• Nenhum resultado encontrado

The impact of picking success on customer purchasing behavior: an example of a Portuguese omni-channel retailer

N/A
N/A
Protected

Academic year: 2021

Share "The impact of picking success on customer purchasing behavior: an example of a Portuguese omni-channel retailer"

Copied!
60
0
0

Texto

(1)

FUNDAÇÃO GETULIO VARGAS

ESCOLA BRASILEIRA DE ADMINISTRAÇÃO PÚBLICA E DE EMPRESAS MESTRADO EXECUTIVO EM GESTÃO EMPRESARIAL

DISSERTAÇÃO APRESENTADA À ESCOLA BRASILEIRA DE ADMINISTRAÇÃO PÚBLICA E DE EMPRESAS PARA OBTENÇÃO DO GRAU DE MESTRE

ILIAS IZIRAREN

Rio de Janeiro - 2019

The impact of picking success on customer purchasing behavior –

an example of a Portuguese omni-channel retailer

(2)

XXI Dados Internacionais de Catalogação na Publicação (CIP)

Ficha catalográfica elaborada pelo Sistema de Bibliotecas/FGV

Iziraren, Ilias

The impact of picking success on customer purchasing behavior : an example of a Portuguese omni-channel retailer / Ilias Iziraren. – 2019.

59 f.

Dissertação (mestrado) - Escola Brasileira de Administração Pública e de Empresas, Centro de Formação Acadêmica e Pesquisa.

Orientadora: Laura Wagner.

Coorientadora: Isabella Vasconcelos. Inclui bibliografia.

1. Sistema de picking. 2. Comércio eletrônico. 2. Comportamento do consumidor. 3. Logística empresarial. 4. Distribuição de mercadorias. I. Wagner, Laura. II.Vasconcelos, Isabella Freitas Gouveia de. III. Escola Brasileira de Administração Pública e de Empresas. Centro de Formação Acadêmica e Pesquisa. IV. Título.

CDD – 658.785 Elaborada por Márcia Nunes Bacha – CRB-7/4403

(3)
(4)

II

Table of contents

Index of abbreviations ... III Index of variables ... IV Abstract ... V-VI

1. Introduction ... 1

2. Literature Review ... 3

2.1. The business model of online retailers... 3

2.2. Online supermarkets / grocery stores ... 5

3. Institutional setup ... 6

4. Hypothesis development ... 8

4.1. The impact of a low picking success on the customer behavior ... 8

4.2. Moderating factors ... 9

5. Data ... 12

5.1. Available data ... 12

6. Identification strategy ... 15

6.1. Matching strategy ... 15

6.2. Using the matching output to implement causal analyses ... 18

7. Estimation results ... 19

7.1. The impact of a low picking success on the customer behavior ... 19

7.2. Moderating factors ... 21

8. Robustness checks and limitations ... 26

8.1. Substitution ... 27

8.2. Threshold variation ... 31

9. Extensions ... 34

9.1. Geographical differences ... 34

9.2. Development of the picking success ... 35

9.3. Basket composition ... 36

10. Conclusions ... 37 References ... VII-X Appendices ... XI-XX

(5)

III

Index of abbreviations

BI = Business Intelligence

CRM = Customer Relationship Management OCE = Online Customer Experience

(6)

IV

Index of variables

Picking Success = Percentage value between 0 and 100, depending on the number of items successfully delivered out of the total order

Treat = Binary variable; 1 for Picking Success > 85%, 0 for Picking Success < 85%

Slot cost = Costs for chosen delivery slot

Delivery Cost = Total costs for delivering the shipment

Days to delivery = Remaining days until the order will be delivered

Fresh items (%) = Percentage of fresh goods in the order

NSKU = Number of stock-keeping-units (number of items per delivery)

Home delivery (%) = Percentage of customers who ordered their delivery to their home

Store delivery (%) = Percentage of customers who ordered their delivery to their local store

Total Purchases = Number of all items ordered by a specific customer

Relative average item price = Difference between the average price of the first order to the average order of the following ones

DDHigh = “Delivery Date high”, 1 if the days to delivery are above average, 0 otherwise

DCostHigh = “Delivery Costs high”, 1 if the delivery costs are above average, 0 otherwise

RDCostsHigh = “Relative Delivery Costs high”, 1 if the delivery costs relative to the total order value are above average, 0 otherwise

DiscountHigh = 1 if the discounts received on the order are high, 0 otherwise

% Poor Delivery = Percentage of deliveries with a Picking Success below 85%

(7)

V

Abstract (English)

Title: The impact of picking success on customer purchasing behavior – an example of a Portuguese omni-channel retailer

Author: Ilias Iziraren

Problem definition: Supermarkets and grocery stores in developed economies are

discovering the internet as a new marketplace via online deliveries. This study reveals the consequences of a low picking success (ratio of ordered to received items) and discusses the magnitude of the product unavailability. The picking success is thereby solely related to the operational performance (i.e. the ability of the seller of delivering the desired items). Academic / practical relevance: A broad product range and thin margins characterize today’s business environment for supermarkets and online retailers; by closely examining on how a poor picking success influences the customers purchasing behavior, important insights are gained regarding the relationship of the online delivery of goods and the respective customer satisfaction and ultimately retention.

Methodology: A dataset consisting of 497,309 order shipments with in total 14,283,175 ordered products has been received through the collaboration with a large Portuguese omni-channel retailer. To establish relationships, causal analyses are used on matched pairs of customers. Thereby, a set of possible explanatory variables is analyzed (such as received discounts, remaining days to delivery etc.) to identify interrelations between the variables and the customer behavior after receiving a delivery with a low picking success.

Results: By analyzing the data it becomes evident that a low picking success leads not only to a lower number of items per shipment afterwards, but also to a lower total number of purchases. Furthermore, customers are thereupon switching to lower-priced product alternatives, anticipating the scarcity of more expensive brand products, which might lead to even lower margins. This study also shows, how discounts can be used in order to (partly) compensate for a low picking success.

Managerial implications: Most importantly, a low picking success has to be avoided to sustainably retain customers. In the short-term, a low picking success could be

compensated through intelligent discounts – this is however not a sustainable method to avoid other side effects of a low picking success (which might be amongst others a deterioration in the reputation, a lower perceived brand value or product substitutions).

Key words: Omni-channel retailer, online delivery, customer behavior, retention, picking

(8)

VI

Abstract (Portuguese)

Título: O impacto da escolha do sucesso no comportamento de compra do cliente - um exemplo de um varejista Omni-Channel Português

Autor: Ilias Iziraren

Definição do problema: Supermercados e mercearias em economias desenvolvidas estão descobrindo a internet como um novo mercado através de entregas on-line. Este estudo revela as conseqüências de um baixo sucesso de picking (razão entre itens pedidos e recebidos) e discute a magnitude da indisponibilidade do produto. O sucesso do picking é, portanto, relacionado apenas ao desempenho operacional (ou seja, a capacidade do

vendedor de entregar os itens desejados).

Relevância acadêmica / prática: Uma ampla gama de produtos e margens pequenas caracterizam o ambiente de negócios atual para supermercados e varejistas on-line; Examinando de perto como um sucesso de picking fraco influencia o comportamento de compra dos clientes, obtêm-se insights importantes sobre o relacionamento da entrega on-line de mercadorias e a respectiva satisfação do cliente e, por fim, a retenção.

Metodologia: Um conjunto de dados consistindo de 497.309 remessas de pedidos com um total de 14.283.175 produtos encomendados foi recebido através da colaboração com um grande varejista omni-channel Português. Para estabelecer relacionamentos, as análises causais são usadas em pares de clientes combinados. Desse modo, um conjunto de

possíveis variáveis explicativas é analisado (como descontos recebidos, dias restantes até a entrega etc.) para identificar inter-relações entre as variáveis e o comportamento do cliente após receber uma entrega com baixo sucesso de picking.

Resultados: Ao analisar os dados, torna-se evidente que um baixo sucesso de picking leva não apenas a um menor número de itens por remessa depois, mas também a um menor número total de compras. Além disso, os clientes estão mudando para alternativas de produtos de preço mais baixo, antecipando a escassez de produtos de marcas mais caras, o que pode levar a margens ainda menores. Este estudo também mostra como os descontos podem ser usados para (parcialmente) compensar um baixo sucesso de picking.

Implicações gerenciais: Mais importante, um baixo sucesso de picking tem que ser evitado para reter clientes de maneira sustentável. No curto prazo, um baixo sucesso de picking poderia ser compensado através de descontos inteligentes - este não é um método

sustentável para evitar outros efeitos colaterais de um baixo sucesso de picking (que pode ser, entre outros, uma deterioração na reputação, uma marca percebida mais baixa valor ou um substitucao de produtos infavoraveis).

Palavras-chave: Supermercado, entrega online, comportamento do cliente, retenção, sucesso de picking, descontos, margens, falha de serviço, varejo on-line, qualidade de serviço

(9)

1

1. Introduction

“I mean, they sent us an e-mail a week after we placed our order telling us that it’s out of stock, and their site was still showing that it’s in stock.” By these words, a customer

emphasized his frustration when purchasing goods online at a large American retailer (Holloway, Beatty, 2003).

Not being able to deliver a product may lead to lost sales or product substitution. While in many industries product availability is shown at the time of purchase, this is not (always) the case for online retailers. Thus, the online retailer agrees on delivering a product upfront but only verifies its availability at a later stage. This timely difference diminishes a consumer’s ability to choose a substitute product and might result in dissatisfaction (Breugelmans et al., 2006).

What are the consequences of product unavailability to the omni-channel retailer in terms of customer retention? Does it have ramifications beyond the immediate loss of paying customers? This is the research question this thesis aims to address in particular,

This study establishes links between the quality of online deliveries regarding the respective picking success and subsequent customer behavior. Furthermore, it offers a set of alternative courses of action for the online retailers to prevent customers from taking on undesirable behaviors even in the case of a low picking success.

For this purpose, a proprietary and rich data set of a Portuguese omni-channel retailer has been obtained and analyzed from the first October 2016 to the 29th of September 2017. This

data set contains 497,309 online delivery shipments of almost 15 million individual product purchases, whereby the respective delivery status indicates whether an item is available, substituted or could not be replaced. The four possible values consist of “Total Picking”, “Partial Picking”, “Substitution” and “Stock Rupture”, which have been quantified in this study to allow for an identification of the underlying relationships. The aim of this study is to understand the consequences of product availability on customer retention.

To address this issue, the shipment data has been aggregated on the customer level and customers with different levels of picking success have been closely examined. Thereby

(10)

2

relationships have been identified immediately on the purchase volume subsequent to a delivery with a low picking success.

The methodology of this paper is to identify relationships between independent and dependent variables of pairs of matched customers. The customers in the matched pair had respectively a good and a bad picking but did not differ in other aspects (i.e. customer purchase characteristics). They customers have been divided into a treatment and a control group (meaning that the treatment group received the treatment of getting a delivery with a low picking success, and the control group did not), which are related to the picking success of their first delivery. The results are interesting: Customers with a low picking success purchase in total around 40 products less after their first delivery, thereby not only the total order amount is reduced but also the purchased items per shipment (they purchase over 5 items less per shipment afterwards). This is hardly surprising given the strong customer opinions on deliveries with a low picking success (see above).

Additionally, moderating factors such as the delivery costs and time, discounts and the geographic location have been taken into account. While delivery costs do play a crucial role, the delivery time is an important moderating factor. Customers who placed their order a long time in advance and still received a delivery with a poor picking success reduce the order quantities by an additional 10 items per delivery as compared to those who ordered only shortly in advance. Discounts tend to be an attractive option to compensate customers, the results indicate that by compensating those who received a low picking success with discounts, the order quantity per shipment could be increased by roughly 4.9 items and thereby almost offset the negative effect.

The study indicates that consumers indeed tend to reduce their purchase volume per shipment as well as in total as consequence to a delivery with a poor picking success. Additionally, they change the purchased product type in a manner which reduces the average price per item, thereby impacting the gross margins of the Portuguese omni-channel retailer. While the product category as well as the delivery costs play a subordinated role as mediating factors, the days to delivery significantly impact the customers acceptance of a delivery with a low picking success. Interestingly, the omni-channel retailer has the possibility to counteract the negative consequences regarding the customer behavior by applying targeted and generous

(11)

3

discounts. The results contribute to the existing research about delivery quality as key dimension in online retailing which is summarized next.

2. Literature review

The study is closely related to inventory availability tracking. The particularity of this setting is that usually online retailers track the inventory status in real time (especially when online and offline delivery are separate). As a result, the customer is doubtlessly informed about the availability of the respective product (to anticipate substitution and lost sales through unavailability). However, the chosen Portuguese omni-channel retailer does not reveal the availability up until placing the order. This enables an unbiased analysis of the customer’s first choice, since the customer does not substitute his favourite item based on availability (as it might also be the case offline, see Figure 1). Thereby, a two-folded approach regarding the literature review is chosen: First, the business model of the online retailers in general will be analyzed. Secondly, a closer look at online supermarkets and grocery stores is taken. 2.1. The business model of online retailers

Online retailers have dealt with a setting where they compete against each other and against established offline retailers. The key issue thereby has been always how to successfully attract or “entice” customers (Chen et al., 2002). In their study, the researches used an online survey to point out that the attitude towards the usage of a virtual store is a crucial success factor for important metrics such as the amount of sales or the customer retention. Enders et al. observed already in the year 2000 that the business models of retailing experience a fundamental change with the emerge of a new generation of Internet-savvy consumers, who simply are not accustomed anymore to shopping in physical stores (Enders et al., 2000). With the course of time, the business model of online retailers has become more sophisticated. As the focus of online retailers has shifted to a more customer-centric design and customer segments are targeted more concisely, business intelligence (BI) and customer relationship management (CRM) are nowadays key influencing factors for the success of the business of online retailers (Phan et al., 2010). By utilizing the example of the online retailer Fingerhut, the authors show on the one hand how nowadays a business can completely fail

(12)

4

by focussing purely on traditional wisdom about price discrimination and switching costs, and on the other hand they show how an online retailer was successfully reborn by implementing a new structure and taxonomy based on BI and CRM. The website is thereby the central point of interaction between potential customers and the organization. Those websites are nowadays often managed through so-called online search intermediaries, such as Amazon or Expedia. Companies have therefore to understand that not only should they promote their product positively on their own website, but they have to appear in high positions at the rankings of those intermediaries, since they are highly relevant for consumer choices (Ursu, 2018).

The business model of online retailers stands and falls with the ability to build a solid and loyal customer base (Rafiq et al., 2013); relationship satisfaction leads to trust. Relationship satisfaction is thereby the main driver of customer loyalty (Rafiq et al., 2013). The authors point out that e-grocery retailers have to focus their relationship investment efforts on relationship satisfaction and affective commitment in order to create strong bonds of loyalty. These results are confirmed by Kim et al. (2009), who illustrate how e-loyalty brings high rates of customer retention and is an effective tool to reduce costs for recruiting new customers, which leads to a long-term profitability of the online retailer with a sustainable business model. They point out how customers base their purchase decision not only on the online retailers’ website appearance and functionality, but especially also on previous reviews of the relevant retailer (Kim et al., 2009). It is clearly stated how e-satisfaction and e-trust eventually leads to e-loyalty.

Product availability information play therefore a crucial role in the customer decision process, this is also underlined in the study from Wagner, Calvo and Cui (2018), which shows how low availability messages significantly increase the sales of the products. Also, relevant for this study is the exploration of the option to order goods and pick them up in the store, as the Portuguese omni-channel retailer offers. It has been shown how this option can increase store sales through cross-selling (Gallino and Moreno, 2014). Another viable option for online retailers is to offer a special high-quality delivery, since product delivery quality influences the consumers shopping decisions (Cui et al., 2018). In this study, the clash between SF Express and Alibaba has been taken as a quantitative example of what happens

(13)

5

if the highest delivery quality is taken out as an option. There is strong evidence found that this removal reduced sales significantly despite sales dispersion, another strong hint for delivery quality as key success factor for online retailers.

2.2.Online Supermarkets / grocery stores

In the year 2000, the differences in purchasing behavior between online and offline supermarket customers has already been questioned by research and management. A study from Degeratu et al. (2000) proposes a general conceptual framework which articulates various factors that might affect the behavior of online and offline customers. The study is designed in a way that it compares consumer choices in Peapod (an online grocery store) and traditional supermarkets belonging to the same chain and operating in the same geographical area (Degeratu et al., 2000). The study supports with its results the fear of many executives, that an increased online exposure leads to an increased focus on the price. Additionally, the results confirm that online promotions are better signs of price reductions, while offline promotions induce larger changes in brand choices. This has important implications for the way how online supermarkets and grocery stores place promotions and work with price discounts. Another interesting insight gained from this study is regarding the importance of brand names; the authors point out that while the brand name might set products apart especially for product categories which are differentiated on brand image, brand names remain less important for functional products, which are largely sold by online supermarkets and grocery stores (Degeratu et al., 2000).

Besides the previously discussed key metrics like perceived value, loyalty, e-satisfaction etc., existing research points out the importance of the metric “fulfilment” for online supermarkets and grocery stores (Marimon et al., 2009). Fulfilment does not only mediate through “perceived value” but the study confirmed the important impact on loyalty and purchasing of this dimension. This is reinforced by previous studies (e.g. Parasuraman et al., 1988) who restated that a company is unable to compete unless it is reliable. Most of the complaints of the customer of the Spanish supermarket of the study of Marimon et al. were related to the dimension reliability/ fulfilment, which demonstrates that this aspect is of special importance for online supermarkets.

(14)

6

Holloway and Beatty explored the consequences of service failure and potential recovery opportunities in more depth (2003). The authors addressed the issue of various types of service failures and described the recovery methods of the companies and their respective success. It is argued that effective complaint management is one of the most valuable tools for building and maintaining customer relationships even in the case of a poor picking success. Firms might commit errors; they should however focus on recovery management as means of increasing customer satisfaction and loyalty (Holloway, Beatty, 2003). By analyzing a large sample of online shoppers, the authors found out that the largest number of service failures were failures regarding the picking success. This includes that a product was either not delivered at all, the wrong product was delivered, or the product was delivered to the wrong address.

3. Institutional setup

To obtain the relevant data for this study, it has been agreed on a partnership with one of the leading retail chains from Portugal, which is spread all over continental Portugal, as well as on Madeira and the Azores. The product portfolio consists of primarily groceries; however, general merchandise is on the rise (products such as Electronics, Textiles etc.). The physical facilities of the Portuguese omni-channel retailer are organized via a dense offline network of regular supermarkets, convenience stores and smaller supermarkets in urban locations and hypermarkets.

For a couple of years, the omni-channel retailer also offers products online. While grocery is the primary market, the retailer also offers goods from other categories, such as electronics, textiles or merchandise. An important fact to mention is that the supermarkets serve online and offline customers at the same time. Customers might choose to receive the purchased goods at their domestic address or ready to pick up at the local store.1

1 There exists a delivery pass; customers can pay a fixed amount to ensure free delivery for a predefined time

(15)

7

Figure 1: Customer journey

Customer journey: The customer might choose to purchase his goods off- or online.

a) In the offline-case, the customer goes to the physical store and checks if his desired items are available.

- In case they are not, the customer might decide to either substitute the missing items himself, or not to purchase the items at all.

b) In the online-case, customers place their order, and enable the supermarket to substitute (or not).

- All the following steps are carried out by the supermarket, the customer can’t further influence his order.

Important to understand is that offline as well as online customers are served by the same stores (there does not exist a central warehouse which might be responsible for online customers only).

(16)

8

4. Hypothesis development

In this section the expectations regarding the consequences of a poor picking success are defined and explained. Also, the potential influence of other factors is discussed and justified by existing research.

4.1.The impact of a low picking success on the customer behavior

During the course of this study, we have learned how high-quality delivery leads to increased sales in online retailing (Cui et al., 2018). This applies not only to the rapidity of online delivery, but especially to the reliability and fulfilment dimension in online retailing (Marimon et al., 2009). As a consequence, one should expect decreasing sales from those customers who received a delivery with a low picking success for the first time.

HYPOTHESIS 1 (Total Purchases). After receiving delivery with a low picking success for

the first time, the customers will order less in total than those customers who received a delivery with a high picking success.

An interesting follow-up discussion is whether the decrease should also refer to the number of items purchased per order. With the emergence of online retailing consumers gained more sovereignty about the purchasing process (D. Grewal et al., 2004). This means that the consumers are able to personalize and customize their orders according to their personal preferences. The variety of information freely available in the web, like product and brand information, peer reviews, price information etc., provides the customer with an enhanced control over the shopping process (Tedeschi, 1999).

This might lead in the concrete case of the analyzed omni-channel retailer to the development that only certain items, which are not easily obtainable at similar stores with a better service quality, are purchased via the online channel of the supermarket. Customers might selectively eliminate items they don’t need on a daily basis, as those are crucially to be delivered on time. Also, if the price is similar to other online supermarkets, the picking success might be the only decision-relevant dimension, which is why customers might substitute items previously on the shopping list from this supermarket.

(17)

9

HYPOTHESIS 2. (Items per delivery) After receiving a delivery with a low picking success

for the first time, the customer will customize his item list in a way that he purchases less items per delivery as those who did not receive a delivery with a low picking success.

Going one step further, one must understand who exactly the potential and existing customers of online grocery stores and supermarkets are. The paper “Consumer response to online grocery shopping” (Morganosky et al., 2000) asks a similar question for the US consumer market and finds out that the majority is younger than 55 years old and has an above average income of $70,000 or more. Those higher-income households, which face usually time and other resource constraints are willing to pay a higher price for the items and additionally a delivery fee in exchange for comfort and availability. The main demographic variable related to buying all groceries online was education (Morganosky et al., 2000), giving a strong hint towards the relevant customer segment.

Similar characteristics can be assumed for the Portuguese customer segment purchasing groceries online, since the delivery costs in some cases make up a significant part of the total shipment value. It can therefore be assumed, that the customers do not strongly focus on purchasing low-priced items, but rather emphasize quality and are willing to pay the price accordingly. However, if those high-quality items, which are usually available in smaller quantities than the house brand of the supermarket with a lower quality perception (Block et al., 2007) are unavailable, the customer might switch to lower-priced alternatives with a higher availability.

HYPOTHESIS 3. (Average price) After receiving a delivery with a low picking success for

the first time, the customer will reduce the average price of the items he purchases subsequently.

4.2. Moderating factors

Besides receiving a poor picking success, there might be other factors influencing the customers’ future purchasing behavior with the analyzed supermarket. Also, there might be factors preventing the customer from reacting negatively to a poor picking success which have to be explored.

(18)

10

Days to Delivery

One of the main moderating factors which could have an impact on the customer behavior might be the time to the delivery. It has to be mentioned thereby that the customer specifies the desired days to delivery; they are not randomly assigned by the supermarket itself. Therefore, it will be assumed subsequently that the store will try to obtain the good up until it needs to be delivered. Intuitively, one could suppose that the longer into the future the desired date of delivery, the more precise would the customer expect the delivery to be. This is confirmed by existing research; customers clearly showed their expectations regarding the days of delivery: “What’s the point of having an online service if you can’t get it delivered within at least 2 weeks! I feel like if something is placed on back order or discontinued, you should be notified immediately.” (Holloway, Beatty, 2003).

Performing poorly on the delivery might therefore be more harmful if the delivery is on a later date in the future. This could be partly related to the described expectation management of the customers, on the other hand customers might be left with more time to react to a delivery with a low picking success if this happens one or only few days after the order placement. This leads to the fourth hypothesis:

HYPOTHESIS 4. (Delivery time) The customers who ordered a significant amount of time

in advance and still receive a delivery with a poor picking success react more negatively than those who order only a short time before the desired delivery date.

Delivery costs

The delivery costs could significantly impact the perception of the quality of the picking success. The more the customers pay in delivery costs, the more precise should they be able to expect their delivery to be. However, the delivery costs certainly relate to a specific delivery slot which has to be accounted for, and they might also differ according to the desired days to delivery. Prior research especially confirms the impact of high relative delivery costs (to the shipment value) on the expected reliability of the delivery (Boyaci et al., 2009). Consumers paying high delivery prices place greater value on receiving a flawless and reliable delivery.

(19)

11

Obtaining a delivery with a poor picking success might therefore be harmful especially in cases where the customers paid high delivery costs. It will be interesting to determine, whether the absolute amount of delivery matters, the amount paid relative to the other customers or the amount relative to the delivery value. This might be a crucial moderating factor in trying to understand the direction and magnitude of the customer’s reaction to receiving a delivery with a poor picking success.

HYPOTHESIS 5. (Delivery fees) Customers who paid a large percentage of delivery fees

react especially negative to a delivery with a poor picking success.

Discounts

One interesting idea to explore is whether obtaining discounts could be responsible for preventing unfavorable customer reactions to a low service quality. Lee and Chen-Yu emphasized in 2018 that price discount affects play an important mediating role in the relationship between the given discount and the customers’ perception. While the pure focus on price discounts might lead to the perception of low-quality products, when the price discount affect served as a mediator, the negative feelings regarding the potential quality of the items could no longer be found (Lee, Chen-Yu, 2018).

Keiser described already in 1988 how a discount of a few percent could make a trivial difference in the commission. It is mentioned how the company should keep track of the issues requiring discussion, as is the poor service quality in the case of the Portuguese omni-channel retailer. When dealing with unsatisfied and demanding customers, the company has to take the initiative and offer constructive solution approaches to their customers without losing a reasonably strong bargaining position (Keiser, 1988). If customers who received large amounts of discounts show lower levels of negative reactions to a poor picking success, this might be an effective bargaining tool in order to compensate for the low delivery quality. HYPOTHESIS 6. (Discounts) After receiving a delivery with a low picking success for the

first time, the customers who received large amounts of discounts do not react as negatively (in terms of reduced items per purchase) as customers do who did not receive discounts.

(20)

12

5. Data

The data used in this study entails purchase information from customers who bought products online from a large Portuguese omni-channel retailer. The data source contains 14,283,175 individual product purchases, reaching from the 1st of October 2016 to the 29th of September

2017.

5.1. Available data

The data base contains purchase information broken down on the individual product level per order line. It contains for each purchased product information about the shipment number, such as the SKU (stock keeping unit, identification for the supermarket), the requested quantity, the discount that could be applied to the item and the unit price, customer information such as the geographical location of the customer, and delivery information such as the delivery date, the store from which the delivery takes place, the order date and a specific time stamp per item. The most interesting variable however is the picking status; 1 means thereby partial picking (less than 100% and more than 0%), 2 means stock rupture and therefore that the ordered item has not been delivered, 3 means total picking and therefore that the item has been delivered successfully and 5 means substitution and therefore that the ordered item has been replaced by a similar product that was available.

The data source therefore contains important metrics which might reveal interesting insights into the customer behavior; customers ordering in different delivery slots, living in different locations and ordering different basket sizes with different prices can be clearly distinguished. The average basket size comprises around 29 items with an average basket price of a little more than 100€. The number of matched customers equals to 1,684 (details on the matching can be found in later sections). On average, around 8% of the delivered items are so-called fresh items and therefore perishable. The total basket value is almost equal to 100€, while on average 23€ of discounts were received per delivery.

(21)

13

Table 1: Summary statistics

Figure 2: Picking success on NSKU

Descriptive Statistics

Variable N Mean Min Max

Slot cost (€) 1,684 5.14 0 15 Delivery Cost (€) 1,684 6.02 0 15 Days to delivery 1,684 1.78 0 30 Fresh items (%) 1,684 8.11 0 36.84 Total Price (€) 1,684 99.67 0 965.04 Discounts (€) 1,684 23.07 0 508.80

NSKU per delivery 1,684 24.86 2 111 Average Price (€) 1,684 3.71 0.8 104.99

Week (unit) 1,684 24.78 1 53

ZIP Code (2 digits) 1,684 29.73 12 94

Home delivery (%) 1,684 0.91 0 1

Store delivery (%) 1,684 0.09 0 1

(22)

14

The number of purchased items increases with a higher picking success – a relationship which is to be explored during this study.

The most important adjustment which was made during the given data analysis was done by retrieving the picking status on the shipment level.

Picking Status: As the picking status has been available on product level and not on shipment level, it has been analyzed how a picking status per shipment can be implemented. To do so, four variables are created within the aggregated file, counting the respective picking status in columns for the individual product shipments. This can then be aggregated on a shipment level, displaying per shipment, how many items were successfully delivered, partially delivered, not at all delivered or substituted. From these numbers, a new variable, the Picking status of the respective shipment, is created. This is done as follows:

▪ For the items which are fully delivered, the Picking status is set equal 100%. ▪ For the items with partial delivery, an equal distribution between a Picking success

>0% and <100% is assumed, meaning that partially delivered items could have a Picking success of e.g. 27% or 81% with the same probability. This leads to an average Picking status of 50%, therefore the Picking status is set accordingly. ▪ Substitution is the most difficult subject of analysis, it is very hard to say whether

customers can be satisfied by substitution. For this reason, the simplifying assumption is made that the customers can only be either fully satisfied or dissatisfied with the substitution, with an equal probability, leading to an average Picking status of 50% ▪ Stock rupture corresponds clearly to a Picking status of 0%

Example: If a shipment contains 6 items with a full delivery, 3 items with a partial delivery, 1 substituted item and 2 items that were not delivered, the Picking Success would be as follows:

𝑥 =6𝑥1 + 3𝑥0.5 + 1𝑥0.5 + 2𝑥0

(23)

15

6. Identification strategy

The main objective of this study is to determine, whether and how the picking success influences the customer behavior and which measures supermarkets can take to prevent negative reactions or to react to them.

6.1.Matching Strategy

The model is to be built in a way that it clearly examines how the picking success influences the future purchasing behavior of the customer.

Various econometric challenges become evident when trying to evaluate the treatment effect in greater depth. First of all, people with a high picking success might be structurally different from people with a low picking success, based on observable characteristics, such as the composition of the basket. Also, the number of products purchased might be far off through the differences in purchasing power. Secondly, a selection bias might occur due to unobservable characteristics which hinder the comparability of the treatment and the control group.

In order to resolve those problems, an individual-level risk-set matching without repetition (Rosenbaum 2010, Chapter 12) is applied, to find an ideal counterpart for each treated customer within the control group. This matching strategy assures comparability in observable characteristics; unobservable characteristics are therefore expected to be comparable as well. The matching procedure therefore creates pairs of very similar customers; addressing the concern that the treated and control population are fundamentally different. Important thereby is that the customers from the treatment and the control group are similar before the treatment takes place (see Table 2). Therefore, the first order of a customer has to be analyzed primarily before the consequences are depicted in a second and further orders.

The analysis is implemented on matched pairs of customers, since a simple regression would not fully avoid biases such as omitted variable biases. The treatment takes place whenever the customer orders online and receives his order2. Customers who received less or equal to

2 Binary approach without matching would lead to a similar direction; however, the causality question would

(24)

16

85% of their purchase and have therefore a picking success of 85% or less are the treatment group, whereas customers who received more than 85% of their purchase are the control group3. Approximately 10% of the customers received 85% or less of their purchase; the

control group makes up 90% of the available data. Therefore, it should be possible to find for each customer of the treatment group an ideal counterpart in the control group to implement the matching successfully.

Various dimensions of variables could be chosen to implement the matching; static and dynamic customers characteristics as well as state variables that might change over time. The most relevant ones to implement the desired matching procedure are the following:

▪ The basket size in items: Customers should be comparable in terms of number of purchased items

▪ The location of the customer: There might exist significant differences between the neighborhoods regarding the picking success

▪ Delivery method: Customers that order the delivery to their home or the store should be matched respectively

▪ The date of the first purchase: The total number of items purchased afterwards could be misinterpreted if the customers are not matched on time; therefore, customers are paired if they purchased in the same week

▪ Days to Delivery: Customers are matched on how many days in advance they place their order

▪ Order Period: As there might exist structural differences between the daytime of the order, customers are matched on whether they order in the morning, afternoon, evening or at night

For each of the variables (except the delivery location, the order period and the days to delivery), help variables need to be created to enable a concrete and precise matching:

(25)

17

▪ The basket size in items: The help variable “quantity” is created, cutting the Numbers of SKU’s (NSKU) into sequences of different basket sizes.

▪ The postal code: Since there are too many distinct postal codes and the matching would therefore be impossible, the help variable “location” is created, which takes the first two numbers of each postal code as approximation for the relevant area. ▪ The week of the first purchase: The help variable “week” is added to the dataset, to

match customers who purchased online for the first time during a comparable time horizon.

The matching occurs at the point of the first purchase of the customer at the Portuguese omni-channel retailer, whereby the same customer should not be used as a control for multiple customers of the treatment group (exact attribute matching).

As mentioned before, the 10% individuals of the treatment group, which is the poor picking success, are matched with the remaining 90% of the control group. The major drawback is that the picking success is a continuous variable which has been artificially transformed in a binary one with a given threshold. This issue will be discussed as well in section 8 and the robustness of the results will be verified with different thresholds.

For each of the customers in the treatment group, various customers of the control group are potential matching partners. To have one exact matched pair, customers from the control are randomly eliminated from the matching possibilities. Those matched pairs are then given a so-called subclass, to be perfectly identifiable. These subclasses are then added to the initial dataset; after doing that, the identified pairs can be compared across all various dimensions before and after the treatment. The matching successfully made customers from treatment and control group comparable ex-ante, which can be shown by the following analysis of the standard effect size (which measures the mean difference in relation to the respective standard deviations) which should be below 0.2:

(26)

18 Matching validation

N Treated mean Treated Sd Control mean Control Sd Standard mean

differences p-value

Week (unit) 1684 32.18 16.07 32.18 16.07 0.00 1.00

Zip code (2 digits) 1684 28.38 14.06 28.38 14.06 0.00 1.00

Days to delivery 1684 1.80 1.80 1.80 1.80 0.00 1.00

Delivery Cost (€) 1684 3.95 2.93 4.01 2.93 -0.02 0.55

SlotCost (€) 1684 4.51 2.73 4.55 2.73 -0.02 0.66

Discounts (€) 1684 18.32 12.02 19.98 15.98 -0.12 0.17 NSKU per delivery 1684 24.15 14.41 22.95 15.03 0.08 0.32

Table 2: Matching validation

As one can observe, the standard mean difference is for each variable well below 0.2 and the p-value never below 0.1, which shows that the sample is sufficiently balanced (observable characteristics of control and treated group are similar). The SMD is a useful alternative to evaluate the quality of the matching especially if the sample size is large. In a next step, it has to be understood how the matching output is to be used.

6.2. Using the matching output to implement causal analyses

By comparing the matched customers in the respective output dimensions, one observes that they are reasonably similar. The identified matching pairs are then retraced via their subclass in the original dataset, to take a look at the purchase behavior after the respective delivery with a high or low picking success. The outcomes are analyzed regarding the various previously defined hypotheses.

To verify the expected causal inference, the treatment effects are measured by using the below model specifications:

𝑌𝑝,𝑖 = α𝑇𝑅𝐸𝐴𝑇𝑝,𝑖 + λ𝑝+ ε𝑝,𝑖 (1)

(27)

19

In this equation, i denotes an index for a particular customer who might have received the treatment or not. p denotes the index for the matching pair (and does therefore not have a numerical understanding). λ𝑝 stands for the pair fixed effect and ε𝑝,𝑖 is the random noise. The dependent variable 𝑌𝑝,𝑖 {Total Purchases, NSKU, etc.) is thereby the object of interest for the analysis. Equation (1) is valid for the main hypotheses 1-3, while equation (2) considers moderating factors and interaction terms as well.

𝑇𝑅𝐸𝐴𝑇𝑝,𝑖 indicates whether the respective customer received a delivery with a low picking success at his first purchase or not and is therefore either equal to 0 or 1. 𝑀𝑂𝐷𝐹𝐴𝐶 is a variable for different independent possible moderating factors, such as the amount of discounts that the customer received. This variable is expressed in a binary format, to enable a meaningful interpretation. The most interesting variable is probably the interaction term between the treatment variable and the respective moderating factor γ𝑇𝑅𝐸𝐴𝑇𝑝,𝑖 x 𝑀𝑂𝐷𝐹𝐴𝐶.

This variable enables to spot differences in the behavior of customers who received a good or poor service and respectively react differently with regard to an influencing moderating factor.

7. Estimation results

The estimation results from the previously established hypotheses are displayed below. The treatment effect itself is closely examined, as well as the previously described moderating factors.

7.1.The impact of a low picking success on the customer behavior

Main results. The previously expected impact of the low picking success on the customer behavior has been evaluated by analyzing the impact of the picking success in the first shipment on the total purchase number (HYPOTHESIS 1. (Total Purchases)). The idea was thereby that the customers reduce through a lower commitment to the delivering company the total transaction value with the company, in terms of total number of purchases as well as in terms of total exposure to the company.

(28)

20

The testing of the hypothesis shows clearly a relationship between the picking success in the first order and the total number of purchases afterwards (Table 3). Customers who received a delivery with a high picking success (>85%) during the first order purchased on average more than 41 items more in total than customers who did not receive an acceptable delivery. This result is highly significant (p<0.05) and also a large part of the variation (R₂ = 67.9%). It can therefore be stated, that the delivery quality of the first delivery strongly impacts on the total amount of items purchased afterwards4.

The next step is to understand, whether this result also applies to the items per delivery (HYPOTHESIS 2. (Items per delivery)). The data confirms this hypothesis (Table 3), customers who received a high service quality at the first delivery purchase 5.514 items more per purchase, which is a remarkable number considering the average basket size of around 25 items (more than 20%!). The result is highly significant (p<0.01) and 26.9% of the adjusted mean variation can be explained by the first delivery.

Lastly, it is to be analyzed how the composition of the basket changes, especially regarding low- and high-priced items (HYPOTHESIS 3. (Average price)) It has been discussed how an unsuccessful delivery might impact on the purchasing behavior of the wealthier customers involved in purchasing groceries online, who might substitute their favorite high-quality items with lower-quality items with a higher availability.

To evaluate the impact of the picking success on the development of the average price, the available data was manipulated in order to create a variable showing the difference in average prices between the first purchase and the further purchases. This variable (called relative average item price = average price of first purchase – average price further purchases) subtracts the average price of all further orders from the average price of the first order; a negative relative average item price would therefore show that the customers purchase higher priced items with the curse of time.

The results indeed confirm the initial hypothesis, the relative average item price is highly significantly negative for the customers who received a delivery with a high picking success (Table 3), meaning that for those customers the average price of the first delivery was by

(29)

21

over 1.80€ lower than the average price of all further deliveries. The direction of this is not surprising, however the magnitude is significant when comparing with the mean of all average prices of 3.71€. Also, a large part of the variance in the relative average item price is explained by the service quality of the first delivery (>65%). The number of observations still correspond to the number of customers thereby.

Table 3: Main results for customer behavior

7.2.Moderating factors Days to Delivery

It has to be determined whether the differences in days to delivery impact on the direction and magnitude of customer behavior after receiving a delivery with a low picking success. The backbone of this analysis is as stated previously the assumption, that customers expect a better delivery the further in the future the desired delivery date (even though they can freely choose their delivery date).

The next step would be to test hypothesis 4 (HYPOTHESIS 4. (Delivery time)). For this purpose, a generalized approach is suitable, illustrating the results of a poor service for customer with a short delivery horizon a poor service for customers with a long delivery horizon on the items purchased in the future.

Hypotheses 1-3 Dependent variable Total Purchases NSKU Relative average item price TREAT 41.126** (19.182) 5.514(0.788) *** -1.822*** (0.287) Observations 1,684 1,684 1,684 R2 0.152 0.766 0.651 Residual Std. Error (df = 539) 315.196 (df = 539) 12.950 (df = 539) 4.724 Note: *p<0.1; **p<0.05; ***p<0.01

(30)

22

Table 4: Days to delivery

The results clearly confirm Hypothesis 4; not only do the customers order less after receiving a delivery with a low picking success, customers order more than twice less afterwards if they ordered a long time in advance for the first delivery. They tend to order more than 10 items less if they ordered a long time in advance for the first delivery and still had a low picking success, a result which is significant on a level of p<0.01. This analysis of the impact of the days to delivery in combination with differences in the service quality level of the first delivery offer crucial insights for the supermarket. The service for those who order a relatively long time in advance has to be improved not only to approach the service level for urgent deliveries, but especially to reduce the negative reactions of those who ordered in advance and received a poor first delivery.

Delivery costs

In order to obtain a full picture of the relationship between the customer reactions to deliveries with a poor picking success and the remaining days to delivery, one has to include the delivery costs into the equation. Those costs might be related to specific slots and days in advance, however, one has to recognize also the relationship of these costs to the expectations generated in the customers ordering the items. The approach will therefore be two-folded to test Hypothesis 5 (HYPOTHESIS 5. (Delivery fees)). The validity of this hypothesis will be analyzed with regard to the total amount of delivery costs, as well as regarding the amount of delivery costs relative to the total basket value.

Hypothesis 4 Dependent variable NSKU Treat 4.065(1.147) *** DDHigh (2.446) 1.126 Treat : DDHigh -10.028(3.177) *** Observations 1,684 R2 0.771 Adjusted R2 0.213 Residual Std. Error 13.826 (df = 451) Note: *p<0.1; **p<0.05; ***p<0.01

(31)

23

The first observation that has been made was that the number of products per shipment is significantly higher the higher the delivery costs, which shows that either the supermarket charges more for a larger delivery volume or that customers themselves are willing to pay higher delivery costs for more items (Table 5). To simplify the analysis, the delivery costs are divided into high and low (above and below the mean delivery costs respectively). Then, the joint impact of a delivery with a poor picking success and high delivery costs are analyzed via an interaction term.

The variables Treat (treatment) and DCostHigh (high delivery costs) have the expected significant impact on the customer behavior of purchasing less and more items per shipment respectively. The direction of the interaction term also indicates, that the customers who received a delivery with a poor picking success and paid high delivery fees are ordering less items per delivery subsequently, this result is however not significant and can therefore not be statistically confirmed.

The next step is to take a look at the delivery costs relative to the total basket value. For this purpose, the variable “RDCostsHigh” (relative delivery costs) is created. In a similar approach, the interaction term is analyzed to identify the impact of the relative delivery costs in combination with a delivery with the picking success on the number of items per purchase:

Table 5: Delivery costs

Hypothesis 5 Dependent variable

NSKU NSKU TREAT 4.267(1.981) *** 6.442(1.420) *** DCostHigh (1)/ RDCostsHigh (2) 6.279*** (2.018) (17.011) -17.136 Interaction term (2.616)3.270 (19.445)0.047 Observations 1,684 1,684 R2 0.767 0.766 Residual Std. Error (df = 451) 13.942 (df = 451) 14.064 Note: *p<0.1; **p<0.05; ***p<0.01

(32)

24

Similar to the absolute delivery costs, the relative delivery costs are not the crucial criterion for customer who received a poor first picking success to reduce their number of items per purchase. The coefficient is though again negative but is far from being statistically significant.

Even though there are no significant impacts stemming from the delivery costs (absolute and relative to the basket value) on the reactions to a delivery with a poor picking success, this result does have important implications for the bigger context of this study. It reinforces the importance of the moderating factor days to delivery independently of the delivering costs. The delivery costs are not the crucial mediating factor which causes the customer to purchase less items per delivery after receiving a delivery with a poor picking success; the days to delivery have a significantly higher explanatory power. This means that customers react more negatively in a case where they chose to receive the delivery at a later date and the picking success is still low than when paying high delivery fees and receiving a delivery with a low picking success.

Discounts

In this section, it has been determined whether the reception of discounts could prevent negative customer reactions to a poor picking success (HYPOTHESIS 6. (Discounts)).

Table 6: Discounts

Hypothesis 6 Dependent variable NSKU Treat 7.445(1.197) *** DiscountHigh (1.570) 1.610 Treat : DiscountHigh -4.980(2.201) ** Observations 1,684 R2 0.773 Adjusted R2 0.290 Residual Std. Error 12.762 (df = 537) Note: *p<0.1; **p<0.05; ***p<0.01

(33)

25

First of all, the impact on the number of items per order was analyzed. To verify the established assumptions about discounts, it has been taken a look at customers that received a poor/ good service and a high/low discount. To implement this comparison, the dummy variable “DiscountHigh” was introduced. This variable indicates whether the customer obtained an above average discount or not. In the next step, the relationship of the picking success and the amount of discounts on the number of items per order has been analyzed. The crucial factor of this analysis was thereby the interaction term between “Treat” and “DiscountHigh”, as this interaction term indicates the reaction of customers who received a delivery with a poor picking success AND a high discount.

The low delivery success has a strong and significant impact on the items per delivery, whereas high discounts per sé do not have an influence. The important interaction term between “Treat” and “DiscountHigh” does however have a positive and significant sign, meaning that customers who were affected by a delivery with a low picking success quality, but were compensated by receiving high discounts did buy on average almost 5 items more per further purchase than customers who were affected by a poor service quality but were not compensated by receiving high discounts. This result is significant on a level of p<5% and has important implications for the alternative courses of action of the supermarket.

The adjusted R₂ is equal the almost 30%, meaning that the delivery quality and the discounts are able to explain almost 30% of the variation in the basket size after the first delivery. An interesting next step is subsequently to explore more on the magnitude of the effect; does the discount only help to satisfy disappointed customers in comparison to other disappointed ones, or goes the effect even so far as undoing completely the negative reactions of customers on a poor first delivery?

To answer this question, a new subset of data was created, only consisting of the customers who received an above average amount of discounts. For this subset, it was analyzed whether the picking success plays a role regarding the impact on the basket size of further purchases. A delivery with a low picking success has in this subset still a negative impact on the basket size, the impact is however smaller (only 3.3 items less) and also less signficant than in the previously analyzed case for customers with high and low discounts (Table 7).

(34)

26

To obtain a full picture of the impact, one should also consider the possibility, that the discounts do not fully compensate for a poor picking success at the first delivery regarding the basket size of the further purchases, but they might be compensating in terms of the frequency of further purchases. Indeed we can confirm this idea, by analyzing the impact of the delivery with a low picking success on the total further purchases, the coefficient is still negative, however not significant anymore and therefore not necessarily different from 0. This means that customers who received a poor picking success and a high amount of discounts do not necessarily purchase less items in total than customers who received a good picking success and a high amount of discounts, which is (depending on the key metric used) a very useful insight for the supermarket. The adjusted R₂ is also very low in this case, meaning that when receiving a large amount of discounts, the variation in Total Purchases can not be very well explained by receiving an acceptable first delivery or not.

Table 7: Main results for high discount customers

Independently if the discounts can completely offset the negative reactions from a poor first delivery, one has to recognize the impressive impact the discounts have in compensating the customers for receiving a delivery with a low picking success.

8. Robustness checks and limitations

In this section, the overall validity of the above results is discussed and verified. The central limitation and an important area of discussion is the picking success, what exactly makes the

Hypothesis 6 Dependent Variable

NSKU Total Purchases TREAT (1.880) 3.328* (48.220) 49.208

Observations 687 687

R2 0.907 0.836

Residual Std. Error (df = 95) 13.027 (df = 95) 334.077 Note: *p<0.1; **p<0.05; ***p<0.01

(35)

27

picking “successful”? The two core issues that can be derived from this difficult question in the concrete case of the Portuguese omni-channel retailer are substitution and the “artificial” thresholds that are defined to declare a picking as successful or not. In the following, both are discussed in more detail and alternatives are given to the stated assumptions in this study to explore on the robustness of the depicted results. More precisely, different thresholds are tested and different numerical values for substitution are suggested.

8.1. Substitution

When the “picking success” variable was described in section 5 (Data), a set of assumptions was taken to quantify how high or low the picking success of a delivery is. This has been based on the “Picking Status”. This variable might take on the four different values of 1 (Partial Picking), 2 (Stock Rupture), 3 (Total Picking) or 5 (Substitution). While the total picking unarguably represents a picking success of 100% and the stock rupture without doubt a picking success of 0%, partial picking and substitution are clearly more difficult to quantify. As there are no more information from the supermarket regarding the partial picking, the logical assumption of equal distribution was taken, meaning that the partial picking success is distributed equally >0% and <100%. If a customer ordered 10 bananas for example, it might be equally likely that he receives 2 or 9 of them, leading to a picking success of 20% or 90% respectively. It is an important limitation of this study, that there is no further quantitative elaboration on the partial picking success, the closest approximation to the actual mean of the partial picking is however the equal distribution as done. Further studies might want to investigate the robustness of the presented results after obtaining further information on the partial picking from the supermarket.

In general, this process also applies to the interpretation of the substitution. During the course of this study the assumption has been taken that 50% of the customers are satisfied if their ordered goods are replaced by similar ones if unavailable. A similar approach of estimating the customers satisfaction and the respective picking success was chosen. However, this assumption might be more off-track than regarding the partial picking. The first reason therefore is that customers might have the option to choose whether or not they allow for substitution. If the information on that would be available on a shipment level, a different quantitative approach for substitution might be chosen (i.e. that all customers who previously

(36)

28

allowed for substitution might perceive the picking success as higher as those who did not wish for substitution). Also, customers might be more or less satisfied regarding the proximity of the desired good to the substituted good.

Various influencing factors might therefore lead to crucial differences in the quantitative picking success related to substitution. To prove the validity of the above displayed results, a two-folded approach is chosen. The robustness is tested for two extreme points. First, it will be assumed that customers consistently dislike substitution and strongly dislike receiving goods they did not primarily order. The associated picking success will therefore be equal to 0%. On the other extreme (especially if the substitution option exists) it might be the case that customers are highly satisfied with receiving meaningful substitutions, such that they are not left without groceries on a particular day. In this case, the associated picking success might be rather equal to 100%.

For both of the depicted extreme cases, robustness tests are performed to verify the validity of the most important results. The most crucial results with the most interesting managerial implications of this study are the following:

▪ Total Purchases: The total number of purchases of customers who received a delivery with a poor picking success is subsequently reduced

▪ Items per purchase: The number of items per purchase decreases after receiving a delivery with a poor picking success for the first time

▪ Average Price: After receiving a delivery with a poor picking success, customers change the composition of the basket in a way that reduces the average price per item ▪ Days to Delivery: Customers who received a delivery with a poor picking success after ordering several days in advance subsequently reduce the number of items per purchase more than those who ordered only shortly in advance

▪ Discounts: Customers who suffered from a delivery with a low picking success can be successfully compensated through discount, meaning that subsequently the number of items per purchase and the total number of purchases is reduced significantly less than in the case where the customers did not receive large discounts.

(37)

29

8.1.1. Substitution = 0%: Robust results

First, the robustness of the previously mentioned results is tested for substitution equalling a picking success of 0%. The matching procedures are thereby implemented analogously to the prior analysis. Since the average picking success however decreases by assigning a value of 0% to the substituted goods, the previously used threshold has to be adapted to allow for a comparable separation into treatment and control group as previously. As the overall picking success changes, the threshold is also modified accordingly.

The impact of a low picking success on the customer behavior

As one can see by the results of the appendix tables 1-3, the direction and rough magnitude of the results does not change. A delivery with a low picking success still leads to a change in the basket composition, such that the average price for those affected by it decreases subsequently (Table 1, Appendix). Also, the average number of items per purchase (Table 2, Appendix) as well as the total number of purchases (Table 3, Appendix) decrease after receiving a delivery with a poor picking success. The results do therefore not change, even if it is assumed that the customers do not value substitution in any kind.

Moderating factors

Besides the immediate impact on the customer behavior, one has to also understand whether the impact on the moderating factors changes. Also, for the moderating factors it seems like the results do not change by altering the picking success value of substitution. Customers still purchase significantly more items per delivery when receiving a high amount of discounts (Table 4, Appendix), and customers also are equally disappointed in the case of ordering a longer time into the future and receiving a delivery with a poor picking success as stated before (Table 5, Appendix). The robustness of the central results from before is therefore ensured in the case of customers not valuing the substitution as assumed previously.

Imagem

Figure 1: Customer journey
Table 3: Main results for customer behavior  7.2.Moderating factors
Table 7: Main results for high discount customers
Table 1      Table 2
+2

Referências

Documentos relacionados

Fonte: CPS/FGV a partir dos microdados do Censo Demográfico 2000 www.fgv.br/cps Nota: A variável omitida é religião Católica * Controlada pelas variáveis: sexo, cor ou raça,

Finally, in the framework of the present dissertation, another statistical analysis was performed in order to identify which could be the minimum HQLA stock amount that the

 Managers involved residents in the process of creating the new image of the city of Porto: It is clear that the participation of a resident designer in Porto gave a

O objetivo do relato é descrever atividade desenvolvida com os acadêmicos da 1ª fase do curso de Psicologia, a fim de ajudá-los com a apropriação de

Ousasse apontar algumas hipóteses para a solução desse problema público a partir do exposto dos autores usados como base para fundamentação teórica, da análise dos dados

The probability of attending school four our group of interest in this region increased by 6.5 percentage points after the expansion of the Bolsa Família program in 2007 and

O soro dos animais vacinados com lipossomo, EBS e proteolipossomos foram coletados semanalmente antes e após a infecção experimental para a detecção da produção de anticorpos IgG,

A planificação da implementação da metodologia Lean recorrendo à formação dos pares, ao envolvimento dos profissionais e o apoio da administração, o rigor científico na