Win More Effectively in Real-Time Bidding

on YouTube ↗26m

" summary

Frames bidding as surplus maximization using probabilistic models and quantile regression (pinball loss) to predict market price distributions, then searches quantiles to pick the bid that maximizes expected surplus. Model uses embeddings and a gated cross network, consumes 600 TB of data, makes 250M predictions/day with <50 ms latency, and in an A/B test keeps win rate stable while achieving 12% lower spend.

AI-written summary · about

" chapters (8)

1. 0:00Introduction and background
2. 3:57Real-time bidding basics
3. 10:25Bidding strategies comparison
4. 13:35Quantile regression approach
5. 17:15Model architecture details
6. 20:51Practical example and results
7. 22:18AB test outcomes
8. 23:21Conclusions and next steps

" related talks

• 27mEconometrics for Learning AgentsVasilis Syrgkanis · Complexity and Simplicity in Economics (Simons Institute workshop, 2015)
• 44mOffer Optimization, a Dual ApproachGonzalo Barrera Borla · PyData Córdoba Argentina 2019
• 30mData Processing in RTB: Fast, Cheap and 98% AccuratePavel Kalaydin · HighLoad++ 2014
• 29mProbabilistic demand forecasting at scaleHagop Dippel · PyData Global 2022
• 29mAI and ML for Computational AdvertisingAbraham Bagherjeiran · Global Artificial Intelligence Conference 2020
• 54mOn How Machine Learning and Auction Theory Power Facebook AdvertisingEric Sodomka · Algorithmic Game Theory and Practice 2015
• 27mModelling the extreme using Quantile RegressionMassimiliano Ungheretti · PyData Global 2020
• 40mOnline Ad Allocation: Robust Optimization for Repeated AuctionsVahab Mirrokni · Optimization and Decision-Making Under Uncertainty (Simons Institute, September 2016)