Optimal Bidding and Experimentation for Multi-Layer Auctions in Online Advertising

Abstract

The digital advertising industry heavily relies on online auctions, which are mostly of first-price type. For first-price auctions, the success of a good bidding algorithm crucially relies on accurately estimating the highest bid distribution based on historical data which is often censored. In practice, a sequence of first-price auctions often takes place through multiple layers, a feature that has been ignored in the literature on data-driven optimal bidding strategies. In this paper, we introduce a two-step algorithmic procedure specifically for this multi-layer first-price auction structure. Furthermore, to examine the performance of our procedure, we develop a novel inference scheme for A/B testing under budget interference (an experimental design which is also often used in practice). Our inference methodology uses a weighted local linear regression estimation to control for the interference incurred by the amount of spending in control/test groups given the budget constraint. Our bidding algorithm has been deployed in a major demand-side platform in the United States. Moreover, in such an industrial environment, our tests show that our bidding algorithm outperforms the benchmark algorithm by 0.5% to 1.5%.

Keywords: advertising science, bidding, causal inference, interference, two-sided marketplace