Синяк А.А., Гизатулин А.М. - Methodological problems modeling dynamic of indicators online auctions

This theme is very actual for modern economy of Ukraine. Some online auctions has appeared in our country during last years (for example, Aukro). We also can have an access to foreign online auctions. In Ukraine this theme has bee researched by Gratz; also it has been discussed by foreign scientists (for example, Lucking-Reiley, Kauffman, Wood).

We contemplate this problem from different points of view:

indetification and quantifiation of new bidding behavior and phenomena, such as bid sniping and bid shilling;
creation of a taxonomy of bidder types;
development of descriptive probabilistic models to capture bidding and bidder activity, as well as bidder behavior in terms of bid timing and amount;
another stream of research focuses on the price evolution during an online auction.

A key reason for the booming of empirical online auctions research is the availability of data: lots and lots of data! However, while data open the door to investigating new types of research questions, they also bring up new challenges. Some of these challenges are related to data volume, while others reflect the new structure of Web data. Both issues pose serious challenges for the empirical researcher.

One major aspect is the combination of temporal and cross-sectional information. Online auctions (e.g., eBay) are a point in case. Online auctions feature two fundamentally different types of data: the bid history and the auction description. The bid history lists the sequence of bids placed over time and as such can be considered a time series. In contrast, the auction description (e.g., product information, information about the seller, and the auction format) does not change over the course of the auction and therefore is cross-sectional information.

The analysis of combined temporal and cross-sectional data poses challenges because most statistical methods are geared only toward one type of data. Moreover, while methods for panel data can address some of these challenges, these methods typically assume that events arrive at equally spaced time intervals, which is not at all the case for online auction data. In fact, Web-based temporal data that are user-generated create nonstandard time series, where events are not equally spaced. In that sense, such temporal information is better described as a process. Because of the dynamic nature of the Web environment, many processes exhibit dynamics that change over the course of the process. On eBay, for instance, prices speed up early, then slow down later, only to speed up again toward the auction end. Classical statistical methods are not geared toward capturing the change in process dynamics and toward teasing out similarities (and differences) across thousands (or even millions) of online processes [1].

Another challenge related to the nature of online auction data is capturing competition between auctions. Consider again the example of eBay auctions. On any given day, there exist tens of thousands of identical (or similar) products being auctioned that all compete for the same bidders. For instance, during the time of writing, a simple search under the keywords “Apple iPod” reveals over 10,000 available auctions, all of which vie for the attention of the interested bidder. While not all of these 10,000 auctions may sell an identical product, some may be more similar (in terms of product characteristics) than others. Moreover, even among identical products, not all auctions will be equally attractive to the bidder due to differences in sellers’ perceived trustworthiness or differences in auction format. For instance, to bidders that seek immediate satisfaction, auctions that are 5 days away from completion may be less attractive than auctions that end in the next 5 minutes. Modeling differences in product similarity and their impact on bidders’ choices is challenging. Similarly, understanding the effect of misaligned (i.e., different starting times, different ending times, different durations) auctions on bidding decisions is equally challenging and solutions are not readily available in classical statistical tools.

Another challenge to statistical modeling is the existence of user networks and their impact on transaction outcomes. Networks have become an increasingly important component of the online world, particularly in the “new web,” Web 2.0, and its network-fostering enterprises such as Facebook, MySpace, and LinkedIn. Networks also exist in other places (although less obviously) and impact transaction outcomes. On eBay, for example, buyers and sellers form networks by repeatedly transacting with one another. This raises the question about the mobility and characteristics of networks across different marketplaces and their impact on the outcome of eCommerce transactions. Answers to these questions are not obvious and require new methodological tools to characterize networks and capture their impact on the online marketplace [2].

The most popular methods of data collection are Web crawling and Web services. These two technologies generate large amounts of rich, high-quality online auction data. The most important step in data analysis is data exploration. While the availability of huge amounts of data often tempts the researcher to directly jump into sophisticated models and methods, we consider that it is of extreme importance to fist understand one’s data, and to explore the data for patterns and anomalies.

Another important facet is the concurrent nature of online auctions and their competition with other auctions. There are a lot of methods for visualizing the degree of auction concurrency as well as its context (e.g., collection period and data volume). Semi-continuous data.is unusual data structure that can often be found in online auctions. These data are continuous but contain several “too-frequent” values.

We consider that there are three types of models of online auctions data. The first – and most basic – model only uses information that is available from within the auction to predict the outcome of that auction. The second model builds upon the fist model and considers additional information about other simultaneous auctions. However, the information on outside auctions is not modeled explicitly. The last – and most powerful – model explicitly measures the effect of competing auctions and uses it to achieve better forecasts [3].

Discussing and researching this problem is very important for Ukraine, because it can stimulate the development of Ukrainian science and economy.

Literature

1. Abraham, C., Cornillion, P. A., Matzner-Lober, E., and Molinari, N. Unsupervised curve-clustering using B-spline. Scandinavian Journal of Statistics, 2003. – 700 р.
2. Banks, D. and Said, Y. Data mining in electronic commerce. Statistical Science, 2006. – 350 р.
3. Dellarocas, C. and Narayan, R. A statistical measure of a population’s propensity to engage in post-purchase online word-of-mouth. Statistical Science, 2006. – 402 р.