Pharmacy Sales Dataset Demand Forecasting: Datasets, Joins and Pipeline

A pharmacy sales dataset is transaction-level retail data: one row per line item dispensed or sold, with a date, a product identifier and usually a quantity and price. Demand forecasting needs exactly that shape, because weekly and seasonal patterns only emerge once you aggregate lines into a store-by-product-by-day grid. Kaggle Datasets - Pharmacy & Drug Retail Search is where Datadory catalogs this material: community-uploaded pharmacy sales transactions in CSV, Parquet and SQLite, tagged demand-forecasting and market-basket.

The catch is that these uploads are static. Kaggle hosts over 1 million public datasets overall, but individual pharmacy uploads range from kilobytes to a few hundred MB of CSV, most covering months to a few years of transactions before going quiet. So a forecast model built on them is a benchmark artifact, not an operating system: you train and compare architectures on the historical lines, then wire live signals around them from sources that do refresh.

That split defines every section below. Of the 16 primary Drug Retail datasets Datadory catalogs, all 16 are free, 11 ship through official APIs, and exactly one - the Kaggle search - carries line-item sales. Everything else earns its place by enriching, validating or refreshing what the sales file cannot tell you.

Treat the gap as a design constraint, not a blocker. Public pharmacy-sales work everywhere on Kaggle-style extracts follows the same recipe: forecast on the static line items, join enrichment from public regulatory and price feeds, and re-score whenever a source refreshes.

The working sequence, using only free inputs:

1. Build the product master with the NDC Directory. The openFDA NDC Directory API lists 137,206 marketed finished drug products with nested package and active-ingredient rows, refreshed daily under commercial delivery terms as a single ~26.8 MB zipped partition. Join your file's product names or barcodes to NDC codes so every SKU inherits a labeler and marketing dates.

1. Normalize messy drug strings with NLM. The Clinical Table Search Service returns RxTerms names with strengths, forms and routes (~9,300 names) plus 2,342 ingredients through a key-free JSON API advised at 25 requests per second - enough to reconcile 'amoxicillin 500mg cap' variants across uploads without registering for anything.

1. Flag discontinued SKUs via Drugs@FDA. The openFDA Drugs@FDA API tracks 29,273 applications back to 1939 with marketing status and therapeutic equivalence codes, refreshed daily on weekdays, so products your model still expects to sell can be marked discontinued or substituted.

1. Suppress forecasts during recalls. The openFDA Drug Enforcement / Recall API holds 17,876 reports since 2004 with classification, recalling firm and distribution pattern, refreshed weekly as one ~3.8 MB JSON file. A Class I recall explains a sales collapse better than any seasonality term.

1. Benchmark honestly. Score candidate models on held-out weeks of the same static extract; the Kaggle upload will not grow under you mid-experiment, which makes results reproducible.