LetTokenBurnHouseholdBench: Evaluating Large Language Models as Predictors of Household Economic Behavior
A benchmark with 32 prediction tasks across 6 U.S. household surveys (21.7M observations) evaluating LLMs on numeric, categorical, and probabilistic economic outcomes.
$ npm i -g lettokenburn-cli$ ltb bench info householdbench
External harness. No Harbor adapter has been generated for this benchmark yet. `ltb bench info` prints the known links.
Runs happen on your machine with your own subscription; only the score summary is pushed. CLI docs
| Harness | Source | Confidence | Link | |||
|---|---|---|---|---|---|---|
| XGBoost | 0.8 | N/A | Independent | medium | Oct 6, 2026 | arxiv.org |
| Fable 5.1 | 0.9 | N/A | Independent | medium | Oct 6, 2026 | arxiv.org |
| Claude Opus 5 | 0.9 | N/A | Independent | medium | Oct 6, 2026 | arxiv.org |
| GPT-5.6 Sol | 0.9 | N/A | Independent | medium | Oct 6, 2026 | arxiv.org |
| GPT-6 Astra | 0.9 | N/A | Independent | medium | Oct 6, 2026 | arxiv.org |
| Claude Opus 4.8 | 0.9 | N/A | Independent | medium | Oct 6, 2026 | arxiv.org |
| Claude Sonnet 5 | 0.9 | N/A | Independent | medium | Oct 6, 2026 | arxiv.org |
| GPT-5.6-terra | 0.9 | N/A | Independent | medium | Oct 6, 2026 | arxiv.org |
| DeepSeek V4 Flash | 1.0 | N/A | Independent | medium | Oct 6, 2026 | arxiv.org |
| Qwen3.6 35B A3B | 1.0 | N/A | Independent | medium | Oct 6, 2026 | arxiv.org |
| GPT-5.6 Luna | 1.0 | N/A | Independent | medium | Oct 6, 2026 | arxiv.org |
| DeepSeek V4 Pro | 1.0 | N/A | Independent | medium | Oct 6, 2026 | arxiv.org |
| Qwen3.6 27B | 1.0 | N/A | Independent | medium | Oct 6, 2026 | arxiv.org |
| Qwen3.5 4B | 1.3 | N/A | Independent | medium | Oct 6, 2026 | arxiv.org |
Every number links to where it was published. Low-confidence rows are muted and need per-cell verification.
HouseholdBench comprises 32 tasks derived from six U.S. household surveys, covering consumption, income, labor, expectations, and housing. It includes baseline and policy-response scenarios using longitudinal data.
Models are evaluated against no-change and gradient-boosted tree baselines. Proprietary LLMs approach tree performance, while fine-tuned open-weight models show significant improvement, reducing numeric error by up to 24%.