Investigation of Online Auctions and Price Determination

(1)

Investigation of Online Auctions and

Price Determination

Juan I. Gesino

11604816

University of Amsterdam

Faculty of Economics and Business

Bachelor Thesis

(2)

Statement of Originality

This document is written by Juan I. Gesino who declares to take full responsibility for the contents of this document.

I declare that the text and the work presented in this document are original and that no sources other than those mentioned in the text and its references have been used in creating it.

The Faculty of Economics and Business is responsible solely for the supervision of com-pletion of the work, not for the contents.

(3)

UNIVERSITY OF AMSTERDAM

Abstract

Faculty of Economics and Business Bachelor Thesis

by Juan I. Gesino

Auctions are a widely accepted method for attaining high efficiency in markets as goods are sold to the bidder with the highest value. Governments all over the world have been using auctions to ensure an efficient allocation of electromagnetic spectrum transmission rights and online auctions have been around for decades. This research studies the results from more than 170 thousand auctions of a popular online auctions website in Europe to try to determine the effect of two common bidding strategies: sniping and jump bidding. The results show that both of these patterns have a significant positive effect on the final price of the auction, meaning that, on average, they tend to raise prices. This research contributes to the already existing literature about online auctions and auctions in general by analysing a substantial amount of auction data from a website that has not been studied before.

(4)

Introduction

Auctions are processes by which a seller offers a good for sale and bidders compete among each other to offer the most attractive price. Such offers are referred to as bids. There are a variety of different types of auctions, each one differs in the rules that need to be followed by both the auctioneer and the bidders, the amount of information that is revealed and the conditions for the final purchase. One particularly appealing feature of auctions is their ability to match sellers with the buyers that are willing to pay the highest amount, even in the case when sellers do not know the precise value (Krishna,

2009). This characteristic is particularly attractive in the case of goods that do not have a specific price in the market, such as art and collectables, but also when buying involves high risk such as natural resources extraction licenses (P. R. Milgrom & Weber,1982). One of the most common types of auctions is the open ascending auction. The main characteristic of these auctions is that the auctioneer announces a starting bid, and bid-ders place increasingly higher bids that are publicly known to all participants (Krishna,

2009). In theory, this type of auction allows participants to use other participants’ bids as signals to make their own decisions. More precisely, at any point in time, all bid-ders know the other bidbid-ders’ willingness to pay. As reported by Krishna (2009), the remarkable result from this mechanism is that neither the winner nor the losers of the auction regret the outcome. The winner knew precisely the valuation of all the other bidders and was willing to place a bid higher than the rest. For the losers, they have the information about the willingness of the ultimate winner but decided not to outbid. A prime example of these auctions is those used for livestock. In this $1.4 trillion industry (Thornton,2010), auctions are used for buying and selling cattle, where the price of each animal is difficult to determine, and the level of uncertainty is high. Understanding the mechanics of these auctions is a relevant aspect in determining the expected revenue for cattle producers as well as the downstream producers and sellers of all the related

(6)

5

products in the complex supply chain they represent (Mintert et al., 1990). Another prevailing example of contemporary auctions is electromagnetic spectrum transmission rights. The Federal Communications Commission (FCC) in the United States carries out such auctions where companies compete among each other to buy the exclusive rights of transmission at a particular frequency. The mechanism is similar in other countries with other governmental agencies such as the Agentschap Telecom in The Netherlands and the Bundesnetzagentur in Germany. In these cases, auctions are used for their ability to efficiently allocate extremely scarce resources, giving the spectrum to those companies that value them the most (Cramton et al., 2002). Lastly, other increasingly relevant auctions are online auction websites. These websites become popular with the start of the internet during the late 90s. Arguably the first auctions website, Onsale, was mainly focused on selling collectable items using online auctions (Lewis,1995;Lucking-Reiley,

2000). This application is similar to Catawiki’s use-case, where auctions are used to sell unique items such as collectables. The advantage for these items is that auctions are used to determine market prices for items that are not readily available to the public. These items can often be challenging to set a price for, but by using an auction, the sellers let all the potential buyers express the amount they would be willing to pay. However, sites that came later than Onsale, such as eBay and Amazon, sold all types of products from unique and collectable products to consumer goods such as electronics and fashion. For these items, auctions work as a way to ensure the most efficient price. By letting bidders suggest the price they are willing to pay for the items, auctions are intended to be efficient; only resulting in a transaction for the price that best suits both parties.

As stated byCramton(1998), finding a way to judge each type of auction and being able to determine the best design for each setting is not only an exceptional ordeal but one that is immensely useful and needed. The problem with comparing different auctions empirically is the lack of comparable scenarios in which different types of auctions are used, but the auctions remain closely similar in order to be compared. From a motivation point of view, determining the factors that influence different kinds of auctions helps the designers implement the most efficient type for each situation.

Understanding auctions not only can help understand the development of the markets previously described but serves as a testing ground for investigating human behaviour in market settings. As such, research into auction theory tries to determine the efficiency of auction markets as well as bidding strategies. This process includes understanding the determinants of an auction’s final price. The purpose of this research is to explore the data collected from a popular European auctions site, Catawiki, and attempt to create a model to help understand how a final price is determined. This model will take into consideration all theoretical aspects of price determination to the extent that they are

(7)

6

available in data, as well as some further data that was collected from the website. The data collected represents a total of around 170 thousand auctions, including more than 2 million individual bids. Although there is already extensive literature about auctions and, more generally, online auctions, this study represents a significant addition to the literature by analysing a completely new website. Furthermore, the amount of data collected represents a unique opportunity to analyse different aspects of the auctions published with various granularities and dimensions.

This research paper continues with a literature review and a theoretical framework in

chapter 2, an explanation of the methodology inchapter 3, the presentation of the results and a discussion about them in chapter 4and finally a conclusion inchapter 5.

(8)

Chapter 2

Theoretical Framework

Auctions have been around for a long time. It is estimated that auctions took place in ancient Greece and Rome, some sources placing them in 2 BC (Andreau, 2018), while other sources place the first auctions even before that, in 500 BC (Krishna,2009). This does not come as a surprise, as stated before, auctions are an incredibly useful instrument for price determination of unique goods. Before globalisation and the widespread of industrialisation, products were mostly handcrafted, and there was less homogeneity among them. In today’s context, auctions are still a big part of our societies, especially considering the impact of spectrum auctions since the beginning of the 90s. In 2019 alone, the U.S. Treasury has collected $ 2,7 billion in spectrum auctions for 5,880 licenses (Auctions Summary, 2019). The use of them has not changed much; however, the operationalisation has shifted to make use of the latest technologies. With the expansion of the internet, auctions become not only more accessible for buyers, but also more attractive for sellers, as the increased participation meant more opportunities to sell goods. As stated by Bajari & Horta¸csu (2004) online auctions sites, such as eBay, have seen considerable growth during the late 90s and early 2000s. This growth has been mainly driven by collectables, antiques, and unique items. The explosion of online auctions has not only benefited buyers and sellers but has opened the door to more and better research into auctions and how people behave during these auctions. Recent research has been exploiting the available information generated by these online websites to explore how different factors and characteristics of online auctions affect the final price. One fundamental aspect to take into account when analysing online auctions is the set of rules and definitions the different websites have. Generally speaking, most online auctions websites can be described as open ascending or English auctions. In English auctions, the auctioneer starts the auction with a low bid and continuously increases it while bidders “drop out” of the auction until the last bidder willing to pay the price

(9)

8

remains. The last bidder remaining is the winner of the auction and the one that will purchase the product (P. Milgrom & Weber, 1982). In online auctions, a variation is used, and instead of the auctioneer raising the bid, bidders compete among each other by placing increasingly higher bids (Lucking-Reiley,2000; Krishna, 2009). The bidder with the highest bid by the end of the auction is the one that purchases the product. In these auctions, during the entire process, all bidding information is visible to all bidders, meaning that everyone can see what others are bidding and how they are responding to each bid (Krishna, 2009). Consequentially, bidders can use this information as part of the information they take into account when placing bids. In contrast to sealed-bid auctions, where sealed-bidders can only make use of their private information to place a bid, English auctions allow a further degree of information aggregation. Furthermore, as outlined by P. Milgrom & Weber (1982), bidders are rewarded for all additional information they possess in what is known as the linkage principle. This suggests that an open ascending auction should lead to more efficient and higher revenue results, as participants can use each others’ bids as part of their information set.

However, one particular concern raised by economists of auction theory is the finite nature of auctions online (Ariely et al., 2005; Lucking-Reiley, 2000). For example, in eBay, auctions have a fixed time limit in which bidders can place their bids. In contrast, in Amazon and Catawiki, auctions get extended if a bid is placed close to the end. When auctions have a fixed limit of time, bidders have an incentive to place their bids as close to the end of the auction as possible. This pattern is usually referred to as ’sniping’. Sniping can result in a lack of information aggregation as bids are not placed until the very end of the auction, and thus signals are not sent. Ariely et al. (2005) designed an experimental setting and investigated the effect that different closing rules have on how people placed bids. In their research, the authors found that auctions in which the time gets extended when a bid is placed close to the end, see less activity at the end of the auction compared to those in which the auction closes immediately at the specified time. Their result is consistent with what is expected from a game-theoretic perspective, as fixed-closing auction incentivises late bidding. At the same time, they have shown that auctions without a fixed closing time result in higher efficiency and revenues. The underlying reason is that as bidders are not incentivised to bid late, they place their bids as time goes by, sending signals to other participants that result in more information aggregation.

Another type of pattern observed in auctions is jump biddings. Jump bidding occurs when a bidder submits a bid that is significantly higher than the amount they need to bid to have the current winning bid. Submitting one of these bids results in a possibly costly strategy as the bid can overshoot the required bid to win the auction and purchase the product. As an example, consider an auction with two participants, A and B,

(10)

9

whose valuation for the product that is being auctioned are $50 and $100 respectively. Implementing a jump bidding strategy might cause participant B to place a bid between $50 and $100. However, this is not optimal as participant A will not bid higher than $50. Given this example, a priori, it might seem that jump bidding is not an adequate strategy to use during an auction. Nevertheless, Avery (1998) has shown how jump bidding can constitute a rational bidding strategy. The explanation for this strategy is that jump bids can be interpreted as signals given by the bidder to indicate that they are willing to undergo aggressive bidding to win the auction. Easley & Tenorio(2004) have studied jump bidding empirically by looking at two different auctions websites and found a substantial number of jump bids as well as an inverse relation between the number of jump bids and the bids.

Furthermore, participants of auctions are susceptible to a competition effect, in which each bidder is competing with all other bidders. This mindset causes different effects that might influence the final price of the auction, such as auction fevers and endowment effects. As described by Heyman et al.(2004), increased competition in online auction websites can cause psychological effects on bidders where they might have a sense of “winning” and “losing” that is tied to a quasi-endowment effect. In their paper, the authors demonstrated how increased competition has a positive effect on the final price of the auction. The quasi-endowment effect takes place as participants that outbid the previous highest bid are endowed with the good momentarily. During this period, the participant has the illusion of owning the product, which in turn increases the valuation for it. If they are outbid, they will be willing to bid again even if this results in a price that is higher than the initial valuation for the product (Knetsch,1989).

During this investigation, all these patterns and behaviours will be incorporated into the final model to try to estimate their effect and establish the role they play in the determination of the final price.

(11)

Chapter 3

Methodology

For the purpose of this research, information about online auctions from the auctions website Catawiki were collected. Catawiki is an online auctions website with a particular focus on unique and collectable items1. This makes the visitors of the website substan-tially different from other auctions websites, such as Amazon or eBay, in the sense that they might be more professional and experienced. At the same time, the items that are up for auction in this website are usually not consumer goods, but unique objects whose exact price is challenging to determine. One of the main selling points of Catawiki is their “experts”. These are people that curate auctions before they are published and work side-by-side with the sellers making sure their products are well published and have accurate descriptions.

3.1 Data Collection

Understanding how data is structured by Catawiki is relevant for understanding the main concepts and definitions. The most relevant note is that in Catawiki, auctions are called “lots”, and these lots are grouped according to a common theme into “auctions”. The auctions are curated by experts who work together with the sellers of each lot to make sure the information provided to the buyers is the most accurate and useful. At the same time, auctions are grouped into categories and subcategories that share the same themes, such as watches, cars, coins, among others.

The data collected represents a sample of auctions and lots together with their corre-sponding bids and categories published between February 28th and June 7th of 2020. This represents a total of 7,640 auctions containing a total of 173,513 lots with 2,599,385

1_{More information about Catawiki can be found in}_{Appendix C}

(12)

11

bids. The data was collected by periodically extracting information from Catawiki’s publicly accessible API endpoints.

Additional information about the data collection process is provided in Appendix A.

3.2 Data Exploration

As a first step to understanding the data collected, an Exploratory Data Analysis (EDA) was conducted. The goal of this EDA is to understand the different variables obtained as well as determine which variables might hold relevant information about the auction’s final price. As an initial approach, Table 3.1 shows general summary statistics for the final price of auctions, the number of bids, the starting bid and the number of bidders. The table shows a high variation for the auction’s final price, with prices ranging from AC1 toAC158,000. This variation can also be seen in the initial bid of auctions, with more than 50% of auctions starting at AC1, but a maximum ofAC95,000.

Price N Bids Start Bid Avg. Bid N Bidders N 176862 186852 176862 176862 186852 Mean 280.37 12.84 53.04 152.37 5.56 SD 1491.14 8.79 543.15 861.88 3.39 Median 98.00 12.00 1.00 48.25 5.00 Min 1.00 0.00 1.00 1.00 0.00 Max 158000.00 75.00 95000.00 97500.00 38.00

Table 3.1: General summary statistics

3.2.1 Categories

In Catawiki, all auctions, and therefore, auctions’ lots, are sorted into categories. These categories group together similar auctions in terms of the items sold. Solely by the names of the categories, it is clear that the focus of Catawiki is collectables and antiques as seen in Table 3.2. The table also shows some underlying characteristics from each category. An example of these is the relatively high final price for cars and motorcycles compared to other categories. However, the most salient aspect is the similarity among categories in terms of the number of bids and the number of bidders. In terms of the number of lots in each category, Figure 3.1 shows the most popular are “Interiors & Decorations”, “Jewellery & Watches”, “Art”, and “Coins Bullion & Stamps”. These categories together represent roughly half of all the lots in the sample.

Additionally, the way the final price of each lot in each category is distributed can be seen in a box plot as the one depicted inFigure 3.2. For visualization purposes, the box plot only displays outliers for a final price value lower than 1000. Even when placing

(13)

12

N Mean Price SD Price Mean N Bids SD N Bids Mean Start Bid Avg. Bid Mean N Bidders Mean Category

Archaeology & Natural History 7775 178.17 557.61 10.74 7.05 32.03 96.75 4.94 Art 19073 196.98 683.92 14.26 9.74 34.05 103.66 5.34 Asian Art & Ethnography 6525 182.79 378.32 12.82 9.25 36.41 98.69 5.02 Books & Comics 11309 125.97 303.11 14.31 8.07 15.68 63.44 5.66 Classic Cars, Motorcycles & Automobilia 3618 2486.13 8904.15 12.25 8.19 443.28 1342.86 5.50 Coins, Bullion & Stamps 18944 136.97 225.57 12.17 7.42 36.36 82.28 5.47 Computers, Games & Cameras 3775 106.95 177.08 11.45 7.75 26.94 61.30 5.12 Diamonds & Gemstones 10584 502.73 2069.71 10.46 9.43 99.12 269.62 4.61 Fashion 7524 212.01 555.42 12.16 8.76 36.73 113.09 5.67 Interiors & Decorations 34569 149.83 287.39 11.86 8.68 27.76 80.96 4.94 Jewellery & Watches 34535 483.66 1063.94 13.27 9.99 94.41 263.59 6.15 Militaria & Weaponry 3293 179.45 409.56 11.89 7.75 40.85 102.01 5.41 Music 3736 151.03 306.67 14.24 8.09 30.71 83.28 6.41 Sports & Events 957 206.42 551.24 15.36 8.58 13.22 93.10 6.62 Toys & Models 9523 83.61 113.36 13.90 6.86 8.52 41.96 6.19 Wine & Whisky 11055 179.36 351.09 15.31 7.33 32.02 102.34 6.90

Table 3.2: Summary statistics grouped by category

Figure 3.1: Number of lots per category

this restriction, it can be seen that the distribution is extremely skewed and presents a large number of outliers. The interpretation for this is that even inside these categories, final prices have high variations.

3.2.2 Auctions’ Duration

Another relevant characteristic is the duration of the auctions themselves. This informa-tion is not revelling of any characteristic about the aucinforma-tion by itself, but it can be used to give perspective to other metrics. At the same time, it is necessary to understand if there are auctions that are systemically lasting longer than others.

FromTable 3.3 it can be seen that the average duration of an auction is around seven days, which is in line with Catawiki’s guidelines. It can also be seen that there are a

(14)

13

Figure 3.2: Box plot showing the distribution of the lot price per category

Duration N 186852 Mean 10922.93 SD 3311.45 Min 3301.00 p25 8763.75 Median 10606.50 p75 13421.08 Max 30677.50

Table 3.3: Summary statistics of auctions’ duration in minutes

few auctions longer than that, with only 25% having a duration longer than nine days and the maximum close to 21 days.

From Table 3.4 it can be concluded that the differences in auction duration are not arising from differences in how categories are instrumented. It is worth mentioning that the Classic Cars, Motorcycles & Automobilia category does seem to have a higher auction duration, with an average of around nine days.

3.2.3 Bids Time Delta

To further understand how these auctions are taking place, it is worthy to identify the time differences, or time delta, between two subsequent bids. This value can have multiple interpretations, but the clearest one is that it describes the level of competition

(15)

14

Mean Duration SD Duration Median Duration Category

Archaeology & Natural History 9924.26 2629.83 9191.57

Art 10750.04 2857.31 10605.92

Asian Art & Ethnography 9277.19 2747.79 9129.70

Books & Comics 10919.34 3289.02 10612.33

Classic Cars, Motorcycles & Automobilia 14147.95 3498.84 14833.75

Coins, Bullion & Stamps 10615.95 2931.89 10595.58

Computers, Games & Cameras 9467.95 2273.98 9178.42

Diamonds & Gemstones 9780.78 3137.88 9164.33

Fashion 12248.00 3287.23 12037.65

Interiors & Decorations 12785.62 3802.52 12073.50

Jewellery & Watches 9858.13 2864.11 10511.83

Militaria & Weaponry 10256.88 2312.02 10573.50

Music 9198.58 2016.47 9151.75

Sports & Events 9950.44 3346.92 10563.37

Toys & Models 11078.37 2902.21 10621.50

Wine & Whisky 11322.37 2549.38 11670.38

Table 3.4: Summary statistics of auctions’ duration in minutes grouped by category

and activity between the bidders. If bids are being placed shortly after each other, this might indicate that bidders are incredibly interested in the product and they are willing to pay the costs of placing bids, such as stop what they are doing to go to the website and place the bid.

Max. Time Delta Min. Time Delta Mean Time Delta Median Time Delta SD Time Delta count 165724 165724 165724 165724 165724 mean 4154.55 40.89 847.78 305.22 1396.21 std 2598.72 274.13 799.89 744.15 1075.93 min 0.07 0.00 0.05 0.00 0.00 25% 2369.36 0.00 384.68 0.50 717.42 50% 3694.14 0.00 622.76 34.50 1134.40 75% 5546.37 0.15 1022.01 241.32 1753.14 max 37435.47 9731.67 13547.06 12428.26 20749.97

Table 3.5: Summary statistics of bids’ time differences in minutes

Table 3.5 illustrates how much time goes by between bids and Figure 3.3 depicts the average time difference distribution. Having the context provided in subsection 3.2.2, it is clear that the range of time differences for these bids is remarkably narrow with a mean delta of about 14 hours and most bids being placed even faster than that (mean median of 5 hours).

3.2.4 Sniping

Given that the last highest bidder can purchase the good that is up for auction, there are incentives for bidders to concentrate activity at the end of the auction. For this reason, it is useful to understand behavior at the end of the auction. To accomplish

(16)

15

Figure 3.3: Distribution of mean time differences between bids in days

this, the number of bids and bidders are recorded for the last 30%, 20%, 10%, and 5% of the auction time in Tables3.6,3.7,3.8,3.9respectively. The tables also include these numbers as a percentage of the final amount. It can be seen that on average, more than 50% of all the bids placed during the auction time, are placed in the last 30% of the auction. Considering from subsection 3.2.2 that the average auction lasts around seven days, this means that, on average, most bids are placed in the last two days of the auction. Considering the last 5% of the auction, which on average represents the last 8 hours of the auction, it can be seen that an average of 31% of the bids are placed.

N Bids N Bidders Bids per Bidder N Bids Percentage N Bidders Percentage

N 186852 186852 186852 176862 176862 Mean 6.46 2.99 1.84 0.51 0.56 SD 5.92 2.20 1.49 0.31 0.29 Min 0.00 0.00 0.00 0.00 0.00 p25 2.00 1.00 1.00 0.26 0.33 Median 5.00 3.00 1.50 0.50 0.56 p75 10.00 4.00 2.50 0.77 0.75 Max 55.00 21.00 34.00 1.00 1.00

Table 3.6: Summary statistics of last 30% of auction time

3.2.5 Jump Bidding

Jump bidding is not something that can be directly measured using bidding data due to the fact that it is not exactly clear when a bid should be considered a jump bid and when it is simply a normal bid. However, this research operationalizes the degree

(17)

16

of jump bidding by analyzing the bid increments between subsequent bids. Table 3.10

shows summary statistics for the bid increments found across all auctions which helps to understand the degree of jump bidding in the sample. Conversely, the table shows that across bidding from all the auctions in the sample, the average of the maximum increment is AC70 and a maximum increment of AC63,0002. Interestingly, there is not a great amount of variation among the mean increments of auctions. The maximum mean increment was ofAC9,700 but the average was at AC24 with a 75th percentile at AC15.

2

The maximum bid increment corresponds to an auction for a Ferrari 512 BBi Coupe 4.9L V12 from 1984. The jump bid was placed almost two days after the last bid ofAC87 thousand, raising the winning bid toAC150 thousand. This was the only bid placed by that bidder and the car was ultimately sold for AC154 thousand a week later.

(18)

17

Max Increment Min Increment Mean Increment Median Increment SD of Increments N 170928 170928 170928 170928 165724 Mean 70.55 7.99 24.29 18.14 21.79 SD 418.90 52.34 116.92 88.83 131.51 Min 1.00 0.00 0.67 1.00 0.00 p25 9.00 1.00 4.21 5.00 2.15 Median 18.00 1.00 6.47 5.00 4.24 p75 47.00 5.00 15.31 10.00 12.57 Max 63000.00 5000.00 9714.29 7499.50 13161.99

Table 3.10: Summary statistics of bids increments

3.3 Data Model

The main goal of this research is to determine patterns, if any, that might affect an auction’s final price. Therefore, a model that describes the final price of the auction is needed. The model implemented is a linear regression model that takes into account all variables described before as independent variables and the final price as the depen-dant variable. Additionally, some variables are used as a control to avoid undesired fluctuations such as those observed by categories of lots that have inherently high final prices.

Formally, the model introduced for this analysis can be described as:

Final Pricei= β0+ β1Xi+ γCi+ ε (3.1)

This model will expose the effect that each of these variables has on the final price of auctions. InTable 3.11 a full list of all variables used and their type can be seen.

Variable Name Description Variable Type

Final Bid The monetary value of the final bid in Euros Dependant Start Bid The monetary value of the first bid submitted in Euros Independent Total Time The total number of days that the auction lasted Independent Start Bid Time The percentage of the auction time that went by before the first bid was placed Independent Final Bid Time The percentage of the auction time that went by before the last bid was placed Independent Number of Bids The total number of bids submitted for the auction Independent Total Bidders The total number of unique participants in the auction Independent Average Bid Increments The average increment between consecutive bids in Euros Independent Lot Mean Final Price The average final price among all auctions inside a lot Control Reserve Price Information about whether the auctions had a reserve price (minimum price) Control Category The Catawiki category for the auction Control

(19)

Chapter 4

Results & Discussion

To estimate the linear regression model introduced in Equation 3.1, an Ordinary Least Squares (OLS) was chosen, including all the variables described inTable 3.11. The result of the OLS estimates can be seen inTable 4.1. The table shows all the variables regress against the final price of each auction. The overall F-statistic shows that the model is significant and the R2 _{of 0.743 indicates a good fit with the sample data. To discard any}

perfect multicollinearity problems between the variables used in the model, a correlation analysis was performed in Appendix B. As each auction belongs to a particular group of other auctions3, and these auctions are related to each other, the errors are cluster to avoid any biases. Furthermore, a category variable was added to control for any effects between categories. In terms of the number of observations, the regression estimation was run with a total of 170,876 auctions. This sample size represents a small drop from the original 173,513 lots in the sample since some auctions are never completed or even do not receive any bids at all. In general terms, the regression appears to be an acceptable specification.

The analysis will be separated among the different effects that are expected to be seen according to what was discussed in chapter 2. The first analysis will be centred on the effect of time and auction ending rules on the final price (sniping), followed by an analysis of the effect of jump bidding.

3

Auctions are referred to as “lots” in Catawiki and auctions are groups of these lots. Auctions are constructed by grouping together lots of similar characteristics. For example, an auction could be named “Exclusive Classic Car Auction” containing 27 different types of exclusive classic cars.

(20)

19

Dep. Variable: Final Bid

Model: OLS

Method: Least Squares R-squared: 0.743 Adj. R-squared: 0.743 F-statistic: 1265. Prob (F-statistic): 0.00 No. Observations: 170876 Df Residuals: 170851 Df Model: 24 Coefficient Start Bid 0.548∗∗∗ (0.069)

Total Time (Days) −3.572∗∗∗

(0.937)

Start Bid Time 15.571

(15.769)

Final Bid Time 163.372∗∗∗

(24.862)

Number of Bids 14.687∗∗∗

(0.796)

Total Bidders 5.955∗∗∗

(1.630)

Average Bid Increments 6.675∗∗∗

(0.530)

Lot Mean Final Bid 0.479∗∗∗

(0.039)

Reserve Price Set 35.310∗

(13.743)

Constant −400.542∗∗∗

(29.992)

Notes:

[1] Standard Errors in parenthesis are robust to cluster correlation at the auction level [2] The regression includes control variables for all auction categories

[3] * p < 0.05, ** p < 0.01, *** p < 0.001

Table 4.1: Regression model results

4.1 Sniping

As discussed in chapter 2, sniping is the strategy of placing bids as close to the closing time of the auction as possible with the objective to outbid the highest bidder without giving them time to respond. This strategy is mostly seen in auctions where there is a strict close rule (Ariely et al.,2005). However, Catawiki implements a rule to avoid the incentive for sniping. If a bid is placed during the last minute of the auction, an extra 1.5 minutes are added to the auction time. The goal of this is to prevent people from sniping. Looking at the results observed inTable 4.1, we can analyse the effect of sniping by looking at the Final Bid Time coefficient. This coefficient captures the effect that placing the last bid close to the end of the auction time has on the final price. This value is represented as a percentage and corresponds to the percentage of the auction time that has went by before the last bid was placed. The coefficient is significant at a

(21)

20

0.1% significance level and has a positive effect. This can be interpreted as having the last bid near the end of the auction will yield a final price higher than auctions where the last bid is placed closer to the start. As an example, if all auctions last 10 minutes and auction A’s last bid was placed 9 minutes after the start, while auction B’s last bid was placed 8 minutes after the start, their Final Bid Time will be 90% and 80% respectively. Following this example, and according to the estimated model, auction A will have, on average, a final price that isAC16.38 higher than auction B, with all other factors remaining equal. To generalise, for each 10% that the last bid is placed close to the end, the auction price rises AC16.38 if all other factors remain equal.

The result observed comes as a surprise considering Catawiki’s rules for automatically extending the auction. It would be expected that this rule reduces the incentives to place bids at the end and thus making this effect irrelevant. However, the sign of the coefficient seems to be aligned with what is expected. Conditional on finding an effect, this effect is expected to be positive. Meaning that prices go up if there is sniping. A possible explanation for this finding is that Catawiki’s rules are not strong enough to reduce the incentives for sniping. It might be the case that a 1.5-minute increase is not enough to keep bidders from trying to place bids at the end of the auction. From a design perspective, the amount of time that Catawiki should add to the auction has to be enough to convince bidders that other bidders will be able to react. At the same time, because the time only gets extended when a bid is placed during the last 60 minutes of the auction, placing a bid 61 minutes before the end will not cause an extension. This gives the other bidders only 1 minute to respond. The ending rules on Amazon, as described in Ariely et al. (2005), are much more strict and provide more substantial incentives against sniping. As described in the paper, Amazon closes the auction whenever 10 minutes have gone by since the last bid.

All in all, given these results, it might not be in Catawiki’s interest to modify the rules. The results show that the final price increases when there is a higher amount of sniping. Given that Catawiki charges a commission that is a percentage of the final price, it is in their interest to make prices as high as possible.

4.2 Jump Bidding

The effect of jump bidding can be determined by analysing coefficient for Average Bid Increments. This coefficient captures the increments in bidding from one bid to the next. Jump bidding occurs when a bid is placed that is remarkably higher than the previous one. In general terms, to win an auction, a bidder needs to bid the smallest increment possible on top of the current bid. However, in practice, it is observed that

(22)

21

bidders sometimes place bids that correspond to a higher-than-needed increment with respect to the previous bid (Easley & Tenorio, 2004; Avery, 1998). Taking this into account, if jump bidding is present in the sample, it is expected to have a positive effect on the final price. The regression results show that an increase ofAC1 on the average bid increments will cause aAC6.67 increase in the final price. This result is remarkable, not only because the effect is significant to a 0.01% significance level, but because of how strong the effect is. Going back to Table 3.10, we can see that the potential effect that jump bidding can have on the final price of the auction is exceptionally relevant. On average, auctions in the sample had a mean increment ofAC24.29. This means that with everything else remaining equal, the effect of jump bidding can be on average aAC162.01 contribution towards the final price. Comparing this value with the average final price from Table 3.1gives a 57.86% potential contribution from jump bidding.

Given that the effect of jump bidding can potentially vary across different categories, several regressions for each of the categories were estimated. Figure 4.1shows the value of the coefficients for each regression normalised to the category’s average final price. Although the coefficients themselves show the marginal contribution in the presence of jump bidding to the final price, the coefficients were normalised to allow for better comparison across categories. For example, aAC6 marginal contribution to the final price is not the same on a AC81,000 Lamborghini Diablo from 1991 compared to a AC320 FC Barcelona jersey signed by Lionel Messi. The figure shows that the effect jump bidding has on the final price is higher for Toys & Models, Sports & Events, Art, and Books & Comics. Although it is difficult to draw a definitive conclusion for these results, they can be explained by the experience and sophistication of bidders in each category. Bidders for the previously mentioned categories might be seen as less experienced and more likely to be involved in na¨ıve jump bidding. In contrast, bidders in the categories of Classic Cars, Motorcycles & Automobilia, Jewellery & Watches, and Diamonds & Gemstones might be more sophisticated, experienced and less likely to place higher-than-needed bids irrationally. However, as mentioned before, there exists a degree of jump bidding that can be the consequence of strategic thinking as developed by Avery

(1998). Therefore, further investigation is needed to determine the experience of the bidders in each category and be able to draw a definitive conclusion.

4.3 Limitations

As with any empirical research, this research has a few limitations that might be the starting point for more robust further research. The amount of data available gives a remarkable opportunity to test patterns observed in auctions and even find new ones. In

(23)

22

Figure 4.1: Jump bidding coefficients for categories regressions normalised to the average lot price. Error bars show the 95% confidence interval for the estimated

coef-ficients

this paper, two of the main effects discussed in the literature were analysed: sniping and jump bidding. However, other patterns, such as quasi-endowment effects and the role of competition are other opportunities that should be further explored. Furthermore, given that the data contains information about the different categories and subcategories for each auction, understanding the differences between these can lead to interesting findings. In terms of the data, one limitation is the lack of information regarding the item’s actual value which can be used to determine profit and efficiency effects. For unique and collectables, such as the ones sold in Catawiki, this is not always possible. For jump bidding specifically, this research cannot determine if the effect found cor-responds to na¨ıve or strategic jump bidding. An attempt to capture this effect was conducted by comparing categories for which bidders are assumed to have different lev-els of expertise. However, being able to track bidders across auctions would be the solution to be able to determine if a bidder is a professional or not. As this information was not present in the data analysed in this research, the question remains open. All

(24)

23

in all, given this sample, there is no current way of controlling for the experience of the bidders.

Finally, a limitation of the presented setup is that some level of endogeneity might be present. From this data, it is not particularly evident that jump bidding causes prices to be higher or if high-priced items cause bidders to place jump bids. At most, what was demonstrated in this research is the direction and the presence of a relation between both sniping and jump bidding. Further research into Catawiki could focus on disentangling these effects. Furthermore, an experimental setting could provide insights about the extent of endogeneity in these two patters.

(25)

Chapter 5

Conclusion

Auctions are a big part of our society, from livestock to antiques, to spectrum trans-mission rights; auctions are a remarkable way of achieving high efficiency. By letting bidders communicate the price they are willing to pay to purchase the good, the goal of auctions is to allow transactions where the buyer is the party that is willing to pay the most. The efficient allocation is a fundamental aspect of why auctions are used in markets of heterogeneous goods. An example of this is Catawiki’s target market of unique and collectable items. During the process of this research, data was collected on more than 170 thousand auctions and aggregating data from more than 2 million bids from Catawiki. These data were used to estimate a linear model that attempts to explain how different factors affect the final price of auctions.

The goal of this research is to gather insights toward possible ways in which the final price of an auction can be determined. More precisely, this research analyses two widespread bidding patterns, sniping and jump bidding, and tries to determine the effect they have on the final price. On the one hand, sniping is a technique where the bidder places their bid close enough to the end of the auction to prevent other bidders from reacting in time. On the other hand, jump bidding is a pattern in which the bidder places a substantially higher bid than the previous winning bid. It has been shown that jump bidding can occur in both settings: a na¨ıve inefficient behaviour or a rational strategy (Avery, 1998). The linear model estimated during this research was used to capture the effects of the two patterns on the final price while also adding more variables as controls. The results show that the model is a good fit for the data and the effects are both significant at a 0.01% significance level. In the case of sniping, the effect is positive, which means that the closer the final bid is to the ending time, the higher the final price will be. This is in line with previous research on online auctions (Ariely et al., 2005). However, the presence of sniping itself is surprising, given the automatic extension rule

(26)

25

that Catawiki has. This result might be an indication of an ill-defined rule, aimed at reducing sniping but not succeeding completely. In terms of jump bidding, this research finds significant evidence that high levels of jump bidding are related to a higher price in the auction. Furthermore, the data shows that these effects are not homogenous across different categories of products, which might be an indication of bidders’ experience and sophistication.

All in all, this research attempts to contribute to the existing literature by analysing a relevant number of auctions from a website that has not been analysed before. There remains an incredible potential for analysing the data and understanding other bidding patterns from it. At the same time, conclusions observed from this research can be further validated using laboratory experiments to improve robustness and provide a definitive answer.

(27)

Appendix A

Data Collection & Storage

This Appendix outlines the process used to collect and store the data used during this research. The goal for this setup was to provide a reliable, scalable and robust system that could be extended to other data collection tasks.

A.1 Infrastructure

The ephemerous and state-less nature of the tasks that had to be executed suggested that a prime candidate was to use a serverless paradigm. This approach also had attractive benefits like ease of deployment, low costs, and scalability. The entire infrastructure was designed to be compatible with all kinds of serverless vendors, but for different reasons, Amazon Web Services (AWS) was used for this specific case. In AWS’ infrastructure serverless compute nodes are referred to as “lambdas”. For the rest of this section, the word lambda will be used to describe this state-less compute nodes.

The approach taken to collect the data was to divide all the data collection tasks that needed to be executed in small pieces that are not dependent on each other. The motivation was to be able to perform tasks in isolation by separation of concerns while at the same time, allowing for parallel computing. The logic to break these tasks apart was to divide them into two factors: entities and life cycle. In this way, tasks associated with different entities should execute in different lambdas, and tasks associated with different stages of the same entity should also be executed in different lambdas.

Each lambda was triggered on schedule with a frequency that was fined tunned to achieve optimal results, with a trade-off in compute time and number of calls. The final result consisted of seven different tasks distributed in seven different lambdas. Frequencies ranged from seven days for the least critical tasks (e.g., updating the list of categories)

(28)

27

to 30 minutes for the most critical ones (e.g., getting the bids). All of these tasks wrote the information obtained to the database (seeA.3).

Figure A.1shows a diagram for the previously described infrastructure, depicting Cloud-Watch as the coordinator and the different lambdas as the distributed computation nodes writing to MongoDB.

A.2 Software

The programming language chosen for this task was JavaScript (Flanagan, 2006) us-ing the Nodejs runtime version 12.16.1, which is the current long-term support (LTS) version. The motivation for this choice was the compatibility between JavaScript and available external libraries for crawling and scrapping websites, given that JavaScript is the language used by web browsers. Whenever available, the extraction of information was done by taking advantage of Catawiki’s internal application programming interfaces (APIs). These APIs are the same used by Catawiki’s own website, but they present two immense advantages compared to scrapping and crawling: reliability and more data. In terms of reliability, although internal APIs can change from time to time, these changes tend to be less frequent than changes in the website’s front-end structure. At the same time, APIs return structured data and even data that might not be accessible through the website itself. An example of this is the access to auctions that have already been closed. After some time, Catawiki does not show in their website auctions that have already been closed. However, through their APIs, information can still be accessed. Additionally, using these APIs increases performance as extraction times are reduced remarkably.

A.3 Storage

To store the data, the non-relational NoSQL database MongoDB (Chodorow,2013) was used. The advantage of this database compared to others is the support for unstructured data and scalability. Storing unstructured data was a requirement for the storage choice as the data collected was not entirely known beforehand and was also subject to change during the process. The flexibility of MongoDB made it a perfect candidate for this kind of work.

(29)

28

A.4 Processing

Processing of data was done in a cloud server to be able to handle both compute and high memory usage. Given the size of the data, the server needed to be powerful enough to be able to fit the high amount of data in RAM while also having CPU power to run computations and transformations on the data. This performance was achieved using an AWS EC2 instance, and the code was written in Python (Van Rossum & Drake Jr,

1995) for reliability and speed. Most data processing took place using statistical libraries for Python such as statsmodels (Seabold & Perktold,2010).

At the same time, to improve performance, all processes were broken apart into inde-pendent pieces of work that could later be joined together. This strategy allowed for parallelisation and using the multiple cores available in the server.

λ: Get Categories λ: Get Auctions λ: Update Auctions λ: Get Lots λ: Update Lots λ: Get Bids λ: Crawl Lots MongoDB CloudWatch Schedule

(30)

Appendix B

Perfect Multicollinearity

To understand if the selected variables for the linear regression present any perfect multicollinearity problems, the correlation between each of the regressors was calcu-lated. Figure B.1shows these correlations employing a Pearson’s correlation coefficient. Although some correlations exist, the matrix shows there is no evidence of perfect mul-ticollinearity.

Figure B.1: Pearson correlation coefficient matrix for all regressors

(31)

Appendix C

Catawiki

Catawiki is an online auctions website with a focus on antiques and collectables. The company started as a community of collectors that collaboratively constructed a database of collectable comic books. Looking for ways to monetise the website, the founders turned Catawiki into an online auctions website (O’Hear, 2014). One of the most cru-cial selling points for Catawiki is their experts’ service. The site has a number of people that personally access the items that are being auctioned to verify their state and va-lidity. This mechanism generates trust with the buyers and guarantees a certain degree of veracity in the description of each of the lots.

Figure C.1: Screenshot of a Catawiki auction selling a number of lots of Roman and Byzantine ancient coins

(32)

31

Figure C.1 shows an auction in Catawiki selling lots of different Roman and Byzantine ancient coins. The entire auction is moderate by two experts that validate that the coins are original. Nowadays, Catawiki offers items from a wide range of categories, from coins and stamps to cars and fossils. Figure C.2 shows the list of categories as advertised on their website.

(33)

References

Andreau, J. (2018, 02). auctions. Oxford University Press. Retrieved

from https://oxfordre.com/classics/view/10.1093/acrefore/9780199381135.001.0001/ acrefore-9780199381135-e-8237

Ariely, D., Ockenfels, A., & Roth, A. E. (2005). An experimental analysis of ending rules in internet auctions. RAND Journal of Economics, 890–907.

Auctions summary. (2019, Nov). Retrieved from https://www.fcc.gov/auctions

-summary

Avery, C. (1998). Strategic jump bidding in english auctions. The Review of Economic Studies, 65 (2), 185–210.

Bajari, P., & Horta¸csu, A. (2004). Economic insights from internet auctions. Journal of Economic Literature, 42 (2), 457–486.

Chodorow, K. (2013). Mongodb: the definitive guide: powerful and scalable data storage. ”O’Reilly Media, Inc.”.

Cramton, P. (1998). Ascending auctions. European Economic Review , 42 (3-5), 745–756. Cramton, P., et al. (2002). Spectrum auctions. Handbook of telecommunications

eco-nomics, 1 , 605–639.

Easley, R. F., & Tenorio, R. (2004). Jump bidding strategies in internet auctions. Management Science, 50 (10), 1407–1419.

Flanagan, D. (2006). Javascript: the definitive guide. ”O’Reilly Media, Inc.”.

Heyman, J. E., Orhun, Y., & Ariely, D. (2004). Auction fever: The effect of opponents and quasi-endowment on product valuations. Journal of interactive Marketing, 18 (4), 7–21.

Knetsch, J. L. (1989). The endowment effect and evidence of nonreversible indifference curves. The american Economic review , 79 (5), 1277–1284.

(34)

References 33

Krishna, V. (2009). Auction theory. Academic press.

Lewis, P. H. (1995, May). Auction of collectibles on the internet. The New York Times, 4. Retrieved from https://www.nytimes.com/1995/05/23/business/company -news-auction-of-collectibles-on-the-internet.html

Lucking-Reiley, D. (2000). Auctions on the internet: What’s being auctioned, and how? The journal of industrial economics, 48 (3), 227–252.

Milgrom, P., & Weber, R. J. (1982). The value of information in a sealed-bid auction. Journal of Mathematical Economics, 10 (1), 105–114.

Milgrom, P. R., & Weber, R. J. (1982). A theory of auctions and competitive bidding. Econometrica: Journal of the Econometric Society , 1089–1122.

Mintert, J. R., Blair, J., Schroeder, T. C., & Brazle, F. (1990). Analysis of factors affecting cow auction price differentials. Southern Journal of Agricultural Economics, 22 (1378-2016-110545), 23–30.

O’Hear, S. (2014, Sep). Online auction house catawiki scores AC10m series b led by accel partners. Retrieved from https://techcrunch.com/2014/09/24/catawiki/

Seabold, S., & Perktold, J. (2010). statsmodels: Econometric and statistical modeling with python. In 9th python in science conference.

Thornton, P. K. (2010). Livestock production: recent trends, future prospects. Philo-sophical Transactions of the Royal Society B: Biological Sciences, 365 (1554), 2853– 2867.

Van Rossum, G., & Drake Jr, F. L. (1995). Python reference manual. Centrum voor Wiskunde en Informatica Amsterdam.

Investigation of Online Auctions and Price Determination