PaperlayerView exact revision on arXiv

Source-linked AI summary

HouseholdBench: Evaluating Large Language Models as Predictors of Household Economic Behavior

Jin Huang, Diego Ferreras Garrucho, Yutong Xie, Walter M. Yuan, Qiaozhu Mei, Chen Lian, Jonathon Hazell

arXiv:2610.07563v1cs.CL

TL;DR

Existing evaluations do not comprehensively cover household surveys and outcomes or test how households respond to changing economic conditions. HouseholdBench uses six public U.S. household surveys to construct 32 tasks, including policy-response tasks testing predictions of household behavior under external economic or policy changes. Across tasks, LLMs still do not outperform XGBoost overall, although Fable 5.1 comes close; fine-tuning and aggregating 16 predictions lets small open-weight models outperform leading frontier LLMs on numeric baseline tasks.

  • Problem

    Existing evaluations do not comprehensively cover household surveys and outcomes or test how households respond to changing economic conditions.

    Source: §1 · p. 1

  • Method

    HouseholdBench uses six public U.S. household surveys to construct 32 tasks, including policy-response tasks testing predictions of household behavior under external economic or policy changes.

    Source: §1 · p. 2, §2.2 · p. 4

  • Results

    Across tasks, LLMs still do not outperform XGBoost overall, although Fable 5.1 comes close; fine-tuning and aggregating 16 predictions lets small open-weight models outperform leading frontier LLMs on numeric baseline tasks.

    Source: §5.2 · p. 9, §6 · p. 10

  • Takeaways & Limitations

    HouseholdBench provides a comprehensive way to assess LLMs as predictors of household behavior using household survey data.

    Source: §6 · p. 10

Abstract

from arXiv · show

Large language models (LLMs) have the potential to meet a key goal in economics: a quantitative model of household decision making, across a variety of settings. Yet existing evaluations cover few surveys and outcomes, and do not study how households adjust to changing economic conditions. We introduce a new evaluation, HouseholdBench, which unites 6 U.S. household surveys and 32 prediction tasks spanning numeric, categorical and probabilistic outcomes, related to consumption, income, labor, expectations, and housing. Using past behavior, demographics and macroeconomic conditions, the tasks test whether LLMs predict behavior, including how households adjust to changes in various policies. We evaluate 13 proprietary and open-weight LLMs against a no-change baseline and a gradient-boosted tree model. Most LLMs outperform the no-change baseline, including for policy response tasks -- with the best model lowering error for numeric outcomes by 12.2%. Across most tasks, gradient-boosted trees rank first; leading proprietary LLMs approach their performance, but open-weight models lag. LLMs exhibit systematic over- and underprediction across different tasks. We identify methods that enable a 4 billion parameter open-weight model to match proprietary models' performance: fine-tuning and aggregating 16 predictions per observation. Improvements generalize to policy-response tasks, which are excluded from fine-tuning. We release our datasets, code, and leaderboard on our website: https://jn-huang.github.io/householdbench

1 INTRODUCTION

The introduction motivates household decision-making as central to economic and policy simulations, then presents HouseholdBench as a broad evaluation of LLM predictions, including responses to policy changes. It reports that LLMs often beat a no-change baseline but generally trail gradient-boosted trees, while fine-tuning and prediction aggregation improve a small open-weight model.

1 INTRODUCTION

  • Prior evaluations cover few surveys and outcomes and do not test household responses to real-world economic policy changes.The introduction identifies these gaps as limiting evidence about realistic policy simulations.Source: §1 · pp. 1–2
  • HouseholdBench combines six U.S. household surveys into 32 prediction tasks spanning consumption, income, labor, expectations, and housing.The tasks include numeric, categorical, and probabilistic outcomes and use past behavior, demographics, and macroeconomic conditions.Source: §1 · p. 2, Figure 1 · p. 2
  • Most LLMs outperform the no-change baseline, including on policy-response tasks for numeric and categorical outcomes, while gradient-boosted trees predict best overall.The tree model is reported as 19% better than the no-change baseline for numeric tasks.Source: §1 · pp. 2–3
  • Fine-tuning Qwen3.5-4B and aggregating predictions reduces numeric error by 24%, and the gains generalize to policy-response tasks excluded from fine-tuning.The fine-tuned, aggregated model also outperforms the strongest LLM.Source: §1 · p. 3
  • LLMs systematically overestimate household outcomes on some tasks and underestimate them on others.The paper also reports that proprietary models approach tree performance while open-weight models lag.Source: §1 · pp. 2–3

2 HOUSEHOLDBENCH

HouseholdBench draws on six U.S. household surveys to construct 32 prediction tasks across household economic behavior and beliefs. It distinguishes baseline forecasts from policy-response predictions and compares its coverage with prior studies.

2.1 DATA

  • The six publicly available U.S. surveys provide rich information on household backgrounds, behavior, preferences, and beliefs; all but Census support longitudinal linking.Longitudinal data let the benchmark track households over time and predict behavioral changes rather than only cross-sectional outcomes.Source: §2.1 · p. 3

2.2 EVALUATION TASKS

  • The benchmark builds 32 tasks across five household-choice topics, with numeric, categorical, and probabilistic survey targets.The topics cover consumption and saving, income and resources, labor and retirement, macroeconomic expectations, and housing and location.Source: §2.2 · p. 3, Table 1 · p. 4
  • Baseline tasks use household background, past outcomes, and macro conditions; policy-response tasks add plausibly external policy, shock, treatment, or scenario variation.The latter test whether models predict behavior or beliefs as the economic environment changes.Source: §2.2 · pp. 3–4, Table 1 · p. 4
  • Each observation pairs a household-period’s demographics, recent history, and macro conditions, plus any policy or shock, with its actual survey response; quarter-based splits test on unseen macro conditions.Samples exclude records with missing targets or recent history and apply survey-specific quality filters and period-level outlier trimming.Source: §2.2 · p. 4

2.3 COMPARISON WITH EXISTING WORK

  • HouseholdBench uniquely spans five topics and as many as six surveys among selected studies; its policy-response tasks also include real-world tax rebates, stimulus payments, unemployment-insurance variation, and job loss.The surveys are publicly available, well documented, and—except for Census—support longitudinal linking.Source: §2.1 · p. 3, §2.3 · p. 5, Table 2 · p. 5

3 EXPERIMENTAL SETTING

The experimental setting compares open-weight and proprietary LLMs with statistical references, then evaluates numeric, categorical, and probability predictions using task-specific metrics and an aggregate ranking.

3.1 MODEL SUITE

  • The suite includes open-weight models ranging from small dense systems to larger mixture-of-experts models, plus GPT and Claude families at different capability tiers.Every model is run under its default inference settings.Source: §3.1 · p. 5
  • XGBoost is trained task-by-task on the same household features given to LLMs, while the no-change baseline carries forward the latest observed value and predicts zero policy response.XGBoost uses pre-tuned settings that outperform its default parameters.Source: §3.1 · pp. 5–6

3.2 METRICS

  • RelMAE divides a model’s absolute error by the no-change baseline’s error on the same samples, with values below 1.0 indicating better performance than that baseline.The normalization supports comparison across tasks with different numerical scales.Source: §3.2 · p. 6
  • Macro-F1 averages category-specific F1 scores with equal weight, which matters when outcome categories differ greatly in prevalence.Source: §3.2 · p. 6
  • Mean total variation measures the distance between predicted and household probability vectors across target bins; it ranges from 0 for perfect agreement to 1 for disjoint mass.Source: §3.2 · p. 6
  • Models receive a unified score from their average rank across all 32 tasks, and final results are assessed with paired i.i.d. bootstrapping using 10,000 samples per task.Lower average ranks indicate better performance.Source: §3.2 · p. 6

4 RESULTS AND DISCUSSION

XGBoost leads the benchmark, while most LLMs beat the no-change baseline on numerical and categorical predictions and leading models also improve on policy-response tasks. LLMs nevertheless show systematic prediction biases, and none improves on the baseline for probability tasks.

4.1 MAIN RESULTS

  • Most LLMs beat no-change on numerical and categorical prediction, but no model improves on it for probability tasks.26.9% of baseline probability targets repeat prior reports, and errors on unchanged reports offset gains on changed reports.Source: §4.1 · p. 6
  • XGBoost ranks first overall (average rank 3.4 versus Fable 5.1’s 3.8); Fable is close, while open-weight models lag.The reported bootstrap intervals for the overall ranks overlap.Source: Table 3 · p. 7, §4.1 · p. 7
  • Fable 5.1 reduces numeric error by 13.6% on policy-response tasks without task-specific training.The figure distinguishes baseline-task results (solid bars) from policy-response results (hatched bars).Source: §4.1 · p. 6, Figure 2 · p. 7

4.2 LLMS’ BIAS ON PREDICTING HOUSEHOLD BEHAVIOR

  • On cons sce shock, the median household reports spending or donating 5% of a permanent 10% income gain, while frontier LLMs predict 30% to 40%.The figure summarizes over- and underestimation across numeric tasks and majority-class overprediction on categorical tasks.Source: Figure 3 · p. 8, §4.2 · p. 8
  • On labor cps jobfind, only half of workers remain unemployed, but LLMs predict unemployment for 94% to 100% of them.The three frontier models also miss every worker who leaves the labor force on both labor tasks.Source: Figure 3 · p. 8, §4.2 · p. 8

4.1 MAIN RESULTS

  • LLM performance shows no large systematic change on numerical tasks after or near models’ knowledge cutoffs.The analysis reports no systematic degradation as cutoffs approach in a subset of tasks with long sample periods.Source: §4.1 · p. 8

5 IMPROVING LLMS’ PREDICTIVE POWER

HouseholdBench's fine-tuning section tests whether supervised adaptation and prediction aggregation can improve open-weight models. Qwen3.5-4B gains from fine-tuning generalize to held-out policy-response tasks, while aggregation further improves fine-tuned models on numeric tasks.

5 IMPROVING LLMS’ PREDICTIVE POWER

  • The section’s approach combines supervised fine-tuning on household behavior data with aggregation over draws to address open-weight models’ performance gap with leading proprietary LLMs.Source: §5 · pp. 8–9

5.2 FINE-TUNING RESULTS

  • With K = 16 draws, aggregation lowers fine-tuned models’ RelMAE by 10% to 13% on numeric baseline tasks; fine-tuned Qwen3.5-4B reaches the best language-model performance on numeric and probability tasks, outperforming Fable 5.1.The table reports SFT results; its K column gives the number of aggregated predictions.Source: Table 4 · p. 9, §5.2 · p. 9

5.1 FINE-TUNING SETUP

  • The setup fine-tunes Qwen3.5-4B and Qwen3.6-27B with LoRA for one epoch on 100,702 training examples sampled from 18 baseline tasks.The policy-response tasks are excluded from fine-tuning.Source: §5.1 · p. 9

5.2 FINE-TUNING RESULTS

  • Aggregation averages numerical and probability predictions and uses majority votes for categorical predictions; on backbone models it lowers numeric RelMAE by only 3% for Qwen3.5-4B and 1% for Qwen3.6-27B.Aggregation does not improve proprietary models’ performance.Source: §5.2 · p. 9
  • On baseline tasks, SFT improves Qwen3.5-4B’s RelMAE by 13%, macro-F1 by 15%, and TV by 34%.For Qwen3.6-27B, SFT improves macro-F1 but worsens RelMAE and TV.Source: §5.2 · p. 9
  • On unseen policy-response tasks, SFT improves Qwen3.5-4B’s RelMAE by 10% and macro-F1 by 30%.Policy-response tasks are held out from fine-tuning to test generalization.Source: §5.2 · p. 9, §5.1 · p. 9

6 CONCLUSION

HouseholdBench evaluates LLM predictions of household outcomes and beliefs across five areas, finding that LLMs still do not outperform XGBoost overall, although Fable 5.1 comes close. The conclusion situates the benchmark among domain-specific household-survey evaluations and broader studies of LLM behavioral simulation.

6 CONCLUSION

  • Small open-weight models with fine-tuning and 16-prediction aggregation outperform leading frontier LLMs on numeric baseline tasks, while LLMs overall still do not outperform XGBoost.Fable 5.1 comes close to XGBoost overall.Source: §6 · p. 10

A RELATED WORK

  • Earlier household-survey evaluations cover particular domains, whereas HouseholdBench evaluates a broader set of economic outcomes and settings and tests predictions against realized outcomes after real economic events.Prior policy-response evaluations cited here are limited to stated responses to information or hypothetical scenarios.Source: A RELATED WORK · p. 17
  • HouseholdBench extends LLM behavioral-simulation research by fine-tuning models with household economic decisions, alongside prior work using experimental choices, economic games, and public-opinion surveys.Source: A RELATED WORK · p. 17

B DATA SOURCES

The data appendix describes the sources and preparation of surveys used to construct HouseholdBench tasks. Most sources link respondents or households across time, while the decennial Census contributes independent cross-sections.

B.1 CONSUMER EXPENDITURE SURVEY: DIARY SURVEY

  • The Consumer Expenditure Survey Diary file combines consumer-unit, individual, diary-week, and expenditure records; the cleaned data contain 2,720,802 consumer-unit-day rows for 205,003 consumer units.Daily expenditures are summed by consumer unit, date, and category, with seven rows per valid diary week.Source: §B.1 · pp. 17–18

B.2 CONSUMER EXPENDITURE SURVEY: INTERVIEW SURVEY

  • The Consumer Expenditure Interview file harmonizes fields across questionnaire changes and contains 1,040,182 interviews linked into 375,509 series of one to four quarterly interviews.Quarterly expenditure is calculated from its three monthly components.Source: §B.2 · p. 18

B.3 CURRENT POPULATION SURVEY: BASIC MONTHLY SURVEY

  • The CPS Basic Monthly file contains 79,330,448 person-month rows from January 1976 through June 2026; longitudinal links require consistent identifying characteristics and survey-rotation positions.Source: §B.3 · p. 18

B.4 CURRENT POPULATION SURVEY: DISPLACED WORKER SUPPLEMENT

  • The Displaced Worker Supplement file contains 1,840,305 person-wave records across 21 supplements from 1984 through 2024, including 95,904 identified displaced workers.The supplement generally runs every two years and records involuntary job loss and subsequent outcomes.Source: §B.4 · pp. 18–19

B.5 UNIVERSITY OF MICHIGAN SURVEYS OF CONSUMERS

  • The Surveys of Consumers file contains 342,345 respondent-month rows from January 1978 through February 2026, linked into 217,641 panels with one to three interviews.Repeat interviews are linked using identifiers for interviews about six and twelve months earlier.Source: §B.5 · p. 19

B.6 SURVEY OF CONSUMER EXPECTATIONS

  • The Survey of Consumer Expectations Core file has 186,660 respondent-month rows for 24,592 identifiers; its eight topical files each contain between 6,809 and 38,105 rows.The panel follows household heads for as long as twelve months, although some identifiers appear in more than twelve months.Source: §B.6 · p. 19

B.7 PANEL STUDY OF INCOME DYNAMICS

  • The PSID file combines harmonized PSID-SHELF data with selected fields from the raw main study, yielding 3,533,082 person-wave rows for 84,121 identifiers across 42 waves.It includes a row for each person and wave, including periods when a person was not interviewed or active in the sample.Source: §B.7 · p. 20

B.8 U.S. DECENNIAL CENSUS

  • The decennial Census file uses six IPUMS USA samples and contains 13,445,846 person rows across six independent cross-sections.The source uses the 1990 and 2000 one-percent samples; each person appears once in a given cross-section.Source: §B.8 · p. 20

B.9 ADDITIONAL DATA SOURCES

The appendix describes the macroeconomic series used alongside survey data and the XGBoost reference models. The models use the same predictors supplied to LLMs, with early stopping based on each task’s evaluation metric.

B.9 ADDITIONAL DATA SOURCES

  • Quarterly national predictors include real GDP, unemployment, CPI, the federal funds rate, house prices, the S&P 500, and mortgage rates, with recent quarters and five-year summaries derived from FRED-QD.The series include revisions in the July 2026 file rather than historical vintages available on each survey date.Source: §B.9 · p. 20
  • The appendix also uses quarterly CPI-U group indexes and, for Census migration tasks, state unemployment, per-capita income growth, and house-price growth.Source: §B.9 · pp. 20–21

C XGBOOST MODELS

  • XGBoost receives the same predictors as the LLM household profiles; task-specific models use literature-informed parameters and stop after 300 rounds without validation improvement.The model selects predictions from the best validation iteration without refitting on combined training and validation samples.Source: C XGBOOST MODELS · p. 21
  • XGBoost trains separate regressions, classifiers, or probability-vector regressions by target type, with early stopping based on MAE, one minus Macro-F1, or total-variation distance, respectively.Probability vectors with more than two outcomes are projected onto the unit simplex.Source: C XGBOOST MODELS · p. 21
  • Across 32 tasks, the training split contains 3,544,247 rows and the validation split 447,819; each task uses at most 400,000 training and 50,000 validation rows.Source: C XGBOOST MODELS · p. 21

D ADDITIONAL RESULTS AND DISCUSSION

Figure D1 organizes model performance relative to no-change by probability target and report-change status. Its continued panels distinguish baseline from policy tasks.

D.1 PERSISTENCE IN HOUSEHOLD PROBABILITY REPORTS

  • Figure D1 compares each model’s mean total-variation difference from no-change across unchanged, changed, and all reports, with rows kept separate by target.Negative differences favor the model; zero indicates equal performance. The three panels for each target share a scale, while scales differ across targets.Source: Figure D1 · p. 22
  • The target rows include household beliefs about inflation, unemployment, savings interest rates, and stock prices, as well as work-finding and mortgage-rate probabilities.The figure also includes a work-finding target for finding acceptable work within 12 months.Source: §D.1 · pp. 23–25
  • Solid bars denote baseline tasks and hatched bars denote the policy task; each model and no-change comparator use the same valid observations.Black bars are paired bootstrap intervals, and the no-change row is omitted.Source: Figure D1 · p. 22, §D.1 · p. 25

D.2 BENCHMARK RESULTS

The appendix figures describe performance comparisons across cutoff samples and over outcome-release time. They specify sample matching and uncertainty intervals but provide no stated performance outcomes.

D.2 BENCHMARK RESULTS

  • Figure D2(a) compares matched pre- and post-cutoff task samples for main-leaderboard models evaluated in both samples, excluding no-change.Bars show equal-task means of raw task metrics; whiskers are pointwise 95% bootstrap intervals.Source: Figure D2 · p. 26
  • Figure D2(b) compares pre-cutoff performance with results on 454 rows from nine tasks released after Claude Fable 5.1’s June 2026 knowledge cutoff.The pre-cutoff sample covers those same nine tasks; RelMAE and total-variation axes are reversed because lower values are better.Source: Figure D2 · p. 26

D.3 PERFORMANCE OVER TIME

  • Figure D3 plots main-leaderboard model performance over outcome-release time, with vertical bars showing pointwise 95% bootstrap intervals.Source: Figure D3 · p. 27

E FINE-TUNING WITH A NUMBER-TOKEN LOSS

The section tests whether number-token loss improves supervised fine-tuning for numeric and probability predictions. It reports further gains over SFT alone, with smaller additional gains after aggregation.

E FINE-TUNING WITH A NUMBER-TOKEN LOSS

F.1 AGGREGATING MODEL PREDICTIONS

The section compares single-draw predictions with aggregates of two or sixteen draws and reports distinct aggregation effects for fine-tuned open-weight and proprietary models. It also records the inference settings used for model runs.

F.1 AGGREGATING MODEL PREDICTIONS

  • Fine-tuned Qwen3.5-4B and Qwen3.6-27B improve on numeric and probability tasks with two aggregated predictions, with larger gains at sixteen.Source: §F.1 · p. 29
  • Aggregating two predictions changes proprietary models’ numeric and probability performance by less than 0.5%, while categorical changes are mostly negative.Source: §F.1 · p. 29

G.1 INFERENCE SETTINGS

  • Qwen models use non-thinking mode with temperature 0.7, top-p 0.8, and top-k 20; GPT and Claude use default reasoning effort, and DeepSeek uses medium.Source: §G.1 · p. 30

G.2 FINE-TUNING SETTINGS

The section describes fine-tuning Qwen3.5-4B and Qwen3.6-27B with shared hyperparameters. Training used four NVIDIA A100 80GB GPUs.

G.2 FINE-TUNING SETTINGS

  • Both Qwen models were fine-tuned using four NVIDIA A100 80GB GPUs.The models share the hyperparameters listed in Table G1.Source: §G.2 · p. 30
  • Table G1 gives hyperparameters shared by Qwen3.5-4B and Qwen3.6-27B.The caption says these settings are shared across both training losses.Source: Table G1 · p. 30, §G.2 · p. 30
  • The listed hyperparameters apply to both training losses.Source: Table G1 · p. 30

H EVALUATION TASKS

This appendix describes HouseholdBench tasks and their targets, predictors, baselines, and sample construction. The examples span consumption, wealth, and spending growth.

H EVALUATION TASKS

  • The appendix details 32 tasks, including their targets, predictors, sample construction, splits, and prompt examples.Source: H EVALUATION TASKS · p. 30

H.1 CONS CEX CATEGORIES

  • CEX categories tasks predict twelve categories of household expenditure over the prior three months in current U.S. dollars.Predictors include spending history, income, household characteristics, demographics, geography, and macroeconomic and category-specific price conditions; the no-change baseline carries forward the prior interview's category value. The sample requires nonmissing targets and current or first-history predictors, income of at least $1,000, total expenditure of at least $250, nonnegative categories, and trims extreme income and expenditure values within year-quarter.Source: §H.1 · pp. 30–31

H.2 CONS CEX TOTAL

  • CEX total tasks predict total, nondurable, and durable household expenditure over the prior three months.Total expenditure excludes specified contributions, insurance, and retirement payments; durable expenditure is total less nondurable. Predictors include expenditure history, income, household characteristics, demographics, geography, and macroeconomic conditions; the baseline uses the previous interview, and the sample applies income and expenditure thresholds, nonnegativity rules, and trimming.Source: §H.2 · pp. 31–32

H.3 CONS PSID WEALTH

  • The PSID wealth task predicts household net worth at the next interview, usually two years later and sometimes one year later.Net worth combines reported assets and subtracts debts; predictors include household resources and history, tenure, household composition, demographics, and economic, financial-market, and housing-market conditions. The baseline carries forward current net worth, and the sample requires the next-interview target and specified current predictors while trimming income and net worth extremes.Source: §H.3 · pp. 32–33

H.4 CONS SCE GROWTH

  • The SCE growth task predicts the percentage change in current monthly spending relative to 12 months earlier, with increases positive and decreases negative.Predictors include household finances and expectations, work status, respondent characteristics, and macroeconomic context; the baseline predicts no change. The sample combines the SCE spending module and core survey, requires complete targets and current predictors, limits respondents to ages 18–100, and trims spending changes within survey month.Source: §H.4 · pp. 33–34

H.5 CONS CEX REBATE01

This CEX task evaluates expenditure predictions in the setting of the 2001 federal income-tax rebates. It uses rebate histories alongside household and economic information.

H.5 CONS CEX REBATE01

  • The task is inspired by Johnson et al. (2006).Source: §H.5 · p. 34
  • The task predicts total, durable, and nondurable expenditure, using the amount of the 2001 federal income-tax rebate received as a predictor.Rebate predictors include current and prior three-month amounts, including zero amounts.Source: §H.5 · pp. 34–35
  • The sample covers CEX interviews from January 2001 through March 2002 while rebate questions were fielded.The no-change baseline carries forward expenditure from the immediately preceding three-month period; sample construction requires specified current predictors and first-history expenditure and rebate values.Source: §H.5 · p. 35

H.6 CONS CEX SSPAY

This section specifies nine household-behavior prediction tasks spanning spending and income outcomes, including expectations and responses to income shocks or job displacement.

H.6 CONS CEX SSPAY

  • The CEX Social Security-payment task predicts daily total, food-at-home, and food-away-from-home spending from recent spending, household characteristics, payment timing, and macroeconomic conditions.Its no-change prediction is the previous day’s expenditure; the sample focuses on households reporting Social Security retirement or disability income around scheduled payment dates.Source: §H.6 · pp. 35–36

H.7 CONS CEX STIMULUS08

  • The CEX stimulus task predicts total, durable, and nondurable expenditure over three months using spending history, household characteristics, macroeconomic conditions, and stimulus-payment amounts.The no-change prediction is expenditure in the preceding three-month period; the sample covers interviews from September 2007 through March 2009 while stimulus questions were fielded.Source: §H.7 · p. 37

H.8 CONS PSID JOBLOSS

  • The PSID job-loss task predicts annual cash food expenditure in the following year from spending and work histories, job-loss circumstances, household characteristics, and macroeconomic conditions.Its no-change prediction is the household’s most recently observed annual cash food expenditure before the job-loss year; the sample includes qualifying job-loss households and controls without qualifying losses.Source: §H.8 · p. 38

H.9 CONS SCE SHOCK

  • The SCE shock task predicts the share of a permanent 10% income gain spent or donated and the share of a permanent 10% income loss covered by reduced spending.Respondents consider both changes, and the no-change prediction assigns none of either change to spending.Source: §H.9 · p. 39

H.10 INCOME MICH FINANCE

  • The Michigan Survey task predicts the signed percentage change in household income expected over the next 12 months, combining the reported direction and size of the expected change.Its no-change prediction is zero income change; predictors include household and national assessments, income, demographics, geography, and macroeconomic conditions.Source: §H.10 · p. 40
  • The Michigan finance task asks respondents to predict the percentage change in their income over the next 12 months.Source: §H.10 · p. 40

H.11 INCOME PSID EARNINGS

  • The PSID task predicts respondents’ annual labor earnings two, four, and ten years after the current interview.The no-change prediction is current earnings; the sample links the same respondents across interviews and requires nonmissing targets and required predictors.Source: §H.11 · pp. 40–41

H.12 INCOME SCE GROWTH

  • The SCE task predicts the percentage change in total household income over 12 months, combining expected direction and size.Its no-change baseline is zero change; current household-finance, health, work, and job-search predictors may be missing.Source: §H.12 · pp. 41–42

H.13 INCOME CPS DISPLACE

  • The CPS displacement task predicts current weekly earnings for workers who lost a job and found another, conditional on reemployment.Predictors include demographics, displacement and lost-job history, reemployment history, and national economic conditions; the no-change baseline is earnings on the lost job.Source: §H.13 · p. 42

H.14 INCOME SCE POLICY

This task predicts respondents’ stated household impact from four hypothetical policy changes over the next 12 months. It uses household finances, expectations, work status, demographics, geography, and recent macroeconomic conditions, together with each policy’s direction.

H.14 INCOME SCE POLICY

  • The task predicts one of five stated impact categories for each of four policy changes: welfare benefits, unemployment benefits, payroll tax, and average income tax.Source: §H.14 · p. 43
  • Predictors include household finances and expectations, work status, demographics, resources, geography, macroeconomic conditions, and whether each policy increases or decreases.Source: §H.14 · p. 43
  • The no-change baseline predicts that none of the four policy changes affects the household.Source: §H.14 · p. 43
  • The sample merges the SCE Public Policy module with Core survey and FRED-QD data, excludes specified missing targets or predictors, and requires valid categories and ages 18–100.Current-finance, inflation-expectation, and work-status predictors may remain missing.Source: §H.14 · p. 43
  • The prompt frames the respondent as an American adult, supplies household and economic context, and specifies the four policy changes and their directions.Source: §H.14 · p. 44
  • The response must be a four-string JSON array in policy order; the example answers “somewhat negative” for all four changes.Source: §H.14 · p. 44

H.15 LABOR CPS JOBFIND

This section defines household prediction tasks spanning labor-market outcomes and expectations, inflation expectations, and housing behavior, drawing on prior behavior and household and economic context to specify outcomes and predictors, with baselines and sample construction described for some tasks.

H.15 LABOR CPS JOBFIND

  • The CPS job-find task predicts whether currently unemployed people are employed, unemployed, or out of the labor force one month later, against a baseline of remaining unemployed.The sample includes people aged 16 or older matched to the next monthly interview.Source: §H.15 · p. 44

H.16 LABOR CPS RETIRE

  • The CPS retirement task predicts whether employed respondents aged 50–75 are retired 12 months later, defining retirement as being out of the labor force and reporting retirement as the reason.The no-change baseline predicts that the person is not retired.Source: §H.16 · p. 45

H.17 LABOR CPS SEPARATION

  • The CPS separation task predicts whether currently employed people are employed, unemployed, or out of the labor force one month later, against a baseline of remaining employed.The sample includes people aged 16 or older matched to the next monthly interview.Source: §H.17 · p. 46

H.18 LABOR SCE OFFER

  • The SCE offer task predicts acceptance or rejection of the first listed offer among respondents with at least one recent job offer; undecided respondents are omitted.Acceptance counts even if the respondent no longer works in the job, and the first listed offer need not be the first received; the baseline predicts rejection.Source: §H.18 · p. 47

H.19 LABOR SCE RISK

  • The SCE risk task predicts stated probabilities of job loss and voluntary departure within 12 months, and finding a job within three months conditional on job loss.Each no-change prediction uses the respondent’s stated probability from one month earlier.Source: §H.19 · p. 48

H.20 LABOR SCE SEARCH

  • The SCE search task predicts job seekers’ stated chances of finding an acceptable job within three and 12 months, using their probabilities from one month earlier as the no-change baselines.The sample is restricted to respondents currently looking for work.Source: §H.20 · p. 49
  • The SCE job-search task elicits probabilities of finding an acceptable job within the next 12 months and within the next 3 months.The requested output is a separate two-outcome probability distribution for each horizon.Source: §H.20 · p. 49

H.21 LABOR CPS DISPLACE

  • The CPS displacement task predicts whether a worker reporting a qualifying past job loss is employed, unemployed, or out of the labor force at the supplement interview.The classification uses reported work, job search, availability, and reasons for not working; predictors include displacement details, demographics, household and geographic characteristics, and recent national economic conditions.Source: §H.21 · pp. 49–50

H.22 LABOR CPS UI

  • The CPS UI task predicts whether currently unemployed people become employed, remain unemployed, or leave the labor force one month later, using state-month unemployment-benefit durations among its predictors.It matches eligible unemployed respondents from January 2008 through August 2014 to the following monthly interview; policy measures are timed to the fifth day of each month, matching the CPS interview week.Source: §H.22 · pp. 50–51

H.23 LABOR PSID ADDEDWORKER

  • The PSID added-worker task predicts a spouse or partner’s annual work hours and labor earnings in the job-loss year and following year.The current-year outcomes may include work before as well as after the other adult’s job loss; the sample links couples observed before, during, and after a single adult’s qualifying job loss.Source: §H.23 · pp. 51–52

H.24 LABOR SCE RESWAGE

  • The SCE reservation-wage task predicts the lowest acceptable hourly wage for a new job and preferred weekly hours, converting reported weekly or annual wage amounts to hourly equivalents.Predictors include work and job-search situation, current or most recent pay and hours, household finances and expectations, demographics, and macroeconomic context.Source: §H.24 · pp. 52–53

H.25 MACRO MICH OUTLOOK

  • The Michigan Survey task predicts signed inflation expectations for the next 12 months and average annual inflation over the next 5 to 10 years, using each respondent’s earlier expectations as predictors.The earlier interview occurred six to eight months before the current response; each expectation combines the reported direction and magnitude of price change.Source: §H.25 · pp. 53–54
  • The Michigan outlook task elicits expected general-price changes over the next 12 months and average annual change over 5 to 10 years.Source: §H.25 · p. 54

H.26 MACRO SCE UNCERTAINTY

  • The SCE uncertainty task predicts five probability vectors: inflation distributions at two horizons and higher unemployment, savings rates, and stock prices; its baseline carries forward last month’s vectors.Predictors include prior expectations, household finances and work status, demographics, and macroeconomic conditions. The sample excludes records missing required targets or predictors and applies probability-validity and age criteria.Source: §H.26 · pp. 54–55

H.27 MACRO SCE REVISION

  • The SCE revision task predicts mean inflation expectations at the twelfth interview, using first-interview expectations and intervening inflation to model revisions.The no-change prediction carries forward first-interview expectations; interviews must be exactly 11 months apart. Predictors compare observed inflation over those months with the earlier expectation converted to an 11-month rate.Source: §H.27 · pp. 55–56

H.28 HOUSE CENSUS MOVE

  • The Census move task classifies whether respondents remain in the same home, move within their state, or move to another state over five years.The no-change baseline predicts staying in the same home; predictors include earlier respondent and state characteristics, state economic conditions, and national macroeconomic context.Source: §H.28 · pp. 56–57

H.29 HOUSE PSID OWNER

  • The PSID ownership task predicts whether current renters own a home at their next interview, usually two years later, against a baseline of remaining renters.Predictors include household resources and histories, composition, demographics, macroeconomic context, and housing-market conditions.Source: §H.29 · p. 57

H.30 HOUSE SCE MOVE

  • The SCE move task predicts respondents’ stated probability of moving to a different primary residence within 12 months, using their prior-month probability as the no-change baseline.Predictors include moving circumstances, household finances and expectations, work and job-search information, demographics, and macroeconomic context.Source: §H.30 · p. 58

H.31 HOUSE SCE FINANCING

  • The SCE financing task is inspired by Fuster & Zafar (2021).Source: §H.31 · p. 59
  • The financing task predicts six dollar answers: maximum home price and down payment under flexible down-payment, mortgage-rate-change, and cash-inheritance scenarios.The experiment starts from a 4.5% or 6.5% mortgage rate; the sample comes from the February 2014 home-financing experiment in Fuster & Zafar (2021)’s replication package.Source: §H.31 · p. 59
  • Predictors combine initial housing choices, household finances and characteristics, preferences, and macroeconomic conditions; the no-change baseline repeats initial price and down payment across scenarios.The model predicts all six scenario answers together.Source: §H.31 · p. 59

H.32 HOUSE SCE LOCKIN

  • The lock-in task predicts a respondent’s 0–100% chance of moving within three years if they can carry their current mortgage rate to a new home.The target is also evaluated as probabilities of moving and not moving; the no-change baseline uses the respondent’s ordinary moving probability.Source: §H.32 · pp. 60–61
  • Lock-in predictors include ordinary moving expectations, mortgage and housing circumstances, household finances, work status, demographics, geography, and macroeconomic conditions.The sample merges the SCE Housing and Core surveys with national economic and mortgage-rate data, retaining respondents with an active mortgage.Source: §H.32 · pp. 60–61
Loading 2610.07563v1…