Reproduce “Online Causal Inference for Advertising in Real-Time Bidding Auctions” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Caio Waisman, Harikesh S. Nair, Carlos Carrion
This research from Kellogg, Northwestern University, and Georgia Institute of Technology introduces an online method for causal inference in real-time bidding advertising, leveraging auction theory to concurrently estimate ad effectiveness and manage experimentation costs. The Bidding Thompson Sampling (BITS) algorithm, when evaluated against various benchmarks, achieved minimal regret and provided accurate Average Treatment Effect estimates in both second-price and first-price auction simulations, often outperforming other adaptive and non-adaptive methods.
Key insights
- In second-price auctions, the optimal bid for an advertiser is exactly equal to the maximum of zero and the Average Treatment Effect (ATE) of the ad, providing direct identification of ATE from optimal bids.
- In first-price auctions, the optimal bid is related to the ATE through a transformation involving the reverse hazard rate of the highest competing bid's distribution, allowing for ATE identification by recovering optimal bids and competitive dynamics.
- Embedding experimental design within the microfoundations of economic theory (auction theory) provides structural insights that enable more effective and economically efficient online experimentation, aligning inference and economic goals.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.