Research by Wolfgang Jank and Galit Shmueli
With the rise of e-Bay, online auctions have become an everyday part of life for many people. In some sense, these auctions are a statistician’s dream, containing a huge amount of publicly available data with literally millions of transactions taking place every day. A main feature of online auction data is the change in dynamics of the bidding process, which can move very slowly at some points during the auction only to have a rapid flurry of bids just before the auction closes.
Wolfgang Jank, associate professor of statistics, and Galit Shmueli, associate professor of statistics, with Shanshan Wang, analyst with DemandTec, a California software firm, have developed a dynamic forecasting model to predict price in online auctions. The model operates during the live auction and forecasts price at a future time point during the auction as well as its final price. The model uses both information available at the start of the auction, such as opening price, product characteristics, and seller reputation, as well as information only available after the auction starts, such as the amount of competition and current price level.
Traditional methods of forecasting prices are hard to apply and are not very accurate for online auctions, because they don’t take into account the dramatic changes in auction dynamics. Price changes in online auctions do not happen at a steady pace. Bids arrive in unevenly spaced time intervals, sometimes coming fast and furious and other times just a trickle.
Jank and Shmueli have accounted for the challenges of unevenly spaced data by treating the price evolution in an auction as a functional object, using a functional data framework with appropriate smoothing techniques. This allows the model to represent the extremely unevenly spaced series of bids in a compact form, estimate price dynamics via the derivatives of smooth functional objects and integrate this dynamic information into the forecaster, and incorporate both static and time-varying information about the auction into the forecasting system.
The authors used data from 190 seven-day auctions of both a high-priced item (a Microsoft Xbox) and a low-priced item (the book Harry Potter and the Half-Blood Prince). For each auction they collected bid history to learn the temporal order and magnitude of the bids. This is the information typically used by traditional forecasting models. But they also collected data on the auction format, the product characteristics, and seller attributes. The study was funded in part by a grant from the National Science Foundation and a grant from the Smith School’s Center for Electronic Markets and Enterprises (CEME).
“The innovative part of this model is not just that it measures not just where the price will end up, but how fast the price is moving along this path,” says Jank. The model produces forecasts with low errors, and it outperforms standard forecasting methods, which severely under-predict price evolution.
The ability to accurately predict online auction dynamics, as well as the final auction price, has the potential to be a useful tool for both sellers and purchasers. The ability to accurately predict the dynamics and final price of online auctions has benefits for both sellers and purchasers. Dynamic price scoring would allow auctions to be ranked by lowest expected price, which would in turn allow purchasers to focus their time and energy on just those few auctions that promise the lowest price. Companies that re-sell materials on eBay, known as eBay stores or trading assistants, could also make use of price forecasting. These stores sell materials on behalf of individuals who do not want to use eBay directly. Ongoing price forecasting would allow these stores to pay their customers for their merchandise before the close of the auction, in effect offering faster liquidation for their customers.
Seller insurance also becomes possible with dynamic price forecasting. “Imagine that you’re a seller planning to auction your iPhone on eBay. The company running the auction could use dynamic forecasting to predict what the final price of the auction would likely be, and give you the option to purchase insurance that would guarantee you would at least receive a certain minimum price through the auction,” says Shmueli. “Because the model is dynamic, taking into account information as the auction is ongoing, they could offer you one price for insurance on the first day, and another on the second or third day, depending on how the auction is going.”
“Explaining and Forecasting Online Auction Data Prices and Their Dynamics Using Funcational Data Analysis” was published in the Journal of Business and Economic Statistics.
|Table of Contents||Next Article|