Core algorithms inside a demand-side platform: bidding, prediction, pacing and fraud

You are interviewing for a product manager role at Rakuten and are asked about the algorithms behind programmatic advertising. Explain the main families of algorithms a demand-side platform (DSP) relies on when it buys ad impressions on behalf of advertisers. For each family, describe the decision it supports, the data it learns from, and what makes it hard to get right. ```hint Follow one impression Trace a single impression from the moment an exchange sends the bid request, through the bid and the auction, to the advertiser's report. List each decision the DSP has to make along the way, then ask what kind of model or optimizer that decision needs. ``` ### Clarifying Questions - Should the answer cover only real-time bidding on open exchanges, or also private marketplace and programmatic guaranteed deals? - What goals do the advertisers optimize for: cost per acquisition or return on ad spend, video completions, or reach at a target cost per thousand impressions? - Do the exchanges the DSP buys from run first-price or second-price auctions? - How much identity signal can we assume, such as logged-in IDs, mobile advertising IDs or third-party cookies, given browser and operating-system privacy limits? ### What a Strong Answer Covers - Where the DSP sits in the programmatic chain and the decisions it makes for every bid request under a tight latency budget. - Prediction models (click, conversion, video completion, viewability) and how their calibrated outputs become a bid value for each campaign goal. - Auction and budget algorithms: bid shading or win-rate modeling, budget pacing, frequency control, and choosing among eligible campaigns and creatives. - Audience and identity models (cross-device matching, lookalikes, purchase intent) and quality controls such as invalid-traffic and brand-safety detection. - Measurement and attribution that separates impressions that caused outcomes from those that only preceded them, plus the product trade-offs among these systems. ### Follow-up Questions - After a major exchange switched from second-price to first-price auctions, an advertiser's cost per conversion on that exchange rose. Which part of the DSP would you examine first, and why? - An advertiser complains that a campaign spent slowly in its first week. How would you explain what happened, and which levers could they change? - As third-party cookies and mobile advertising IDs become less available, which of these algorithms degrade most, and where would you invest?