Dynamic Ad Valuation In Real Time Bidding

Abstract

The online advertising world runs millions of daily real time auctions within milliseconds. The current framework in the auction focuses on modeling individual user behavior. Such analysis is sporadic, complex, and difficult to analyze beyond simplified performance metrics. This thesis instead aims to value advertisements on a holistic level across the aggregate bidding behavior. Inspired by financial models, the ad value distribution is modeled upon information contained in the auctions and pay price rather than a specific user profile. To create an ad valuation, two models are implemented and contrasted. One is a rolling model with a functional log-normal form that assumes information is contained in the submitted bids. The other is a gradient boosted tree model, which leverages the large amount of information in the real time auction logs. These valuations are applied to set an optimal auction reserve price and to inform bidding strategies.