LemeBase's picture
Add all-sets.zip (all published files in one zip)
f7168ac verified
|
Raw History Blame Contribute Delete
8.05 kB
metadata
license: cc-by-4.0
pretty_name: Hackathon retail sets
task_categories:
  - time-series-forecasting
tags:
  - retail
  - demand-forecasting
  - inventory
configs:
  - config_name: set-a
    data_files:
      - split: articles
        path: set-a/articles.parquet
      - split: stock_snapshot
        path: set-a/stock_snapshot.parquet
      - split: weekly
        path: set-a/weekly.parquet
  - config_name: set-b
    data_files:
      - split: articles
        path: set-b/articles.parquet
      - split: basket_promotions
        path: set-b/basket_promotions.parquet
      - split: baskets
        path: set-b/baskets.parquet
      - split: promotion_articles
        path: set-b/promotion_articles.parquet
      - split: promotion_stores
        path: set-b/promotion_stores.parquet
      - split: promotions
        path: set-b/promotions.parquet
      - split: sales_lines
        path: set-b/sales_lines.parquet
      - split: stock
        path: set-b/stock.parquet
      - split: stock_events
        path: set-b/stock_events.parquet
  - config_name: set-c
    data_files:
      - split: daily
        path: set-c/daily.parquet
  - config_name: set-d
    data_files:
      - split: sales
        path: set-d/sales.parquet
      - split: stock
        path: set-d/stock.parquet
  - config_name: test-shop
    data_files:
      - split: history
        path: test-shop/history.parquet

Hackathon retail sets

Four retail datasets for a forecasting hackathon: Set A, Set B, Set C and Set D. Each set is one shop chain; they share no stores or articles. Values are as published by the sources; LemeBase renamed columns to shared plain names and dropped name columns (changes per set below). Files in a set have different columns, so load one file at a time, e.g. load_dataset("LemeBase/hackathon-retail-sets", "set-a", split="weekly").

Everything in one file: all-sets.zip (23 MB): set-a to set-d and test-shop exactly as published here, with ATTRIBUTION.md and LICENSE.txt. Or: huggingface-cli download LemeBase/hackathon-retail-sets --repo-type dataset.

Set A (set-a)

Weekly units sold and an in-stock flag per store × article, the article and store hierarchy (codes only), and stock on hand for one week.

File Rows Columns
articles.parquet 599 store_id, article_id, product_group, division, department, department_group, store_format, format
stock_snapshot.parquet 599 store_id, article_id, week_start, on_hand_start, units, missed_units, on_hand_end, in_transit_1w, in_transit_2w
weekly.parquet 98,835 store_id, article_id, week_start, units, in_stock

Licence: CC BY 4.0. Changed by LemeBase: Reshaped from one column per week to one row per week; columns renamed; cost columns of the stock snapshot dropped.

Set B (set-b)

Synthetic grocery data: baskets, sales lines, articles, stock levels and stock events, promotions.

File Rows Columns
articles.parquet 1,500 article_id, category, subcategory, unit_size, price, unit_cost
basket_promotions.parquet 68,573 transaction_id, line_no, promotion_id, mechanic, savings
baskets.parquet 167,823 transaction_id, customer_id, store_id, timestamp, subtotal, tax, total, items, payment_type, weekday
promotion_articles.parquet 3,099 promotion_id, article_id
promotion_stores.parquet 149 promotion_id, store_id
promotions.parquet 264 bundle_qty, buy_qty, can_stack, category, end_date, get_qty, get_type, mechanic, percent_off, priority, promotion_id, scope_type, start_date
sales_lines.parquet 504,000 transaction_id, line_no, article_id, units, unit_price, line_subtotal, line_discount, line_total
stock.parquet 7,970 article_id, store_id, on_hand, reorder_point, price_override_cents, active
stock_events.parquet 26,215 timestamp, store_id, article_id, delta, reason

Licence: CC BY 4.0. Changed by LemeBase: Columns renamed; product names, brands, store names and addresses, the customer table and derived tables dropped; text categories, unit sizes and weekdays replaced by codes.

Set C (set-c)

One row per store × article × day: units sold, sales, discount, stock on hand, receipt and sale recency, phasing-out flag.

File Rows Columns
daily.parquet 491,857 date, store_id, article_id, division, category, subcategory, segment, is_return, units, sales_amount, cogs, discount_amount, discount_pct, on_hand, is_audit_anomaly, unit_cost, unit_price, latest_on_hand, stock_status, sales_type, last_receipt_date, last_sold_date, days_since_receipt, days_since_sale, is_phasing_out

Licence: CC BY 4.0. Changed by LemeBase: Columns renamed; text categories replaced by codes.

Set D (set-d)

Daily sales lines per store × article with sales type and returns, and stock on hand per store × article over date ranges.

File Rows Columns
sales.parquet 125,751 date, sales_type, is_return, return_reason, supplier_id, article_id, division, category, subcategory, segment, store_id, sales_channel, units, sales_amount, cogs, transactions
stock.parquet 284,755 start_date, end_date, stock_status, supplier_id, article_id, division, category, subcategory, segment, store_id, store_type, sales_channel, on_hand, stock_value_at_price, stock_value_at_cost, unit_price, unit_cost

Licence: CC BY 4.0. Changed by LemeBase: Columns renamed; product descriptions dropped; product hierarchy, supplier, sales channel, store type, sales type and stock status text replaced by codes.

Test shop (test-shop)

One real shop, not one of the four sets: the shop the hackathon scores your forecast on. This folder holds its weekly history only; the 13 weeks that follow are hidden.

File Rows Columns
history.parquet 415,272 article_id, week_start, units, on_hand
  • One row per article and week (week_start is a Monday). Every article has every week; units 0 = no sale.
  • units: units sold that week, in the shop's own selling unit (a pack, a bottle, a piece).
  • on_hand: stock at the end of the week as the shop's stock ledger recorded it, carried forward through weeks without a movement; empty before the article's first ledger entry. The ledger does not add up (opening stock plus receipts minus issues matches the recorded stock on only about 56% of rows) and some values are negative, so treat it as a noisy signal.
  • Dates are moved by a fixed number of whole weeks (weekdays kept); article ids are random; names and categories are removed.
  • To forecast: the 13 Mondays after the last history week (2023-04-24 to 2023-07-17), the same for every article. Upload format and scoring: the Submit tab, https://lemebase.com/hackathon/submit.

Licence: CC BY 4.0. Changed by LemeBase: aggregated to weeks, re-keyed, dates moved, columns dropped.

Shared column names

Column Meaning
date day
week_start Monday of the week
timestamp time of the event
store_id store
article_id article (SKU)
units units sold (negative = return where is_return is set)
on_hand units in stock
in_stock article was in stock that week
division product hierarchy level 1 (code)
category product hierarchy level 2 (code)
subcategory product hierarchy level 3 (code)
segment product hierarchy level 4 (code)
unit_price selling price per unit
unit_cost cost per unit
sales_amount sales value
cogs cost of goods sold
supplier_id supplier (pseudonymous id)
stock_status price status of the stock (code)
sales_type sales type code, e.g. FULL_PRICE, PROMO, MARKDOWN_TIER_1, CLEARANCE (codes vary by set)

Other columns keep their meaning from the source and a plain snake_case name. Text values are codes (e.g. DIV-1, CAT-11, SCAT-011); within a set the same value always has the same code, and each code sits under the same parent as in the source.


Sources and licences