16

RESEARCH PAPER

Determining profit-optimizing return policies – a two-step approach on data from taobao.com

Wenyan Zhou1 & Oliver Hinz1

# Institute of Information Management, University of St. Gallen 2015

Abstract Selecting an optimal return policy requires taking into account two effects: the potential positive effect on sales and the potential negative effect of higher costs. We propose a two-step model, in which we first utilize a robust regression to explain purchase behavior, and then apply a zero-inflated neg- ative binominal regression to model the return behavior. We apply this model to data from the most important online plat- form in China and obtain three main findings. First, the adop- tion of return policies results in increased sales, while reputa- tion works as a moderator in this process. Second, good rep- utation and traditional customer friendly return policies (like the Seven-Day Return policy) can significantly increase the number of returns, while more guarantee credibility (enhanced by Guarantee Money) is related to fewer returns. Taken to- gether, both the Seven-Day Return policy (profit increase of + 0.29 %) and Guarantee Money (profit increase of +0.016 % per Yuan guarantee) ultimately increase firms’ profit.

Keywords Return policy . Return behavior . Zero-inflated negative binominal regression . Seven-Day Return policy .

Guarantee credibility

JEL classification M3

Introduction

High return rates are a global problem for online retailers, threatening their business model in the long run. According to the Wall Street Journal, a third of all Internet transactions are returned by shoppers and the return rate is still increasing (Banjo 2013). High return rates cause substantial costs (re- verse logistics cost, product depreciation, management of re- turn process and so on) (Blanchard 2005). ASOS Chief Exec- utive Nick Robertson stated that a 1% decrease in return rates would immediately increase profits by 10 million pounds ($16 million, approximately 30 % of their net income in 2012) (Thomason 2013). In the U.S., product returns cost manufac- turers and retailers approximately $100 billion annually due to lost sales and reverse logistics, reducing profits by 3.8 % on average per retailer or manufacturer (Blanchard 2007).

We analyse the effect of different return policies that small and medium-sized online shops can implement. The situation is especially challenging for them because they typically have only limited liquidity and relatively high labour costs which make them more fragile and high return rates constitute a significant challenge for them. High uncertainty in online environments (compared to traditional brick-and-mortar shops) and intense competition force them to offer attractive return policies to attract customers. For these reasons, it is crucial for small and medium- sized online shops to evaluate whether their return policies are beneficial and which policies are better with respect to profits. As our analyses rest on data provided by Taobao.com, the results can directly be applied by retailers that are active on Taobao.com. Moreover the results may also hold for small and medium-sized online shops that face a fierce competition.

The online purchase process in general can be modeled as two separate decisions (Bechwati and Siegal 2005; Wood 2001): customers’ decision to order and, upon receipt, their decision to keep or return. The effects of return policy can

Responsible Editor: Andreja Pucihar

* Wenyan Zhou wenyan.zhou@stud.tu-darmstadt.de

Oliver Hinz hinz@wi.tu-darmstadt.de

1 Institute of information management, TU Darmstadt/Electronic Markets, Hochschulstraße 1, 64289 Darmstadt, Germany

DOI 10.1007/s12525-015-0198-6 Electron Markets (2016) 26:1 3–1 40 1

Received: 25 July 2014 /Accepted: 6 August 2015 /Published online: 16 September 2015

persist from the pre-purchase to the post-purchase phase (Wood 2001). In the first phase, customers cannot experience and as- sess the actual quality of the ordered products, which constitutes an information asymmetry. Return policies aim at countering this asymmetry and have given online retailing a huge boost (Banjo 2013). Tolerant return policies act as a signal that in- duces customers to perceive higher quality and lower risk (Glover and Benbasat 2010) in a product. As a result, customers could spend less time on considering whether to buy or not, which ultimately eases their purchase decision.

In the second phase (post-purchase), return policies can in- crease customers’satisfaction and maintain long-lasting relation- ships (Pizzutti and Fernandes 2010), because they can easily get their money back if they are unsatisfied. Return behavior can be attributed to two causes: one is the gap between expectations and actual product quality (fit problem), and the other is opportunis- tic or planned behavior. The first one can be improved by offer- ing accurate descriptions as well as services like user-generated product evaluations. But the second cause depends more on customer personality and the return policies themselves. Prior research has shown that relatively restricted return policies – for example, charging Bhassle^ fees (Davis et al. 1998) and restocking fees (Shulman et al. 2009), or granting only condi- tional return guarantees (i.e., only solving verifiable problems) (Chu et al. 1998) – can effectively reduce the return rate.

Overall, tolerant return policies can generate respectable in- creases in sales, but a higher return rate could, in turn, lead to substantial costs in terms of reverse logistics, depreciation and additional labor effort. Therefore, both scholars and practitioners have tried to optimize return policies and find the balance be- tween increased sales and higher return rates using models or experiments. However, empirical studies are limited. The only related study at shop level (using shops’ operating data) is Davis et al. (1998), who investigated 133 retailers with various return policies and found that retailer return policies vary with how quickly a product is consumed, the salvage value (i.e., the re- tailer’s ability to resell the product or obtain credit from its suppliers) of returned merchandise, and whether there are op- portunities to cross-sell or substitute other items when returns occur. Other related studies explore the balance using models. For example, Anderson et al. (2009) established a model to measure the purchase and return of apparel items from the cus- tomer perspective and estimated the model with 987 customer records from a mail-order catalog company.

In this paper, we focus on the research question: How do various return policies (customer friendly return policies and guarantee credibility) affect customer purchase and return be- havior for small and medium-sized online shops? For this purpose we suggest a two-step approach: First we determine the impact of the return policy on sales, and second we exam- ine the influence of the return policy on return behavior. We collected data for 600 shops on the most important online marketplace in China, Taobao.com. The combined gross

merchandise volume of Taobao Marketplace and the affiliated Tmall.com exceeds 1 trillion yuan (~132 billion EUR) in 2013 (Alibaba 2013). The data offer the opportunity to analyze the effect of return policies as the sellers can choose from various return policies on Taobao.

Due to the characteristics of our data, we apply a robust regression model to deal with minor concerns about the po- tential failure to meet assumptions, such as normality, heteroscedasticity, or observations that exhibit large residuals, leverage, or influence. Further, we estimate a zero-inflated negative binomial regression (ZINB) to deal with excessive zeroes in the return rate. We also control for guarantee credi- bility (Suwelack et al. 2011) and reputation (Roggeveen et al. 2014) and their potential interaction effects.

Literature review on return policies

A service guarantee is a promise by a company to compensate the customer in some way if the defined level of service de- livered is not fully met (Sum et al. 2002); examples include money-back guarantees and lowest-price guarantees. Prior re- search (e.g., Su and Zhang (2009)) has suggested that return policies can be analyzed in three different dimensions: 1) re- turn deadlines; 2) consumer effort required (in terms of bring- ing back original receipts and filled-in return forms); and 3) extent of return coverage (extent of money back due to ship- ping charges, inventory holding charges and re-stocking fees).

During the purchase decision process, a tolerant return policy – which entails longer deadlines, less required effort for returning and more coverage – provides an effective signal that reduces uncertainty (Heiman et al. 2001) and heightens per- ceived quality (Boulding and Kirmani 1993; Moorthy and Srinivasan 1995). Signaling theory (Kirmani and Rao 2000) proposes an explanation for the conditions under which a guar- antee reliably signals quality and influences consumer choice (Boulding and Kirmani 1993; Erevelles et al. 2001). That is, when consumers are confronted with information asymmetry about true product quality, they try to assess the magnitude of the penalty faced by the seller when the product’s actual quality is lower than its promised quality (Boulding and Kirmani 1993). This penalty results from both direct return costs (amount of return, expenses) and indirect reputational implications. Further penalties involve the return rate, which is influenced by the perceived quality of the market mechanism used to detect mis- leading applications of the signal (e.g., ease of understanding, interest in comparison) (Heiman et al. 2001). Meanwhile, Desmet (2013) found that this relationship is moderated by brand, price, and the relationship between customer and retailer.

While previous research on return policy’s marketing sig- nals has mainly focused on quality signals, Suwelack et al. (2011) revealed that the credibility of a return policy (guaran- tee credibility) is a key mediator between the return policy and

W. Zhou, O. Hinz104

customer purchasing behavior. Specifically, the more credibil- ity customers feel the more purchase intention will they have. Guarantee credibility also can work as a trust signal (Mavlanova and Benbunan-Fich 2010) which can influence customer return behavior.

In the return decision process, a tolerant return policy how- ever also increases product return rates for customers in remote purchase environments (Wood 2001). Anderson et al. (2009) used utility theory to point out that customers return items only if the net utility (Net utility=Deterministic utility+Fit of trans- action+Return cost) is negative. Specifically, according to the two-step process, the utility of the product (both pre-purchase utility and post-purchase utility) for a specific customer is de- rived from three parts: (1) Deterministic utility, which is known to the firm and the customer at the time of purchase; (2) fit of transaction, which is unknown to the firm but known to the customer after the purchase (Petersen and Kumar 2009); and (3) return cost, which includes the shipping fee (Frischmann et al. 2012), time and labor cost. More lenient policies impose less return costs for the customer, thereby increasing return intention. But other scholars have treated the service guarantee (return policy) as a positive factor for return behavior in the long run, namely in offering a continuous positive effect on employees’ motivation and ability to learn from service failure, which can increase service quality and indirectly reduce cus- tomer intention to return indirectly (Dutta et al. 2007).

While there exists a vast array of literature on the optimal way to design return policies (e.g., Padmanabhan and Png (1997)), only a few papers (e.g., Che (1996)) have examined the impact of return policies on both the pre-purchase and the post-purchase steps (Janakiraman and Ordóñez 2012), and the empirical research on the second step is especially lacking. Table 1 shows the related empirical studies. We can conclude that 1) most empirical studies work on the customer level (using customers’ purchase history or conducting experi- ments), examining the behavior of a number of customers from one company or platform; 2) most of the studies apply simple OLS regressions when analyzing data; and 3) there is no study that focuses on the effects of guarantee certainty on return behavior. Our paper aims to close these gaps by ana- lyzing real data using a zero-inflated negative binominal re- gression, for the purpose of measuring the influence of return policies on both, sales and returns.

Hypotheses and conceptual model

To investigate the return policy’s effect on the whole purchase process, we propose a two-stage conceptual model. We ana- lyze the effect of return policies and guarantee credibility on both, the purchase and the return stage (See Fig. 1). We control for reputation as another important quality signal for online shopping (Zhang et al. 2012).

Customer friendly return policies work as quality signal and risk control measure when customers make their purchase decision. Because of high return costs, the adopted customer friendly return policy is a costly indicator for the seller’s con- fidence in their own product quality. Prospective buyers as- sume that sellers do not adopt policies that lead to losses. So the products’ quality is expected to be higher or closer to their description on the websites, which can lead to a higher fit of expectations and actual product quality. Moreover, customer friendly return policies can reduce the cost of retracting a bad decision and thus enables consumers upfront to purchase while maintaining some flexibility. If consumers have the choice between two equivalent alternatives, they are likely to choose the alternative with the better return policy.

The return policy also influences the after-purchase phase. Return policies without shipping and restocking fees directly decrease return costs and return policies with lenient return deadlines giving the buyers more time to reconsider the pur- chase. Customer friendly return policies can thus reduce abso- lute return costs (money, time and labor cost). As Anderson et al. (2009) pointed out that a customer will choose to return products if the net utility is negative and lower return costs ultimately lead to a lower net utility. We thus state hypothesis 1 and 2.

H1: Customer friendly return policies are positively re- lated with sales. H2: Customer friendly return policies are related with more returns.

We define guarantee credibility in our paper as the credi- bility of a return policy which can certainly influence the buyers’ two-stage decisions. Usually, credibility is difficult to improve or measure (Zhuo et al. 2013), but online shops can easily boost the credibility of return policies directly via support measures – for example, by adopting Guarantee Mon- ey (e.g., in Taobao.com), which ensures that the return policy can be enforced or sellers can be punished if they do not implement their return policies correctly. At Taobao.com sellers can allocate an arbitrary amount of money (Guarantee Money) that the platform operator can use to compensate un- satisfied buyers if there is dispute between buyer and seller. On the one hand, guarantee credibility is thus a strong signal for high quality and credibility of the return policy of the focal sellers. Thus, we suggest that guarantee credibility could also be a factor that directly affects customer purchasing behavior. For example, large firms could use guarantee policies that small firms cannot afford to imitate. If low-quality firms fol- low though, they will suffer significant financial losses when they are caught cheating. On the other hand, guarantee credi- bility can effectively propel return policies forward which makes it easier for buyers to return their products. Similar as return policy, guarantee credibility could significantly reduce

Determining profilt-optimizing return policies 105

the absolute value of return cost. So we derive hypotheses 3 and 4 as follows:

H3: Guarantee credibility is positively related with sales. H4: Guarantee credibility is related to more returns.

Reputation is another important quality signal (Zhang et al. 2012). Purohit and Srivastava (2001) provided a classification scheme to differentiate between transient and durable signals. Durable signals, such as reputation, are classified as Bhigh- scope^ signals, meaning that the cue has evolved over time and cannot be changed easily. More transient cues, such as re- turn policies and guarantee credibility, are classified as Blow- scope^ signals, meaning that they are fairly easy to change

and weigh relatively less as a signal of quality in contrast to a high-scope signal (Purohit and Srivastava 2001). Meanwhile, Roggeveen et al. (2014) found that moderately incongruent sig- nals can be combined to enhance evaluations. In particular, if a firm's reputation and policies are complementary, this can posi- tively moderate the effect of return policies on customer behav- ior. However, reputation only stands for the past records of a firm; it cannot guarantee the firm’s future behavior. Based on these insights we propose hypotheses 5 and 6 as follows.

H5: Reputation moderates the effect of return policy on sales. H6: Reputation moderates the effect of guarantee credi- bility on sales.

Data and model

Taobao.com belongs to the Alibaba group and is the largest global online B2C and C2C platform. The China based plat- form started their business in May 2003 and had 500 million registered users by the end of 2013 (Ye et al. 2013). Taobao is a bilateral market, and also can be categorized as a third-party market place (Timmers 1998). Like on eBay, registered users can act both, as sellers and buyers, on Taobao. The platform

H1

H7H6

H3

H2

H1Customer friendly

return policy

Guarantee credibility

Reputation

Sales

Returns

Fig. 1 Conceptual model

Table 1 Summary of prior empirical research focusing on customers’ product return behavior

Studies Method Data Main contributions

Hess and Mayhew (1997)

Hazard model and empirical evidence

1000 customers during 4 years Using hazard model to predict return behavior.

Davis et al. (1998) OLS regression 133 stores in the Sacramento, California area

Retailers prefer a low-hassle return policy when 1) its products benefit are only realized in long term; 2) there are opportunities for cross selling; and 3) the salvage value from returned merchandise are high.

Wood (2001) Conjoint experiments 68 undergraduate students in the main study

Lenient return policy increases purchase rates and product return rates via signal effect.

Bechwati and Siegal (2005)

Labor experiments 87 undergraduate students in Study 1; 117 undergraduate students in Study 2

Introduces a framework of the mechanisms underlying product returns. Customer return choices are different when they facing disconfirming information.

Anderson et al. (2009) Developing econometric model and empirical evidence

987 customers of a mail-order catalog company

Provide empirical evidence that return policy gives customers an option value that is measurable; add the option value to model how different return policies affect firm profits.

Petersen and Kumar (2009)

Seemingly unrelated regression, TobitModel and empirical evidence

Transactions information of 1572 customers between January 1998 and August 2004 in a B2C company

Empirically demonstrate the role of product returns in the exchange process and show how product return behavior affects future customer and firm behavior.

Bonifield et al. (2010) Regression; Field experiments

141 of the e-retailers listed on BizRate.com; 290 consumers at an e-tail site

Show a correlation between perceived quality of online retailers and product return policy leniency. The shopping experience moderates the relationship between leniency and consumer reaction.

Janakiraman and Ordóñez (2012)

Labor experiment 245 participants from a large public university in the southwestern U.S.

Decreasing the product return deadline has the counterintuitive effect of leading to an increase in product return rates in some cases.

W. Zhou, O. Hinz106

offers unified ordering, a delivery and payment system but sellers have flexibility to choose their own guarantee strategy. Another characteristic of Taobao.com is that most of sellers are small and medium-sized shops in China that do not sell well- known brands and only focus on one special category like cloth- ing and shoes. Although Taobao has started to offer its service in other countries as well, the main market is still China. According to the 3-months Alexa traffic rankings, Taobao.com is ranked 9th and ebay.com is ranked 21st worldwide (Retrieved on Dec. 8, 2014). The blossoming of the online market in China has stimulated much interest among marketers. However, there is still limited research that attempts to understand customer behavior on platforms like Taobao.com.

On Taobao, customers can leave reviews (e.g., comments: good, normal, negative; scores (1–5, 5=very good) for descrip- tion, service and delivery) after purchasing; the default value is Bgood^ if customers do not leave feedback within 15 days, but scores have no default values. All information can be found on the ratings pages, including customer reviews and each shop’s return rate. We randomly choose 600 online shops, but we could not retrieve all the necessary data for eight of them, leaving a total of 592 shops in our sample. The shops’ catego- ries are mainly clothing, shoes, bags, digital products, books and others (see Table 2). To control the effect of product types on customer behavior (Bae and Lee 2011), we use industry data (e.g., Industry Return Rate) as a control variable in our models.

The return rate is the proportion of returns in relation to sales in the observation period. We also calculated the return rate in 2014 and found that the changes are not substantial (see Table 2). The two main return policies adopted by sellers on the platform are BConsumer Guarantee^ and BSeven-Day Return^. BConsumer Guarantee^ means that customers can re- turn or change products within 15 days if there is a quality problem or if the product description does not match the re- ceived product. Moreover, BGuarantee Money^ includes a spe- cial form of a money back guarantee: Taobao.com administra- tors guarantee that the shops offering BConsumer Guarantee^ will return the customer’s funds immediately–a feature that makes the return process both smoother and more credible. The BSeven-Day Return^ policy means that customers can re- turn the product for any reason within 7 days. We use two dummy variables to model whether shops adopt these two re- turn policies or not. Table 3 explains all variables used in this paper.

To examine the tradeoff between the impact of return pol- icies on sales and the one on returns, we suggest two steps. In the first step, we apply a robust regression model to estimate the influence of return policies on sales. In the second step, we estimate a zero-inflated negative binomial regression (ZINB) to examine the impact of return policies on the return rate, a dependent variable with excessive zeroes. The robust regres- sion model is as follows:

Sale si ¼ α0 þ α1* Reputation þ α2* Seven Day Return þ α3* Guarantee Money þ α4* Customer Guarantee þ α5* Description þ α6* Industry Return Rate þ α7* Return Rate þ α8* Reputation * Seven Day Return þ α9*Reputation*Guarantee Money þ α10*Reputation*Customer Guarantee þ ε

ð1Þ

In the second step, we need to account for the fact that the return rate is heavily skewed (see Fig. 2), with the majority of the shops experiencing very few returns. About 45.5 % of the shops encounter no returns at all. This might be due to two reasons: 1) there are some new shops that have no transaction records at all; and 2) some shops have customers who are so satisfied that they do not make use of the return option. As a result, we cannot use a regular regression model, so we instead apply a zero-inflated negative binomial regression (ZINB). ZINB models have two stages: one stage is used to estimate the zero-values (non-returns in this case) and the other is used to estimate the actual returns in absolute numbers (Mwalili et al. 2008). All cases are used in both analyses, but they are weighted based on the results of the model’s logistic component (see Equation 6). Using this model, we can well explain the zero-inflation, so long as two conditions of ZINB are considered: 1) target behavior is not always happening (we have zeroes and non-zeroes in the dataset); and 2) target behavior has to be any integer, including zero (Heilbron 1994). To meet these conditions, we use the integer of return rate times 100 (we

labeled BReturn100^, e.g., if the return rate is 7.034 %, the BReturn100^ is 7.) and use return rate times sales (we labeled BReturn Number^) as two separated dependent variables.

While the data in this study are not true count data, the chosen model is appropriate for two reasons: first, the data are distrib- uted exclusively on the non-negative integers and tend to show heteroskedasticity (exactly like true count data); second, the data appear to be a result of mixture models (i.e., two separate parts cause the need for zero-inflation) (Simons et al. 2006). As such, even though the data are technically not generated by a count process, the resulting distribution has the expected characteris- tics of a count process, and thus a count model is appropriate.

The ZINB distribution is a mixture distribution assigning a mass of p to Bextra^ zeroes and a mass of (1− ω) to a negative binomial distribution, where 0 ≤ ω ≤ 1. McLachlan and Peel (2004) noted that the negative binomial distribution is a con- tinuous mixture of Poisson distributions, which allows the Poisson mean to be gamma distributed; in this way, over- dispersion is modeled. We will compare the results of the

Determining profilt-optimizing return policies 107

ZINB with the outcomes of a zero-inflated Poisson model (ZIP) later on to demonstrate why we do not use a ZIP model in our study. The negative binomial distribution is given by

P Y ¼ 0ð Þ ¼ ω ð2Þ

P YeNegative Binomial λ; αð Þð Þ ¼ 1−ω ð3Þ

yielding the following distribution of counts:

P 0ð Þ ¼ ω þ 1−ωð Þ*F 0jλð Þ ð4Þ

P kð Þ ¼ 1−ωð Þ*F kjλð Þ ð5Þ

where F represents the reference distribution (negative bino- mial with fixed parameter α), represents the predicted proba- bility of being always-zero, modeled by the logistic compo- nent of the model, and represents the predicted mean of the negative binomial component of the model.

Reputation, Seven-Day Return, Guarantee Money, Customer Guarantee, Description and Industry Return Rate were included as independent variables in both components of the models. Thus, the two-part model was parameterized s

ωi ¼ e

β0 þ β1*Reputation þ β2*SevenDayReturn þ β3*GratanteeMoney þ β4*CustomerGuarantee þβ5*Description þ β6*IndustryReturnrate

1 þ e β0 þ β1*Reputation þ β2*SevenDayReturn þ β3*GratanteeMoney þ β4*CustomerGuarantee

þβ5*Description þ β6*IndustryReturnRate ð6Þ

λi ¼ e γ0 þ γ1*Reputation þ γ2*Seven Day Return þ γ3*Gratantee Money þ γ4*Customer Guarantee

þ γ5*Description þ γ6*Industry Return Rate ð7Þ

Table 2 Descriptive statistics

Summary of dependent variables

Dependent variable Min Median Max Mean

Sales 0 21 8324 116.7

Return rate 0 % 1.5 % 150%a 7 %

Accumulation of return rate

Return rate 0 % 5 % 10 % 20 % 30 % 50 % Mean

Accumulation 45.5 % 67.8 % 84.5 % 91.3 % 95.7 % 97.7 % 7 %

Adoption of various return policies

Return policy Seven-Day Return Consumer Guarantee Guarantee Money

Ratio of adoption 47.7 % 88.7 % 76.6 % (Range: 0 – 10,000 Yuan)

Return rate in different industries in 2012 and 2014

Industry Home Accessory Health Food Jewel Accessory

Return rate in 2012 0 % 1.77 % 2.28 %

Return rate in 2014 1.62 % 2.72 % 3.03 %

Industry Digital Products Book and Media Clothes and Shoes

Return rate in 2012 2.31 % 5.31 % 5.53 %

Return rate in 2014 4.02 % 5.48 % 6.58 %

a Return behavior can happen within 7 days after purchasing, so there can be lagged effects. A return rate>100 % means that customers returned products that they have bought in the last 7 days of the previous month which can in this case exceed the sales in the focal month. For example, shop A sold 10 products in May, one of them was returned in May, three of them were returned in June. Then in the next month, Shop A only sold 2 products which the buyers kept but as a result the return rate in June is 150 %

W. Zhou, O. Hinz108

We then use the software package R to estimate both the robust regression and ZINB model.

Results

We start with a Vuong test and find that a zero-inflated nega- tive binomial (ZINB) regression model is superior over a zero- inflated Poisson model (ZIP) (t-value=19.16929, p<0.001) and a standard negative binomial model (t-value=4.945271, p<0.001). The estimated parameters and the p-values of the models are presented in Tables 5. It is clear that return policies can significantly affect both sales and returns, but various

policies function in different ways. Reputation is another im- portant factor working in both a direct and an indirect way.

Columns 2 and 3 in Table 4 indicate the main factors influencing sales, while columns 5 and 6 present a model that includes the interaction between the shops’ reputations and their return policies. According to the data, good reputation and the Seven-Day Return policy significantly increase sales (p<0.05) which supports H1. Other return policies like Guar- antee Money and Customer Guarantee also have positive ef- fects, but this result is not robust across Model 1 and 2. Hence, we only find partly support for H3. With regard to the mod- erator effects, better reputation reduces the influence of tradi- tional return policies (Seven-Day Return policy and Customer Guarantee) on sales (p<0.05), but bolsters the influence of guarantee credibility (Guarantee Money) on sales (p<0.05), thus supporting H5 and H6. The reason might be that tradi- tional return policies and reputation both work as quality sig- nals, thereby acting as substitutes. From a customer’s perspec- tive, sellers with a better reputation are likely to offer generous return policies, and it follows that sellers offering better return policies must be more confident about their services and prod- ucts. However, guarantee credibility extends the reliability of quality signals, thereby serving as a complement to the histor- ical reputation. Guarantee Money is specifically a support measure for future behavior, guaranteeing that buyers can get their money back immediately if necessary.

Table 5 presents the results derived from fitting the ZINB regression model to the return data. The R-square is relatively low because we cannot consider customer-specific factors like the customers’ personality traits.

Return number, Reputation and Guarantee Money remain significant in both parts of the ZINB model. Specifically, good reputation can reduce the possibility of certain non-returns

Table 3 Measurement of variables

Variables Measurement

Reputation Total amount of good comments since business’ start.

Seven-Day Return Customers can return or change their products for any reason within 7 days, coded as dummy variable.

Customer Guarantee Customers can return or change their products within 15 days if there are quality or service (e.g., description misunderstandings) problems, coded as dummy variable.

Guarantee Money Collected by platform operator in advance to pay money back to customers. Sellers can choose the amount of Guarantee Money they allocate to the platform operators.

Description Score Scores provided by customers after receiving products to describe whether the information offered by shoppers matches the real product(s); the range of the score is 1 (very bad) to 5 (very good).

Return Rate Number of returns divided by sales in the specific month.

Return Number Returns in absolute number.

Return100 Return rate * 100.

Sales Number of sales (operationalized by the number of good, normal and bad comments).

Industry Return Rate Average return rate in this shop’s main business’ industry.

0

50

100

150

200

250

300

0% 25% 50% 75% 100% 125% 150%

F r e q

u e n

c y

Return Rate in %

Fig. 2 Histogram of return rate in %

Determining profilt-optimizing return policies 109

Table 4 Impact of return policies on sales

Robust Regression

Dependent Variable Sales

Model 1 (Without interaction) Model 2 (With interaction)

Coef. P-value Coef. P-value

Intercept 23.018 0.271 α0 −5.624 0.872 Reputation 0.005*** 0.000 α1 0.041*** 0.000

Seven-Day Return 8.970*** 0.001 α2 16.757*** 0.000

Guarantee Money 0.019*** 0.000 α3 0.001 0.656

Customer Guarantee −0.582 0.887 α4 11.520** 0.005 Description −4.646 0.543 α5 0.225 0.974 Industry Return Rate 111.770 0.241 α6 135.383 0.126

Return Rate −13.347 ▪ 0.059 α7 −13.738* 0.035 Reputation* Seven-Day Return α8 −0.007*** 0.000 Reputation* Guarantee Money α9 0.000*** 0.000

Reputation* Customer Guarantee α10 −0.039*** 0.000 AIC 8,862.610 8,389.974

R-square 0.4398 0.7323

*** p<0.001, ** p<0.01, * p<0.05, ▪ p<0.1. Significant parts are marked in bold, n=592

Table 5 Results of zero-inflated negative binominal regression model

Zero-inflated Negative Binomial Regression

Dependent Variable Model 3: Return Number Model 4: Return100

Count model coefficients

Coef. P-value e ß or γ Coef. P-value e ß or γ

Intercept ß0 25.870e+01 *** 0.000 1.719e+11 7.004e+00 ▪ 0.086 1.101e+03

Reputation ß1 4.813e-05 *** 0.000 1.000e+00 −1.016e-05 0.852 9.999e-01

Seven-Day Return ß2 3.464e-01 0.110 1.414e+00 −4.989e-01* 0.005 6.072e-01 Guarantee Money ß3 −2.297e-04

▪ 0.074 9.998e-01 −3.060e-04* 0.011 9.997e-01 Customer Guarantee ß4 −1.249e-02 0.969 9.876e-01 6.307e-02 0.807 1.065e+00 Description ß5 −5.101e+00*** 0.000 6.092e-03 −8.347e-01 0.291 4.340e-01 Industry Return Rate ß6 4.885e+00 0.512 1.322e+02 −4.533e-01 0.879 3.130e-01

Zero-inflation model coefficients

Intercept γ0 7.484e+00 0.104 1.778e+03 −1.130e+00 0.747 3.231e-01 Reputation γ1 −7.038e-04*** 0.000 9.993e-01 −6.138e-04*** 0.000 9.994e-01 Seven-Day Return γ2 −4.643e-01 0.472 6.286e-01 −3.945e-01 0.198 6.740e-01 Guarantee Money γ3 −2.486e-03*** 0.000 9.975e-01 −1.649e-03*** 0.000 9.984e-01 Customer Guarantee γ4 −4.334e-01 0.520 6.483e-01 −1.193e-01 0.756 8.875e-01 Description γ5 −9.311e-01 0.327 3.941e-01 8.154e-01 0.256 2.260e+00 Industry Return Rate γ6 −4.620e+01** 0.008 8.614e-21 −3.635e+01*** 0.000 1.634e-16

AIC 2,475.544 2,931.722

R-square 0.0339 0.0142

*** p<0.001, ** p<0.01, * p<0.05, ▪ p<0.1. Significant parts are marked in bold, n=592

W. Zhou, O. Hinz110

(γ1=−7.038e-04, p<0.01) and increase the return number if shops already have return records (ß1=4.813e-05, p<0.01). The reason is that sellers with a better reputation (more past sales) are more likely to face an opportunist (e.g., purchase, use and return of clothing on purpose). And then due to the large number of sales, even with the same return rate, the absolute return number increases. So in both, the zero- inflation and count model, reputation – as we have operation- alized it – increases the return number. Meanwhile, it appears that a higher amount of Guarantee Money reduces the number of returns, likely because it speaks to the return policy’s cred- ibility. Specifically, if a shop were to increase its Guarantee Money by one Yuan, the expected return number in 1 month would decrease by a factor of exp(ß3)=0.998 (p<0.1) while holding all other variables constant. But at the same time, the odds that a shop has certain non-returns would decrease by a factor of exp (γ3)=0.997 (p<0.01). In other words, the higher the Guarantee Money, the less likely shops have certain non- returns. On the other hand, the Seven-Day Return policy only works in the negative binomial part. By adopting this return policy, shops’ expected return number is exp(ß2)=1.414 (p= 0.110) times the expected return number for other shops. These results definitely support H2 and H4.

To our surprise, the Industry Return Rate influences return behavior only in the zero-inflation model (γ6=4.620e+01, p= 0.008). In other words, whether shops have return records largely depends on industry characteristics, while specific re- turn numbers depend mainly on the shops’service or customer fit (as outlined before in our paper).

We also examine the return rate (Return100) (see Model 4 in Table 5, R-square=0.0142). Although the ZINB model for the return number (see Model 3 in Table 5, R-square=0.0339) fits the data better (according to AIC and R-square value), examining the return rate does yield some valuable informa- tion. Specifically, Reputation and Seven-Day Return policy stand out as the major differences: we find that the effect of reputation on returns is not robust. The second one shows that the adoption of the Seven-Day Return policy would increase the return number, but at the same time decrease the return rate (ß2=−4.989e-01, p<0.01). In other words, this return policy increases sales and returns concurrently, but the influence on sales is stronger than on returns, thereby creating the possibil- ity for higher profits.

Counterfactual simulation

Using a counterfactual simulation, we highlight the managerial insights that emerge from the results of our two-step analysis. Managers usually want to acquire better knowledge about how return policies affect both profit and return cost in order to make informed decisions. According to McKinsey’s report in 2013, Chinese e-retailers realize margins of 8–10 % of earnings

before interest, taxes, and amortization, which are slightly higher than the average margin for physical retailers (Dobbs et al. 2013). Meanwhile, return behavior can reduce profits by 3.8 % on average (Blanchard 2007). By integrating the results from model 1 in Table 4 and model 3 in Table 5, we can calculate the margin for return policies and their impact on profits. Adopting the Seven-Day Return policy, for instance, can produce an additional margin gain of +0.29 % compared to shops without this policy (see the Appendix for the detailed calculation). Meanwhile, adopting the Guarantee Money policy would increase profits. Increasing Guarantee Money by 1 Yuan would increase profits by +0.016 %. Although the effect is most likely not truly linear, we carefully conclude that an additional 100 Yuan of Guarantee Money could increase profits by + 1.6 %. As a result, we suggest that retailers on online platform like Taobao should choose effective return policies, both tradi- tional policies (e.g., Seven-Day Return policy) and polices that increase the guarantee credibility.

Conclusions

To help small and medium-sized online shop optimize their return policy, our paper proposes a holistic view on the deci- sion process and takes the double-sided influence (on both, sales and returns) of return policies into account. We find that both adopting customer-friendly return policies and increasing guarantee credibility can significantly increase profits for small and medium-sized online shops.

Although some effects seem obvious, especially those pertaining to the first purchasing phase, the existence of a dis- tinct second phase, i.e., the decision to return or keep a product, leads to more complex effects. The adoption of return policies and higher guarantee credibility results in increased sales, while reputation works as a moderator that decreases the influence of traditional return policies and increases the influence of guar- antee credibility on sales. In the second step, we applied a zero- inflated negative binominal regression model and found that reputation, guarantee credibility, and the average return rate in the particular industry contribute to whether shops have return records at all. For those shops with return records, good repu- tation and traditional return policies (like Seven-Day Return policy) can significantly increase the return number, while higher guarantee credibility and a better product description could reduce it. It is also interesting to note how the impact of return policies differs between the return number and the return rate. For the non-zero part, reputation and return policies en- hance the return number, but not the return rate, which means that sellers would benefit from enhancing their reputations and adopting any available return policies.

Our results are relevant for the large number of actors on Taobao and can potentially be generalized to small and medium-sized online shops that face a similar situation like

Determining profilt-optimizing return policies 111

actors on Taobao. Moreover other platform operators like eBay could revisit the portfolio of return policies that sellers on these platforms can offer.

Theoretical implications

The theoretical contributions of this work are threefold. First, guarantee credibility not only works as a mediator between return policy and perceived risk, as it ultimately increases the customer’s willingness to pay (Suwelack et al. 2011), but seems to also be a new dimension of return policies that di- rectly affects customers’ purchasing or returning behavior. Dimensions of return policy should thus not be restricted to return deadline, customer effort and return coverage (Posselt et al. 2008; Su and Zhang 2009), but should also include guarantee credibility.

Second, the results of this study show that the return policy and guarantee credibility work in different ways. In this paper, we treat reputation as a moderator and find opposing moder- ating effects relative to return policy and guarantee credibility. Specifically, traditional return policies work as a quality signal that is substitutable for reputation, while guarantee credibility works as signal quality that is complementary to reputation.

Third, our study adopts a zero-inflated negative binominal regression model to explain online return behavior. We find evidence that this model is better suited than other regular regression models because ZINB can analyze both zero- values (current non-returns in this case) and counts parts among actual returns separately. The existence of differences between analyzing these two parts shows that this model might be useful for obtaining deeper insights into return behavior.

Managerial implications

Our paper suggests that operators of online platform like ebay.com and Taobao.com should offer both traditional return policies and policies that increase the guarantee credibility. In this way, small and medium-sized online sellers on these plat- forms can increase sales as well as decrease costs incurred from customer return behavior, thereby ultimately increasing profits. Now, most global online platforms do not offer guar- antee credibility policies, although such policies provide clear benefit for all parties. For example, Guarantee Money is a Bwin-win-win^ policy insofar as it offers better service to customers, stimulates sales, and increases the turnover for platform operators. In this paper we suggest a two-step ap- proach for examining the influence of return policies on sales and returns. The proposed approach combines robust regres- sion and ZINB regression, and allows users to easily estimate the results with available transactional data. This approach

proves to be valuable for making informed decisions about the optimal return policies for each shop. Finally, we suggest that managers do not focus only on return numbers, but also pay more attention to changes in the return rate because some return policies might increase the absolute return number but not the return rate itself.

For small and medium-sized online sellers that do not sell over a platform, customer friendly return policies and guaran- tee credibility seem to be even more important. Because with- out the umbrella function of the third-party market place, cus- tomers face an even higher level of uncertainty. As a result, more lenient return policies seem to be promising.

Limitations and future research

This study only examines return policies and guarantee cred- ibility in the online environment. Significant differences be- tween online and offline markets limit the generalizability of our results. For example, offline purchases do not suffer from severe information asymmetry and the fit between expectation and actual outcomes is higher. This leads to lower return in- tentions after purchases in brick-and-mortar stores. So our results cannot be transferred blindly to offline markets. Future research could compare the effect of return policies between online and offline markets.

Moreover, the insights generated by our results may be limited due to cultural differences. Almost all the shops on Taobao run their business in China and the buyers typically come from China as well, so a cultural bias may exist in our study too. We plan to test for culture influences on customer return behavior in the future.

Finally, the R-Squares of our models are relatively low. This might be due to the missing information on customer personality which likely also effects return behavior. A com- bination of data on the shop and on buyers’ level seems promising.

Acknowledgments The authors gratelfully acknowledge the financial support from the China Scholarship Council (No. 201307080002).

Appendix

Change for sales (ΔS): Change due to return policy [α2 or α3] / Average number of sales [116.726]

Change for returns (ΔR): Change due to return policy [ß2 or ß3]* Average return rate (with return record) [0.134]

Change for return probability (RP): Change due to return policy [γ2 or γ3]* Non-return rate [45.5 %]

Profit (P)=(1+ΔS)*[1-ΔR*(1- RP)]* Margin - ΔR*(1- RP)* Return cost

W. Zhou, O. Hinz112

where Margin is 10 % (Dobbs et al. 2013) of revenue, while Return cost is 3.8 % (Blanchard 2007) of revenue.

Profit of normal sellers (Pnormal)=1*(1–7.4 %)*10 %– 7.4 %*3.8 %=8.9788 %

Profit of sellers using Seven-Day Return (P7-day)=(1+ 0.0768) * [1–0.1895 * (1–45.7 %)] * 10 % - 0.1895 * (1– 45.7 %)*3.8 %=9.269 %

Increase in Profit from by Seven-Day Return Policy (ΔP7- day)=P7-day - Pnormal=0.29 %

Profit of sellers using Guarantee Money (Pmoney)=(1+ 0.00016) * [1–0.13397 * (1–45.58 %)] *10 % - 0.13397 * (1–45.58 %) *3.8 %=8.995 %

Increase in Profit from by Guarantee Money Policy (ΔP- money)=Pmoney - Pnormal=0.016 %

References

Alibaba. (2013). Alibaba group company overview. Alibaba Group. Retrieved Oct. 21, 2013, from http://www.alibabagroup.com/en/ about/overview

Anderson, E. T., Hansen, K., & Simester, D. (2009). The option value of returns: Theory and empirical evidence. Marketing Science, 28(3), 405–423.

Bae, S., & Lee, T. (2011). Product type and consumers’ perception of online consumer reviews. Electronic Markets, 21(4), 255–266.

Banjo, S. (2013). Rampant returns plague e-retailers. The Wall Street Journal. Retrieved May 15, 2014, from http://online.wsj.com/ n e w s / a r t i c l e s / SB10001424052702304773104579270260683155216.

Bechwati, N. N., & Siegal, W. S. (2005). The impact of the prechoice process on product returns. Journal of Marketing Research, 42(3), 358–367.

Blanchard, D. (2005). Moving forward in reverse. Logistics Today, 46(7), 1–8.

Blanchard, D. (2007, Aprial 10). Supply chains also work in reverse. Industry Week. Retrieved May 15, 2014, from http://www. industryweek.com/articles/supply_chains_also_work_in_reverse_ 13947.aspx.

Bonifield, C., Cole, C., & Schultz, R. L. (2010). Product returns on the internet: a case of mixed signals? Journal of Business Research, 63(9–10), 1058–1065.

Boulding, W., & Kirmani, A. (1993). A consumer-side experimen- tal examination of signaling theory: do consumers perceive warranties as signals of quality? Journal of Consumer Research, 20(1), 111–123.

Che, Y.-K. (1996). Customer return policies for experience goods. Journal of Industrial Economics, 44(1), 17–24.

Chu, W., Gerstner, E., & Hess, J. D. (1998). Managing dissatisfaction how to decrease customer opportunism by partial refunds. Journal of Service Research, 1(2), 140–155.

Davis, S., Hagerty, M., & CGerstner, E. (1998). Return policies and the optimal level of Bhassle^. Journal of Economics and Business, 50(5), 445–460.

Desmet, P. (2013). How retailer money-back guarantees influence con- sumer preferences for retailer versus national brands. Journal of Business Research, 67(9), 1971–1978.

Dobbs, R., Chen, Y., Orr, G., Manyika, J., Chui, M., & Chang, E. (2013). China’s e-tail revolution. McKinsey Global Institute.

Dutta, S., Biswas, A., & Grewal, D. (2007). Low price signal default: an empirical investigation of its consequences. Journal of the Academy of Marketing Science, 35(1), 76–88.

Erevelles, S., Roy, A., & Yip, L. S. (2001). The universality of the signal theory for products and services. Journal of Business Research, 52(2), 175–187.

Frischmann, T., Hinz, O., & Skiera, B. (2012). Retailers’ use of shipping cost strategies: free shipping or partitioned prices? International Journal of Electronic Commerce, 16(3), 65–88.

Glover, S., & Benbasat, I. (2010). A comprehensive model of perceived risk of e-commerce transactions. International Journal of Electronic Commerce, 15(2), 47–78.

Heilbron, D. C. (1994). Zero‐altered and other regression models for count data with added zeros. Biometrical Journal, 36(5), 531–547.

Heiman, A., McWilliams, B., & Zilberman, D. (2001). Demonstrations and money-back guarantees: market mechanisms to reduce uncer- tainty. Journal of Business Research, 54(1), 71–84.

Hess, J. D., & Mayhew, G. E. (1997). Modeling merchandise returns in direct marketing. Journal of Interactive Marketing, 11(2), 20–35.

Janakiraman, N., & Ordóñez, L. (2012). Effect of effort and deadlines on consumer product returns. Journal of Consumer Psychology, 22(2), 260–271.

Kirmani, A., & Rao, A. R. (2000). No pain, no gain: a critical review of the literature on signaling unobservable product quality. Journal of Marketing, 64(2), 66–79.

Mavlanova, T., & Benbunan-Fich, R. (2010). Counterfeit products on the internet: the role of seller-level and product-level information. International Journal of Electronic Commerce, 15(2), 79–104.

McLachlan, G., & Peel, D. (2004). Finite mixture models. New York: Wiley. Moorthy, S., & Srinivasan, K. (1995). Signaling quality with a money-

back guarantee: the role of transaction costs. Marketing Science, 14(4), 442–466.

Mwalili, S. M., Lesaffre, E., & Declerck, D. (2008). The zero-inflated negative binomial regression model with correction for misclassifi- cation: an example in caries research. Statistical Methods in Medical Research, 17(2), 123–139.

Padmanabhan, V., & Png, I. P. (1997). Manufacturer’s return policies and retail competition. Marketing Science, 16(1), 81–94.

Petersen, J. A., & Kumar, V. (2009). Are product returns a necessary evil? Antecedents and consequences. Journal of Marketing, 73(3), 35–51.

Pizzutti, C., & Fernandes, D. (2010). Effect of recovery efforts on con- sumer trust and loyalty in e-tail: a contingency model. International Journal of Electronic Commerce, 14(4), 127–160.

Posselt, T., Gerstner, E., & Radic, D. (2008). Rating e-tailers’ money- back guarantees. Journal of Service Research, 10(3), 207–219.

Purohit, D., & Srivastava, J. (2001). Effect of manufacturer reputation, retailer reputation, and product warranty on consumer judgments of product quality: a cue diagnosticity framework. Journal of Consumer Psychology, 10(3), 123–134.

Roggeveen, A. L., Goodstein, R. C., & Grewal, D. (2014). Improving the effect of guarantees: the role of a retailer’s reputation. Journal of Retailing, 90(1), 27–39.

Shulman, J. D., Coughlan, A. T., & Savaskan, R. C. (2009). Optimal restocking fees and information provision in an integrated de- mand–supply model of product returns. Manufacturing & Service Operations Management, 11(4), 577–594.

Simons, J. S., Neal, D. J., & Gaher, R. M. (2006). Risk for marijuana- related problems among college students: an application of zero- inflated negative binomial regression. The American Journal of Drug and Alcohol Abuse, 32(1), 41–53.

Su, X., & Zhang, F. (2009). On the value of commitment and availability guarantees when selling to strategic consumers. Management Science, 55(5), 713–726.

Sum, C. C., Lee, Y. S., Hays, J. M., & Hill, A. V. (2002). Modeling the effects of a service guarantee on perceived service quality using alternating conditional expectations (ace). Decision Sciences, 33(3), 347–384.

Determining profilt-optimizing return policies 113

Suwelack, T., Hogreve, J., & Hoyer, W. D. (2011). Understanding money-back guarantees: cognitive, affective, and behavioral out- comes. Journal of Retailing, 87(4), 462–478.

Thomason, E. (2013, Sepetember 27). Online retailers go hi-tech to size up shoppers and cut returns. Reuters. Retrieved May 15, 2014, from http://www.reuters.com/article/2013/09/27/net-us-retail-online- returns-idUSBRE98Q0GS20130927

Timmers, P. (1998). Business models for electronic markets. Electronic Markets, 8(2), 3–8.

Wood, S. L. (2001). Remote purchase environments: the influence of return policy leniency on two-stage decision processed. Journal of Marketing Research, 32(2), 157–169.

Ye, Q., Xu, M., Kiang, M., Wu, W., & Sun, F. (2013). In-depth analysis of the seller reputation and price premium relationship: a comparison between ebay us and taobao china. Journal of Electronic Commerce Research, 14(1), 1–10.

Zhang, X., Luo, J., & Li, Q. (2012). Do different reputation systems provide consistent signals of seller quality: a canonical correlation investigation of Chinese c2c marketplaces. Electronic Markets, 22(3), 155–168.

Zhuo, J., Wei, J., Liu, L., Koong, K., & Miao, S. (2013). An examination of the determinants of service quality in the Chinese express industry. Electronic Markets, 23(2), 163– 172.

W. Zhou, O. Hinz114

Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.