Timeout analysis in real-time bidding for ad-tech company

Abstract

Real-time bidding is the popular way to exchange online ads through instantaneous auctions. To ensure best user-experience and highest profitability, there need to be a time limit to prevent excessively slow auctions. That limit is called timeout value. The timeout value should be chosen carefully to balance the benefits of users, advertisers and publishers. This study analyzes bidding data collected by an ad-tech company to determine which factors strongly affect the timeout result and whether those effects are positive or negative. Relevant academic papers in the field are scrutinized to find out promising factors which should be examined further. Data of those factors will be visualized in the descriptive analysis to give general ideas about the variables and their distribution. After that, random forest and logistic regression models are used to rank the factors and estimate their significance on bid timeout result. The study finds out that number of networks, browsers and bid networks are categories that have strongly relevant factors. Some other categories like geo-location, device families and placement types also have certain relationship with the timeout result. Therefore, the paper suggests that timeout values should be set with high level of details, for example, per browser, per device family or per geo-location, to reflect the effects of those factors on timeout result.