Evaluation Design
Talkat Bridging Prediction and Intervention Problems in Social Systems (Simons Institute, 2026) · 2026-01-12..01-18·
summary
Evaluation scores are statistical claims about value and should state the estimand and inference assumptions; common metrics hide cost and prevalence uncertainty so accuracy implicitly assumes equal misclassification costs and matching prevalence while AUC averages over score-driven thresholds. The work formalizes regret under varying cost and prevalence, proposes bounded-threshold Brier evaluations that average over an expert-specified interval using a uniform prior, warns against value-hacking, and ships BrierTools to compute these comparisons.
AI-written summary · about
chapters (9)
- 0:00Evaluation design critique
- 3:49Goodhart’s Law pitfalls
- 6:30Hacking value in design
- 10:08Real-world evaluation variability
- 16:14Regret formalism for classifiers
- 17:24Critique of accuracy and AUC
- 20:29Bounded Brier score approach
- 22:59Handling deployment uncertainty
- 24:29Q&A
related talks
46mThe Measure and Mismeasure of Fairness
40mAUC is Worthless: Lessons to Transition From Academic to Business Data Science
33mThe Silent Crash:Why Your RAG Evaluation Metrics Are Lying to You
37mSelection bias: The elephant in the room
28mMeasuring Model Fairness
36mThe Social Cost of Strategic Classification
35mEnsuring Reliable Evaluation Systems for Your ML
30mMy forecast is better than yours! What does that even mean?