householdbench/householdbench-eval

HouseholdBench: evaluation sample

Code and rebuild instructions · Website and leaderboard · Full data for the 17 tasks · Paper

This dataset holds the test prompts scored in the paper for the 17 tasks whose source data we may redistribute: 3,200 pre-cutoff and 1,500 post-cutoff prompts, 4,700 of the paper's 8,579. The other 15 tasks use IPUMS, PSID, Michigan Surveys of Consumers or ICPSR data, which we may not redistribute. To evaluate on them, download those files under your own access and rebuild the tasks with the code repository, which builds the test prompts for all 32 tasks.

A score on this dataset covers 17 of 32 tasks, so it is not comparable with the 32-task leaderboard on the website.

Tasks

Task	Topic	Survey	Mode	Target type	Test, pre-cutoff	Test, post-cutoff	Where to get it
cons_cex_categories	Consumption and saving	CEX	Baseline	Numeric	200	–	This dataset
cons_cex_rebate01	Consumption and saving	CEX	Policy response	Numeric	200	–	This dataset
cons_cex_sspay	Consumption and saving	CEX	Policy response	Numeric	200	–	Rebuild with the code
cons_cex_stimulus08	Consumption and saving	CEX	Policy response	Numeric	200	–	This dataset
cons_cex_total	Consumption and saving	CEX	Baseline	Numeric	200	–	This dataset
cons_psid_jobloss	Consumption and saving	PSID	Policy response	Numeric	200	–	Rebuild with the code
cons_psid_wealth	Consumption and saving	PSID	Baseline	Numeric	200	–	Rebuild with the code
cons_sce_growth	Consumption and saving	SCE	Baseline	Numeric	200	200	This dataset
cons_sce_shock	Consumption and saving	SCE	Policy response	Numeric	200	200	This dataset
house_census_move	Housing and location	Census	Baseline	Categorical	200	–	Rebuild with the code
house_psid_owner	Housing and location	PSID	Baseline	Categorical	200	–	Rebuild with the code
house_sce_financing	Housing and location	SCE	Policy response	Numeric	100	–	This dataset
house_sce_lockin	Housing and location	SCE	Policy response	Probabilistic	100	–	This dataset
house_sce_move	Housing and location	SCE	Baseline	Probabilistic	200	200	This dataset
income_cps_displace	Income dynamics and household resources	CPS	Policy response	Numeric	200	–	Rebuild with the code
income_mich_finance	Income dynamics and household resources	Michigan	Baseline	Numeric	200	200	Rebuild with the code
income_psid_earnings	Income dynamics and household resources	PSID	Baseline	Numeric	200	–	Rebuild with the code
income_sce_growth	Income dynamics and household resources	SCE	Baseline	Numeric	200	200	This dataset
income_sce_policy	Income dynamics and household resources	SCE	Policy response	Categorical	200	–	This dataset
labor_cps_displace	Labor supply, job search, and retirement	CPS	Policy response	Categorical	200	–	Rebuild with the code
labor_cps_jobfind	Labor supply, job search, and retirement	CPS	Baseline	Categorical	200	200	Rebuild with the code
labor_cps_retire	Labor supply, job search, and retirement	CPS	Baseline	Categorical	200	200	Rebuild with the code
labor_cps_separation	Labor supply, job search, and retirement	CPS	Baseline	Categorical	200	200	Rebuild with the code
labor_cps_ui	Labor supply, job search, and retirement	CPS	Policy response	Categorical	200	–	Rebuild with the code
labor_psid_addedworker	Labor supply, job search, and retirement	PSID	Policy response	Numeric	79	–	Rebuild with the code
labor_sce_offer	Labor supply, job search, and retirement	SCE	Baseline	Categorical	200	–	This dataset
labor_sce_reswage	Labor supply, job search, and retirement	SCE	Policy response	Numeric	200	–	This dataset
labor_sce_risk	Labor supply, job search, and retirement	SCE	Baseline	Probabilistic	200	200	This dataset
labor_sce_search	Labor supply, job search, and retirement	SCE	Baseline	Probabilistic	200	100	This dataset
macro_mich_outlook	Macroeconomic expectations	Michigan	Baseline	Numeric	200	200	Rebuild with the code
macro_sce_revision	Macroeconomic expectations	SCE	Policy response	Numeric	200	200	This dataset
macro_sce_uncertainty	Macroeconomic expectations	SCE	Baseline	Probabilistic	200	200	This dataset
In this dataset (17 tasks)					3,200	1,500
Rebuild with the code (15 tasks)					2,879	1,000
Paper total (32 tasks)					6,079	2,500

Use

from datasets import load_dataset

ds = load_dataset("householdbench/householdbench-eval", "cons_sce_growth")  # splits: pre_cutoff, post_cutoff

The code repository explains the record fields and how to download, query and score these prompts.

Known issues

• Two prompt templates contain wording errors that every evaluated model saw, so they are kept to preserve comparability: "about the same than" (SCE household-finance questions) and a doubled "Your household your household" (cons_sce_growth).
• macro_sce_revision record IDs in this release drop the __m01_YYYYMM__m12_YYYYMM suffix used during the paper's evaluation runs; the records are otherwise identical.

License

The source surveys keep their own terms; no single license covers the data. LICENSE.md lists each source's terms and the attribution notices you must keep when you use or share these files. The Survey of Consumer Expectations (SCE) license requires these notices:

Some survey questions were taken or adapted from the Survey of Consumer Expectations, © 2013-2019 Federal Reserve Bank of New York (FRBNY). The SCE questions are available without charge at https://www.newyorkfed.org/microeconomics/sce and may be used subject to license terms posted there. FRBNY did not participate in or endorse HouseholdBench, and FRBNY disclaims any responsibility or legal liability for the administration of the survey and the analysis and interpretation of data collected.

Source: Survey of Consumer Expectations, © 2013-2019 Federal Reserve Bank of New York (FRBNY). The SCE data are available without charge at https://www.newyorkfed.org/microeconomics/sce and may be used subject to license terms posted there. FRBNY disclaims any responsibility or legal liability for this analysis and interpretation of Survey of Consumer Expectations data.

Citation

@article{huang2026householdbench,
  title   = {{HouseholdBench}: Evaluating Large Language Models as Predictors of Household Economic Behavior},
  author  = {Huang, Jin and Ferreras Garrucho, Diego and Xie, Yutong and Yuan, Walter M. and Mei, Qiaozhu and Lian, Chen and Hazell, Jonathon},
  journal = {arXiv preprint arXiv:2610.07563},
  year    = {2026}
}