Mystique: A Budget Pacing System for Performance Optimization in Online Advertising
Online advertising plays a pivotal role in sustaining the accessibility of free content on the Internet, serving as a primary revenue source for websites and online services. This dynamic marketplace sees advertisers allocating budgets and competing for the opportunity to present ads to users engaging with web pages, online services, and mobile apps. Modern online advertising often employs first-price auctions to determine ad placements. Yet, conducting auctions as isolated events in a greedy manner, may lead to sub-optimal results, necessitating some form of budget pacing. Traditionally, budget pacing has been achieved through hard throttling, where ads or campaigns are selectively made eligible for each auction using a biased coin-toss with a specified probability (or pacing-signal). More recently, the pacing signal has been leveraged to soft throttle ads, and is used as a multiplicative factor on their bids, thus enabling participation in all auctions but with potentially modified bids.In this study, we introduce Mystique, a "soft" throttling-based budget pacing system. Mystique operates on two levels: it utilizes spending data to establish a daily target spending curve for each campaign, and continuously updates a pacing signal to align the actual spending with this curve. Our offline evaluation in a complexsimulated marketplace, demonstrates Mystique's ability to outper-form several baseline algorithms, enabling budget depletion while securing more opportunities. Mystique has been in production for several years now, serving a major native advertising marketplace,and successfully pacing over one billion USD annually.
