Get the data

Set A

Weekly sales with in-stock flags. When an article is out of stock, sales show only what was available. CC BY 4.0.

Set B

Daily transactions. Fully synthetic, and under a year of history. CC BY 4.0.

Set C

Daily sales and stock per store × article. 96.0% of store × article days have no sales. CC BY 4.0.

Set D

Daily sales, stock recorded by date range. 99.3% of store × article days have no sales, and under a year of history. CC BY 4.0.

Download all five (zip, 23 MB): Set A to Set D and the test shop history, as Parquet.

Or: huggingface-cli download LemeBase/hackathon-retail-sets --repo-type dataset

How the baselines work

AutoARIMA

  1. Each article's own weekly sales
  2. One small model per article (no season)
  3. 13 weeks per article

Top-down

  1. All articles added up: one weekly total
  2. One model of the total: trend + season (52 weeks with two years of history, else 4)
  3. 13 weeks of total
  4. Split to articles by their share of the last 13 weeks

Both: an article with no sales in the last 8 weeks forecasts 0.

Set A

Weekly sales

Download
13-week holdout 1,500 2,000 2,500 3,000 units / week 15 Jan 2024 12 Feb 2024 11 Mar 2024 8 Apr 2024 6 May 2024 3 Jun 2024 Naive (same as last week) Top-down (MSTL) AutoARIMA Actual

Actual AutoARIMA Naive (same as last week) Top-down (MSTL)

Set B

Daily sales

Download
13-week holdout 18,000 19,000 20,000 21,000 units / week 2 Jun 2025 30 Jun 2025 28 Jul 2025 25 Aug 2025 22 Sep 2025 20 Oct 2025 Naive (same as last week) Top-down (MSTL) AutoARIMA Actual

Actual AutoARIMA Naive (same as last week) Top-down (MSTL)

Set C

Daily sales

Download
13-week holdout 100 200 300 400 units / week 24 Nov 2025 22 Dec 2025 19 Jan 2026 16 Feb 2026 16 Mar 2026 13 Apr 2026 Naive (same as last week) Top-down (MSTL) AutoARIMA Actual

Actual AutoARIMA Naive (same as last week) Top-down (MSTL)

Set D

Daily sales

Download
13-week holdout 2,000 4,000 6,000 8,000 units / week 24 Nov 2025 22 Dec 2025 19 Jan 2026 16 Feb 2026 16 Mar 2026 13 Apr 2026 Naive (same as last week) Top-down (MSTL) AutoARIMA Actual

Actual AutoARIMA Naive (same as last week) Top-down (MSTL)