On Machine Learning for Prediction and Prioritization in the Allocation of Scarce Societal Resources
Talkat Bridging Prediction and Intervention Problems in Social Systems (Simons Institute, 2026) · 2026-01-12..01-18·
summary
Shows that ML is most useful as measurement not as final allocator: BART models estimate household re‑entry probabilities from St. Louis intake data (2007–2014), then weekly optimal matching lowers re‑entry about 5% relative while helping 35% of households and harming 21%. A 520‑person Mechanical Turk study finds ~2/3 are outcome‑oriented; caseworker decisions are predictable by a short tree (AUC ~0.76) but a gradient boosting model reaches AUC 0.945, and discretionary moves target less vulnerable clients.
AI-written summary · about
chapters (13)
- 0:00Introduction
- 1:05Nevada student risk prediction
- 4:46AI roles in resource allocation
- 8:08Economics vs computer science
- 9:32Local Justice theory
- 10:38Homelessness allocation systems
- 13:12Prioritization criteria
- 16:53Bureaucratic scoring
- 17:59Matching interventions
- 20:37AI predictions vs decisions
- 25:28Caseworker discretion
- 32:33Political constraints
- 34:24Acknowledgments
related talks
12mOn the Use of ML for Blackbox System Performance Prediction
42mOperationalizing predictive analytics for homelessness prevention in Los Angeles
34mMachine Learning to know if you R coming
27mData Science and ML with an emphasis on Banking...
21mDemystifying Machine Learning for Programmatic UA & Retargeting
25mAutomating Operations with ML
25mThe Uncanny Valley of ML
44mHandling predictive uncertainty in Machine Learning