# Datasheet — CARD (Causal Recovery of Demand)

*Template: Gebru et al., "Datasheets for Datasets" (CACM 2021).*

## Motivation

**For what purpose was the dataset created?** To test whether demand models that fit observed sales well can
also recover causal price response, substitution structure, and counterfactual outcomes when prices and
promotions are endogenous. The benchmark pairs a synthetic retail transaction panel with product description
texts that carry the true substitution geometry, so it also measures whether a model extracts economic
structure from text.

**Who created it?** Juwon Hong, Minha Hwang, and Venkatesh Shankar; maintained by Juwon Hong.

## Composition

**What do the instances represent?** Store-product-week transaction records from a synthetic facial-tissue
category, plus one marketing-copy description per product. The data are generated by a known
data-generating process; no real transactions are included.

**Structure.** A factorial grid of 4 cells (2 demand-model families × endogeneity on/off). Every cell is the
full market: 40 products, 731 stores, 156 weeks (140 public training weeks + 16 holdout-context weeks with
quantities withheld).

**Files per cell (public).** `transactions_train_public.csv` and `transactions_holdout_context_public.csv`
(`product_id`, `store_id`, `week`, `units`, `dollars`, `price`, `promo_flag`, `promo_cost`,
`supply_cost_proxy`; quantities withheld in the holdout context), `products_public.csv` (`product_id`,
`product_text`, `brand_code`), `stores_public.csv` (`store_id`, `market`, `chain`), and
`counterfactual_sweep_context_public.csv` (`intervention_id`, join keys, `baseline_price`,
`intervention_price`). Market and brand identifiers are pseudonymized; store demographics are withheld. The
authoritative enumeration with per-file SHA-256 hashes is each cell's `release/MANIFEST.json`.

**Is anything withheld?** Yes, deliberately: the substitution answer key, the hidden ground-truth
elasticities and counterfactual outcomes, the hidden full-length panel, all structural shocks, and the store
demographics. The dev seed ships its scoring truth (`hidden/`) for local iteration; eval-seed truth stays
with the maintainer. The generator is withheld during the evaluation phase under a published SHA-256
commitment (`GENERATOR_COMMITMENT.md`).

**Does the dataset contain personal data, confidential data, or offensive content?** No. All records are
synthetic; the texts are machine-written marketing copy for facial tissue. Store identifiers are synthetic;
market and brand codes are pseudonymized. No human subjects, no PII.

## Collection process (generation process)

There is no collection process; the dataset is simulated. The numeric generator is calibrated to moments of
the licensed IRI academic retail panel and TDLinx/Spectra store demographics, plus public CDC FluView
influenza curves; the licensed sources enter only through aggregate calibration targets and per-product
anchors whose publication their licenses permit. The product texts were written by a large language model
from controlled per-product briefs and frozen with a published checksum. The substitution answer key is
built from the texts by a committee of four pinned sentence encoders. The construction is documented at the
equation level in the paper appendix.

## Preprocessing / cleaning / labeling

Released prices are rounded to realistic price endings; observed counts reflect permanent store assortments
plus a missing-not-at-random zero mechanism matching retail sparsity (public files carry positive-sale rows
only). Causal answer keys are computed on the carried support before those observation layers, so the ground
truth is unaffected. No other cleaning is applied.

## Uses

**Intended.** Benchmarking demand models: sales forecasting on the holdout window, elasticity recovery,
counterfactual demand response under price interventions, and the value of reading the product text.
Classical econometric, ML, and multimodal/foundation-model approaches are all in scope.

**Out of scope.** The data support no conclusions about the real facial-tissue market, any real retailer or
brand, or consumer behavior; the texts are not a general product-text corpus; the panel is not suitable for
training production pricing systems.

## Distribution

Data: Hugging Face [`jean-jsj/CARD`](https://huggingface.co/datasets/jean-jsj/CARD), CC BY 4.0, archived
under DOI [10.57967/hf/9681](https://doi.org/10.57967/hf/9681). Evaluation and scoring code:
[github.com/jean-jsj/CARD](https://github.com/jean-jsj/CARD), Apache-2.0.

## Maintenance

**Maintainer and contact:** Juwon Hong (GitHub: `jean-jsj`; issues on the code repository).
**Versioning:** releases are immutable and identified by the `benchmark_version` stamp in each cell's
`release/scoring_params.json`; corrections publish as a new dataset version, never as silent overwrites.
Scoring math is frozen between minor versions of the `card-metrics` package, so scores stay comparable.
Hidden evaluation-seed truth is disclosed when a seed is retired and is verifiable against the published
commitments.