Está en la página 1de 11

An Empirical Analysis of Algorithmic Pricing

on Amazon Marketplace

Le Chen Alan Mislove Christo Wilson
Northeastern University Northeastern University Northeastern University
Boston, MA USA Boston, MA USA Boston, MA USA

ABSTRACT tings due to lack of data (e.g., competitors’ prices) and physical
The rise of e-commerce has unlocked practical applications for al- constraints (e.g., manually relabeling prices on products). In con-
gorithmic pricing (also called dynamic pricing algorithms), where trast, e-commerce is unconstrained by physical limitations, and col-
sellers set prices using computer algorithms. Travel websites and lecting real-time data on customers and competitors is straightfor-
large, well known e-retailers have already adopted algorithmic pric- ward. Travel websites are known to use personalized pricing [25],
ing strategies, but the tools and techniques are now available to while some e-retailers are known to automatically match competi-
small-scale sellers as well. tors prices [40, 17].
While algorithmic pricing can make merchants more competi- While algorithmic pricing can make merchants more compet-
tive, it also creates new challenges. Examples have emerged of itive and potentially increase revenue, it also creates new chal-
cases where competing pieces of algorithmic pricing software inter- lenges. First, poorly implemented pricing algorithms can inter-
acted in unexpected ways and produced unpredictable prices [37], act in unexpected ways and even produce unexpected results, es-
as well as cases where algorithms were intentionally designed to pecially in complex environments populated by other algorithms.
implement price fixing [5]. Unfortunately, the public currently lack For example, two competing dynamic pricing algorithms inadver-
comprehensive knowledge about the prevalence and behavior of al- tently raised the price of a used textbook to $23M on Amazon [37];
gorithmic pricing algorithms in-the-wild. reporters have noted that similar algorithmic pricing also exists in
In this study, we develop a methodology for detecting algorith- day-to-day commodities [9]. Second, dynamic pricing algorithms
mic pricing, and use it empirically to analyze their prevalence and can implement collusive strategies that harm consumers. For exam-
behavior on Amazon Marketplace. We gather four months of data ple, the US Justice Department successfully prosecuted several in-
covering all merchants selling any of 1,641 best-seller products. dividuals who implemented a price fixing scheme on Amazon using
Using this dataset, we are able to uncover the algorithmic pricing algorithms [5]. Unfortunately, regulators and the public currently
strategies adopted by over 500 sellers. We explore the characteris- lack comprehensive knowledge about the prevalence and behavior
tics of these sellers and characterize the impact of these strategies of algorithmic pricing algorithms in-the-wild.
on the dynamics of the marketplace. In this study, our goal is to empirically analyze deployed algo-
rithmic pricing strategies on Amazon Marketplace. Specifically,
we want to understand what algorithmic pricing strategies are used
1. INTRODUCTION by participants in the market, how prevalent these strategies are,
For the last several years, growth in e-commerce has massively and ultimately how they impact customer experience. We chose
outpaced growth among traditional retailers. For example, while to focus on Amazon for three reasons: first, Amazon is the largest
retail sales shrank 1.3% in the first quarter 2015 in the US, e- e-commerce destination in the US and Europe [16]. Second, Ama-
commerce grew 3.7% [21]. Although e-commerce only accounts zon is a true marketplace populated by third-party sellers, as well
for around 7.3% of the overall $22 trillion in global retail spending as Amazon itself. Third, Amazon’s platform provides APIs that are
projected for 2015, this percentage is projected to rise to 12.4% by specifically designed to facilitate algorithmic pricing [1].
2019 [27]. Furthermore, these overall figures mask the dispropor- To implement our study, we develop a novel methodology to col-
tionate gains of e-commerce in specific sectors, such as apparel, lect data and uncover sellers that are likely using algorithmic pric-
media, and office supplies. ing. We collect four months of data from 1,641 of the most popular
The rise of e-commerce has unlocked practical applications for products on Amazon. We gather information about the top-20 sell-
algorithmic pricing (sometimes referred to as dynamic pricing al- ers of each product every 25 minutes, including the sellers’ prices,
gorithms or Revenue/Yield Management). Algorithmic pricing ratings, and other attributes. We use this data to analyze changes
strategies are challenging to implement in traditional retail set- in price over time, as well as compare the attributes of sellers. We
focus on top selling products because they tend to have multiple
sellers, and thus are likely to exhibit more competitive dynamics.
We begin by analyzing the algorithm underlying Amazon’s Buy
Box. This algorithm determines, for a given product being sold by
many sellers, which of the sellers will be featured in the Buy Box on
Copyright is held by the International World Wide Web Conference Com- the product’s landing page (i.e., which seller is the “default” seller).
mittee (IW3C2). IW3C2 reserves the right to provide a hyperlink to the As shown in Figure 1, customers use the Buy Box to add products
author’s site if the Material is used in electronic media.
WWW 2016, April 11–15, 2016, Montréal, Québec, Canada.
to their cart; sellers not selected for the Buy Box are relegated to
ACM 978-1-4503-4143-1/16/04. a separate webpage. The precise features and weights used by the

Other products have a $0. a seller-specified representation 2. Amazon 2. http://personalization. 6]: strategic and successful. and depending on the product category. § 5 ucts (books. Overall.34 We begin by briefly introducing Amazon Marketplace.99/item sold. The vast majority of cat- egories have a 15% referral fee. 3P sellers can opt to handle logistics (inventory. shipping method. etc. However.05/lb fee for standard ship- ing.) themselves. 44% of online shoppers navigate directly to Amazon to make purchases. we determine that. when they do not offer the lowest price for a given product).g. sellers may become “Pro Merchants” for $39. site [3].35/product. Because of these cus- In summary. BACKGROUND of an item) per month must pay $0.. ping. we make our code and data available at uct sold. which push may lead to customer dissatisfaction [9]. as well as a wide range of When customers purchase products from Amazon. These rapidly fluctuating prices digital devices (Kindle e-readers.58 for extremely heavy. Using Machine Learning (ML). and 3P Sellers and FBA. We develop a technique to detect sellers likely using algo. and identify 543 such sellers. we identify over 500 sellers who are very likely using algorithmic pricing. The Buy Box is shown on every product . 2.0005 per active SKU. (FBA) program.ccs.99/month. zon also functions as a marketplace for third parties. these fees influence the dynamic pricing Amazon. imum referral fees of $1-$2/item. among all claims to have 2M Third-Party (3P) sellers worldwide who sold the variables we can observe. customers towards Amazon’s shopping apps. zon Marketplace. DVDs. To identify pricing algorithms. Further. and finally small media items. we compare the characteristics of algorithmic and non- algorithmic sellers. Amazon also enforces min- Outline. Amazon earned $89B in revenue in 2014. 4. and are more likely to be fea. the Buy Box.04-$10.Buy Box algorithm are unknown [13]. phones etc. rather than using search engines or 1. algorithmic sellers “win” the Buy Box more frequently (even boasts 244M active customers [22]. plus on the features of the market that are salient to algorithmic pric- variable per pound shipping fees ranging from $0.10/lb for expedited shipping. video.45 + $0.1 Amazon Marketplace As we discuss in § 5. Referral Fee: Amazon assesses a referral fee on each prod- To facilitate further study. Ama- frequency of price changes. 1.neu. We focus packing fee per product depending on its size and type. and use correlative analysis to identify specific sellers whose prices track the target price over time.). 3. We present a comprehensive overview of dynamics on Ama. etc. we treat the target price of each product (e. including Third-Party (3P) sellers. § 7 presents related work and § 8 concludes. and involves five 4. larly shaped items. Although Amazon began as an on- line bookstore. in the US and Europe [27]. we also Prime membership program. representing 40% of all items sold via the web- tant feature used by the Buy Box algorithm to select sellers. Finally. 2B items in 2014. this work makes the following contributions: tomer retention efforts. For our purposes. § 2 covers background on Amazon and the Amazon Marketplace.65 + $0. 2.g. showing they are components [4. founded in 1994. FBA Fee: Sellers that use FBA must pay a $1. Media prod- algorithm that Amazon uses to select the Buy Box winner. Overall. shipping. returns. Closing Fee: Amazon’s closing fees vary based on product and § 3 covers our data collection methodology. in which case Amazon handles all logistics. which Amazon inspires fierce loyalty among customers through their may further contribute to their feedback scores. or they can join the Fulfilled By Amazon 3. or tured in the Buy Box. but that customer feedback and ratings are also used. tablets. which gives customers free 2-day observe that the lowest price and the Buy Box for products with shipping (or better) as well as unlimited access to digital streams for algorithmic sellers are significantly more volatile than for products $99/year. and product weight.. 5. it now sells over 20 categories of physical products 2. low prices are the most impor. to be more successful than non-algorithmic sellers: they offer fewer products. e-books). or $0. Listing Fee: High-volume sellers that list more than 2M Stock Keeping Units (SKUs. The remainder of this paper is organized as follows. The fee structure for 3P sellers is complicated. visiting competing online retailers [35]. In addition to acting as a merchant. and § 6 explores the characteristics and impact of these sellers. and do so through the Buy Box. The fees vary between 6-45% of the total sale price.39 for ing. the lowest advertised price or Amazon’s price) as a time series. Seller Fee: “Individual” sellers must pay $0. they have much higher levels of feedback than other sellers. including the characteristics of sellers. § 4 explores the category. but receive significantly higher amounts of feed- back (suggesting they have much higher sales volumes).) have a flat fee of $1. presents our algorithm for detecting sellers using algorithmic pric. understanding the Buy Box algorithm is important because sellers may choose dynamic pricing strategies that maximize their chance of being selected by the algorithm. we examine the dynamic pricing strategies used by sellers in Amazon Marketplace. rithmic pricing. irregu- the APIs offered by Amazon Marketplace Web Services. We observe that algorithmic sellers appear Figure 1: An example Buy Box on Amazon. they typically digital goods (e. to $124. We explore the properties of these sellers.2 The Buy Box (even fresh food in select cities [15]). Next. is the largest e-commerce website strategies used by 3P sellers. Amazon’s success is further bolstered by their branded without any algorithmic sellers. yet the algorithm is of criti- cal importance since 82% of sales on Amazon go through the Buy Box [38]. downloadable and streaming music.

Appeagle. the Buy uct’s price equal to the lowest competing offer. We use the data from the product pages to analyze the Buy Box algorithm in § 4. Instead. their prices. signed to facilitate dynamic pricing.1 MWS also has a subscription API. describe our data collection process. this tool Box algorithm is a function B(P ) → pi . The most sophisticated of these tools is the Ama. offered for a product (or less. with pi ∈ P . In these cases. These ser- the features used by the Buy Box algorithm (e.45 0 $6. Figure 4: An example New Offers page on Amazon. which is a set of APIs for programatically interfacing with the marketplace. This tool allows a 3P seller to set a prod- being offered by n sellers with prices P = {p1 . Specifically. MWS includes 3. and their reviews (number of reviews and average score).47 Price CDF Seller Page 40 $6. we used web scraping to obtain information on the ac- tive sellers and their prices. Unfortunately. Each product. To achieve this goal. As shown only adjusts the product’s price once: if the lowest price changes in Figure 1.44 1min 15mins 1hr 3hrs 1day 07:00 08:00 09:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00 17:00 Interval Between Changes Timeline (Minutes) Figure 2: Frequency of page updates. page on Amazon: it contains the price of the product. Figure 1 shows an example Buy Box. DATA COLLECTION we use Machine Learning (ML) to examine the Buy Box algorithm The goal of our study is to analyze the dynamic pricing strate- in-depth in § 4. Feed- ing that 82% of sales on Amazon go through it [38]. it is unknown whether this feature list is sophisticated dynamic pricing strategies. and RepricerExpress leverage MWS to the underlying algorithm the focus of much speculation by 3P sell. and we leave them to future work. prices. this information is paginated into 10 3P sellers per page. many products on Amazon are sold by multiple sellers. and a button to purchase the receive near real-time price updates for specified products. Unfortunately. and changing We would ideally have liked to use the Amazon Marketplace Web Services (MWS) API to collect the seller and price informa- tion for products. Seller Platforms. In this section. Thus. listing all sellers for a 2 In this study. we Amazon offers an array of tools to help 3P sellers manage prod. we only focus sellers who offer new items. the seller’s price is not automatically reduced as well. · · · . their shipping costs. managing inventory. zon Marketplace Web Service (MWS). Figure 3: Examples of price jitter (highlighted with arrows) in the Buy Box on a product page. update includes aggregated information about the lowest 20 prices However. if there are fewer than 20 offers). we crawled the New Offers page2 (the page that is linked to in Figure 1 if one clicks on “2 new”. used given product.49 80 $6. In addition to scraping the New Offers pages for products. . Amazon also has a web-based price match- seller’s offer is displayed in the Buy Box. including specific challenges uct inventory. strategies offered by these services in greater detail in § 5. or what the weights of the features are. RepriceIt. 100 $6. we also scraped the product pages themselves.g. it is not surpris. pn }. the name of the seller. shown in Figure 4). that we needed to overcome to obtain useful. offer subscription-based services for 3P sellers that combine in- ers [13]. items are not covered. offers from other sellers are relegated to a separate again. we describe below how we handle cases where there are more than 10 3P sell- ers. Companies like Sellery. webpage (an example is shown in Figure 4).46 Buybox Price 20 Buybox Seller $6. Because “winning” the Buy Box is so critical for making sales on Amazon.3 Amazon Marketplace Web Service a large number of products—in the marketplace. that allows sellers to information. sellers may use dynamic pricing strategies that give them an advantage with respect to being chosen by the algorithm. we found that the API did not meet our requirements for two reasons: the API does not return the identity of 3P sellers (just their chosen price). However. and the API is heavily rate-limited. if product is ing tool for 3P sellers [8]. shipping prices. gies being used by sellers on Amazon.. We discuss the types of complete.1 Obtaining Sellers and Prices functions for listing products. 1 Amazon’s documentation stipulates that sellers may only update prices every 20 minutes [2]. representative data.48 60 $6. This has made visor. Although Amazon has released some information about ventory management with dynamic pricing capabilities. a proprietary Amazon algorithm determines which In addition to MWS. we require longitudinal data about sellers and their prices—ideally for 2. shipping vices enable any merchant to easily become a 3P seller and leverage options and speed) [7]. The capabilities of MWS are clearly de- Given the prominent placement of the Buy Box. Formally. 3. for each product we ex- amine. This page lists all 3P sellers.

2014 and December 8. our between number of products and crawling frequency. 3 Second.158 products changes happen either on very short timescales (< 1 minute. likely from a public listing of all Amazon products. the overall characteristics of prices and sellers price change. with a long tail of update times. For this crawl. we randomly selected 5 products from the best-seller August 11.000 best- products3 and crawled their product page and the first 2 seller pages selling products to study. the customer is 5 not able to add the product to their cart at the old price). https://archive. To explore reduces the amount of data we need to collect while still capturing the origins of this high level of dynamicity. some specific times. we set up an Amazon Individual Seller account. when we refer to “price”. To cover more products at longer intervals. We conduct a second crawl between Specifically. We choose to only download the first two pages of sellers. and the as we found the sellers who change their prices often (suggesting Buy Box seller in Figure 2.4 3. Altogether there are 1. listed several products. Beauty. etc. (containing the Buy Box) and the first two pages of 3P sellers (typ- We first examine how frequently sellers’ prices change and how ically. as ex- We therefore choose a crawling frequency of every 25 minutes. as well as somewhat lower prices. we select a crawling frequency. we turn to selecting the products to study. ferent products. As shown in Figure 2. we downloaded all seller pages.4 Limitations Using these results. containing the 20 sellers with the lowest frequently the Buy Box is updated. Amazon states that it can take up to 15 minutes for all (details of these Figures are discussed in the next section). we need to decide how frequently we will crawl each page. As a tradeoff There are two noteworthy limitations to our dataset. First Crawl (Crawl1). Amazon 4 may alter pages for Prime users. Although the exact number of products come from 23 departments from Amazon. -1 is anti-correlation).. reflect what the majority of Amazon users see. Second Crawl (Crawl2). product price and the number of sellers in Figures 5 and 6. We conduct products by price. Electronics. systems to converge to the new price. we create a high-resolution dataset that will help to il. a customer cannot add the item to their shopping cart (i. users cannot search or sort by base price alone. we are referring to the total price. as the best-selling products change over time: there We observe that the price appears to change five times in this are 196 products in common between the two typically to highlight sellers and To further verify these results. It is important to note that the New Offers 3. all Amazon’s customers [23]. downloading only the first two pages massively previous price change. we randomly sampled 2. the Buy Box price of an example product in Figure 3 (we observed similar behavior for other products. the Buy Box price. Recall that we ing the lowest-cost shipping option) for each seller.5 We compare the Amazon inconsistencies) or very long timescales (> 30 minutes). our dataset should accurately downloadable music. 100 100 100 Crawl1 80 80 80 Crawl2 60 60 60 CDF CDF CDF Crawl1 Crawl1 40 Crawl2 40 Crawl2 40 20 Random 20 Random 20 0 0 0 $1 $10 $100 $1000 1 10 102 103 -1 -0. First. download the product page (containing the Buy Box). However. we crawled data from Amazon using browsers that were http://www.. pected. 2015 and September 21. estimates place it at around 20–40% of best-seller products (we exclude digital goods such as e-books.5 1 Product Price Number of Sellers per Product Spearman's ρ Figure 5: Cumulative distribution of product Figure 6: Cumulative distribution of number of Figure 7: Correlation between price and rank (1 is prices.790 Prime members is unknown. but that old prices sometimes briefly reappear after a in Figures 5 and 6. Our first crawl was conducted between 3. and downloaded both the product page (covering up to 20 sellers) once per minute for 3 days. but do not include them due to It is important to note that the first and second crawls cover dif- space constraints). most briefly quantify this bias.5 0 0. such as Appliances. 2014. This result is likely due to Amazon’s distributed in- are very similar between the two crawls despite the time difference frastructure. likely to have such sellers. state. We select 1. We observe that the updates are surpris- dynamic pricing algorithms) were within the first two pages 96% ingly dynamic: 40% of price changes occur within a minute of the of the time. but did not To do . even though a customer may see an outdated price. As shown timeseries. crawl. we choose to dataset is biased (by design) towards best-selling products. so we focus on best-selling products We do this as Amazon uses total price when users explicitly sort since they are likely to have many competing sellers. and changed their prices at products that are eligible for expedited Prime shipping. we observe that our best-sellers show many more sellers than random products. We found that when prices are in an inconsistent of our analysis and conclusions may not extend to Prime users. We plot the cumulative distri- prices). rather than actual price changes by sellers. and gift cards). price includ. Thus. Thus. Best-seller not logged-in to Prime accounts. Throughout the are aiming to study dynamic pricing.2 Determining Crawling Frequency September 15. but not always. not all products are equally paper. we plot a timeseries of most of the “interesting” behavior. two separate crawls that have different characteristics. We select 837 best- Because 3P sellers (and Amazon) can change their price at any selling products that had at least two sellers at the beginning of the time. bution of inter-update times for sellers. luminate the tradeoff between crawling resolution and frequency. Calculating Prices. Thus. the very rapid price “jitters” are likely caused by transient inconsistencies in Amazon’s infrastructure. 2015. Thus. sellers per product.e. perfect correlation.3 Selecting Products page lists both the base price and the total price (i.

we model dynamic pricing strategies. price is not the In the previous section. In contrast. We collect the rankings for all products in our dataset. Second. we examine the relationship between seller rank (from the tance (calculated as the average of the Gini index among all splits in New Offers page) and the winner of the Buy Box. RF is an ensemble classifier that achieves low bias and low Buy Box winner: the seller winning the Buy Box is constant for variance by aggregating the decisions from a large number of low- 31% of products. i. This includes sales volume. To facilitate our analysis. other classifiers.2 Behavior of the Buy Box Algorithm Unfortunately. However. and the list of sellers sorted by price. we conduct an in-depth investigation of the features and In this section. We immediately see that important seller features. the Buy Box as a prediction problem. vector. since it may influence how sellers choose Model and Features. such as kernel-based SVM.. Figure 9 shows the trees). and thus cannot Next. Amazon returns sellers sorted by price). the rating and feedback of a seller come win the Buy Box. RF is an ideal classifier for dreds of changes. and follow up by characterizing the dynamics and be- havior of the Buy Box. Is the Product FBA?: true if the seller uses FBA. each of which is characterized by a feature contains Buy Box winners and seller rankings. Rank zero means they are the first seller in the list. around 20% of products have correlation <1. as well as speculation Amazon’s systems take additional seller attributes into account (be. most products see a number of changes: only 13% of products have Classifier Selection. results show that Amazon’s systems take additional characteristics uate sellers. we see fewer changes to the Box. we construct a feature vector for 4. First. then Spearman’s ρ 6. and cal. our task because it outputs interpretable measures of feature impor- Next. or many more than one per day. As shown in Figure 7. we examine the empirical behavior of the Buy Box. for a product Note that in this section. (i. quantify their impact on the Buy Box algorithm. are harder to interpret. In both cases. Thus. Furthermore. Positive Feedback: positive feedback percentage for the sellers returned by Amazon. Average Rating: average customer rating6 of the seller. Contrary to our expectations. some products even experience hun- bagging samples [36]. Amazon explicitly ranks all sellers for each product on the New Offers page. for each product in our dataset. response time to customer inquiries. Is Amazon the Seller?: true if the seller is Amazon. sellers at different ranks. rate of returns and refunds. Price Difference to the Lowest: difference between the ner is not necessarily the seller who is ranked the highest.. per product. for many products. . the Buy Box win. we only use data from Crawl2. Given our dataset. However. for products with different numbers of sellers. seller.1 Seller Ranking each seller containing the following seven features: As shown in Figure 4.e. Box algorithm is crucial. This result gives us our first clue that According to Amazon’s documentation. Feedback Count: total feedback count for the seller. algorithm [13. we demonstrated that the Buy Box algo- only factor that impacts ranking and selection for the Buy Box). as we ing with dynamics over time. other features are possibly used by the Buy Box sides price) when making decisions. suggesting that our data does capture the many dataset over the six weeks of observation. culate Spearman’s Rank Correlation (ρ) between the ordered list of 4. will equal 1. Amazon does not al- 7. Understanding the Buy of the features and weights used by the Buy Box algorithm. the Buy Box winner correlated decision trees with different feature combinations and and price is highly dynamic. 7]. If the lists perfectly correspond 5. Surprisingly. from 3P sellers. seller’s price and the current lowest price for the product. are chosen by the algorithm. these We begin our analysis by exploring how Amazon’s systems eval. and shipping times. Price Ratio to the Lowest: ratio between the seller’s price how Amazon chooses to weigh various seller features.3 Algorithm Features and Weights uses non-trivial strategies to evaluate sellers (i. 1.1. 4. we use Machine Learning (ML) to try to infer some weights that drive the Buy Box algorithm. the fraction of sellers at different ranks that “win” the Buy Box. even without these features we are able to achieve high to the price and seller in the Buy Box for all products in the Crawl2 prediction accuracy. and the current lowest price of the product. THE BUY BOX rank sellers solely on prices (see Figure 7). Figure 8 plots the number of changes will show. rithm uses features beyond just price to select the Buy Box winner. Specifically. start. and there is a long tail of sellers at higher ranks from customer surveys asking for a star rating (0–5 stars) and spe- that win. Recall that we have already shown that Amazon does not cific questions about the customer’s experience. while 50% of products sifier to predict whether a specific seller of a product wins the Buy have more than 14 changes. we first examine the seller ranking algorithm as it offers clues as to 2.. algorithm. only 60% of the top-ranked sellers 6 Fully described in § 6. 4.e. 3. Thus. we briefly examine Amazon’s seller ranking beyond price into account when evaluating sellers. since it offered by n sellers. However. Taken together. we observe that Amazon 4. We leverage a Random Forest (RF) clas- static Buy Box prices over the entire period. ways sort sellers by price. 100 100 100 Win Rate (%) 80 80 80 Accuracy 60 60 60 CDF 40 Price 40 40 Prediction 20 20 20 Baseline: Lowest Price Seller Baseline: Lowest Rank 0 0 0 0 1 10 100 1000 0 5 10 15 19 1 5 10 15 20 Number of Changes Rank Number of Sellers Figure 8: Cumulative distribution of changes to Figure 9: Probability of winning the Buy Box for Figure 10: Buy Box winner prediction accuracy the Buy Box price and winner.e. we cannot measure these features. our goal is to predict which seller will be chosen to occupy the Buy Box.

we build a detection algorithm that tries of sellers for a given product. r).05 as algorithmic pricing candidates. (t1 . unless otherwise stated. Recall that Note that when we construct the three target price time series. We are able to gain some visibility into this algorithm. repricing software for Amazon Marketplace.8 and choose three tar- pressive predictive power of our RF classifier. as external observers https://sellerengine. p1 ). r). However. Overall. Figure 10 shows the accuracy of our RF classi.03 Is the Product FBA? 0. p1 ). the seller’s positive feedback and feedback count are also important metrics for winning the Buy We also construct three target price time series. pamzn m )} vantage. and feedback count are all important features. . they set and the timing of changes suggest algorithmic control. corresponding to Box. (t1 . We hypothesize that sellers using algorithmic pricing are likely ure 10 also depicts the accuracy of two naïve baseline classifiers. we observe that Amazon’s algorithm for choosing the Once we have constructed the time series corresponding to winner of the Buy Box is a combination of a number of undocu- (s. (t1 . However. 50–60% accuracy. since FBA confers free shipping for Prime users. and that the mag- suggest that sellers who use algorithmic strategies to maintain low nitude of the price changes are relatively constant. Sellery provides one more option: matching to the aver- 7 Our dataset includes at least 1K samples at each rank. amzn amzn AM ZNr = {(t0 . etc). · · · .e. while the price relative to the competitor with the lowest price. Both baselines only achieve first define several target prices that the seller could match against. pm )} portant than other features. sellers where the prices 100% accuracy in the 1-seller case. we calculate the similarity between Sr and LOWr . p0 ). and the Feature Weights. p1 ). i. with the results indicating that price. sellers (and not their usage of the Amazon Marketplace API. thus only a subset of seller/product pairs include AM ZNr . Higher weights mean that the For a given seller/product pair (s. as determined by our RF classifier.7 (the empirical cutoff of a strong positive correlation) and p-value ≤ 0. pm )} member. back. we lack ground truth on which sellers are using algo- validation) for all products in Crawl2. · · · . (tm . seller feed. 2N Dr = {(t0 . lowest price for r. fier at predicting the winner of the Buy Box (using 10-fold cross. based on two different change thresholds.1 Methodology To put the accuracy results of our classifier in perspective. the “Amazon is the Seller” feature. (tm . This exclusion rule also prevents features. We note that doing so is non-trivial. These results When ρ is large. We mark pairs prices relative to their competitors are likely to gain a large advan- tage in the struggle to win the Buy Box. p1 ). such as ourselves are only able to measure the prices offered by Evaluation. the lowest price plow i . Thus. We use a change threshold of 20 Table 1: Relative importance of different features in winning the Buy Box. p0 ). (t1 .33 Lowest Price 726 544 426 408 Positive Feedback 0. 8 ketplace. and Amazon’s which contradicts conventional wisdom about how the Buy Box price pamzn i for r at each time ti : algorithm functions [13]. Interestingly. this strategy is neither likely to be useful for results are statistically significant. This makes sense intuitively: for example. Sr = {(t0 . Obviously. mented features and weights. p2nd 2nd 2nd 0 ). DYNAMIC PRICING DETECTION tuitively. It is possible that Amazon scores highly in these missing lowest price for r at each time ti . as a function of the number rithm pricing. winning the Buy Box. it is trivial to achieve to locate sellers that behave like “bots”. we caution that this does not necessarily mean that Ama- zon has tilted the Buy Box algorithm in their favor. Finally. we observe that using FBA has low importance. and also highlights the im. nor being competitive among the sellers. Amazon’s price.36 Sellers Products Sellers Products Price Ratio to the Lowest 0.7 get prices for each product: lowest price. (tm . pm )} Although we observe that “being Amazon” does confer some ad. it is possible that FBA is an important factor in cases where the customer is an Amazon Prime LOWr = {(t0 . the 2nd lowest price p2nd i . we other chooses the lowest ranked seller. 5. · · · . As shown of the prices pi offered by the seller at time ti : in Table 1. 9 In fact. In- 5.10 2nd Lowest Price 721 494 425 370 Feedback Count 0. with ρ ≥ 0. Feature Weight Threshold = 10 Threshold = 20 Strategy Price Difference to the Lowest 0. fier achieves 75–85% accuracy even in the most challenging cases with many sellers. (tm . plow low low 0 ). it chooses the lowest ranked one). it means that the price changes contained in the pair of time series occur at the same moments. The final step in our methodology is to filter our candidates. Therefore. we examine the weights calculated second lowest price. For example. However. · · · . thus the age price. which manifests in our classifier as additional weight in us from comparing Amazon’s prices against themselves. We motivate our selection of target prices by examining popular ture used by the Buy Box algorithm.02 Table 2: Number of sellers and products with detected algorithmic pricing. and AM ZNr (respectively) using Spearman’s Rank Correlation. however. 2N Dr . note that Amazon does not sell all products in our dataset. a seller who al- est price will win the Buy Box (if there are multiple sellers offering ways wants to offer the lowest on a specific product must set their the same lowest price. Moreover.10 Amazon Price 297 277 176 183 Is Amazon the Seller? 0. we Amazon’s Buy Box documentation states that the algorithm uses exclude the prices offered by s. we see that the classi.. to base their prices at least partially on the prices of other sell- One baseline classifier always predicts that the seller with the low.9 for each feature by our RF classifier. we construct a time series feature is more predictive of who will win the Buy Box. if a seller exhibiting high correlation with the target price We now turn to detecting algorithmic pricing on Amazon Mar. if s always offers the several features that we are unable to measure.06 Total 918 678 543 513 Average Rating 0. then LOWr will actually contain the second- ume [7]. Fig. the two price-based features are significantly more im. which reconfirms that price is not the sole fea. such as sales vol. ers.

set their price within $1 of the lowest price. the same wholesale cost for a product. if the lowest price is likely to often match (i. However. we see that 2. For example.1. 2500 100 100 To Amazon To Amazon Number of Sids 2000 Crawl1 80 To Lowest 80 To Lowest 1500 Crawl2 60 60 CDF CDF 1000 40 40 500 20 20 0 0 0 1 10 20 30 40 50 60 70 80 -$60 -$40 -$20 $0 $20 $40 $60 -200 -100 -50 0 50 100 200 Change Threshold Price Change Amount Price Change Percentage (%) Figure 11: Number of algorithmic sellers detected Figure 12: CDF of absolute price differences be. To determine the gap between the prices offered by the suspected olds. Thus. Figure 13: CDF of relative price differences be- with different change thresholds.5 $16 Price Price $14 $9 Algo Seller Seller 2 Seller 4 Seller 1 Seller 3 $12 Algo Seller Seller1 Seller 2 $8. we plot the relative difference between . we define the change all unique numbers of sellers and products we detect. number of products they sell that we detect with change thresholds The fact that algorithmic sellers matching to Amazon tend to of 10 and 20. As described in their algorithmic pricing than Amazon’s price. In the case threshold as the minimum number of price changes that must occur when the change threshold is 20. tween algorithmic sellers and target prices. Conversely. price in this plot (we ignore the second lowest price in this plot as matching to the second lowest price is very similar to matching to 5. to the second-lowest price as well. in the remainder of the paper. we plot two figures. However. time. we observe that the number of sellers decreases algorithmic sellers and the target prices. this provides more evi. tween algorithmic sellers and target prices.e. § 2. To see if we can observe algorithmic sellers that include exclusive. We observe that algorithmic sellers who match to the Now that we have described our methodology. then it is possi. In this table. The Total line shows the over- ble that the correlation is coincidental. also makes a large number of price changes. For example. As expected.. these fees in their prices. doing algorithmic pricing when we apply different change thresh. However. we merge the sellers and products from charge higher prices may be due to the required commission fees Crawl1 and Crawl2 and present the total unique numbers. the 3P seller must charge a ing the overall lowest price (and 2nd lowest price) as the target for higher price to maintain the same profit margin. this is 38% of all Figure 11 shows the number of sellers that we consider to be sellers that have ≥20 changes for at least one product they sell. Unless otherwise ure 12 examines the absolute difference between the algorithmic stated.4% of all sellers in in a time series Sr for us to consider s as using algorithmic pricing. $10 $10 Algo Amazon Seller 2 Seller 1 Seller 3 $9 $9 Price Price $8 $8 $7 Algo Amazon Seller1 Seller 2 Seller 3 $7 $6 10/30 10/31 11/01 11/02 11/03 11/04 11/05 11/06 11/07 11/08 11/09 08/23 08/24 08/25 08/26 08/27 08/28 08/29 08/30 08/31 09/01 Timeline (Year 2014) Timeline (Year 2014) Figure 16: Example of Amazon (in red) setting a premium over the lowest Figure 17: Example of Amazon (in red) matching to the lowest price over price of all other sellers. if a 3P seller and Amazon share We immediately observe that many more sellers appear to be us. it is im. we choose 20 as our change sellers’ prices and the corresponding target prices. est when Amazon is the lowest price. our dataset use algorithmic pricing. Amazon’s commission fees are around 15% for most product portant to note the different strategies are not necessarily mutually categories. Fig- rapidly as we increase the change threshold. $20 $10 $18 $9. We separate threshold since it represents a conservative threshold that is in the sellers matching to the lowest price and sellers matching Amazon’s “knee” of the distribution. that Amazon charges. prices offer by two other sellers. correlate strongly with) number of price changes in the time series is small. a seller matches to both Amazon and low.5 $10 11/12 11/13 11/14 11/15 11/16 11/17 11/18 11/19 11/20 11/22 11/24 11/26 11/28 11/30 12/02 12/04 12/06 12/08 12/10 Timeline (Year 2014) Timeline (Year 2014) Figure 14: Example of 3P seller (in red) matching the lowest price of all Figure 15: A second example of 3P seller (in red) matching the lowest other sellers. we briefly exam. only 40% of Table 2 shows the number of algorithmic pricing sellers and the algorithmic sellers are within $1 of Amazon’s price.2 Algorithmic Pricing Sellers the lowest). lowest price are very close to the lowest price: 70% of these sellers ine the set of sellers that we find to be doing algorithmic pricing. and a seller matching to the dence that the seller is using algorithmic pricing.

In the figure. sellers matching Amazon’s price and the overall lowest price. percentage of feedback that is positive (0–100). ing algorithmic pricing. back percentage for all sellers in our dataset. Note that the vertical rithmic seller is willing to sell the product for as low as $12. the anomalies in Figure 18 around 1 month are artifacts caused by the majority of the time they sell at prices up to 40% higher. Note that algorithmic sellers on the Buy Box? Amazon does not display these stats for sellers with insufficient feedback (typically new sellers). given this compressed value range. As shown in Figure 15. As we show lowest price is able to sell the product well above their reserve momentarily. respectively). We observe that algo- lowest price across all other sellers. Note that the inven- in Fig 17. and the algorith. we can see four rithmic sellers are active in the marketplace for significantly longer other sellers that offer the lowest price over time. Figure 20: Percentage of positive feedback for al- for algorithmic and non-algorithmic sellers. we observe several cases where the seller offering the while it is only 15 days for a non-algorithmic seller. we see very different behavior between algorithmic role as a seller separately. we are interested in answering the following questions: and also provide feedback about whether their experience with the (1) How do the business practices of algorithmic sellers compare to seller was positive or negative. as shown in Figure 19. the algorithmic seller always matches sales volume. above tory for algorithmic sellers includes all products they sell. However. Given our crawling 5. role as a merchant and the host of the marketplace. we examine its As expected. Although we the algo.3 Price Matching Examples methodology. we compare and contrast Next. we conducted a separate crawl that exhaustively Fourth. periods of time than non-algorithmic sellers. Figure 16 shows a case where Inventory. Second. To obtain com- in time. first time we observe a seller offering that product. we compare the total Amazon (in red) chooses their price to be a premium above the number of products sold by algorithmic and non-algorithmic sell- lowest price of all other sellers. not just which they match the lowest price. We begin by examining the lifespan of can observe a number of sellers who choose a new price that is seller/product pairs in our dataset. First. perhaps focusing on items that they can obtain in bulk at At this point. we observe cases that Amazon adopts more complex collected the entire inventory for 100 randomly selected algorith- pricing strategies than simply matching lowest prices. we Product Lifespan. algorithmic sellers sell fewer unique products by a large margin. Third. On Amazon. dian seller/product lifetime for an algorithmic seller is 30 days. Amazon presents each seller’s av- non-algorithmic sellers? (2) What fraction of market dynamics are erage rating (0–5). respectively. algo- of products. customers may rate sellers on a 0–5 scale particular. we have identified the sellers who are likely us. the algorithmic sellers’ prices and the target prices in Figure 13. the are three other sellers that offer the lowest price at different points dataset may not contain all products sold by sellers. the me- mic seller (in red) always quickly matches their price. Note that this gorithmic sellers have slightly higher positive feedback than non- list of characteristics is not comprehensive: as mentioned in § 4. our data suggests that algorithmic sellers have a high price. We observe that al- feedback volume. where we detected algorithmic pricing. we observe many cases where Amazon itself appears to be employing algorithmic pricing. and ends the last time we observe that seller offering the product. etc. but below which they sell the specific products where we detect algorithmic pricing. 100 100 Algo 100 Algo Algo 80 80 Non-Algo 80 Non-Algo Non-Algo 60 60 60 CDF CDF CDF 40 40 40 20 20 20 0 0 0 25mins 1d 10d 2mon 0 1 10 102 103 104 105 0 20 40 60 80 100 Lifespan Number of Products Positive Feedback (%) Figure 18: Distribution of seller/product lifetimes Figure 19: Number of products sold by algorith. gorithmic and non-algorithmic sellers. we observe that there ers. we examine the feedback received by sellers from cus- the characteristics of algorithmic and non-algorithmic sellers. that they have a large amount of inventory. but that Amazon is almost always slightly more expensive. Next. ANALYSIS products. In this section. and ranking in the seller page. so the long lifespans of their products further suggest the lowest price from the other two sellers. Surprisingly. As shown mic and non-algorithmic sellers. the shortest possible lifespan is 25 minutes. shipping time. low wholesale prices. Since Crawl1 and Crawl2 focused on best selling products. In the figure. algorithmic sellers. we examine the following four Figure 20 shows the cumulative distribution of positive feed- seller-level characteristics: lifespan of products. Feedback.1 Business Practices following analysis (this only filter out 5% and 15% of algorithmic To compare the general business practices between algorithmic and non-algorithmic sellers in our dataset. and Rank. different lengths of Crawl1 and Crawl2.3. respectively. and thus we ignore them in the 6. while the We conclude this section by showing a few example products longest are 3 and 1 months for Crawl1 and Crawl2. In tomers. For example. The lifespan of a pair begins the 15–30% above Amazon’s price. mic and non-algorithmic sellers. product at a small premium relative to the lowest price. likely caused by algorithmic sellers? and (3) What is the impact of and total amount of feedback on the New Offers pages. sellers and non-algorithmic sellers. Since Amazon itself plays a dual rithmic sellers’ positive feedback advantage is more significant. This suggests that al- gorithmic sellers tend to specialize in a relatively small number of 6. . almost all sellers have greater than we do not have access to several seller features such as return rate 80% positive feedback. inventory size. Figure 14 shows an Figure 18 shows the distribution of seller/product lifespans for example where a 3P seller has a clear strategy to always match the both algorithmic and non-algorithmic sellers. Amazon appears to have a ceiling at around $9. plete inventories.

we this percentage increases to 96%. if we compare the products that have Taken together. Crawl” is the average percentage of products sold by Amazon dur- However. Second. We observe that for non- a seller separately. . products they sell. if we examine the algorithmic sellers. However. Figure 21 examines the amount of feedback received by algo. even Amazon tends to be ranked lower for products where back (suggesting they may have higher sales volumes). Note that neither CDF goes to 100% because Finally. products that are sold by Amazon. non-algorithmic sellers. However. our results show that algorithmic sellers exhibit sellers doing algorithmic pricing to products without such sellers.2 Price and Buy Box Dynamics sellers. they still dominate the mar- more common than not. As we observed in Figure 7. the rank of algorithmic sellers on the ketplace. Pairs with at least one price change are sale does appear to change over time. rank in the presence/absence of algorithmic sellers. of which have <20 changes during our crawl. “Overall” means that Ama- This suggests that either there is no competitive pressure to insti- zon sold the product at some point during our crawls. we observe that Amazon ranks highly for almost all aggressive about soliciting customer feedback. followed by price and product changes in the Buy Box. offering around 75% of all best-selling products over time. we observe a ing each snapshot. it makes sense to examine their role as of price changes per seller/product pair. algorithmic sellers winning the Buy Box. while “Each gate changes. prices change 100 or even 1000 times. Figure 24 shows the distribution of number and host of the marketplace. New Offers page for algo and non-algo sellers. and they acquire significantly larger amounts of positive feed. As shown in Figure 22. algorithmic sellers could have much higher sales volume than list is roughly ordered from least-to-highest price. Figure 22: Cumulative distribution of rank on the Figure 23: Cumulative distribution of Amazon’s rithmic and non-algorithmic sellers.10 and there is a long tail of products whose ket. or the seller is unsophisticated. we plot Figure 23. The feedback and sales volume characteristics of algorithmic sellers put them at an advantage with respect to non-algorithmic 6. algorithmic sellers are active. they participate in the marketplace for longer periods of products without such sellers. Amazon as a Seller. the price never changes for 65% of the pairs. algorithmic sellers could be more Overall. Table 3 shows the percentage of best-selling algorithmic sellers. in compete. 100 100 100 Non-Algo Algo Win Rate (%) 80 80 Price 80 60 60 Sid 60 Non-Algo CDF 40 Algo CDF 40 Sid 40 20 20 Price 20 Non-Algo Algo 0 0 0 0 1 10 100 1000 0 1 10 100 1000 0 5 10 15 19 Number of Price Changes Number of Changes Rank Figure 24: Number of price changes per Figure 25: Number of changes in the Buy Box for Figure 26: Probability of winning the Buy Box for seller/product pair. but in the products where they do we see that Amazon ranks in the top 5 sellers 88% of the time. In this case we observe a stark for their low prices. which contrast: algorithmic sellers acquire significantly greater amounts shows the the rank of Amazon as a seller on the New Offers page of feedback. products with and without algorithmic sellers. time. Although Amazon’s inventory of products for- very different distribution. There are two reasons why this could be happening: for every product at each crawl. this plot includes all the products sold by the algorithmic sellers. significantly different characteristics than non-algorithmic sellers: we see a dramatic difference. To demonstrate this. when comparing all the products in Next. algorithmic sellers are much more active in the marketplace. In products with algorithmic sellers. algorithmic sellers. which is the change Each Crawl 62% 63% threshold we set to detect those algorithmic sellers. Unsurprisingly. Besides having an enormous inventory. Algorithmic Non-Algorithmic 10 Overall 72% 73% Readers may notice that the number of price changes for algo- rithmic sellers in Figure 24 goes below 20. the first. Thus. First. Amazon is also known rithmic and non-algorithmic sellers. Given Amazon’s dual role as merchant Price Changes. some Table 3: Percentage of products with Amazon as one of the listed sellers. 100 100 100 Algo 80 80 80 Non-Algo 60 60 60 CDF CDF CDF 40 40 Algo 40 Algo 20 20 Non-Algo 20 Non-Algo 0 0 0 1 102 104 106 0 5 10 15 19 0 5 10 15 19 Number of Ratings Rank Absolute Rank Figure 21: Amount of feedback received for algo. algo and non-algo sellers at different ranks. we look into price changes for each seller/product New Offers page tends to be significantly higher than that of non- pair. we move on to examining the dynamics of Amazon Mar- their respective inventories. we compare the frequency of algorithmic and non- there are cases where sellers do not appear in the top 20 list. they sell fewer unique products.

[30] Finally. we examine the impact of algorithmic sellers 8. and tence of systemic racial discrimination on AirBnB [24]. CONCLUDING DISCUSSION on the Buy Box. Second.12 Ar- Auditing E-Commerce Algorithms. Any opinions. Although we do not observe any ion spam and artificial ratings to manipulate the reputation sys. Clearly. First. However. them vulnerable to manipulation and fraud conducted by unscrupu. customers. [20] studied Uber’s surge pric- ing algorithm and revealed that an implementation bug was caus. we observe cases where algorithmic sellers change prices tens or even hundreds of Theoretical Work on Price Competition. there are also caveats introduced by algorithmic pric- sellers. Similarly. findings. as previously noted. revealed the exis. in markets. Figure 25 compares the number of seller (labeled E-commerce marketplaces have changed many aspects of how as Sid) and price changes we observe in Buy Boxes for products goods are bought and sold. Minkus et al. However. [34] discovered per. The Buy Box Issues in Online Marketplace. and not always positive. [26. function of the dependencies among goods/services offered by the However. algorithm exacerbates the disparity between algorithmic and non- tomers convenience. and a vast inventory of products. these practices. 10] extend the traditional price competition model ing. CNS-1319019. the existence against other competitors in the market. PayPal). As expected. In this paper. understood. markets that include competing algorithmic and non-algorithmic tomers to a great deal of volatility. user-friendly automation platforms like Sellery pricing game in the marketplace. serves. 20% of products with. RELATED WORK enue than non-algorithmic sellers.g. which would be difficult for a human to maintain the Internet and computation technologies. these services have made that have algorithmic sellers. this makes the shopping experience more complicated for are becoming increasingly ubiquitous on online marketplaces.e. Furthermore. as it creates a largely winner-take-all market- However. of these issues in our data. algorithmic sellers tend to set their prices greater than or such sellers in our data set.. Recently. this result is quite interesting: as shown in Figures 12 tecting using a target price time series.neu. which suggests an arms race that may ter- price competition in marketplaces where equilibria rarely exist. [39] conduct an empirical analysis on rithms pushing prices to unrealistic heights [37] and being used to the Seller Reputation Escalation (SRE) ecosystem that provides a implement price fixing [5]. confusing and undesirable [9]. winning the Buy . the anonymous nature of online marketplaces makes place where the Buy Box winner receives the vast majority of sales. the impact of these algorithms on users are often poorly essary to truly understand how these factors impact customers. there are documented cases of algo- tems on online marketplace. and CHS-1408345. it is not clear what the impact of dynamic pricing is on reveal that attackers can correlate the highly sensitive user informa. With easy access to times per day. with the ultimate goal of increasing transparency of Several empirical studies have also show that online market. especially in heterogeneous products with algorithmic sellers. conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of the NSF. Unfortunately. [12] models with algorithmic sellers. this exposes cus. Furthermore. major e-commerce sites. As shown in Figure 26. [11. [19. Edelman et al. it is challenging for non-algorithmic sellers to compete proposed by Bertrand [14] to combinatorial settings. 11 Available at http://personalization.ccs. As shown in Figures 14–17. Finally. using computer algorithms to automati- products with algorithmic sellers experience many more price and cally price goods) a realistic possibility for even small-scale sell- seller changes in the Buy Box: for example. the presence of algorith- tion from public profiles in eBay’s feedback system with their so. 33] study insincere sellers that generate opin. even though al. ers.. [25. Given the importance of winning the search community. this is indeed the We collected large-scale data on products and sellers on Amazon case: algorithmic sellers are more likely to win the Buy Box at all Marketplace. algorithmic pricing (i. These sellers are playing a of cost-effective. places can cause privacy issues for consumers. pricing receive more feedback and win the Buy Box more fre- quently. Online marketplaces bring cus. low prices. Acknowledgements ing users to receive out-of-date pricing information. they manage to win line marketplaces. Automated algorithms guably. 32. algorithmic sellers. unintentional market distortions. which gives rise to the need for merchants to payment providers (e. We view our efforts to detect dynamic shill-purchasing service for escalating business’ reputations on the pricing as the first step towards long-term monitoring of algorithms Taobao online marketplace. and products that do not. the impact of algorithmic pricing on marketplaces out algorithmic sellers have zero price changes. equal to the lowest price for a product. increasing automation opens the door to intentional and lous parties. e-commerce sellers are over time—especially one attempting to manage many products able to adjust their prices automatically by setting algorithmic rules simultaneously—but is trivially automated. although more quantitative and qualitative work is nec- However. and we make our code and data available to the re- ranks except for the top 12 http://camelcamelcamel. versus only 2% for and customers is not yet understood. Next. third-party price monitoring tools like CamelCamelCamel. 14] model a pricing game and Feedvisor is a win for sellers. minate with all serious sellers adopting automation. Our findings illustrate the power of algorithmic pricing in on- gorithmic sellers do not offer the lowest prices. mic sellers does not necessarily push item prices down to their re- cial network profiles on Facebook.11 We found that algorithmic sellers can be de- Buy Box. we took the first steps towards detecting and quan- Next.Buy Box. Sellers we identified as using algorithmic the Buy Box anyway due to their feedback and sales volume. and we identify over 500 and 13. 28. we examine whether algorithmic sellers are successful at tifying sellers using algorithmic pricing on Amazon Marketplace. uncovered instances of price discrimination and price steering on This research was supported in part by NSF grants CNS-1054233. algorithmic pricing can sonal information and detailed shopping habits leaking from online cause prices to fluctuate rapidly. likely suggesting higher sales volumes and thus more rev- 7. 29] We thank the anonymous reviewers for their helpful comments. which they may perceive to be sellers. 31. 18. especially smaller merchants who played by the sellers and study the properties of its equilibria as a lack a dedicated programming staff. customers.

D. A.adweek. AND W EIDENHAMER .655. Digital discrimination: The case Amazon-Sellers-Sold-Record-Setting-2-Billion-Items. Adweek.pdf. AND N ISAN . In Machine https://www.eweek. V. L. S. [32] NASH . D. reputation manipulation: The emergence of [19] C HAWLA . [24] E DELMAN . In ’14. U. The Wall Street Journal. J des Savants 67.. builds-its-european-marketplace. 2015. [16] B ROHAN . I know what you’re buying: Privacy SB10000872396390444914904577617333130724846. EMNLP ’13. [38] TAFT.wsj. L. B. [7] Increase your chances of winning the buy box. Retailer. EC ’13. Cryptography (2015). T. The Bargaining Measuring price discrimination and steering on [5] Former e-commerce executive charged with price fixing in the e-commerce web sites. Coming soon: Toilet paper priced Emerging Networking Experiments and Technologies (2013). prevalence of deception in online review communities. KDD ’13. M ISLOVE . L IU . T.. Japan. Amazon. [26] L I . sociale and of recherches sur les principles mathematiques de la [34] P REIBUSCH . International Conference on World Wide Web (2015).. AND L AOUTARIS . Data Mining (2013).com. In Proceedings of the 15th ACM/USENIX Internet amazon.. http://www. antitrust division’s first online marketplace prosecution.businesswire. 499âĂŞ508. AND L AOUTARIS . The buy box cheat sheet. com/en_US/dev_guide/index. 2 (1883).geekwire. N IU . Estimating the [13] B ENMAYER . April 2015. Feedvisor.698. November 2014. H SU .com/en_US/feeds/ [23] D URYEE . September 2015. Amazon. Crowd-assisted search for price discrimination in e-commerce: ref=hp_left_cn?ie=UTF8&nodeId=200832850. Internet online-marketplace. J. Global e-commerce set to grow 25% in 2015. http://recode.. Proceedings of the Fourteenth ACM Conference on Electronic [31] M UKHERJEE . [33] OTT.algorithm/. October 2014. [37] S UTTER .price. May 2014. WWW ’15. Business Wire. PAES L EME B. M. A. AND G HOSH . B. AND M ATTIOLI . The price of anarchy in bertrand https://www. J. C. Price A.internetretailer.census. Amazon. V. http://docs. AND ROSS . [12] BABAIOFF . L UCIER .. J. M. 2015. [27] L INDNER . HotNets-XI. http://docs.S.geekwire. R. Proceedings of the 21st International Conference on World Wide Web [14] B ERTRAND . October 2015. Amazon buy box: The internet’s $80 billion sales coupled-walmarts-price-matching-leads-ridiculously-cheap.... [20] C HEN CoNEXT ’13. [21] D E NALE . theorie des richesses. Identifying manipulated offerings Department of Justice. L. WANG .developer..internetretailer. N ISAN .html/ N. Geek Wire. AND C ARDIE . prime members worldwide.html?nodeId=1161240.. E RRAMILLI . T. Amazon is absolutely eviscerating other retailers online. S. F.. October 2015).com/2015/amazonfresh-delays. G. In https://www. J.wsj. https://www. J. G UNES . AND L UCA . Price competition using behavioral footprints. Amazon hoax coupled with walmart’s price shipping. Spotting opinion spammers [11] BABAIOFF . In Proceedings of the 24th competition in networks. AND W ILSON .html/ commerce-set-grow-25-2015. WWW ’12.. [3] Amazon sellers sold record-setting more than 2 billion items [18] C HAWLA ..93 – plus [17] C ASTILLO . M.. on review portals. [29] M IKIANS submitfeed. In Proceedings of the 23rd SIGKDD International Conference on Knowledge Discovery and International Conference on World Wide Web (2014). http://www. R.amazonservices. http://www. [2] Amazon marketplace web service (amazon mws) documentation – http://www. R. 299year-prime-fresh-rollout-for-another-3-months/. 2012.. EC ’09. Random forests. U. January 2015. G YARMATI . http: 80-million-high-spending-prime-members-worldwide/. www/data/pdf/ec_current. In Proceedings of the 19th ACM in online combinatorial markets.justice.. [6] Fulfillment fees for orders on amazon. M. G. wal-mart. Census Bureau . AND B ERENDT. C ASTELLANOS . In Privacy Enhancing Technologies (2014).. L IU . In Financial [15] B executive-charged-price-fixing-antitrust-divisions-first. N. Tech. Retailer. In Proceedings of the Ninth ACM Conference on [9] A NGWIN .. //www. Econometrica 18. Learning (2001). AND N IU . D. C. WANG . M. [36] S TATISTICS .cnn. Internet eviscerating-other-retailers-online-new-survey-shows/. N. Purchase details leaked to paypal (short paper). X.developer. [8] Match low price.html. Commerce (2013). M ISLOVE . PETS [10] BABAIOFF . Bertrand reputation-escalation-as-a-service. Measurement Conference (IMC’15) (Tokyo. May 2014. customers-past-year/.. T. In Proceedings of the Sixteenth 155–162. D.html. Re/code. 2015. J. Geek Wire. 2015. M. GeekWire. ACM Conference on Economics and Computation (2015). http://www. Amazon..9. http://www. August 2015. D. September 2015. 2015. H. e-commerce retail sales. tweaks subscription model in california. (2012). Detecting price and search discrimination on the internet. WWW ’14. A. REFERENCES News.html/ Proceedings of the 11th ACM Workshop on Hot Topics in Networks ref=hp_rel_topic?ie=UTF8&nodeId=201687830. July 2015. new survey shows. ?nodeId=201119410. P EETZ . K. A. amazon-buy-box-the-internets-80-billion-sales-button. L IU . of airbnb. [25] H ANNAK .amazonservices. [4] Fees and pricing. D. [22] D URYEE .. R. W ILSON . The Wall Street Journal. First results. T. worldwide in 2014. Quarterly match-policy-to-include-amazon-wal-mart-1443636271. like airline tickets. 2015. W. CNN. ps4s-161527. Target expands online price-match policy to include Hood of Uber. Amazon may have up to 80 million high-spending Feeds_SubmitFeed. AND S TAVROU . Peeking Beneath the [40] Z IOBRO .. B. S. J. In In Proceedings of the Conference on Empirical http://www. http://www. eWeek.. M. In Proceedings of the 10th ACM Conference on Electronic [39] X U . 2 (1950). In http://feedvisor. AND ROUGHGARDEN . AND com/gp/help/customer/display. C. J. Amazon.. Amazon adds 30 million customers in the past year. N. AND H ANCOCK . C ARDIE . K UMAR .. June 2015. A. P. E RRAMILLI . L AZER .... G YARMATI . M.. S OELLER . D.. H. Book review of theorie mathematique de la richesse (2012). April 2015. https://www. 2nd quarter matching leads to ridiculously cheap [1] Amazon marketplace web service (amazon mws) documentation.. L. J. Amazon seller lists book at $23. April 2011. Amazon. M. Amazonfresh delays $299/year prime fresh rollout. E-commerce Commerce (2009). F.. breaches on ebay. 2014. T. AND B REIMAN . Methods in Natural Language Processing (2015).com/articles/ [30] M INKUS . Amazon builds up its european marketplace. AND S IVAN .com/2015/04/30/amazon. Bertrand networks. TECH/web/04/25/amazon. fluctuations and welfare guarantees. EC ’ http://www. A. AND PAES L EME .. [35] R AY. [28] M IKIANS .