Prompt
How do I source a dataset for demand forecasting with historical time series data?
Latest observation
To source a dataset for demand forecasting with historical time series data, you have a few good paths depending on whether you want to practice, prototype, or build a production model.
1) Use public datasets
If you need a ready-made dataset, look for datasets with:
- a timestamp
- a target variable like sales/demand/units sold
- ideally multiple products, locations, or stores
- enough history to capture seasonality
Good places to search:
- Kaggle
- UCI Machine Learning Repository
- Google Dataset Search
- GitHub repositories
- Open data portals from governments or retailers
Common public demand forecasting datasets:
- M5 Forecasting – Accuracy (Walmart sales by product/store)
- Rossmann Store Sales
- Corporación Favorita Grocery Sales
- Store Item Demand Forecasting Challenge
- Instacart Market Basket (more for purchase behavior than direct demand)
Search terms to use:
- “sales forecasting dataset”
- “retail demand time series dataset”
- “historical demand by SKU store”
- “promotional sales time series”
2) Build one from internal business data
If this is for a real use case, the best source is often your own operational data:
- ERP / inventory system
- POS (point-of-sale) transactions
- order management system
- e-commerce sales logs
- warehouse shipments
- replenishment / stock movement data
Typical fields you want:
- date or datetime
- product ID / SKU
- store / region / channel
- units sold or demand quantity
- price
- promotions / discounts
- stockouts / inventory availability
- holidays / events / weather if relevant
Important: make sure “demand” is not just “sales” if stockouts happened, because sales may understate true demand.
3) Buy or license data
If you need industry-specific or higher-quality data:
- data vendors
- market research firms
- retail analytics providers
- industry consortia
- cloud marketplace datasets
This is useful if you need:
- more granular geographic data
- external variables
- benchmark data
- less cleaning effort
4) Create synthetic or simulated data
If you’re learning or testing pipelines, you can generate synthetic demand series with:
- trend
- seasonality
- promotions
- random shocks
- intermittent demand
This is good for:
- architecture testing
- model prototyping
- privacy-safe demos
5) What makes a dataset good for demand forecasting?
Check for:
- time granularity: daily, weekly, monthly
- history length: enough to capture seasonality
- missing data: gaps should be understood
- stockouts: identify censored demand
- external drivers: price, promo, holidays, weather
- multiple series: many SKUs/stores usually improve modeling options
6) Quick checklist before using a dataset
Ask:
- What exactly is the target: sales, orders, consumption, or demand?
- Is the data complete and time-ordered?
- Are there censored periods from stockouts?
- Is the forecast horizon clear?
- Are there covariates available for future periods?
7) If you want, I can help you find one
If you tell me:
- your domain (retail, manufacturing, energy, etc.)
- preferred granularity (daily/weekly/monthly)
- whether you want public or real-world data
- and your tool stack (Python, SQL, etc.)
…I can recommend specific datasets and even give you a download + preprocessing workflow.