification performance on a dataset (the validation set)

that had pre-assigned clusters. The clustering stage included four main sub-steps as illustrated

in Figure 2.

Figure 2. The Four Sub-Steps Followed in Clustering

Attributes selection. It’s essential to note that only numerical variables were used for

clustering since measuring distances between numerical data points is meaningful, while it is

not possible to measure distance between categorical ones. The eight attributes we used were:

lifecycle, manufacturer's suggested retail price (MSRP), average unit retail price (AUR) over

style lifecycle, average store count over style lifecycle and monthly sales over style lifecycle.

22
Table 7. List of the Variables Considered by Each Step of the Three-Step Model

Process Name Variable Name Variable Category

Clustering Lifecycle Lifecycle
MSRP Price and Promotion
Average AUR Price and Promotion
Average Store Count Store
Retail Sales Units Sales Units
Classification Lifecycle Lifecycle
MSRP Price and Promotion
AUR Price and Promotion
Store Count Store
Fiscal Year Calendar
Fiscal Month Calendar
Lifecycle Month Calendar
Lifecycle Start Month Calendar
Color Group Product
Basic Material Product
Gender Product
Category Product

Cut Product
Cluster Number Cluster
Prediction Lifecycle Lifecycle
MSRP Price and Promotion
AUR Price and Promotion
Store Count Store
Fiscal Year Calendar
Fiscal Month Calendar
Lifecycle Month Calendar
Lifecycle Start Month Calendar
Color Group Product
Basic Material Product
Gender Product
Category Product
Cut Product
Cluster Number Cluster
Average Sales Sales Units

23
Data normalization. To avoid the high level of influence that some variables like sales may

have over the others, the eight numerical variables were converted to the same scale by

subtracting the average attribute value from each member data point, then dividing it by the

standard deviation of the same attribute.

High dimensionality reduction. After normalizing the data, we used the t-SNE algorithm to

reduce data dimensionality in preparation for clustering.

k-Means clustering. Once the data were normalized and the data dimensionality were lowered

to two components only, we ran k-Means clustering algorithm to partition the data records into

k number of clusters.

3.5.2.2 Classification
By the end of the clustering stage, cluster numbers were assigned to the records of both training

and validation sets. Next, the classification stage was initiated to create a link between the

styles with pre-assigned cluster from the training set and brand-new styles from the validation

and testing sets. The classification drivers were both the categorical attributes and the numerical

attributes (except sales). The classification stage had three sub-steps, as illustrated by Figure 3.

Figure 3. The Three Sub-Steps Followed in Classification

Attribute selection. Besides the categorical and numerical variables that were preselected in

24
the feature selection process, the cluster numbers that resulted from the clustering stage were

also used in classification. Cluster numbers were treated as a target variable as the objective

was to match the records from the validation and testing sets with the clusters from the training

set.

Classification. Regression trees, random forests and SVM were the algorithms used for the

purpose of classification.

Accuracy evaluation. To evaluate the results of the three classification algorithms, we simply

compared the clusters allocated to the validation set against the pre-assigned clusters that

resulted from the clustering step.

3.5.2.3 Prediction
As the name indicates, the objective of the prediction stage is to predict the future sales for the

brand-new styles in the validation and testing sets. As illustrated in Figure 4, prediction had

three sub-steps.

Figure 4. The Three Sub-Steps Followed in Prediction

Attributes Selection. The variables used for prediction were the same ones used in

classification. Additionally, a new variable, average sales, was calculated for every record of

the training, validation and testing sets.

Prediction. To predict the sales for the products in the validation and testing sets, the following
25
five algorithms were tested and later compared for accuracy: regression trees, random forests,

neural networks, k-NN and linear regression.

Test & Score. Refer to Section 3.6.

3.6 Performance Measurement

We used two performance metrics for our forecasting models: forecast accuracy and bias. We

measured forecast accuracy using Weighted Mean Absolute Percentage Error (WMAPE) and

bias using Weighted Mean Percentage Error (WMPE). Absolute forecast error was first

calculated for each record at style-month level and then WMAPE was computed at both the

style-month and style-lifecycle level for reporting model performance. For seasonal products,

since the lifecycle is around 1 – 4 months, normally the entire purchase quantity is confirmed

prior to the beginning of the season. Therefore, we are interested in knowing the forecast

accuracy for the whole season instead of individual month. Equation 1 was used to calculate

the absolute forecast error for each record. We then used Equation 2 to calculate the forecast

accuracy at either the monthly or lifecycle level by aggregating the MAPE of each style-month

or style-lifecycle weighted by the sales units. We finally used Equation 3 to calculate the

forecast bias in a similar fashion.

𝐴𝑏𝑠𝑜𝑙𝑢𝑡𝑒 𝐹𝑜𝑟𝑒𝑐𝑎𝑠𝑡 𝐸𝑟𝑟𝑜𝑟 = |𝐹𝑜𝑟𝑒𝑐𝑎𝑠𝑡𝑒𝑑 𝑆𝑎𝑙𝑒 − 𝐴𝑐𝑡𝑢𝑎𝑙 𝑆𝑎𝑙𝑒𝑠| [1]

∑𝑛𝑖=1|𝐹𝑜𝑟𝑒𝑐𝑎𝑠𝑡𝑒𝑑 𝑆𝑎𝑙𝑒𝑠 − 𝐴𝑐𝑡𝑢𝑎𝑙 𝑆𝑎𝑙𝑒𝑠| [2]

𝐹𝑜𝑟𝑒𝑐𝑎𝑠𝑡 𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦 (𝑊𝑀𝐴𝑃𝐸) =
∑𝑛𝑖=1 𝐴𝑐𝑡𝑢𝑎𝑙 𝑆𝑎𝑙𝑒𝑠

∑𝑛𝑖=1 𝐹𝑜𝑟𝑒𝑐𝑎𝑠𝑡𝑒𝑑 𝑆𝑎𝑙𝑒𝑠 − 𝐴𝑐𝑡𝑢𝑎𝑙 𝑆𝑎𝑙𝑒𝑠 [3]

𝐹𝑜𝑟𝑒𝑐𝑎𝑠𝑡 𝐵𝑖𝑎𝑠 (𝑊𝑀𝑃𝐸) =
∑𝑛𝑖=1 𝐴𝑐𝑡𝑢𝑎𝑙 𝑆𝑎𝑙𝑒𝑠

26
Since the company did not keep records of previous forecasts, the baseline forecast accuracy

was estimated using sell-in data, assuming shipment for a season was equal to its demand

forecast. It is not feasible to allocate shipments to sales at a monthly level, so the baseline

forecast accuracy was estimated on a lifecycle level for each style. If we consider only sales at

the full price status, the WMAPE is over 100%. If both full price and markdown sales are

considered, the WMAPE is 16%. In this case, both numbers are not directly comparable to our

model results. However, this is the best reference we have regarding the company’s forecasting

performance.

27
4. Results
This section reports and analyzes the results of feature selection and the two types of

forecasting models: general model and three-step model.

4.1 General Model

Based on the results of recursive feature elimination, the model with 12 variables resulted in

the lowest error as shown in Figure 5. The list of variables in the order of variable importance

is shown in Table 8. These variables were used to build the subsequent forecasting models

while the remaining two variables (category and sub-category) were dropped. The top six

attributes account for the majority of the variances. Store count, month and lifecycle month are

the top three attributes. Among the product attributes, gender, basic material and color group

are the top three attributes.

Table 8. List of Attributes Selected for Model Building

Importance Rank Attribute Attribute Category

1 Store Count Store
2 Month Calendar
3 Lifecycle Month Lifecycle
4 Gender Product
5 AUR Price and Promotion
6 Year Calendar
7 Basic Material Product
8 MSRP Price and Promotion
9 Color Group Product
10 Lifecycle Lifecycle
11 Cut Description Product
12 Product Class Description Product

28
Figure 5. Cross Validation Error by Number of Attributes

The results in terms of forecast accuracy of the four individual models using regression trees,

random forests, k-NN and neural networks and the two ensemble models using the median and

average of individual outputs are shown in Table 9, while the forecast bias are shown in Table

10. MAPE were calculated on both the style-month and style-lifecycle level.

Considering the individual models, random forests gives the best predictive performance on

the validation data with the highest accuracy and lowest bias. It achieved 37% WMAPE on the

style-lifecycle level and 47% on the style-month level with a negative bias of 2%. Although

the regression trees model has a slightly higher WMAPE and also tends to under-forecast, it

provides better interpretability and visually gives insights into which predictor variables are

more significant. Store count appears at the top of the tree, indicating that it is the most

important attribute in predicting demand. Examining the first three layers of the tree, we can

see that store count, month and lifecycle month are at the top in terms of feature importance.

This is in line with our findings in the feature selection process. k-NN also gives reasonably

good results in terms of forecast accuracy. However, it is worth noting that k-NN will only

predict results within the range of the training data, since it is simply searching for the nearest

29
k neighbors and predicting sales to be the average of those of the nearest neighbors. Therefore,

it may not work as well if the new data are not in the same range as the training data. As for

the ensemble methods, taking the median and average of the individual model outputs yields a

better forecast accuracy with a WMAPE of 35%. Neural networks show the worst performance

in terms of both accuracy and bias, with a 49% WMAPE and a positive bias of 19%, indicating

a tendency to over-forecast.

Table 9. Forecast Accuracy of the General Model

Regression Random k-NN Neural Median Average

Trees Forests Networks
WMAPE 38% 37% 39% 49% 35% 35%
(Lifecycle)
WMAPE 49% 47% 50% 56% 45% 45%
(Monthly)

Table 10. Forecast Bias of the General Model

Regression Random k-NN Neural Median Average

Trees Forests Networks
WMPE -12% -2% -2% +19% -2% +1%
(Lifecycle)
WMPE -12% -2% -2% +19% -2% +1%
(Monthly)

4.2 Three-Step Model

This section discusses the results from each of the stages followed in the three-step model:

clustering, classification and prediction.

4.2.1 Clustering and Classification

To determine the right number of clusters that fit our data, we used the silhouette score

(silhouette coefficient). The silhouette score is a method of interpretation and validation of

30
consistency within clusters of data. It measures and compare the mean intra-cluster distance to

the mean nearest-cluster distance for each data point within a cluster. The silhouette score

ranges between -1 (wrong clustering) and +1 (best value) with 0 indicating overlapping clusters.

For our dataset, the number of clusters that revealed the highest silhouette score was seven as

illustrated in Table 11.

Table 11. The Silhouette Score for Different k Number of Clusters

Number of Clusters Silhouette Score

2 0.378
3 0.422
4 0.481
5 0.469
6 0.487
7 0.489
8 0.463
9 0.441
10 0.427

In addition to the silhouette score, we took an additional measure to verify which number of

clusters work best for our dataset. Using the training and validation sets, we ran a classification

exercise for each k from the table above, then compared the clusters that resulted from

classification against those that resulted from clustering. The number of clusters that revealed

the best classification match was five, with an overall accuracy of 93%. Looking up the

silhouette score for five versus seven clusters, the difference is minimal. Figure 6 shows the

classification results using five clusters and SVM algorithm. The number 590 is the total

number of records included in the validation set. The vertical axis represents the number of

records that were pre-assigned to each of the five clusters (C1 to C5) based on clustering while

the horizontal axis is the number of records allocated to each of the five clusters based on

31
classification.

Figure 6. The Confusion Matrix Resulted from Using Five Clusters and SVM Algorithm

The best performing classification algorithm was SVM. Table 12 compares the classification

accuracy based on the three algorithms used.

Table 12. Comparison of the Classification Accuracy by Algorithm

Algorithm Name Overall Accuracy

SVM 93%

Random Forests 89%

Regression Trees 71%

Additional analyses were performed on the validation set to better understand how clusters

were allocated based on lifecycle, sales volume, AUR and store count. The main insights

uncovered from these analyses are displayed in Table 13. The analyses showed that lifecycle

length was a clear distinguishing driver between the five clusters. Each cluster included one

lifecycle length, except Cluster 4, which included styles with mixed lifecycles. However, the

styles that were included in Cluster 4 seemed to have relatively smaller sales volume, smaller

store count and higher AUR compared to the other clusters. Both clusters C1 and C3 included

styles with a three-months lifecycle. However, C3 had smaller sales volume and lower AUR
32
compared to C1.

Table 13. Characteristics of Styles Distribution among Clusters

Cluster Number of Lifecycle Average Monthly Average Store
Records (Months) Sales (Units) AUR Count
C1 108 3 1298 $38 88
C2 112 2 953 $30 79
C3 84 3 839 $24 83
C4 106 2, 3, 4 462 $44 37
C5 180 4 958 $31 82

4.2.2 Prediction
Running the prediction algorithms on both validation and testing sets resulted in relatively

different results in terms of best performing algorithms. However, the forecast accuracy of the

ensemble methods for both datasets was much closer. Overall, the testing set had slightly better

forecast accuracy and worse forecast bias compared to the validation set. Figure 7 and Figure

8 display the forecast accuracy and forecast bias for each of the two datasets.

Starting with the validation set, both k-NN and random forests delivered the highest forecast

accuracy on a style-lifecycle level with a WMAPE of 37%. However, k-NN had no forecast

bias compared to 4% under-forecast bias by random forests. Neural networks were third-best

in forecast accuracy on a style-lifecycle level with 39% WMAPE and worst in forecast bias

with -27% WMPE (under-forecast). Finally, regression trees had 40% of WMAPE on a style-

lifecycle level and 12% over-forecast bias. The two ensemble methods, average and median,

delivered the best overall results in forecast accuracy on both a style-month (43%) and a style-

lifecycle (34%) levels with a forecast bias around 1% (under-forecast).

For the testing set, regression trees delivered the best performance with a forecast accuracy of

31% (style-lifecycle level) and 2% of forecast bias (over-forecast). Linear regression showed
33
a very close performance with 31% of forecast accuracy (style-lifecycle level) and 3% of

forecast bias (over-forecast). The neural networks’ performance in the testing set was slightly

better compared to the validation set with a forecast accuracy of 33% (style-lifecycle level).

However, the neural networks’ forecast bias (23% under-forecast) was still relatively high

compared to the other four algorithms. Random forests and k-NN had the lowest forecast

accuracy (around 36%) while they had a forecast bias of 7% and 18% (under-forecast),

respectively. Like in the validation set, the ensemble methods also delivered the highest overall

forecast accuracy with 30% WMAPE on a style-lifecycle level. However, their forecast bias

was relatively high (around 8% under-forecast) compared to the regression trees and linear

regression algorithms.

Figure 7. Forecast Accuracy and Bias of the Three-Step Model (Validation Set)

34
Figure 8. Forecast Accuracy and Bias of the Three-Step Model (Testing Set)

In addition to the analyses above, we clicked down into cluster level to understand how the

model performs from one cluster to another. For the validation set, random forests was the best

performer in clusters C2, C3 and C5 with a forecast accuracy of 36%, 37% and 34%,

respectively (see Table 14). The forecast bias for random forests was -11% (C1), -1% (C3) and

+3% (C5). It’s essential to note that these three clusters (C1, C3 and C5) share relatively similar

store count and average monthly sales. They only differ in the length of lifecycle. k-NN was

the best performer in cluster C1 with a forecast accuracy of 28% and a forecast bias of 4%

(over-forecast). For cluster C4, regression trees performed best with a forecast accuracy of 45%

and forecast bias of 30% (under-forecast). The relatively bad performance in cluster C4 could

be linked to the complexity of this cluster including multiple lifecycle lengths, low monthly

sales volume and high AUR on average.

The testing set included three clusters: C1, C3 and C4. Those were the clusters that resulted

from the classification stage. Similar to the validation set, random forests and regression trees

delivered the highest forecast accuracy and lowest forecast bias in cluster C3 and cluster C4

(see Table 15). In cluster C1, k-NN revealed the best forecast accuracy (28%) and lowest

35
forecast bias (4% over-forecast).

On a cluster level, the performance of the two ensemble methods didn’t always outperform the

performance of the individual algorithms.

Table 14. Best Performing Algorithm by Cluster (Validation Set)

Cluster Best-Performing Forecast Forecast Best-Performing Forecast Forecast

Algorithm Accuracy Bias Ensemble Accuracy Bias
C1 k-NN 28% +4% Average 26% -4%
C2 Random Forests 36% -11% Median 38% -10%
C3 Random Forests 37% -1% Median 33% 0%
C4 Regression Trees 45% -30% Median 51% -35%
C5 Random Forests 34% +3% Average 33% +14%

Table 15. Best Performing Algorithm by Cluster (Testing Set)

Cluster Best-Performing Forecast Forecast Best-Performing Forecast Forecast

Algorithm Accuracy Bias Ensemble Accuracy Bias
C1 Linear Regression 28% -11% Median/Average 29% -22%
C3 Random Forests 32% +6% Average 33% +11%
C4 Regression Tree 39% 0% Median 39% 0%

36
5. Discussion
This section discusses implications of the model results, then moves on to describe limitations

of our model, and finally outlines some future research opportunities.

5.1 Implications

In evaluating the suitability of the models for the sponsoring company, the ease of

implementation was considered in addition to the models’ predictive performance. As a starting

point, the general model serves as a good framework for an immediate implementation,

outperforming the company’s current forecasting model in terms of forecast accuracy and bias.

Among the different methods tested in the general model, the ensemble methods (median and

average) and random forests gave the best predictive performance, thus are the methods that

we recommend using when implementing the general model.

The three-step model through the clustering and classification stages offers visibility into the

underlying factors that impact demand. With this model, forecasting can be customized to

deliver best possible results based on product characteristics such as planned lifecycle, store

number and retail price. Regression tree is what we recommend applying on complex clusters

with multiple lifecycle lengths. Random forests is the algorithm we recommend using on

clusters with mono lifecycle, while k-NN and linear regression are what we recommend using

on similar clusters but with higher sales volume and AUR.

5.2 Limitations
Due to the limitation in the inventory data available, lost sales were not considered in building

our forecast models. Inventory data were provided at the monthly and style level. As a result,

we only have one snapshot of the inventory level for each month and at an aggregated level
37
across all sizes of a style. It is therefore not feasible to estimate lost sales, which is needed to

reflect true demand not captured in POS data.

Since the company did not keep records of previous forecasts, direct comparison between our

model performance and the company’s current forecast performance is not possible. Going

forward, we recommend that the company keep track of forecast history so forecasting

performance can be measured and improvement areas can be identified accordingly.

The intended product lifecycle was estimated based on the POS data by counting the number

of consecutive months with full price sales records for a particular style. In practice, the

intended product lifecycle will have to be pre-determined in order to be inputted into the

forecasting model. There may be some difference between intended and actual product

lifecycle.

Store count, which refers to the number of stores that a style is carried in, was estimated using

sales record. In the case where inventory is available but there are no sales, store count will be

overestimated. In practice, the pre-determined store count should be used as an input since the

store count based on actual sales will not be available at the time of forecast.

Due to the complexity of the promotional data provided, price promotions were not used as an

explicit attribute, rather they were embedded as a change in the AUR. Price promotions play a

major role in driving demand and capturing this component explicitly may help improve the

achieved forecast accuracy.

5.3 Future Research

In future research, the company should consider incorporating lost sales, a more accurate
38
measure of intended product lifecycle, as well as store count in building and evaluating the

forecast model. The company may also consider extending future research to cover shorter

forecast horizons and higher data granularity. While this project focuses on roughly a five-

month range forecast for placing orders to manufacturers, there are opportunities to dive deeper

into the data at the store and weekly level, for the purpose of store inventory allocation and size

curve analysis. In addition, the relationship between price and demand could be studied for

price optimization.

39
6. Conclusion
Our research project proposed a methodology that offers two different forecasting models

based on machine learning techniques. These models will enable the company to achieve better

forecast accuracy compared to the current performance by considering store count, lifecycle,

calendar and product attributes simultaneously.

The data pre-processing phase of the proposed methodology is an important stage that

facilitates the formation of the inputs to the models. The feature engineering process helps

create new variables that bring additional value to demand interpretation. The feature selection

process allows us to gain insights into the importance of the different predictor variables and

their influence on forecast accuracy. Another value proposition of this phase is the possibility

of using, processing and delivering value out of the categorical variables that have always been

considered a challenge when it comes to forecasting demand in the fashion industry.

When it comes to the models, the general model serves as a starting point for easy

implementation of the machine learning forecasting framework. The three-step model

involving clustering, classification and prediction enables the company further to visualize the

relationship between predictor variables and customize the forecasting approaches accordingly.

Finally, the project opens doors for further research that possibly cover store inventory

allocation, size curve analysis and price optimization.

40
References
B., & Thomassey, S. (2016). Intelligent demand forecasting systems for fast fashion.
Information Systems for the Fashion and Apparel Industry, 145-161. doi:10.1016/b978-0-08-
100571-2.00008-7

Carbonneau, R., Laframboise, K., & Vahidov, R. (2008). Application of machine learning
techniques for supply chain demand forecasting. European Journal of Operational Research,
184(3), 1140-1154. doi:10.1016/j.ejor.2006.12.004

Chase, C. W., Jr. (2017). Machine Learning is Changing Demand Forecasting. Journal of
Business Forecasting,43-45.

Choi, T., Hui, C., Liu, N., Ng, S., & Yu, Y. (2014). Fast fashion sales forecasting with limited
data and time. Decision Support Systems, 59, 84-92. doi:10.1016/j.dss.2013.10.008

Lu, C. (2014). Sales forecasting of computer products based on variable selection scheme and
support vector regression. Neurocomputing, 128, 491-499.
doi:10.1016/j.neucom.2013.08.012

Kaya M., Yeşil E., Dodurka M.F., Sıradağ S. (2014) Fuzzy Forecast Combining for Apparel
Demand Forecasting. In: Choi TM., Hui CL., Yu Y. (ads) Intelligent Fashion Forecasting
Systems: Models and Applications. Springer, Berlin, Heidelberg

Pillo, G. D., Latorre, V., Lucidi, S., & Procacci, E. (2016). An application of support vector
machines to sales forecasting under promotions. 4Or, 14(3), 309-325. doi:10.1007/s10288-
016-0316-0

Rashad, A., & Spraggon, S. (2013). Assembling the crystal ball: using demand signal
repository to forecast demand (Unpublished master's thesis).

Shmueli, G., Bruce, P. C., Yahav, I., Patel, N. R., & Lichtendahl, K. C. (2018). Data mining
for business analytics concepts, techniques, and applications in R. Hoboken, NJ, USA: Wiley.

Thomassey, S. (2010). Sales forecasts in clothing industry: The key success factor of the
supply chain management. International Journal of Production Economics, 128(2), 470-483.
doi:10.1016/j.ijpe.2010.07.018

41
Vhatkar, S., & Dias, J. (2016). Oral-Care Goods Sales Forecasting Using Artificial Neural
Network Model. Procedia Computer Science, 79, 238-243. doi:10.1016/j.procs.2016.03.031

Wong, W., & Guo, Z. (2010). 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.
doi:10.1016/j.ijpe.2010.07.008

42