Research and Advances
Computing Applications Virtual extension

When Snipers Become Predators: Can Mechanism Design Save Online Auctions?

Posted
  1. Introduction
  2. Emerging Bidding Agents
  3. A Natural Experiment With a Sniping Agent
  4. Implications
  5. Recommendations
  6. References
  7. Author
  8. Footnotes
  9. Figures
  10. Tables

Online auction houses such as eBay, and to an extent Ubid, have become household names in today’s networked economy. They attract legions of buyers and sellers and have evolved from a bare-bones architecture designed solely to match buyers and sellers to a sophisticated virtual marketplace with value-added services such as feedback mechanisms, payment processing, and escrow services. In this article we take a critical look at the mechanism design choices the auctioneers have made. We predict that an emerging technological force, in the form of ubiquitous open-source smarter bidding agents, is likely to have a significant influence on their future practice. This position is substantiated through the use of common terminology in the subfield of micro-economics called mechanism design.

This field deals with incentives and how they can be aligned to practically implement good systemic solutions to problems involving several self-interested economic agents. We discuss the mechanics of current bidding technology on eBay, compare the architecture of two relatively more sophisticated third-party bidding agents, and report on a field experiment that deployed one of these agents on eBay. We illustrate conditions where such agents could lower economic efficiency, promote collusive bidding, and ultimately push electronic auctions towards a precipice. We finish by demonstrating how an understanding of economic market mechanisms, through the introduction of news auctions rules, can help avoid this precipice.

Consider the auction rules of eBay, the single largest online auction house. On eBay, bidding is done through the use of a proxy bidding agent. eBay’s bidding core is designed to bring out the effect of a second price auction, allocating the goods to the highest bidder at a bid increment above the second-highest bid. In its Web site, eBay encourages bidders to submit their maximum willingness to pay the first time they bid, with the assurance that the proxy bidding agent will register a bid that is only one bid-increment above the next lowest bid.

Observe that, from a bidding strategy perspective, eBay’s proxy bidding agent does not allow the bidder to specify the time she wishes to place her bid. At the time of a bid, the agent accepts only her current maximum willingness to pay for the item and deploys immediately after that information is specified. Yet, real bidders in online auctions differ significantly in the timing of their bids and this has a significant impact on their expected payoff [4].

On eBay, if you talk to any of the regulars or read the popular press, a recurring theme persists [6]. Sniping, or very late bidding, it appears, is what you need to do to avoid falling short on eBay. The main rationale given by this strategy’s proponents is that it avoids bidding wars. This rationale is consistent with economic theory that recognizes the lack of a dominant strategy in the multiple equilibria space of continuous time auctions. This behavior is accentuated by eBay’s use of a hard closing time as a part of its auction design.

In an interesting and detailed investigation of just over 1000 auctions on eBay, Roth and Ockenfels observed that 18% of the bids were made in the last 60 seconds [11]. They argue that given a significant likelihood of not being able to submit a last minute bid, sniping is consistent with equilibrium, perfectly rational behavior in both private and common-value auctions. The likelihood of a last minute bid not going through could be either due to the fact that they do not get transmitted or registered through the Internet, or perhaps bidders simply forget to come back at the last minute due to some other activity, such as walking the dog, or a variety of other plausible reasons. Interestingly, between Roth and Ockenfels’ 2001 analysis and the present time, other technological developments in the form of newer sniping bidding agents have occurred. As described in the next section, these widely available free-auction sniping agents are bypassing the bidding structure that is the core of eBay’s design. If widely deployed they could turn a majority of auctions on eBay into de facto sealed-bid, one-shot auctions.

Back to Top

Emerging Bidding Agents

An often overlooked aspect of participating in online auctions is that bidding is a costly activity. Easley and Tenorio point out that the dimensions of this cost include the search cost of determining a valuation, the time cost of placing bids and monitoring the auction, and the overall opportunity cost of foregoing other alternative feasible solutions, say a posted price [7]. Automated agents, defined as “software entities that have been given sufficient autonomy and intelligence to enable them to carry out specified tasks with little or no human supervision,” are an obvious technology that could reduce bidding costs [9]. However, one has to be careful about considering the underlying incentives of who provides these agents. Incentive alignment is an emerging theme in information systems (IS) design. Ba, Stallaert and Whinston illustrate the importance of considering the interests and incentives of the users for relatively mature IS areas such as decision support systems, knowledge management, and e-business supply chain coordination [2]. This issue is particularly important, yet relatively unexplored, in the area of autonomous software agents.

eBay’s proxy bidding agent and Ubid’s bid butler are examples of agents that are provided by the auctioneer whose incentives are typically more closely aligned with that of the seller. In many cases, such as on Ubid, the auctioneer is the seller. eBay’s revenue is directly proportional to the selling price of the items it helps clear. A sign of a maturing market will be when this anomaly is recognized and bidding agents that explicitly have their incentives aligned with that of the buyer emerge.

Such is the case with bidding on eBay. Recall that eBay’s proxy bidding agent does not allow the bidder to specify the time of placing a bid. In that sniping seems to be a preferred strategy on eBay, a plethora of sniping tools have emerged that help bidders adopt this strategy. A simple Google search revealed several free as well as other commission-based sniping tools, far more than can be listed in the space of this article. We contrast two popular sniping agents from an architectural and incentive perspective.

Sniping agent architecture and incentive alignment. We define the term sniping agent as a bidding agent that allows a bidder to specify her maximum willingness to pay as well as the time at which she desires to place her bid. Both these features are visible in Figure 1 which displays the user interface of an open-source, Java-based applet, JBidWatch, available from www.jbidwatcher.com/ and a Web site called Cniper.com. Sniping is accomplished, for instance, in the latter through a pull-down menu that allows bidders to specify how many seconds before the close of the auction they wish their bid to be submitted. While both these agents provide the same sniping functionality, the differences arise in the way they install and run, and consequently on their potential incentive alignment.

JbidWatch is more technical to install, in that it requires the Java Runtime Environment (JRE) to be loaded on a user’s machine. It is a Java program that runs on a user’s machine and is devoid of any need to access a centralized server. Its repository of information is local in nature and contains a given user’s bidding activity. In contrast, Cniper.com runs on a server-side PHP (a recursive acronym for Hypertext PreProcessor) script and thereby has the ability to aggregate information across its users on a centralized database. It is reasonable to expect that the average consumer is going to be more comfortable with using an agent such as Cniper.com. On the other hand, JbidWatch’s source code is freely available and general enough for a reasonably sophisticated Java programmer to extend it to replicate other bidding strategies, particularly those for more complex auctions. For instance, consider the case of multi-unit auctions that are prevalent on Ubid. Typically, several bidders win a given multi-unit auction, and, because it is a discriminatory (pay-your-bid) pricing scheme, there can be a significant spread in the lowest and highest winning bid. Obviously, a rational consumer would maximize her surplus by winning at the lowest possible winning level, but achieving this consistently is not a trivial task [4].

Back to Top

A Natural Experiment With a Sniping Agent

To test the performance of the sniping agent JbidWatch we worked with an experienced eBay participant who has maintained an eBay identity since May 1998. This eBayer has an excellent reputation reflected in 162 positive (140 buy and 22 sell), 0 negative and two neutral ratings. These ratings have been provided by 102 unique individuals indicating a broad consensus on his reputation. The bidder has always maintained a sniping strategy, but as is apparent in Table 1, her sniping effectiveness has been greatly influenced by external non-monetary factors. The following table shows the number of auctions this bidder bid on and the winning percentage, separated into three categories of time.

The first column reflects a manual sniping period when the bidder was single as well as the period in which she got married but before she had children. The second column shows a constrained manual sniping period, the period after she had children but before she had the use of a bidding agent. The last column reflects the agent-based sniping period, after she started using JbidWatch. Anecdotal evidence, apparent in the sharp increase in number of auctions per-time-period she bid on, suggests that the sniping agent technology reduced her bidding costs. While the difference in winning percentages of the manual sniping and the agent-based sniping period are statistically insignificant, it is useful to examine the details of the transactions using the agent. The bidder claims “never to have been out-sniped with the agent.” This was confirmed by our examination of the data. The price range of the final bids of the auctions she bid on were from $11 to $32. Of the 20 she won, the average winning price was $16, and ranged from $12–$27. Of the 11 she lost, 6 of them were lost by $0.50 and in the other 5, the final price was significantly ($4–$7) higher then her maximum bid. Thus, the bidder lost only when her maximum bid was too low or when another bidder had a higher willingness-to-pay.

By all accounts, sniping using the agents seems to perform at least as well as dedicated manual sniping and the bidder nets a positive experience with the open-source bidding agent. The period of constrained sniping reflects how non-monetary factors, such as early parent-hood, can impact the outcomes of such auctions.

Back to Top

Implications

For single item auctions that run on eBay, the most immediate implication of an (anticipated) widespread adoption of newer bidding agents such as JbidWatch is that it converts the auction into a single-shot sealed bid auction. From eBay’s perspective it also bypasses the proxy bidding agent mechanism, reducing it to simply a search engine, with bidding outsourced to agents that act on behalf of bidders. Viewed simplistically, this in itself may not appear to be troubling. However, the long history of markets suggest that a key role markets fulfill is to provide institutional stability. Bakos lists the provision of an institutional infrastructure, such as a commercial code, contract law, monitoring and enforcement, to be a central function of electronic markets [3]. With this background let me stretch the imagination a bit and consider the technological and incentive implications of potentially opportunistic behavior using the emerging bidding agent technologies.

Multiple sniping agents could clog servers. First we consider a technological issue. Consider the not so unlikely scenario of several sniping agents all of which have been instructed to bid during the 10th last second of an auction. This could manifest itself in a mini denial-of-service attack on a localized portion of the listing server, resulting in none of the bids going through, and an inefficient allocation. Sellers would not realize the maximum possible revenue and the highest bidding consumer would not be allocated the item. The consequence is economic inefficiency. Consider next the case of incentive misalignment or opportunism.

Predatory and collusive agents. One of the primary objectives of good mechanism design is to prevent predatory and collusive behavior [8]. eBay has had a long and continuing history of dealing with opportunistic behavior from both the buyer and the sellers (see [12} for a recent update). The 2002 Internet Fraud Complaint Center’s annual report (available at www1.ifccfbi.gov/) cites online auction fraud, such as shill-bidding or non-delivery of items, as the single most reported offense on the Internet, making up 42.8% of all reported fraud.

Not surprisingly trust is an important issue in the context of virtual non face-to-face monetary transactions. Recent research (see databases.si.umich.edu/reputations/ for a collection of related articles) suggests that eBay’s feedback-based reputation and trust system are key ingredients in the price formation process [10]. Resnick et al. indicate a price difference of 7.6% of the selling price between an eBay seller with extremely high reputation and a newcomer with little reputation [10]. So where does the emerging agent technology fit into this picture?

The question arises which if any sniping agent should one trust? Should it be those that are capable of aggregating bid information from hundreds of buyers, many of whom may be bidding on the same auction, and may have specified the same last minute (second) bidding instance? What priority rules are the agents using and what is their feedback rating? Or should it be the one that provides Java source code that could be embedded into any Web document so that any well-meaning bidder could pretend to be a sniping agent, pay a few dollars to list on Google, and discover the entire demand curve of a big-ticket item to their bidding advantage? And where in the process is the market-maker, the lone star of the dot-com rubble? Consider the following simple, perhaps hypothetical (one can never be too sure its not going on), case of collusive behavior.

Collusive bidding against a repeat seller. One of the great themes that have driven the online auction bandwagon is that of individuals who have quit their day jobs to become the so-called “power sellers” on eBay. They depend on eBay for their livelihood, and eBay gets a recurring revenue stream from sellers with great reputations. Many sellers specialize in certain products, say Maclaren strollers, and list these day after day. It is easy to show that if there is a significant likelihood that a product is going to be auctioned repeatedly, and there exist third-party agents that can aggregate bids across several individuals, then an autonomous agent, acting on behalf of the buyers, can reduce the expected price of the product by colluding against the seller. Such tacit collusion, known as bid rotation, could be implemented by assessing the marginal bid outside the colluding set of bidders, tagging those within the colluding set who have bid higher as potential winners, and placing successive reduced bids equal to an increment above the outside marginal bid for those tagged. With minimal delay in consumption this would, on average, reduce the expected price the colluders would pay and hurt the power-sellers ability to price-discriminate. Armed with key valuation information, the agent technology could do this implicitly, without consent and with mathematical precision.

Back to Top

Recommendations

It is anybody’s guess as to the likelihood of the above-mentioned scenarios. Suffice it to say that in the midst of ongoing major corporate ethical meltdowns (Enron, Tyco and Anderson to name a few), one cannot underestimate the reach of opportunistic behavior. The central message of this article is that a mechanism design approach that can be incorporated at little cost, can go a long way in reducing the risks of the participating agents in electronic marketplaces. These would not only include the buyers and the sellers, but the market-makers themselves.

eBay’s best response, and perhaps its ticket to relevancy, is going to be to make a single-shot sealed bid auction de jure instead of de facto. It would do well to adopt incentive-compatible second-price sealed bid Vickrey auction format where the highest bidder wins and pays a price equivalent to the second highest bid. In such auctions truth-telling is a dominant strategy, and fair allocations are guaranteed so that the highest valuing consumer gets the object at its marginal value.

While we have used eBay as a representative market, the issue of smart agent-based bidding and incentive alignment is pertinent to other markets, for instance B2B, as well. In their most general form, such markets are two-sided (exchanges), multi-item, multi-unit, and multi-attribute (price is not the only dimension that matters) in nature. Agents have a role in lowering the cost of preparing, submitting, and monitoring a complex bid. Intuitively, specialized B2B markets are likely to have a lesser number of participants than typical B2C markets. Coupled with the fact that the agents are likely to have minimal coordination costs among themselves, this provides a strong incentive for collusion. Clearly further research, dealing with these more complex B2B markets, is required to determine conditions when agents are more likely to maximize their informational rents from the mechanism through collusion.

Ultimately, it is going to be in everybody’s best interest if the online auctioneers offer a full portfolio of auction design choices. To begin with, for continuous (non-sealed bid) auctions, they should offer the seller the choice of using a hard or a soft closing in the form of a going, going, gone period. In addition, they should offer a choice of the ascending English, the descending Dutch, and the sealed-bid second-price Vickrey auction. As the old saying goes, let the market decide which is the suitable mechanism for a variety of product, buyer, and domain combinations.

The outcomes of auctions are not supposed to depend on how fast one can type in a bid, or on the speed of one’s Internet connection. For that matter, they are not supposed to rest on whether one needs to take the dog for a walk.

Back to Top

Back to Top

Back to Top

Back to Top

Figures

F1 Figure 1. Sniping agent interfaces: JbidWatch (left) runs on a remote client whereas Cniper.com uses a centralized architecture.

Back to Top

Tables

T1 Table 1. Sniping history and success rate.

Back to top

    1. Ba, S. and Pavlou, P. Evidence of the effect of trust building technology in electronic markets: Price premiums and buyer behavior. MIS Quarterly (forthcoming); www.sba.uconn.edu/users/sulin/ (2002).

    2. Ba, S., Stallaert, J., and Whinston, A. B. Introducing a third dimension in information systems design: The case for incentive alignment. Information Systems Research 12, 3 (2001), 225–239.

    3. Bakos, Y. The emerging role of electronic marketplaces on the Internet. Commun. ACM 41, 8 (1998), 35–42.

    4. Bapna, R., Goes, P., and Gupta, A. Online auctions: Insights and analysis. Commun. ACM 44, 11 (2001), 42–50.

    5. Bapna, R., Goes, P., and Gupta, A. Analysis and design of business-to-consumer online auctions. Management Science 49, 1 (2003), 85–101.

    6. Coursey, D. How to snipe on eBay. ZDNET News (2000); zdnet.com.com/2100-11-522481.html.

    7. Easley, R. F., and Tenorio, R. Jump-bidding strategies in Internet auctions. Notre Dame working paper (2002); www.nd.edu/~reasley/vita.html.

    8. Klemperer, P. What really matters in auction design. Journal of Economic Perspectives (forthcoming); http://www.paulklemperer.org/ (2002).

    9. Nwana, H., Rosenschein, J., Sandholm, T., Sierra, C., Maes, P., and Guttmann, R. Agent-mediated electronic commerce: Issues, challenges and some viewpoints. Proceedings of the 2nd International Conference on Autonomous Agents, Minneapolis, MN (May 9–13, 1998), 189–196.

    10. Resnick, P., Zeckhauser, R., Swanson, J., and Lockwood, K. The value of reputation on eBay: A controlled experiment. Working paper (2002); www.si.umich.edu/~presnick/papers/postcards/.

    11. Roth, A. E., and Ockenfels, A. Last minute bidding and the rules for ending second-price auctions: Theory and evidence from a natural experiment on the Internet. Working paper (2001); www.economics.harvard.edu/~aroth/alroth.html.

    12. Wolverton, T. Fraud lingers despite eBay efforts. CNET News (2002); news.com.com/2100-1017-940427.html.

    Partial support for this research was also provided by TECI—the Treibick Electronic Commerce Initiative, OPIM/School of Business, University of Connecticut.

Join the Discussion (0)

Become a Member or Sign In to Post a Comment

The Latest from CACM

Shape the Future of Computing

ACM encourages its members to take a direct hand in shaping the future of the association. There are more ways than ever to get involved.

Get Involved

Communications of the ACM (CACM) is now a fully Open Access publication.

By opening CACM to the world, we hope to increase engagement among the broader computer science community and encourage non-members to discover the rich resources ACM has to offer.

Learn More