Strategic Returns Prevention in E-Commerce: Simulating Financial and

Environmental Outcomes Through Agent-Based Modeling

Marie Niederlaender

1 a

, Urs Liebau

2 b

, Yajing Chen

2 c

, Emil Breustedt

2 d

, Saad Driouech

1 e

and Dirk Werth

1 f

August-Wilhelm Scheer Institut, Uni Campus D 5 1, Saarbr

ucken, Germany

August-Wilhelm Scheer Institut, Center for Digital Greentech, Clausthal-Zellerfeld, Germany

{ﬁrstname.lastname}@aws-institut.de

Keywords:

Agent-Based Simulation, Customer Returns Behavior, Return Prevention, Sustainable Supply Chain.

Abstract:

Product returns pose an environmental and ﬁnancial burden on manufacturers and online retailers worldwide,

especially in the fashion sector. Over 50% of all ordered garments end up being returned, which gives rise

to an ongoing search for approaches to successfully manage returns or to avoid returns in the ﬁrst place. For

both approaches, an accurate prediction of returns can be useful, since it allows for an improved inventory risk

assessment and strategic reselling of garments, while also providing crucial information on common drivers

of return rates. This study focuses on preventive strategies in the context of customers placing selection

orders in online shops. An Agent based approach provides insight into the outcomes of three different return

prevention strategies, which are compared with the original outcome of real world data from a German clothing

manufacturer selling garments for special occasions. The four outcomes are analysed in terms of their ﬁnancial

and environmental impact, utilising common life cycle assessment strategies.

1 INTRODUCTION

The global fashion e-commerce market has continu-

ously grown each year over the past years alongside

the growth of online retail in general. With fashion

e-commerce being forecast to reach over US 781.5

billion in 2024 and an estimate to reach US 1.6 tril-

lion by 2030, the fashion sector is the largest B2C

e-commerce market segment to date (Statista, 2024).

Alongside the increasing revenue comes the ﬁnan-

cial and environmental burden of an increasing vol-

ume of product returns. In 2022, most product returns

were associated with the fashion sector, for example

in Germany, a share of 91% of returns were fash-

ion items (Forschungsgruppe Retourenmanagement,

2022). Due to the additional costs that returns im-

pose on online retailers due to the necessary repro-

cessing steps before reselling the items, many busi-

nesses opt to send returned items to landﬁlls. In 2021

https://orcid.org/0009-0008-1935-821X

https://orcid.org/0009-0002-6825-8342

https://orcid.org/0009-0002-3599-4140

https://orcid.org/0009-0004-0688-510X

https://orcid.org/0000-0002-2445-1098

https://orcid.org/0000-0003-2115-6955

in Germany, approximately 17 million returned items

were disposed, while other European countries ex-

hibit even higher disposal rates (Forschungsgruppe

Retourenmanagement, 2022). Another issue with re-

turns is posed by the large amount of parcel shippings,

where each returned package generates an average

of 1.5 kg of CO

equivalents (Forschungsgruppe Re-

tourenmanagement, 2022), contributing to the fashion

sector being among the top three of most polluting in-

dustries. Minimising returns overall has the largest

impact in reducing environmental harm and ﬁnancial

strain. Strategies can be twofold - either pursuing

preventive strategies with the main goal of avoiding

returns in the ﬁrst place, or reactive strategies aim-

ing at improving the handling of returns in terms of

sustainability, time and cost (Deges, 2021). In either

case, it is important to understand the drivers of re-

turns. The reasons for consumers to return items are

diverse and may include, but not be limited to the fol-

lowing reasons: Unmet expectations with regards to

look or quality, personal preferences, dissatisfaction

with the ﬁt, the wrong size, too much time passed be-

tween order and delivery, receiving the incorrect item,

ordering for someone else. Therefore, it can be advan-

tageous for retailerst to know when which items will

likely be returned. The return reason can also be an

Niederlaender, M., Liebau, U., Chen, Y., Breustedt, E., Driouech, S. and Werth, D.

Strategic Returns Prevention in E-Commerce: Simulating Financial and Environmental Outcomes Through Agent-Based Modeling.

DOI: 10.5220/0013180600003890

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 17th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2025) - Volume 1, pages 453-462

ISBN: 978-989-758-737-5; ISSN: 2184-433X

453

important indicator of the item’s condition upon re-

turn. A large quantity of fashion returns are related

to size or selection of items. The latter case is given

when customers order multiple versions of the same

item or multiple items from the same category (like

multiple dresses) with the intention of buying only

one or a few of them. This very common behavior in

the fashion sector is referred to as bracketing (Bim-

schleger et al., 2019).

This study focuses on an Agent based simulation

of preventive strategies based on the targeting of se-

lection orders by utilising awareness-raising methods

in order to have customers make alterations to their

shopping card in favor of the environment. This re-

search is part of a larger scope of research that aims at

developing an AI-based recommendation system for

the prediction of returns and improved return manage-

ment. This part of the research aims at presenting a

blueprint for estimating the possible ﬁnancial and en-

vironmental impact that can be achieved using the re-

turn predictions and their associated return reasons to

target speciﬁc return driving behavior before an order

has been placed. The ﬁrst section references related

work in the context of this study. The subsequent sec-

tion describes the methods used for the setup of the

Agent based approach and the methods used to infer

the environmental impact from the outputs. Third, we

analyse the outcomes of the simulation in general as

well as from an environmental perspective based on

material consumption.

2 RELATED WORK

This section presents a selection of work related to the

topics of this study, including sustainability research

in the fashion industry, development of size ﬁnding

tools and returns prediction models, as well as Agent

based approaches to customer behavior simulation in

an online shop setting. Yang et al. (2017) show a

comprehensive overview of the sustainability efforts

and shortcomings in the fashion industry, touching on

different topics ranging from the selection of sustain-

able materials to the effects of free returns and the

lack of credibility of sustainability claims made by

fashion companies. In the context of size recommen-

dation, Eshel et al. (2021) propose a deep learning

framework based on transformers for the size predic-

tion across different clothing segments (Mens, Wom-

ens, Kids, Unisex clothing) and categories (Tops, Bot-

toms, Dress/Skirt, Footware). In addition, utilising

the size predicting features for enhancing eBay’s sim-

ilar items recommendation service. Nestler et al.

(2021) propose SizeFlags, a probabilistic Bayesian

model based on weakly annotated large-scale data

from customers for prediction of the most suitable

size. In this study, the subjectivity of size percep-

tion is emphasised, stating that the ’true’ size of a cus-

tomer often remains unknown and can vary greatly by

external factors including changes in physique over

time and the mindset around what ﬁts best.

When it comes to prediction of customer behavior,

Hummel et al. (2011) present an Agent based simula-

tion in order to predict customer behavior in an online

shop setting when the preferred payment method is

not available, taking into account gender speciﬁc pref-

erences of payment methods from real-world data.

Another recent study compares ﬁve different classical

Machine Learning Algorithms for the predictions of

returns, highlighting the importance of features rep-

resenting bracketing behavior and ordering habits of

customers (Niederlaender et al., 2024). Agent-based

modeling can be used to simulate how targeting spe-

ciﬁc order behaviors, like size-related and selection-

related returns, can inﬂuence shopping cart composi-

tion and ultimately impact broader factors such as ma-

terial consumption, costs, and proﬁts. This bottom-up

approach allows for a detailed analysis of individual

customer behavior and the effectiveness of prevention

strategies. By incorporating real-world data and sim-

ulating customer interactions with size ﬁnding tools

and targeted messages, this model can capture the het-

erogeneity of customer behavior and assess the sus-

tainability implications of different strategies.

3 METHODS

This section presents the methods utilised for the esti-

mation of achievable ﬁnancial and environmental im-

pact when employing strategies for the avoidance of

returns and is composed of two steps. First, the data

input, setup and output of the Agent-based simulation

are described in subsection 3.1. Based on the out-

put generated, ﬁnancial aspects are analysed. An en-

vironmental analysis is conducted in the second step

and the methodology for the determination of relevant

parameters based on the output are described in sub-

section 3.2.

3.1 Simulation Setup

This simulation showcases three different return

prevention strategies pertaining to selection orders,

where a customer ordered multiples of the same cloth-

ing item in several sizes or a customer ordered mul-

tiple items from the same clothing category (for ex-

ample dresses, pants, jackets). The latter case may

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

454

Figure 1: Flowchart representing the decision making process for each agent based on the initial shopping cart. Each of the

four approaches (Initial, Normal, Targeted, Incentive) is simulated separately with the ground truth shopping cart data as a

starting point.

not always indicate bracketing behavior, however, the

dataset is a deciding factor.

The Input Data. This simulation is based on real-

world sales and returns data of a German manufac-

turer of festive dresses and garments for special oc-

casions. Therefore, the assumption that the major-

ity of customers only intend to keep one festive gar-

ment from a category, one festive dress for example,

is considered reasonable. The overall return rate of

the shop is 72%, which may be caused by the speciﬁ-

cation on special occasions. The tabular data contains

orders in the time span from March 1st 2023 to Febru-

ary 29 2024, where each order can be identiﬁed by a

unique order ID. Further columns include the size of

the items, a unique article ID, clothing category, the

color, ﬁt and style of the garment, but also the mate-

rial composition in percent and the weight of the top

layer fabric per square meter in grams. The latter two

columns formed the basis for the calculations of mate-

rial consumption, which in turn have been used in the

environmental analysis. For items where the weight

of the top layer fabric was not given a value, the av-

erage weight over all top layer fabrics was assigned.

To estimate material consumption, for each cloth-

ing category the average square meters of total fab-

ric needed, including multiple layers were estimated

to be: Cocktail dress 3.0 m

; Stola 1.0 m

; Evening

gown 4.5m

; Day dress 2.75m

; Dress 3.0m

; Jump-

suit 3.5 m

; Bolero 1.25m

; Skirt (midi) 2.0 m

; Top

1.25m

; Blouse 2.0 m

; Corset 1.25 m

; Skirt (long)

3.0m

; Pants 2.0 m

; Jacket 2.25 m

; Skirt (short)

1.25m

The weight of the fabric per square meter of fabric

was approximated to match the given weight of the

top layer fabric provided in the data. Due to the large

amount of orders giving rise to averaging effects, the

over- and underestimation of heavier or lighter gar-

ments were considered to be negligible. From the

total weight estimation of the garment, subsequently

the total amount of fabric components in the garments

were estimated using the given material composition.

Materials used as fabric components are given in Ta-

ble 1.

Input Parameters. For each order in the real world

data, an agent was created to act as an entity which al-

ters the contents of the shopping basket as a result of

the returns prevention strategies in place. Depending

on the simulations scenarios, which are explained in

the next paragraph, different input parameters drive

the agents proneness to act a certain way. For the

commonly known scenario where a size ﬁnding tool

is in place but no active measures are taken, the

Sizeﬁnder Adoption Rate S

normal

determines the prob-

ability with which a customer agent uses the size ﬁnd-

ing tool in a situation where they would place a size

selection order instead in the ground truth shopping

cart. For each agent, S

normal

was drawn from a nor-

mal distribution with mean 0.2 and a standard devia-

tion of 0.05, which makes some customer agents more

prone to use the tool than others. The mean value and

standard deviation for the Sizeﬁnder Adoption Rate

were determined by interviewing online fashion re-

tailers which have their own size ﬁnding tool in place

in their online shop. Based on this, an exemplary as-

sessment of a person’s Sensitivity to react to a targeted

Popup due to the contents of their shopping basket in-

dicating size or category selection orders was made.

An increased proneness to react compared to the sim-

ple presence of a size ﬁnding tool is considered real-

istic, since it raises the awareness of the customer re-

garding their own shopping behavior. Therefore, the

Targeted Popup Sensitivity S

target

for each agent was

Strategic Returns Prevention in E-Commerce: Simulating Financial and Environmental Outcomes Through Agent-Based Modeling

455

drawn from a normal distribution with mean 0.4 and a

standard deviation of 0.1. Similarly, a person’s sensi-

tivity to react to a version of the targeted popup which

is not solely based on raising awareness but also on an

incentive to act on the recommendations is considered

to be even higher. Incentives might include offering

free shipping or free returns in the case of ordering

only what the size ﬁnding tool recommends. Thus,

the Incentive Sensitivity S

incentive

for each agent was

drawn from a normal distribution with mean 0.6 and

a standard deviation of 0.1. The size ﬁnding tool was

set to correctly suggest the best size for the customer

in 80% of cases, so the accuracy a of the tool was set

to be a = 0.8. To track the total costs present in the

context of the shipping and return management, the

cost for shipping a parcel (either as return or as ship-

ping) were set to be EUR 5, where in the context of

this simulation, the return cost is borne by the retailer,

the outward shipping cost is borne by the customer.

The costs for reprocessing an incoming return parcel

were estimated to be EUR 10 on average, which is

determined by the interviews conducted with online

fashion retailers.

Simulation Scenarios. First, in order for the simu-

lation results to generalise better, for each order three

agents were created, which creates potentially differ-

ent outcomes for each unique order. The observed

scenarios are illustrated as a ﬂow chart in Figure 1.

Second, three separate strategies are simulated in-

dependently from each other and compared to the

ground truth order data, which is shown as the ’Do

Nothing’ or ’Initial’ state scenario. In the ﬁrst strat-

egy, the ’Normal’ approach, the simulation is set up to

mimick the prevention of size related selection orders

due to the presence of a size ﬁnding tool on the web-

site. The possible positive impact on non-selection or-

ders due to better size estimation is not considered in

this study, meaning that non-selection orders remain

unaltered throughout the entire simulation. In the

’Normal’ scenario, customer agents may or may not

use the size ﬁnding tool depending on their personal

preference determined by S

normal

. Consequently, de-

pending on the correctness of the tool in each particu-

lar case, determined by the sizeﬁnder accuracy a, the

customer agent may either keep one item in the cor-

rect size, or they may keep the item in the incorrect

size.

The second returns prevention strategy, referred

to as the ’Targeted’ approach, the customer agent is

made aware of the potential to use the size ﬁnder in

the case of a size related selection in their shopping

cart. In that case, they may or may not use the tool,

depending on their sensitivity to such an awareness-

raising strategy given by S

target

. If the customer in

turn has an order which can be classiﬁed as a selection

order based on ordering n ≥ 2 items from the same

clothing category, the customer agent is also made

aware of the potential returns caused and the effects

that coincide with it. Depending on the customers

sensitivity to the message S

target

, the customer agent

proceeds to remove up to n − 1 items of this clothing

category from the shopping cart. The third strategy

is the same as the second, with a key difference: The

agent is given an incentive to act according to the rec-

ommendations. The eagerness to react to the offer-

ing of an incentive like free returns or free shipping

for each agent is given by their incentive sensitivity

incentive

Outputs and Implementation. The simulation was

implemented completely using python with com-

monly used libraries like pandas and the Agent based

modeling library Mesa. After each simulation sce-

nario has been run, the potentially altered shopping

cart including returns and non-returns are saved, as

well as the total value of the cart and the value of the

returned items in the cart. The information if the cus-

tomer agent received any advantageous offer from the

Incentive case is also saved as an output. The total

amount of outward shippings and return shippings is

tracked along with the total return processing cost.

3.2 Environmental Analysis

This chapter discusses the methodology used in this

study to analyse the environmental impacts under dif-

ferent scenarios. Sustainability assessments are in-

complete without consideration of environmental, one

of the 3 pillars of sustainable development (European-

Commission, 2024).The goal of this sustainability as-

sessment is to explore how variations in material con-

sumption across different scenarios impact the envi-

ronment.

3.2.1 Methodology Selection

Life Cycle Assessment (LCA), a method commonly

used in the literature to quantify the potential envi-

ronmental impacts of a product system, was adopted.

LCA involves four phases: Goal and Scope, Life Cy-

cle Inventory (LCI), Life Cycle Impact Assessment

(LCIA), and Interpretation. Since this study does not

involve a complete LCA, it focuses solely on the po-

tential environmental impacts of materials under dif-

ferent scenarios. Therefore, this study concentrates

on two key phases of LCA: LCI and LCIA ISO 14040,

ISO 14044).

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

456

As part of the LCI phase, environmental data

related to raw materials is extracted from the LCI

database, focusing on the environmental impact of

fabrics during its production. In addition to categoris-

ing these data, the LCIA method is used to classify the

impacts of different materials under various scenarios

into speciﬁc impact categories. This categorization

allows for a better understanding of the environmen-

tal impacts of each material under different scenarios.

(ISO 14040, 2006)(ISO 14044, 2006)

3.2.2 LCI Database and LCIA Method

In LCA studies related to raw materials, there are vari-

ations in the selection of LCI databases across the lit-

erature. Table 1 presents different LCI databases and

LCIA methods used in different studies that have con-

ducted LCAs on the same type of fabrics as this study.

From Table 1, it is evident that the majority of rel-

evant literature has selected the ecoinvent database, as

their LCI database. Therefore, this study also chooses

to use the ecoinvent database as the data source for

analyzing the environmental impacts of different fab-

rics.

In contrast to the selection of LCI databases, there

is a signiﬁcant variation in the choice of LCIA meth-

ods among previous studies. These differences arise

from variations in research locations and objectives.

For example, IPCC primarily provides results for the

Global Warming Potential (GWP) category (IPCC,

2022), while USEtox focuses on toxicity research in

the United States and includes only categories re-

lated to toxicity (USEtox, 2024). Many articles have

utilised the CML method, however, CML has not

been updated since 2017 (Leiden-University, 2016).

Given these uncertainties, this study has chosen to

adopt the ReCiPe 2016 method, as it includes 18 mid-

point categories and is continuously updated (ReCiPe,

2016).

3.2.3 Analysis of Environmental Impacts

As shown in Table 1, the variations in LCIA meth-

ods have also led to differences in the selection of

impact categories by different authors. The ReCiPe

2016 method selected for this study encompasses var-

ious environmental impact pathways (ReCiPe, 2016).

However, the environmental impacts of different fab-

ric can vary signiﬁcantly (Velden et al., 2014)(Parvez

et al., 2018)(Wu, 2020)(Shena et al., 2010)(Guo et al.,

2021)(Manteco, 2022)(Shuang et al., 2023)(Fangli

et al., 2021)(Gomez-Campos et al., 2020)(Marek

et al., 2023). To better demonstrate how changes

in the ordering process can impact the environment

from the perspective of fabric production and the 6

most relevant categories to the textile industry were

selected for further focus. Carbon footprint (CC),

Terrestrial Ecotoxicity (TEtox), and Human Toxicity

(HT, both cancer and non-cancer) were selected be-

cause they are the Top 3 most frequently examined

categories in other studies, as shown in Table 1. Ad-

ditionally, Freshwater Ecotoxicity (FEtox) and Fossil

Fuel Depletion were included. First, FEtox is consis-

tently highlighted in EU article as one of the three ma-

jor environmental issues in textile production, due to

the large volumes of water pollution generated during

the process (European-Parliament, 2024). Second,

excessive fossil fuel consumption in fabric produc-

tion contributes signiﬁcantly to environmental degra-

dation (Zaman et al., 2023). Reducing fossil fuel use

has been identiﬁed as an important factor in mak-

ing textile production more sustainable (Zaman et al.,

2023).

This study will analyze how different scenarios

of fabric usage affect six environmental impact cat-

egories. By identifying the fabric with the greatest

environmental impact, this research will support sus-

tainable decision-making.

Normen Clature Environmental Analysis. The

Parameters for interpretation of environmental im-

pacts are described by the following Normen Cla-

ture: AD: Abiotic depletion; AP: Acidiﬁcation po-

tential; AWARE: Available water remaining; CLCD:

Chinese life cycle database; CC: Climate change; ET:

Eutrophication; ETO: Ecotoxicity; EDIP: Environ-

mental design of industrial products; FD: Fine dust;

FPM: Fine particulate matter formation; FE: Fresh-

water eutrophication; FEtox: Freshwater ecotoxicity;

GWP: Global warming potential; HT: Human toxi-

city; HTc: Human toxicity, cancer; HTnc: Human

toxicity, non-cancer; IPCC: International panel on cli-

mate change; IR: Ionising radiation; LU: Land use;

ME: Marine eutrophication; MEtox: Marine ecotox-

icity, RD: Mineral, fossil and renewable resource de-

pletion; OD: Ozone depletion; OF: Ozone formation;

OFh: Ozone formation, Human health; PM: Partic-

ulate matter; POF: Photochemical ozone formation;

POxF: Photochemical oxidant formation; PED: Pri-

mary energy demand; SOD: Stratospheric ozone de-

pletion; TA: Terrestrial acidiﬁcation; TE: Terrestrial

ecosystems; TEtox: Terrestrial ecotoxicity; WU: Re-

source depletion-water; USEtox: United nation envi-

ronment program and society of environmental toxi-

cology chemistry; WC: Water consumption

Strategic Returns Prevention in E-Commerce: Simulating Financial and Environmental Outcomes Through Agent-Based Modeling

457

Table 1: Summary of LCI databases and LCIA methods used with impact categories from various literature sources.

Material type LCI database LCIA method Impact categories References

Nylon Ecoinvent

IPCC, USEtox, ReCiPe,

IMPACT 2002+, CML,

Ecopoints 97, Eco-

indicator 99, Green-

house gas protocol

GWP, ETO, AP, ET,

POxF, FD, TE, FE, SOD,

IR, OFh, FPM, OF, TA,

MEtox, HTc, HTnc, LU,

(Velden et al.,

2014)(Parvez et al.,

2018)

Polyester

Gabi,

Ecoinvent

EDIP, IPCC, USE-

tox, ReCiPe, IMPACT

2002+, CML

GWP, WC, ETO, AP, ET,

POxF, FD, AD, OD, HT,

FEtox, Tetox

(Velden et al., 2014)

(Wu, 2020) (Shena et al.,

2010)

Elastane Ecoinvent

IPCC, USEtox,

ReCiPe,IMPACT 2002+

GWP, ETO, AP, ET,

POxF, FD, MEtox; WC

(Velden et al., 2014)

Viscos

CLCD,

Ecoinvent

CML, IPCC

AD, PED, WU, AP,

GWP, ET, CC, OD, HT,

FEtox, Tetox, POxF, AP

(Shena et al., 2010) (Guo

et al., 2021) (Manteco,

2022)

Cotton Ecoinvent

CML, IMPACT 2002+,

ReCiPe, EDIP, USEtox,

Eco indicator 99, foot-

print, Australian impact

method, Australian Indi-

cator Set V3, IPCC

CC, OD, TA, FE, FE-

tox, HT, WC, AD, Tetox,

POxF, AP, ET

(Shena et al., 2010)

(Manteco, 2022)

(Shuang et al., 2023)

(Fangli et al., 2021)

Flax

Ecoinvent,

Agribalse

ILCD 2001+

CC, OZ, PM, IR, POF,

AP, FE, ME, RD

(Gomez-Campos et al.,

2020)

Metallic fab-

rics

Ecoinvent

IPCC, USEtox 2,

AWARE

GWP, HTc, HTnc, FE-

tox, WC

(Marek et al., 2023)

4 DISCUSSION OF RESULTS

4.1 Comparison of the Simulation

Outcomes

The relative changes in the simulation outcomes com-

pared to the initial ground truth case are summarised

in Table 2. The absolute values are compared in Fig-

ure 2. It becomes clear that the difference between the

targeted approach and the normal approach is not very

pronounced. The normal approach often saves more

resources than the targeted approach, despite the lat-

ter’s broader scope. While the targeted approach re-

duces total returns, it also decreases the number of

items kept. The incentive approach, while effective

in reducing returns, might impact sales. Additional

costs associated with incentives should be considered.

While the overall reduction in material consumption

is similar across cases, it’s not directly proportional to

the reduction in returns. This is due to varying prod-

uct compositions and return reasons. Further research

is needed to optimize pop-up timing and placement,

as well as address complex cases involving multiple

high-risk items. Due to the high individuality of the

carts as well as the unknown sensitivity of customers

to small changes in size und unknown personal prefer-

ences, the decision process of customers remain pre-

Table 2: Relative changes of the different scenarios (N =

Normal, T = Targeted, I = Incentive) compared to the initial

scenario.

Outcome

[%]

Total Returns -0.67 -0.69 -9.14

Total Non Returns -1.93 -1.84 -10.4

Return Shippings -0.20 -0.17 -1.90

Outward Shippings -0.24 -0.23 -0.72

Sum Price Returns -0.67 -0.71 -9.46

Sum Price Non Returns -2.01 -1.97 -11.1

Return Shipping Costs -0.20 -0.17 -1.90

Total Outward Shipping

Costs

-0.24 -0.23 -0.72

Return Processing Costs -0.67 -0.69 -9.14

Nylon -0.79 -1.59 -5.90

Polyester -1.12 -1.06 -9.68

Metallic Fabric -0.93 -0.95 -6.16

Elastane -1.23 -1.16 -8.75

Viscose -0.83 -1.37 -4.41

Cotton -1.18 -1.49 -11.3

Flax -0.92 0.00 -1.84

dominantly difﬁcult to comprehend. Online retailers

are therefore encouraged to explore their customers

sensitivities S

normal

, S

target

and S

incentive

in the con-

text of their own online shop. Further, the sensitivi-

ties S

target

and S

incentive

may vary depending on the re-

turn category, for example making customers on aver-

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

458

Figure 2: Comparison of simulation outcomes for the four scenarios, where Initial represents the ground truth data.

age more susceptible to removing size selection items

than removing category selection items. This differ-

entiation needs further investigation.

4.2 Impact of Outcomes on Relevant

Environmental Parameters

A key objective of this research is to provide an as-

sessment and classiﬁcation of potential environmen-

tal impacts that may arise due to changes in the or-

dering process. For this reason, the results from the

various scenarios in Figure 4 were analysed for their

environmental impacts using the environmental anal-

ysis method described (see subsection 3.2). As out-

lined in the methodology, the focus was placed on

the following six impact categories: Climate Change

(CC), Freshwater Aquatic Ecotoxicity (FAE), Ter-

restrial Ecotoxicity (TE), Human Toxicity, cancer

(HTC), Human Toxicity, non-cancer (HTNC), Fos-

sil Depletion (FT). The results are presented in two

main sections. An overview of the absolute values in

the scenarios is described. Then the relative share of

the individual materials for the respective impact cat-

egory is discussed in more detail and it is described

whether there are changes in the scenarios compared

to the “Initial” scenario.

4.2.1 Absolute Summed Impacts per Scenario

Figure 3 illustrates the impact of the different cloth-

ing sales scenarios (Initial, Targeted, Incentive, and

Figure 3: Absolute values of different impact metrics for

different ordering scenarios.

Normal) on the six environmental metrics. The units

are not the same for all metrics. Care should therefore

be taken when comparing values across metrics. The

chart highlights the signiﬁcant impact that different

sales strategies can have on the environment. Out-

comes for other metrics are shown in the Appendix.

The ”Incentive” scenario consistently outperforms the

others across all six metrics, indicating that it is the

most sustainable option. The percentage decline in

Strategic Returns Prevention in E-Commerce: Simulating Financial and Environmental Outcomes Through Agent-Based Modeling

459

each metric is between 7 and 9 percent. The high-

est percentage decrease is in Fossil depletion with 8.6

percent and the lowest in Freshwater aquatic ecotox-

icity with 7.5 percent.The Target scenario also low-

ers the impact compared to Normal and Initial. The

highest decrease compared to the Initial Scenario is

in Fossil depletion with 1.2 percent and the lowest

Freshwater aquatic ecotoxicity with one percent. The

CO2 equivalent can be considered in more detail as

an example. The difference between the initial sce-

nario with 253 tons and the incentive scenario with

232 tons corresponds to savings of 21 tons. By way

of comparison, a return ﬂight of 3000 miles (e.g. from

Boston to London and back) emits around one ton of

CO2 per passenger(United States Environmental Pro-

tection Agency, 2018). Freshwater aquatic ecotoxic-

ity has about a ton of difference between Initial with

13t and Incentive with 12t 1.4 DCB equivalent. The

target of a maximum pollutant concentration of 75 mi-

crograms of p-DCB per liter of drinking water set by

the United States Environmental Protection Agency

(2009) (EPA) can be used for reference. Both exam-

ples were used to better illustrate the quantities given.

Overall, it can be seen that, compared to the initial

scenario, only the incentive scenario has a clear inﬂu-

ence and would have an impact reduction of around

8 percent in most impact categories. Target and Nor-

mal also show a reduction in the impact categories,

but this is signiﬁcantly smaller.

4.2.2 Relative Proportion of Individual

Materials per Impact Category

Figure 4: Percentage share of the analysed fabrics of the

total impact in the various metrics.

Figure 4 illustrates the relative impact of different fab-

ric materials on the same six environmental metrics as

in Figure 3. The values are totaled over the 4 scenar-

ios. The material composition of the clothes signiﬁ-

cantly inﬂuences the environmental footprint. The use

of PES (Polyester) and MFT (Metallic Fiber) consis-

tently contributes to a higher impact across all met-

rics compared to other fabrics, despite the fact that

the total consumption of MFT is just a small share

of the total material consumption. PES with percent-

ages between 65 and 33 and MTF between 63 and

29 percent. Nylon and Polyurethane generally exhibit

lower environmental impacts. Viscos, Cotton and

Flax demonstrate the lowest impacts relatively. This

analysis highlights the signiﬁcance of fabric compo-

sition in sustainable clothing consumption. While in-

terventions in the ordering process can have a pos-

itive impact, substantial improvements require more

signiﬁcant changes. Focusing on PES and MFT com-

ponents offers the greatest potential for reducing en-

vironmental impact.

5 CONCLUSION AND OUTLOOK

This work proposed an Agent-Based approach for

the estimation of ﬁnancial and environmental out-

comes when implementing return prevention strate-

gies in fashion e-commerce. Alternative order out-

comes when targeting customers based on size or cat-

egory related selection orders and intervening with

targeted popups or incentives were explored and com-

pared to ground truth order data from a German online

retailer, as well as the scenario of a size ﬁnding tool

in place but not raising awareness about its presence.

Depending on customers unique sensitivity to these

strategies, returns can be prevented by reducing selec-

tion orders, having a positive ﬁnancial impact through

saved return shipping and processing cost, but also

preventing some amount of orders that would have

not been returned in the initial scenario. However, this

aspect needs further investigation taking into account

the possibility of follow-up orders after returning the

ﬁrst time. One key objective of this work is to classify

the sustainability effects that could arise as a result

of changes in the ordering process. For this reason,

the sustainability effects of the results from the var-

ious scenarios from Figure 4. were examined using

the sustainability analysis method described in sub-

section 3.2. In a ﬁrst step, the impact of the outcome

from the initial scenario was determined. In a sec-

ond step, the initial scenario was compared with the

changed outcomes of the different scenarios, which

were also evaluated with regard to their sustainabil-

ity effects. In a third step, these sustainability assess-

ments are used to compare the various impacts of the

scenarios. Future work is suggested to investigate the

impact of targeting other return types besides selec-

tion order induced returns, while taking varying sen-

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

460

sitivities dependent on the return type into account.

The potential for cart abandonment should be consid-

ered. Real-world customer responses to these strate-

gies must be studied. Online retailers are encouraged

to analyze their own sales and return data to estimate

the potential impact on ﬁnancial and environmental

sustainability. Expanding these strategies to other e-

commerce types could offer new insights. AI and ML

can help predict returns and inform decision-making

for efﬁcient and sustainable return processing. A re-

turn prediction system could identify high-risk orders

and items, enabling targeted interventions to prevent

returns. By analyzing customer return history, per-

sonalized recommendations can be made to reduce re-

turns without discouraging legitimate ones.

ACKNOWLEDGEMENTS

This research was funded in part by the Ger-

man Federal Ministry of Education and Research

(BMBF) under the project OptiRetouren (grant num-

ber 01IS22046B). It is a joint project of the August-

Wilhelm Scheer Institut, INTEX, HAIX and h+p.

August-Wilhelm Scheer Institut is mainly entrusted

with conducting research in AI for forecasting returns

volume and for recommendations based on AI.

REFERENCES

Bimschleger, C., Patel, K., and Leddy, M. (2019). Bringing

it back: Retailers need a synchronized reverse logis-

tics strategy. Technical report, Deloitte Development

LLC.

Deges, F. (2021). Retourencontrolling im online-handel.

Controlling – Zeitschrift f

ur erfolgsorientierte Un-

ternehmenssteuerung, 2/2021:61–68.

Eshel, Y., Levi, O., Roitman, H., and Nus, A. (2021). Pre-

size: predicting size in e-commerce using transform-

ers. In Proceedings of the 44th international ACM

SIGIR conference on research and development in in-

formation retrieval, pages 255–264.

European-Commission (2024). Sustainable development.

European-Parliament (2024). The impact of textile pro-

duction and waste on the environment (infographics).

Technical report.

Fangli, C., Pinghua, X., Xiang, J., Laili, W., and Jiang, C.

(2021). A review: life cycle assessment of cotton tex-

tiles. Special Issue on Circular Economy. Industria

textila, (72):1.

Forschungsgruppe Retourenmanagement (2022). Ergeb-

nisse des europ

aischen retourentachos ver

offentlicht.

https://www.retourenforschung.de/info-

ergebnisse-des-europaeischen-retourentachos-

veroeffentlicht.html. Online; accessed 2023-01-26.

Gomez-Campos, A., Vialle, C., Rouilly, A., Sablayrolles,

C., and Hamelin, L. (2020). Flax ﬁber for technical

textile: a life cycle inventory. Journal of Cleaner Pro-

duction.

Guo, S., Li, X., Zhao, R., and Gong, Y. (2021). Comparison

of life cycle assessment between lyocell ﬁber and vis-

cose ﬁber in china. The International Journal of Life

Cycle Assessment, 26:1545-1555.

Hummel, A., Kern, H., and D

ohler, A. (2011). An agent-

based simulation of payment behavior in e-commerce.

In Multiagent System Technologies: 9th German Con-

ference, MATES 2011, Berlin, Germany, October 6-7,

2011. Proceedings 9, pages 41–52. Springer.

IPCC (2022). Climate change 2022, mitigation of climate

change. Technical report.

ISO 14040 (2006). Environmental management-life cycle

assessment-principles and framework. International

Organization for Standardization.

ISO 14044 (2006). Environmental management-Life cy-

cle assessment-Requirements and guidelines, volume

14044. International Organization for Standardiza-

tion.

Leiden-University (2016). Handbook on life cycle assess-

ment.

Manteco (2022). Life cycle assessment on mwool® by

manteco. Technical report, Manteco.

Marek, S., Marzena, W.-A., and Ganczewski, G. (2023).

Striving for a less toxic production of metallized tex-

tiles - environmental impact assessment. Journal of

Cleaner Production.

Nestler, A., Karessli, N., Hajjar, K., Weffer, R., and Shir-

vany, R. (2021). Sizeﬂags: reducing size and ﬁt re-

lated returns in fashion e-commerce. In Proceedings

of the 27th ACM SIGKDD Conference on Knowledge

Discovery & Data Mining, pages 3432–3440.

Niederlaender, M., Lodi, A., Gry, S., Biswas, R., and

Werth, D. (2024). Garment returns prediction for ai-

based processing and waste reduction in e-commerce.

In Proceedings of the 16th International Conference

on Agents and Artiﬁcial Intelligence - Volume 2:

ICAART, pages 156–164. INSTICC, SciTePress.

Parvez, M. M. A., Huda, N., Farjana, S. H., and Candace,

L. (2018). Environmental proﬁle evaluations of piezo-

electric polymers using life cycle assessment. IOP

Conference Series: Earth and Environmental Science.

ReCiPe (2016). Pr

e sustainability - method for the life cy-

cle impact assessment (lcia). https://pre-sustainability.

com/articles/recipe/. Accessed: 2024-10-15.

Shena, L., Worrell, E., and Patel, M. K. (2010). Environ-

mental impact assessment of man-made cellulose ﬁ-

bres. Resources, Conservation and Recycling. 55:260-

274.

Shuang, C., Lisha, Z., Lirong, S., Huang, H. Q., Ying, Z.,

Li, L. X., Xiangyu, Y., Yi, L., and Laili, W. (2023).

A systematic review of the life cycle environmental

performance of cotton textile products. Science of the

Total Environment, 883.

Statista (2024). Fashion ecommerce report

2023. https://www.statista.com/topics/9288/

fashion-e-commerce-worldwide/#topicOverview.

Online; accessed 2024-09-30.

Strategic Returns Prevention in E-Commerce: Simulating Financial and Environmental Outcomes Through Agent-Based Modeling

461

United States Environmental Protection Agency (2009).

Drinking water contaminants: p-dichlorobenzene (p-

dcb). Accessed: 2024-10-15.

United States Environmental Protection Agency (2018).

Emission factors for greenhouse gas inventories. Ac-

cessed: 2024-10-15.

USEtox (2024). The usetox model — usetox®.

Velden, N. M., Patel, M. K., and Vogtl

ander, J. G. (2014).

Lca benchmarking study on textiles made of cotton,

polyester, nylon, acryl, or elastane. . The International

Journal of Life Cycle Assessment, 19(2), 331–356.

Wu, Z. (2020). Haode evaluating the life-cycle environ-

mental impacts of polyester sports t-shirts. IOP Con-

ference Series: Earth and Environmental Science.

Yang, S., Song, Y., and Tong, S. (2017). Sustainable re-

tailing in the fashion industry: A systematic literature

review. Sustainability, 9(7):1266.

Zaman, S., Umair, M., and Javid, A. (2023). Sustainable

production practices in textiles.

APPENDIX

Table 3: Outcomes of Initial and Normal impact metrics for

different ordering scenarios.

Category Initial Normal Unit

Climate change

incl biogenic

carbon

253733.28 251091.32 kg CO2-

Eq.

Freshwater

aquatic ecotoxi-

city

13889.52 13750.29 kg 1.4

DCB eq.

Marine aquatic

ecotoxicity

18070.87 17889.51 kg 1.4

DCB eq.

Terrestrial eco-

toxicity

1019839.03 1009264.53 kg 1.4

DCB eq.

Freshwater Eu-

trophication

125.70 124.46 kg P eq.

Marine Eutroph-

ication

45.70 45.23 kg N eq.

Terrestrial Acid-

iﬁcation

695.41 688.14 kg SO2

eq.

Photochemical

Ozone Forma-

tion. Ecosys-

tems

565.56 559.55 kg NOx

eq.

Ionizing Radia-

tion

21460.72 21245.99 kBq Co-

60 eq.

Human toxicity

cancer

14643.98 14493.22 kg 1.4-

DB eq.

Human toxicity

non-cancer

263596.40 260927.14 kg 1.4-

DB eq.

Fine Particulate

Matter Forma-

tion

301.25 298.08 kg

PM2.5

eq.

Photochemical

Ozone Forma-

tion. Human

Health

518.20 512.70 kg NOx

eq.

Table 3: Outcomes of Initial and Normal impact metrics for

different ordering scenarios (cont.).

Category Initial Normal Unit

Stratospheric

Ozone Depletion

0.784 0.775 kg CFC-

11 eq.

Land use 4984.98 4932.55 Annual

crop eq.

Fossil depletion 96251.62 95233.25 kg oil

eq.

Metal depletion 2748.16 2720.18 kg Cu

eq.

Freshwater Con-

sumption

3667.45 3628.48 m3

Table 4: Outcomes of Targeted and Incentive impact met-

rics for different ordering scenarios.

Category Targeted Incentive Unit

Climate change

incl biogenic

carbon

251091.30 232872.84 kg CO2-

Eq.

Freshwater

aquatic ecotoxic-

ity

13751.61 12859.24 kg 1.4

DCB eq.

Marine aquatic

ecotoxicity

17891.43 16726.44 kg 1.4

DCB eq.

Terrestrial eco-

toxicity

1009505.05 937694.95 kg 1.4

DCB eq.

Freshwater Eu-

trophication

124.46 116.57 kg P eq.

Marine Eutrophi-

cation

45.20 42.06 kg N eq.

Terrestrial Acidi-

ﬁcation

688.10 637.73 kg SO2

eq.

Photochemical

Ozone Forma-

tion. Ecosystems

559.62 516.92 kg NOx

eq.

Ionizing Radia-

tion

21248.71 19875.78 kBq Co-

60 eq.

Human toxicity

cancer

14495.31 13481.50 kg 1.4-

DB eq.

Human toxicity

non-cancer

260960.88 243528.13 kg 1.4-

DB eq.

Fine Particulate

Matter Formation

298.10 275.83 kg

PM2.5

eq.

Photochemical

Ozone Forma-

tion. Human

Health

512.76 473.89 kg NOx

eq.

Stratospheric

Ozone Depletion

0.776 0.710 kg CFC-

11 eq.

Land use 4929.10 4564.97 Annual

crop eq.

Fossil depletion 95245.71 88053.60 kg oil

eq.

Metal depletion 2719.68 2536.61 kg Cu

eq.

Freshwater Con-

sumption

3625.23 3347.68 m3

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

462