FRESH RETAIL · DEMAND BENCHMARK

FreshRetailNet-LT Leaderboard

FreshRetailNet-LT is a fresh retail dataset with hourly stockout annotations for studying latent demand recovery and demand forecasting.

Compare methods on two tasks: recovering demand under stockouts and forecasting the next seven days. Forecasting methods may use raw sales, a separate recovery model, or an integrated end-to-end approach.

Hourly stockout annotationsDemand recovery7-day forecasting

Reviewed community leaderboard

Loading…

Community forecasting methods, independently evaluated and approved by the maintainers.

RankModelRecoveryTeamWAPE ↓WPE ≈ 0MAE ↓MSE ↓Scored records

Download forecasting data

Download historical data for training and context, and the labeled evaluation set for the July 14–20, 2025 forecasting task.

Parquet format. The evaluation file includes ground-truth labels for local evaluation. Historical context is provided separately in the training file.

Forecasting input rules and prediction format
  • Forecast dates: July 14–20, 2025. Use historical observations dated before July 14, 2025 as context.
  • Feature usage: All provided features may be used except evaluation-period sale_amount, hours_sale, hours_stock_status, and stock_hour6_22_cnt. These answers and their derivatives must not be used as model inputs, training or tuning targets. Historical sales and stockout annotations before July 14, 2025 may be used. Supplied evaluation weather observations are allowed.
  • Prediction format: Prepare one row per store, product, and forecast date. Use the same normalized sales scale as the dataset. The four columns are store_id, product_id, dt, and prediction.
  • Local scoring: The browser pre-evaluator and maintainer review script require complete target coverage, then score records satisfying psd >= 0 and stock_hour6_22_cnt == 0, and report WAPE, WPE, MAE, and MSE. Here psd is each store-product’s mean raw daily sales from June 2 through July 13, 2025; it is precomputed in the public scoring reference. Submit predictions for all target records; the evaluator applies the scoring mask.
  • Public benchmark: Evaluation labels are publicly available. This is not a hidden-label test. Use browser pre-evaluation below, then send your files through the external submission channel for manual review. Browser checks do not upload or save submissions.

Check your forecasting predictions

Select a CSV to calculate scores in your browser. Your predictions stay on your device during this check; selecting a file does not submit it.

Loading evaluation reference…

Use historical sales directly, without a separate demand recovery step.

Columns: store_id,product_id,dt,prediction. Include all evaluation records; predictions must be finite and non-negative, on the dataset’s normalized sales scale.



Submit for review

Please send your prediction CSV, downloaded JSON report, and a short method description to:

Submission email: Loading…

Include your team name, model, recovery method (if applicable), training data, and feature usage. Results will be published after independent evaluation and manual review.

Evaluation protocol & metrics
Paper & citation

FreshRetailNet-LT: A Stockout-Annotated Censored Demand Dataset for Latent Demand Recovery and Forecasting in Fresh Retail.

Initial baseline results: manuscript Tables II–III. Community results: manually reviewed submissions.