Description

What this does: Produces robust weekly orders by pooling information across your product hierarchy while preserving SKU‑specific signals. Especially effective for sparse/intermittent SKUs.

Method: Negative Binomial GLM with Fourier seasonality → Empirical Bayes SKU adjustments → CV‑tuned shrinkage → per‑segment overdispersion (alpha) → posterior aggregation over protection period → base‑stock ordering.

Robustness: Handles collinearity and convergence via fallbacks (simplified formulas, L2‑regularized NB/Poisson). Auto‑resolves common file paths and validates required inputs.

Inputs required:

• data/Week 0 - Master.csv (must contain Division, Department, ProductGroup, plus Store, Product)
• Long sales (Store, Product, Week, SalesQty) — e.g., exported from eda_order1.ipynb
• Week 0 state (End Inventory, In Transit W+1, In Transit W+2)
• Submission index template

Outputs:

• submissions/orders_hierarchical_eb.csv (competition‑ready)
• Diagnostics: artifacts/hierarchical/figs/ (shrinkage weights, naive vs EB mean/std)
• Demand stats: artifacts/hierarchical/demand_stats_hierarchical.csv

Notes:

• Runtime ~15–45s for 599 SKUs (includes CV).
• If shrinkage appears weak (weights ~0), increase SHRINK_SCALE or review hierarchy quality.
• Everything runs top‑to‑bottom inside notebooks/hierarchical_bayes.ipynb.