The University of Southampton

Academic Year (2021/2022)

Faculty Of Social Science

Southampton Business School

MSc Dissertation

Modelling Product Returns Behaviour at Level of Consumers

(ERGO reference number:76487)

(Student registration number:32865023)

Presented for MSc. Business Analytics and Finance

This project is entirely the original work of student registration number 32865023.
Where
material is obtained from published or unpublished works, this has been fully
acknowledged by citation in the main text and inclusion in the list of references

Word Count:14866
Contents

1. Introduction ................................................................ 1
2. Literature Review ........................................................... 5
2.1. Product Returns Behavior ............................................ 6
2.1.1. Reasons For Product Returns ..................................... 6
2.1.2. Return Fraud .................................................... 13
2.1.3. Return Policy and Actions to Reduce Return ...................... 14
2.2. Quantitative Analytics Model of Product Returns ...................... 16
3. Methodology............................................................... 21
3.1 Understanding A Problem and Final Goal ............................. 22
3.2 Data Collection ..................................................... 22
3.3 Data preparation and preprocessing .................................. 23
3.3.1 Methods Introduction ............................................ 23
3.3.2 Implementation procedure ....................................... 24
3.4 Modeling and Testing ............................................... 30
3.4.1 methods introduction ............................................ 30
3.4.2 Implementation procedure ....................................... 32
3.5 Model optimization .................................................. 34
3.5.1 Methods introduction ............................................ 34
3.5.2 implementation procedure of optimization ......................... 35
4. Results .................................................................... 40
4.1 Results of Data Preprocessing ....................................... 40
4.2 Modeling Results ................................................... 41
4.2.1 Logistic Regression ............................................. 41
4.2.2 Results of naïve Bayes ........................................... 43
4.2.3 Results of Neural Network Classifier ............................... 43
4.3 Model Optimization.................................................. 44
4.3.1 Optimized Logistic Regression ................................... 44
4.3.2 Optimized Naïve Bayes Model .................................... 45
4.3.3 Optimized Neural Network ........................................ 46
4.3.4 Comparison of Predictive Methods ............................... 46
5. Discussion ................................................................ 47
5.1 Discussion of Three Machine Learning Models ........................ 47
5.2 The Limitations of This Study ........................................ 48
6 Conclusion ................................................................ 51
6.1 Main conclusion .................................................... 51
6.2 Suggestions For Reducing Product Returns .......................... 52
6.3 Suggestions For Future Research .................................... 53
References .................................................................... 54
Abstract

Due to the continuous increase of online sales, the number of product returns has also

increased significantly, which has also generated corresponding return costs. For retailers,

the factors affecting the return of products need to be clear, so that timely responses and

measures should be taken. From the perspective of demographic characteristics, this study

explores the impact of demographic characteristics, including purchasing power and average

age, on product returns. In other words, at the customer level, what characteristics of

customers are more inclined to return. In the process of data analysis, we use Logical

Regression and Naive Bayes, which are suitable for binary classification problems, and

Neural Networks that can perform well in various analyses. We use these three models to

predict the two independent variables, purchasing power and average age, and the

dependent variables (binary data: 1 means return, 0 means no return). The results suggest

that these two features do have an impact on product return.

Keywords: Product Return, Demographic Characteristics, Customer, Machine Learning

 Dissertation

1. Introduction
Nowadays, customers can buy any product anywhere as online sales grows increasingly in

these years. From 2017 to 2021, online sales grew up from £295.9 billion to £465.4 billion in

Europe (Online Shopping Behavior in The United Kingdom (UK),2021). The e-commerce

market has increased rapidly the last couple of years. E-commerce players offer free shipping

and hassle-free returns because of the high competition in e-commerce, and because of this,

the number of product returns increased dramatically. Compared with offline shopping, online

shopping cannot provide the real product so that when making a purchase decision, a

customer cannot physically inspect, feel or touch a product to resolve issues associated with

it. (Yan and Cao, 2017). In order to discover if a product is of interest to them, they may

review the product description, view an image of the product, or rate the product on the

website. Moreover, customers may be able to realize the value of a product and avoid the

purchase of an unsatisfactory product as a result of product quality or product performance

(Yoo and Seung, 2014). However, it is still a possibility to buy an unsuitable product to return

it. Most retail businesses provide free delivery and multiple ways for returning items -

including returning items bought online to a physical store without charge. In the Financial

Times (Ram, 2016), it was reported that certain businesses reported up to 70% returns, which

is higher than the typical returns rate of 20–40%.

Due to escalating rates of product returns, most firms have no choice but to incur these extra

costs. Online purchasing has resulted in an increase in product returns, which show no signs

of slowing down. There is no doubt that product returns lead to many operation and profit

problems. For example, transaction costs, repair costs, supply chain management,

inventory management and so on, such measures bringing to great management costs to

enterprises. A growing number of studies have examined the costs of product returns and

reverse logistics over the last 20 years. Existing research show costs of handling returns are

much higher than that of delivering products even costs for returns are more than the cost to

manufacture (Frei et al., 2020). According to another study by Cullen et al. (2013), even if a

1
 Dissertation

company reported a return rate of 70%, each item sold in such a case is effectively causing

the company to make a loss each time (Cullen et al., 2013).

We must highlight that the various return routes make it challenging for companies to track

products sold in stores or online, which is a big challenge to supply chain and inventory

management and need a longer time to solve. There are many disciplinary approaches to the

product returns problem, including customer behavior control, marketing, and advertising,

purchasing, supply chain management, customer service, analytics, and strategic operations

management, as well as circular economy, product design, material science, and waste

management (Frei et al., 2016).

In order to attract more loyal customers and provide competitive customer service, many

online stores offer unconditional product returns and will not charge a fee for returning, which

results in more and more product returns. However, even return fraud appears as a loose

return policy. Return fraud is the act of buying goods with the intention of returning them,

which is a form of informal or illegitimate borrowing. For example, some fashion blogger post

outfit pictures on social media attracting many young girls to simulate their actions by

illegitimate borrowing cloth and necessaries on the online stores; It has even been reported

that some sports enthusiasts purchase large televisions to observe a sporting event over the

summer holidays and then return them (Frei et al., 2016). By abusing the return policy,

customers can return a product they bought with the intention to return it at the time of

purchase. In doing so, compared with abusive customers who just pay at little or no cost no

matter what in aspects of physical, experiential, or financial, retailers have to cost much for

this return action even disrupting their management system. For instance, According to

estimates, retailers in the United States alone incur $5.6 billion in costs each year because

of return abuse (Ketzenberg et al.,2020). Some retailers acknowledged such issues but

struggled to avoid them due to the dilemma of rejecting improper returns while providing high

levels of customer services (Jack et al.,2019). Some major retailers, like Victoria's Secret,

manage their return process by using third-party customer profiling service (Ketzenberg et

al.,2020). Returning merchandise requires the checkout clerk to scan the customer's ID along

2
 Dissertation

with the original transaction receipt. A third-party service provider combines that information

with the customer's transaction history to recommend whether to reject, warn, or accept the

return (Speights and Rittman, 2015). A lack of academic studies supporting the

implementation of better returns systems is present despite the recent increase in the

academic literature on product returns.

On another side, increasing product returns leads to increased transportation and product

waste, both of which affect the environment, including continual demand for products, placing

strain on the Earth's resources, and resulting in issues like plastic waste in the food chain,

and the possibility of catastrophic effects on marine and human life. Additionally, over the

years, the widespread use of palm oil, wood, plastic, coconuts, and several other items over

the years has led to the decimation of our rain forests, the elimination of biodiversity, the

degradation of air quality, and health risks for humans (Baden and Frei, 2021). Deforestation,

habitat loss, loss of biodiversity, pollution, congestion, and toxic waste are also environmental

impacts of the production and transport of goods (Sala et al.,2019). After COVID-19 broke

out in 2020, we were encouraged to consume to revive the economy, causing shops to close,

high streets to empty, jobs to disappear, and GDP to plunge. Paradoxically, the more you

buy, the more returns you will get, which will also cause more waste of resources and

environmental pollution. There is no doubt that product returns not only have a lot of profit

impact on retailers, but also pose a great challenge to environmental protection and

sustainable development. This is a reason why we focus on product return so that more

research can be applied to realize product return and customers’ psychology and to provide

effective suggestions to retailers to reduce costs and protect the environment.

In order to reduce product return, retailers have applied much measures and many research

focuses on this specific. One of the most important is to identify who will return products and

why customers return them. In this research, the goal is to address that confusion in terms of

customer’s product return behavior rather than in terms of product, retailer, and manufacturer.

Specifically, this paper focus on the effects of customers’ demographic characteristics such

as age and purchasing power.

3
 Dissertation

It is a diverse set of reasons that customers return products. However, it is very difficult for

retailers to obtain data on demographic characteristics, so that it is limited to analyzing the

impact of demographic characteristics on product returns. Our research is to identify the

demographic characteristics that are manifest in customers who are more likely to return

products, and explain which factor is most significant to product return. Therefore, we use the

Logistic Regression model, Naïve Bayes and Neural Network to quantitative analyze the

consumer personality of who is used to product return by using the data of second-hand

electronic products in Europe and apply qualitative analysis to advise retailers on how to

reduce return rate for decreasing cost of return and protecting the environment. Specifically,

this paper attempt to achieve the goal of the study by considering the following research

question:

1 In terms of demographic characteristics such as age and purchasing power , which factor

might affect customer’s return the products?

2 Which factor has greater influence?

To answer these questions, any such Machine Learning model will need to (1) identify

demographic characteristics that might have an impact on product returns and(2) make

classifications or predictions of customers over time with a certain level of accuracy.

In this section, we briefly introduce why we select this field to research and clarify what

question we can solve and what technique will be used. In the next section, we give a detailed

overview of the literature on product returns and identify areas that have not been studied.

Meanwhile, we will review and summarize the research progress of product return and find

out the current research gaps to help us better fill the gaps in this field. In the third section,

the methodology adopted in this paper will be outlined. Including data selection, data cleaning,

data preprocessing, model selection, model evaluation and model optimization. The purpose

of this research was to explore the potential of Machine Learning, which can learn from data

and make predictions. To create a robust and resilient return process, Machine Learning can

be a successful technology that utilizes the enormous data capacity of the customers.

4
 Dissertation

From the fourth to the sixth section, we summarize the analysis results and gain some

valuable insights based on the data. We will extract the factors that affect the return behavior

from the analysis results and put forward some basic solutions for retailers. However, in terms

of actual creativity, it may be necessary to refer to specific issues. Furthermore, we explain

the limitations of this research and critically think about the strengths and weaknesses of this

research and the direction of future development for reference.

2. Literature Review

The literature review is divided into two parts, first of all, it summarizes the research on the

influence of consumer return behavior so far, including various behaviors and factors that

affect returns. Such as product quality, product price, lower than expected and so on. On the

other hand, this part summarizes the existing research conclusions and future research

directions from the return policy, return fraud, demographic characteristics, and product

categories. For the problems that have been found, of course, corresponding solutions should

be put forward. Consequently, we found out what methods retailers have taken to deal with

and solve various problems arising from returns. The second part is to summarize the

quantitative analysis methods that have been used in product returns so far. Suitable analysis

tools and analysis results can be summarized according to previous studies. This is essential

for us to use quantitative analysis methods in this study.

Product Returns refer to a process in which customers take previously purchased

merchandise back to a retailer, from which they receive a refund, an exchange, or store credit

as a result. Product return is critical not only in terms of retailers because of uncertainty

related to price, demand, and quality of the product, but also in terms of suppliers, customers

and the environment. First, the return will reduce profits tosome degree, which must influence

the bargaining power of retailers in the whole industry. Second, product return will waste

customers’ energy and money so as to lose their customer for this brand and retailer. Third,

the process of product return will generate many wastes of packages and transportation

5
 Dissertation

which is a big challenge for our environment. Due to so many influences in different aspects,

researchers focus on this field early and have got some achievements.

Reasons for product return can be categorized into 3 types in the product lifecycle, including

manufacturing, distribution and customer returns. Manufacturing return refers to returns from

surplus materials damage, scrap and so on. Distribution returns are initiated from external

return sources, such as product recalls, damage return, wrong delivered the product, and

stock adjustments. A product can already be returned from the manufacturer to the raw

material producer. Of course, product returns might happen from retailers to their supplier

(Ambilkar et al., 2021). This research, however, does not specifically focus on return flows

from manufacturers, as they occur less frequently than returns from customers. Customers

return, which this research focuses, indicate a return process from the end of the consumer,

like product failure, lower product quality as expectated, unsuitable products, wrong delivered

product, damaged package, and fraud return. Returns may occur at every stage of product

sales.

2.1. Product Returns Behavior

2.1.1. Reasons For Product Returns

According to research (Wood, 2001), when deciding to purchase online, the decision involves

two steps: either ordering or not ordering, which involves a high degree of uncertainty

because the customer cannot assess the quality of the product (Bonifield et al., 2010). In the

second phase, the customer decides whether to keep or return the items (Wood, 2001).

Therefore, the customer will inspect all aspects of product information they can get. In an

online purchase, the retailer describes product information on the website, like price, quality,

appearance, and usage. Customers make a purchase decisions based on these information

but cannot experience a real product before receiving it. A decision that whether they accept

this product will be made when they receive the package (Teo and Yeong, 2003).

Consequently, online retailers strive toward two goals to maximize sales: a high order
6
 Dissertation

intention during the first stage and a low return intention during the second stage

(Gelbrich,2017).

The simplest return behavior is when the product cannot fulfill needs. Petersen and Kumar

(2009), 5% of all returns are because of defective products or incorrectly sold products.

According to Pei and Paswan (2018), they suggested several reasons may lead customers

to return a product if it is unsatisfactory, either internally (e.g., change of mind due to

impulsiveness) or externally (e.g., the product does not meet expectations, perceived risk of

keeping the product, negative attitude by social group). Therefore, in this research,

Confirmation/Disconfirmation can be used to explain return behaviors that are caused by

unsatisfactory purchases.

Quality and price of products are related to the possibility of return rate. Customer satisfaction

is reduced when low-quality products and services are provided, which will result in product

returns. Meanwhile, providing high-quality products and services is rewarded with higher

selling prices (Kirmani and Rao, 2000). Another study finding (Jiang et al., 2005) indicates

that customer satisfaction after delivery has a much higher influence on both overall customer

satisfaction and repurchase intentions than satisfaction at checkout, and that price perception

directly influences customer satisfaction and repurchase intentions. Meanwhile, A reduction

in price (Anderson et al. 2009; Petersen and Kumar 2009) makes it less likely for products to

be returned from a marketing-mix perspective. A decrease in customer demand can also be

a result of an increase in price, especially when the consumer demand for a particular product

is price-sensitive (Li et al., 2013). Recently study (Fan et al.,2022) indicates that the majority

of returns even occur for other reasons, not defective products.

Even though The reasons for product return mentioned above, such as quality, failure to meet

demand and express delivery problems, have been improved by improving quality and

enhancing the role of online feedback, consumer product returns have been on the rise

(D'Innocenzio, 2011), which suggests that dissatisfaction, product failure, or dishonest intent

are not the only reasons for consumer product returns. Hence, the topic of product return is
7
 Dissertation

focused persistently. At the same time, insight research (Petersen and Kummar, 2012) shows

that taking into account the characteristics of products and customers, If expectations are

high, there is a higher likelihood of purchasing and returning the product; if expectations are

low, there is a lower likelihood of purchasing and returning the product. Similarly, some

customers have buyer’s remorse and change their minds after purchase and as a result, they

return the products.

In terms of information related to products, Cuffie (2020) suggested that providing better

descriptions of these characteristics could shrink the gap between customer expectations

and the reality of the product offered. Additionally, providing customers with high-tech tools

to bring them closer to the product would help decrease the number of returns. Minnema et

al. (2016) suggested product reviews are related to product return. More specifically, a

positive review of a product will not result in an exactly positive impact on a potential customer.

Therefore, retailers shouldn’t just encourage very satisfied customers to write reviews.

Because excessive positive reviews increase the probability of purchase, the negative impact

on the probability of product returns cannot be offset. By knowing how many other people

have experienced the product, a buyer can be less uncertain about the product itself (Babić

et al.,2016). Lower uncertainty then leads to lower return probability. In other words, review

volume is expected to lower return probability. Researchers (Sahoo et al.,2018) found that

unbiased online reviews improve consumer purchasing decisions, which reduces returns;

biased reviews result in more returns. In the meantime, they observe that consumers are

more likely to write negative reviews when they return products than if they don't return them.

Another finding (Minnema et al.,2016) shows that there is an increased return rate for

products whose displayed average rating is higher than their true rating when the displayed

average rating is discontinuous.

It is especially important to consider the return policy when selling online since more than 70

percent of online consumers consider return policies when making purchase decisions (Su,

2009). A return policy that gives the consumer compensation for returns can boost consumer

8
 Dissertation

demand and subsequently increase sales, resulting in an increase in returns and a higher

cost of returns. Consumer responses to different return policies have been examined in prior

studies. Despite lenient return policies, Wood (2001) finds that purchases increase without

returns increasing. Using a direct sale model, Mukhopadhyay and Setoputro (2004) examine

how pricing and return policies have an impact on the purchase and return decisions of

customers. Return quantity is determined by a return policy, and product and service quality

are not considered. More specifically in return policy detail, longer deadlines, according to

Janakiraman and Ordónez (2012), result in consumers delaying or postponing their return

decisions. It is possible for online retailers to implement a restrictive return policy as a method

of reducing returns. Customers' return tendency is reduced by measures such as reshipment

fees (Petersen and Kumar, 2009). However, it could deter customers from ordering in the

first place since they anticipate costly reversals (Wood, 2001). Janakiraman et al. (2016)

report that retailers avoid restrictive policies because of this side effect. Nonetheless, a

lenient policy may result in an increase in returns, resulting in high costs for online retailers.

In terms of consuming experience, Bechwati and Siegal (2005) suggested consumers'

involvement and choosing alternative products influence the likelihood of product returns.

Unless a store provides high-quality products at competitive prices, it is more likely that

products will be returned. As Lee and Yi (2017) demonstrate, gift-with-purchase promotions

are associated with fewer returns of consumer products, hypothesizing that consumers are

less likely to return products that are sold with free gifts. Walsh et al. (2016) illuminate product

return rates are correlated with online retailers' reputations. A conclusion can be drawn from

two experiments that show that reputation reduces return rates. The finding also shows that

the strength of the relationship between reputation and product returns is influenced by

shopping frequency. In the last step of the purchase process, one of the last opportunities for

e-retailers to influence consumers' return behavior is through delivery package cues

(Garretson and Burton, 2005). In comparison to what remains in consumers' memories (i.e.,

the stimuli displayed at the time of purchase), delivery packages probably carry clearer and

fresher information. Finally, Zhou et al. (2018) explored the cognitive-emotional response
9
 Dissertation

process of consumers after opening the package. According to this research, pleasure plays

a crucial role in consumers' return choices.

On the other hand, Studies that focus on the characteristics of the consumers ordering the

products are far fewer in number. Research on online shopping product returns (Cheung

2003; Chang et al., 2005; Cheung et al.,2005; Zhou et al., 2007) has shown that demographic

characteristics, such as gender, age, education, and income, influence the likelihood of the

products being returned. The results indicate that these four variables are related to product

return. Makkonen et al. (2021) focus on four demographic characteristics (i.e., gender, age,

education, and income) as well as payment method preference. It is more likely to return a

product when paying with a credit card. Meanwhile, among women, it is a greater probability

of return frequency than among men, while the odds of return frequency decreased with age.

Yan and Cao (2017) confirmed another point, an argument that the payment method

influences the return of products as well. Due to the “buy-now-pay-later” mentality associated

with credit cards, the researchers explain this finding with the fact that impulsive consumption

behaviour is more likely to result, as well as a lower threshold of returns because there has

been no exchange of money yet. For different categories, Clothing, and shoes, for example,

are more likely to be returned by women and younger consumers (Deloitte, 2019), while

consumer electronics are more likely to be returned by men and older consumers (Deloitte,

2019).

Table 2-1：Summary of Researches of Product Return

Name Time Primary Content

Petersen and
2009 About 5% of the goods are returned due to quality problems.
Kumar

10
 Dissertation

internal drivers (i.e., variety seeking, impulsiveness, perceived

uniqueness, level of morality, and self-monitoring) and
Pei and external drivers (i.e., product compatibility, returning cost,
2018
Paswan perceived risk, the complexity of procedure, and social group
influence) are driven reasons to return products (Pei and
Paswan, 2018.).

Kirmani and
2000 low-quality products and services will result in product returns.
Rao

A reduction in price makes it less likely for products to be

Anderson et al. 2009
returned from a marketing-mix perspective.

When the price is expected to drop before purchase,

Bandi et al. 2018 customers tend to choose a payment method that is easy to
return.

the majority of returns even occur for other reasons, not

Fan et al. 2021
defective products.

dissatisfaction, product failure, or dishonest intent are not the

D'Innocenzio 2011 only reasons for consumer product returns. However, the
influence of other factors is not clearly stated.

Petersen and higher expectations should lead to higher purchase and return
2012
Kummar probabilities.

providing customers with high-tech tools to bring them closer

Khasawneh 2020
to the product would help decrease the number of returns.

Minnema et al. 2016 product reviews are related to product return.

Babic et al. 2016 review volume is expected to lower the return probability.

11
 Dissertation

unbiased online reviews improve consumer purchasing

Sahoo et al. 2018
decisions, which reduces returns.

products with higher average long-term ratings have a higher

Minnema et al. 2016
purchase and lower return probability.

Return compensation will lead to more product sales and

Su 2009
more returns.

A loose return policy will lead to an increase in product sales

Wood 2001
but will not result in an increase in products.

The return policy determines the quantity returned. And

Janakiraman Longer deadlines, according to Janakiraman and Ordónez
2019
and Ordónez (2019), result in consumers delaying or postponing their return
decisions.

Petersen and Customers' return tendency is reduced by measures such as

2009
Kummar reshipment fees.

Syrdal and
2016 retailers avoid restrictive policies because of this side effect.
Freling

Bechwati and
2005 Substitutes will affect the possibility of customer return.
Siegal

Lee and Yi 2017 Promotion products are less likely to be returned.

Product return rates are correlated with online retailers'

Albrecht et al. 2016
reputations.

delivery package cues have an impact on product return. In

Garretson and
2005 other words, good packaging will reduce the possibility of
Burton
customer return.

12
 Dissertation

Makkonen et Women are more likely to return goods than men. And as the
2021
al. age decreases, the probability of return increases gradually.

payment method influences the return of products as the

Yan and Cao 2017 probability of return is high under impulse consumption, and the
payment method convenient for return is more favorable.

2.1.2. Return Fraud

Return fraud has attracted researchers' attention in recent years. Fraud return is a critical

factor that results in product returns since the customer who fraud return plans to return it

when experiencing the value of product. It is imperative to study consumers' product return

behavior from an ethical perspective due to unethical behavior becoming an everyday matter

in the workplace, marketplace, society, and even the academic scene (Craciun, 2006). As

the first one to examine product returns from an ethical perspective, a study (Schmidt et al.

1999) use the term of “deshopping” and define it as the deliberate return of goods for reasons

other than actual faults in the product. Using the term 'retail borrowing', Piron and Young

(2000) study the effect of gender, income, and the economic status of the borrower on the

behavior of retail lenders in order to identify unethical behavior. In Johnson and Rhee (2008),

consumer traits, demographic characteristics, and social groups are studied in relation to

merchandise borrowing, and the results show clear agreement with those reported by Piron

and Young. The findings of Harris (2008) demonstrate a relationship between demographic

factors such as age, sex, and level of education and psychographic factors such as the prior

experience of fraudulent return and knowledge of returning rules and regulations. Resulting

from this paper, it was found that fraud is more likely to occur among younger, female

consumers with a lower levels of education, a conclusion that sexes, ages, and educational

levels have an impact on fraud return. Consequently, in order to understand consumers'

behavior relating to product returns, a comprehensive understanding is necessary, as well as

13
 Dissertation

empirical data supporting the findings. This research (Harris, 2008) also demonstrates that

there are eight psychographic factors linked to fraudulent returning tendency: past experience

of fraudulent returns, level of public self-consciousness, knowledge of returning rules and

regulations, consumer anomie, attitude toward complaining, social norms, consumption-

related, thrill-seeking needs, and consumers' perceived impact on returning. To improve this

condition, retailers are temptated to prevent return abuse by charging customers a return fee.

Table 2-2：Summary of Researches of Return Fraud

Name Time Primary Content

Use the term of “deshopping” and define it as the deliberate
Craciun 2006 return of goods for reasons other than actual faults in the
product.
Use the term “retail borrowing” to investigate unethical
Piron and
2000 behavior and study gender effects and income effects on
Young
retail borrowing
Johnson and
2008 Results show highly consistent with Piron and Young
Rhee
Fraud is more likely to occur among younger, female
consumers with lower levels of education. Furthermore, it
Harris 2008
has been found that fraud returning tendencies are
influenced by eight psychographic factors as well.

In the study, we focus on demographic characteristics of product return. Specifically, we will

explore whether the return rate will correspondingly change as the purchasing power and

age change. Based on the purchasing power in each county and the average age in each

town data in Germany, we can obtain some insights into the consumers’ habits and

characteristics in Europe so that some measurements can be provided to the retail industry.

2.1.3. Return Policy and Actions to Reduce Return

Many firms have taken some measures to fit their product return management strategy and

thus reduce the return rate. For example, the return window at Wal-Mart is 90 days, with

some exceptions, and the return window at Dell is 21 days with a 15% restocking fee. As

14
 Dissertation

another example, outdoor gear retailer REI has a "no-questions-asked" return policy, one of

the most accommodating in the industry (Grind, 2013). Customers who abuse Best Buy's

return policy are blacklisted and charged a 15% restocking fee (Boyle, 2006). Stock et al.

(2002) points out that several companies are managing supply chains to simplify consumer

returns in order to combat this problem. By outsourcing the return process to reverse logistics

specialists, reducing costs by simplifying the return process, and redistributing returned

merchandise, some profits can be salvaged. Even more, some companies implement more

restrictive return policies like penalties. However, penalties imposed on customers who return

a product can cause negative emotions such as regret, resulting in inaction (Bower and

Maxham, 2012). Customers already feel negative emotions when a product doesn't meet

their expectations. Further increases in this level may result from restrictive return handling.

As a result of this disadvantage, a restrictive policy would seem highly unfeasible. Therefore,

a way of rewarding instead of punishment is proposed (Gelbrich,2017). Keep rewards are

defined as promotion strategies that offer an incentive to customers for keeping the ordered

items (e.g., free shipping on their next purchase) while allowing lenient return policies. A high

return rate of lenient policy is improved by adding a promotional component that may reduce

the return tendency.

Except for adapting the return policy to reduce product return, some techniques to describe

information about the product have been applied. Bechwati and Siegal (2005) mentioned that

the information provided by retailers affects the ability of customers to adequately evaluate

products before purchasing. Retailers have taken some technique to describe clearly about

the product. For instance, to help customers make better decisions and to avoid return costs,

retailers have invested in technologies like zoom features. Furthermore, Online Customer

Reviews (OCRS) contribute to forming customer expectations before purchase (Chen and

Xie, 2008), and may affect return rates. Furthermore, based on online review information,

some retailers provide more information about products to customer. A study from De et al.

(2013) examined the effectiveness of different online product inspection technologies in

reducing product returns. On the basis of detailed information regarding how customers use
15
 Dissertation

technology before purchase, they demonstrate that using an online zoom tool leads to fewer

product returns and that using alternative images of a product leads to higher returns.

Table 2-3：Summary of Researches of Return Policy

Name Time Primary Content

Wal-Mart is 90 days and the return window at Dell is 21 days
Grind 2013 with a 15% restocking fee. And outdoor gear retailer REI has
a "no-questions-asked" return policy.

Customers who abuse Best Buy's return policy are blacklisted

Boyle 2016
and charged a 15% restocking fee.

Stock, Speh, Several companies are managing supply chains to simplify

2002
and Shear consumer returns in order to combat this problem.

Penalties imposed on customers who return products may

Bower and
2012 lead to negative emotions such as regret, resulting in more
Maxham
returns.

Rewarding(e.g., free shipping on their next purchase) can

Gelbrich 2017
better reduce product returns than punishment.

Bechwati The information provided by retailers is related to purchasing

2005
and Siegal and product return.

Online Customer Reviews (OCRS) can provide more

Chen and
2008 information about products, thus resulting in reducing in
Xie
product returns.

Using an online zoom tool is associated with fewer product

De et al. 2013
returns.

2.2. Quantitative Analytics Model of Product Returns

Forecasting models are typically classified as qualitative forecasting models or quantitative

Forecasting models (Zhou et al.,2016). The qualitative forecasting models are generally

subjective and are mostly based on the opinions and judgments of experts. There is general

use for such types of methods when there is little to no historical data available on which to

base a forecast, or when there is very little data available. On the other hand, in quantitative
16
 Dissertation

forecasting, the data available is used to make predictions about the future and a statistical

association is presented between past and present values, based on the patterns in the data.

In short, A subjective judgment is used for the former if historical data are unavailable, while

a more practical approach is used for the latter. Quantitative forecasting methods include

time series methods, such as moving averages and linear regressions, as well as measures

of causality and econometric models, such as regression analysis or autoregressive moving

averages with exogenous input and neural network (Salehzadeh et al.,2020).

According to Kumar and Yamaoka (2007), their research shows that dynamic regression

models are a good choice when data are in a wide range variety. With the combination of

DEA/linear regression and moving averages, Potdar and Rogers (2012) developed a model

to forecast consumer electronics product returns using reason-codes. To forecast product

returns, they are taking consumer behavior and turning it into meaningful data. In contrast,

Alexandra et al. (2016) use Holt’s and ARIMA methods to forecast the future returns of 36

products, which indicates a higher accuracy. Ma and Kim (2016) applied autoregressive

statistical models to predict return quantity and time. There is also a recommendation here

for the use of Gaussian distributions when counts are large (e.g., the data for reusable bottles

presented in this study), as these distributions provide a good fit to the data. In this way, the

total amount of returns within a certain time can be estimated, but individual return actions

cannot be predicted.

Secondly, for the Machine Learning model, Using Mahalanobis feature extraction, Urbanke

et al. (2015) predict return rates based on product features (e.g., brand, color, size), customer

attributes (e.g., past return rates), and basket information, including platform, payment

method, and the total number of items, an algorithm is applicable to product return business

since most customer databases contain categorical data. However, Typically, the required

information is available only after customers have completed their online shopping journey,

so this method is not designed for customer-product level prediction. The information related

to customers when they search on the online stores, such as what they like, consumption

17
 Dissertation

level, and what products are in their purchase cart, is significant to analyze what personalities

are inclined to return the product. Moreover, the historical purchase and return records for

products in the past can be highly valuable sources of information but can be challenging to

integrate in a principled way in order to predict future returns. The work by Zhu et al. (2018)

focuses on modeling customer online shopping behaviors and predicting their return actions

through the integration of the rich information that comes from the purchase and return history

of products (e.g., return history, purchase-no-return behavior, and similarity between

customers and products) by using HyGraph, which is a local random walk algorithm with a

fixed running time based on the size of the output clusters, rather than the entire graph.

Zhou and Xie (2016) demonstrate that their study is the first model that has been developed

to forecast the quantity, time, and probability of product return and remanufacturing by using

the GERT stochastic network analysis technique. A data-driven model was developed by Cui

et al. (2020) using detailed operational information on each product and information about

the retailer to predict return volume by the retailer, product type and period. LASSO yields a

predictive model achieving the best prediction accuracy for future return volume. Stacking

and Vote algorithms from EML algorithms are used by Tüylü et al. (2022) to estimate product

return rates, indicating that the EML algorithms can be used to predict product return rates.

In another research (Dzyabura et al.,2018), a Convolutional Neural Network and color

histograms are employed to extract information from images using machine learning, then

using this information in a gradient boosted regression tree prediction model. By incorporating

visual characteristics into the model, the accuracy of predicting return rates is increased by

an impressive 37% compared to models that do not include images. From Ketzenberg et al.

(2020), several well-performing Machine Learning methods, such as Logistic Regression,

Support Vector Machines, Random Forests, and Neural Networks are applied to measure

return abuse, predictive models could achieve 99.6% accuracy.

According to previous research, Machine Learning methods such as Logistic Regression,

Random Forest, Neural Network, Decision Tree, EML method and NLP have been applied
18
 Dissertation

to data analysis in this field. In this study, other Machine Learning methods such as Bayesian,

and other optimization methods are mainly used to predict.

Table 2-4: Summary of Data Analytics Methods of Product Return

Name Time Primary Content Analytics Methods

Kumar and The Regression model is

2007 Dynamic regression models
Yamaoka suited for a number of data.

The incoming returns are

split into different categories
using reason codes. The
computation part of this
model uses a combination of
two approaches, namely the Data envelopment analysis
Potdar and extreme point approach and (DEA) as a first step combined
2012
Rogers the central tendency with linear regression while the
approach. The two central tendency approach
approaches are employed
separately for different types
of reason codes, and then
the combined results are
analyzed.

To forecast the future returns

Alexandra
2016 of 36 products, which Holt’s and ARIMA
et al.
indicates a higher accuracy.

Ma and To predict return quantity and

2016 Autoregressive statistical models
Kim time.

Predict return rates based on The approach proposed in this

product features (e.g., brand, study is to maximize the
Urbanke et
2015 color, size), customer Mahalanobis distance of the
al.
attributes (e.g., past return observation from the expected
rates), and basket value under the null hypothesis,
19
 Dissertation

information, including thus minimizing the probability

platform, payment method, that the null hypothesis is true
and the total number of
items.

Incorporating the rich

information in the purchase
The Lograph Algorithm is based
and return history of a
on random walks and its running
product into a model for
Zhu et al. 2018 time is determined by the size of
predicting customer online
the output cluster rather than the
shopping behavior and
entire graph
predicting their return
actions.
Developed a data-driven
model for predicting return
Cui et al. 2020 volume at the retailer, Data-driven mode
product type and period
levels.

To estimate product return

Stacking and Vote algorithms
Tüylü et al. 2022 rates and this model shows a
from EML algorithms
good ability to predict.

The accuracy of predicting

return rates is increased by A convolutional neural network
Dzyabura
2018 an impressive 37% and color histograms are
et al.
compared to models that do employed to extract information
not include images.
To predict return abuse and Logistic regression, support
Ketzenberg
2020 predictive models could vector machines, random
et al.
achieve 99.6% accuracy. forests, and neural networks

2.3. Research Innovation

We have found that limited by the product return data, there are few relevant studies on the

demographic characteristics of product returns. And the existing research focuses on North

America and Asia. However, due to the influence of economic level, regional culture and

education level, the consumption habits in different places are different. Compared with the

conservative consumption habits in Asia and the free consumption concept in North America,

20
 Dissertation

Europe has great uniqueness in consumption behavior. At the same time, Europe's economic

level is relatively developed, and its culture is both traditional and open. Therefore, the impact

of demographic characteristics on product returns may be different from that of other regions.

In this study, we focus on the impact of demographic characteristics, including purchasing

power and average age, on product returns in Europe. In addition, in terms of data analysis

methods, statistical analysis methods, logistic regression and other basic research methods

are used in the publications of product returns about demographic characteristics. Moreover,

we not only use Logistic Regression in Machine Learning, a model that is suitable for binary

classification data but also use Naive Bayesian classification and Neural Network classifier.

These three analysis methods are more rational for the data applied in this study. The

independent variable is continuous variables, and the dependent variable is binary data: 1

means the product has been returned, and 0 means the product has not been returned.

According to the data analysis results, we can know the conclusion, that is, whether the return

of products will change with the change in consumers' age and purchasing power.

Furthermore, one of the goals of this study is to offer some advice to the retail industry in

Europe according to our quantitative analysis and qualitative analysis. We hope that the

conclusion we get is valuable and could be used for practical application.

3. Methodology

In this section, we follow the process of data mining, the data mining has been broken down

into six steps: business understanding, data understanding, data preparation, modeling,

evaluation, and deployment. Specifically, subsection 1 is understanding a problem and the

final goal. Subsection 2, data collection, introduces the data source and data type. Subsection

3 discusses data cleaning and data preprocessing: including processing missing values and

outliers, data standardization, data discretization and features selection. Subsection 4

discusses model selection, including Logistic Regression, Naive Bayesian and Neural

Networks, and model evaluation, including ROC and confusion matrix. In the last part, due to

the dataset and model problems found in the previous part, the data and model are optimized.

21
 Dissertation

We focus on the analysis of the third part of modeling and evaluation and the fourth part of

model optimization, as this part is the focus of this paper. However, the performance of the

model before optimization is limited, so the optimization process is significant.

3.1 Understanding A Problem and Final Goal

The return of products will be affected by many factors, such as the factors of the product

itself, the factors of retailers, and even the problems of breakage during sales and

transportation. At the same time, the return of products is also affected by the characteristics

of consumers, such as income, gender, education, etc. In this study, we mainly explore the

impact of demographic characteristics on product returns. Specifically, when the age or the

purchasing power changes, how the frequency of consumer returns will change. Explore the

relationship between these two features and product returns.

3.2 Data Collection

The original data comes from the sales data of a second-hand product sales website in

Europe. The retailer mainly sells electronic products, including computers, mobile phones,

televisions, and other products. The data content includes:

Table 3-1: Data Explanation

Name Explanation

zip code zip code of the consumer address

consumer ID values corresponding to each customer are unique

product ID values corresponding to each product are unique

invoice ID corresponding to each product sold are unique values

payment methods include PayPal, eBay payment, and Amazon payment

Sales channels online and offline

Reasons for return including product quality, price and change of mind.

22
 Dissertation

Package type DHL and Amazon Prime

from January 2020 to June 2022, which are the sales
the date of
data after covid19 outbreak, so the impact of covid19
purchasing
can be removed.
Although the data includes the return reason, there is no reason related to demographic

characteristics in the return reason, so it cannot be used directly. Therefore, we matched the

purchasing power of German counties and the average age of towns by zip code.

The original dataset needs to be matched with the purchasing power data and the average

age data. We can conduct one-to-one accurate matching to counties through postal code,

but some counties have high similarities in names, so we can't judge the specific location, a

total of 854 (less than 1%). Similarly, there is a similar situation when matching the average

age, and there are 7531 pieces of data (less than 5%) of the average age that cannot be

identified.

3.3 Data preparation and preprocessing

3.3.1 Methods Introduction

[Link] Zscore Normalization

In data analysis, many independent variables that need to be analyzed are at the same level,

so that different trend ranges of different units can be compared. Otherwise, it will cause

difficulties in the analysis work and even affect the accuracy of later modeling.

This method normalizes the data based on the mean and standard deviation of the original

data. Then the standard deviation is processed according to the following formula:

𝑍!" = (𝑋!" − 𝑋! ) /𝑆! ,

where: Zij is the value of the variable after standardization; Xij is the actual variable value, 𝑆!

is the standard deviation (Saranya and Manikandan, 2013).

23
 Dissertation

Then, reverse the sign before the indicator. The standardized variable value fluctuates

around 0. A value greater than 0 indicates that it is above the average level, and a value less

than 0 indicates that it is below the average level.

[Link] Weight of Evidence (WoE)

The data is discretized by WoE. Weights of evidence (WoE) measures the relative risk

associated with an attribute category (Fan et al.,2011). The higher the weight of evidence

(in favour of being good), the lower the risk for that category.

WoE=ln(pro_goodcategory/pro_badcategory),

Where pro_goodcategory=number of goodcategory/number of goodstotal

pro_badcategory= number of badscategory/number of badstotal

if pro_goodcategory> pro_badcategory, WoE >0

if pro_goodcatego< pro_badcategory, WoE >0

[Link] Information Value

Information Value (IV) is a step to select feature, that is, among all variables, select the

independent variable that is important to predict the dependent variable. IV is a measure of

predictive power used to assess the appropriateness of the classification and select

predictive variables (Wang et al., 2020).

IV = Σ((pro_good#$%&'()* _pro_bad#$%&'()* ) ∗ WoE#$%&'()* )

Rule of thumb，

IV< 0.02: unpredictive; 0.02 – 0.1: weak; 0.1-0.3: medium; IV > 0.3: strong.

3.3.2 Implementation procedure

[Link] Data description

After data cleaning, There are 162663 observations and four variables, including Customer

ID, Purchasing Power, Average Age, and Return Value. Variables are shown below，
24
 Dissertation

Table 3-2: Features

Name Explanation
Corresponding to each customer are unique values,
Customer ID
object variables, dependent variables
The purchasing power of each county in Germany,
Purchasing Power
numerical variable, dependent variables
The average age of each town in Germany,
Average Age
numerical variable, dependent variables
Whether a customer return product, 1 means return
Return Value and 0 means do not return, binary variable,
independent variables

Table 3-3: Data Description

Purchasing Power Avg Age Return Value
count 161769 155332 162663
mean 24.881 44.283 0.098664
std 2.461 1.973 0.298211
min 19.901 27.900 0
25% 23.503 43.500 0
50% 23.694 43.600 0
75% 26.252 45.200 0
max 37.705 56.200 1
According to the data description, the purchasing power of each county in Germany is from

19.9k to 37.7k and the mean purchasing power of each county is 24.9k. The average age of

each town is from 27.9 to 56.2 years old, but mostly around 43-45 years old since the 25

percentile to 75 percentile of average age lies in this area.

Figure 3-1: Histogram Plot of Variables

25
 Dissertation

According to the histogram, purchasing power mostly lies on 24k, and the data shows the

normal distribution. Meanwhile, the average age is mostly around 44, which suggests normal

distribution as well.

Figure 3-2: Pair Plot of Variables

The scatter chart can let us know whether there are patterns in the data; It can be seen from

the figure that the scatter plot data of purchasing power and average age are unevenly

distributed, which may mean that the variables are not related. Therefore, we can put them

into the Machine Learning algorithm since many Machine Learning model is based on the

assumption of uncorrelation between features.

The independent variables are regarded as inputs and the dependent variables are regarded

as the outputs that depend on the inputs. By using Supervised Machine Learning algorithms,

this paper will analyze a number of observations and try to mathematically express the

dependence between inputs, purchasing power, and average age, and outputs, product

return.

[Link] Data preprocessing

26
 Dissertation

According to the table of data descriptions, it indicates that there are missing values in these

data since the count number is less than 162663. There are 894 missing values in purchasing

power and 7331missing value in average age. Meanwhile, there are only two independent

variables, and the influence of each variable on the dependent variable is 50%, so the

absence of any one variable will have a great influence on the dependent variable; In addition,

the missing value of purchasing power in each county accounts for less than 1%, and the

missing value of average age in each town is less than 5%. After deleting these missing

values, the objectivity of the data and the correctness of the results will not lead to wrong

analysis conclusions. The output results of the Machine Learning model can still reflect the

real situation of the data. Therefore, the observation corresponding to the missing value can

be removed when processing the missing value. Finally, the dataset includes 154840

observations after missing values are handled.

Table 3-4: Data Description After Dropping Missing Value

Purchasing Power Avg Age Return Value

count 154840 154840 154840
mean 24.966 44.273 0.097
std 2.458 1.965 0.296
min 19.901 27.9 0
25% 23.503 43.5 0
50% 23.934 43.6 0
75% 26.326 45.1 0
max 37.705 56.2 1
According to the plot shown below, it is suggested that there are no obvious outliers as the

figure shows an obvious bulge When the purchasing power is about 33, and there is a small

bulge around 35, which means that the purchasing power of some counties is 33K and 35K,

and there is no "long tail" after that, meaning that no obvious outliers.

27
 Dissertation

Figure 3-3: Distribution of Purchasing Power

Similarly, the figure shown below does not have a “long tail”, which indicates no obvious

outlier so that the variable of average age does not need to be handled.

Figure 3-4: Distribution of Average Age

[Link] Data Normalization and Discretization

In this study, the average age of each town and the purchasing power of each county are not

within the same order of magnitude and cannot be directly compared. After normalization,

the data are mapped in the same interval, so that the variables can be compared directly.

Before WoE, the data needs to be coarse binning first. Data binning splits up the value range

of continuous variables into separate intervals or bands, such as binning age data to 18-

25,26-32,33-40, and so on. Meanwhile, binning data merge values of discrete variables into

28
 Dissertation

a smaller number of categories. However, in this study, we just need to adjust

continuous(numerical) variables since the average age of each town and purchasing power

of each county are numerical variables. After manual adjustment, the binning of the

purchasing power of each county and the average age of each town are adjusted to [0,1],

and the trend of variables is monotonically increasing.

Figure 3-5: Data Binning of Average Age

Figure 3-6: Data Binning of Purchasing Power

The WoE can indicate the prediction ability of the box for the dependent variable. If the WoE

is positive and the value is larger, the probability of bad users of the box is higher; and if the

negative value of the WoE is larger, the probability of good users of the box is higher.

Therefore, the WoE has good business interpretability.

29
 Dissertation

The general standard is that when the IV value is greater than 0.3, the variable has a strong

prediction ability; When 0.1 < IV < 0.3, the prediction ability of this variable is general. When

IV < 0.1, the predictive ability of this variable is weak.

The credit card dataset is split into two halves one training set and the other testing set. In

this study, we chose the ratio of 7:3 for the training dataset over the testing dataset.

3.4 Modeling and Testing

In the dataset of product returns, there are two values for the classification of transactions

which means that it is a binary classification problem where transactions are classified either

as return (1) or non-return (0). After preprocessing the data, the classifiers are trained using

the training data to evaluate the methods. In this study of classification techniques, We study

several typically competitive, well-performing machine learning methods that include Logistic

Regression, Naïve Bayes and Neural Networks

3.4.1 methods introduction

[Link] Logistic regression

Logistic regression is to fit the data by linear regression, and then use the logic function to

predict the classification results. It is especially suitable for classification, when the dependent

variable is dichotomized (0 / 1, true / false, yes / no), logistic regression will show a good

fitting result. In this study, the dependent variable return value includes two cases: 1 and 0.

1 means that the product has been returned, and 0 means that the product has not been

returned. Hilbe (2009) introduces Logistic regression models in detail.

Let y = 1 indicate the thing happens, y = 0 indicate that the thing does not happen, the

probability of happing is P (y = 1) = P1, the probability of not happening is P(y = 0) = P0 = 1-

P1

Then odds= P1/1- P1= P1/ P0, Odds refers to the probability of happening compared with

nonhappening

Logit=log(odds)=ln (P1/ P0)= β0+β1X1

30
 Dissertation

when the independent variable change by 1 unit, the odds change eβ1 times.

[Link] Naïve Bayes Model

The Bayesian classification algorithm is the general name of a large class of classification

algorithms and takes the probability that the sample may belong to a certain class as the

classification basis. Naïve Bayes’ Theorem is a mathematical formula used for calculating

conditional probabilities. A conditional probability is a probability that an event will occur after

another event has taken place. See more detail of the Naïve Bayes model in the book of

Johnson(2022).

The formula is

P(Y|X) = P(X|Y) * P(Y) / P(X)

Where

P(Y|X) is called posterior probability, meaning how often A happens given that B happens

P(X|Y) represents how often X happens given that Y happens

P(Y) represents how likely Y is on its own

P(X) represents how likely X is on its own

If P (Y=1| X)>P (Y=0| X), class as happening; otherwise, class as not happening.

[Link] Neural Network

There is no explicit formula or algebraic expression for neural network regression. The

process of training the Neural Network is an optimization problem, that is, to find out which

parameters make the model work best. Chollet (2021) indicates that the observed input data

is x, the output is y, and the predicted value of the model is f (x). When initializing the Neural

Network model, the parameters of f (x) are random values. For the observation value x, the

random parameters are randomly generated.

In this study, there are two classes, the Softmax function is,

f(x)=1/exp(β1X1+β0), which is common as an output function of probability estimations,

31
 Dissertation

where β1 is the coefficient of x1, β0 is constant

This process is called forward propagation, and it is a process that the input (observation

value) is calculated layer by layer (including linear calculation and nonlinear activation) to

obtain the predicted value of the output value. To evaluate the accuracy of the predicted

value, a loss function is required. A common loss function is the sum of squares of residuals

loss = 𝑙𝑜𝑠𝑠 = ∑ 𝑖 ( 𝑦?! − 𝑦! ) , .

The optimal parameters can be obtained by reducing the loss of the loss function. This

process is called back propagation. In essence, the process of back propagation is to

calculate the gradient value of each parameter for the loss function, and then adjust it

according to the gradient value. Theoretically, convergence can be achieved quickly.

[Link] Confusion Matrix And Receiver Operating Characteristic Curve (ROC)

In prediction analysis, the confusion matrix is a two row and two column table composed of

false positions, false negatives, true positions and true negatives. True positives are the

cases that are predicted as positive and in reality, they are positive as well. True negative is

the cases that are considered as negative in advance. False positives are cases that are

expected to be positive but turn out to be. False negative is one that appears to be negative

but is actually positive (Tze-Wey, 2003).

the value of ROC means: "the ROC is equal to the probability that a randomly chosen positive

example is ranked higher than a randomly chosen negative example(Hand and Till ,2001)."

The value of ROC is on [0.5,1]. In the case of ROC > 0.5, the closer the value of ROC is to

1, the better the diagnostic effect.

3.4.2 Implementation procedure

[Link] Logistic regression

Let y = 1 indicate that the product is returned, y = 0 indicate that the product is not returned,

the probability of product return is p (y = 1) = P1, the probability of not returning is P(y = 0) =

P0 = 1- P1, X1 = purchasing power, X2 = average age

32
 Dissertation

Then odds= P1/1- P1= P1/ P0 Odds refers to the probability of product return compared with

nonreturn Logit=log(odds)=ln(P1/ P0)= β0+β1X1+β2X2

Since both X1 and X2 are continuous variables, when the independent variable changes by

1 unit, the odds change eβ1 和 eβ2times.

[Link] Naïve Bayes model

According to the fundamental Naïve Bayes assumption, we assume that

No two features are dependent on each other, which means there is no correlation between

each variable ‘purchasing power’, ‘average age’. Meanwhile, according to table 3-2, it is

evidently indicate that is true.

Each feature is equally influential, suggesting that each variable represents the same weight.

In this study,

P(Y|X1X2) = P(X1X2|Y)*P(Y)/P(X1X2)=P(X1|Y)P(X2|Y) *P(Y)/P(X1X2)

Where
P(Y|X1X2) is a probability that product has been returned, given average purchasing power

and average age of a place.

P(X1|Y) is a probability of a customer who has specific purchasing power return products

P(X2|Y) is a probability of a customer who is a specific age power return product

P(Y) is probability that product has been returned

P(X1X2) is a probability of a customer who has a specific purchasing power and in a specific

average age
If P (Y=1| X1X2)>P (Y=0| X1X2), class as product return; otherwise, class as not product return

In this study, because the variable characteristics are binomial distribution, we chose

Bernoulli Bayesian model to fit the data.

[Link] Neural Network

When initializing the Neural Network model, the parameters of f (x) are random values.

In this study, there are two classes, the softmax function is f(x)=1/exp(β1X1+β2X2+β0), which

is common as output function of probability estimations, where

β1 is the coefficient of X1, X1 is independent variable of purchasing power in each county

33
 Dissertation

β2 is the coefficient of X2, X2 is independent variable of average age in each town

β0 is constant

loss function is the sum of squares of residuals 𝑙𝑜𝑠𝑠 = ∑ 𝑖 ( 𝑦?! − 𝑦! ) , .where 𝑦?! is the true

value and 𝑦! is predicted value (Janocha and Czarnecki, 2017).

3.5 Model optimization

3.5.1 Methods Introduction

[Link] Polynomial expansion

Polynomial expansion is the process of extending features to a polynomial space, which is

represented by the n-degree combination of the original dimensions (James et al., 2015).

Polynomial expansion improves the input variable to idempotent transformation, which helps

to better reveal the important relationship between the input variable and the target variable.

Sometimes these features can improve modeling performance, although at the cost of adding

thousands or even millions of additional input variables. Polynomial features create new input

features based on existing features. For example, if the dataset has an input feature x, the

polynomial feature will be to add a new feature (column), where the value is calculated by

squaring the value in X, for example, X2. Ghaith and Li (2020) pointed out that the prediction

performance of the model was good after the variables were processed by Polynomial

Expansion.

[Link] Decision Tree Binning

The decision-making process of a decision tree is essentially a series of if / then statements,

which make decisions through machine learning. decision trees have an important property:

they are mutually exclusive and complete. This means that for each sample, there is and only

one path from the root node to a leaf node. At the same time, because a series of conditional

judgments are constantly made on the features, the decision tree can also be understood as

the solution of the conditional probability of 𝑃 (Y𝑖| 𝑋𝑖). To construct the decision tree, the

algorithm iterates all possible questions, finds the one with the largest amount of information

34
 Dissertation

for the target variable, divides the data set into two parts, and repeats this process until the

end. Therefore, In the first step, the decision tree needs to be divided into nodes. In the

second Step: the condition for a node to stop dividing into leaf nodes is that all samples in

the node belong to the same category. That is, nodes are "pure". Therefore, when selecting

features for partitioning, the purer the nodes, the better the features. We use entropy to

measure the "purity" of a group of samples. We recommend Rabuzin et al. (2014)

International Journal of Data Warehousing and Mining to learn more about the discretization

method.

For the discrete random variable X, its distribution is:

𝑃(𝑋=𝑋𝑖) = 𝑝𝑖,𝑖=1,2,3...𝑘

Information entropy is
- -

H(X) = E pilnpi , E pi = 1
./0 ./0

For binary classification,

p1+p2=1

𝐻=−( p1𝑙𝑛 p1+ p2𝑙𝑛p2)

Then，
∂H dpj dpj 1
= −lnpi − 1 − lnpj − = ln N − 1O
∂pi dpi dpi pi
0
When ,≤𝑝𝑖≤1，𝐻′(𝑝𝑖)≤0，𝐻 decreasing gradually.

0
When 0≤𝑝𝑖≤,, 𝐻′(𝑝𝑖)≤0，𝐻 increasing gradually.

The results of d When 𝑝1=1/2. Information entropy is maximum. For binary classification,

the smaller the entropy is, the better the result is.

3.5.2 implementation procedure of optimization

[Link] feature transformation

According to the results of model analysis, among the three models, the ROC score is

between 0.6 and 0.65, and that there is no big difference between different models. Therefore,

it can be explained that there are some problems in the process of machine learning due to
35
 Dissertation

the problems in the data itself. In reviewing the original data and data analysis, we found the

following problems: 1) in the original data, only less than 10% of the return data, that is, the

return value of 1 is less than 10%, and the return value of 0 is more than 90%. This

phenomenon led to data imbalance. 2) There are too few features, which is not conducive to

machine learning. In the data used in this study, there are only two characteristics, purchasing

power and average age, and many models with excellent performance, such as neural

networks, are suitable for data with a large amount of observations and more data features.

In the process of optimizing the model, we first reduce the amount of data and increase the

number of features to reduce overfitting. We only selected the data of the first half of 2022,

with a total of 33543 observations.

Thus, we suppose that the independent variable, X1=purchasing power, X2=average age,

and the additional variable by Polynomial expansion X3 = X1 * X2, X4 =X1 * X1, X5 = X2 * X2,

X6 = X1 / X2

36
 Dissertation

Figure 3-7：Scatter of Variables

There is no obvious relationship between purchasing power and average age but purchasing

power and average age have a certain relationship with the newly added variables

respectively, because the newly added variables are set based on purchasing power and

average age.

[Link] Data Discretizing

When discretizing the data, we use decision tree binning. This method has two advantages:

1) pruning prevents overfitting. Pruning includes pre pruning and post pruning. The former

controls the depth of the tree or the number of nodes by setting thresholds for continuous

variables, and operates before the nodes are divided, thereby preventing overfitting. The

latter is to examine non-leaf nodes from the bottom up. If replacing this internal node with a

leaf node can improve the generalization ability of the decision tree, replace it. 2) In order to

prevent the training decision tree from being too biased towards some categories due to too

many samples in some categories of the training set. The algorithm calculates the weight by

itself, and the category with a small sample size will have a higher weight.

Decision tree binning results are show below,

Figure 3-8: Entropy of X1 Figure 3-9:Entropy of X2

37
 Dissertation

Figure 3-10: Entropy of X3 Figure 3-11: Entropy of X4

Figure 3-12: Entropy of X5 Figure 3-13: Entropy of X6

In the optimization, we added dummy variables. Dummy variables are 0 or 1 to indicate

whether a category appears. Because after the data is discretized, using only discrete

numbers to represent categories may affect the accuracy of subsequent regression or

classification. After deleting the missing values after discretization of the decision tree and

adding dummy variables, there are 20185 observed values

[Link] Logistic Regression

Let y = 1 indicate that the product is returned, y = 0 indicate that the product is not returned,

the probability of product return is p (y = 1) = P1, the probability of not returning is P(y = 0) =

P0 = 1-P1, X1 = purchasing power, X2 = average age， X3 = X1 * X2, X4 =X1 * X1, X5 = X2

* X2, and X6 = X1 / X2

Then odds= P1/1- P1= P1/ P0 Odds refers to the probability of product return compared with

nonreturn
38
 Dissertation

Since the Decision Tree Binning requires dummy variables. Therefore, in the process of

discretization of variables, each variable is divided into different "boxes", and the number of

variables is adjusted to 30 variables. Therefore, the six variables we assume cannot get the

coefficients, but only the coefficients of the variables after discretization. Therefore, it is

impossible to directly judge the specific impact of the two characteristics, purchasing power

and average age, on product returns, but this model can be used to judge whether returns

will occur at the customer level.

[Link] Naïve Bayes Model

The independent variable X1，X2，X3，X4 and X5 dependent variable Y are consistent with those

mentioned above，then
P(Y|X1X2X3X4X5) = P(X1X2X3X4X5|Y) * P(Y)/P(X1X2X3X4X5)

=P(X1|Y) P(X2|Y)P(X3|Y)P(X4|Y) P(X5|Y) *P(Y)/P(X1X2 X3X4 X5)

Where
P(Y|X1X2 X3X4X5 X6) is a probability that product has been returned, given average

purchasing power and average age of a place and satisfy the other four additional variables.

P(X1|Y) is a probability of a customer who has specific purchasing power return products

P(X2|Y) is a probability of a customer who is a specific age power return product

P(X3|Y), P(X4|Y), P(X5|Y), P(X6|Y) is a probability of a customer who satisfies the other four

additional variables respectively return products.

P(Y) is probability that product has been returned

P(X1X2 X3X4 X5 X6) is a probability of a customer who has a specific purchasing power

and in a specific average age and satisfies the other four additional variables.
If P(Y=1| X1X2X3X4 X5 X6)>P(Y=0| X1X2X3X4 X5 X6), class as product return; otherwise class

as not product return.

[Link] Neural network

As this is a two-class problem, the last layer is adjusted to sigmoid as the activation function,

which is more suitable than softmax function. Sigmoid function, that is f(x)=1/(1+e-x). Neural

networks are complex: the functions of each layer are different, and the results are obtained

39
 Dissertation

after many iterations. Therefore, we cannot determine the relationship between independent

variables and dependent variables through specific coefficients. Instead, a model can be

established to judge whether customers will return goods from the perspective of customers.

4. Results
4.1 Results of Data Preprocessing
Based on the results, the table indicates that the minimum boxes of the average age of each

town and the minimum boxes of the purchasing power of each county are negative, and the

minimum boxes of the average age of each town are greater than the minimum boxes of the

purchasing power of each county, indicating that there are more good users of the minimum

boxes of the purchasing power of each county. Similarly, the maximum boxes of the average

age of each town and the maximum boxes of the purchasing power of each county are both

negative and positive, and the maximum boxes of the average age of each town are smaller

than the maximum boxes of the purchasing power of each county, indicating that there are

more bad users in the maximum boxes of the purchasing power of each county. Similarly, at

75percentile, there are more bad users in purchasing power of each county.

Table 4-1: Test Dataset Description After WoE

Index Return Value Avg Age_WoE Purchasing Power_WoE

count 46452 46452 46452

mean 0.097 -0.020 -0.050

std 0.296 0.223 0.349

min 0 -0.189 -0.323

25% 0 -0.189 -0.323

50% 0 -0.189 -0.323

75% 0 0.230 0.390

max 1 0.297 0.404

Table 4-2: Train Dataset Description After Woe

40
 Dissertation

Index Return Value Avg Age_WoE Purchasing Power_WoE

count 108388 108388 108388

mean 0.097 -0.020 -0.050

std 0.296 0.223 0.349

min 0 -0.189 -0.323

25% 0 -0.189 -0.323

50% 0 -0.189 -0.323

75% 0 0.230 0.390

max 1 0.297 0.404

According to the result, IV of purchasing power is around 0.126, and IV of average age is

0.052, which suggests that purchasing power shows the medium predictable capability and

average age is weak predictable. In other words, the influence of average age on product

return is not obvious.

Table 4-3: information value of variables

variable info_value

purchasing power_WoE 0.126382263

avg age_ WoE 0.051629219

However, according to business insights, it is generally believed that age is related to returns.

And as there are few data features, removing the features may have an impact on the model

effect, so we do not delete variables with IV values lower than 0.1.

4.2 Modeling Results

4.2.1 Logistic Regression

Logit=log(odds)=ln (P1/ P0)= 0.50989886+1.777191X1+1.429618X2

the coefficient of purchasing power is 1.777, and the coefficient of average age is 1.430.

When X1 variable (purchasing power) is changed by one unit and other variables remain

unchanged, the odds increase e1.777191 times. Similarly, when the X2 variable (average age)

is changed by one unit and other variables remain unchanged, the odds increase e1.429618
41
 Dissertation

time. Furthermore, the average age has a greater impact on the prediction of whether the

product is returned.

Table 4-4: Parameter of Logistic Regression

Column Coefficient

purchasing power_ WoE 1.777191

avg age_ WoE 1.429618

intercept 0.50989886

In order to understand the fitting effect of the model, we need to evaluate the logistic

regression. Confusion matrix shows that TP is 0.64, FP is 0.36, FN is 0.46 and TN is 0.54.

And classification accuracy is 0.59, classification error is 0.41, sensitivity is 0.58 and

specificity is 0.60, meaning that this model has a not very strong performance and does not

fit well with dataset.

Figure 4-1: Confusion Matrix of Logistic Regression

It is generally considered that ROC exceeding 0.75 is acceptable. However, the ROC is 0.624,

the result shows that this model does not fit very well with the data set. So, we need to

optimize the model in the next subsection.

42
 Dissertation

Figure 4-2：ROC of Logistic Regression

4.2.2 Results of naïve Bayes

Similarly, we evaluate Naïve Bayes models. ROC describes the relationship between TPR

and FPR. The larger the area enclosed by ROC curve and coordinate graph boundary, the

better the model; The ROC is 0.622, a result showing that the method has a low prediction

ability, and this model does not fit well with the dataset.

Figure 4-3：ROC of Naïve Bayes Classifier

4.2.3 Results of Neural Network Classifier

The ROC is 0.628, similarly with ROC of naïve bayes, a result shows that the method has

not much good performance, and this model does not fit well with the dataset. Meanwhile,

this model does not show better than logistic regression and naïve bayes.

43
 Dissertation

Figure 4-4：ROC of Neural Network Classifier

Although there are empirical rules and heuristics to determine the network structure, there

are no known optimal decision rules (Brownlee, 2018). Therefore, we will improve the data in

the preprocessing of the original data, rather than optimizing in the neural network model

itself.

4.3 Model Optimization

4.3.1 Optimized Logistic Regression

After the optimization of the model, the ROC is 0.783, indicating that the fitting effect of this

model has been greatly improved compared with that before the optimization, but better

results will be obtained if better data sets are used.

Figure 4-5：ROC of Optimized Logistic Regression

The accuracy is 0.9, indicating that the model has better prediction ability than before

optimization. However, the generalization ability still has great limitations, that is to say, if the

data set is changed to test the Machine Learning model, the results need to be discussed.
44
 Dissertation

4.3.2 Optimized Naïve Bayes Model

Classification accuracy is 0.55, classification error is 0.45, sensitivity is 0.99 and specificity is

0.13, meaning that this model has a not good performance and there is around 45%

possibility of the wrong prediction.

Figure 4-6：Confusion Matrix Of Naïve Bayes Classifier

Furthermore, the ROC is 0.743, the result shows that the dataset fit is better than the naïve

Bayes model before optimization. However, the model still has the above-mentioned

problems, including the imbalance of data caused by too few data features, and the

generalization ability needs to be improved. Similarly，this model can be used to judge

whether returns will occur at the customer level.

Figure 4-7：ROC of Naïve Bayes Classifier

45
 Dissertation

4.3.3 Optimized Neural Network

classification accuracy is 0.94, classification error is 0.06, sensitivity is 0.94 and specificity is

0.71, meaning that this model has a relatively strong performance but there is only around

6% possibility of the wrong prediction.

Figure 4-8：Confusion Matrix of Neural Network

The ROC is 0.801, indicating that compared with logistic regression and naïve bayes, the

Neural Network model has relatively good predictability, but the defects still need to be solved

in future research.

Figure 4-9：ROC Of Neural Network

4.3.4 Comparison of Predictive Methods

According to the evaluation results of each machine learning model, the Neural Network

model has high accuracy and low errors, and the model has reached a relatively ideal state,

that is, the prediction of unreturned products is unreturned, and the prediction of returned

products is returned. The naive Bayesian model has low accuracy and high errors. The model
46
 Dissertation

is not ideal, that is, it predicts that there are more returned goods and more returned goods.

The accuracy of the logistic regression model is high, but the specificity cannot be predicted,

which indicates that there are problems in prediction, and the confusion matrices FP and TN

have defects.

Table 4-5: Performance of Three Model

Logistic Regression Naïve Bayes Neural Network

Classification
0.93 0.55 0.94
Accuracy
Classification Error 0.07 0.45 0.06
Sensitivity 0.93 1.00 0.94
Specificity 0.13 0.71
ROC 0.783 0.743 0.802
According to the value of ROC, it can be directly seen that the prediction effect of neural

network model is better, and the accuracy of Naïve Bayes prediction is the lowest. Neural

Network showed the optimal performance for all the data promotions as compared to Naïve

Bayes and Logistic Regression.

The ROC results of the three models are not much different, which indicates that there are

certain defects in the data itself, and the impact of these shortcomings cannot be avoided

through different models. Although the prediction ability of the optimized model has been

greatly improved, and the accuracy rate has reached a higher level. However, each index

has its own bias, so even the evaluation results cannot fully explain the actual performance

of the model.

5. Discussion

5.1 Discussion of Three Machine Learning Models

Among the three machine learning models selected, each machine learning algorithm has its

unique advantages and is suitable for the data type used in this study. That's why we chose

these models. For logistic regression, it performs well for simple datasets. In the datasets

used in this study, there are only two independent variables and one binary dependent

variable. However, this dataset is suited for logistic regression and also will not reduce the

47
 Dissertation

performance of model. And Logistic regression is applicable when the dependent variable is

a binary variable and the independent variables are categorical or continuous (Ershadi and

Omidzadeh, 2018). 2) A Logistic Regression model is less likely to be over-fitted but it can

be overfitted in high dimensional datasets.

In terms of Naïve Bayes, advantages include 1) Naïve Bayes is based on the independence

assumption where training is very easy and fast. It requires considering each attribute in each

class separately. It is a straightforward test involving the use of tables and calculating

conditional probabilities according to a normal distribution. Even violating the independence

assumption, the performance of Naive Bayes is comparable to the performance of most

advanced classifiers (Gladence et al., 2015). 2) A naive Na|ve Bayes model has a higher

asymptotic error than logistic regression, but naïve Bayes converges faster and approaches

the higher error of logistic regression, which means that for an infinite training data set, logistic

regression should be superior to the original Naïve Bayes because it has a lower error.

However, due to the limited amount of data, Naïve Bayes may outperform regression since

it requires less data to achieve optimal performance. (Witteveen et al., 2018). 3) In order to

obtain good results from Naive Bayes classifiers, it is necessary to collect a large number of

records. In this dataset，we have more than 160k observations. (Aidaroos et al., 2010).

In terms of neural network,1) In contrast to linear relation algorithms, neural network

algorithms have the advantage of high precision, high precision, and high reliability when

there is uncertainty regarding variable relationships and distribution forms between data or

when complex systems cannot express the relationship between input and output data with

general relation. 2) A small amount of knowledge about the problem is sufficient to achieve

positive results, which is noteworthy due to the fact that this model does not require a great

deal of specific information about variables (Bennett et al., 2013). 3) self-adjusting ability to

a given set of data (Sharghi et al., 2018).even if the dataset has been handling not enough

to fit an efficient model, neural network model will adjust predictable capacity automatically

to get a higher accuracy.

5.2 The Limitations of This Study

48
 Dissertation

This study has the following obvious limitations. First, the data analysis basis of the paper is

based on a German retail company. However, in fact, the consumption habits of each country

or region will be greatly different under the influence of various lifestyle. For example, Asian

countries are used to a more conservative consumption habit. Therefore, their purchasing

habits are more cautious so as to fewer product problems, and less frequency of products

returns. On the contrary, in Europe, consumption habits are completely different, as

consumption habits are more radical.

Second, the analysis is based on the data of second-hand electronic products, which

suggests the kind of product is single. At the same time, the reasons for returning electronic

products and consumables are certainly not same, and the impact of demographic

characteristics on product returns is different from other products.

Third, the frequency of product returns may be the fact that if a person tends to return

infrequently, the reasons for these rare returns may be related to some serious problems in

the ordered product, such as failure or damage during delivery. On the contrary, if a person

tends to return relatively frequently, the reason for the return is less likely to be related to the

actual problems of the product, but more likely to be related to the mismatch between the

product and personal needs, desires, or expectations. Therefore, the returned products can

be classified, and the factors that really affect the product quality can be excluded for further

analysis. In this study, since the product types cannot be matched with specific transaction,

the influence of this reason cannot be excluded. In other words, if a second-hand product is

of poor quality, the probability of returning is very high. Otherwise, the probability of returning

is small.

Fourth, in our data, the average purchasing power and average age of the region are matched

by the zip code of individuals who purchase and return goods. The purchasing power and

age of each region are the average value of that region, which represents the relative status

of the region, that is, the overall purchasing power of the region is high, but it does not rule

out that individuals have low purchasing power. As for age, the data we use can only show

that the region as a whole is younger or older, but it cannot show that individuals must be like

49
 Dissertation

this. Therefore, the data itself is biased. Meanwhile, although we have considered the

problems of the data itself as much as possible in the optimization, such as the overfitting

problem caused by too many observations and too few features, which leads to poor

generalization effect. However, from the optimized ROC, the fitting results and prediction

accuracy of the data are only 0.75-0.8, indicating that the fitting and merging of the model

have not reached the optimal state.

Fifth, the selected model has its own advantages and limitations In terms of logistic regression,

it is quite sensitive to noise and overfitting. Especially, in this dataset there are Few

independent variables, but many observations.

In terms of the Naïve Bayes classifier, 1) The Naive Bayes model is capable of generating

classification bias, since the influence of these two attributes may be overvalued and the

influence of other attributes may be undervalued (Aidaroos et al., 2010). 2) vanishing values

can also be explained by combining several small probabilities together (e.g. 0.053). In this

study, the average age and average purchasing power are matched by the zip code. However,

in some regions, the product is purchased only twice, but the data set of the whole day is

more than 160K, so the probability of occurrence is very low, thus vanishing value will happen.

In terms of the Neural Network, 1) it is impossible to establish a suitable model for the

purposes of business decisions due to the lack of standard or fixed rules for governing the

design and development of appropriate models (Abrahart et al., 2012). The inability to

incorporate knowledge acquired from existing physical laws into ANNs also serves as a

drawback to their use. 2) Overparametrization and overfitting problems are common

problems among artificial neural networks (Adeyemo et al., 2018; Sayagavi and Charhate,

2017), especially in the absence of appropriate input selection and early stopping techniques.

To sum up, this study has certain limitations, which may lead to the analysis results not fully

consistent with the actual situation. For example, in this study, we found that purchasing

power and average age do not have a particularly large impact on product returns. Relatively

speaking, age has a greater impact on product returns. However, if we use the data of other

product types, such as clothes, cosmetics, and food., the data of other regions may deviate

50
 Dissertation

from our results. In addition, we generally think that Logistic Regression and Naive Bayes are

suitable for binary data, and Neural Networks often have better prediction results because

they have no fixed formula mechanism, but in practical problems, it needs to be analyzed

according to specific conditions. In future related research, more suitable models can be

selected to predict based on more comprehensive data and other appropriate analysis

methods, such as ABM decision making model.

6 Conclusion

6.1 Main conclusion

In this study, we mainly analyze the influence of purchasing power and average age on

product return. Specifically, 1) do the two factors have an impact on product return? 2) Which

factor has a greater impact on product returns? Purchasing power in essence reflects the

disposable income of consumers; The influence of average age indicates whether there is a

direct relationship between age and product return. For the first point, the study found that

the two factors have an impact on product return at some level. In other words, changes in

purchasing power will influence product returns. In summary, as the purchasing power and

age increase will increase customers' willingness to return goods. This conclusion is

inconsistent with the research conclusion of makkonen et al. (2021). Specifically, as the age

increases, there is a tendency to reduce returns. makkonen et al. (2021) also explored the

relationship between income and return but did not get a clear conclusion. For the second

point, the influence of purchasing power on product return is greater than that of age. Among

the coefficients of the logistic regression model, the coefficient of X1 (purchasing power)

is1.78, and the coefficient of X2 (average age) is 1.43, indicating that purchasing power has

a greater impact on the dependent variable Y.

The influence of age on product returns can be explained by the great differences in

consumption habits among consumers of different ages. For older consumers, they have

higher requirements for products, and they are not good at exploring products through online

information, which may lead to more returns. On the contrary, for young people, the frequency

of online shopping is high, and the value of consumption of young people is relatively low.
51
 Dissertation

They have a greater tolerance for accepting secondary consumption and will choose less

returns to avoid trouble. The impact of purchasing power on returns may be related to income

level. For customer with strong purchasing power, the consumption level is usually high.

Compared with customer with low purchasing power, if they also buy the same type of

products, people with strong purchasing power are willing to spend more money on the same

type of products, which will naturally make it easier to buy appropriate products, thus reducing

the possibility of returns. On the contrary, consumers with low purchasing power may choose

products with lower cost performance because they want to buy cheap products, which will

make it easier to return the goods.

6.2 Suggestions For Reducing Product Returns

The significance of this study is that results can help retailers identify the demographic

characteristics that are more prone to frequent returns and apply the analysis results to solve

the problem of product returns, and ultimately reduce the return cost. Consequently, we give

some suggestions based on the outcome found. For example, if there is an obvious

correlation between product return and age, different return policies are used for specific

customers. If the customer is older than 50 years old, a higher service fee is charged for

return. For customers aged 30-50, a part of the service fee is charged for return. For

customers aged 20-30, it can be returned for free. Customers can also be "credit rating" on

the consumption platform. Customers who frequently return goods have low credit scores,

and such customers will be charged a certain fee when returning goods. For different age

groups, you can also send message prompts at the time of customer shopping checkout,

such as "your return records are too many, which may affect your shopping experience on

the shopping platform and implement a stricter return strategy. Please return with caution.

Some studies have shown that appropriate prompts can play a deterrent role. Of course, a

very important judgment of the retailers is what age group the customers belong to and what

consumption level they have. These can also be calculated according to consumption records

and combined with machine learning algorithms.

52
 Dissertation

6.3 Suggestions For Future Research

We have explained the limitations of this study in detail in the previous part. In particular, in

our data, the average purchasing power and average age of the region are matched by the

zip code of individuals who purchase and return goods. There is a deviation between the

average value and the real value of the actual individual, so there is a deviation in the data

itself. If actual data, for example through questionnaire surveys, can be obtained in the future,

the bias can be avoided, and analysis results will be more accurate. In addition, the data itself

has too few features, too many observations, and a small proportion of data with a return

value of 1 (indicating return). Although some reasonable methods can be used to adjust and

achieve significant improvements, if we want to improve the generalization ability, apply it to

more data, and get consistent conclusions and judgments, we still need to get more efficient

dataset. Future research should focus on the diversity of data features that need to be

improved and the imbalance of dependent variables so that more data features can be used

to improve the generalization effect. Based on the characteristics of some models, we

selected three models that we thought were suitable, but the results were not good enough.

finally, if there is no problem with the data itself, researchers can try other more effective

models to get a reasonable result and widely to deployment.

53
 Dissertation

References

Abrahart, R.J., Anctil, F., Coulibaly, P., Dawson, C.W., Mount, N.J., See, L.M., Shamseldin,

A.Y., Solomatine, D.P., Toth, E. and Wilby, R.L. (2012) ‘two decades of anarchy? Emerging

themes and outstanding challenges for neural network river forecasting’. Progress in Physical

Geography, 36(4), pp.480-513. doi:10.1177/0309133312444943

Adeyemo, J., Oyebode, O. and Stretch, D. (2018) River flow forecasting using an improved

artificial neural network. Available at ：

[Link] (Accessed: 10 July 2022).

Al-Aidaroos, K.M., Bakar, A.A. and Othman, Z. (2010) ‘Naive Bayes variants in classification

learning’. In 2010 international conference on information retrieval & knowledge management

(CAMP),pp. 276-281. IEEE. doi: 10.1109/INFRKM.2010.5466902.

Ambilkar, P., Dohale, V., Gunasekaran, A. and Bilolikar, V. (2022) ‘Product returns

management: a comprehensive review and future research agenda’. International Journal of

Production Research, 60(12), pp.3920-3944. doi: 10.1080/00207543.2021.1933645

Anderson, E.T., Hansen, K. and Simester, D. (2009) ‘The option value of returns: Theory and

empirical evidence’. Marketing Science, 28(3), pp.405-423. doi:10.1287/mksc.1080.0430

Babić Rosario, A., Sotgiu, F., De Valck, K. and Bijmolt, T.H. (2016) ‘The effect of electronic

word of mouth on sales: A meta-analytic review of platform, product, and metric

factors’. Journal of Marketing Research, 53(3), pp.297-318. doi:10.1509/jmr.14.0380

54
 Dissertation

Baden, D. and Frei, R. (2021) ‘Product Returns: An Opportunity to Shift towards an Access-

Based Economy?’. Sustainability, 14(1), p.410. doi:10.3390/su14010410

Bandi, C., Moreno, A., Ngwe, D. and Xu, Z. (2018) Opportunistic returns and dynamic pricing:

Empirical evidence from online retailing in emerging markets. Harvard business school

marketing unit working paper, (19-030), pp.19-030. Available

at:[Link] (Accessed: 17 August

2022).

Bechwati, N.N. and Siegal, W.S. (2005) ‘The impact of the prechoice process on product

returns’. Journal of Marketing Research, 42(3), [Link].1509/jmkr.2005.42.3.358

Bennett, N.D., Croke, B.F., Guariso, G., Guillaume, J.H., Hamilton, S.H., Jakeman, A.J.,

Marsili-Libelli, S., Newham, L.T., Norton, J.P., Perrin, C. and Pierce, S.A.(2013)

‘Characterising performance of environmental models’. Environmental Modelling &

Software, 40, pp.1-20. doi:10.1016/[Link].2012.09.011

Bonifield, C., Cole, C. and Schultz, R.L. (2010) ‘Product returns on the internet: a case of

mixed signals?’. Journal of Business Research, 63(9-10), pp.1058-1065.

doi:10.1016/[Link].2008.12.009

Bower, A.B. and Maxham III, J.G. (2012) ‘Return shipping policies of online retailers:

Normative assumptions and the long-term consequences of fee and free returns’. Journal of

Marketing, 76(5), pp.110-124. doi:10.1509/jm.10.0419

Boyle, M. (2006) Best Buy’s giant gamble. Available at:

[Link] (Accessed: 19 August

2022).

55
 Dissertation

Chang, M. K., Cheung, W., & Lai, V. S. (2005) ‘Literature derived reference models for the

adoption of online shopping’. Information & Management, 42(4), 543–

[Link].1016/[Link].2004.02.006.

Chen, Y. and Xie, J. (2008) ‘Online consumer review: Word-of-mouth as a new element of

marketing communication mix’, Management Science, vol. 54, no. 3, pp. 477-491.

doi:10.1287/mnsc.1070.0810

Cheung, C. M. K., Chan, G. W. W., & Limayem, M. (2005) ‘A critical review of online

consumer behavior: Empirical research’. Journal of Electronic Commerce in Organizations,

3(4), 1–19. doi:10.4018/jeco.2005100101

Chollet, F. (2021) Deep learning with Python. Simon and Schuster.

Chan, G., Cheung, C., Kwong, T., Limayem, M. and Zhu, L. (2003) Online consumer behavior:

a review and agenda for future research. Available

at:[Link] (Accessed: 9

September 2022).

Craciun, G.M. (2006) Mood effects on ordinary unethical behavior. University of South

Carolina.

Cuffie, H.G., Najar, R.I. and Khasawneh, M.T. (2020) Topic Modeling for Customer Returns

Retail Data. Available

at:[Link]

origsite=gscholar&cbl=51908 (Accessed: 9 September 2022).

Cui, H., Rajagopalan, S. and Ward, A.R. (2020) ‘Predicting product return volume using

machine learning methods’. European Journal of Operational Research, 281(3), pp.612-627.

doi:10.1016/[Link].2019.05.046

56
 Dissertation

Cullen, J., Tsamenyi, M., Bernon, M. and Gorst, J.(2013) ‘Reverse logistics in the UK retail

sector: A case study of the role of management accounting in driving organisational

change’. Management Accounting Research, 24(3), pp.212-227.

doi:10.1016/[Link].2013.01.002

D'Innocenzio, A. and Beck, R. (2011) Wal-Mart, humbled king of retail, plots rebound.

Available at: [Link]

07-02-2011/ (Accessed: 11 September 2022).

De, P., Hu, Y. and Rahman, M.S. (2013) ‘Product-oriented web technologies and product

returns: An exploratory study’. Information Systems Research, 24(4), pp.998-

[Link].1287/isre.2013.0487

Dzyabura, D., El Kihal, S. and Ibragimov, M.(2018) Leveraging the power of images in

predicting product return rates. Available at:

[Link]

(Accessed: 11 September 2022).

Ershadi, M.J. and Omidzadeh, D. (2018) Customer validation using hybrid logistic regression

and credit scoring model: a case study. Available

at:[Link]

[Link] (Accessed: 11 July 2022).

Fan, D., Cui, X.M., Yuan, D.B., Wang, J., Yang, J. and Wang, S. (2011) ‘Weight of evidence

method and its applications and development’. Procedia Environmental Sciences, 11,

pp.1412-1418. doi:10.1016/[Link].2011.12.212

Fan, H., Khouja, M. and Zhou, J. (2022) ‘Design of win-win return policies for online

retailers’. European Journal of Operational Research, 301(2), pp.675-693. doi:

10.1016/[Link].2021.11.030

57
 Dissertation

Fawcett, T. (2006) ‘An introduction to ROC analysis’. Pattern Recognition Letters, 27(8),

[Link]: [Link]

Fontana, R., Luciano, E. and Semeraro, P. (2021) ‘Model risk in credit risk’, Mathematical

Finance, 31(1), pp. 176–202. doi: 10.1111/mafi.12285.

Frei, R., Bines, A., Lothian, I. and Jack, L. (2016) ‘Understanding reverse supply

chains’. International Journal of Supply Chain and Operations Resilience, 2(3), pp.246-266

Available at:[Link] (Accessed: 20 June 2022)

Frei, R., Jack, L. and Krzyzaniak, S.A.(2020) ‘Sustainable reverse supply chains and circular

economy in multichannel retail returns’. Business Strategy and the Environment, 29(5),

pp.1925-1940. Available at:[Link]

(Accessed: 20 June 2022)

Gareth, J., Daniela, W., Trevor, H. and Robert, T. (2013) An introduction to statistical learning:

with applications in R. Spinger. Available

at:[Link] (Accessed: 25June 2022)

Garretson, J.A. and Burton, S.(2005) ‘The role of spokescharacters as advertisement and

package cues in integrated marketing communications’. Journal of Marketing, 69(4), pp.118-

132. doi:10.1509/jmkg.2005.69.4.118

Gelbrich, K., Gäthke, J. and Hübner, A. (2017) ‘Rewarding customers who keep a product:

How reinforcement affects customers' product return decision in online retailing’. Psychology

& marketing, 34(9), pp.853-867. doi:10.1002/mar.21027

Gelbrich, K., Gäthke, J. and Hübner, A. (2017) ‘Rewarding customers who keep a product:

How reinforcement affects customers' product return decision in online retailing’. Psychology

& marketing, 34(9), pp.853-867. doi:10.1002/mar.21027

58
 Dissertation

Ghaith, M. and Li, Z. (2020) ‘Propagation of parameter uncertainty in SWAT: A probabilistic

forecasting method based on polynomial chaos expansion and machine learning’. Journal of

Hydrology, 586, p.124854. doi: 10.1016/[Link].2020.124854

Gladence, L.M., Karthi, M. and Anu, V.M.(2015) ‘A statistical comparison of logistic

regression and different Bayes classification methods for machine learning’. ARPN Journal

of Engineering and Applied Sciences, 10(14), pp.5947-5953. Available at:

[Link]

L/publication/282921131_A_statistical_comparison_of_logistic_regression_and_different_b

ayes_classification_methods_for_machine_learning/links/570228d408aea6b7746a8689/A-

statistical-comparison-of-logistic-regression-and-different-bayes-classification-methods-for-

[Link] (Accessed: 31June 2022).

Grind, K. (2013) ‘Retailer REI ends era of many happy returns’. Wall Street Journal, 16.

Available

at:[Link]

(Accessed: 31June 2022).

Harris, L.C.(2008) ‘Fraudulent return proclivity: an empirical analysis’. Journal of

Retailing, 84(4), pp.461-476. doi:10.1016/[Link].2008.09.003

Hilbe, J.M. (2009) Logistic regression models. Chapman and hall/CRC.

Jack, L., Frei, R. and Krzyzaniak, S.A.C. (2019) Buy online, return to store: the challenges

and opportunities of product returns in a multichannel environment. Available

at:[Link] (Accessed: 29 June 2022).

Janakiraman, N. and Ordóñez, L. (2012) ‘Effect of effort and deadlines on consumer product

returns’. Journal of Consumer Psychology, 22(2), pp.260-271.

doi:10.1016/[Link].2011.05.002

59
 Dissertation

Jiang, P. and Rosenbloom, B.(2005) ‘Customer intention to return online: price perception,

attribute-level performance, and satisfaction unfolding over time’. European journal of

marketing, 39(1/2), pp.150-174. doi:[Link]

Johnson, A.A., Ott, M.Q. and Dogucu, M.(2022) Bayes Rules!: An Introduction to Applied

Bayesian Modeling. CRC Press.

Janocha, K. and Czarnecki, W.M.(2017) On loss functions for deep neural networks in

classification. Available at: [Link] (Accessed: 29 July 2022).

Johnson, K.K. and Rhee, J.(2008) ‘AN INVESTIGATION OF CONSUMER TRAITS AND

THEIR RELATIONSHIP TO MERCHANDISE BORROWING WITH

UNDERGRADUATES’. Journal of Family & Consumer Sciences Education, 26(1). Available

at:[Link] (Accessed: 16 July 2022).

Ketzenberg, M.E., Abbey, J.D., Heim, G.R. and Kumar, S.(2020) ‘Assessing customer return

behaviors through data analytics’. Journal of Operations Management, 66(6), pp.622-645.

doi:10.1002/joom.1086

Kirmani, A. and Rao, A.R.(2000) ‘No pain, no gain: A critical review of the literature on

signaling unobservable product quality’. Journal of marketing, 64(2), pp.66-79.

doi:10.1509/jmkg.64.2.66.18000

Kumar, S. and Yamaoka, T.(2007) ‘System dynamics study of the Japanese automotive

industry closed loop supply chain’. Journal of Manufacturing Technology Management.

Available at:

[Link]

(Accessed: 16 July 2022).

60
 Dissertation

Lee, S. and Yi, Y.(2017) ‘Seize the Deal, or Return It Losing Your Free Gift”: The Effect of a

Gift-With-Purchase Promotion on Product Return Intention’. Psychology & Marketing, 34(3),

pp.249-263. doi: [Link]

Li, J., He, J. and Zhu, Y.(2018, July) ‘E-tail product return prediction via hypergraph-based

local graph cut’. In Proceedings of the 24th ACM SIGKDD International Conference on

Knowledge Discovery & Data Mining, pp. 519-527. doi:10.1145/3219819.3219829

Li, Y., Xu, L. and Li, D.(2013) ‘Examining relationships between the return policy, product

quality, and pricing strategy in online direct selling’. International Journal of Production

Economics, 144(2), pp.451-460. doi:[Link]

Ma, J. and Kim, H.M.(2016) ‘Predictive model selection for forecasting product

returns’. Journal of Mechanical Design, 138(5), [Link]: 10.1109/ITMC.2018.8691285

Makkonen, M., Frank, L. and Kemppainen, T.(2021) The effects of consumer demographics

and payment method preference on product return frequency and reasons in online shopping.

Available at: [Link] (Accessed: 16 July 2022).

Meng, J. and Li, H.(2017) ‘An efficient stochastic approach for flow in porous media via sparse

polynomial chaos expansion constructed by feature selection’. Advances in Water

Resources, 105, pp.13-28. doi: [Link]

Minnema, A., Bijmolt, T.H., Gensler, S. and Wiesel, T.(2016). ‘To keep or not to keep: Effects

of online customer reviews on product returns’. Journal of retailing, 92(3), pp.253-267.

doi:10.1016/[Link].2016.03.001

Mukhopadhyay, S.K. and Setoputro, R.(2005) ‘Optimal return policy and modular design for

build-to-order products’. Journal of Operations Management, 23(5), pp.496-506.

doi:[Link]

61
 Dissertation

Pei, Z. and Paswan, A.(2018) ‘CONSUMERS'LEGITIMATE AND OPPORTUNISTIC

PRODUCT RETURN BEHAVIORS IN ONLINE SHOPPING’. Journal of Electronic

Commerce Research, 19(4), pp.301-319. Available

at:[Link] (Accessed: 16 July 2022).

Petersen, J.A. and Kumar, V.(2010) ‘Can product returns make you money?’. MIT Sloan

Management Review, 51(3), p.85. Available at:[Link]

content/uploads/sites/4/2017/02/Can-Returns-Make-You-Money_White-[Link]

(Accessed: 20 July 2022).

Piron, F. and Young, M.(2000) ‘Retail borrowing: insights and implications on returning used

merchandise’. International Journal of Retail & Distribution Management.

doi:10.1108/09590550010306755

Potdar, A. and Rogers, J.(2012) ‘Reason-code based model to forecast product

returns’. Foresight. doi:10.1108/14636681211222393

Rabuzin, K., Varazdin, C., Karthika, S., Tamil Nadu, I., Bose, S., Kannan, A. and Keyvanpour,

M.R.(2014) International Journal of Data Warehousing and Mining. [Link]

[Link]/[Link]?tid%3D106857%26ptid%3D91422%26ctid%3D15%26t%3Dtable+of+c

ontents (Accessed: 20 July 2022).

Ram, A. (2016) ‘UK retailers count the cost of returns’. Financial Times. Available at:

[Link] (Accessed 20July

2022)

Sahoo, N., Dellarocas, C. and Srinivasan, S. (2018) ‘The impact of online product reviews on

product returns’. Information Systems Research, 29(3), pp.723-738.

doi:10.1287/isre.2017.0736

62
 Dissertation

Sala, S., Benini, L., Beylot, A., Castellani, V., Cerutti, A., Corrado, S., Crenna, E., Diaconu,

E., Sanyé-Mengual, E., Secchi, M. and Sinkko, T.(2019) ‘Consumption and Consumer

Footprint: methodology and results’. Indicators and Assessment of the Environmental Impact

of European Consumption. Luxembourg. Available

at:[Link] (Accessed 20July 2022)

Salehzadeh, R., Tabaeeian, R.A. and Esteki, F.(2020) ‘Exploring the consequences of

judgmental and quantitative forecasting on firms' competitive performance in supply

chains’. Benchmarking: An International Journal, 27(5), pp.1717-1737.

doi:[Link]

Saranya, C. and Manikandan, G. (2013) ‘A study on normalization techniques for privacy

preserving data mining’. International Journal of Engineering and Technology (IJET), 5(3),

pp.2701-2704. Available at:

[Link]

(Accessed 20July 2022)

Sayagavi, V. and Charhate, S. (2017) ‘Applications of soft tools to solve hydrological

problems for an integrated Indian catchment’. International Journal of Water Resources and

Environmental Engineering, 9(7), pp.150-161. Available

at:[Link]

rhate%2C+2017&btnG= (Accessed 28 July 2022)

Schmidt, R.A., Sturrock, F., Ward, P. and Lea-Greenwood, G.(1999) ‘Deshopping–the art of

illicit consumption’. International Journal of Retail & Distribution Management. doi:

[Link]

Sharghi, E., Nourani, V., Soleimani, S. and Sadikoglu, F.(2018) ‘Application of different

clustering approaches to hydroclimatological catchment regionalization in mountainous

63
 Dissertation

regions, a case study in Utah State’. Journal of Mountain Science, 15(3), pp.461-484.

doi:10.1016/[Link].2014.09.001

Speights, D. and Rittman, T.(2015) Fighting return fraud during the holiday season. White

Paper, The Retail Equation. Available at: [Link]

season (accessed 25 November 2015).

Stock, J., Speh, T. and Shear, H.(2002) ‘Many happy (product) returns’. Harvard business

review,80(7), pp.16-17. Available at:[Link] (accessed

25 July 2022).

Su, X.(2009) ‘Consumer returns policies and supply chain performance’. Manufacturing &

Service Operations Management, 11(4), pp.595-612. doi:10.1287/msom.1080.0240

Teo, T.S. and Yeong, Y.D. (2003) ‘Assessing the consumer decision process in the digital

marketplace’. Omega, 31(5), pp.349-363. doi:10.1016/S0305-0483(03)00055-0

Tüylü, A.N.A. and Eroglu, E. (2022) ‘The prediction of product return rates with ensemble

machine learning algorithms’. Journal of Engineering Research.

doi: [Link]

Urbanke, P., Kranz, J. and Kolbe, L.(2015) Predicting product returns in e-commerce: the

contribution of mahalanobis feature extraction. Available

at:[Link]

Kranz/publication/283270887_Predicting_Product_Returns_in_E-

Commerce_The_Contribution_of_Mahalanobis_Feature_Extraction/links/5720d84c08aead2

6e721322b/Predicting-Product-Returns-in-E-Commerce-The-Contribution-of-Mahalanobis-

[Link]

64
 Dissertation

Walsh, G., Albrecht, A.K., Kunz, W. and Hofacker, C.F.(2016) ‘Relationship between online

retailers’ reputation and product returns’. British Journal of Management, 27(1), pp.3-20.

doi :[Link]

Wang, W., Lesner, C., Ran, A., Rukonic, M., Xue, J. and Shiu, E.(2020, April) ‘Using small

business banking data for explainable credit risk scoring’. In Proceedings of the AAAI

Conference on Artificial Intelligence,Vol. 34, No. 08, pp. 13396-13401. doi::

[Link]

Wang, Y., Anderson, J., Joo, S.J. and Huscroft, J.R.(2019) ‘The leniency of return policy and

consumers’ repurchase intention in online retailing’. Industrial Management & Data Systems.

doi:[Link]

Witteveen, A., Nane, G.F., Vliegen, I.M., Siesling, S. and IJzerman, M.J.(2018) ‘Comparison

of logistic regression and Bayesian networks for risk prediction of breast cancer

recurrence’. Medical decision making, 38(7), pp.822-833.

doi:[Link]

Wood, S.L.(2001) ‘Remote purchase environments: The influence of return policy leniency

on two-stage decision processes’. Journal of Marketing Research, 38(2), pp.157-169.

doi:10.1509/jmkr.38.2.157.18847

Yan, R. and Cao, Z.(2017) ‘Product returns, asymmetric information, and firm

performance’. International Journal of Production Economics, 185, pp.211-222.

doi :[Link]

Yoo, S.H.(2014) ‘Product quality and return policy in a supply chain under risk aversion of a

supplier’. International Journal of Production Economics, 154, pp.146-155.

doi:[Link]

65
 Dissertation

Zhou, L., Dai, L., & Zhang, D. (2007) ‘Online shopping acceptance model – A critical survey

of consumer factors in online shopping’. Journal of Electronic Commerce Research, 8(1), 41–

62. Available at:[Link] (accessed

25 July 2022).

Zhou, L., Xie, J., Gu, X., Lin, Y., Ieromonachou, P. and Zhang, X.(2016) ‘Forecasting return

of used products for remanufacturing using Graphical Evaluation and Review Technique

(GERT)’. International Journal of Production Economics, 181, pp.315-324.

doi:[Link]

Zhou, W., Hinz, O. and Benlian, A.(2018) ‘The impact of the package opening process on

product returns’. Business Research, 11(2), [Link]:[Link]

017-0055-x

Zhu, Y., Li, J., He, J., Quanz, B.L. and Deshpande, A.A.(2018) ‘July. A Local Algorithm for

Product Return Prediction in E-Commerce’. In IJCAI ,pp. 3718-3724. Available

at:[Link]

m_for_Product_Return_Prediction_in_E-Commerce/links/5cf504ce299bf1fb18538ae3/A-

[Link] (accessed 25 July 2022).

66