Econometrics for Learning Agents

on YouTube ↗27m

" summary

Claim: bidders are best-modeled as vanishing-regret learning agents, not Nash equilibria, so values can be inferred from bid histories. Method: convert regret inequalities into a convex rationalizable set, recover it via its support function which reduces inference to a single curve of increase-in-cost vs increase-in-click probability. Data and results: applied to nine frequent changers on Bing, point estimates use the smallest multiplicative regret; average observed bids are ~60% of predicted values, many listings show near-zero regret while others reach ~50%; simulations with multiplicative-weight learners recover true values.

AI-written summary · about

" chapters (23)

1. 0:00Basic econometric task
2. 1:06Requires behavioral assumption
3. 1:39Behavior is far from stable
4. 2:22Sophisticated bidding tools
5. 2:57Learning agent model
6. 4:05No-regret assumption
7. 6:32Why no-regret?
8. 7:41High-level approach
9. 9:10Sponsored search auction setting
10. 13:12Examples from real data
11. 13:26Characterizing the rationalizable set
12. 14:54Inference of rationalizable set
13. 17:02Data Description
14. 17:36Point prediction in real data
15. 18:07Distribution of bid shading: Average bid / Predicted Value
16. 18:38Distribution of smallest rationalizable multiplicative regret
17. 19:18Correlation: regret vs bid shading
18. 20:24Setup
19. 20:46Rationalizable Sets over Time
20. 22:12Distribution of Prediction Errors
21. 22:41Distribution of Predicted Values
22. 23:17Conclusion
23. 23:38Next Steps/Directions - Dynamic analysis of nationalizable sets

" related talks

• 43mLearning about AuctionsAvrim Blum · Uncertainty in Computation 2016
• 48mRobust Inference for Games via Theoretical GuaranteesDenis Nekipelov · Algorithmic Game Theory and Practice 2015
• 01h 17mOnline Reinforcement Learning and RegretSean Sinclair, Christina Lee Yu · Data-Driven Decision Processes Boot Camp 2022
• 15mFundamental Limits Of Learning In Data-Driven ProblemsYanjun Han · Meet the Fellows Welcome Event
• 04mOn gradient-based methods for finding game-theoretic equilibriaMichael Jordan · O'Reilly Artificial Intelligence Conference 2019 - San Jose
• 09mMechanism Design Via Machine Learning: Overfitting, Incentives, and PrivacyEllen Vitercik · Meet the Fellows Welcome Event
• 26mWin More Effectively in Real-Time BiddingAn Tran · bit summit 2025
• 48mOutposts Between Worst- and Average Case Analysis: A Case Study in Auction DesignTim Roughgarden · Uncertainty in Computation 2016