Bidder Feedback in First-Price Auctions for Video Advertising

TL;DR

This paper extends the concept of minimum-bid-to-win feedback from first-price display advertising auctions to complex combinatorial video ad auctions, providing theoretical characterizations and empirical insights into bidding dynamics.

Contribution

It generalizes the minimum-bid-to-win concept to combinatorial auctions and characterizes the set of optimal bid updates using bicooperative game theory, with empirical analysis on real data.

Findings

Characterization of bid updates as the core of a bicooperative game

Linear programming formulation for the assignment problem case

Empirical analysis of bidding dynamics and convergence in video ad auctions

Abstract

In first-price auctions for display advertising, exchanges typically communicate the "minimum-bid-to-win" to bidders after the auction as feedback for their bidding algorithms. For a winner, this is the second-highest bid, while for losing bidders it is the highest bid. In this paper we investigate the generalization of this concept to general combinatorial auctions, motivated by the domain of video advertising. In a video pod auction, ad slots during an advertising break in a video stream are auctioned all at once, under several kinds of allocation constraints such as a constraint on total ad duration. We cast the problem in terms of computing bid updates (discounts and raises) that maintain the optimality of the current allocation. Our main result characterizes the set of joint bid updates with this property as the core of an associated bicooperative game. In the case of the assignment problem--a special case of video pod auctions--we provide a linear programming characterization of this bicooperative core. Our characterization leads to several candidates for a generalized minimum-bid-to-win. Drawing on video pod auction data from a real ad exchange, we perform an empirical analysis to understand the bidding dynamics they induce and their convergence properties.