Core Concepts
Simulating advertiser behavior using online learning algorithms can provide useful insights into complex real-world ad auction environments, where standard equilibrium analysis is infeasible.
Abstract
The paper presents a simulation-based approach to modeling advertiser behavior in online ad auctions, which departs from the canonical auction models in several key ways:
Advertiser values and click-through rates can depend on the user's search query, but advertisers can only partially target their bids to specific queries.
Advertisers do not know the number, identity, and precise value distribution of competing bidders.
Advertisers only receive partial, aggregated feedback about auction outcomes.
Payment rules are only partially known to bidders.
The authors demonstrate that, despite these complexities, simulating advertisers as agents governed by no-regret learning algorithms can still yield useful insights. Specifically:
In multi-query environments, soft-floor reserve prices can improve revenues compared to standard auction formats, even with symmetric bidder types.
In single-query asymmetric environments, soft-floors do not outperform suitably chosen standard reserve prices.
The simulation-based approach can be used to infer advertiser value distributions from observed bid data, without relying on equilibrium assumptions.
The authors provide implementation code and illustrate their approach using both synthetic and real-world data from an e-commerce website.
Stats
The paper does not contain any explicit numerical data or statistics. The key insights are derived from simulation results.
Quotes
The paper does not contain any direct quotes.