# Retail demand forecasting - sktime use case

Model demand uncertainty across every product, store, and region with sktime, using probabilistic and distributional forecasts for retail planning.

Canonical page: https://www.sktime.net/use-cases/retail-forecasting/

## Plan retail demand with ranges.

sktime models demand across products, stores, and regions with probabilistic forecasts, so teams can see risk and keep hierarchy totals coherent.

## Trust Signals

- Open source: Free to run, fully auditable, and no per-seat or per-SKU license.
- Proven methods: The same approaches used on the M5, Walmart, and Rossmann forecasting problems.
- Built by the maintainers: You talk to the people who actually build sktime, not a reseller.

## Forecasts that don't agree leak money at every level.

- Cash frozen on the shelf: Buffer every series to feel safe and capital sits as inventory you have already paid for, then gets marked down or written off.
- Demand you never observe: Run lean and empty shelves censor demand: observed sales understate what shoppers wanted, and the lost basket never appears in the data.
- Planner days lost to reconciling: Teams burn hours forcing bottom-up and top-down numbers to meet in spreadsheets, and they still rarely agree.

## Uncertainty-aware demand forecasts pay for themselves.

When you know the full demand distribution, you size inventory to real risk instead of a single point estimate. Most of the savings land in stock decisions, the rest in the hours and licenses you stop spending.

- Less safety stock: Probabilistic forecasts size buffers to real risk, not gut feel. (↓)
- Fewer markdowns & waste: Order closer to true demand, so less stock ends up discounted or dumped. (↓)
- Recovered stockout demand: Catch the demand you used to miss when shelves ran empty. (↑)
- Forecasting license fees: sktime is open source. No per-seat or per-SKU vendor bill. (€0)
- Manual reconciliation hours: Hierarchy-aware workflows keep totals coherent when planning needs them. (freed)

Directional benefits only. Your real numbers come from a short scoping call on your own data.

## Built for real retail planning.

- Full probabilistic outputs: Intervals, quantiles, variance, and predictive distributions for risk-aware planning.
- Global & panel models: One model trained across many related series, instead of one model each.
- Real-world drivers: Add holidays, promotions, prices, store IDs, and SNAP purchase-day flags from M5/Walmart: dates when SNAP purchases are allowed in CA/TX/WI stores.
- Hierarchy when you need it: Bottom-up, top-down, and optimal reconciliation keep product, store, and regional totals coherent.

## Three steps from your data to a forecast you trust.

- Scoping call: We map your series, your hierarchy, and where the numbers hurt most today.
- Pilot on your data: We model a real slice of your demand data, quantify uncertainty, and keep hierarchy totals coherent where needed.
- Run in your planning: Roll it into your S&OP cadence, with the maintainers a message away.

## The three things planning leads check first.

### Do we need a data-science team?

No. We help stand it up on your data, and your planners work with the forecasts, not the code.

### Is open source safe for enterprise?

sktime is permissively licensed, auditable, and widely used in production. Nothing is locked behind a vendor.

### How fast can we see value?

A short pilot on a slice of your demand data shows the lift before you commit to anything.

## Bring uncertainty-aware forecasts into your planning.

sktime is free and open-source. When you want this running on your own demand data, the people behind it can help.