dev/complex_covariance_probit_endogenous_seed001/release/release_notes.md

Causal Demand Benchmark First Pipeline Run

This release folder contains the public-facing files for one benchmark cell: complex market, covariance_probit DGP, endogeneity on, seed 001.

Files

• public/products_public.csv: public SKU metadata, family-level text, structured attributes, and image specs.
• public/transactions_train_public.csv: positive-sales (units>0) SKU-store-week transaction rows with anonymized store IDs and public supply-cost proxy. Stores carry a permanent subset of the 40-SKU catalog; carriage is inferable from row presence (>=1 positive train-window row) but is NOT released as an explicit matrix.
• public/counterfactual_context_public.csv: price-policy intervention context.

Hidden DGP parameters, store covariates, calibration quantities, and counterfactual truth are intentionally outside this public release folder.

covariance_probit cells use a two-stage structural model: (1) a purchase-incidence model sets the total category demand per (store, week) via market_size_st = M_0 * Total_Households[s] * seasonality_t * exp(rho*CV_st) * exp(lambda*G_mt + eps), where CV_st = log sum_j exp(Vbar_jst) is the category value (inclusive value of the same brand-choice utilities; Bucklin & Gupta 1992), calibrated against total IRI units; (2) conditional choice is J-only over the J inside products (no outside option). The exogenous inputs public/households_public.csv and public/seasonality_index_public.csv are released; the realized market-size proxy is retired (option A) so the hidden price response cannot be divided out. Complex cells include rendered package-style PNG images under public/images/.

Selection

• SKU count: 40
• SKU cumulative IRI revenue share: 0.8017
• Synthetic stores: 731
• Synthetic weeks: 156

CARD packaging (v0.1)

• Two legacy counterfactual context files (DGP-construction fixtures the scorer ignores) are not included in this release.
• stores_public.market and products_public.brand_code are pseudonymized (consistent across cells and future seeds).
• stores_public.household_count is withheld (store demographics stay with the hidden store covariates); store scale is inferable only from the sales rows.
• release/scoring_params.json carries the scoring config (family, eval-window length, version).