Prediction and Welfare in Ad Auctions

We study how standard auction objectives in sponsored search markets are affected by refinement in the prediction of ad relevance (click-through rates). As the prediction algorithm takes more features into account, its predictions become more refined; a natural question is whether this is desirable from the perspective of auction objectives. Our focus is on mechanisms that optimize for a convex combination of economic efficiency and revenue, and our starting point is the observation that the objective of such a mechanism can only improve with refined prediction, making refinement in the best interest of the search engine. We demonstrate that the impact of refinement on market efficiency is not always positive; nevertheless we are able to identify natural – and to some extent necessary – conditions under which refinement is guaranteed to also improve economic efficiency. Our main technical contribution is in explaining how refinement changes the ranking of advertisers by value (efficiency-optimal ranking), moving it either towards or away from their ranking by virtual value (revenue-optimal ranking). These results are closely related to the literature on signaling in auctions.