Working Paper: CEPR ID: DP18009
Authors: Francesco Decarolis; Gabriele Rovigatti; Michele Rovigatti; Ksenia Shakhgildyan
Abstract: Artificial Intelligence Algorithms differ in their capabilities depending on the type of available data. We explore how this dimension informs two key design features: memory and updating (or learning) rule. We apply this insight to the case of online search auctions, where platforms control the type of data given to advertisers about their rivals’ bids. Simulated experiments with asymmetric bidders reveal that, when less detailed information is available to train the algorithms, the auctioneer revenues improve substantially. This might explain why hosting platforms have recently reduced the information disclosed, an industry trend known as data obfuscation. Finally, we explain how our findings are linked to dynamic strategies and to the possibility of calculating counterfactuals, as well as to the responsiveness of the algorithms to the actions of other players.
Keywords: auctions; procurement; collusion; data privacy
JEL Codes: C73; D82; D83; D18
Edges that are evidenced by causal inference methods are in orange, and the rest are in light blue.
| Cause | Effect |
|---|---|
| data availability (C81) | auction outcomes (D44) |
| less detailed information (Y50) | auctioneer revenues (D44) |
| not revealing competitor bids (D44) | auctioneer revenues (D44) |
| not revealing competitor bids (D44) | decline in advertiser rewards (D49) |
| full information scenario (D83) | higher rewards for advertisers (M52) |
| no information scenario (D89) | lower advertiser rewards (D44) |