Prompt
Help me find a dataset for demand forecasting with enough history to backtest properly
Latest observation
Oct 16, 2026
OpenAI APIWeb search: off
Here are good dataset options for demand forecasting with enough history to do real backtesting. I’ll group them by what they’re best for.
Best public datasets with long history
1) M5 Forecasting (Walmart)
- What it is: Daily unit sales for thousands of products across Walmart stores
- History: ~5.5 years of daily data
- Granularity: Product × store × day
- Why it’s good: Very popular for forecasting benchmarks; rich enough for proper rolling-origin backtests
- Where: Kaggle — “M5 Forecasting - Accuracy”
2) M4 / M3 datasets
- What it is: Large collection of time series from many domains
- History: Varies; some series are long enough for backtesting
- Granularity: Mixed
- Why it’s good: Great for benchmarking forecasting methods, though not pure “demand” data
- Where: M4 competition data is available via forecasting repositories / Kaggle mirrors
3) Rossmann Store Sales
- What it is: Daily sales for pharmacy stores
- History: About 2 years
- Granularity: Store × day
- Why it’s good: Good for retail demand forecasting and feature-rich modeling
- Limitation: History is shorter than M5, but still usable for backtesting if you use smaller validation windows
- Where: Kaggle
4) Corporación Favorita Grocery Sales
- What it is: Daily unit sales for grocery items
- History: Several years
- Granularity: Store × item × day
- Why it’s good: Very large, realistic retail demand dataset
- Where: Kaggle
5) Dominick’s Finer Foods / Dunnhumby Retail data
- What it is: Grocery/retail purchase histories
- History: Multi-year
- Granularity: Household/item or store/item depending on version
- Why it’s good: Strong for demand and promotion-sensitive forecasting
- Where: Often available via research repositories or packaged academic releases
If you need very long time series
6) Retailer scanner datasets / academic sales panels
Examples:
- Dunnhumby
- Dominick’s
- Favorita
- M5
These are usually the best for proper backtesting because they have:
- many time periods
- repeated seasonality
- intermittent promotions
- hierarchical structure
If you want easy access and simple setup
7) Kaggle retail datasets
Search terms:
- “sales forecasting”
- “retail demand”
- “store item demand”
- “grocery sales”
- “Walmart sales”
Some of these are smaller but good for prototyping.
What counts as “enough history” for backtesting?
For demand forecasting, a good rule of thumb is:
- At least 2 full seasonal cycles for the main seasonality
- weekly seasonality: 8–12 weeks minimum
- annual seasonality: ideally 2+ years
- For robust backtesting:
- 3–5 rolling windows
- horizon based on your use case, e.g. 7, 14, 28, or 56 days
So if your data is daily and seasonal yearly, aim for 2+ years, preferably 4+ years.
My recommendations
If you want the safest choices:
- M5 Forecasting — best all-around public benchmark for demand forecasting
- Corporación Favorita — excellent large-scale retail demand data
- Rossmann — simpler but useful
- Dunnhumby/Dominick’s — great if you can access them
If you tell me:
- your preferred frequency (daily/weekly/monthly)
- industry (retail, spare parts, energy, etc.)
- whether you need public only or can use proprietary/paid data
…I can narrow it down to the best 3 datasets and suggest a backtesting scheme.