HouseholdBench tests 13 LLMs on predicting household economic behavior; GBDT still wins
daveholtz · x · 2026-10-20
Researchers from University of Michigan, LSE, UC Berkeley and MobLab introduce HouseholdBench, a benchmark testing whether LLMs can predict how households spend, work and respond to policy.
Setup:
- Unites 6 U.S. household surveys and 32 prediction tasks spanning numeric, categorical and probabilistic outcomes across consumption, income, labor, expectations and housing
- Conditions on past behavior, demographics and macroeconomic conditions, including policy-response tasks
Findings:
- 13 proprietary and open-weight LLMs evaluated: most beat a no-change baseline, with the best model cutting numeric-outcome error by 12.2%
- Gradient-boosted trees rank first on most tasks; leading proprietary LLMs approach them, open-weight models lag
- LLMs show systematic over- and under-prediction across tasks
- Fine-tuning plus aggregating 16 predictions per observation lets a 4B open-weight model match proprietary models, generalizing even to held-out policy-response tasks
Datasets, code and a leaderboard are released.
More from Research
- Community pushes Quasi-RH zero-free record to θ ≤ 0.874810282, Lean proofs pending — IgorCarron · 2026-10-21
- 3DRAE: Any 3D Scene in 1K Tokens — Diffusion in 3D Latent Space for Scene Generation — Dongxu Wei · 2026-10-21
- Deepak Pathak: There Is No Perfect Data Source for Robotics — chris_j_paxton · 2026-10-21
- AI models cluster into three personality families that cross lab boundaries, finds AA — ArtificialAnlys · 2026-10-21
- Discrete action representations match or beat continuous ones for robot policies — GuanyaShi · 2026-10-21
- ETSAM adapts Segment Anything 2 to treat 3D scans as video, scoring 0.812 on membrane segmentation — bravo_abad · 2026-10-21