Internet Auctions with Artificial Adaptive Agents

The growing number of auction sites on the internet enable game theorists to ask strategic questions on rationality of the observed bidding behavior. The most popular of them (eBay, Amazon.com, AuctionWatch.com and Yahoo!, etc.) operate under similar sets of rules with seemingly small differences. They implement types of ascending-bid auctions and can strategically differ from a single-round sealed-bid second-price auction.\t The major difference between auction formats is about their ending procedures (as described in the paper). Most of the bidding activity is observed in the final hour of the auctions. Hence, the seemingly small difference in auction ending can cause later bids in the eBay auctions than Amazon.com. In this paper, I try to investigate the evolution of bidding patterns in internet auctions. I investigate evolutionary stability of late and multiple bidding in the private-value and common-value frameworks. I use adaptive artificial agent markets in the analysis. I seek similarities between the simulation data and actual human bidding behavior. I implement discrete finite time, sequential and repeated auctions in our simulations. In this paper, I show that common- and private-value auctions can evolutionarily lead to multiple and late bidding. I observe that artificial agent late bidding is in much more frequency in eBay auctions than Amazon.com in the private-value format (i.e. computer auctions). With common values (i.e. antique auctions), expert artificial bidders bid later more frequently than naive ones. The eBay auctions stage more frequent late bidding in the eBay auctions than in Amazon.com when there is a single naive bidder. Otherwise, auctions in Amazon.com cause more late bidding. I observe that eBay auctions generate in general less average revenue for sellers. They also cause more average profit for bidders in the private-value model. For the common-value model, bidders are indifferent betIen the two formats: each can dominate the other for different number of bidders. I also test the robustness of the results under different sets of parameters.\t