Dynamic pricing in online retail: Yay or Nay?

(1)

Dynamic pricing in online retail:

Yay or Nay?

By:

(2)

2

Master Thesis

Dynamic pricing in online retailing and its effect on sales

By

ILSE GRIËT BRUNINK

University of Groningen

Faculty of Economics and Business

MSc Marketing Intelligence & Management

13

th

_{of January 2020}

Populierenlaan 1-71 9741 HA Groningen (06) 36328687 i.g.brunink@student.rug.nl Student number S3522423

Supervisor: Prof. Dr. T.H.A. Bijmolt Second supervisor: R. Hars MSc.

Examiner: Prof. Dr. F. Eggers

University of Groningen Faculty of Economics & Business

(3)

3

Summary

Dynamic pricing is a hot topic nowadays in e-commerce. The airline industry first implemented it in the 1970s for revenue management and yield management. Every industry in the world nowadays adopts this strategy of revenue maximization through dynamic pricing.

Because of the quick rise of dynamic pricing in the online retail market in the Netherlands, it is interesting to take a look at what dynamic pricing means for a company in terms of pricing and sales. Dynamic pricing is for many companies, especially online retailers, a rather new topic. Dynamic pricing generates many data, and it is sometimes difficult for companies to analyze and understand it.

Hence there is not much empirical research on the effect of dynamic pricing on unit sales. Consequently, take into account the pricing of competitors. Therefore the following research questions were established: What is the effect of dynamic pricing on the unit sales per day? Various sub-questions were created to specify the problem.

Several authors found out that dynamic pricing is favorable for firms in terms of sales and revenue maximization issues as well as better purchase prediction of customers. Besides, based on the theory, several hypotheses were created for testing. The dependent variable is the unit sales per day. Independent variables are the daily product price, minimum competitors price, maximum competitors price, average competitors price, the number of competitors, price variance, the total number of price changes, and the total number of price changes of competitors.

(4)

(5)

5

Preface

Before starting this master thesis about dynamic pricing, I first want to provide some acknowledgment to some of the people who helped me during this semester.

First of all, I would like to thank my supervisor Prof. Dr. Tammo Bijmolt, for his useful advice and critical feedback during this semester and the thesis meetings.

Also, I would like to thank Rikke Nielsen and her colleagues of Frank.nl for providing me with the data for this thesis.

I would also like to express my gratitude for my fellow students in my thesis group and the students of the Hamburg University for the four great weeks in Groningen and Hamburg we spend together. We did not only have fun, but we also learned a lot from each other — also, the professors of Hamburg University for sharing their knowledge and thoughts about my thesis.

Also, I would like to thank my friends for their interest in my study in the recent years.

Lastly, I would like to express my deepest gratitude to my family. My mom, dad, and my sister. They have been my main pillars in the past two and a half years. When I started the pre-master, I never thought I would make it at all. They had always believed in me that I could do it, and when I was utterly done with it, they gave me the strength to continue. Thank you.

I hope you will enjoy reading my thesis.

Ilse Brunink

(6)

6

1. Introduction

1.1. Background

Since the rise of online retailers, prices are not as static anymore as they used to be. Today it is possible to change prices multiple times per day. Dynamic pricing strategies have been successfully applied across multiple industries, such as the airline industry, hotels, and car rental agencies (Kannan & Kopalle, 2001). The airline industry started with the use of dynamic pricing strategies since the 1970s to maximize their revenue, which is called revenue management or yield management. The airline industry used a form of pricing that uses price elasticities. The optimum number of products is offered to the right number of customers at the best price (Maglaras & Meissner, 2006). Basically, every industry in the world nowadays adopts this strategy of revenue maximization through dynamic pricing. Due to the Internet, alternative pricing strategies have been applied for selling goods and services online. Such as internet-based auctions for commodity suppliers (Gerstgrasser, 2019). The internet increases the ease and efficiency of applying auctions to other products like electronics or clothing. Furthermore, for products sold at posted prices usually were changed weekly or even longer. Now the products can be sold online for posted prices that can be changed daily or even multiple times per day. Pricing is becoming more and more dynamic. Because of the efficiency of the internet, it is possible to adjust prices more quickly based on the fact that it is easier to measure the demand and track competitive prices (Jayaraman & Baker, 2003).

(9)

9

1.2. Problem

Because of the quick rise of dynamic pricing in the online retail market in the Netherlands, it is interesting to take a look at what dynamic pricing means for a company in terms of pricing and sales. Dynamic pricing is for many companies, especially online retailers, a rather new topic. Dynamic pricing generates many data, and it is sometimes difficult for companies to analyze and understand it. It requires new programs and tools, as well as people that have the competencies for work with data (Kannan & Kopalle, 2001).

1.3. Research questions

The main goal of this thesis is to provide insights into dynamic pricing and its effect on sales. Primarily, what is the effect of the dynamic pricing on the unit sales per day online? Furthermore, how can this be translated into a better optimization of dynamic pricing? The main problem statement: "What is the effect of dynamic pricing on the unit sales per day?". To answer the problem statement, the researcher created several sub-questions are:

1. What is dynamic pricing?

2. Which dynamic pricing strategies are there?

3. What is the effect of competition on dynamic pricing?

4. What models can be used to measure the effect of dynamic pricing on unit sales per day? 5. If the prices of a product are dynamic, how will this influence the relationship between

price and sales?

1.4. Relevance

Pricing is important in marketing is part of the marketing mix. Having the right pricing strategy is crucial for an organization but is also highly complex (Gijsbrechts, 1993). Especially when the firm is dealing with much competition offering relatively the same type of products or product categories (Willart, 2015). Even though there is multiple research available about pricing, there is a gap in the literature about dynamic pricing. Besides, there is not much empirical research on the effect of dynamic pricing on unit sales. Hence, take into account the pricing of competitors. This thesis will provide new insights into dynamic pricing in a Dutch online retail market.

(10)

10 spread over four primary markets: Living & Interior, Sport & Outdoor, Kitchen & Dining, and Toys & Hobby. They started in the Netherlands, but are now active throughout Europe in 15 countries. Frank.nl believes that online shopping has no limits and should be available to everyone. Therefore, their online stores are well-organized and straightforward and offer a wide range of more than 300,000 products (Frank.nl, n.d.).

1.4. Outline of the research

(11)

11

2. Theory

This chapter includes the conceptual framework of this research. Then, the literature is reviewed on dynamic pricing to get a greater understanding of the topic of dynamic pricing, including hypotheses.

(12)

12

2.1. Dynamic pricing

Dynamic pricing is often defined as a pricing strategy in which prices change over time, across customers, and/or products (Kannan & Kopalle, 2001). An innovative pricing mechanism allows adjusting prices with the use of the gathered data about the customer, such as where they live, what they buy, and how much they spend (Weiss & Mehrotra, 2001). Also, it allows companies to price discriminate on an individual customer level to maximize profits as well as their consumer surplus (Weisstein, Monroe, & Kukar-Kinney, 2013).

The change from physical stores to internet stores websites have changed the way of pricing of products. Kannan and Kopalle (2001) refer to this as the physical chain and the virtual chain. The physical chain was, for a long time, the dominant model. Information flowed alongside the product and was nothing more than a supporting element and not a source of value. In the new internet-era, data and information is more available and accessed more easily. It creates this information-based environment, even the product itself can be developed dynamically as well as priced dynamically. Therefore, dynamic pricing can be based on individual online behavior, and companies can micromanage their marketing and pricing strategies. Nowadays, by the use of cookies, data is retrievable about the website information from the Internet users' information drive about the past interactions with the website. Also, you can track with clickstream technology the users' path when they viewed and clicked on an advertisement that linked to the website, the different pages on the website, or links to other websites (Weiss & Mehrotra, 2001). The downside of this tracking is individual online behavior for pricing is that it can cause price discrimination.

2.2. Dynamic pricing strategies

(13)

13 When a retailer has a competition-based dynamic pricing strategy, always monitor the competitor's price changes and use this as an input for their prices (Fisher, Gallino, & Li, 2018). For example, a retailer can ask several Euro or some percent lower or higher than the target competitor or a competitor with the lowest price. When competitors change their prices, the retailer must answer several questions: Should the retailer respond to the price change? If so, to which competitor? How much should the retailer respond? And to which products? Most strategies are characterized by competition due to limited demand information. For a successful strategy, frequent price adjustment and anticipation of the market dynamics are vital (Schlosser & Boissier, 2018). Optimizing price adjustment can be very time consuming and complicated; therefore, in competitive markets, competition-based dynamic pricing strategy is most applicable and is usually applied by e-commerce.

2.3. Dynamic pricing effect on sales

As stated previously, companies use dynamic pricing to maximize profits as well as their consumer surplus. Several authors researched the effect of dynamic pricing on sales.

Gupta and Pathak (2014) used a machine learning framework for predicting purchases by online customers based on dynamic pricing. In this case, the purchase decision was a binary variable. From the individual customers, they created 6 clusters, and for each cluster, a dynamic price range was determined using a regression. Lastly, the purchase decision was prediction using logistic regression. They found out that when the perfect price range determination leads to a better revenue generation compared to products with a fixed price. Moreover, it will lead to a lower error rate in the prediction of purchase behavior (Gupta & Pathak, 2014). So, dynamic pricing can lead to higher revenue and a higher prediction of customer behavior.

(14)

(15)

15

Table 1: Overview theory dynamic pricing effect on sales

Author Dynamic pricing definition DV Outcome Between customers Time series data Competition included Gupta & Pathak (2014) Offering different prices to customers at different times Dynamic pricing leads to better purchase prediction Yes No No Maglaras & Meissner (2006) 2006 Fluctuations in the price of products, separate pricing strategies for product Comparing multiple models for multiple product pricing No No Yes Kopalle, Mela, & Marsh (1999) 1999 Change in price due to promotions Dynamic pricing has a positive effect on sales No Yes Yes

2.4. Price, size, and frequency of price changes

(16)

16 price dynamically, it has a positive effect on the unit sales of a retailer. It may indicate that frequent price changes is has a positive effect on unit sales. Therefore, the following hypotheses have been established for the size and frequency of price changes on the effect of unit sales.

H1: the number of price changes has a positive effect on the unit sales per day H2: the variance ratio in price has a positive effect on the unit sales per day

As mentioned before, according to Kopalle, Mela, & Marsh (1999), a temporary decrease in price can increase price sensitivity among customers. Price sensitivity is the overall reaction of consumers to the price of the product and how they feel about it (Ramirez & Goldsmith, 2009). When a consumer does not buy the product immediately, they can check the next day the product and its price again, it is possible that the price has gone up. Which will lead to not purchasing the product. The response in price is measured in price elasticities. Price elasticities measure how sensitive the demand for a product response is when there is a change in price. Price elasticity is characterized as a percentage change in the quantity of the response demand to a 1% change in their price. Price elasticities are usually negative (Huang et al., 2018).

According to the theory above of price elasticities, an increase (decrease) in price leads to a decrease (increase) of demand. So when the price goes up, fewer units will be sold. Therefore, the following hypothesis is established:

H3: Price has a negative effect on the unit sales per day

Even though the expected effect is that price has a negative effect on unit sales, it is also expected that dynamic pricing has a positive effect on sales. Therefore, if a product is priced more dynamic, this will affect the relationship between price and unit sales. Therefore, the following hypotheses are created:

H4: there is a moderation effect of the price variance of a product on the effect between price and the unit sales per day

(17)

17

2.5. Competition

The more competitors, the more choice, the more it affects the unit sales. Also, consumers are getting more aware of different retail providers, which leads to the rise of price comparison websites (Ronayne, 2019). Because of these price comparison websites, firms are getting affected by the price changes of competitors and the entry of more competitors. A result of more competition is the increasing price competition pressure on firms by the price sensitivity of the shopper. The more competition, the more likely that the consumer chooses a company that offers a better deal. For firms, it is always a trade-off for charging prices for products (Papi, 2018). When consumers are searching for prices, they always have the aspirational price in their head. Firms can charge the highest price possible what consumers see as their most aspirational price, which will decrease competition in price. Consumers will not search for alternatives. On the other hand, firms can also reduce their price always lower than their components to induce price comparison. The last is usually the case. With price variations and firms always try to set prices lower than their competition, the searching for alternatives increases by consumers.

Based on the research above, the more competitors there are, the more likely a consumer will search for a better price at a competitor. Therefore, the following hypothesis has been established.

H6: the number of competitors has a negative effect on the unit sales per day

When the price of one product is changed, it usually affects the demand for substitute products. Cross price elasticities is a tool to measure the price elasticities concerning competitive products. Cross price elasticity is the percentage change in the quantity of the response demand to a change in prices of other related products (Huang et al., 2018). For cross-price elasticities, it is essential to distinguish between products that are complements or substitutes (Graves & Sexton, 2009). Complement products are that the more use of one product needs more use of another product. Complement products usually have negative cross-price elasticities.

(18)

18 indicates when a price increases of a flight, the choice to take the train is very minimal for consumers. The cross-price elasticity is positive (0.0008) but almost zero (Gama, 2017).

Another example is cigarettes and e-cigarettes. According to Grace & Kivell (2015), the cross elasticities for e-cigarettes and regular cigarettes is 0.16. If the market price for tobacco cigarettes will increase two times the current market price, 30% of the participants will buy the e-cigarette. The closer the substitutes so the more identical to the product, the higher the cross-price elasticity will be.

When competitors increase their prices, the demand will decrease for the products. According to cross-price elasticities, when a retailer increases its price of a product, the demand for this product will increase for the competitors. Therefore, the following hypotheses have been established for the effect of dynamic pricing on the unit sales per day:

H7: the prices of competitors have a positive effect on the unit sales per day

(19)

19

3. Design of the empirical study

This chapter contains an explanation about the research design, the data, the variables, and the modeling used for the data. Also, some descriptive statistics and graphs for visualization.

3.1. Dynamic pricing at Frank.nl

Frank.nl using a dynamic pricing tool called Omnia. Their dynamic pricing strategy is that there is not a clear strategy. Prices are changed once a day, usually during the night. It is also possible that the price stays the same for an extended period. The prices are based on the competitors and usually match the lowest-priced competitors. It is possible to give weights to each competitor based on the importance of price considerations. However, Frank.nl does not do that. Hence, Frank.nl is following a competitors-based dynamic pricing strategy. It is also possible to set fixed rules of the pricing, such as setting a fixed price or always 10% under the lowest competitors. It is also possible that some products are not dynamic prices due to the interference of the supplier, and the suggested retail price must be maintained. Omnia gets its data from Google shopping and does not scrap the websites of the competitors.

3.2. Research design

The research design of this thesis is conclusive. The research is about testing hypothesis and to examine relationships. Furthermore, the data analysis is quantitative data and based on secondary data. The researcher did not collect the data (Malhotra, 2010). Lastly, the type of conclusive research design is a descriptive research design. Descriptive research is about describing something, usually market characteristics, functions, relationships, and make predictions.

3.3. Data

Data analysis is performed to answer the research questions. The data analysis is performed in the software R, which is a software program, as well as a programming language. Besides, R is an open-source software and free of charge.

(20)

20 2018 until the 22nd_{of October 2019. The prices of products are in the data, including the suggested}

retail price (SRP), daily prices of the competitors of each product. Of these 300 products, a selection of products was made that have at least 100 units sold in one year. From the dataset of 300 products, only the products that had at least 100 or more units sold over 365 days were selected. This came to a total of 143 products. Of these products, several products with incomplete prices removed from the dataset. The sample for this research is 135 products.

Of these products, graphs were made to visualize the prices for each day to consider if the pricing was dynamic or static. Figure 2 shows a very dynamic priced product. The red line indicates how many price changes there are during the 365 days. Figure 3 also shows a dynamic priced product, but this product does not change almost daily but varies more weekly overtime, which is also a dynamic priced product.

(21)

21

Figure 3: Example of a moderate dynamic priced product

Figure 4 below indicates a more static product. This product has only two price changes in one year.

Figure 4: Example of highly static priced product

(22)

22

Figure 5: Example of dynamic priced product influenced by season

3.3.1. Variables

The dataset contains several variables. A complete overview of the variables is outlined in table 1.

Table 2: Overview of all variables

Variable Specification Description

Unit sales per day Dependent variable The unit sales per day per product over 365 days Daily product price

(Price.Frank.nl)

Independent variable The daily price per product over 365 days

Minimum competitors price (min_price)

Independent variable The price of the lowest competitor per product on a specific day

Maximum competitors price (max_price)

Independent variable The price of the highest competitors per product on a specific day

Number of competitors (number_competitors)

Independent variable The total number of

(23)

23 competitors, the total number of that day is zero.

Average temperature (avg_temp) Control variable The average temperature per day

Season Control variable Autumn, summer, winter, and

spring

Variance ratio (variance_ratio) Independent variable The price variance of each product. Calculate by (price variance/mean price) of the daily product price

Total number of price changes (total_price_changes)

Independent variable The total number of price changes per products Category dummies

(CAT1…CAT39)

Control variable Dummy variables for each category

Total number of price changes of the competitors

(total_price_changes_competition)

Independent variable The total number of price changes per products of the competitors. Based on the average competitors' price.

The dependent variable is the unit sales per day. Independent variables are the daily product price, minimum competitors price, maximum competitors price, average competitors price, the number of competitors, price variance, the total number of price changes, and the total number of price changes of competitors.

(24)

24

Figure 6: Total number of price changes of the product in the dataset

The data contains many missing values in the prices of the competitors. Also, there are over 1,000 competitors in the dataset. Modeling each competitor is much work, and therefore the competitors' price variables were to be created by taking the average, minimum, and maximum prices of the competitors. Of all the competitors, the average competitors, minimum competitors' price, and maximum competitors' prices were calculated for each day per product. Furthermore, for the ‘number of competitors’ variable, it is the case when there is not a price of the competitors for that day, the competitor will not be seen as a competitor for that specific day. One of the reasons is that the program Omnia is getting their prices from Google shopping and not directly from the websites of the competitors. When a product and its price is not on Google shopping, the product may be out of stock, the competitors are not offering the product anymore, or the competitors did not offer that product yet.

3.3.2. Imputation of variables

(25)

25 the most optimal values. Mainly because the minimum competitors' price was sometimes higher than the maximum competitors' prices for the imputed variables, this is not possible, and it is also not consistent with the non-missing values. Furthermore, the average competitors' price should be between the minimum and maximum competitors' prices. The k-nearest neighbor imputation method was the only method that accounted for all that and, therefore, the most suitable imputation method for this dataset.

3.3.3. Control variables

In addition to the independent variables, also control variables are included in the dataset. The 'average temperature' is the average of the minimum and maximum temperature for that day. These values are from the database of the KNMI (Koninklijk Nederlands Meteorologisch Instituut), which is the Royal Dutch Meteorological Institute. According to Steinker, Hoberg, & Thonemann (2017), the weather has a significant effect on online sales. Sun, rain, and temperature all have a significant effect on online daily sales. Especially during the summer, weekends, and when there is extreme weather. Also, the control variable 'season' is included in the dataset as well. According to Kumar, Gaidarev, and Woo (2004), the definition of seasonality is the underlying demand of a specific type of product group as a function of time of the year that is independent of other external factors such as changes in price, inventory, and promotions. Seasonality is also consistent from year to year. For e-commerce, in the fourth quarter of the year, the highest sales are usually made (Dehning, Richardson, Urbaczewski, & Wells, 2004). Also, all kinds of different products are in the dataset spread over multiple categories. Some products are more seasonal based, such as swimming pools.

(26)

26

Figure 7: Total unit sales per month

Lastly, dummy variables for the product categories are included as well. In total, the products are divided into 38 categories. The list of the categories, including explanation, is in appendix A.

3.3.4. Correlation

(27)

27

Figure 8: Correlation matrix of variables

3.4. Modeling count data

The dependent variables in this research is a count variable. According to Hilbe (2014), a count variable is a that has nonnegative integer values ranging from zero to some higher undermined value. It can range from zero to infinity but generally to the maximum value of the count data.

3.4.1. Different count models

Most count models are based on two probability functions - the Poisson and the Negative Binomial Probability Density Function. Additional models that are also considered necessary for modeling count data are the Poisson inverse Gaussian model (PIG), Greene's three-parameter Negative Binomial P (NB-P), and generalize Poisson (GP) model. These distributions are all closely related (Hilbe, 2014).

(28)

28 model fails to adjust for overdispersion. So the standard errors are biased, and the results cannot be trusted because they may appear significant when it is, in fact, not significant. Important is for the data to do a test for over - and under dispersion.

The most obvious and most used solution for dealing with Poisson overdispersion or underdispersion is to use a Negative Binomial model. It has an extra parameter, compared to the Poisson model, called the negative binomial dispersion parameter. However, it is not usable when there is an underdispersion.

3.4.2. Zero’s included in count data

Another aspect to take into consideration when modeling count data are zero counts. Poisson, Negative Binomial, and PIG distributions all assume the possibility of zero counts even if there are may not be any in the data. If zero counts are not possible for modeling that the PDF needs to be adjusted for excluding zero counts. These kinds of models called Zero-truncated models. Data with a disproportionate amount of zero counts is another problem. The expected percentage of zero counts based on the Poisson PDF is under 1%. When there is a low mean and a high percentage of zero counts, the model needs to be adjusted accordingly. It is possible to use either a two-part hurdle model or a mixture model like Zero-inflated Poisson or a Zero-inflated Negative Binomial (Hilbe, 2014).

3.4.3. Assessing model fit

(29)

29

3.4.5. Model specification

Below, the equation of the count model is outlined, as well as what each abbreviation is meaning.

𝑆_𝑖 = exp (𝛼_𝑖) + exp (𝛽₁𝑃_𝑖) + exp (𝛽₂𝑀𝐼𝐶𝑃_𝑖) + exp (𝛽₃𝑀𝐴𝐶𝑃_𝑖) + exp ( 𝛽₄𝐴𝐶𝑃_𝑖) + exp ( 𝛽5𝑁𝐶𝑖) + exp (𝛽6𝑉𝑅𝑖) + exp (𝛽7𝑇𝑃𝐶𝑖) + exp (𝛽8𝑇𝑃𝐶𝐶𝑖)

+ exp (𝛽₉𝐴𝑉𝐺𝑇) + exp (𝛽₁₀𝑆𝑆) + exp (𝛽₁₁𝐶𝐴𝑇1 … … + 𝛽₄₉𝐶𝐴𝑇39) + 𝜀_𝑖

𝑆_𝑖 = Sales in units of product i

𝛼 = Intercept

𝑃𝑖 = Daily product price of product i

𝑀𝐼𝐶𝑃_𝑖 = Minimum competitors daily price of product i 𝑀𝐴𝐶𝑃𝑖 = Maximum competitors daily price of product i

𝐴𝐶𝑃_𝑖 = Average competitors daily price of product i

𝑁𝐶_𝑖 = Number of daily competitors of product i

𝑉𝑅_𝑖 = Variance ratio of product i

𝑇𝑃𝐶𝑖 = Total number of price changes of product i

𝑇𝑃𝐶𝐶_𝑖 = Total number of competitors price changes of product i

𝐴𝑉𝐺𝑇 = Average daily temperature

𝑆𝑆 = Season: autumn, winter, spring, or summer

𝐶𝐴𝑇1 … … + 𝐶𝐴𝑇39) = Dummy variables of the 39 product categories

(30)

30

4. Results

In this chapter, the results of the count models are outlined. First, the chapter will start with the dispersion test is performed to assess which count model to use. Then, the different count models are estimated, as well as the assessment fit to consider which model has the best fit. Lastly, the direct effects and interaction effects and the hypotheses were tested.

4.1. Model choice

To test whether which distribution is the right one to use, Poisson or Negative Binomial, a dispersion test needs to be performed. The first step is to look at the mean and the variance for the count variable ‘unit sales per day.’ To use a Poisson distribution, the mean and the variance needs to be equal to each other. The mean is 0.4703, and the variance is 15.092. The mean and variance are not equal, and therefore, the Poisson distribution cannot be used on this type of data. Furthermore, the variance is much higher than the mean, which indicates that there is the assumption of overdispersion. Hence, a dispersion test has been performed for the Poisson model as well to test this assumption. The dispersion test is significant with a p-value of 0.0005, which indicates that there is overdispersion. Also, the alpha is 54.67218, which is very high. The higher the alpha, the higher the overdispersion.

The variable of unit sales per day contains many zero's. In total, out of the 49,275 observations, 39,090 are zero's, which is 79.33%. Furthermore, it is possible to have zero observations. Therefore, the Zero-inflated or Zero-altered/Hurdle Negative Binomial model is the right model to use for this data. Both count models, which include zeros, are run in R. By the use of the information criteria and likelihood ratio, the best model will be determined.

(31)

31 fit of the different zero-inflated models. The first model performs better compared to the null model according to the AIC and log-likelihood.

Table 3: Assessment fit of the Zero-inflated Negative Binomial model

Model df Log-likelihood AIC

ZI-NB model without multicollinearity

19 -38930.8 77891.60

ZI-NB null model 3 -39684.6 79375.21

Table 4 shows the assessment fit of the Hurdle/Zero-altered models. This model could handle all the dummy variables for the product categories. Also, this model is checked for multicollinearity between the independent variables. Again, multiple variables appeared to be highly correlated. The remaining variables are daily product price, the number of competitors, maximum competitors price, variance ratio, season, average temperature, the total number of price changes, and the total price changes of the competitors. The log-likelihood estimation, as well as the AIC, both indicate that the hurdle model without multicollinearity is the best performing model. Furthermore, the hurdle model performs better than the zero-inflated models.

Table 4: Assessment fit of Hurdle Negative Binomial model

Model df Log-likelihood AIC

Hurdle model without multicollinearity

103 -36652.7 73503.41

Hurdle null model 3 -39390.67 78787.33

4.2. Direct effects

(32)

32

Table 5: Output of direct effects of the Hurdle Negative Binomial model

Independent variable Estimates Standard error p-value Zero hurdle model (logit)

Intercept

Daily product price 0.0001 0.0004 0.1274

Number of competitors 0.018 0.0015 0.0001

Maximum competitors price 0.0005 0.0001 0.0001

Total price changes 0.0034 0.0008 0.0001

Variance ratio 0.2617 0.0324 0.0001

Total price changes competitors -0.0017 0.0003 0.0001 Control variables Average temperature -0.0158 0.0033 0.0001 Summer 0.3458 0.0418 0.0001 Autumn 0.1581 0.0328 0.0001 Winter -0.0857 0.0368 0.0198 Count model Intercept

Daily product price -0.002 0.0007 0.0001

Number of competitors 0.006 0.0031 0.0556

Maximum competitors price 0.0001 0.0002 0.3717

Total price changes 0.008 0.0015 0.0001

Variance ratio 0.267 0.0644 0.0001

(33)

33

Table 6: Exponent coefficients of the direct effects of the Hurdle Negative Binomial model

Independent variable Exponent coefficients Count model

Exponent coefficients Zero hurdle model (logit)

Intercept 1.0070 2.1156

Daily product price 0.9975*** 1.0006

Number of competitors 1.0059* 1.0177***

Maximum competitors price 1.0001 1.0005***

Total price changes 1.0077*** 1.0033***

Variance ratio 1.3060*** 1.2991***

Total price changes competitors 0.9965*** 0.9984***

*_p<0.1;**_p<0.05;***_p<0.01

The hurdle model is interpreted into two parts. The first part is the logit model. This means the effect of whether there are positive or zero counts. Meaning when there is a unit sold or not. The second model is the count model for all positive counts (unit sales). The baseline odds of having a positive count versus zero (actual sales versus zero unit sales) is 2.1156. Given that the unit sales are positive, the average unit sales are 1.0070.

The total price changes have a positive and significant effect in the logit model (p=0.0001). The more price changes a product has, the more likely there will be a unit sold. The odds of unit sales are increased by 1.0033 times by a unit increase in total price changes, which is 100*(1.0033-1) = 0.33%. In the count model, the total price changes also have a positive and significant effect (p=0.0001). If the price changes of a product and increases by one unit, the odds are increased by 1.0077 times, which is 100*( 1.0077-1)= 0.77% on days where there is at least one unit sold. H1 is supported.

(34)

34 in price increases by one unit, the odds will increase by 1.3060, which is 100*(1.3060-1) = 30.60%, on days were at least one unit is sold. H2 is supported.

For the logit, the daily price does not have a significant effect on unit sales per day. This means the price does not affect whether there is a purchase or not. For the count model, the daily product price has a significant negative effect on the unit sales per day (p= 0.0001). This means when the daily price of a product increases, the fewer sales will be made on days when there is at least one unit sold. When the daily price increase by one unit, the odds of the unit will decrease by 0.9975, which is 100*(0.9975-1) = 0.25%. H3 is partially supported.

The number of competitors has a positive and significant effect on unit sales per day in the logit model (p=0.0000). Meaning, the number of competitors has a positive effect on whether there is a product sold or not. The odds of unit sales are increased by 1.0177 times by a unit increase in the number of competitors, which is 100*(1.0177-1) = 1.77%. In the count model, the number of competitors does not have a significant effect on unit sales per day. H6 is not supported.

The minimum and average competitors' prices were deleted from the model due to high correlation with each other and with the daily product price. The maximum competitors' price did not correlate and remained in the model. The maximum competitors' price has a positive and significant effect on the unit sales per day in the logit model (p=0.0001). The higher the price of the competitors, the more likely that customers would purchase a product. The odds of unit sales are increased by 1.0005 times by a one-unit increase in the maximum competitors' price, which is 100*(1.0005-1) = 0.05%. For the count model, the maximum competitors' price is not significant. H7 is partially supported.

(35)

35

4.3. Interaction effects

After the direct effect model estimation, the interaction effects were included in the model as well. Below the results of the interaction effects will be described. The output of the interaction effect is outlined in table 7.

Table 7: Output of interaction effects of the Hurdle Negative Binomial model

Independent variable Estimates Standard error p-value Zero hurdle model (logit)

Daily product price * variance ratio

0.0005 0.0007 0.4941

Daily product price * total price changes

0.0005 Not estimated Not estimated

Count model

Daily product price * variance ratio

0.0011 0.0919 0.3683

Daily product price * total price changes

0.0001 Not estimated Not estimated

The first interaction effect is the effect of the variance ratio on the effect between daily product price and unit sales per day. The interaction effect of price variance on the effect of the daily product price on the unit sales per day appears not to be significant in both the count model and in the logit model. So the variance in price does not affect the relation between daily product price and unit sales per day. H4 is not supported.

(36)

(37)

37

5. Conclusion and discussion

The focus of this thesis was to provide insights into the effect of dynamic pricing on the unit sales per day for online retailing. The hypotheses were tested by the use of a Hurdle Negative Binomial model. An overview of all the hypotheses is outlined in table 8.

Table 8: Hypotheses table

Hypotheses Accept/reject

H1: the number of price changes has a positive effect on the unit sales per day

Accept

H2: the variance ratio in price has a positive effect on the unit sales per day

Accept

H3: Price has a negative effect on the unit sales per day Partially accept

H4: there is a moderation effect of the price variance of a product on the effect between price and the unit sales per day

Reject

H5: there is a moderation effect of the number of price changes on the effect between price and the unit sales per day

N.A.

H6: the number of competitors has a negative effect on the unit sales per day

Reject

H7: the prices of competitors have a positive effect on the unit sales per day

Partially accept

H8: the number of price changes of competitors has a positive effect on the unit sales per day

Reject

(38)

38 By changing the price of a product, it is vital to be aware of the price sensitivity of consumers. A temporary decrease in price will results in the high price sensitivity of consumers when the prices will increase again. Also, according to price elasticities, an increase is price will result in a decrease in demand. The results of the analysis are that price indeed has a negative effect on whether the number of units of a product will be sold. Also, it was tested if the variance of price may have a moderation effect on the effect of price and unit sales. However, this effect proofed to be insignificant. Therefore, the variance in price does not change in the negative effect price has on sales.

Concerning what was proposed by the theory, the number of competitors seems not to be having a negative effect on the unit sales per day. It even seems to have a positive effect on whether a product will be sold or not. Several aspects can plays or role here such as the online retailer in this research have mainly a lower price than the competitors or the choices of alternatives are too much and consumers may experience the effect of choice overload. However, these are just assumptions and can be explored by future research.

The minimum competitors' price and the average competitors' price were deleted from the model due to multicollinearity issues. Therefore, the analysis of the prices of competitors is based on the maximum competitors' price. The maximum competitors' price appears to have a positive effect on the unit sales per day on whether a product will be sold or not. This effect is in line with the proposed theory about cross-price elasticities. If the prices of a product of competitors increases, the demand for that product will decrease, and consumers will search for alternative retailers to purchase the product. Therefore, the own demand of that product of a retailer will increase. The total number of price changes appears to have a negative effect on the unit sales instead of positive, which was proposed by the theory. The more price changes a competitor has, the more it affects the unit sales per day. From these results, it can be said that the dynamic pricing of products also has a positive effect on competitors, and therefore this will result in a negative effect on the unit sales per day.

(39)

39 price changes will increase unit sales. Unfortunately, the price will still have a negative effect on sales; not even dynamic pricing can change this relationship.

5.1. Managerial implications

First of all, the outcome of the study suggests that the dynamic pricing of products is favorable for an online retailer. Therefore, product managers of online retailers should adopt a dynamic pricing strategy for their products. Three aspect needs to be taken into consideration when pricing products dynamically, namely the frequency of price change, the size of the price change, and the prices of the competition.

More substantial price changes appear to have a bigger effect on unit sales than smaller frequent price changes. More substantial price changes have a bigger effect, but the downside is that it will increase the price sensitivity of the customer. A higher decrease in price seems favorable to sell a product, but when the price increases again will increase the price sensitivity. Products can only be sold at a low price and will turn unprofitable for retailers. It should be a trade-off between big or small price changes and frequent or infrequently price changes.

In addition, dynamic pricing should also be important to keep up with competitors. The study suggests that if competitors dynamically price their product, it will affect the unit sales of their firm in a negative way. Does the competition price dynamically, but do you keep the prices static? Then there is a good chance that you will miss out on sales. Keep track of the prices of competitors is crucial in highly competitive markets and is easy to do for an online retailer.

5.2. Limitations and suggestions for further research

This research has some limitations regarding the data and the analysis. Therefore, these limitations should be kept in mind when reading the results.

First of all, several variables had to been removed from the model due to multicollinearity by having high VIF-scores. The average competitors' price and the minimum competitors' price could not be included in the analysis due to the high multicollinearity of the daily product price of the online retailer.

(40)

40 created. Still, as much the researcher tried to make the categories specific, there are still some minor differences within the product categories. Furthermore, this research uses a count model and uses product categories to capture the individual effects of product categories. For further research, it is also possible to use a panel model to estimate the results. The panel model will also account for time series data.

Another limitation is that some values are imputed. The prices of competitors contain a lot of missing values. Therefore, the three variables average competitors price, maximum competitors price, and the minimum competitors' price were created to deal with the missing values. However, the dataset still contains missing values on several days. Thus, these missing values were imputed. Imputation is always risky to do because it can create bias results.

The model is not accounted for price endogeneity. Meaning that the price does not only affect sales, but the price is also affected by the sales in return. The sales may be high and have a high price because there is more demand for a specific product. This effect is not included in the research. For further research, it is suggested to account for price endogeneity in the model to create more precise results

(41)

41

Reference list

Dehning, B., Richardson, V. J., Urbaczewski, A., & Wells, J. D. (2004). Reexamining the value relevance of E-Commerce initiatives. Journal of Management Information Systems, 21(1), 55–82. https://doi.org/10.1080/07421222.2004.11045788

Elmaghraby, W., Elmaghraby, W., & Keskinocak, P. (2003). Dynamic pricing in the presence of inventory considerations: Research overview, current practices, and future directions Identification and Allocation of Increased-Risk Encephalitis Organs View project Responsible Operations View project Dynamic Pricing i. 49(10), 1287–1309. https://doi.org/10.1109/EMR.2003.24939

Fisher, M., Gallino, S., & Li, J. (2018). Competition-based dynamic pricing in online retailing: a methodology validated with field experiments. Management Science, 64(6), 2496–2514. https://doi.org/10.1287/mnsc.2017.2753

Gama, A. (2017). Own and cross-price elasticities of demand for domestic flights and intercity trains in the U.S. Transportation Research Part D: Transport and Environment, 54, 360– 371. https://doi.org/10.1016/j.trd.2017.06.010

Gerstgrasser, M. (2019). Reverse auctions are different from auctions. Information Processing Letters, 147, 49–54. https://doi.org/10.1016/j.ipl.2019.03.006

Gijsbrechts, E. (1993). Research in Consumer Ome Recent Developments. International Journal of Research in Marketing Intern. J. of Research in Marketing IO, 10(2), 115–151. Retrieved from https://pure.uvt.nl/portal/files/933360/pricesan.pdf

Gupta, R., & Pathak, C. (2014). A machine learning framework for predicting purchase by online customers based on dynamic pricing. Procedia Computer Science, 36(C), 599–605.

https://doi.org/10.1016/j.procs.2014.09.060

Huang, J., Gwarnicki, C., Xu, X., Caraballo, R. S., Wada, R., & Chaloupka, F. J. (2018). A comprehensive examination of own- and cross-price elasticities of tobacco and nicotine replacement products in the U.S. Preventive Medicine, 117(April), 107–114.

(42)

42 Jayaraman, V., & Baker, T. (2003). The internet as an enabler for dynamic pricing of goods.

IEEE Transactions on Engineering Management, 50(4), 470–477. https://doi.org/10.1109/TEM.2003.820134

Kannan, P. K., & Kopalle, P. K. (2001). Dynamic pricing on the Internet: Importance and implications for consumer behavior. International Journal of Electronic Commerce, 5(3), 63–83. https://doi.org/10.1080/10864415.2001.11044211

Kopalle, P. K., Mela, C. F., & Marsh, L. (1999). The dynamic effect of discounting on sales: Empirical analysis and normative pricing implications. Marketing Science, 18(3), 317–332. https://doi.org/10.1287/mksc.18.3.317

Maglaras, C., & Meissner, J. (2006). Dynamic pricing strategies for multiproduct revenue management problems. Manufacturing and Service Operations Management, 8(2), 136– 148. https://doi.org/10.1287/msom.1060.0105

Papi, M. (2018). Price competition with satisficing consumers. International Journal of Industrial Organization, 58, 252–272. https://doi.org/10.1016/j.ijindorg.2017.09.001 Ramirez, E., & Goldsmith, R. (2009). Some antecedents of price sensitivity. Journal of

Marketing Theory and Practice, 17(3), 199–213. https://doi.org/10.2753/MTP1069-6679170301

Ronayne, D. (2019). Price Comparison Websites David Ronayne No : 1056 Warwick Economics Research Papers Price Comparison Websites. Warwick Economics Research Papers,

2015(October 2015). Retrieved from

https://warwick.ac.uk/fac/soc/economics/research/workingpapers/2015/twerp_1056b_ronay ne.pdf

Schlosser, R., & Boissier, M. (2018). Dealing with the dimensionality curse in dynamic pricing competition: Using frequent repricing to compensate imperfect market anticipations.

Computers and Operations Research, 100, 26–42. https://doi.org/10.1016/j.cor.2018.07.011 Steinker, S., Hoberg, K., & Thonemann, U. W. (2017). The Value of Weather Information for

(43)

43 Weiss, R. M., & Mehrotra, A. K. (2001). on the Internet and the Future of E-Commerce T H E L

E S S O N S O F H I S T O RY : A I R L I N E I N D U S T RY. New York, (May). Weisstein, F. L., Monroe, K. B., & Kukar-Kinney, M. (2013). Effects of price framing on

consumers’ perceptions of online dynamic pricing practices. Journal of the Academy of Marketing Science, 41(5), 501–514. https://doi.org/10.1007/s11747-013-0330-0

Willart, S. P. C. (2015). Price competition in retailing: The importance of the price density function. Journal of Retailing and Consumer Services, 26, 125–132.

(44)

44

Appendices

Appendix A

Product category Explanation

cat1 hooks and hangers

cat2 Bike accessories

cat3 bicycle saddle

cat4 kitchen appliances (electronic)

cat5 Storage

cat6 Swimming pool accessories

cat7 bicycle pumps

cat8 Batteries

cat9 Garden

cat10 Backpacks

cat11 Weather stations

cat12 Heaters

cat13 Toys

cat14 Cool boxes

cat15 Binoculars

cat16 Hobby

cat17 Clocks

cat18 Tableware

cat19 Telescope accessories

cat20 Microscopes

cat21 Brands

cat22 Cutting boards

cat23 Sport

cat24 DIY jobs

cat25 Exterior lighting

cat26 Airconditioning

cat27 Interior lighting

cat28 Globes

cat29 Spare filters

cat30 Knives

cat31 Thermometers

cat32 Swimming pools

cat33 Table textiles

cat34 Wine

cat35 Hammocks

(45)

45

cat37 Dehumidifiers

cat38 Furniture

(46)

46

Dynamic pricing in online retail: Yay or Nay?